[bugfix]: keep sdpa metadata masks 2D; lift to 4D only inside the torch-sdpa impl

SDPAMetadataBuilder.build() lifted 2D [batch, key_len] padding masks to 4D
[batch, 1, 1, key_len], but the metadata is cross-backend: HYWorld and
HunyuanVideo15 build SDPAMetadata while the layer selector picks FLASH_ATTN
when installed. flash_attn.py's equal-length branch padded
attn_mask.shape[1] assuming 2D, so a 4D mask produced a garbage-length pad
and an out-of-bounds gather (CUDA device-side assert, SIGABRT in the CI
transformer tests: test_hyworld crashed, test_matrixgame2/test_wanvideo
never ran).

Store the mask 2D in the builder (the convention every in-tree
producer/consumer assumes) and do the 2D -> [batch, 1, 1, key_len] lift
inside _normalize_attn_mask_for_sdpa, the torch-sdpa-only consumer, so
torch.sdpa still does not misread 2D as its [query_len, key_len]
broadcast. The int -> bool coercion is unchanged. Composes with the
consumer-side normalization in flash_attn.py (98116070): metadata carries
the shared 2D convention, and each impl reshapes for its own kernel.

Adds a CPU-only regression test for the contract and wires it into the CI
unit lane explicitly (fastvideo/tests/attention/ has no lane of its own).

.buildkite/performance-benchmarks/tests/wan-t2v-1.3b.json

@@ -1,9 +1,9 @@
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
   "config_schema_version": 2,
-  "workload_id": "wan-t2v-1.3b",
-  "variant_id": "canonical",
-  "benchmark_version": 1,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "description": "Wan2.1 T2V 1.3B inference performance",
   "model": {
     "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",

apps/performance_dashboard/frontend/src/App.tsx

@@ -1,6 +1,7 @@
 import { useEffect, useMemo, useState } from "react";
 
-import { fetchSummary, fetchTrends, refreshData, RunSource, SummaryResponse, TrendGroup, TrendPoint } from "./api";
+import { fetchSummary, fetchTrends, refreshData } from "./api";
+import type { CohortValue, RunSource, SummaryResponse, TrendGroup, TrendPoint } from "./api";
 
 const METRIC_KEYS = ["latency", "throughput", "memory", "text_encoder_time_s", "dit_time_s", "vae_decode_time_s"];
 const RUN_SOURCES: Array<{ value: "" | RunSource; label: string }> = [
@@ -109,6 +110,61 @@ function metricLabel(metricKey: string) {
   return METRIC_DEFINITIONS[metricKey]?.label ?? metricKey;
 }
 
+type CohortFields = {
+  model_id: string;
+  gpu_type: string;
+  workload_id: CohortValue;
+  variant_id: CohortValue;
+  benchmark_version: CohortValue;
+  recipe_fingerprint: CohortValue;
+  hardware_profile_id: CohortValue;
+  software_profile_id: CohortValue;
+};
+
+function cohortValue(value: CohortValue) {
+  if (value === null || value === undefined || value === "") {
+    return "legacy";
+  }
+  return String(value);
+}
+
+function shortCohortValue(value: CohortValue) {
+  const text = cohortValue(value);
+  if (text === "legacy" || text.length <= 14) {
+    return text;
+  }
+  return text.slice(0, 12);
+}
+
+function cohortKey(cohort: CohortFields) {
+  return [
+    cohort.model_id,
+    cohort.gpu_type,
+    cohortValue(cohort.workload_id),
+    cohortValue(cohort.variant_id),
+    cohortValue(cohort.benchmark_version),
+    cohortValue(cohort.recipe_fingerprint),
+    cohortValue(cohort.hardware_profile_id),
+    cohortValue(cohort.software_profile_id)
+  ].join("|");
+}
+
+function cohortTitle(cohort: CohortFields) {
+  const workload = cohortValue(cohort.workload_id);
+  const variant = cohortValue(cohort.variant_id);
+  const version = cohortValue(cohort.benchmark_version);
+  const versionLabel = version === "legacy" ? version : `v${version}`;
+  return `${workload} / ${variant} / ${versionLabel}`;
+}
+
+function cohortDetail(cohort: CohortFields) {
+  return [
+    `recipe ${shortCohortValue(cohort.recipe_fingerprint)}`,
+    shortCohortValue(cohort.hardware_profile_id),
+    shortCohortValue(cohort.software_profile_id)
+  ].join(" | ");
+}
+
 function formatMetricValue(metricKey: string, value: number | null | undefined, tooltip = false) {
   const definition = METRIC_DEFINITIONS[metricKey];
   if (!definition) {
@@ -171,7 +227,9 @@ function TrendChart({ group, metricKey }: { group: TrendGroup; metricKey: string
         top: `${(activePoint.y / height) * 100}%`
       }
     : undefined;
-  const ariaLabel = `${metricLabel(metricKey)} trend for ${group.model_id} on ${group.gpu_type}`;
+  const ariaLabel = `${metricLabel(metricKey)} trend for ${group.model_id} on ${group.gpu_type}, ${cohortTitle(
+    group
+  )}`;
 
   return (
     <div className="chart-shell">
@@ -419,7 +477,7 @@ export default function App() {
       <section className="panel">
         <div className="panel-header">
           <h2>Latest Status</h2>
-          <span>{latestRows.length} model/GPU groups</span>
+          <span>{latestRows.length} comparison cohorts</span>
         </div>
         {latestRows.length === 0 ? (
           <div className="empty">No records match the selected filters.</div>
@@ -432,6 +490,7 @@ export default function App() {
                   <th>Recomputed</th>
                   <th>Model</th>
                   <th>GPU</th>
+                  <th>Cohort</th>
                   <th>Commit</th>
                   <th>Source</th>
                   <th>Baseline</th>
@@ -446,7 +505,7 @@ export default function App() {
               </thead>
               <tbody>
                 {latestRows.map((row) => (
-                  <tr key={`${row.model_id}-${row.gpu_type}`}>
+                  <tr key={cohortKey(row)}>
                     <td>
                       <span className={`badge ${row.status}`}>{row.status}</span>
                     </td>
@@ -457,6 +516,12 @@ export default function App() {
                     </td>
                     <td>{row.model_id}</td>
                     <td>{row.gpu_type}</td>
+                    <td>
+                      <div className="cohort-cell">
+                        <strong>{cohortTitle(row)}</strong>
+                        <span>{cohortDetail(row)}</span>
+                      </div>
+                    </td>
                     <td>{shortSha(row.commit_sha)}</td>
                     <td>
                       <span className={`source-badge source-${row.run_source}`}>{runSourceLabel(row.run_source)}</span>
@@ -495,11 +560,13 @@ export default function App() {
           ) : (
             trends.map((group) =>
               METRIC_KEYS.map((metricKey) => (
-                <article className="trend-card" key={`${group.model_id}-${group.gpu_type}-${metricKey}`}>
+                <article className="trend-card" key={`${cohortKey(group)}-${metricKey}`}>
                   <div>
                     <h3>{metricLabel(metricKey)}</h3>
                     <p>
                       {group.model_id} | {group.gpu_type}
+                      <span>{cohortTitle(group)}</span>
+                      <span>{cohortDetail(group)}</span>
                     </p>
                   </div>
                   <TrendChart group={group} metricKey={metricKey} />

apps/performance_dashboard/frontend/src/api.ts

@@ -13,6 +13,17 @@ export type MetricValue = {
   precision: number;
 };
 
+export type CohortValue = string | number | null;
+
+export type ComparisonCohort = {
+  workload_id: CohortValue;
+  variant_id: CohortValue;
+  benchmark_version: CohortValue;
+  recipe_fingerprint: CohortValue;
+  hardware_profile_id: CohortValue;
+  software_profile_id: CohortValue;
+};
+
 export type SummaryRow = {
   model_id: string;
   gpu_type: string;
@@ -34,7 +45,7 @@ export type SummaryRow = {
   build_id: string;
   job_id: string;
   metrics: Record<string, MetricValue>;
-};
+} & ComparisonCohort;
 
 export type RunSource = "pr" | "local" | "scheduled_main" | "unknown";
 
@@ -68,13 +79,13 @@ export type TrendPoint = {
   build_id: string;
   job_id: string;
   metrics: Record<string, number | null>;
-};
+} & ComparisonCohort;
 
 export type TrendGroup = {
   model_id: string;
   gpu_type: string;
   points: TrendPoint[];
-};
+} & ComparisonCohort;
 
 export type TrendsResponse = {
   groups: TrendGroup[];

apps/performance_dashboard/frontend/src/styles.css

@@ -149,6 +149,11 @@ h3 {
   font-size: 0.82rem;
 }
 
+.trend-card p {
+  display: grid;
+  gap: 2px;
+}
+
 .stat strong {
   display: block;
   margin-top: 8px;
@@ -186,7 +191,7 @@ h3 {
 
 table {
   width: 100%;
-  min-width: 1120px;
+  min-width: 1260px;
   border-collapse: collapse;
 }
 
@@ -209,6 +214,25 @@ td {
   font-size: 0.9rem;
 }
 
+.cohort-cell {
+  display: grid;
+  gap: 2px;
+}
+
+.cohort-cell strong,
+.trend-card p span {
+  color: #1b2836;
+  font-size: 0.78rem;
+  font-weight: 700;
+}
+
+.cohort-cell span,
+.trend-card p span + span {
+  color: #607080;
+  font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", monospace;
+  font-size: 0.72rem;
+}
+
 .badge {
   display: inline-flex;
   align-items: center;

docker/Dockerfile

@@ -68,7 +68,7 @@ ARG FLASH_ATTN_WHEEL_RELEASE_ARM64=https://github.com/mjun0812/flash-attention-p
 # cutlass-4.4 `cute.core.ThrMma` API, which crashes on the cutlass-dsl 4.5 that
 # flashinfer/quack pull in. After the wheel install we overlay this cutlass-4.5-safe
 # upstream cute (flash-attn-4) so the image runs FA4 instead of the FA2 fallback.
-ARG FA4_CUTE_REF=940cd9680f3315f2f06b43ab5bea2c2cf2d96806
+ARG FA4_CUTE_REF=82d6441eec5d4dfec120153db2c0145ae855a083
 
 # Provided automatically by BuildKit/buildx (e.g. "amd64" / "arm64") and used to
 # select the prebuilt flash-attn wheel. Empty under a plain `docker build` without
@@ -161,7 +161,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
         uv pip install flash-attn==${FLASH_ATTN_VERSION} --no-build-isolation; \
     fi
 
-# Overlay the cutlass-4.5-safe upstream FA4 cute (FA4_CUTE_REF) over the
+# Overlay the CuTe-DSL-4.6-compatible upstream FA4 cute (FA4_CUTE_REF) over the
 # wheel/source one so the image runs FA4, not the FA2 fallback. This pulls the FA4
 # runtime stack (cutlass-dsl, quack-kernels, apache-tvm-ffi, torch-c-dlpack-ext) --
 # the same deps the [dreamverse] extra already installs in CI; the installed torch

docs/contributing/performance_benchmarks.md

@@ -201,9 +201,9 @@ configs and remain loadable. New or migrated configs should use
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
   "config_schema_version": 2,
-  "workload_id": "wan-t2v-1.3b",
-  "variant_id": "canonical",
-  "benchmark_version": 1
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2
 }
 ```
 
@@ -214,21 +214,25 @@ metadata that make the measured workload explicit:
 
 | Field | Purpose |
 |---|---|
-| `workload_id` | Stable benchmark family, such as `wan-t2v-1.3b`. |
-| `variant_id` | Intentional recipe family, such as `canonical`. |
+| `workload_id` | Stable benchmark family, such as `wan-t2v`. |
+| `variant_id` | Intentional recipe family, including model size and parallelism config, such as `1.3b-sp2`. |
 | `benchmark_version` | Version of the measurement protocol and comparison policy. |
 
 If a config declares `config_schema_version: 2`, loading fails clearly when any
 required v2 identity field is missing. If v2 identity or metadata fields are
 added without `config_schema_version: 2`, loading also fails so partial
 migrations do not silently run as v1 configs. Optional v2 metadata fields
-reserved for follow-up work, such as `recipe`, `metric_threshold_policy`, and
-`quality_metadata`, must be JSON objects when present.
+reserved for follow-up work, such as `metric_threshold_policy` and
+`quality_metadata`, must be JSON objects when present. (`recipe` is emitted
+by the harness and is not config-declarable.)
 
 Recipe fingerprinting, hardware/software profile IDs, exact-identity
-comparison, metric-specific threshold policy behavior, promoted baselines, and
-dashboard regrouping are separate follow-up changes. Until those land, rolling
-baseline comparison remains keyed by `(model_id, gpu_type)`.
+comparison, and dashboard cohort grouping land with this change: v2 records
+compare only within their identity cohort, and a record that opens a NEW
+cohort is marked `baseline_status: "initialized_new_cohort"` (regression
+gating starts once that cohort accumulates history). Legacy v1 configs keep
+the `(model_id, gpu_type)` rolling-baseline comparison. Metric-specific
+threshold policies and promoted baselines remain separate follow-ups.
 
 ### Raw record (`results/perf_*.json`)
 
@@ -238,9 +242,9 @@ Written by `test_inference_performance.py`. One file per benchmark run.
 {
   "benchmark_id": "wan-t2v-1.3b-2gpu",
   "config_schema_version": 2,
-  "workload_id": "wan-t2v-1.3b",
-  "variant_id": "canonical",
-  "benchmark_version": 1,
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "model_short_name": "Wan2.1-T2V-1.3B-Diffusers",
   "device": "NVIDIA L40S",
   "num_gpus": 2,
@@ -270,7 +274,43 @@ Written by `test_inference_performance.py`. One file per benchmark run.
   "timestamp": "2026-05-08T22:00:00+00:00",
   "text_encoder_time_s": 2.141,
   "dit_time_s": 8.437,
-  "vae_decode_time_s": 3.208
+  "vae_decode_time_s": 3.208,
+  "recipe": {
+    "recipe_schema_version": 1,
+    "benchmark": {
+      "benchmark_id": "wan-t2v-1.3b-2gpu",
+      "workload_id": "wan-t2v",
+      "variant_id": "1.3b-sp2",
+      "benchmark_version": 2
+    },
+    "model": { "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" },
+    "init_kwargs": { "num_gpus": 2, "sp_size": 2, "tp_size": 1 },
+    "generation_kwargs": { "height": 480, "width": 832, "num_frames": 45 },
+    "inputs": { "prompt_count": 1, "prompt_sha256": ["<measured-prompt-sha256>"] },
+    "attention": { "requested_backend": "FLASH_ATTN", "resolved_backend": "FLASH_ATTN" }
+  },
+  "recipe_fingerprint": "<sha256>",
+  "hardware_profile": {
+    "device_type": "cuda",
+    "gpu_count": 2,
+    "gpus": [{ "name": "NVIDIA L40S", "memory_gb": 48, "compute_capability": "8.9" }],
+    "interconnect": "none_or_partial"
+  },
+  "hardware_profile_id": "hw-<sha256-prefix>",
+  "software_profile": {
+    "python": "3.12",
+    "pytorch": "2.12",
+    "cuda": "13.0",
+    "packages": {
+      "fastvideo_kernel": "0.3.2",
+      "flashinfer": "0.2.11",
+      "nvidia_cutlass_dsl": "4.5.0",
+      "triton": "3.4.1"
+    }
+  },
+  "software_profile_id": "sw-<sha256-prefix>",
+  "environment_metadata": { "env": { "IMAGE_VERSION": "py3.12-cuda13.0.0" } },
+  "environment_fingerprint": "env-<sha256-prefix>"
 }
 ```
 
@@ -282,6 +322,9 @@ result, used as the rolling-baseline source of truth.
 ```jsonc
 {
   "model_id": "wan-t2v-1.3b-2gpu",
+  "workload_id": "wan-t2v",
+  "variant_id": "1.3b-sp2",
+  "benchmark_version": 2,
   "timestamp": "2026-05-08T22:00:00+00:00",
   "commit_sha": "<full sha>",
   "gpu_type": "NVIDIA L40S",
@@ -298,6 +341,10 @@ result, used as the rolling-baseline source of truth.
       "gated": true
     }
   },
+  "recipe_fingerprint": "<sha256>",
+  "hardware_profile_id": "hw-<sha256-prefix>",
+  "software_profile_id": "sw-<sha256-prefix>",
+  "environment_fingerprint": "env-<sha256-prefix>",
   "success": true
 }
 ```
@@ -310,6 +357,18 @@ dashboard lists skipped plots for metric series that have no non-null values.
 Records missing both `run_source` and `baseline_eligible` are treated as legacy
 successful main/full-suite uploads and remain eligible for rolling baselines.
 
+New records compare only against the same `model_id`, `gpu_type`,
+`workload_id`, `variant_id`, `benchmark_version`, `recipe_fingerprint`,
+`hardware_profile_id`, and `software_profile_id` cohort.
+`environment_metadata` and `environment_fingerprint` are audit data and are not
+part of the comparison key.
+The recipe prompt digests describe the prompts actually measured by the
+benchmark run; extra configured prompts are ignored unless the benchmark runner
+executes them.
+Software profile package cohorts keep exact versions for relevant
+attention/kernel packages, including FastVideo kernels, FlashAttention,
+FlashInfer, Cutlass DSL, SageAttention, Triton, and xFormers when installed.
+
 ## Environment variable reference
 
 | Variable | Default | Used by | Purpose |
@@ -364,7 +423,7 @@ When the rolling-baseline phase runs, it emits:
      "benchmark_id": "<unique-id>",
      "config_schema_version": 2,
      "workload_id": "<stable-workload-id>",
-     "variant_id": "canonical",
+     "variant_id": "<variant, e.g. 1.3b-sp2>",
      "benchmark_version": 1,
      "model": { "model_path": "...", "model_short_name": "..." },
      "init_kwargs": { "num_gpus": 1, ... },
@@ -391,7 +450,9 @@ When the rolling-baseline phase runs, it emits:
 
    Legacy v1 configs without `config_schema_version` still load, but should not
    gain v2 identity or metadata fields until they are migrated to
-   `config_schema_version: 2`.
+   `config_schema_version: 2`. For v2 configs, `workload_id`, `variant_id`,
+   and `benchmark_version` are part of the comparison key; benchmark runs
+   fail if any of these identity fields are missing.
 
 2. The pytest test auto-discovers all configs — no test code needed. CI
    picks it up on the next `/test performance` run.

docs/design/inference_schema_parity_inventory.yaml

@@ -191,6 +191,9 @@ surfaces:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig
           - fastvideo.configs.pipelines.gen3c.Gen3CInferenceConfig
+      color_correction_strength:
+        sources:
+          - fastvideo.configs.pipelines.dreamx_world.DreamXWorld5BARPipelineConfig
       default_camera_rotation:
         sources:
           - fastvideo.configs.pipelines.gen3c.Gen3CConfig

docs/inference/support_matrix.md

@@ -58,6 +58,8 @@ pipeline initialization and sampling.
 | FastWan2.1 T2V 1.3B | `FastVideo/FastWan2.1-T2V-1.3B-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
 | FastWan2.2 TI2V 5B Full Attn* | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | 720P | ⭕ | ⭕ | ⭕ | ✅ | ⭕ |
 | Wan2.2 TI2V 5B | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 720P | ⭕ | ⭕ | ✅ | ⭕ | ⭕ |
+| DreamX-World 5B Cam | `FastVideo/DreamX-World-5B-Cam-Diffusers` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
+| DreamX-World 5B AR | `FastVideo/DreamX-World-5B-Diffusers` | 704px1280p | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Lucy Edit Dev 5B*** | `decart-ai/Lucy-Edit-Dev` | 480P | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ |
 | Wan2.2 T2V A14B | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |
 | Wan2.2 I2V A14B | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | 480P<br>720P | ❌ | ❌ | ✅ | ⭕ | ⭕ |

docs/training/trackers.md

@@ -0,0 +1,91 @@
+# Training Trackers
+
+FastVideo can send training metrics and validation media to Weights & Biases
+or SwanLab. Tracking runs only on global rank 0, and local tracker files are
+stored under `<output_dir>/tracker`.
+
+## Supported Trackers
+
+| Value | Backend | Installation |
+|-------|---------|--------------|
+| `wandb` | Weights & Biases | Included with FastVideo |
+| `swanlab` | SwanLab | Install the optional `swanlab` dependency |
+| `none` | Disable external tracking | No additional package |
+
+You can enable more than one backend, for example `trackers: [wandb, swanlab]`.
+Metrics and validation media are converted to the artifact type required by
+each backend.
+
+## Install SwanLab
+
+For a published FastVideo installation, install the SwanLab extra:
+
+```bash
+uv pip install "fastvideo[swanlab]"
+```
+
+For an editable source checkout, include the same extra during installation:
+
+```bash
+uv pip install -e ".[swanlab]"
+```
+
+If FastVideo is already installed, you can install the compatible SDK directly:
+
+```bash
+uv pip install "swanlab>=0.6.7"
+```
+
+Authenticate once before starting a training run:
+
+```bash
+swanlab login
+```
+
+See the [SwanLab login documentation](https://docs.swanlab.cn/en/api/cli-swanlab-login.html)
+for non-interactive and self-hosted setups.
+
+## Configure Tracking
+
+Select SwanLab in the YAML config used by the modular training framework:
+
+```yaml
+training:
+  checkpoint:
+    output_dir: outputs/my_run
+  tracker:
+    trackers: [swanlab]
+    project_name: my_project
+    run_name: my_run
+```
+
+To log to both supported services:
+
+```yaml
+training:
+  tracker:
+    trackers: [wandb, swanlab]
+    project_name: my_project
+    run_name: my_run
+```
+
+An empty or omitted `trackers` list selects W&B when `project_name` is set.
+Use an explicit `none` entry to disable external tracking:
+
+```yaml
+training:
+  tracker:
+    trackers: [none]
+```
+
+## Validation Videos
+
+SwanLab currently accepts GIF video artifacts. FastVideo converts validation
+MP4 files and in-memory video arrays to GIF automatically before logging them.
+For video files, FastVideo uses the sampling frame rate supplied by the caller,
+or the source file's frame rate when no value is supplied. In-memory arrays use
+the frame rate supplied by the caller. Both forms fall back to 16 FPS when no
+frame rate is available.
+
+For details about configuring validation callbacks, see
+[Training Infrastructure](train_infra.md#callbacks-pluggable-hooks).

docs/training/train_infra.md

@@ -161,6 +161,9 @@ training:
     decay_interval_steps: 0
 ```
 
+See [Training Trackers](trackers.md) to configure Weights & Biases or SwanLab,
+including SwanLab installation and authentication.
+
 ### `callbacks` — Pluggable hooks
 
 Callbacks run at specific points in the training loop (before/after optimizer

examples/inference/basic/basic_dreamx_world.py

@@ -0,0 +1,64 @@
+import os
+
+from fastvideo import VideoGenerator
+
+
+OUTPUT_PATH = os.getenv("DREAMX_WORLD_OUTPUT_PATH", "video_samples_dreamx_world")
+
+
+def _env_int(name: str, default: int) -> int:
+    return int(os.getenv(name, str(default)))
+
+
+def _env_float(name: str, default: float) -> float:
+    return float(os.getenv(name, str(default)))
+
+
+def main():
+    model_name = os.getenv("DREAMX_WORLD_MODEL_DIR", "FastVideo/DreamX-World-5B-Cam-Diffusers")
+    generator = VideoGenerator.from_pretrained(
+        model_name,
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=True,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=False,
+        override_pipeline_cls_name="DreamXWorldPipeline",
+    )
+
+    prompt = os.getenv(
+        "DREAMX_WORLD_PROMPT",
+        "A cinematic first-person drive through a futuristic coastal city at "
+        "sunrise, reflective glass towers, clean streets, soft volumetric light.",
+    )
+    image_path = os.getenv(
+        "DREAMX_WORLD_IMAGE_PATH",
+        "https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG",
+    )
+
+    kwargs = {
+        "output_path": OUTPUT_PATH,
+        "save_video": os.getenv("DREAMX_WORLD_SAVE_VIDEO", "1") != "0",
+        "height": _env_int("DREAMX_WORLD_HEIGHT", 480),
+        "width": _env_int("DREAMX_WORLD_WIDTH", 832),
+        "num_frames": _env_int("DREAMX_WORLD_NUM_FRAMES", 161),
+        "num_inference_steps": _env_int("DREAMX_WORLD_STEPS", 30),
+        "guidance_scale": _env_float("DREAMX_WORLD_GUIDANCE", 5.0),
+        "action_list": os.getenv("DREAMX_WORLD_ACTIONS", "w,d,w").split(","),
+        "action_speed_list": [
+            float(value)
+            for value in os.getenv("DREAMX_WORLD_ACTION_SPEEDS", "4.0,2.0,4.0").split(",")
+        ],
+    }
+    if image_path:
+        kwargs["image_path"] = image_path
+
+    try:
+        generator.generate_video(prompt, **kwargs)
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_kandinsky5_i2v.py

@@ -0,0 +1,37 @@
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_kandinsky5_i2v"
+
+IMAGE_PATH = "assets/girl.png"
+
+
+def main():
+    generator = VideoGenerator.from_pretrained(
+        "kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers", 
+        # "kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers"
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+        # image_encoder_cpu_offload=False,
+    )
+
+    prompt = (
+        "A woman stands up and walks away"
+    )
+    _ = generator.generate_video(
+        prompt,
+        image_path=IMAGE_PATH,
+        output_path=OUTPUT_PATH,
+        save_video=True,
+        height=1024,
+        width=1024,
+        num_frames=121,
+    )
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/basic_kandinsky5_t2v.py

@@ -0,0 +1,37 @@
+from fastvideo import VideoGenerator
+
+OUTPUT_PATH = "video_samples_kandinsky5_t2v"
+
+def main():
+    generator = VideoGenerator.from_pretrained(
+        "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
+        # "kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers"
+        # "kandinskylab/Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers"
+        num_gpus=1,
+        use_fsdp_inference=False,
+        dit_cpu_offload=False,
+        vae_cpu_offload=False,
+        text_encoder_cpu_offload=True,
+        pin_cpu_memory=True,
+        # image_encoder_cpu_offload=False,
+    )
+
+    prompt = (
+        "A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes "
+        "wide with interest. The playful yet serene atmosphere is complemented by soft "
+        "natural light filtering through the petals. Mid-shot, warm and cheerful tones."
+    )
+    _ = generator.generate_video(prompt, output_path=OUTPUT_PATH, save_video=True,height=512, width=768, num_frames=121)
+
+    prompt2 = (
+        "A majestic lion strides across the golden savanna, its powerful frame "
+        "glistening under the warm afternoon sun. The tall grass ripples gently in "
+        "the breeze, enhancing the lion's commanding presence. The tone is vibrant, "
+        "embodying the raw energy of the wild. Low angle, steady tracking shot, "
+        "cinematic.")
+    _ = generator.generate_video(prompt2, output_path=OUTPUT_PATH, save_video=True, height=512, width=768, num_frames=121)
+
+
+if __name__ == "__main__":
+    main()

fastvideo/attention/backends/flash_attn.py

@@ -338,7 +338,11 @@ class FlashAttentionImpl(AttentionImpl):
                 )
 
             qkv = torch.stack([query, key, value], dim=2)
-            attn_mask_padded = F.pad(attn_mask, (qkv.shape[1] - attn_mask.shape[1], 0), value=True)
+            key_padding_mask = _key_padding_mask_from_attn_mask(attn_mask, attn_mask.shape[-1]).to(device=query.device)
+            if key_padding_mask.shape[-1] > qkv.shape[1]:
+                raise ValueError("Invalid key padding mask length for FLASH_ATTN: "
+                                 f"expected at most {qkv.shape[1]}, got {key_padding_mask.shape[-1]}")
+            attn_mask_padded = F.pad(key_padding_mask, (qkv.shape[1] - key_padding_mask.shape[-1], 0), value=True)
             output = flash_attn_no_pad(qkv, attn_mask_padded, causal=False, dropout_p=0, softmax_scale=None)
         elif self.nvfp4_fa4:
             output = self._forward_nvfp4(query, key, value)

fastvideo/attention/backends/nabla.py

@@ -0,0 +1,147 @@
+# SPDX-License-Identifier: Apache-2.0
+"""NABLA block-sparse flex-attention backend (Kandinsky5 "nabla" checkpoints).
+
+The block mask is data-dependent: nablaT_v2 mean-pools 64-token blocks of the
+fractal-ordered sequence, thresholds the softmaxed block map, and ORs it with a
+precomputed spatio-temporal-window (STA) mask carried on the attention
+metadata. The mask spans the full sequence, so this backend does not support
+sequence parallelism — use it via LocalAttention only.
+"""
+
+import math
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+
+try:
+    from torch.nn.attention.flex_attention import BlockMask, flex_attention
+    flex_attention = torch.compile(flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs")
+    CAN_USE_FLEX_ATTN = True
+except ImportError:
+    CAN_USE_FLEX_ATTN = False
+
+from fastvideo.attention.backends.abstract import (
+    AttentionBackend,
+    AttentionImpl,
+    AttentionMetadata,
+    AttentionMetadataBuilder,
+)
+
+
+def nablaT_v2(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    sta: torch.Tensor,
+    thr: float = 0.9,
+) -> "BlockMask":
+    q = q.transpose(1, 2).contiguous()
+    k = k.transpose(1, 2).contiguous()
+
+    # Map estimation
+    B, h, S, D = q.shape
+    s1 = S // 64
+    qa = q.reshape(B, h, s1, 64, D).mean(-2)
+    ka = k.reshape(B, h, s1, 64, D).mean(-2).transpose(-2, -1)
+    map = qa @ ka
+
+    map = torch.softmax(map / math.sqrt(D), dim=-1)
+    # Map binarization
+    vals, inds = map.sort(-1)
+    cvals = vals.cumsum_(-1)
+    mask = (cvals >= 1 - thr).int()
+    mask = mask.gather(-1, inds.argsort(-1))
+
+    mask = torch.logical_or(mask, sta)
+
+    # BlockMask creation
+    kv_nb = mask.sum(-1).to(torch.int32)
+    kv_inds = mask.argsort(dim=-1, descending=True).to(torch.int32)
+    return BlockMask.from_kv_blocks(torch.zeros_like(kv_nb), kv_inds, kv_nb, kv_inds, BLOCK_SIZE=64, mask_mod=None)
+
+
+class NablaAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "NABLA_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["NablaAttentionImpl"]:
+        return NablaAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["NablaAttentionMetadata"]:
+        return NablaAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["NablaAttentionMetadataBuilder"]:
+        return NablaAttentionMetadataBuilder
+
+
+@dataclass
+class NablaAttentionMetadata(AttentionMetadata):
+    # Block-level STA window mask [1, 1, S/64, S/64], precomputed once per run.
+    sta_mask: torch.Tensor = None  # type: ignore[assignment]
+    # Cumulative-probability threshold for block-map binarization.
+    P: float = 0.9
+    visual_shape: tuple[int, int, int] = (0, 0, 0)
+
+
+class NablaAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self) -> None:
+        pass
+
+    def prepare(self) -> None:
+        pass
+
+    def build(
+        self,
+        current_timestep: int,
+        sta_mask: torch.Tensor,
+        P: float,
+        visual_shape: tuple[int, int, int],
+        **kwargs: Any,
+    ) -> NablaAttentionMetadata:
+        return NablaAttentionMetadata(
+            current_timestep=current_timestep,
+            sta_mask=sta_mask,
+            P=P,
+            visual_shape=visual_shape,
+        )
+
+
+class NablaAttentionImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        softmax_scale: float,
+        causal: bool = False,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        if not CAN_USE_FLEX_ATTN:
+            raise RuntimeError("NABLA attention requires torch.nn.attention.flex_attention, "
+                               "which is unavailable in this PyTorch build.")
+        if causal:
+            raise ValueError("NABLA attention does not support causal masking.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: NablaAttentionMetadata,
+    ) -> torch.Tensor:
+        # q/k/v: [B, S, heads, head_dim], fractal-ordered by the model; S % 64 == 0.
+        block_mask = nablaT_v2(query, key, attn_metadata.sta_mask, thr=attn_metadata.P)
+        return flex_attention(
+            query=query.transpose(1, 2),
+            key=key.transpose(1, 2),
+            value=value.transpose(1, 2),
+            block_mask=block_mask,
+        ).transpose(1, 2)

fastvideo/attention/backends/sdpa.py

@@ -2,6 +2,7 @@
 
 import torch
 from dataclasses import dataclass
+from torch.nn import functional as F
 from fastvideo.attention.backends.abstract import (  # FlashAttentionMetadata,
     AttentionBackend, AttentionImpl, AttentionMetadata, AttentionMetadataBuilder)
 from fastvideo.logger import init_logger
@@ -19,7 +20,7 @@ class SDPABackend(AttentionBackend):
 
     @staticmethod
     def get_name() -> str:
-        return "SDPA"
+        return "TORCH_SDPA"
 
     @staticmethod
     def get_impl_cls() -> type["SDPAImpl"]:
@@ -49,9 +50,51 @@ class SDPAMetadataBuilder(AttentionMetadataBuilder):
             current_timestep: int,
             attn_mask: torch.Tensor,
     ) -> SDPAMetadata:
+        # Store the mask exactly as passed. The metadata is cross-backend:
+        # call sites (HYWorld, HunyuanVideo15) build SDPAMetadata while the
+        # layer's selector may pick FLASH_ATTN, and the shared convention for
+        # padding masks is the tokenizer-style 2D [batch, key_len]. Any
+        # reshaping for torch.sdpa happens inside the SDPA impl
+        # (_normalize_attn_mask_for_sdpa).
         return SDPAMetadata(current_timestep=current_timestep, attn_mask=attn_mask)
 
 
+def _normalize_attn_mask_for_sdpa(
+    attn_mask: torch.Tensor | None,
+    query: torch.Tensor,
+    key: torch.Tensor,
+) -> torch.Tensor | None:
+    if attn_mask is None:
+        return None
+
+    attn_mask = attn_mask.to(device=query.device)
+    # F.scaled_dot_product_attention only accepts bool or float masks;
+    # tokenizers commonly produce int64 0/1 padding masks.
+    if attn_mask.dtype != torch.bool and not attn_mask.dtype.is_floating_point:
+        attn_mask = attn_mask != 0
+
+    key_len = key.shape[-2]
+    if attn_mask.shape[-1] > key_len:
+        raise ValueError("Invalid attention mask length for SDPA: "
+                         f"expected at most {key_len}, got {attn_mask.shape[-1]}")
+    if attn_mask.shape[-1] < key_len:
+        # Front-pad as "attend": double-stream layouts (HYWorld) prepend
+        # non-text tokens the tokenizer mask does not cover.
+        valid_value = True if attn_mask.dtype == torch.bool else 0.0
+        attn_mask = F.pad(attn_mask, (key_len - attn_mask.shape[-1], 0), value=valid_value)
+
+    if attn_mask.dim() == 2:
+        # In-tree producers pass 2D [batch, key_len] padding masks; lift to a
+        # broadcastable [batch, 1, 1, key_len] here so torch.sdpa does not
+        # reinterpret 2D as its documented [query_len, key_len] broadcast.
+        return attn_mask[:, None, None, :]
+    if attn_mask.dim() == 3:
+        return attn_mask[:, None, :, :]
+    if attn_mask.dim() == 4:
+        return attn_mask
+    raise ValueError(f"Unsupported attention mask shape for SDPA: {attn_mask.shape}")
+
+
 class SDPAImpl(AttentionImpl):
 
     def __init__(
@@ -82,6 +125,7 @@ class SDPAImpl(AttentionImpl):
 
         attn_mask = attn_metadata.attn_mask if (attn_metadata is not None
                                                 and hasattr(attn_metadata, "attn_mask")) else None
+        attn_mask = _normalize_attn_mask_for_sdpa(attn_mask, query, key)
         attn_kwargs = {
             "attn_mask": attn_mask,
             "dropout_p": self.dropout,

fastvideo/attention/layer.py

@@ -252,6 +252,7 @@ class LocalAttention(nn.Module):
                  causal: bool = False,
                  supported_attention_backends: tuple[AttentionBackendEnum, ...]
                  | None = None,
+                 default_backend: AttentionBackendEnum | None = None,
                  **extra_impl_args) -> None:
         super().__init__()
         if softmax_scale is None:
@@ -262,7 +263,10 @@ class LocalAttention(nn.Module):
             num_kv_heads = num_heads
 
         dtype = get_compute_dtype()
-        attn_backend = get_attn_backend(head_size, dtype, supported_attention_backends=supported_attention_backends)
+        attn_backend = get_attn_backend(head_size,
+                                        dtype,
+                                        supported_attention_backends=supported_attention_backends,
+                                        default_backend=default_backend)
         impl_cls = attn_backend.get_impl_cls()
         self.attn_impl = impl_cls(num_heads=num_heads,
                                   head_size=head_size,

fastvideo/attention/selector.py

@@ -84,8 +84,9 @@ def get_attn_backend(
     dtype: torch.dtype,
     supported_attention_backends: tuple[AttentionBackendEnum, ...]
     | None = None,
+    default_backend: AttentionBackendEnum | None = None,
 ) -> type[AttentionBackend]:
-    return _cached_get_attn_backend(head_size, dtype, supported_attention_backends)
+    return _cached_get_attn_backend(head_size, dtype, supported_attention_backends, default_backend)
 
 
 @cache
@@ -94,6 +95,7 @@ def _cached_get_attn_backend(
     dtype: torch.dtype,
     supported_attention_backends: tuple[AttentionBackendEnum, ...]
     | None = None,
+    default_backend: AttentionBackendEnum | None = None,
 ) -> type[AttentionBackend]:
     # Check whether a particular choice of backend was
     # previously forced.
@@ -112,6 +114,12 @@ def _cached_get_attn_backend(
         if backend_by_env_var is not None:
             selected_backend = backend_name_to_enum(backend_by_env_var)
 
+    # Layer-level default (e.g. a checkpoint that requires a specific sparse
+    # backend). Lower precedence than the global force and the env var, so
+    # users can still override it.
+    if selected_backend is None and default_backend is not None:
+        selected_backend = default_backend
+
     # get device-specific attn_backend
     from fastvideo.platforms import current_platform
 

fastvideo/configs/models/dits/__init__.py

@@ -1,5 +1,6 @@
 from fastvideo.configs.models.dits.cosmos import CosmosVideoConfig
 from fastvideo.configs.models.dits.cosmos2_5 import Cosmos25VideoConfig
+from fastvideo.configs.models.dits.dreamx_world import DreamXWorldARConfig, DreamXWorldConfig
 from fastvideo.configs.models.dits.flux_2 import Flux2Config
 from fastvideo.configs.models.dits.hunyuangamecraft import HunyuanGameCraftConfig
 from fastvideo.configs.models.dits.hunyuanvideo import HunyuanVideoConfig
@@ -13,7 +14,7 @@ from fastvideo.configs.models.dits.hyworld import HYWorldConfig
 from fastvideo.configs.models.dits.kandinsky5 import Kandinsky5VideoConfig
 
 __all__ = [
-    "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "CosmosVideoConfig",
-    "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig", "HYWorldConfig", "Kandinsky5VideoConfig",
-    "MagiHumanVideoConfig", "StableAudioConfig", "Flux2Config"
+    "HunyuanVideoConfig", "HunyuanVideo15Config", "HunyuanGameCraftConfig", "WanVideoConfig", "DreamXWorldConfig",
+    "DreamXWorldARConfig", "CosmosVideoConfig", "Cosmos25VideoConfig", "LongCatVideoConfig", "LTX2VideoConfig",
+    "HYWorldConfig", "Kandinsky5VideoConfig", "MagiHumanVideoConfig", "StableAudioConfig", "Flux2Config"
 ]

fastvideo/configs/models/dits/dreamx_world.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: Apache-2.0
+from dataclasses import dataclass, field
+
+from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+from fastvideo.configs.models.dits.wanvideo import WanVideoArchConfig
+
+
+@dataclass
+class DreamXWorldArchConfig(WanVideoArchConfig):
+    """DreamX-World DiT config with camera PRoPE control fields."""
+
+    add_control_adapter: bool = True
+    cam_method: str | None = "prope"
+    attn_compress: int = 1
+    cam_self_attn_layers: tuple[int, ...] | None = None
+
+
+@dataclass
+class DreamXWorldConfig(DiTConfig):
+    arch_config: DiTArchConfig = field(default_factory=DreamXWorldArchConfig)
+
+    prefix: str = "Wan"
+
+
+@dataclass
+class DreamXWorldARArchConfig(DreamXWorldArchConfig):
+    """DreamX-World-5B autoregressive causal DiT config."""
+
+    model_type: str = "ti2v"
+    patch_size: tuple[int, int, int] = (1, 2, 2)
+    text_len: int = 512
+    text_dim: int = 4096
+    freq_dim: int = 256
+    attn_compress: int = 4
+    cam_self_attn_layers: tuple[int, ...] | None = tuple(range(30))
+    local_attn_size: int = 12
+    sink_size: int = 3
+    num_frames_per_block: int = 3
+    rope_cache_policy: str = "block_relativistic"
+    # The official AR checkpoint (AMAP-ML/DreamX-World ``model.safetensors``)
+    # already uses FastVideo's native key names and the converter copies the
+    # tensors verbatim, so every rule is an identity. The rules enumerate the
+    # full state-dict surface of ``DreamXWorldARTransformer3DModel`` (norm1 /
+    # norm2 / head.norm are affine-free and have no parameters).
+    param_names_mapping: dict = field(
+        default_factory=lambda: {
+            r"^patch_embedding\.(.*)$": r"patch_embedding.\1",
+            r"^text_embedding\.([02])\.(.*)$": r"text_embedding.\1.\2",
+            r"^time_embedding\.([02])\.(.*)$": r"time_embedding.\1.\2",
+            r"^time_projection\.1\.(.*)$": r"time_projection.1.\1",
+            r"^blocks\.(\d+)\.self_attn\.(q|k|v|o)\.(.*)$": r"blocks.\1.self_attn.\2.\3",
+            r"^blocks\.(\d+)\.self_attn\.norm_(q|k)\.weight$": r"blocks.\1.self_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.cross_attn\.(q|k|v|o)\.(.*)$": r"blocks.\1.cross_attn.\2.\3",
+            r"^blocks\.(\d+)\.cross_attn\.norm_(q|k)\.weight$": r"blocks.\1.cross_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.cam_self_attn\.(q_proj|k_proj|v_proj|out_proj)\.(.*)$": r"blocks.\1.cam_self_attn.\2.\3",
+            r"^blocks\.(\d+)\.cam_self_attn\.norm_(q|k)\.weight$": r"blocks.\1.cam_self_attn.norm_\2.weight",
+            r"^blocks\.(\d+)\.norm3\.(.*)$": r"blocks.\1.norm3.\2",
+            r"^blocks\.(\d+)\.ffn\.([02])\.(.*)$": r"blocks.\1.ffn.\2.\3",
+            r"^blocks\.(\d+)\.modulation$": r"blocks.\1.modulation",
+            r"^head\.head\.(.*)$": r"head.head.\1",
+            r"^head\.modulation$": r"head.modulation",
+        })
+    reverse_param_names_mapping: dict = field(default_factory=dict)
+    lora_param_names_mapping: dict = field(default_factory=dict)
+
+
+@dataclass
+class DreamXWorldARConfig(DreamXWorldConfig):
+    arch_config: DiTArchConfig = field(default_factory=DreamXWorldARArchConfig)
+
+    prefix: str = "Wan"

fastvideo/configs/models/dits/kandinsky5.py

@@ -2,14 +2,24 @@
 from dataclasses import dataclass, field
 
 from fastvideo.configs.models.dits.base import DiTArchConfig, DiTConfig
+from fastvideo.platforms import AttentionBackendEnum
+
+
+def _is_kandinsky5_transformer_block(n: str, m) -> bool:
+    return ("text_transformer_blocks" in n or "visual_transformer_blocks" in n) and n.split(".")[-1].isdigit()
 
 
 @dataclass
 class Kandinsky5ArchConfig(DiTArchConfig):
-    _fsdp_shard_conditions: list = field(default_factory=lambda: [
-        lambda n, m:
-        ("text_transformer_blocks" in n or "visual_transformer_blocks" in n) and n.split(".")[-1].isdigit()
-    ])
+    _fsdp_shard_conditions: list = field(default_factory=lambda: [_is_kandinsky5_transformer_block])
+
+    # NABLA block-sparse attention for attention_type="nabla" checkpoints, plus
+    # the dense backends every DiT supports.
+    _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
+        AttentionBackendEnum.NABLA_ATTN,
+        AttentionBackendEnum.FLASH_ATTN,
+        AttentionBackendEnum.TORCH_SDPA,
+    )
 
     # Native FastVideo implementation uses the same parameter names as diffusers
     # except FFN internals: Diffusers FFN uses `in_layer/out_layer`, while

fastvideo/configs/models/encoders/base.py

@@ -43,11 +43,16 @@ class TextEncoderArchConfig(EncoderArchConfig):
     require_processor: bool = False
 
     def __post_init__(self) -> None:
-        self.tokenizer_kwargs = {
+        # update_model_arch re-runs __post_init__ after pipeline configs may
+        # have customized tokenizer_kwargs (e.g. kandinsky5/gen3c/longcat set
+        # "padding"); rebuilding the dict here would silently wipe those
+        # customizations, so only fill in defaults for keys not already set.
+        defaults = {
             "truncation": True,
             "max_length": self.text_len,
             "return_tensors": "pt",
         }
+        self.tokenizer_kwargs = defaults | self.tokenizer_kwargs
 
 
 @dataclass

fastvideo/configs/pipelines/__init__.py

@@ -1,10 +1,12 @@
 from fastvideo.configs.pipelines.base import PipelineConfig
 from fastvideo.configs.pipelines.cosmos import CosmosConfig
 from fastvideo.configs.pipelines.cosmos2_5 import Cosmos25Config
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BARPipelineConfig, DreamXWorld5BCamPipelineConfig
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.hunyuangamecraft import HunyuanGameCraftPipelineConfig
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
+from fastvideo.configs.pipelines.kandinsky5 import Kandinsky5I2VConfig, Kandinsky5T2VConfig
 from fastvideo.configs.pipelines.matrixgame2 import MatrixGame2I2V480PConfig
 from fastvideo.configs.pipelines.matrixgame3 import MatrixGame3I2V720PConfig
 from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
@@ -16,5 +18,6 @@ __all__ = [
     "HunyuanConfig", "FastHunyuanConfig", "HunyuanGameCraftPipelineConfig", "PipelineConfig", "Hunyuan15T2V480PConfig",
     "Hunyuan15T2V720PConfig", "WanT2V480PConfig", "WanI2V480PConfig", "WanT2V720PConfig", "WanI2V720PConfig",
     "SelfForcingWanT2V480PConfig", "LucyEditDevConfig", "CosmosConfig", "Cosmos25Config", "LTX2T2VConfig",
-    "HYWorldConfig", "MatrixGame2I2V480PConfig", "MatrixGame3I2V720PConfig", "get_pipeline_config_cls_from_name"
+    "DreamXWorld5BCamPipelineConfig", "DreamXWorld5BARPipelineConfig", "HYWorldConfig", "Kandinsky5T2VConfig",
+    "Kandinsky5I2VConfig", "MatrixGame2I2V480PConfig", "MatrixGame3I2V720PConfig", "get_pipeline_config_cls_from_name"
 ]

fastvideo/configs/pipelines/dreamx_world.py

@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World-5B-Cam FastVideo model configuration helpers."""
+
+from collections.abc import Callable
+from dataclasses import dataclass, field
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits.dreamx_world import (DreamXWorldARArchConfig, DreamXWorldARConfig,
+                                                        DreamXWorldArchConfig, DreamXWorldConfig)
+from fastvideo.configs.models.encoders import BaseEncoderOutput, T5Config
+from fastvideo.configs.models.encoders.t5 import T5ArchConfig
+from fastvideo.configs.models.vaes import WanVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig
+from fastvideo.configs.pipelines.wan import LucyEditDevConfig, t5_postprocess_text
+
+
+def make_dreamx_world_5b_cam_dit_config() -> DreamXWorldConfig:
+    """Return the DreamX-World DiT config matching DreamX-World-5B-Cam."""
+    return DreamXWorldConfig(arch_config=DreamXWorldArchConfig(
+        num_attention_heads=24,
+        attention_head_dim=128,
+        in_channels=48,
+        out_channels=48,
+        ffn_dim=14336,
+        num_layers=30,
+        cross_attn_norm=True,
+        qk_norm="rms_norm_across_heads",
+        add_control_adapter=True,
+        cam_method="prope",
+        attn_compress=1,
+        cam_self_attn_layers=None,
+    ))
+
+
+def make_dreamx_world_5b_ar_dit_config() -> DreamXWorldARConfig:
+    """Return the DreamX-World-5B autoregressive causal DiT config."""
+    return DreamXWorldARConfig(arch_config=DreamXWorldARArchConfig(
+        model_type="ti2v",
+        num_attention_heads=24,
+        attention_head_dim=128,
+        in_channels=48,
+        out_channels=48,
+        ffn_dim=14336,
+        num_layers=30,
+        cross_attn_norm=True,
+        qk_norm=True,
+        add_control_adapter=True,
+        cam_method="prope",
+        attn_compress=4,
+        cam_self_attn_layers=tuple(range(30)),
+        local_attn_size=12,
+        sink_size=3,
+        num_frames_per_block=3,
+    ))
+
+
+def make_dreamx_world_5b_cam_vae_config() -> WanVAEConfig:
+    """Return the Wan2.2 48-channel VAE config used by DreamX-World-5B-Cam."""
+    return LucyEditDevConfig().vae_config
+
+
+def make_dreamx_world_5b_cam_text_encoder_config() -> T5Config:
+    """Return the UMT5-XXL text encoder config used by DreamX-World-5B-Cam."""
+    return T5Config(
+        arch_config=T5ArchConfig(
+            vocab_size=256384,
+            d_model=4096,
+            d_kv=64,
+            d_ff=10240,
+            num_layers=24,
+            num_decoder_layers=None,
+            num_heads=64,
+            relative_attention_num_buckets=32,
+            dropout_rate=0.0,
+            text_len=512,
+            feed_forward_proj="gelu",
+            is_encoder_decoder=False,
+        ),
+        prefix="umt5",
+    )
+
+
+@dataclass
+class DreamXWorld5BCamPipelineConfig(PipelineConfig):
+    """Pipeline config for the first-scope DreamX-World-5B-Cam mode."""
+
+    dit_config: DiTConfig = field(default_factory=make_dreamx_world_5b_cam_dit_config)
+    vae_config: VAEConfig = field(default_factory=make_dreamx_world_5b_cam_vae_config)
+    text_encoder_configs: tuple[EncoderConfig,
+                                ...] = field(default_factory=lambda: (make_dreamx_world_5b_cam_text_encoder_config(), ))
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
+                                  ...] = field(default_factory=lambda: (t5_postprocess_text, ))
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", ))
+    flow_shift: float | None = 3.0
+    ti2v_task: bool = True
+    expand_timesteps: bool = True
+    vae_tiling: bool = False
+    vae_sp: bool = False
+    vae_precision: str = "fp32"
+    vae_decode_precision: str | None = "bf16"
+    dit_precision: str = "bf16"
+
+    def __post_init__(self) -> None:
+        self.vae_config.load_encoder = True
+        self.vae_config.load_decoder = True
+        self.dit_config.expand_timesteps = self.expand_timesteps
+
+
+@dataclass
+class DreamXWorld5BARPipelineConfig(DreamXWorld5BCamPipelineConfig):
+    """Pipeline config for DreamX-World-5B autoregressive forcing."""
+
+    dit_config: DiTConfig = field(default_factory=make_dreamx_world_5b_ar_dit_config)
+    flow_shift: float | None = 5.0
+    ti2v_task: bool = True
+    is_causal: bool = True
+    dmd_denoising_steps: tuple[int, ...] = (1000, 750, 500, 250)
+    warp_denoising_step: bool = True
+    context_noise: float = 0.1
+    num_frames_per_block: int = 3
+    color_correction_strength: float = 1.0
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.dit_config.expand_timesteps = True

fastvideo/configs/pipelines/kandinsky5.py

@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+
+from fastvideo.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from fastvideo.configs.models.dits import Kandinsky5VideoConfig
+from fastvideo.configs.models.encoders import BaseEncoderOutput, CLIPTextConfig
+from fastvideo.configs.models.encoders.reason1 import Reason1Config
+from fastvideo.configs.models.vaes import HunyuanVAEConfig
+from fastvideo.configs.pipelines.base import PipelineConfig, preprocess_text
+
+# Byte-exact copy of the upstream Kandinsky5/diffusers template, including the
+# "promt"/"scren" typos: the checkpoints were trained with this exact system  # codespell:ignore promt,scren
+# prompt, and ENCODE_START_IDX below is the tokenized length of everything
+# before the user prompt. Fixing the typos shifts user content to index 127
+# and mis-conditions every generation.
+KANDINSKY5_PROMPT_TEMPLATE = "\n".join([
+    "<|im_start|>system\nYou are a promt engineer. Describe the video in detail.",  # codespell:ignore promt
+    "Describe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.",
+    "Describe the location of the video, main characters or objects and their action.",
+    "Describe the dynamism of the video and presented actions.",
+    "Name the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.",  # codespell:ignore scren
+    "Describe the visual effects, postprocessing and transitions if they are presented in the video.",
+    "Pay attention to the order of key actions shown in the scene.<|im_end|>",
+    "<|im_start|>user\n{}<|im_end|>",
+])
+KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX = 129
+
+
+def kandinsky5_qwen_preprocess_text(prompt: str) -> str:
+    if not prompt.strip():
+        prompt = "."
+    return KANDINSKY5_PROMPT_TEMPLATE.format(prompt)
+
+
+def kandinsky5_qwen_postprocess_text(outputs: BaseEncoderOutput,
+                                     mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    if outputs.hidden_states is None:
+        raise RuntimeError("Kandinsky5 Qwen prompt embeddings require hidden_states.")
+    hidden_states = outputs.hidden_states[-1]
+    prompt_embeds = hidden_states[:, KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX:]
+    mask = mask[:, KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX:]
+    if prompt_embeds.shape[1] == 0:
+        prompt_embeds = hidden_states[:, -1:]
+        mask = torch.ones((mask.shape[0], 1), dtype=mask.dtype, device=mask.device)
+    return prompt_embeds, mask
+
+
+def kandinsky5_clip_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    if outputs.pooler_output is None:
+        raise RuntimeError("Kandinsky5 CLIP pooled output is required.")
+    return outputs.pooler_output
+
+
+@dataclass
+class Kandinsky5T2VConfig(PipelineConfig):
+    """Kandinsky-5.0 Lite text-to-video pipeline configuration."""
+
+    dit_config: DiTConfig = field(default_factory=Kandinsky5VideoConfig)
+    vae_config: VAEConfig = field(default_factory=HunyuanVAEConfig)
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(default_factory=lambda: (Reason1Config(), CLIPTextConfig()))
+    preprocess_text_funcs: tuple[Callable[[str], Any], ...] = field(
+        default_factory=lambda: (kandinsky5_qwen_preprocess_text, preprocess_text))
+    postprocess_text_funcs: tuple[Callable[..., Any], ...] = field(
+        default_factory=lambda: (kandinsky5_qwen_postprocess_text, kandinsky5_clip_postprocess_text))
+
+    dit_precision: str = "bf16"
+    vae_precision: str = "bf16"
+    text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", "bf16"))
+    text_encoder_max_lengths: tuple[int, ...] = field(
+        default_factory=lambda: (KANDINSKY5_PROMPT_TEMPLATE_ENCODE_START_IDX + 512, 77))
+
+    flow_shift: float | None = 5.0
+    vae_tiling: bool = True
+
+    def __post_init__(self) -> None:
+        if len(self.text_encoder_configs) != 2:
+            raise ValueError(f"Kandinsky5 pipeline requires exactly 2 text encoders (qwen and clip), "
+                             f"but got {len(self.text_encoder_configs)} encoder(s).")
+        if len(self.text_encoder_precisions) != 2:
+            raise ValueError("Kandinsky5 pipeline requires exactly 2 text encoder precisions, "
+                             f"but got {len(self.text_encoder_precisions)}.")
+        if len(self.text_encoder_max_lengths) != 2:
+            raise ValueError("Kandinsky5 pipeline requires exactly 2 text encoder max lengths, "
+                             f"but got {len(self.text_encoder_max_lengths)}.")
+
+        self.vae_config.load_encoder = False
+        self.vae_config.load_decoder = True
+
+        qwen_cfg = self.text_encoder_configs[0]
+        qwen_cfg.arch_config.output_hidden_states = True
+        qwen_cfg.arch_config.tokenizer_kwargs.update({
+            "padding": True,
+            "truncation": True,
+            "return_tensors": "pt",
+        })
+
+        clip_cfg = self.text_encoder_configs[1]
+        clip_cfg.arch_config.tokenizer_kwargs.update({
+            "padding": "max_length",
+            "max_length": 77,
+            "truncation": True,
+            "add_special_tokens": True,
+            "return_tensors": "pt",
+        })
+
+
+@dataclass
+class Kandinsky5I2VConfig(Kandinsky5T2VConfig):
+    """Kandinsky-5.0 image-to-video pipeline configuration."""
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        # I2V needs the VAE encoder to encode the conditioning image.
+        self.vae_config.load_encoder = True

fastvideo/distributed/parallel_state.py

@@ -23,6 +23,7 @@ If you only need to use the distributed environment without model parallelism,
  you can skip the model parallel initialization and destruction steps.
 """
 import contextlib
+import gc
 import os
 import pickle
 import weakref
@@ -1006,6 +1007,13 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
     if shutdown_ray:
         import ray  # Lazy import Ray
         ray.shutdown()
+    # Actually free GPU memory, as the name promises. FSDP-wrapped modules sit
+    # in reference cycles (module <-> FSDP state <-> hooks), so dropping the
+    # last user reference does not free their parameters until a gc pass runs.
+    # Without this, back-to-back model-load tests in one pytest session OOM.
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
 
 
 def get_same_node_ranks(pg: ProcessGroup | StatelessProcessGroup, source_rank: int = 0) -> list[int]:

fastvideo/models/dits/dreamx_world.py

@@ -0,0 +1,511 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+from typing import Any
+
+import torch
+import torch.nn as nn
+
+from fastvideo.configs.models.dits.dreamx_world import DreamXWorldConfig
+from fastvideo.distributed.communication_op import (
+    sequence_model_parallel_all_gather_with_unpad,
+    sequence_model_parallel_shard)
+from fastvideo.distributed.parallel_state import get_sp_world_size
+from fastvideo.layers.layernorm import RMSNorm
+from fastvideo.layers.linear import ReplicatedLinear
+from fastvideo.layers.rotary_embedding import get_rotary_pos_embed
+from fastvideo.models.dits.wanvideo import (LayerNormScaleShift,
+                                            PatchEmbed,
+                                            WanTimeTextImageEmbedding,
+                                            WanTransformer3DModel,
+                                            WanTransformerBlock)
+from fastvideo.platforms import AttentionBackendEnum, current_platform
+from fastvideo.attention import LocalAttention
+from fastvideo.distributed.parallel_state import get_sp_world_size
+from fastvideo.layers.quantization import QuantizationConfig
+from fastvideo.models.dits.base import BaseDiT
+
+
+def _dreamx_invert_se3(transforms: torch.Tensor) -> torch.Tensor:
+    assert transforms.shape[-2:] == (4, 4)
+    rot_inv = transforms[..., :3, :3].transpose(-1, -2)
+    out = torch.zeros_like(transforms)
+    out[..., :3, :3] = rot_inv
+    out[..., :3, 3] = -torch.einsum("...ij,...j->...i", rot_inv,
+                                    transforms[..., :3, 3])
+    out[..., 3, 3] = 1.0
+    return out.to(dtype=transforms.dtype)
+
+
+def _dreamx_lift_k(intrinsics: torch.Tensor) -> torch.Tensor:
+    assert intrinsics.shape[-2:] == (3, 3)
+    out = torch.zeros(intrinsics.shape[:-2] + (4, 4),
+                      device=intrinsics.device,
+                      dtype=intrinsics.dtype)
+    out[..., :3, :3] = intrinsics
+    out[..., 3, 3] = 1.0
+    return out
+
+
+def _dreamx_invert_k(intrinsics: torch.Tensor) -> torch.Tensor:
+    assert intrinsics.shape[-2:] == (3, 3)
+    out = torch.zeros_like(intrinsics)
+    out[..., 0, 0] = 1.0 / intrinsics[..., 0, 0]
+    out[..., 1, 1] = 1.0 / intrinsics[..., 1, 1]
+    out[..., 0, 2] = -intrinsics[..., 0, 2] / intrinsics[..., 0, 0]
+    out[..., 1, 2] = -intrinsics[..., 1, 2] / intrinsics[..., 1, 1]
+    out[..., 2, 2] = 1.0
+    return out.to(dtype=intrinsics.dtype)
+
+
+def _dreamx_apply_tiled_projmat(feats: torch.Tensor,
+                                matrix: torch.Tensor) -> torch.Tensor:
+    batch, num_heads, seq_len, feat_dim = feats.shape
+    proj_dim = matrix.shape[-1]
+    assert feat_dim % proj_dim == 0
+
+    if matrix.shape[1] == seq_len:
+        feats = feats.view(batch, num_heads, seq_len, feat_dim // proj_dim,
+                           proj_dim)
+        out = torch.einsum("btij,bntpj->bntpi", matrix, feats)
+        return out.reshape(batch, num_heads, seq_len, feat_dim)
+
+    cameras = matrix.shape[1]
+    assert seq_len > cameras and seq_len % cameras == 0
+    feats = feats.reshape(batch, num_heads, cameras, -1,
+                          feat_dim // proj_dim, proj_dim)
+    out = torch.einsum("bcij,bncpkj->bncpki", matrix, feats)
+    return out.reshape(batch, num_heads, seq_len, feat_dim)
+
+
+def _dreamx_prope_qkv(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                      viewmats: torch.Tensor, intrinsics: torch.Tensor):
+    batch, num_heads, seq_len, head_dim = q.shape
+    cameras = viewmats.shape[1]
+    assert q.shape == k.shape == v.shape
+    assert viewmats.shape == (batch, cameras, 4, 4)
+    assert intrinsics.shape == (batch, cameras, 3, 3)
+    assert head_dim % 4 == 0
+
+    intrinsics_norm = torch.zeros_like(intrinsics)
+    intrinsics_norm[..., 0, 0] = intrinsics[..., 0, 0]
+    intrinsics_norm[..., 1, 1] = intrinsics[..., 1, 1]
+    intrinsics_norm[..., 2, 2] = 1.0
+
+    proj = torch.einsum("...ij,...jk->...ik",
+                        _dreamx_lift_k(intrinsics_norm), viewmats)
+    proj_t = proj.transpose(-1, -2).to(dtype=viewmats.dtype)
+    proj_inv = torch.einsum(
+        "...ij,...jk->...ik",
+        _dreamx_invert_se3(viewmats),
+        _dreamx_lift_k(_dreamx_invert_k(intrinsics_norm)),
+    ).to(dtype=viewmats.dtype)
+
+    q = _dreamx_apply_tiled_projmat(q, proj_t)
+    k = _dreamx_apply_tiled_projmat(k, proj_inv)
+    v = _dreamx_apply_tiled_projmat(v, proj_inv)
+    return q, k, v, proj
+
+
+class DreamXPropeSelfAttention(nn.Module):
+    """DreamX-World parallel PRoPE camera self-attention branch."""
+
+    def __init__(self,
+                 dim: int,
+                 attn_dim: int,
+                 num_heads: int,
+                 qk_norm: str | bool = True,
+                 eps: float = 1e-6,
+                 quant_config: QuantizationConfig | None = None,
+                 prefix: str = ""):
+        super().__init__()
+        assert attn_dim % num_heads == 0
+        self.attn_dim = attn_dim
+        self.num_heads = num_heads
+        self.head_dim = attn_dim // num_heads
+        self.qk_norm = qk_norm
+
+        self.q_proj = ReplicatedLinear(dim,
+                                       attn_dim,
+                                       quant_config=quant_config,
+                                       prefix=f"{prefix}.q_proj")
+        self.k_proj = ReplicatedLinear(dim,
+                                       attn_dim,
+                                       quant_config=quant_config,
+                                       prefix=f"{prefix}.k_proj")
+        self.v_proj = ReplicatedLinear(dim,
+                                       attn_dim,
+                                       quant_config=quant_config,
+                                       prefix=f"{prefix}.v_proj")
+        self.out_proj = ReplicatedLinear(attn_dim,
+                                         dim,
+                                         quant_config=quant_config,
+                                         prefix=f"{prefix}.out_proj")
+
+        if qk_norm == "rms_norm":
+            self.norm_q = RMSNorm(self.head_dim, eps=eps)
+            self.norm_k = RMSNorm(self.head_dim, eps=eps)
+        elif qk_norm in (True, "rms_norm_across_heads"):
+            self.norm_q = RMSNorm(attn_dim, eps=eps)
+            self.norm_k = RMSNorm(attn_dim, eps=eps)
+        elif qk_norm is False:
+            self.norm_q = nn.Identity()
+            self.norm_k = nn.Identity()
+        else:
+            raise ValueError(f"Unsupported qk_norm for DreamX PRoPE: {qk_norm}")
+
+        nn.init.zeros_(self.out_proj.weight)
+        if self.out_proj.bias is not None:
+            nn.init.zeros_(self.out_proj.bias)
+
+        self.attn = LocalAttention(
+            num_heads=num_heads,
+            head_size=self.head_dim,
+            dropout_rate=0,
+            softmax_scale=None,
+            causal=False,
+            supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN,
+                                          AttentionBackendEnum.TORCH_SDPA))
+
+    def forward(self, hidden_states: torch.Tensor,
+                y_camera: dict[str, torch.Tensor]) -> torch.Tensor:
+        if get_sp_world_size() > 1:
+            # The transformer shards the sequence before the block loop and
+            # this branch uses LocalAttention (no all-to-all): under
+            # sequence parallelism each rank would attend only within its
+            # own shard — silently wrong output. Fail loudly until this
+            # path is ported to DistributedAttention and validated.
+            raise NotImplementedError(
+                "DreamXPropeSelfAttention does not support sequence "
+                "parallelism yet (LocalAttention on a sharded sequence "
+                "corrupts output). Run with sp_size=1.")
+        batch_size, seq_len, _ = hidden_states.shape
+
+        query, _ = self.q_proj(hidden_states)
+        key, _ = self.k_proj(hidden_states)
+        value, _ = self.v_proj(hidden_states)
+
+        if self.qk_norm == "rms_norm":
+            query = query.view(batch_size, seq_len, self.num_heads,
+                               self.head_dim)
+            key = key.view(batch_size, seq_len, self.num_heads, self.head_dim)
+            query = self.norm_q(query)
+            key = self.norm_k(key)
+        else:
+            query = self.norm_q(query).view(batch_size, seq_len,
+                                            self.num_heads, self.head_dim)
+            key = self.norm_k(key).view(batch_size, seq_len, self.num_heads,
+                                        self.head_dim)
+
+        value = value.view(batch_size, seq_len, self.num_heads, self.head_dim)
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        query, key, value, output_projection = _dreamx_prope_qkv(
+            query,
+            key,
+            value,
+            viewmats=y_camera["viewmats"],
+            intrinsics=y_camera["K"],
+        )
+
+        out = self.attn(query.transpose(1, 2), key.transpose(1, 2),
+                        value.transpose(1, 2))
+        out = _dreamx_apply_tiled_projmat(out.transpose(1, 2),
+                                          output_projection).transpose(1, 2)
+        out = out.flatten(2)
+        out, _ = self.out_proj(out)
+        return out
+
+
+class DreamXWorldTransformerBlock(WanTransformerBlock):
+
+    def __init__(self,
+                 dim: int,
+                 ffn_dim: int,
+                 num_heads: int,
+                 qk_norm: str = "rms_norm_across_heads",
+                 cross_attn_norm: bool = False,
+                 eps: float = 1e-6,
+                 added_kv_proj_dim: int | None = None,
+                 supported_attention_backends: tuple[AttentionBackendEnum, ...]
+                 | None = None,
+                 quant_config: QuantizationConfig | None = None,
+                 prefix: str = "",
+                 add_control_adapter: bool = True,
+                 cam_method: str | None = "prope",
+                 attn_compress: int = 1,
+                 cam_self_attn_layers: tuple[int, ...] | None = None,
+                 layer_idx: int | None = None):
+        super().__init__(dim, ffn_dim, num_heads, qk_norm, cross_attn_norm,
+                         eps, added_kv_proj_dim,
+                         supported_attention_backends, quant_config, prefix)
+        self.cam_self_attn = None
+        add_cam_attn = add_control_adapter and cam_method == "prope"
+        if add_cam_attn and cam_self_attn_layers is not None:
+            add_cam_attn = layer_idx in cam_self_attn_layers
+        if add_cam_attn:
+            if num_heads % attn_compress != 0 or dim % attn_compress != 0:
+                raise ValueError("DreamX attn_compress must divide dim and num_heads")
+            self.cam_self_attn = DreamXPropeSelfAttention(
+                dim,
+                dim // attn_compress,
+                num_heads // attn_compress,
+                qk_norm=qk_norm,
+                eps=eps,
+                quant_config=quant_config,
+                prefix=f"{prefix}.cam_self_attn")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        freqs_cis: tuple[torch.Tensor, torch.Tensor],
+        original_seq_len: int,
+        y_camera: dict[str, torch.Tensor] | None = None,
+    ) -> torch.Tensor:
+        if hidden_states.dim() == 4:
+            hidden_states = hidden_states.squeeze(1)
+        orig_dtype = hidden_states.dtype
+
+        if temb.dim() == 4:
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
+                self.scale_shift_table.unsqueeze(0) + temb.float()).chunk(
+                    6, dim=2)
+            shift_msa = shift_msa.squeeze(2)
+            scale_msa = scale_msa.squeeze(2)
+            gate_msa = gate_msa.squeeze(2)
+            c_shift_msa = c_shift_msa.squeeze(2)
+            c_scale_msa = c_scale_msa.squeeze(2)
+            c_gate_msa = c_gate_msa.squeeze(2)
+        else:
+            e = self.scale_shift_table + temb.float()
+            shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
+                6, dim=1)
+        assert shift_msa.dtype == torch.float32
+
+        norm_hidden_states = (self.norm1(hidden_states.float()) *
+                              (1 + scale_msa) + shift_msa).to(orig_dtype)
+        query, _ = self.to_q(norm_hidden_states)
+        key, _ = self.to_k(norm_hidden_states)
+        value, _ = self.to_v(norm_hidden_states)
+
+        if self.norm_q is not None:
+            query = self.norm_q(query)
+        if self.norm_k is not None:
+            key = self.norm_k(key)
+
+        query = query.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
+        key = key.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
+        value = value.squeeze(1).unflatten(2, (self.num_attention_heads, -1))
+
+        attn_output, _ = self.attn1(
+            query,
+            key,
+            value,
+            original_seq_len,
+            freqs_cis=freqs_cis,
+        )
+        attn_output = attn_output.flatten(2)
+        attn_output, _ = self.to_out(attn_output)
+        attn_output = attn_output.squeeze(1)
+        if self.cam_self_attn is not None and y_camera is not None:
+            attn_output = attn_output + self.cam_self_attn(
+                norm_hidden_states, y_camera)
+
+        null_shift = null_scale = torch.tensor([0], device=hidden_states.device)
+        norm_hidden_states, hidden_states = self.self_attn_residual_norm(
+            hidden_states, attn_output, gate_msa, null_shift, null_scale)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
+
+        attn_output = self.attn2(norm_hidden_states,
+                                 context=encoder_hidden_states,
+                                 context_lens=None)
+        norm_hidden_states, hidden_states = self.cross_attn_residual_norm(
+            hidden_states, attn_output, 1, c_shift_msa, c_scale_msa)
+        norm_hidden_states, hidden_states = norm_hidden_states.to(
+            orig_dtype), hidden_states.to(orig_dtype)
+
+        ff_output = self.ffn(norm_hidden_states)
+        hidden_states = self.mlp_residual(hidden_states, ff_output, c_gate_msa)
+        hidden_states = hidden_states.to(orig_dtype)
+
+        return hidden_states
+
+
+class DreamXWorldTransformer3DModel(WanTransformer3DModel):
+    _fsdp_shard_conditions = DreamXWorldConfig()._fsdp_shard_conditions
+    _compile_conditions = DreamXWorldConfig()._compile_conditions
+    _supported_attention_backends = DreamXWorldConfig(
+    )._supported_attention_backends
+    param_names_mapping = DreamXWorldConfig().param_names_mapping
+    reverse_param_names_mapping = DreamXWorldConfig().reverse_param_names_mapping
+    lora_param_names_mapping = DreamXWorldConfig().lora_param_names_mapping
+
+    def __init__(self, config: DreamXWorldConfig, hf_config: dict[str,
+                                                                  Any]) -> None:
+        BaseDiT.__init__(self, config=config, hf_config=hf_config)
+        self.quant_config = config.quant_config
+
+        inner_dim = config.num_attention_heads * config.attention_head_dim
+        self.hidden_size = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.in_channels = config.in_channels
+        self.out_channels = config.out_channels
+        self.num_channels_latents = config.num_channels_latents
+        self.patch_size = config.patch_size
+        self.text_len = config.text_len
+
+        assert config.num_attention_heads % get_sp_world_size() == 0, f"The number of attention heads ({config.num_attention_heads}) must be divisible by the sequence parallel size ({get_sp_world_size()})"
+
+        self.patch_embedding = PatchEmbed(in_chans=config.in_channels,
+                                          embed_dim=inner_dim,
+                                          patch_size=config.patch_size,
+                                          flatten=False)
+        self.condition_embedder = WanTimeTextImageEmbedding(
+            dim=inner_dim,
+            time_freq_dim=config.freq_dim,
+            text_embed_dim=config.text_dim,
+            image_embed_dim=config.image_dim,
+        )
+        self.blocks = nn.ModuleList([
+            DreamXWorldTransformerBlock(
+                inner_dim,
+                config.ffn_dim,
+                config.num_attention_heads,
+                config.qk_norm,
+                config.cross_attn_norm,
+                config.eps,
+                config.added_kv_proj_dim,
+                self._supported_attention_backends,
+                quant_config=config.quant_config,
+                prefix=f"{config.prefix}.blocks.{i}",
+                add_control_adapter=config.add_control_adapter,
+                cam_method=config.cam_method,
+                attn_compress=config.attn_compress,
+                cam_self_attn_layers=config.cam_self_attn_layers,
+                layer_idx=i)
+            for i in range(config.num_layers)
+        ])
+        self.norm_out = LayerNormScaleShift(inner_dim,
+                                            norm_type="layer",
+                                            eps=config.eps,
+                                            elementwise_affine=False,
+                                            dtype=torch.float32,
+                                            compute_dtype=torch.float32)
+        self.proj_out = nn.Linear(
+            inner_dim, config.out_channels * math.prod(config.patch_size))
+        self.scale_shift_table = nn.Parameter(
+            torch.randn(1, 2, inner_dim) / inner_dim**0.5)
+        self.gradient_checkpointing = False
+        self.__post_init__()
+
+    def forward(self,
+                hidden_states: torch.Tensor,
+                encoder_hidden_states: torch.Tensor | list[torch.Tensor],
+                timestep: torch.LongTensor,
+                encoder_hidden_states_image: torch.Tensor | list[torch.Tensor]
+                | None = None,
+                guidance=None,
+                y_camera: dict[str, torch.Tensor] | None = None,
+                **kwargs) -> torch.Tensor:
+        orig_dtype = hidden_states.dtype
+        if encoder_hidden_states is not None and not isinstance(
+                encoder_hidden_states, torch.Tensor):
+            encoder_hidden_states = encoder_hidden_states[0]
+        if isinstance(encoder_hidden_states_image,
+                      list) and len(encoder_hidden_states_image) > 0:
+            encoder_hidden_states_image = encoder_hidden_states_image[0]
+        else:
+            encoder_hidden_states_image = None
+
+        batch_size, _, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        d = self.hidden_size // self.num_attention_heads
+        rope_dim_list = [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)]
+        freqs_cos, freqs_sin = get_rotary_pos_embed(
+            (post_patch_num_frames, post_patch_height, post_patch_width),
+            self.hidden_size,
+            self.num_attention_heads,
+            rope_dim_list,
+            dtype=torch.float32 if current_platform.is_mps() else torch.float64,
+            rope_theta=10000)
+        freqs_cis = (freqs_cos.to(hidden_states.device).float(),
+                     freqs_sin.to(hidden_states.device).float())
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+        hidden_states, original_seq_len = sequence_model_parallel_shard(
+            hidden_states, dim=1)
+
+        if timestep.dim() == 2:
+            ts_seq_len = timestep.shape[1]
+            timestep = timestep.flatten()
+        else:
+            ts_seq_len = None
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep,
+            encoder_hidden_states,
+            encoder_hidden_states_image,
+            timestep_seq_len=ts_seq_len)
+        if ts_seq_len is not None:
+            timestep_proj = timestep_proj.unflatten(2, (6, -1))
+        else:
+            timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            if encoder_hidden_states is not None:
+                encoder_hidden_states = torch.concat(
+                    [encoder_hidden_states_image, encoder_hidden_states], dim=1)
+            else:
+                encoder_hidden_states = encoder_hidden_states_image
+
+        if current_platform.is_mps() or current_platform.is_npu():
+            encoder_hidden_states = encoder_hidden_states.to(orig_dtype)
+
+        assert encoder_hidden_states.dtype == orig_dtype
+
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states,
+                    timestep_proj, freqs_cis, original_seq_len, y_camera)
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states,
+                                      timestep_proj, freqs_cis, original_seq_len,
+                                      y_camera=y_camera)
+
+        if temb.dim() == 3:
+            shift, scale = (self.scale_shift_table.unsqueeze(0) +
+                            temb.unsqueeze(2)).chunk(2, dim=2)
+            shift = shift.squeeze(2)
+            scale = scale.squeeze(2)
+        else:
+            shift, scale = (self.scale_shift_table +
+                            temb.unsqueeze(1)).chunk(2, dim=1)
+
+        hidden_states = self.norm_out(hidden_states, shift, scale)
+        hidden_states = sequence_model_parallel_all_gather_with_unpad(
+            hidden_states, original_seq_len, dim=1)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(batch_size, post_patch_num_frames,
+                                              post_patch_height,
+                                              post_patch_width, p_t, p_h, p_w,
+                                              -1)
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        return output
+
+
+EntryClass = DreamXWorldTransformer3DModel

fastvideo/models/dits/dreamx_world_ar.py

@@ -0,0 +1,920 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World autoregressive causal DiT.
+
+Adapted from DreamX-World's Apache-2.0
+``wan/modules/causal_camera_model_2_2_prope_infinity.py``. The implementation is
+kept native to FastVideo: no production import from DreamX, Diffusers, or
+Transformers is required.
+"""
+
+from typing import Any
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fastvideo.configs.models.dits.dreamx_world import DreamXWorldARConfig
+from fastvideo.layers.linear import ReplicatedLinear
+from fastvideo.models.dits.base import BaseDiT
+from fastvideo.models.dits.dreamx_world import (_dreamx_apply_tiled_projmat,
+                                               _dreamx_prope_qkv)
+
+
+def attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+    # Deliberately raw SDPA rather than fastvideo.attention.LocalAttention:
+    # (1) LocalAttention dispatches through the attention-backend registry, so
+    #     FLASH_ATTN could be selected and its kernel is not bit-identical to
+    #     torch SDPA — the AR KV-cache rollout must stay numerically frozen;
+    # (2) LocalAttention requires an active ForwardContext, which direct
+    #     transformer invocations (parity tests) do not set;
+    # (3) the sibling causal model keeps raw SDPA in the same KV-cache window
+    #     path (matrixgame2/causal_model.py).
+    # Sequence-parallel gap: this model never shards the sequence; run with
+    # sp_size=1 (see fastvideo/layers/AGENTS.md on documenting raw SDPA).
+    q_bhld = q.transpose(1, 2)
+    k_bhld = k.transpose(1, 2)
+    v_bhld = v.transpose(1, 2)
+    out = F.scaled_dot_product_attention(q_bhld, k_bhld, v_bhld, dropout_p=0.0)
+    return out.transpose(1, 2)
+
+
+def prope_qkv(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+              viewmats: torch.Tensor, Ks: torch.Tensor):
+    q, k, v, output_projection = _dreamx_prope_qkv(q, k, v, viewmats, Ks)
+
+    def apply_fn_o(x: torch.Tensor) -> torch.Tensor:
+        return _dreamx_apply_tiled_projmat(x, output_projection)
+
+    return q, k, v, apply_fn_o
+
+
+def sinusoidal_embedding_1d(dim, position):
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float64)
+    sinusoid = torch.outer(
+        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    return torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+
+
+def rope_params(max_seq_len, dim, theta=10000):
+    assert dim % 2 == 0
+    freqs = torch.outer(
+        torch.arange(max_seq_len),
+        1.0 / torch.pow(theta,
+                        torch.arange(0, dim, 2).to(torch.float64).div(dim)))
+    return torch.polar(torch.ones_like(freqs), freqs)
+
+
+class WanRMSNorm(nn.Module):
+    """Kept private instead of fastvideo.layers.layernorm.RMSNorm.
+
+    The official DreamX-World ``model_2_2.py`` computes the RMS statistics in
+    the *input* dtype — the upstream code has the fp32 upcast explicitly
+    commented out (``# return self._norm(x.float())...``). FastVideo's RMSNorm
+    always normalizes in fp32, which is not bit-identical under bf16, so the
+    verbatim implementation stays.
+    """
+
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.dim = dim
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return self._norm(x).type_as(x) * self.weight
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
+
+
+class WanLayerNorm(nn.LayerNorm):
+    """Kept private instead of fastvideo.layers.layernorm.FP32LayerNorm.
+
+    The official DreamX-World ``model_2_2.py`` normalizes in the *input* dtype
+    (no ``x.float()`` upcast, unlike Wan2.1). FP32LayerNorm casts input and
+    affine params to fp32, which is not bit-identical under bf16, so the
+    verbatim implementation stays.
+    """
+
+    def __init__(self, dim, eps=1e-6, elementwise_affine=False):
+        super().__init__(dim, elementwise_affine=elementwise_affine, eps=eps)
+
+    def forward(self, x):
+        return super().forward(x).type_as(x)
+
+
+class WanCrossAttention(nn.Module):
+    def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.q = ReplicatedLinear(dim, dim)
+        self.k = ReplicatedLinear(dim, dim)
+        self.v = ReplicatedLinear(dim, dim)
+        self.o = ReplicatedLinear(dim, dim)
+        self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+        self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+
+    def _kv(self, context, b, n, d):
+        k, _ = self.k(context)
+        k = self.norm_k(k).view(b, -1, n, d)
+        v, _ = self.v(context)
+        v = v.view(b, -1, n, d)
+        return k, v
+
+    def forward(self, x, context, context_lens, crossattn_cache=None):
+        b, n, d = x.size(0), self.num_heads, self.head_dim
+        q, _ = self.q(x)
+        q = self.norm_q(q).view(b, -1, n, d)
+
+        if crossattn_cache is not None:
+            if not crossattn_cache["is_init"]:
+                crossattn_cache["is_init"] = True
+                k, v = self._kv(context, b, n, d)
+                crossattn_cache["k"] = k
+                crossattn_cache["v"] = v
+            else:
+                k = crossattn_cache["k"]
+                v = crossattn_cache["v"]
+        else:
+            k, v = self._kv(context, b, n, d)
+
+        x = attention(q, k, v)
+        x = x.flatten(2)
+        out, _ = self.o(x)
+        return out
+
+
+def block_relativistic_rope(x, grid_sizes, freqs, start_frame=0, relative_frame_indices=None):
+    """
+    Apply Block-Relativistic RoPE to input tensor.
+    Adapted from Infinity-RoPE (https://arxiv.org/abs/2511.20649).
+
+    Args:
+        x: Input tensor [B, L, num_heads, head_dim]
+        grid_sizes: Tensor [B, 3] containing (F, H, W)
+        freqs: RoPE frequencies
+        start_frame: Starting frame index for sequential RoPE
+        relative_frame_indices: Optional tensor [F] specifying explicit frame indices
+                               for Block-Relativistic RoPE. Overrides start_frame if provided.
+    """
+    n, c = x.size(2), x.size(3) // 2
+    freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)
+
+    output = []
+    for i, (f, h, w) in enumerate(grid_sizes.tolist()):
+        seq_len = f * h * w
+        x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape(
+            seq_len, n, -1, 2))
+
+        if relative_frame_indices is not None:
+            frame_indices = relative_frame_indices.long()
+            freqs_temporal = freqs[0][frame_indices].view(f, 1, 1, -1).expand(f, h, w, -1)
+        else:
+            freqs_temporal = freqs[0][start_frame:start_frame + f].view(f, 1, 1, -1).expand(f, h, w, -1)
+
+        freqs_i = torch.cat([
+            freqs_temporal,
+            freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+            freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+        ], dim=-1).reshape(seq_len, 1, -1)
+
+        x_i = torch.view_as_real(x_i * freqs_i).flatten(2)
+        x_i = torch.cat([x_i, x[i, seq_len:]])
+        output.append(x_i)
+
+    return torch.stack(output).type_as(x)
+
+
+class CausalWanSelfAttention(nn.Module):
+    """Self-attention with KV cache and Block-Relativistic RoPE for causal inference."""
+
+    def __init__(self, dim, num_heads, local_attn_size=6, sink_size=1,
+                 qk_norm=True, eps=1e-6):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.local_attn_size = local_attn_size
+        self.sink_size = sink_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+        self.max_attention_size = 39600 if local_attn_size == -1 else local_attn_size * 880
+
+        self.q = ReplicatedLinear(dim, dim)
+        self.k = ReplicatedLinear(dim, dim)
+        self.v = ReplicatedLinear(dim, dim)
+        self.o = ReplicatedLinear(dim, dim)
+        self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+        self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+
+    def forward(self, x, seq_lens, grid_sizes, freqs, kv_cache,
+                current_start=0, cache_start=None, sink_recache_after_switch=False):
+        """
+        Args:
+            x: Shape [B, L, C]
+            seq_lens: Shape [B]
+            grid_sizes: Shape [B, 3] containing (F, H, W)
+            freqs: RoPE frequencies [1024, head_dim / 2]
+            kv_cache: Dict with 'k', 'v', 'global_end_index', 'local_end_index'
+            current_start: Current position in the global token sequence
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+        if cache_start is None:
+            cache_start = current_start
+
+        q, _ = self.q(x)
+        q = self.norm_q(q).view(b, s, n, d)
+        k, _ = self.k(x)
+        k = self.norm_k(k).view(b, s, n, d)
+        v, _ = self.v(x)
+        v = v.view(b, s, n, d)
+
+        frame_seqlen = math.prod(grid_sizes[0][1:]).item()
+        num_new_frames = grid_sizes[0][0].item()
+        current_end = current_start + q.shape[1]
+        sink_tokens = self.sink_size * frame_seqlen
+        kv_cache_size = kv_cache["k"].shape[1]
+        num_new_tokens = q.shape[1]
+
+        cache_update_info = None
+        is_recompute = current_end <= kv_cache["global_end_index"].item() and current_start > 0
+
+        if self.local_attn_size != -1 and (current_end > kv_cache["global_end_index"].item()) and (
+                num_new_tokens + kv_cache["local_end_index"].item() > kv_cache_size):
+            # === ROLLING MODE: cache full, evict oldest non-sink tokens ===
+            num_evicted_tokens = num_new_tokens + kv_cache["local_end_index"].item() - kv_cache_size
+            num_rolled_tokens = kv_cache["local_end_index"].item() - num_evicted_tokens - sink_tokens
+            local_end_index = kv_cache["local_end_index"].item() + current_end - \
+                kv_cache["global_end_index"].item() - num_evicted_tokens
+            local_start_index = local_end_index - num_new_tokens
+
+            temp_k = kv_cache["k"].detach().clone()
+            temp_v = kv_cache["v"].detach().clone()
+            temp_k[:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                temp_k[:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+            temp_v[:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                temp_v[:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+
+            write_start_index = max(local_start_index, sink_tokens) if is_recompute else local_start_index
+            roped_offset = max(0, write_start_index - local_start_index)
+            write_len = max(0, local_end_index - write_start_index)
+            if write_len > 0:
+                temp_k[:, write_start_index:local_end_index] = k[:, roped_offset:roped_offset + write_len]
+                temp_v[:, write_start_index:local_end_index] = v[:, roped_offset:roped_offset + write_len]
+
+            # Block-Relativistic RoPE: query uses window-relative indices
+            query_relative_indices = torch.arange(
+                self.local_attn_size - num_new_frames, self.local_attn_size, device=q.device)
+            roped_query = block_relativistic_rope(
+                q, grid_sizes, freqs, relative_frame_indices=query_relative_indices).type_as(v)
+
+            # Block-Relativistic RoPE: cached K uses position-in-window indices
+            num_cache_frames = local_end_index // frame_seqlen
+            cache_relative_indices = torch.arange(0, num_cache_frames, device=k.device)
+            cache_grid_sizes = grid_sizes.clone()
+            cache_grid_sizes[0, 0] = num_cache_frames
+            roped_temp_k = block_relativistic_rope(
+                temp_k[:, :local_end_index].view(b, num_cache_frames, frame_seqlen, n, d).flatten(1, 2),
+                cache_grid_sizes, freqs, relative_frame_indices=cache_relative_indices).type_as(v)
+
+            cache_update_info = {
+                "action": "roll_and_insert",
+                "sink_tokens": sink_tokens,
+                "num_rolled_tokens": num_rolled_tokens,
+                "num_evicted_tokens": num_evicted_tokens,
+                "local_start_index": local_start_index,
+                "local_end_index": local_end_index,
+                "write_start_index": write_start_index,
+                "write_end_index": local_end_index,
+                "new_k": k[:, roped_offset:roped_offset + write_len],
+                "new_v": v[:, roped_offset:roped_offset + write_len],
+                "current_end": current_end,
+                "is_recompute": is_recompute
+            }
+        else:
+            # === DIRECT INSERT MODE: cache not yet full ===
+            local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
+            local_start_index = local_end_index - num_new_tokens
+
+            temp_k = kv_cache["k"].detach().clone()
+            temp_v = kv_cache["v"].detach().clone()
+
+            write_start_index = max(local_start_index, sink_tokens) if is_recompute else local_start_index
+            if sink_recache_after_switch:
+                write_start_index = local_start_index
+            roped_offset = max(0, write_start_index - local_start_index)
+            write_len = max(0, local_end_index - write_start_index)
+            if write_len > 0:
+                temp_k[:, write_start_index:local_end_index] = k[:, roped_offset:roped_offset + write_len]
+                temp_v[:, write_start_index:local_end_index] = v[:, roped_offset:roped_offset + write_len]
+
+            # RoPE with relative indices (growing sequentially before cache fills)
+            current_frame_in_window = local_start_index // frame_seqlen
+            query_relative_indices = torch.arange(
+                current_frame_in_window, current_frame_in_window + num_new_frames, device=q.device)
+            roped_query = block_relativistic_rope(
+                q, grid_sizes, freqs, relative_frame_indices=query_relative_indices).type_as(v)
+
+            num_cache_frames = local_end_index // frame_seqlen
+            cache_relative_indices = torch.arange(0, num_cache_frames, device=k.device)
+            cache_grid_sizes = grid_sizes.clone()
+            cache_grid_sizes[0, 0] = num_cache_frames
+            roped_temp_k = block_relativistic_rope(
+                temp_k[:, :local_end_index].view(b, num_cache_frames, frame_seqlen, n, d).flatten(1, 2),
+                cache_grid_sizes, freqs, relative_frame_indices=cache_relative_indices).type_as(v)
+
+            cache_update_info = {
+                "action": "direct_insert",
+                "local_start_index": local_start_index,
+                "local_end_index": local_end_index,
+                "write_start_index": write_start_index,
+                "write_end_index": local_end_index,
+                "new_k": k[:, roped_offset:roped_offset + write_len],
+                "new_v": v[:, roped_offset:roped_offset + write_len],
+                "current_end": current_end,
+                "is_recompute": is_recompute
+            }
+
+        # Attention: sink tokens + local window
+        if sink_tokens > 0:
+            local_budget = self.max_attention_size - sink_tokens
+            k_sink = roped_temp_k[:, :sink_tokens]
+            v_sink = temp_v[:, :sink_tokens]
+            if local_budget > 0:
+                local_start_for_window = max(sink_tokens, local_end_index - local_budget)
+                k_local = roped_temp_k[:, local_start_for_window:local_end_index]
+                v_local = temp_v[:, local_start_for_window:local_end_index]
+                k_cat = torch.cat([k_sink, k_local], dim=1)
+                v_cat = torch.cat([v_sink, v_local], dim=1)
+            else:
+                k_cat = k_sink
+                v_cat = v_sink
+            x = attention(roped_query, k_cat, v_cat)
+        else:
+            window_start = max(0, local_end_index - self.max_attention_size)
+            x = attention(
+                roped_query,
+                roped_temp_k[:, window_start:local_end_index],
+                temp_v[:, window_start:local_end_index])
+
+        x = x.flatten(2)
+        x, _ = self.o(x)
+        return x, (current_end, local_end_index, cache_update_info)
+
+
+class CausalPropeSelfAttention(nn.Module):
+    """PRoPE self-attention with optional KV cache for camera-controlled inference."""
+
+    def __init__(self, dim, attn_dim, num_heads, window_size=(-1, -1),
+                 local_attn_size=-1, sink_size=0, qk_norm=True, eps=1e-6):
+        assert dim % num_heads == 0
+        assert attn_dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.attn_dim = attn_dim
+        self.num_heads = num_heads
+        self.head_dim = attn_dim // num_heads
+        self.local_attn_size = local_attn_size
+        self.sink_size = sink_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+        self.window_size = window_size
+        self.max_attention_size = 39600 if local_attn_size == -1 else local_attn_size * 880
+
+        self.q_proj = ReplicatedLinear(dim, attn_dim)
+        self.k_proj = ReplicatedLinear(dim, attn_dim)
+        self.v_proj = ReplicatedLinear(dim, attn_dim)
+        self.out_proj = ReplicatedLinear(attn_dim, dim)
+
+        self.norm_q = WanRMSNorm(attn_dim, eps=eps) if qk_norm else nn.Identity()
+        self.norm_k = WanRMSNorm(attn_dim, eps=eps) if qk_norm else nn.Identity()
+
+        nn.init.zeros_(self.out_proj.weight)
+        nn.init.zeros_(self.out_proj.bias)
+
+    def forward(self, x, cam_viewmats, cam_K, seq_lens, grid_sizes, freqs,
+                kv_cache=None, current_start=0, cache_start=None,
+                sink_recache_after_switch=False, cache_update_policy="commit_detached"):
+        """
+        Args:
+            x: Shape [B, L, C]
+            cam_viewmats: Camera view matrices
+            cam_K: Camera intrinsics
+            kv_cache: Optional KV cache dict. When None, runs full attention over current chunk.
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+        if cache_start is None:
+            cache_start = current_start
+
+        q, _ = self.q_proj(x)
+        q = self.norm_q(q).view(b, s, n, d)
+        k, _ = self.k_proj(x)
+        k = self.norm_k(k).view(b, s, n, d)
+        v, _ = self.v_proj(x)
+        v = v.view(b, s, n, d)
+
+        # Apply PRoPE (Positional Rotary Position Embedding from camera parameters)
+        q_t, k_t, v_t, apply_fn_o = prope_qkv(
+            q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2),
+            viewmats=cam_viewmats, Ks=cam_K)
+        proped_q = q_t.transpose(1, 2)
+        proped_k = k_t.transpose(1, 2)
+        proped_v = v_t.transpose(1, 2)
+
+        if kv_cache is None:
+            # No cache: full attention over current chunk
+            x_out = attention(proped_q, proped_k, proped_v)
+        else:
+            # KV cache mode with rolling cache support
+            frame_seqlen = math.prod(grid_sizes[0][1:]).item()
+            num_new_tokens = s
+            current_end = current_start + num_new_tokens
+            sink_tokens = self.sink_size * frame_seqlen
+            kv_cache_size = kv_cache["k"].shape[1]
+            is_recompute = (current_end <= kv_cache["global_end_index"].item()) and (current_start > 0)
+
+            if self.local_attn_size != -1 and (current_end > kv_cache["global_end_index"].item()) and (
+                    num_new_tokens + kv_cache["local_end_index"].item() > kv_cache_size):
+                # === ROLLING MODE ===
+                num_evicted_tokens = num_new_tokens + kv_cache["local_end_index"].item() - kv_cache_size
+                num_rolled_tokens = kv_cache["local_end_index"].item() - num_evicted_tokens - sink_tokens
+                local_end_index = kv_cache["local_end_index"].item() + current_end - \
+                    kv_cache["global_end_index"].item() - num_evicted_tokens
+                local_start_index = local_end_index - num_new_tokens
+
+                if cache_update_policy != "none":
+                    with torch.no_grad():
+                        kv_cache["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                            kv_cache["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+                        kv_cache["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                            kv_cache["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+
+                    write_start_index = max(local_start_index, sink_tokens) if is_recompute else local_start_index
+                    roped_offset = max(0, write_start_index - local_start_index)
+                    write_len = max(0, local_end_index - write_start_index)
+                    if write_len > 0:
+                        with torch.no_grad():
+                            kv_cache["k"][:, write_start_index:local_end_index] = proped_k[:, roped_offset:roped_offset + write_len].detach()
+                            kv_cache["v"][:, write_start_index:local_end_index] = proped_v[:, roped_offset:roped_offset + write_len].detach()
+            else:
+                # === DIRECT INSERT MODE ===
+                local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
+                local_start_index = local_end_index - num_new_tokens
+
+                if cache_update_policy != "none":
+                    write_start_index = max(local_start_index, sink_tokens) if is_recompute else local_start_index
+                    if sink_recache_after_switch:
+                        write_start_index = local_start_index
+                    roped_offset = max(0, write_start_index - local_start_index)
+                    write_len = max(0, local_end_index - write_start_index)
+                    if write_len > 0:
+                        with torch.no_grad():
+                            kv_cache["k"][:, write_start_index:local_end_index] = proped_k[:, roped_offset:roped_offset + write_len].detach()
+                            kv_cache["v"][:, write_start_index:local_end_index] = proped_v[:, roped_offset:roped_offset + write_len].detach()
+
+            # Attention: sink tokens + local window
+            if sink_tokens > 0:
+                local_budget = self.max_attention_size - sink_tokens
+                k_sink = kv_cache["k"][:, :sink_tokens].detach()
+                v_sink = kv_cache["v"][:, :sink_tokens].detach()
+                if local_budget > 0:
+                    local_start_for_window = max(sink_tokens, local_end_index - local_budget)
+                    k_local = kv_cache["k"][:, local_start_for_window:local_end_index].detach()
+                    v_local = kv_cache["v"][:, local_start_for_window:local_end_index].detach()
+                    k_cat = torch.cat([k_sink, k_local], dim=1)
+                    v_cat = torch.cat([v_sink, v_local], dim=1)
+                else:
+                    k_cat = k_sink
+                    v_cat = v_sink
+                x_out = attention(proped_q, k_cat, v_cat)
+            else:
+                window_start = max(0, local_end_index - self.max_attention_size)
+                x_out = attention(
+                    proped_q,
+                    kv_cache["k"][:, window_start:local_end_index].detach(),
+                    kv_cache["v"][:, window_start:local_end_index].detach())
+
+            if not is_recompute and cache_update_policy != "none":
+                kv_cache["global_end_index"].fill_(current_end)
+                kv_cache["local_end_index"].fill_(local_end_index)
+
+        # Apply inverse PRoPE
+        x = apply_fn_o(x_out.transpose(1, 2)).transpose(1, 2)
+        x = x.flatten(2)
+        x, _ = self.out_proj(x)
+        return x
+
+
+class CausalWanAttentionBlock(nn.Module):
+
+    def __init__(self, dim, ffn_dim, num_heads, local_attn_size=-1, sink_size=0,
+                 qk_norm=True, cross_attn_norm=False, eps=1e-6, **kwargs):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.local_attn_size = local_attn_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        self.add_control_adapter = kwargs.get('add_control_adapter', False)
+        self.cam_method = kwargs.get('cam_method')
+        self.attn_compress = kwargs.get('attn_compress', 1)
+        self.layer_idx = kwargs.get('layer_idx')
+        cam_self_attn_layers = kwargs.get('cam_self_attn_layers')
+
+        # layers
+        self.norm1 = WanLayerNorm(dim, eps)
+        self.self_attn = CausalWanSelfAttention(
+            dim, num_heads, local_attn_size, sink_size, qk_norm, eps)
+        self.norm3 = WanLayerNorm(
+            dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WanCrossAttention(dim, num_heads, (-1, -1), qk_norm, eps)
+        self.norm2 = WanLayerNorm(dim, eps)
+        # nn.Linear (not ReplicatedLinear) on purpose: the official checkpoint
+        # stores these as positional Sequential keys (ffn.0 / ffn.2) that the
+        # copy-only converter and the strict-load tests require verbatim, and
+        # ReplicatedLinear's (out, bias) tuple return cannot compose inside
+        # nn.Sequential without renaming the state-dict surface.
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, ffn_dim), nn.GELU(approximate='tanh'),
+            nn.Linear(ffn_dim, dim))
+
+        # PRoPE self-attention branch for camera control
+        add_cam_attn = self.add_control_adapter and self.cam_method == 'prope'
+        if add_cam_attn and cam_self_attn_layers is not None:
+            add_cam_attn = self.layer_idx in cam_self_attn_layers
+        if add_cam_attn:
+            self.cam_self_attn = CausalPropeSelfAttention(
+                dim, dim // self.attn_compress, num_heads,
+                local_attn_size=local_attn_size, sink_size=sink_size,
+                qk_norm=qk_norm, eps=eps)
+
+        # modulation
+        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+
+    def forward(self, x, e, seq_lens, grid_sizes, freqs, context, context_lens,
+                kv_cache, crossattn_cache=None, current_start=0, cache_start=None,
+                cam_viewmats=None, cam_K=None, sink_recache_after_switch=False,
+                cache_update_policy="commit_detached"):
+        num_frames, frame_seqlen = e.shape[1], x.shape[1] // e.shape[1]
+        e = (self.modulation.unsqueeze(1) + e).chunk(6, dim=2)
+
+        # self-attention
+        attn_input = (self.norm1(x).unflatten(
+            dim=1, sizes=(num_frames, frame_seqlen)) * (1 + e[1]) + e[0]).flatten(1, 2)
+        y, cache_update_info = self.self_attn(
+            attn_input, seq_lens, grid_sizes, freqs, kv_cache,
+            current_start, cache_start, sink_recache_after_switch)
+
+        # PRoPE camera attention (parallel branch)
+        if hasattr(self, 'cam_self_attn') and cam_viewmats is not None and cam_K is not None:
+            prope_kv_cache = None
+            if kv_cache is not None and "prope_k" in kv_cache:
+                prope_kv_cache = {
+                    "k": kv_cache["prope_k"],
+                    "v": kv_cache["prope_v"],
+                    "global_end_index": kv_cache["prope_global_end_index"],
+                    "local_end_index": kv_cache["prope_local_end_index"],
+                }
+            y = y + self.cam_self_attn(
+                attn_input, cam_viewmats, cam_K, seq_lens, grid_sizes, freqs,
+                kv_cache=prope_kv_cache, current_start=current_start,
+                cache_start=cache_start, cache_update_policy=cache_update_policy)
+
+        x = x + (y.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * e[2]).flatten(1, 2)
+
+        # cross-attention & FFN
+        x = x + self.cross_attn(self.norm3(x), context, context_lens,
+                                crossattn_cache=crossattn_cache)
+        y = self.ffn(
+            (self.norm2(x).unflatten(dim=1, sizes=(num_frames, frame_seqlen))
+             * (1 + e[4]) + e[3]).flatten(1, 2))
+        x = x + (y.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * e[5]).flatten(1, 2)
+
+        return x, cache_update_info
+
+
+class CausalHead(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = WanLayerNorm(dim, eps)
+        self.head = ReplicatedLinear(dim, out_dim)
+        self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5)
+
+    def forward(self, x, e):
+        num_frames, frame_seqlen = e.shape[1], x.shape[1] // e.shape[1]
+        e = (self.modulation.unsqueeze(1) + e).chunk(2, dim=2)
+        x, _ = self.head(
+            self.norm(x).unflatten(dim=1, sizes=(num_frames, frame_seqlen))
+            * (1 + e[1]) + e[0])
+        return x
+
+
+class DreamXWorldARTransformer3DModel(BaseDiT):
+    """DreamX-World-5B autoregressive causal transformer."""
+
+    _fsdp_shard_conditions = DreamXWorldARConfig()._fsdp_shard_conditions
+    _compile_conditions = DreamXWorldARConfig()._compile_conditions
+    _supported_attention_backends = DreamXWorldARConfig()._supported_attention_backends
+    param_names_mapping = DreamXWorldARConfig().param_names_mapping
+    reverse_param_names_mapping = DreamXWorldARConfig().reverse_param_names_mapping
+    lora_param_names_mapping = DreamXWorldARConfig().lora_param_names_mapping
+    _no_split_modules = ["CausalWanAttentionBlock"]
+
+    def __init__(self, config: DreamXWorldARConfig, hf_config: dict[str, Any]) -> None:
+        super().__init__(config=config, hf_config=hf_config)
+
+        model_type = config.model_type
+        patch_size = config.patch_size
+        text_len = config.text_len
+        in_dim = config.in_channels
+        dim = config.hidden_size
+        ffn_dim = config.ffn_dim
+        freq_dim = config.freq_dim
+        text_dim = config.text_dim
+        out_dim = config.out_channels
+        num_heads = config.num_attention_heads
+        num_layers = config.num_layers
+        local_attn_size = config.local_attn_size
+        sink_size = config.sink_size
+        qk_norm = bool(config.qk_norm)
+        cross_attn_norm = config.cross_attn_norm
+        eps = config.eps
+        add_control_adapter = config.add_control_adapter
+        cam_method = config.cam_method
+        attn_compress = config.attn_compress
+        cam_self_attn_layers = config.cam_self_attn_layers
+
+        assert model_type in ['t2v', 'i2v', 'ti2v']
+        self.model_type = model_type
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.local_attn_size = local_attn_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings — nn.Linear inside nn.Sequential on purpose: the official
+        # checkpoint keys are positional (text_embedding.0/.2, time_embedding.0/.2,
+        # time_projection.1) and must load verbatim (see ffn comment above).
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size)
+        self.text_embedding = nn.Sequential(
+            nn.Linear(text_dim, dim), nn.GELU(approximate='tanh'),
+            nn.Linear(dim, dim))
+        self.time_embedding = nn.Sequential(
+            nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim))
+        self.time_projection = nn.Sequential(
+            nn.SiLU(), nn.Linear(dim, dim * 6))
+
+        # transformer blocks
+        self.blocks = nn.ModuleList([
+            CausalWanAttentionBlock(
+                dim, ffn_dim, num_heads, local_attn_size, sink_size,
+                qk_norm, cross_attn_norm, eps,
+                add_control_adapter=add_control_adapter,
+                cam_method=cam_method,
+                attn_compress=attn_compress,
+                layer_idx=layer_idx,
+                cam_self_attn_layers=cam_self_attn_layers)
+            for layer_idx in range(num_layers)
+        ])
+        for layer_idx, block in enumerate(self.blocks):
+            block.self_attn.layer_idx = layer_idx
+            block.self_attn.num_layers = self.num_layers
+
+        # head
+        self.head = CausalHead(dim, out_dim, patch_size, eps)
+
+        # RoPE frequencies
+        assert (dim % num_heads) == 0 and (dim // num_heads) % 2 == 0
+        d = dim // num_heads
+        self.freqs = torch.cat([
+            rope_params(1024, d - 4 * (d // 6)),
+            rope_params(1024, 2 * (d // 6)),
+            rope_params(1024, 2 * (d // 6))
+        ], dim=1)
+
+        self.num_attention_heads = num_heads
+        self.attention_head_dim = dim // num_heads
+        self.hidden_size = dim
+        self.in_channels = in_dim
+        self.out_channels = out_dim
+        self.num_channels_latents = out_dim
+        self.init_weights()
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        self.__post_init__()
+
+    def forward(self, x=None, t=None, context=None, seq_len=None, y=None, y_camera=None,
+                kv_cache=None, crossattn_cache=None, current_start=0,
+                cache_start=0, cache_update_policy="commit_detached",
+                hidden_states=None, encoder_hidden_states=None, timestep=None, **kwargs):
+        """
+        Causal inference with KV caching.
+        See Algorithm 2 of CausVid (https://arxiv.org/abs/2412.07772).
+
+        Args:
+            x: List of input video tensors [C_in, F, H, W]
+            t: Timestep tensor [B, L]
+            context: List of text embeddings [L, C]
+            seq_len: Maximum sequence length for positional encoding
+            y: Optional conditional video inputs (I2V mode)
+            y_camera: Camera parameters dict {'viewmats': ..., 'K': ...}
+            kv_cache: List of KV cache dicts per transformer block
+            crossattn_cache: List of cross-attention cache dicts
+            current_start: Current position in global token sequence
+            cache_start: Cache start position
+            cache_update_policy: Cache update strategy ('commit_detached' or 'none')
+
+        Returns:
+            Stacked output tensors [B, C_out, F, H/8, W/8]
+        """
+        if x is None and hidden_states is not None:
+            x = [sample for sample in hidden_states]
+        if t is None and timestep is not None:
+            t = timestep
+        if context is None and encoder_hidden_states is not None:
+            if isinstance(encoder_hidden_states, torch.Tensor):
+                context = [sample for sample in encoder_hidden_states]
+            else:
+                context = encoder_hidden_states
+        if seq_len is None:
+            if torch.is_tensor(t):
+                seq_len = int(t.shape[1]) if t.dim() > 1 else int(t.numel())
+            elif x is not None:
+                sample = x[0]
+                seq_len = (sample.shape[1] // self.patch_size[0]) * (sample.shape[2] // self.patch_size[1]) * (sample.shape[3] // self.patch_size[2])
+        if x is None or t is None or context is None or seq_len is None:
+            raise ValueError("DreamXWorldARTransformer3DModel requires x/t/context/seq_len or FastVideo aliases")
+
+        device = self.patch_embedding.weight.device
+        if self.freqs.is_meta or self.freqs.device != device:
+            d = self.dim // self.num_heads
+            self.freqs = torch.cat([
+                rope_params(1024, d - 4 * (d // 6)),
+                rope_params(1024, 2 * (d // 6)),
+                rope_params(1024, 2 * (d // 6)),
+            ], dim=1).to(device)
+
+        if y is not None:
+            x = [torch.cat([u, v], dim=0) for u, v in zip(x, y, strict=True)]
+
+        # patch embedding
+        x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
+        grid_sizes = torch.stack(
+            [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
+        x = [u.flatten(2).transpose(1, 2) for u in x]
+        seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
+        assert seq_lens.max() <= seq_len
+        x = torch.cat(x)
+
+        # time embedding
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).type_as(x))
+        e0 = self.time_projection(e).unflatten(
+            1, (6, self.dim)).unflatten(dim=0, sizes=t.shape)
+
+        # text embedding
+        context_lens = None
+        context = self.text_embedding(
+            torch.stack([
+                torch.cat(
+                    [u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
+                for u in context
+            ]))
+
+        # camera parameters
+        if y_camera is not None and isinstance(y_camera, dict):
+            cam_viewmats = y_camera['viewmats']
+            cam_K = y_camera['K']
+        else:
+            cam_viewmats = None
+            cam_K = None
+
+        block_kwargs = dict(
+            e=e0, seq_lens=seq_lens, grid_sizes=grid_sizes, freqs=self.freqs,
+            context=context, context_lens=context_lens,
+            cam_viewmats=cam_viewmats, cam_K=cam_K,
+            cache_update_policy=cache_update_policy,
+        )
+
+        cache_update_infos = []
+        for block_index, block in enumerate(self.blocks):
+            block_kwargs.update({
+                "kv_cache": kv_cache[block_index] if kv_cache is not None else None,
+                "crossattn_cache": crossattn_cache[block_index] if crossattn_cache is not None else None,
+                "current_start": current_start,
+                "cache_start": cache_start,
+            })
+            x, block_cache_update_info = block(x, **block_kwargs)
+            if kv_cache is not None:
+                cache_update_infos.append((block_index, block_cache_update_info))
+
+        # Apply deferred cache updates
+        if kv_cache is not None and cache_update_infos and cache_update_policy != "none":
+            self._apply_cache_updates(kv_cache, cache_update_infos)
+
+        # head & unpatchify
+        x = self.head(x, e.unflatten(dim=0, sizes=t.shape).unsqueeze(2))
+        x = self.unpatchify(x, grid_sizes)
+        return torch.stack(x)
+
+    def _apply_cache_updates(self, kv_cache, cache_update_infos):
+        """Apply deferred cache updates collected from all transformer blocks.
+
+        For Block-Relativistic RoPE, this stores un-roped K values in the cache.
+        RoPE is applied dynamically during attention based on each token's current
+        relative position in the sliding window.
+        """
+        with torch.no_grad():
+            for block_index, (current_end, local_end_index, update_info) in cache_update_infos:
+                if update_info is not None:
+                    cache = kv_cache[block_index]
+
+                    if update_info["action"] == "roll_and_insert":
+                        sink_tokens = update_info["sink_tokens"]
+                        num_rolled_tokens = update_info["num_rolled_tokens"]
+                        num_evicted_tokens = update_info["num_evicted_tokens"]
+                        write_start_index = update_info.get("write_start_index", update_info["local_start_index"])
+                        write_end_index = update_info.get("write_end_index", update_info["local_end_index"])
+                        new_k = update_info["new_k"].detach()
+                        new_v = update_info["new_v"].detach()
+
+                        cache["k"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                            cache["k"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+                        cache["v"][:, sink_tokens:sink_tokens + num_rolled_tokens] = \
+                            cache["v"][:, sink_tokens + num_evicted_tokens:sink_tokens + num_evicted_tokens + num_rolled_tokens].clone()
+
+                        if write_end_index > write_start_index and new_k.shape[1] == (write_end_index - write_start_index):
+                            cache["k"][:, write_start_index:write_end_index] = new_k
+                            cache["v"][:, write_start_index:write_end_index] = new_v
+
+                    elif update_info["action"] == "direct_insert":
+                        write_start_index = update_info.get("write_start_index", update_info["local_start_index"])
+                        write_end_index = update_info.get("write_end_index", update_info["local_end_index"])
+                        new_k = update_info["new_k"].detach()
+                        new_v = update_info["new_v"].detach()
+
+                        if write_end_index > write_start_index and new_k.shape[1] == (write_end_index - write_start_index):
+                            cache["k"][:, write_start_index:write_end_index] = new_k
+                            cache["v"][:, write_start_index:write_end_index] = new_v
+
+                is_recompute = False if update_info is None else update_info.get("is_recompute", False)
+                if not is_recompute:
+                    kv_cache[block_index]["global_end_index"].fill_(current_end)
+                    kv_cache[block_index]["local_end_index"].fill_(local_end_index)
+
+    def unpatchify(self, x, grid_sizes):
+        """Reconstruct video tensors from patch embeddings."""
+        c = self.out_dim
+        out = []
+        for u, v in zip(x, grid_sizes.tolist(), strict=True):
+            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
+            u = torch.einsum('fhwpqrc->cfphqwr', u)
+            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size, strict=True)])
+            out.append(u)
+        return out
+
+    def init_weights(self):
+        """Initialize model parameters using Xavier initialization."""
+        for m in self.modules():
+            if isinstance(m, (nn.Linear, ReplicatedLinear)):
+                nn.init.xavier_uniform_(m.weight)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+        nn.init.xavier_uniform_(self.patch_embedding.weight.flatten(1))
+        for m in self.text_embedding.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, std=.02)
+        for m in self.time_embedding.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, std=.02)
+
+        nn.init.zeros_(self.head.head.weight)
+
+
+EntryClass = DreamXWorldARTransformer3DModel

fastvideo/models/dits/hunyuanvideo15.py

@@ -699,8 +699,8 @@ class IndividualTokenRefinerBlock(nn.Module):
         if mask is None:
             attn_output = self.attn(q, k, v) 
         else:
-            from fastvideo.attention.backends.flash_attn import FlashAttnMetadataBuilder
-            attn_metadata = FlashAttnMetadataBuilder().build(
+            from fastvideo.attention.backends.sdpa import SDPAMetadataBuilder
+            attn_metadata = SDPAMetadataBuilder().build(
                 current_timestep=0,
                 attn_mask=mask,
             )

fastvideo/models/dits/hyworld/hyworld.py

@@ -173,8 +173,11 @@ class HYWorldDoubleStreamBlock(MMDoubleStreamBlock):
         img_v_prope = img_v_prope.permute(0, 2, 1, 3)  # [batch, seqlen, num_heads, head_dim]
         # end hyworld
 
-        from fastvideo.attention.backends.flash_attn import FlashAttnMetadataBuilder
-        attn_metadata = FlashAttnMetadataBuilder().build(
+        # The metadata only carries the text padding mask; the executing
+        # attention kernel is still whatever the layer's selector picked
+        # (flash-attn when installed), not necessarily torch SDPA.
+        from fastvideo.attention.backends.sdpa import SDPAMetadataBuilder
+        attn_metadata = SDPAMetadataBuilder().build(
             current_timestep=0,
             attn_mask=encoder_attention_mask,
         )
@@ -192,14 +195,9 @@ class HYWorldDoubleStreamBlock(MMDoubleStreamBlock):
             )
         
         # begin hyworld
-        # attention with prope
-        from fastvideo.attention.backends.flash_attn import FlashAttnMetadataBuilder
-        attn_metadata_prope = FlashAttnMetadataBuilder().build(
-            current_timestep=0,
-            attn_mask=encoder_attention_mask,
-        )       
+        # attention with prope (same text mask, so reuse the metadata)
         # NOTE: Do NOT pass freqs_cis to prope attention - HY-WorldPlay does not apply RoPE to prope
-        with set_forward_context(current_timestep=0, attn_metadata=attn_metadata_prope):
+        with set_forward_context(current_timestep=0, attn_metadata=attn_metadata):
             img_attn_prope, _ = self.attn(
                 img_q_prope,
                 img_k_prope,

fastvideo/models/dits/kandinsky5.py

@@ -10,13 +10,21 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from fastvideo.attention import LocalAttention
+from fastvideo.attention.backends.nabla import CAN_USE_FLEX_ATTN, flex_attention, nablaT_v2
 from fastvideo.configs.models.dits import Kandinsky5VideoConfig
 from fastvideo.layers.layernorm import LayerNormScaleShift
 from fastvideo.layers.linear import ReplicatedLinear
 from fastvideo.layers.mlp import MLP
+from fastvideo.logger import init_logger
 from fastvideo.models.dits.base import BaseDiT
 from fastvideo.platforms import AttentionBackendEnum
 
+logger = init_logger(__name__)
+
+if not CAN_USE_FLEX_ATTN:
+    logger.warning("torch.nn.attention.flex_attention is unavailable in this PyTorch build; "
+                   "Kandinsky5 NABLA sparse attention (Pro checkpoints) cannot be used.")
+
 FRACTAL_PIXEL_SIZE = 8
 _ARCH_CONFIG_DEFAULTS = Kandinsky5VideoConfig().arch_config
 
@@ -263,10 +271,9 @@ class Kandinsky5Modulation(nn.Module):
 
 
 def _apply_rotary(x: torch.Tensor, rope: torch.Tensor) -> torch.Tensor:
-    orig_dtype = x.dtype
     x_ = x.reshape(*x.shape[:-1], -1, 1, 2).to(torch.float32)
     x_out = (rope * x_).sum(dim=-1)
-    return x_out.reshape(*x.shape).to(orig_dtype)
+    return x_out.reshape(*x.shape).to(x.dtype)
 
 
 class Kandinsky5Attention(nn.Module):
@@ -277,6 +284,7 @@ class Kandinsky5Attention(nn.Module):
         head_dim: int,
         supported_attention_backends: tuple[AttentionBackendEnum, ...] | None,
         prefix: str = "",
+        use_nabla: bool = False,
     ):
         super().__init__()
         assert num_channels % head_dim == 0
@@ -306,6 +314,17 @@ class Kandinsky5Attention(nn.Module):
             causal=False,
             supported_attention_backends=supported_attention_backends,
         )
+        # NABLA checkpoints get a second attention layer whose backend defaults
+        # to NABLA_ATTN; FASTVIDEO_ATTENTION_BACKEND still overrides it.
+        self.nabla_attention = None
+        if use_nabla:
+            self.nabla_attention = LocalAttention(
+                num_heads=self.num_heads,
+                head_size=head_dim,
+                causal=False,
+                supported_attention_backends=supported_attention_backends,
+                default_backend=AttentionBackendEnum.NABLA_ATTN,
+            )
 
     def forward(
         self,
@@ -339,27 +358,54 @@ class Kandinsky5Attention(nn.Module):
             key = _apply_rotary(key, rotary_emb).type_as(key)
 
         if sparse_params is not None:
-            raise NotImplementedError(
-                "Sparse attention is not yet supported for Kandinsky5 in FastVideo."
-            )
+            if self.nabla_attention is None:
+                raise RuntimeError("sparse_params passed to an attention layer built without use_nabla; "
+                                   "this checkpoint/config combination is inconsistent.")
+            try:
+                # Backend impl reads sta_mask/P from the forward-context
+                # attention metadata built by the denoising stage.
+                hidden_states = self.nabla_attention(query, key, value)
+            except AssertionError as exc:
+                # Standalone parity tests call the model without a pipeline
+                # forward context; run the NABLA kernel directly.
+                if "Forward context is not set" not in str(exc):
+                    raise
+                attn_mask = nablaT_v2(query, key, sparse_params["sta_mask"], thr=sparse_params["P"])
+                hidden_states = flex_attention(
+                    query=query.transpose(1, 2),
+                    key=key.transpose(1, 2),
+                    value=value.transpose(1, 2),
+                    block_mask=attn_mask,
+                ).transpose(1, 2)
+        else:
+            try:
+                hidden_states = self.local_attention(query, key, value)
 
-        try:
-            hidden_states = self.local_attention(query, key, value)
-        except AssertionError as exc:
-            # LocalAttention requires pipeline forward context. Standalone
-            # parity tests call the model directly, so fallback to Torch SDPA.
-            if "Forward context is not set" not in str(exc):
-                raise
-            query = query.transpose(1, 2)
-            key = key.transpose(1, 2)
-            value = value.transpose(1, 2)
-            hidden_states = F.scaled_dot_product_attention(query,
-                                                           key,
-                                                           value,
-                                                           attn_mask=None,
-                                                           is_causal=False)
-            hidden_states = hidden_states.transpose(1, 2)
-        hidden_states = hidden_states.flatten(2)
+            except AssertionError as exc:
+                # LocalAttention requires pipeline forward context. Standalone
+                # parity tests call the model directly, so fallback to Torch SDPA.
+                if "Forward context is not set" not in str(exc):
+                    raise
+
+                query_shape = query.shape[:-2]
+                key_shape = key.shape[:-2]
+                query = query.reshape(query_shape[0], -1, self.num_heads,
+                                      query.shape[-1]).transpose(1, 2)
+                key = key.reshape(key_shape[0], -1, self.num_heads,
+                                  key.shape[-1]).transpose(1, 2)
+                value = value.reshape(key_shape[0], -1, self.num_heads,
+                                      value.shape[-1]).transpose(1, 2)
+                hidden_states = F.scaled_dot_product_attention(
+                    query,
+                    key,
+                    value,
+                    attn_mask=None,
+                    is_causal=False,
+                )
+                hidden_states = hidden_states.transpose(1, 2).reshape(
+                    *query_shape, self.num_heads, -1)
+                
+        hidden_states = hidden_states.flatten(-2, -1)
 
         hidden_states, _ = self.out_layer(hidden_states)
         return hidden_states
@@ -476,7 +522,8 @@ class Kandinsky5TransformerDecoderBlock(nn.Module):
                  head_dim: int,
                  supported_attention_backends: tuple[AttentionBackendEnum, ...]
                  | None = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 use_nabla: bool = False):
         super().__init__()
         self.visual_modulation = Kandinsky5Modulation(time_dim, model_dim, 9)
 
@@ -491,7 +538,8 @@ class Kandinsky5TransformerDecoderBlock(nn.Module):
             model_dim,
             head_dim,
             supported_attention_backends=supported_attention_backends,
-            prefix=f"{prefix}.self_attention")
+            prefix=f"{prefix}.self_attention",
+            use_nabla=use_nabla)
 
         self.cross_attention_norm = LayerNormScaleShift(
             model_dim,
@@ -624,7 +672,8 @@ class Kandinsky5Transformer3DModel(BaseDiT):
                                               arch.ff_dim,
                                               head_dim,
                                               self._supported_attention_backends,
-                                              prefix=f"{config.prefix}.visual_transformer_blocks.{i}")
+                                              prefix=f"{config.prefix}.visual_transformer_blocks.{i}",
+                                              use_nabla=arch.attention_type == "nabla")
             for i in range(arch.num_visual_blocks)
         ])
 
@@ -694,6 +743,7 @@ class Kandinsky5Transformer3DModel(BaseDiT):
                                                   scale_factor)
         to_fractal = sparse_params[
             "to_fractal"] if sparse_params is not None else False
+        
         visual_embed, visual_rope = fractal_flatten(visual_embed, visual_rope,
                                                     visual_shape,
                                                     block_mask=to_fractal)
@@ -724,6 +774,7 @@ class Kandinsky5Transformer3DModel(BaseDiT):
 
         if return_dict:
             return Kandinsky5TransformerOutput(sample=x)
+        
         return x
 
     def materialize_non_persistent_buffers(self, device: torch.device,

fastvideo/models/encoders/qwen2_5_vl_custom.py

@@ -467,6 +467,18 @@ class Qwen2_5_VisionTransformerPretrainedModel(nn.Module):
         return hidden_states
 
 
+def _compute_default_rope_parameters(config, device=None, seq_len=None, **kwargs):
+    # transformers>=5 removes the "default" entry from ROPE_INIT_FUNCTIONS and
+    # moves rope_theta inside rope_parameters; replicate the 4.x default init.
+    rope_params = getattr(config, "rope_parameters", None) or getattr(config, "rope_scaling", None) or {}
+    base = rope_params.get("rope_theta", getattr(config, "rope_theta", 10000.0))
+    head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+    partial_rotary_factor = getattr(config, "partial_rotary_factor", 1.0)
+    dim = int(head_dim * partial_rotary_factor)
+    inv_freq = 1.0 / (base**(torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
+    return inv_freq, 1.0
+
+
 class Qwen2_5_VLRotaryEmbedding(nn.Module):
     def __init__(self, config: Qwen2_5_VLConfig, device=None):
         super().__init__()
@@ -479,7 +491,14 @@ class Qwen2_5_VLRotaryEmbedding(nn.Module):
         self.original_max_seq_len = config.max_position_embeddings
 
         self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        if self.rope_type in ROPE_INIT_FUNCTIONS:
+            self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        elif self.rope_type == "default":
+            # transformers>=5 drops the "default" entry from ROPE_INIT_FUNCTIONS.
+            self.rope_init_fn = _compute_default_rope_parameters
+        else:
+            raise KeyError(f"Unsupported rope_type '{self.rope_type}'; available: "
+                           f"{['default', *ROPE_INIT_FUNCTIONS]}")
 
         inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
         self.register_buffer("inv_freq", inv_freq, persistent=False)
@@ -948,6 +967,16 @@ QWEN2_5_VL_ATTENTION_CLASSES = {
 # If FlashAttention2 is not available, transparently fall back to SDPA.
 if not is_flash_attn_2_available():
     QWEN2_5_VL_ATTENTION_CLASSES["flash_attention_2"] = Qwen2_5_VLSdpaAttention
+else:
+    # transformers>=5 only resolves the flash-attn functions when the model
+    # preloads them via its attention interface; this module bypasses
+    # PreTrainedModel, so _flash_attention_forward(implementation=None) raises
+    # unless we preload here.
+    try:
+        from transformers.modeling_flash_attention_utils import lazy_import_flash_attention
+        lazy_import_flash_attention("flash_attention_2")
+    except (ImportError, ValueError):
+        QWEN2_5_VL_ATTENTION_CLASSES["flash_attention_2"] = Qwen2_5_VLSdpaAttention
     
 class Qwen2_5_VLDecoderLayer(nn.Module):
     def __init__(self, config: Qwen2_5_VLConfig, layer_idx: int):
@@ -1035,7 +1064,7 @@ class Qwen2_5_VLModel(nn.Module):
     def __init__(self, config: Qwen2_5_VLConfig):
         super().__init__()
         self.config = config
-        self.padding_idx = config.pad_token_id
+        self.padding_idx = getattr(config, "pad_token_id", None)
         self.vocab_size = config.vocab_size
 
         self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
@@ -1408,6 +1437,18 @@ class Qwen2_5_VLCausalLMOutputWithPast(ModelOutput):
     rope_deltas: Optional[torch.LongTensor] = None
 
 
+def _flatten_text_config(config):
+    # transformers>=5 stops forwarding text-model attributes (hidden_size,
+    # vocab_size, rope_scaling, ...) from the composite Qwen2_5_VLConfig to
+    # config.text_config; this module reads them from the top level.
+    text_config = getattr(config, "text_config", None)
+    if text_config is not None:
+        for key, value in text_config.to_dict().items():
+            if not hasattr(config, key):
+                setattr(config, key, value)
+    return config
+
+
 class Qwen2_5_VLForConditionalGenerationSimple(nn.Module):
     _tied_weights_keys = ["lm_head.weight"]
     config_class = Qwen2_5_VLConfig
@@ -1415,6 +1456,7 @@ class Qwen2_5_VLForConditionalGenerationSimple(nn.Module):
 
     def __init__(self, config):
         super().__init__()
+        config = _flatten_text_config(config)
         self.config = config
         self.visual = Qwen2_5_VisionTransformerPretrainedModel(config.vision_config)
 

fastvideo/models/loader/component_loader.py

@@ -320,6 +320,19 @@ class TextEncoderLoader(ComponentLoader):
                 f"text encoder index {idx} out of range for text_encoder_configs (len={len(encoder_configs)}), model_path={model_path}"
             )
         encoder_config = encoder_configs[idx]
+        if (
+            model_config.get("architectures") == ["CLIPModel"]
+            and isinstance(model_config.get("text_config"), dict)
+        ):
+            valid_arch_fields = {
+                f.name for f in dataclasses.fields(encoder_config.arch_config)
+            }
+            model_config = {
+                key: value
+                for key, value in deepcopy(model_config["text_config"]).items()
+                if key in valid_arch_fields
+            }
+            model_config["architectures"] = ["CLIPTextModel"]
         encoder_config.update_model_arch(model_config)
         if idx < 0 or idx >= len(encoder_precisions):
             raise IndexError(
@@ -404,7 +417,9 @@ class TextEncoderLoader(ComponentLoader):
             else:
                 loaded_weights: set[str] = model.load_weights(
                     self._get_all_weights(
-                        model, model_path, to_cpu=use_cpu_offload
+                        model,
+                        model_path,
+                        to_cpu=fastvideo_args.text_encoder_cpu_offload,
                     )
                 )  # type: ignore
 

fastvideo/models/registry.py

@@ -11,7 +11,7 @@ import tempfile
 from abc import ABC, abstractmethod
 from collections.abc import Callable, Set
 from dataclasses import dataclass, field
-from functools import lru_cache
+from functools import cache, lru_cache
 from typing import NoReturn, TypeVar, cast
 
 import cloudpickle
@@ -32,6 +32,8 @@ _TEXT_TO_VIDEO_DIT_MODELS = {
     "HYWorldTransformer3DModel":
     ("dits", "hyworld", "HYWorldTransformer3DModel"),
     "WanTransformer3DModel": ("dits", "wanvideo", "WanTransformer3DModel"),
+    "DreamXWorldTransformer3DModel": ("dits", "dreamx_world", "DreamXWorldTransformer3DModel"),
+    "DreamXWorldARTransformer3DModel": ("dits", "dreamx_world_ar", "DreamXWorldARTransformer3DModel"),
     "CausalWanTransformer3DModel": ("dits", "causal_wanvideo", "CausalWanTransformer3DModel"),
     "CosmosTransformer3DModel": ("dits", "cosmos", "CosmosTransformer3DModel"),
     "Cosmos25Transformer3DModel": ("dits", "cosmos2_5", "Cosmos25Transformer3DModel"),
@@ -48,6 +50,8 @@ _TEXT_TO_VIDEO_DIT_MODELS = {
 _IMAGE_TO_VIDEO_DIT_MODELS = {
     # "HunyuanVideoTransformer3DModel": ("dits", "hunyuanvideo", "HunyuanVideoDiT"),
     "WanTransformer3DModel": ("dits", "wanvideo", "WanTransformer3DModel"),
+    "DreamXWorldTransformer3DModel": ("dits", "dreamx_world", "DreamXWorldTransformer3DModel"),
+    "DreamXWorldARTransformer3DModel": ("dits", "dreamx_world_ar", "DreamXWorldARTransformer3DModel"),
     "CausalWanTransformer3DModel": ("dits", "causal_wanvideo", "CausalWanTransformer3DModel"),
     "MatrixGame2WanModel": ("dits", "matrixgame2", "MatrixGame2WanModel"),
     "CausalMatrixGame2WanModel": ("dits", "matrixgame2", "CausalMatrixGame2WanModel"),
@@ -141,7 +145,7 @@ _LEGACY_FAST_VIDEO_MODELS = {
 MODELS_PATH = os.path.dirname(__file__)
 
 
-@lru_cache(maxsize=None)
+@cache
 def _discover_and_register_models() -> dict[str, tuple[str, str, str]]:
     discovered_models: dict[str, tuple[str, str, str]] = {}
     for root, dirs, files in os.walk(MODELS_PATH):
@@ -156,7 +160,7 @@ def _discover_and_register_models() -> dict[str, tuple[str, str, str]]:
 
             filepath = os.path.join(root, filename)
             try:
-                with open(filepath, "r", encoding="utf-8") as f:
+                with open(filepath, encoding="utf-8") as f:
                     source = f.read()
                 tree = ast.parse(source, filename=filename)
 

fastvideo/performance/hf_store.py

@@ -268,16 +268,31 @@ def load_records_for_model(
     model_id: str,
     gpu_type: str | None = None,
     *,
+    workload_id: str | None = None,
+    variant_id: str | None = None,
+    benchmark_version: str | None = None,
+    recipe_fingerprint: str | None = None,
+    hardware_profile_id: str | None = None,
+    software_profile_id: str | None = None,
     last_n: int | None = None,
     successful_only: bool = True,
     baseline_eligible_only: bool = False,
 ) -> list[dict[str, Any]]:
-    """Return records for a specific *model_id*, optionally filtered by GPU.
+    """Return records for a specific *model_id*, optionally filtered by cohort.
 
     Args:
         local_dir: Root directory previously populated by :func:`sync_from_hf`.
         model_id: Matches the ``model_id`` field inside each JSON record.
         gpu_type: When set, only records whose ``gpu_type`` matches are returned.
+        workload_id: When set, only records from the same workload are returned.
+        variant_id: When set, only records from the same workload variant are returned.
+        benchmark_version: When set, only records from the same benchmark version are returned.
+        recipe_fingerprint: When set, only records from the same benchmark
+            recipe are returned.
+        hardware_profile_id: When set, only records from the same hardware
+            cohort are returned.
+        software_profile_id: When set, only records from the same software
+            cohort are returned.
         last_n: When set, return only the most recent *n* records (after all
             other filters). Useful for sliding-window baseline calculations.
         successful_only: Passed through to :func:`load_records`.
@@ -299,6 +314,18 @@ def load_records_for_model(
     if gpu_type is not None:
         records = [r for r in records if r.get("gpu_type") == gpu_type]
 
+    identity_filters = {
+        "workload_id": workload_id,
+        "variant_id": variant_id,
+        "benchmark_version": benchmark_version,
+        "recipe_fingerprint": recipe_fingerprint,
+        "hardware_profile_id": hardware_profile_id,
+        "software_profile_id": software_profile_id,
+    }
+    for key, expected in identity_filters.items():
+        if expected is not None:
+            records = [r for r in records if str(r.get(key)) == str(expected)]
+
     if last_n is not None:
         records = records[-last_n:]
 

fastvideo/performance_dashboard/service.py

@@ -17,6 +17,15 @@ from fastvideo.performance.hf_store import is_baseline_eligible_record, safe_flo
 from fastvideo.performance.metric_policy import regression_delta, resolve_metric_policies
 
 Record = dict[str, Any]
+CohortKey = tuple[str, str, str, str, str, str, str, str]
+COMPARISON_COHORT_KEYS = (
+    "workload_id",
+    "variant_id",
+    "benchmark_version",
+    "recipe_fingerprint",
+    "hardware_profile_id",
+    "software_profile_id",
+)
 
 
 def parse_timestamp(value: Any) -> datetime | None:
@@ -77,15 +86,44 @@ def record_metadata(record: Record) -> Record:
     }
 
 
-def group_by_model_gpu(records: list[Record]) -> dict[tuple[str, str], list[Record]]:
-    groups: dict[tuple[str, str], list[Record]] = defaultdict(list)
+def record_comparison_metadata(record: Record) -> Record:
+    return {key: record.get(key) or "" for key in COMPARISON_COHORT_KEYS}
+
+
+def comparison_cohort_key(record: Record) -> CohortKey:
+    return (
+        str(record.get("model_id") or "unknown"),
+        str(record.get("gpu_type") or "unknown"),
+        str(record.get("workload_id") or ""),
+        str(record.get("variant_id") or ""),
+        str(record.get("benchmark_version") or ""),
+        str(record.get("recipe_fingerprint") or ""),
+        str(record.get("hardware_profile_id") or ""),
+        str(record.get("software_profile_id") or ""),
+    )
+
+
+def group_by_comparison_cohort(records: list[Record]) -> dict[CohortKey, list[Record]]:
+    groups: dict[CohortKey, list[Record]] = defaultdict(list)
     for record in records:
-        model_id = str(record.get("model_id") or "unknown")
-        gpu_type = str(record.get("gpu_type") or "unknown")
-        groups[(model_id, gpu_type)].append(record)
+        groups[comparison_cohort_key(record)].append(record)
     return {key: sorted(value, key=record_sort_key) for key, value in groups.items()}
 
 
+def comparison_sort_key(record: Record) -> CohortKey:
+    return comparison_cohort_key(record)
+
+
+def latest_row_sort_key(row: Record) -> tuple[Any, ...]:
+    return (row["status"] != "fail", *comparison_sort_key(row))
+
+
+def group_identity(key: CohortKey) -> tuple[str, str]:
+    model_id = key[0]
+    gpu_type = key[1]
+    return model_id, gpu_type
+
+
 def baseline_value(records: list[Record], metric_key: str) -> float | None:
     values = [safe_float(record.get(metric_key)) for record in records]
     values = [value for value in values if value is not None]
@@ -99,7 +137,8 @@ def build_latest_summary(records: list[Record],
                          baseline_window: int = 5,
                          run_source: str | None = None) -> list[Record]:
     rows: list[Record] = []
-    for (model_id, gpu_type), group in group_by_model_gpu(records).items():
+    for key, group in group_by_comparison_cohort(records).items():
+        model_id, gpu_type = group_identity(key)
         latest_candidates = group
         if run_source:
             latest_candidates = [record for record in group if record_run_source(record) == run_source]
@@ -156,6 +195,7 @@ def build_latest_summary(records: list[Record],
             "timestamp": latest.get("timestamp"),
             "commit_sha": latest.get("commit_sha"),
             **record_metadata(latest),
+            **record_comparison_metadata(latest),
             "success": success,
             "baseline_n": len(baseline_records),
             "worst_regression_pct": worst_regression,
@@ -166,12 +206,13 @@ def build_latest_summary(records: list[Record],
             "metrics": metrics,
         })
 
-    return sorted(rows, key=lambda row: (row["status"] != "fail", row["model_id"], row["gpu_type"]))
+    return sorted(rows, key=latest_row_sort_key)
 
 
 def build_trends(records: list[Record]) -> list[Record]:
     trends: list[Record] = []
-    for (model_id, gpu_type), group in group_by_model_gpu(records).items():
+    for key, group in group_by_comparison_cohort(records).items():
+        model_id, gpu_type = group_identity(key)
         points = []
         for record in group:
             metric_policies = resolve_metric_policies(record.get("regression_thresholds"))
@@ -179,6 +220,7 @@ def build_trends(records: list[Record]) -> list[Record]:
                 "timestamp": record.get("timestamp"),
                 "commit_sha": record.get("commit_sha"),
                 **record_metadata(record),
+                **record_comparison_metadata(record),
                 "success": bool(record.get("success", True)),
                 "metrics": {
                     policy.key: safe_float(record.get(policy.key))
@@ -189,6 +231,7 @@ def build_trends(records: list[Record]) -> list[Record]:
         trends.append({
             "model_id": model_id,
             "gpu_type": gpu_type,
+            **record_comparison_metadata(group[-1]),
             "points": points,
         })
-    return sorted(trends, key=lambda trend: (trend["model_id"], trend["gpu_type"]))
+    return sorted(trends, key=comparison_sort_key)

fastvideo/pipelines/basic/dreamx_world/__init__.py

@@ -0,0 +1,28 @@
+# SPDX-License-Identifier: Apache-2.0
+from fastvideo.configs.pipelines.dreamx_world import (
+    DreamXWorld5BARPipelineConfig,
+    DreamXWorld5BCamPipelineConfig,
+    make_dreamx_world_5b_ar_dit_config,
+    make_dreamx_world_5b_cam_dit_config,
+    make_dreamx_world_5b_cam_text_encoder_config,
+    make_dreamx_world_5b_cam_vae_config,
+)
+from fastvideo.pipelines.basic.dreamx_world.dreamx_world_ar_pipeline import DreamXWorldARPipeline
+from fastvideo.pipelines.basic.dreamx_world.dreamx_world_pipeline import DreamXWorldPipeline
+from fastvideo.pipelines.basic.dreamx_world.stages import (
+    DREAMX_Y_CAMERA_KEY,
+    DreamXWorldCameraConditioningStage,
+)
+
+__all__ = [
+    "DREAMX_Y_CAMERA_KEY",
+    "DreamXWorld5BARPipelineConfig",
+    "DreamXWorld5BCamPipelineConfig",
+    "DreamXWorldCameraConditioningStage",
+    "DreamXWorldARPipeline",
+    "DreamXWorldPipeline",
+    "make_dreamx_world_5b_ar_dit_config",
+    "make_dreamx_world_5b_cam_dit_config",
+    "make_dreamx_world_5b_cam_text_encoder_config",
+    "make_dreamx_world_5b_cam_vae_config",
+]

fastvideo/pipelines/basic/dreamx_world/ar_denoising.py

@@ -0,0 +1,219 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World autoregressive causal denoising stage."""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+from tqdm.auto import tqdm
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.basic.dreamx_world.stages import DREAMX_Y_CAMERA_KEY
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.stages.denoising import DenoisingStage
+
+
+class DreamXWorldARCausalDenoisingStage(DenoisingStage):
+    """Official DreamX AR-forcing denoising loop with KV cache."""
+
+    _AR_NOISE_SEED_OFFSET = 1_000_003
+
+    def __init__(self, transformer, scheduler, pipeline=None, vae=None) -> None:
+        super().__init__(transformer=transformer, scheduler=scheduler, pipeline=pipeline, vae=vae)
+        self.num_transformer_blocks = len(self.transformer.blocks)
+        self.num_frame_per_block = int(getattr(self.transformer, "num_frame_per_block", 3))
+        self.local_attn_size = int(getattr(self.transformer, "local_attn_size", 12))
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        assert batch.latents is not None, "latents must be prepared before DreamX AR denoising"
+        assert batch.prompt_embeds, "prompt embeds must be prepared before DreamX AR denoising"
+        latents = batch.latents
+        device = latents.device
+        target_dtype = torch.bfloat16
+        autocast_enabled = device.type == "cuda" and not fastvideo_args.disable_autocast
+
+        frame_seq_length = (latents.shape[-2] // self.transformer.patch_size[1]) * (latents.shape[-1] //
+                                                                                    self.transformer.patch_size[2])
+        timesteps = torch.tensor(
+            tuple(getattr(fastvideo_args.pipeline_config, "dmd_denoising_steps", (1000, 750, 500, 250))),
+            dtype=torch.long,
+        ).cpu()
+        if getattr(fastvideo_args.pipeline_config, "warp_denoising_step", True):
+            self.scheduler.set_timesteps(1000)
+            scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(), torch.tensor([0], dtype=torch.float32)))
+            timesteps = scheduler_timesteps[1000 - timesteps]
+        timesteps = timesteps.to(device)
+
+        if latents.shape[2] % self.num_frame_per_block != 0:
+            raise ValueError("DreamX AR latent frames must be divisible by num_frame_per_block")
+
+        y_camera = batch.extra.get(DREAMX_Y_CAMERA_KEY, batch.extra.get("y_camera"))
+        if isinstance(y_camera, dict):
+            y_camera = {
+                k: v.to(device=device, dtype=target_dtype) if torch.is_tensor(v) else v
+                for k, v in y_camera.items()
+            }
+
+        if batch.image_latent is not None and batch.image_latent.shape[1] == latents.shape[1]:
+            latents[:, :, :batch.image_latent.shape[2]] = batch.image_latent.to(device=device, dtype=latents.dtype)
+
+        kv_cache = self._initialize_kv_cache(latents.shape[0], target_dtype, device, frame_seq_length)
+        crossattn_cache = self._initialize_crossattn_cache(latents.shape[0], target_dtype, device)
+        prompt = batch.prompt_embeds[0]
+        if torch.is_tensor(prompt):
+            prompt = prompt.to(device=device, dtype=target_dtype)
+            context = [sample for sample in prompt]
+        else:
+            context = prompt
+
+        num_blocks = latents.shape[2] // self.num_frame_per_block
+        start = 0
+        first_frame_mask = torch.ones_like(latents)
+        first_frame_mask[:, :, 0] = 0
+        base_generator = batch.generator[0] if isinstance(batch.generator, list) else batch.generator
+        noise_generator = self._make_noise_generator(base_generator, device)
+
+        with tqdm(total=num_blocks * len(timesteps), desc="DreamX AR denoising", leave=False) as progress:
+            for _ in range(num_blocks):
+                current_num_frames = self.num_frame_per_block
+                block_latents = latents[:, :, start:start + current_num_frames]
+                noisy_input = block_latents.clone()
+                mask_block = first_frame_mask[:, :, start:start + current_num_frames]
+                camera_block = self._slice_camera(y_camera, start, current_num_frames)
+
+                for idx, current_timestep in enumerate(timesteps):
+                    timestep = torch.full(
+                        (latents.shape[0], current_num_frames * frame_seq_length),
+                        int(current_timestep.item()),
+                        device=device,
+                        dtype=torch.long,
+                    )
+                    if start == 0:
+                        timestep[:, :frame_seq_length] = 0
+                    with torch.autocast(device_type="cuda", dtype=target_dtype, enabled=autocast_enabled):
+                        denoised = self.transformer(
+                            hidden_states=block_latents.to(target_dtype),
+                            encoder_hidden_states=torch.stack(context).to(target_dtype),
+                            timestep=timestep,
+                            y_camera=camera_block,
+                            kv_cache=kv_cache,
+                            crossattn_cache=crossattn_cache,
+                            current_start=start * frame_seq_length,
+                        )
+                    denoised = denoised.to(latents.dtype)
+                    if idx < len(timesteps) - 1:
+                        next_timestep = torch.full((latents.shape[0], current_num_frames),
+                                                   int(timesteps[idx + 1].item()),
+                                                   device=device,
+                                                   dtype=torch.long)
+                        noise_kwargs = {"device": device, "dtype": denoised.dtype}
+                        if noise_generator is not None:
+                            noise_kwargs["generator"] = noise_generator
+                        noise = torch.randn(denoised.permute(0, 2, 1, 3, 4).shape, **noise_kwargs)
+                        block_btchw = self.scheduler.add_noise(
+                            denoised.permute(0, 2, 1, 3, 4).flatten(0, 1),
+                            noise.flatten(0, 1),
+                            next_timestep.flatten(),
+                        ).unflatten(0, (latents.shape[0], current_num_frames))
+                        block_latents = block_btchw.permute(0, 2, 1, 3, 4)
+                        block_latents = block_latents * mask_block + noisy_input * (1 - mask_block)
+                    else:
+                        block_latents = denoised * mask_block + noisy_input * (1 - mask_block)
+                    progress.update()
+
+                latents[:, :, start:start + current_num_frames] = block_latents
+                self._update_context_cache(block_latents, context, camera_block, kv_cache, crossattn_cache, start,
+                                           frame_seq_length, target_dtype, autocast_enabled,
+                                           float(getattr(fastvideo_args.pipeline_config, "context_noise", 0.1)))
+                start += current_num_frames
+
+        batch.latents = latents
+        return batch
+
+    def _make_noise_generator(self, generator: torch.Generator | None, device: torch.device) -> torch.Generator | None:
+        if generator is None:
+            return None
+        if getattr(generator, "device", None) == device:
+            return generator
+        seed = int(generator.initial_seed()) + self._AR_NOISE_SEED_OFFSET
+        return torch.Generator(device=device).manual_seed(seed)
+
+    @staticmethod
+    def _context_noise_timestep(context_noise: float) -> int:
+        if 0.0 < context_noise <= 1.0:
+            return int(context_noise * 1000)
+        return int(context_noise)
+
+    def _slice_camera(self, y_camera: Any, start: int, num_frames: int):
+        if not isinstance(y_camera, dict):
+            return y_camera
+        return {
+            "viewmats": y_camera["viewmats"][:, start:start + num_frames],
+            "K": y_camera["K"][:, start:start + num_frames],
+        }
+
+    def _initialize_kv_cache(self, batch_size: int, dtype: torch.dtype, device: torch.device,
+                             frame_seq_length: int) -> list[dict[str, Any]]:
+        size = self.local_attn_size * frame_seq_length if self.local_attn_size != -1 else 18480
+        heads = self.transformer.num_attention_heads
+        head_dim = self.transformer.attention_head_dim
+        cam_self_attn = next(
+            (getattr(block, "cam_self_attn", None)
+             for block in self.transformer.blocks if getattr(block, "cam_self_attn", None) is not None),
+            None,
+        )
+
+        caches = []
+        for _ in range(self.num_transformer_blocks):
+            cache = {
+                "k": torch.zeros(batch_size, size, heads, head_dim, dtype=dtype, device=device),
+                "v": torch.zeros(batch_size, size, heads, head_dim, dtype=dtype, device=device),
+                "global_end_index": torch.tensor([0], dtype=torch.long, device=device),
+                "local_end_index": torch.tensor([0], dtype=torch.long, device=device),
+            }
+            if cam_self_attn is not None:
+                cam_heads = int(cam_self_attn.num_heads)
+                cam_head_dim = int(cam_self_attn.head_dim)
+                cache.update({
+                    "prope_k":
+                    torch.zeros(batch_size, size, cam_heads, cam_head_dim, dtype=dtype, device=device),
+                    "prope_v":
+                    torch.zeros(batch_size, size, cam_heads, cam_head_dim, dtype=dtype, device=device),
+                    "prope_global_end_index":
+                    torch.tensor([0], dtype=torch.long, device=device),
+                    "prope_local_end_index":
+                    torch.tensor([0], dtype=torch.long, device=device),
+                })
+            caches.append(cache)
+        return caches
+
+    def _initialize_crossattn_cache(self, batch_size: int, dtype: torch.dtype, device: torch.device):
+        heads = self.transformer.num_attention_heads
+        head_dim = self.transformer.attention_head_dim
+        return [{
+            "k": torch.zeros(batch_size, 512, heads, head_dim, dtype=dtype, device=device),
+            "v": torch.zeros(batch_size, 512, heads, head_dim, dtype=dtype, device=device),
+            "is_init": False,
+        } for _ in range(self.num_transformer_blocks)]
+
+    def _update_context_cache(self, block_latents: torch.Tensor, context: Any, camera_block: Any,
+                              kv_cache: list[dict[str, Any]], crossattn_cache: list[dict[str, Any]], start: int,
+                              frame_seq_length: int, target_dtype: torch.dtype, autocast_enabled: bool,
+                              context_noise: float) -> None:
+        timestep = torch.full(
+            (block_latents.shape[0], block_latents.shape[2] * frame_seq_length),
+            self._context_noise_timestep(context_noise),
+            device=block_latents.device,
+            dtype=torch.long,
+        )
+        with torch.autocast(device_type="cuda", dtype=target_dtype, enabled=autocast_enabled):
+            self.transformer(
+                hidden_states=block_latents.to(target_dtype),
+                encoder_hidden_states=torch.stack(context).to(target_dtype),
+                timestep=timestep,
+                y_camera=camera_block,
+                kv_cache=kv_cache,
+                crossattn_cache=crossattn_cache,
+                current_start=start * frame_seq_length,
+            )

fastvideo/pipelines/basic/dreamx_world/camera_conditioning.py

@@ -0,0 +1,228 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+from scipy.interpolate import interp1d
+from scipy.spatial.transform import Rotation, Slerp
+
+_ACTION_TO_MOTION = {
+    "w": "forward",
+    "a": "left",
+    "d": "right",
+    "s": "backward",
+    "j": "left_rot",
+    "l": "right_rot",
+    "i": "up_rot",
+    "k": "down_rot",
+}
+_TRANSLATION_BASE_UNIT = 1.0
+_ROTATION_BASE_UNIT = 10.0
+_INTRINSIC_ROW = [0.8, 0.5, 0.5, 0.5]
+
+
+@dataclass
+class DreamXCamera:
+    fx: float
+    fy: float
+    cx: float
+    cy: float
+    w2c_mat: np.ndarray
+
+    @property
+    def c2w_mat(self) -> np.ndarray:
+        return np.linalg.inv(self.w2c_mat)
+
+    @classmethod
+    def from_pose_row(cls, row: list[float]) -> DreamXCamera:
+        w2c_mat = np.eye(4, dtype=np.float64)
+        w2c_mat[:3, :] = np.asarray(row[7:], dtype=np.float64).reshape(3, 4)
+        return cls(
+            fx=float(row[1]),
+            fy=float(row[2]),
+            cx=float(row[3]),
+            cy=float(row[4]),
+            w2c_mat=w2c_mat,
+        )
+
+
+def _translation_step(motion_type: str, current_pose: dict[str, np.ndarray], value: float, duration: int) -> np.ndarray:
+    if motion_type in ("forward", "backward"):
+        yaw = np.radians(current_pose["rotation"][1])
+        pitch = np.radians(current_pose["rotation"][0])
+        forward = np.array([-math.sin(yaw) * math.cos(pitch), math.sin(pitch), math.cos(yaw) * math.cos(pitch)])
+        direction = 1 if motion_type == "forward" else -1
+        return forward * value * direction / duration
+    if motion_type in ("left", "right"):
+        yaw = np.radians(current_pose["rotation"][1])
+        right = np.array([math.cos(yaw), 0.0, math.sin(yaw)])
+        direction = -1 if motion_type == "left" else 1
+        return right * value * direction / duration
+    return np.zeros(3)
+
+
+def _rotation_step(motion_type: str, value: float, duration: int) -> np.ndarray:
+    if not motion_type.endswith("rot"):
+        return np.zeros(3)
+    axis = motion_type.split("_")[0]
+    rotation = np.zeros(3)
+    if axis == "left":
+        rotation[1] = value
+    elif axis == "right":
+        rotation[1] = -value
+    elif axis == "up":
+        rotation[0] = -value
+    elif axis == "down":
+        rotation[0] = value
+    return rotation / duration
+
+
+def _euler_to_quaternion(angles: np.ndarray) -> list[float]:
+    pitch, yaw, roll = np.radians(angles)
+    cy = math.cos(yaw * 0.5)
+    sy = math.sin(yaw * 0.5)
+    cp = math.cos(pitch * 0.5)
+    sp = math.sin(pitch * 0.5)
+    cr = math.cos(roll * 0.5)
+    sr = math.sin(roll * 0.5)
+    return [
+        cy * cp * cr + sy * sp * sr,
+        cy * sp * cr + sy * cp * sr,
+        sy * cp * cr - cy * sp * sr,
+        cy * cp * sr - sy * sp * cr,
+    ]
+
+
+def _quaternion_to_rotation_matrix(quaternion: list[float]) -> np.ndarray:
+    qw, qx, qy, qz = quaternion
+    return np.array([
+        [1 - 2 * (qy**2 + qz**2), 2 * (qx * qy - qw * qz), 2 * (qx * qz + qw * qy)],
+        [2 * (qx * qy + qw * qz), 1 - 2 * (qx**2 + qz**2), 2 * (qy * qz - qw * qx)],
+        [2 * (qx * qz - qw * qy), 2 * (qy * qz + qw * qx), 1 - 2 * (qx**2 + qy**2)],
+    ])
+
+
+def _pose_rows_from_actions(action_seq: list[str], action_speed_list: list[float], duration: int) -> list[list[float]]:
+    if len(action_seq) != len(action_speed_list):
+        raise ValueError("action_seq and action_speed_list must have the same length")
+
+    positions: list[np.ndarray] = []
+    rotations: list[np.ndarray] = []
+    current_pose = {
+        "position": np.array([0.0, 0.0, 0.0]),
+        "rotation": np.array([0.0, 0.0, 0.0]),
+    }
+
+    for action_id, speed in zip(action_seq, action_speed_list, strict=True):
+        motion_types = [_ACTION_TO_MOTION[key] for key in list(action_id)]
+        translation_step = np.zeros(3)
+        rotation_step = np.zeros(3)
+        for motion_type in motion_types:
+            translation_step += _translation_step(motion_type, current_pose,
+                                                  float(speed) * _TRANSLATION_BASE_UNIT, duration)
+            rotation_step += _rotation_step(motion_type, float(speed) * _ROTATION_BASE_UNIT, duration)
+
+        segment_positions = []
+        segment_rotations = []
+        for index in range(1, duration + 1):
+            segment_positions.append(current_pose["position"] + translation_step * index)
+            segment_rotations.append(current_pose["rotation"] + rotation_step * index)
+        current_pose["position"] = segment_positions[-1].copy()
+        current_pose["rotation"] = segment_rotations[-1].copy()
+        positions.extend(segment_positions)
+        rotations.extend(segment_rotations)
+
+    rows: list[list[float]] = [[0.0] + _INTRINSIC_ROW + [0.0, 0.0] +
+                               [1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0]]
+    for index, (position, rotation) in enumerate(zip(positions, rotations, strict=False)):
+        rotation_matrix = _quaternion_to_rotation_matrix(_euler_to_quaternion(rotation))
+        translation = -rotation_matrix @ position
+        extrinsic = np.hstack([rotation_matrix, translation.reshape(3, 1)])
+        rows.append([float(index)] + _INTRINSIC_ROW + [0.0, 0.0] + extrinsic.flatten().tolist())
+    return rows
+
+
+def _interpolate_camera_poses(
+    cameras: list[DreamXCamera],
+    src_indices: np.ndarray,
+    tgt_indices: np.ndarray,
+) -> list[DreamXCamera]:
+    if len(cameras) <= 1:
+        return [cameras[0]] * len(tgt_indices) if cameras else []
+    src_rot_mat = np.array([camera.w2c_mat[:3, :3] for camera in cameras])
+    src_trans_vec = np.array([camera.w2c_mat[:3, 3] for camera in cameras])
+
+    dets = np.linalg.det(src_rot_mat)
+    flip_handedness = dets.size > 0 and np.median(dets) < 0.0
+    if flip_handedness:
+        flip_mat = np.diag([1.0, 1.0, -1.0]).astype(src_rot_mat.dtype)
+        src_rot_mat = src_rot_mat @ flip_mat
+
+    trans = interp1d(src_indices, src_trans_vec, axis=0, kind="linear", bounds_error=False,
+                     fill_value="extrapolate")(tgt_indices)
+    quats = Rotation.from_matrix(src_rot_mat).as_quat().copy()
+    for index in range(1, len(quats)):
+        if np.dot(quats[index], quats[index - 1]) < 0:
+            quats[index] = -quats[index]
+    rot = Slerp(src_indices, Rotation.from_quat(quats))(tgt_indices).as_matrix()
+    if flip_handedness:
+        rot = rot @ flip_mat
+
+    ref = cameras[0]
+    result = []
+    for index in range(len(tgt_indices)):
+        w2c_mat = np.eye(4, dtype=np.float64)
+        w2c_mat[:3, :] = np.hstack([rot[index], trans[index].reshape(3, 1)])
+        result.append(DreamXCamera(ref.fx, ref.fy, ref.cx, ref.cy, w2c_mat))
+    return result
+
+
+def _relative_c2w_poses(cameras: list[DreamXCamera]) -> np.ndarray:
+    abs_w2cs = [camera.w2c_mat for camera in cameras]
+    abs_c2ws = [camera.c2w_mat for camera in cameras]
+    target_cam_c2w = np.eye(4, dtype=np.float64)
+    abs2rel = target_cam_c2w @ abs_w2cs[0]
+    poses = [target_cam_c2w] + [abs2rel @ c2w for c2w in abs_c2ws[1:]]
+    return np.asarray(poses, dtype=np.float32)
+
+
+def _invert_se3(transforms: torch.Tensor) -> torch.Tensor:
+    rotation_inv = transforms[..., :3, :3].transpose(-1, -2)
+    output = torch.zeros_like(transforms)
+    output[..., :3, :3] = rotation_inv
+    output[..., :3, 3] = -torch.einsum("...ij,...j->...i", rotation_inv, transforms[..., :3, 3])
+    output[..., 3, 3] = 1.0
+    return output
+
+
+def build_dreamx_camera_condition(
+    action_seq: list[str],
+    action_speed_list: list[float],
+    *,
+    num_frames: int,
+    height: int,
+    width: int,
+    dtype: torch.dtype = torch.float32,
+    device: torch.device | str = "cpu",
+) -> dict[str, torch.Tensor]:
+    del height, width  # DreamX-World-5B-Cam uses fixed normalized intrinsics.
+    duration = math.ceil(num_frames / len(action_seq))
+    rows = _pose_rows_from_actions(action_seq, action_speed_list, duration)[:num_frames]
+    cameras = [DreamXCamera.from_pose_row(row) for row in rows]
+
+    latent_frame_count = 1 + (len(cameras) - 1) // 4
+    src_indices = np.arange(len(cameras), dtype=np.float64)
+    tgt_indices = np.linspace(0, len(cameras) - 1, latent_frame_count)
+    cameras = _interpolate_camera_poses(cameras, src_indices, tgt_indices)
+
+    c2ws = torch.as_tensor(_relative_c2w_poses(cameras), dtype=dtype, device=device)
+    viewmats = _invert_se3(c2ws)
+
+    intrinsics = torch.zeros((latent_frame_count, 3, 3), dtype=dtype, device=device)
+    intrinsics[:, 0, 0] = 969.6969696969696 / (960.0 * 2)
+    intrinsics[:, 1, 1] = 969.6969696969696 / (540.0 * 2)
+    intrinsics[:, 2, 2] = 1.0
+    return {"viewmats": viewmats, "K": intrinsics}

fastvideo/pipelines/basic/dreamx_world/config.py

@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Compatibility exports for DreamX-World pipeline configs."""
+
+from fastvideo.configs.pipelines.dreamx_world import (
+    DreamXWorld5BARPipelineConfig,
+    DreamXWorld5BCamPipelineConfig,
+    make_dreamx_world_5b_ar_dit_config,
+    make_dreamx_world_5b_cam_dit_config,
+    make_dreamx_world_5b_cam_text_encoder_config,
+    make_dreamx_world_5b_cam_vae_config,
+)
+
+__all__ = [
+    "DreamXWorld5BARPipelineConfig",
+    "DreamXWorld5BCamPipelineConfig",
+    "make_dreamx_world_5b_ar_dit_config",
+    "make_dreamx_world_5b_cam_dit_config",
+    "make_dreamx_world_5b_cam_text_encoder_config",
+    "make_dreamx_world_5b_cam_vae_config",
+]

fastvideo/pipelines/basic/dreamx_world/dreamx_world_ar_pipeline.py

@@ -0,0 +1,67 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World-5B autoregressive pipeline entrypoint."""
+
+from fastvideo.api.sampling_param import SamplingParam
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BARPipelineConfig
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
+from fastvideo.pipelines import ComposedPipelineBase, LoRAPipeline
+from fastvideo.pipelines.basic.dreamx_world.ar_denoising import DreamXWorldARCausalDenoisingStage
+from fastvideo.pipelines.basic.dreamx_world.stages import (
+    DreamXWorldCameraConditioningStage,
+    DreamXWorldImageVAEEncodingStage,
+)
+from fastvideo.pipelines.stages import (
+    ConditioningStage,
+    DecodingStage,
+    InputValidationStage,
+    LatentPreparationStage,
+    TextEncodingStage,
+    TimestepPreparationStage,
+)
+
+logger = init_logger(__name__)
+
+
+class DreamXWorldARPipeline(LoRAPipeline, ComposedPipelineBase):
+    """DreamX-World-5B autoregressive causal camera pipeline."""
+
+    _required_config_modules = ["text_encoder", "tokenizer", "vae", "transformer", "scheduler"]
+    pipeline_config_cls = DreamXWorld5BARPipelineConfig
+    sampling_params_cls = SamplingParam
+
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(shift=fastvideo_args.pipeline_config.flow_shift)
+        self.modules["scheduler"].set_timesteps(1000)
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
+        self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage())
+        self.add_stage(stage_name="prompt_encoding_stage",
+                       stage=TextEncodingStage(
+                           text_encoders=[self.get_module("text_encoder")],
+                           tokenizers=[self.get_module("tokenizer")],
+                       ))
+        self.add_stage(stage_name="conditioning_stage", stage=ConditioningStage())
+        self.add_stage(stage_name="timestep_preparation_stage",
+                       stage=TimestepPreparationStage(scheduler=self.get_module("scheduler")))
+        self.add_stage(stage_name="latent_preparation_stage",
+                       stage=LatentPreparationStage(
+                           scheduler=self.get_module("scheduler"),
+                           transformer=self.get_module("transformer", None),
+                       ))
+        self.add_stage(stage_name="image_latent_preparation_stage",
+                       stage=DreamXWorldImageVAEEncodingStage(vae=self.get_module("vae")))
+        self.add_stage(stage_name="dreamx_camera_conditioning_stage", stage=DreamXWorldCameraConditioningStage())
+        self.add_stage(stage_name="denoising_stage",
+                       stage=DreamXWorldARCausalDenoisingStage(
+                           transformer=self.get_module("transformer"),
+                           scheduler=self.get_module("scheduler"),
+                           vae=self.get_module("vae"),
+                           pipeline=self,
+                       ))
+        self.add_stage(stage_name="decoding_stage", stage=DecodingStage(vae=self.get_module("vae"), pipeline=self))
+        logger.info("DreamXWorldARPipeline initialized with autoregressive causal denoising")
+
+
+EntryClass = DreamXWorldARPipeline

fastvideo/pipelines/basic/dreamx_world/dreamx_world_pipeline.py

@@ -0,0 +1,78 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World video pipeline entrypoint."""
+
+from fastvideo.api.sampling_param import SamplingParam
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
+from fastvideo.pipelines import ComposedPipelineBase, LoRAPipeline
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BCamPipelineConfig
+from fastvideo.pipelines.basic.dreamx_world.stages import DreamXWorldCameraConditioningStage
+from fastvideo.pipelines.stages import (
+    ConditioningStage,
+    DecodingStage,
+    DenoisingStage,
+    InputValidationStage,
+    LatentPreparationStage,
+    TextEncodingStage,
+    TimestepPreparationStage,
+)
+
+logger = init_logger(__name__)
+
+
+class DreamXWorldPipeline(LoRAPipeline, ComposedPipelineBase):
+    """DreamX-World-5B-Cam pipeline with native FastVideo camera conditioning."""
+
+    _required_config_modules = ["text_encoder", "tokenizer", "vae", "transformer", "scheduler"]
+    pipeline_config_cls = DreamXWorld5BCamPipelineConfig
+    sampling_params_cls = SamplingParam
+
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(shift=fastvideo_args.pipeline_config.flow_shift)
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
+        self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage())
+
+        self.add_stage(
+            stage_name="prompt_encoding_stage",
+            stage=TextEncodingStage(
+                text_encoders=[self.get_module("text_encoder")],
+                tokenizers=[self.get_module("tokenizer")],
+            ),
+        )
+
+        self.add_stage(stage_name="conditioning_stage", stage=ConditioningStage())
+
+        self.add_stage(
+            stage_name="timestep_preparation_stage",
+            stage=TimestepPreparationStage(scheduler=self.get_module("scheduler")),
+        )
+
+        self.add_stage(
+            stage_name="latent_preparation_stage",
+            stage=LatentPreparationStage(
+                scheduler=self.get_module("scheduler"),
+                transformer=self.get_module("transformer", None),
+            ),
+        )
+
+        self.add_stage(stage_name="dreamx_camera_conditioning_stage", stage=DreamXWorldCameraConditioningStage())
+
+        self.add_stage(
+            stage_name="denoising_stage",
+            stage=DenoisingStage(
+                transformer=self.get_module("transformer"),
+                transformer_2=self.get_module("transformer_2", None),
+                scheduler=self.get_module("scheduler"),
+                vae=self.get_module("vae"),
+                pipeline=self,
+            ),
+        )
+
+        self.add_stage(stage_name="decoding_stage", stage=DecodingStage(vae=self.get_module("vae"), pipeline=self))
+
+        logger.info("DreamXWorldPipeline initialized with native camera conditioning")
+
+
+EntryClass = DreamXWorldPipeline

fastvideo/pipelines/basic/dreamx_world/presets.py

@@ -0,0 +1,58 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World model family pipeline presets."""
+
+from fastvideo.api.presets import InferencePreset, PresetStageSpec
+
+_NEGATIVE_PROMPT_CN = ("色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，"
+                       "静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，"
+                       "多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，"
+                       "形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，"
+                       "背景人很多，倒着走")
+
+_DENOISE_STAGE = PresetStageSpec(
+    name="denoise",
+    kind="denoising",
+    description="DreamX-World camera-conditioned denoising pass",
+    allowed_overrides=frozenset({
+        "num_inference_steps",
+        "guidance_scale",
+    }),
+)
+
+DREAMX_WORLD_5B_CAM = InferencePreset(
+    name="dreamx_world_5b_cam",
+    version=1,
+    model_family="dreamx_world",
+    description="DreamX-World 5B camera-control video generation",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 480,
+        "width": 832,
+        "num_frames": 161,
+        "fps": 16,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 30,
+        "negative_prompt": _NEGATIVE_PROMPT_CN,
+    },
+)
+
+DREAMX_WORLD_5B_AR = InferencePreset(
+    name="dreamx_world_5b_ar",
+    version=1,
+    model_family="dreamx_world",
+    description="DreamX-World 5B autoregressive camera-control generation",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 704,
+        "width": 1280,
+        "num_frames": 1005,
+        "fps": 16,
+        "guidance_scale": 1.0,
+        "num_inference_steps": 4,
+        "negative_prompt": _NEGATIVE_PROMPT_CN,
+    },
+)
+
+ALL_PRESETS = (DREAMX_WORLD_5B_CAM, DREAMX_WORLD_5B_AR)

fastvideo/pipelines/basic/dreamx_world/stages.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World pipeline stages."""
+
+from __future__ import annotations
+
+import torch
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.stages.base import PipelineStage
+from fastvideo.pipelines.stages.validators import VerificationResult
+
+from fastvideo.pipelines.basic.dreamx_world.camera_conditioning import (
+    build_dreamx_camera_condition, )
+
+DREAMX_Y_CAMERA_KEY = "dreamx_y_camera"
+
+logger = init_logger(__name__)
+
+
+class DreamXWorldCameraConditioningStage(PipelineStage):
+    """Build PRoPE camera conditioning for DreamX-World-5B-Cam."""
+
+    def forward(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> ForwardBatch:
+        del fastvideo_args
+        if DREAMX_Y_CAMERA_KEY in batch.extra:
+            return batch
+
+        action_seq = batch.extra.get("dreamx_action_seq", batch.action_list)
+        action_speed_list = batch.extra.get("dreamx_action_speed_list", batch.action_speed_list)
+        if action_seq is None:
+            action_seq = ["w"]
+        if action_speed_list is None:
+            action_speed_list = [4]
+
+        if isinstance(action_seq, str):
+            action_seq = [action_seq]
+        if isinstance(action_speed_list, int | float):
+            action_speed_list = [action_speed_list]
+        if len(action_speed_list) == 1 and len(action_seq) > 1:
+            action_speed_list = list(action_speed_list) * len(action_seq)
+        action_speed_list = [float(speed) for speed in action_speed_list]
+
+        height = int(batch.height) if batch.height is not None else 704
+        width = int(batch.width) if batch.width is not None else 1280
+        num_frames = int(batch.num_frames)
+        dtype = batch.latents.dtype if torch.is_tensor(batch.latents) else torch.float32
+        device = batch.latents.device if torch.is_tensor(batch.latents) else "cpu"
+
+        y_camera = build_dreamx_camera_condition(
+            list(action_seq),
+            action_speed_list,
+            num_frames=num_frames,
+            height=height,
+            width=width,
+            dtype=dtype,
+            device=device,
+        )
+        batch.extra[DREAMX_Y_CAMERA_KEY] = {key: value.unsqueeze(0) for key, value in y_camera.items()}
+        return batch
+
+    def verify_output(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> VerificationResult:
+        del fastvideo_args
+        result = VerificationResult()
+        y_camera = batch.extra.get(DREAMX_Y_CAMERA_KEY)
+        result.add_check("dreamx_y_camera", y_camera, lambda value: isinstance(value, dict))
+        if isinstance(y_camera, dict):
+            result.add_check("dreamx_y_camera.viewmats", y_camera.get("viewmats"), torch.is_tensor)
+            result.add_check("dreamx_y_camera.K", y_camera.get("K"), torch.is_tensor)
+        return result
+
+
+class DreamXWorldImageVAEEncodingStage(PipelineStage):
+    """Encode the conditioning image into the first-frame latent.
+
+    Official AR-forcing flow (AMAP-ML/DreamX-World inference_ar_forcing.py):
+    the input image is resized, normalized to [-1, 1], VAE-encoded
+    deterministically, and written into frame 0 of the noise — the causal
+    denoiser then treats frame 0 as clean context. This stage produces
+    ``batch.image_latent`` ([B, C, 1, H_lat, W_lat]); the injection into
+    the latents happens in DreamXWorldARCausalDenoisingStage.
+    """
+
+    def __init__(self, vae) -> None:
+        super().__init__()
+        self.vae = vae
+
+    def forward(
+        self,
+        batch: ForwardBatch,
+        fastvideo_args: FastVideoArgs,
+    ) -> ForwardBatch:
+        if batch.pil_image is None:
+            # No conditioning image: the causal denoiser falls back to
+            # running from pure noise (frame 0 uninitialized). Warn loudly —
+            # this pipeline is registered I2V and the official flow always
+            # forces from a frame.
+            logger.warning("DreamXWorldARPipeline called without an input image; "
+                           "first-frame context will be noise (T2V-style). Pass an "
+                           "image for the official AR-forcing behavior.")
+            return batch
+
+        from fastvideo.platforms import get_local_torch_device
+        from fastvideo.utils import PRECISION_TO_TYPE
+
+        device = get_local_torch_device()
+        image = batch.pil_image
+        if not isinstance(image, torch.Tensor):
+            import numpy as np
+            import PIL.Image
+            assert isinstance(image, PIL.Image.Image)
+            width = batch.width if isinstance(batch.width, int) else batch.width[0]
+            height = batch.height if isinstance(batch.height, int) else batch.height[0]
+            image = image.convert("RGB").resize((width, height), PIL.Image.Resampling.LANCZOS)
+            arr = torch.from_numpy(np.asarray(image)).float().permute(2, 0, 1) / 255.0
+            image = (arr - 0.5) / 0.5  # official Normalize([0.5], [0.5])
+            image = image.unsqueeze(0)  # [1, C, H, W]
+        if image.dim() == 4:
+            image = image.unsqueeze(2)  # [B, C, 1, H, W]
+        elif image.dim() == 5:
+            image = image[:, :, :1]
+        image = image.to(device=device, dtype=torch.float32)
+
+        vae_dtype = PRECISION_TO_TYPE[fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = (vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
+        self.vae = self.vae.to(device)
+        with torch.autocast(device_type="cuda", dtype=vae_dtype, enabled=vae_autocast_enabled):
+            if not vae_autocast_enabled:
+                image = image.to(vae_dtype)
+            encoder_output = self.vae.encode(image)
+
+        # Official encode_to_latent is deterministic ((mean - mu) / sigma per
+        # channel); the posterior mean + shift/scale is the FastVideo
+        # equivalent of that normalization.
+        latent = encoder_output.mean
+        if getattr(self.vae, "shift_factor", None) is not None:
+            shift = self.vae.shift_factor
+            latent = latent - (shift.to(latent.device, latent.dtype) if isinstance(shift, torch.Tensor) else shift)
+        scale = self.vae.scaling_factor
+        latent = latent * (scale.to(latent.device, latent.dtype) if isinstance(scale, torch.Tensor) else scale)
+
+        batch.image_latent = latent
+        return batch
+
+    def verify_input(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        return result

fastvideo/pipelines/basic/kandinsky5/__init__.py

@@ -0,0 +1 @@
+# SPDX-License-Identifier: Apache-2.0

fastvideo/pipelines/basic/kandinsky5/kandinsky5_i2v_pipeline.py

@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
+from fastvideo.pipelines.stages.input_validation import InputValidationStage
+from fastvideo.pipelines.stages.kandinsky5 import (
+    Kandinsky5DecodingStage,
+    Kandinsky5DenoisingStage,
+    Kandinsky5ImageEncodingStage,
+    Kandinsky5LatentPreparationStage,
+    Kandinsky5NormalizationStage,
+)
+from fastvideo.pipelines.stages.text_encoding import TextEncodingStage
+from fastvideo.pipelines.stages.timestep_preparation import TimestepPreparationStage
+
+
+class Kandinsky5I2VPipeline(ComposedPipelineBase):
+    """Kandinsky-5.0 image-to-video pipeline."""
+
+    _required_config_modules = [
+        "scheduler",
+        "text_encoder",
+        "text_encoder_2",
+        "tokenizer",
+        "tokenizer_2",
+        "transformer",
+        "vae",
+    ]
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
+        self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage())
+
+        self.add_stage(
+            stage_name="text_encoding_stage",
+            stage=TextEncodingStage(
+                text_encoders=[
+                    self.get_module("text_encoder"),
+                    self.get_module("text_encoder_2"),
+                ],
+                tokenizers=[
+                    self.get_module("tokenizer"),
+                    self.get_module("tokenizer_2"),
+                ],
+            ),
+        )
+
+        self.add_stage(
+            stage_name="timestep_preparation_stage",
+            stage=TimestepPreparationStage(scheduler=self.get_module("scheduler")),
+        )
+
+        self.add_stage(
+            stage_name="latent_preparation_stage",
+            stage=Kandinsky5LatentPreparationStage(
+                scheduler=self.get_module("scheduler"),
+                transformer=self.get_module("transformer"),
+            ),
+        )
+
+        # Runs AFTER latent preparation so the initial noise is the seeded
+        # generator's first draw (official kandinsky-5 RNG order); this stage
+        # then samples the image latent and places it into the latents.
+        self.add_stage(
+            stage_name="image_encoding_stage",
+            stage=Kandinsky5ImageEncodingStage(vae=self.get_module("vae")),
+        )
+
+        self.add_stage(
+            stage_name="denoising_stage",
+            stage=Kandinsky5DenoisingStage(
+                transformer=self.get_module("transformer"),
+                scheduler=self.get_module("scheduler"),
+            ),
+        )
+
+        self.add_stage(
+            stage_name="normalization_stage",
+            stage=Kandinsky5NormalizationStage(),
+        )
+
+        self.add_stage(
+            stage_name="decoding_stage",
+            stage=Kandinsky5DecodingStage(vae=self.get_module("vae"), pipeline=self),
+        )
+
+
+EntryClass = Kandinsky5I2VPipeline

fastvideo/pipelines/basic/kandinsky5/kandinsky5_pipeline.py

@@ -0,0 +1,74 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase
+from fastvideo.pipelines.stages.input_validation import InputValidationStage
+from fastvideo.pipelines.stages.kandinsky5 import (
+    Kandinsky5DecodingStage,
+    Kandinsky5DenoisingStage,
+    Kandinsky5LatentPreparationStage,
+)
+from fastvideo.pipelines.stages.text_encoding import TextEncodingStage
+from fastvideo.pipelines.stages.timestep_preparation import TimestepPreparationStage
+
+
+class Kandinsky5T2VPipeline(ComposedPipelineBase):
+    """Kandinsky-5.0 Lite text-to-video pipeline."""
+
+    _required_config_modules = [
+        "scheduler",
+        "text_encoder",
+        "text_encoder_2",
+        "tokenizer",
+        "tokenizer_2",
+        "transformer",
+        "vae",
+    ]
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs) -> None:
+        self.add_stage(stage_name="input_validation_stage", stage=InputValidationStage())
+
+        self.add_stage(
+            stage_name="text_encoding_stage",
+            stage=TextEncodingStage(
+                text_encoders=[
+                    self.get_module("text_encoder"),
+                    self.get_module("text_encoder_2"),
+                ],
+                tokenizers=[
+                    self.get_module("tokenizer"),
+                    self.get_module("tokenizer_2"),
+                ],
+            ),
+        )
+
+        self.add_stage(
+            stage_name="timestep_preparation_stage",
+            stage=TimestepPreparationStage(scheduler=self.get_module("scheduler")),
+        )
+
+        self.add_stage(
+            stage_name="latent_preparation_stage",
+            stage=Kandinsky5LatentPreparationStage(
+                scheduler=self.get_module("scheduler"),
+                transformer=self.get_module("transformer"),
+            ),
+        )
+
+        self.add_stage(
+            stage_name="denoising_stage",
+            stage=Kandinsky5DenoisingStage(
+                transformer=self.get_module("transformer"),
+                scheduler=self.get_module("scheduler"),
+            ),
+        )
+
+        self.add_stage(
+            stage_name="decoding_stage",
+            stage=Kandinsky5DecodingStage(vae=self.get_module("vae"), pipeline=self),
+        )
+
+
+EntryClass = Kandinsky5T2VPipeline

fastvideo/pipelines/basic/kandinsky5/presets.py

@@ -0,0 +1,165 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Kandinsky-5 model family pipeline presets."""
+
+from fastvideo.api.presets import InferencePreset, PresetStageSpec
+
+_NEGATIVE_PROMPT = ("Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, "
+                    "deformed, walking backwards")
+
+_DENOISE_STAGE = PresetStageSpec(
+    name="denoise",
+    kind="denoising",
+    description="Main Kandinsky-5 denoising pass",
+    allowed_overrides=frozenset({
+        "num_inference_steps",
+        "guidance_scale",
+    }),
+)
+
+KANDINSKY5_T2V_LITE_5S = InferencePreset(
+    name="kandinsky5_t2v_lite_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Lite T2V 5s",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 50,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_T2V_LITE_DISTILLED_5S = InferencePreset(
+    name="kandinsky5_t2v_lite_distilled_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Lite T2V Distilled 5s",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 1.0,
+        "num_inference_steps": 16,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_T2V_PRO_5S = InferencePreset(
+    name="kandinsky5_t2v_pro_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Pro T2V 5s",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 50,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_T2V_PRO_DISTILLED_5S = InferencePreset(
+    name="kandinsky5_t2v_pro_distilled_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Pro T2V Distilled 5s",
+    workload_type="t2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 1.0,
+        "num_inference_steps": 16,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_I2V_LITE_5S = InferencePreset(
+    name="kandinsky5_i2v_lite_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Lite I2V 5s",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 50,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_I2V_PRO_5S = InferencePreset(
+    name="kandinsky5_i2v_pro_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Pro I2V 5s",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 5.0,
+        "num_inference_steps": 50,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_I2V_LITE_DISTILLED_5S = InferencePreset(
+    name="kandinsky5_i2v_lite_distilled_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Lite I2V Distilled 5s",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 1.0,
+        "num_inference_steps": 16,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+KANDINSKY5_I2V_PRO_DISTILLED_5S = InferencePreset(
+    name="kandinsky5_i2v_pro_distilled_5s",
+    version=1,
+    model_family="kandinsky5",
+    description="Kandinsky-5.0 Pro I2V Distilled 5s",
+    workload_type="i2v",
+    stage_schemas=(_DENOISE_STAGE, ),
+    defaults={
+        "height": 512,
+        "width": 768,
+        "num_frames": 121,
+        "fps": 24,
+        "guidance_scale": 1.0,
+        "num_inference_steps": 16,
+        "negative_prompt": _NEGATIVE_PROMPT,
+    },
+)
+
+ALL_PRESETS = (KANDINSKY5_T2V_LITE_5S, KANDINSKY5_T2V_LITE_DISTILLED_5S, KANDINSKY5_T2V_PRO_5S,
+               KANDINSKY5_T2V_PRO_DISTILLED_5S, KANDINSKY5_I2V_LITE_5S, KANDINSKY5_I2V_LITE_DISTILLED_5S,
+               KANDINSKY5_I2V_PRO_5S, KANDINSKY5_I2V_PRO_DISTILLED_5S)

fastvideo/pipelines/stages/__init__.py

@@ -33,6 +33,8 @@ from fastvideo.pipelines.basic.ltx2.stages import (  # noqa: F401
 from fastvideo.pipelines.stages.matrixgame2_denoising import MatrixGame2CausalDenoisingStage
 from fastvideo.pipelines.stages.matrixgame3_denoising import MatrixGame3DenoisingStage
 from fastvideo.pipelines.stages.hyworld_denoising import HYWorldDenoisingStage
+from fastvideo.pipelines.stages.kandinsky5 import (Kandinsky5DecodingStage, Kandinsky5DenoisingStage,
+                                                   Kandinsky5LatentPreparationStage)
 from fastvideo.pipelines.stages.gamecraft_denoising import GameCraftDenoisingStage
 from fastvideo.pipelines.stages.gen3c_stages import (Gen3CCFGPolicyStage, Gen3CConditioningStage, Gen3CDenoisingStage,
                                                      Gen3CLatentPreparationStage)
@@ -65,6 +67,9 @@ __all__ = [
     "MatrixGame2CausalDenoisingStage",
     "MatrixGame3DenoisingStage",
     "HYWorldDenoisingStage",
+    "Kandinsky5DecodingStage",
+    "Kandinsky5DenoisingStage",
+    "Kandinsky5LatentPreparationStage",
     "GameCraftDenoisingStage",
     "Gen3CCFGPolicyStage",
     "Gen3CConditioningStage",

fastvideo/pipelines/stages/denoising.py

@@ -191,6 +191,19 @@ class DenoisingStage(PipelineStage):
             },
         )
 
+        dreamx_y_camera = batch.extra.get("dreamx_y_camera", batch.extra.get("y_camera"))
+        if isinstance(dreamx_y_camera, dict):
+            dreamx_y_camera = {
+                key: value.to(device=local_device, dtype=target_dtype) if torch.is_tensor(value) else value
+                for key, value in dreamx_y_camera.items()
+            }
+        dreamx_camera_kwargs = self.prepare_extra_func_kwargs(
+            self.transformer.forward,
+            {
+                "y_camera": dreamx_y_camera,
+            },
+        )
+
         for key in ("flux2_txt_ids", "flux2_img_ids"):
             value = batch.extra.get(key)
             if torch.is_tensor(value):
@@ -242,7 +255,11 @@ class DenoisingStage(PipelineStage):
             # the image latent instead of appending along the channel dim
             assert batch.image_latent is None, "TI2V task should not have image latents"
             assert self.vae is not None, "VAE is not provided for TI2V task"
+            vae_device = next(self.vae.parameters()).device
+            self.vae = self.vae.to(local_device)
             z = self.vae.encode(batch.pil_image).mean.float()
+            if getattr(fastvideo_args, "vae_cpu_offload", False):
+                self.vae = self.vae.to(vae_device)
             if (hasattr(self.vae, "shift_factor") and self.vae.shift_factor is not None):
                 if isinstance(self.vae.shift_factor, torch.Tensor):
                     z -= self.vae.shift_factor.to(z.device, z.dtype)
@@ -495,6 +512,7 @@ class DenoisingStage(PipelineStage):
                             **pos_cond_kwargs,
                             **action_kwargs,
                             **camera_kwargs,
+                            **dreamx_camera_kwargs,
                             **timesteps_r_kwarg,
                             **flux2_id_kwargs,
                         )
@@ -537,6 +555,7 @@ class DenoisingStage(PipelineStage):
                                     **neg_cond_kwargs,
                                     **action_kwargs,
                                     **camera_kwargs,
+                                    **dreamx_camera_kwargs,
                                     **timesteps_r_kwarg,
                                     **flux2_id_kwargs,
                                 )

fastvideo/pipelines/stages/kandinsky5.py

@@ -0,0 +1,545 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from __future__ import annotations
+
+import contextlib
+from typing import Any
+
+import PIL
+import torch
+from diffusers.utils.torch_utils import randn_tensor
+from tqdm.auto import tqdm
+
+from fastvideo.attention.backends.nabla import NablaAttentionMetadataBuilder
+from fastvideo.distributed import get_local_torch_device
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.forward_context import set_forward_context
+from fastvideo.logger import init_logger
+from fastvideo.models.loader.component_loader import TransformerLoader, VAELoader
+from fastvideo.models.vaes.common import ParallelTiledVAE
+from fastvideo.models.vision_utils import normalize, numpy_to_pt, pil_to_numpy, resize
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.stages.base import PipelineStage
+from fastvideo.pipelines.stages.decoding import DecodingStage
+from fastvideo.pipelines.stages.encoding import EncodingStage
+from fastvideo.pipelines.stages.validators import StageValidators as V
+from fastvideo.pipelines.stages.validators import VerificationResult
+from fastvideo.utils import PRECISION_TO_TYPE
+
+logger = init_logger(__name__)
+
+
+class Kandinsky5LatentPreparationStage(PipelineStage):
+
+    def __init__(self, scheduler, transformer) -> None:
+        super().__init__()
+        self.scheduler = scheduler
+        self.transformer = transformer
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        if batch.height is None or batch.width is None:
+            raise ValueError("height and width must be provided for Kandinsky5.")
+        height = int(batch.height)
+        width = int(batch.width)
+        num_frames = int(batch.num_frames)
+
+        temporal_ratio = fastvideo_args.pipeline_config.vae_config.arch_config.temporal_compression_ratio
+        spatial_ratio = fastvideo_args.pipeline_config.vae_config.arch_config.spatial_compression_ratio
+        patch_size = fastvideo_args.pipeline_config.dit_config.arch_config.patch_size
+
+        if num_frames % temporal_ratio != 1:
+            num_frames = num_frames // temporal_ratio * temporal_ratio + 1
+            batch.num_frames = num_frames
+
+        required_divisor_h = spatial_ratio * patch_size[1]
+        required_divisor_w = spatial_ratio * patch_size[2]
+        if height % required_divisor_h != 0 or width % required_divisor_w != 0:
+            raise ValueError(f"Kandinsky5 height must be divisible by {required_divisor_h} and width by "
+                             f"{required_divisor_w}; "
+                             f"got height={height}, width={width}.")
+
+        # NABLA sparse attention (Pro checkpoints) reshapes the post-patch grid
+        # into 8x8 blocks; validate here instead of crashing mid-denoise after
+        # all the encoding work is done.
+        arch_cfg = getattr(self.transformer, "config", None) or fastvideo_args.pipeline_config.dit_config.arch_config
+        if getattr(arch_cfg, "attention_type", "regular") == "nabla":
+            nabla_divisor_h = required_divisor_h * 8
+            nabla_divisor_w = required_divisor_w * 8
+            if height % nabla_divisor_h != 0 or width % nabla_divisor_w != 0:
+                raise ValueError(f"Kandinsky5 NABLA checkpoints require height divisible by {nabla_divisor_h} and "
+                                 f"width divisible by {nabla_divisor_w}; "
+                                 f"got height={height}, width={width}.")
+
+        if isinstance(batch.prompt, list):
+            batch_size = len(batch.prompt)
+        elif batch.prompt is not None:
+            batch_size = 1
+        else:
+            batch_size = batch.prompt_embeds[0].shape[0]
+        batch_size *= batch.num_videos_per_prompt
+
+        dtype = PRECISION_TO_TYPE[fastvideo_args.pipeline_config.dit_precision]
+        device = get_local_torch_device()
+        num_latent_frames = (num_frames - 1) // temporal_ratio + 1
+        num_channels = getattr(
+            self.transformer,
+            "in_visual_dim",
+            fastvideo_args.pipeline_config.dit_config.arch_config.in_visual_dim,
+        )
+        shape = (
+            batch_size,
+            num_latent_frames,
+            height // spatial_ratio,
+            width // spatial_ratio,
+            num_channels,
+        )
+
+        if isinstance(batch.generator, list) and len(batch.generator) != batch_size:
+            raise ValueError(f"generator list length {len(batch.generator)} does not match batch size {batch_size}.")
+
+        visual_cond = getattr(self.transformer, "visual_cond", False)
+
+        if batch.latents is None:
+            latents = randn_tensor(shape, generator=batch.generator, device=device, dtype=dtype)
+            if hasattr(self.scheduler, "init_noise_sigma"):
+                latents = latents * self.scheduler.init_noise_sigma
+        else:
+            valid_shapes = [shape]
+            if visual_cond:
+                valid_shapes.append((*shape[:-1], num_channels * 2 + 1))
+            if tuple(batch.latents.shape) not in valid_shapes:
+                raise ValueError(f"Provided latents shape {list(batch.latents.shape)} does not match expected "
+                                 f"Kandinsky5 latent shape(s): {[list(s) for s in valid_shapes]}.")
+            latents = batch.latents.to(device=device, dtype=dtype)
+
+        if visual_cond and latents.shape[-1] == num_channels:
+            cond = torch.zeros_like(latents)
+            cond_mask = torch.zeros(
+                (*latents.shape[:-1], 1),
+                device=latents.device,
+                dtype=latents.dtype,
+            )
+            latents = torch.cat([latents, cond, cond_mask], dim=-1)
+
+        # I2V image conditioning is placed by Kandinsky5ImageEncodingStage,
+        # which runs AFTER this stage so the initial noise is the generator's
+        # first draw (matching the official kandinskylab/kandinsky-5 order).
+        batch.latents = latents
+        batch.raw_latent_shape = (
+            batch_size,
+            num_channels,
+            num_latent_frames,
+            height // spatial_ratio,
+            width // spatial_ratio,
+        )
+        return batch
+
+    def verify_input(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("num_frames", batch.num_frames, V.positive_int)
+        result.add_check("height", batch.height, V.positive_int)
+        result.add_check("width", batch.width, V.positive_int)
+        result.add_check("latents", batch.latents, V.none_or_tensor)
+        return result
+
+    def verify_output(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("latents", batch.latents, [V.is_tensor, V.with_dims(5)])
+        return result
+
+
+class Kandinsky5DenoisingStage(PipelineStage):
+
+    def __init__(self, transformer, scheduler) -> None:
+        super().__init__()
+        self.transformer = transformer
+        self.scheduler = scheduler
+
+    @staticmethod
+    def _scale_factor(height: int, width: int) -> tuple[float, float, float]:
+        if 480 <= height <= 854 and 480 <= width <= 854:
+            return (1.0, 2.0, 2.0)
+        return (1.0, 3.16, 3.16)
+
+    @staticmethod
+    def _text_rope_pos(mask: torch.Tensor, device: torch.device) -> torch.Tensor:
+        seq_len = int(mask.sum(1).max().item())
+        return torch.arange(seq_len, device=device)
+
+    @staticmethod
+    def fast_sta_nabla(
+        T: int,
+        H: int,
+        W: int,
+        wT: int = 3,
+        wH: int = 3,
+        wW: int = 3,
+        device: torch.device | str = "cuda",
+    ) -> torch.Tensor:
+        """
+        Create a sparse temporal attention (STA) mask for efficient video generation.
+
+        This method generates a mask that limits attention to nearby frames and spatial positions, reducing
+        computational complexity for video generation.
+
+        Args:
+            T (int): Number of temporal frames
+            H (int): Height in latent space
+            W (int): Width in latent space
+            wT (int): Temporal attention window size
+            wH (int): Height attention window size
+            wW (int): Width attention window size
+            device (str): Device to create tensor on
+
+        Returns:
+            torch.Tensor: Sparse attention mask of shape (T*H*W, T*H*W)
+        """
+        max_extent = int(torch.tensor([T, H, W], device=device).amax().item())
+        r = torch.arange(0, max_extent, 1, dtype=torch.int16, device=device)
+        mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs()
+        sta_t, sta_h, sta_w = (
+            mat[:T, :T].flatten(),
+            mat[:H, :H].flatten(),
+            mat[:W, :W].flatten(),
+        )
+        sta_t = sta_t <= wT // 2
+        sta_h = sta_h <= wH // 2
+        sta_w = sta_w <= wW // 2
+        sta_hw = (sta_h.unsqueeze(1) * sta_w.unsqueeze(0)).reshape(H, H, W, W).transpose(1, 2).flatten()
+        sta = (sta_t.unsqueeze(1) * sta_hw.unsqueeze(0)).reshape(T, T, H * W, H * W).transpose(1, 2)
+        return sta.reshape(T * H * W, T * H * W)
+
+    def get_sparse_params(self, sample: torch.Tensor, device: torch.device) -> dict[str, Any] | None:
+        """
+        Generate sparse attention parameters for the transformer based on sample dimensions.
+
+        This method computes the sparse attention configuration needed for efficient video processing in the
+        transformer model.
+
+        Args:
+            sample (torch.Tensor): Input sample tensor
+            device (torch.device): Device to place tensors on
+
+        Returns:
+            Dict: Dictionary containing sparse attention parameters
+        """
+        assert self.transformer.config.patch_size[0] == 1
+        _, T, H, W, _ = sample.shape
+        T, H, W = (
+            T // self.transformer.config.patch_size[0],
+            H // self.transformer.config.patch_size[1],
+            W // self.transformer.config.patch_size[2],
+        )
+        if self.transformer.config.attention_type == "nabla":
+            sta_mask = self.fast_sta_nabla(
+                T,
+                H // 8,
+                W // 8,
+                self.transformer.config.attention_wT,
+                self.transformer.config.attention_wH,
+                self.transformer.config.attention_wW,
+                device=device,
+            )
+
+            sparse_params = {
+                "sta_mask": sta_mask.unsqueeze_(0).unsqueeze_(0),
+                "attention_type": self.transformer.config.attention_type,
+                "to_fractal": True,
+                "P": self.transformer.config.attention_P,
+                "wT": self.transformer.config.attention_wT,
+                "wW": self.transformer.config.attention_wW,
+                "wH": self.transformer.config.attention_wH,
+                "add_sta": self.transformer.config.attention_add_sta,
+                "visual_shape": (T, H, W),
+                "method": self.transformer.config.attention_method,
+            }
+        else:
+            sparse_params = None
+
+        return sparse_params
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        if batch.timesteps is None:
+            raise ValueError("timesteps must be prepared before Kandinsky5 denoising.")
+        if batch.latents is None:
+            raise ValueError("latents must be prepared before Kandinsky5 denoising.")
+        if not fastvideo_args.model_loaded["transformer"]:
+            loader = TransformerLoader()
+            self.transformer = loader.load(fastvideo_args.model_paths["transformer"], fastvideo_args)
+            fastvideo_args.model_loaded["transformer"] = True
+
+        device = get_local_torch_device()
+        target_dtype = PRECISION_TO_TYPE[fastvideo_args.pipeline_config.dit_precision]
+        autocast_enabled = target_dtype != torch.float32 and not fastvideo_args.disable_autocast
+        latents = batch.latents
+        num_channels = getattr(
+            self.transformer,
+            "in_visual_dim",
+            fastvideo_args.pipeline_config.dit_config.arch_config.in_visual_dim,
+        )
+
+        prompt_embeds = batch.prompt_embeds[0].to(device=device, dtype=target_dtype)
+        pooled = batch.prompt_embeds[1].to(device=device, dtype=target_dtype)
+        if batch.prompt_attention_mask is None or not batch.prompt_attention_mask:
+            raise ValueError("Kandinsky5 requires Qwen prompt attention masks.")
+        text_rope_pos = self._text_rope_pos(batch.prompt_attention_mask[0].to(device), device)
+
+        neg_prompt_embeds = None
+        neg_pooled = None
+        negative_text_rope_pos = None
+        if batch.do_classifier_free_guidance and batch.negative_prompt_embeds:
+            neg_prompt_embeds = batch.negative_prompt_embeds[0].to(device=device, dtype=target_dtype)
+            neg_pooled = batch.negative_prompt_embeds[1].to(device=device, dtype=target_dtype)
+            if batch.negative_attention_mask is None or not batch.negative_attention_mask:
+                raise ValueError("Kandinsky5 requires Qwen negative attention masks for CFG.")
+            negative_text_rope_pos = self._text_rope_pos(batch.negative_attention_mask[0].to(device), device)
+
+        height = int(batch.height)
+        width = int(batch.width)
+        temporal_ratio = fastvideo_args.pipeline_config.vae_config.arch_config.temporal_compression_ratio
+        spatial_ratio = fastvideo_args.pipeline_config.vae_config.arch_config.spatial_compression_ratio
+        num_latent_frames = (int(batch.num_frames) - 1) // temporal_ratio + 1
+        visual_rope_pos = [
+            torch.arange(num_latent_frames, device=device),
+            torch.arange(height // spatial_ratio // 2, device=device),
+            torch.arange(width // spatial_ratio // 2, device=device),
+        ]
+        scale_factor = self._scale_factor(height, width)
+
+        sparse_params = self.get_sparse_params(latents, device)
+
+        # I2V keeps the first (conditioning) frame fixed during denoising.
+        # Key off the actual image conditioning, not transformer.visual_cond:
+        # official T2V checkpoints also ship visual_cond=True, and skipping
+        # frame 0 for them leaves it as undenoised noise.
+        cond_frames = 1 if batch.image_latent is not None else 0
+
+        trajectory_timesteps: list[torch.Tensor] = []
+        trajectory_latents: list[torch.Tensor] = []
+
+        with tqdm(total=batch.num_inference_steps, desc="Kandinsky5 Denoising") as progress_bar:
+            for i, timestep in enumerate(batch.timesteps):
+                if hasattr(self, "interrupt") and self.interrupt:
+                    break
+
+                t_expand = timestep.unsqueeze(0).repeat(latents.shape[0]).to(device=device, dtype=target_dtype)
+                attn_metadata = None
+                if sparse_params is not None:
+                    attn_metadata = NablaAttentionMetadataBuilder().build(
+                        current_timestep=i,
+                        sta_mask=sparse_params["sta_mask"],
+                        P=sparse_params["P"],
+                        visual_shape=sparse_params["visual_shape"],
+                    )
+                autocast_ctx = (torch.autocast(device_type="cuda", dtype=target_dtype, enabled=autocast_enabled)
+                                if device.type == "cuda" else contextlib.nullcontext())
+                with set_forward_context(current_timestep=i, attn_metadata=attn_metadata,
+                                         forward_batch=batch), autocast_ctx:
+                    pred_velocity = self.transformer(
+                        hidden_states=latents.to(dtype=target_dtype),
+                        encoder_hidden_states=prompt_embeds,
+                        pooled_projections=pooled,
+                        timestep=t_expand,
+                        visual_rope_pos=visual_rope_pos,
+                        text_rope_pos=text_rope_pos,
+                        scale_factor=scale_factor,
+                        sparse_params=sparse_params,
+                        return_dict=True,
+                    ).sample
+
+                    if neg_prompt_embeds is not None and neg_pooled is not None:
+                        uncond_pred_velocity = self.transformer(
+                            hidden_states=latents.to(dtype=target_dtype),
+                            encoder_hidden_states=neg_prompt_embeds,
+                            pooled_projections=neg_pooled,
+                            timestep=t_expand,
+                            visual_rope_pos=visual_rope_pos,
+                            text_rope_pos=negative_text_rope_pos,
+                            scale_factor=scale_factor,
+                            sparse_params=sparse_params,
+                            return_dict=True,
+                        ).sample
+                        pred_velocity = uncond_pred_velocity + batch.guidance_scale * (pred_velocity -
+                                                                                       uncond_pred_velocity)
+
+                latents[:, cond_frames:, :, :, :num_channels] = self.scheduler.step(
+                    pred_velocity[:, cond_frames:],
+                    timestep,
+                    latents[:, cond_frames:, :, :, :num_channels],
+                    return_dict=False,
+                )[0]
+
+                if batch.return_trajectory_latents:
+                    trajectory_timesteps.append(timestep)
+                    # latents is mutated in place, so snapshot a channels-first copy.
+                    trajectory_latents.append(latents[..., :num_channels].permute(0, 4, 1, 2, 3).cpu())
+
+                if i == len(batch.timesteps) - 1 or (i + 1) % self.scheduler.order == 0:
+                    progress_bar.update()
+
+        if trajectory_latents:
+            batch.trajectory_latents = torch.stack(trajectory_latents, dim=1)
+            batch.trajectory_timesteps = torch.stack(trajectory_timesteps, dim=0).cpu()
+
+        batch.latents = latents[:, :, :, :, :num_channels]
+        return batch
+
+    def verify_input(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("latents", batch.latents, [V.is_tensor, V.with_dims(5)])
+        result.add_check("prompt_embeds", batch.prompt_embeds, V.min_list_length(2))
+        return result
+
+
+class Kandinsky5DecodingStage(DecodingStage):
+
+    def __init__(self, vae: ParallelTiledVAE, pipeline=None) -> None:
+        super().__init__(vae=vae, pipeline=pipeline)
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        if batch.latents is None:
+            raise ValueError("latents must be available before Kandinsky5 decoding.")
+        # Kandinsky5 latents are channels-last [B, T, H, W, C]; the base stage
+        # (and the trajectory latents recorded by the denoising stage) work
+        # channels-first.
+        batch.latents = batch.latents.permute(0, 4, 1, 2, 3).contiguous()
+        return super().forward(batch, fastvideo_args)
+
+
+class Kandinsky5ImageEncodingStage(EncodingStage):
+    """Encode the conditioning image into a VAE latent for I2V."""
+
+    def __init__(self, vae: ParallelTiledVAE, pipeline=None) -> None:
+        super().__init__(vae=vae)
+
+    @staticmethod
+    def _preprocess(image, height: int, width: int) -> torch.Tensor:
+        if isinstance(image, PIL.Image.Image):
+            image = resize(image, height, width)
+            image = numpy_to_pt(pil_to_numpy(image))  # always lands in [0, 1]
+            return normalize(image)  # [0, 1] -> [-1, 1]
+        # Tensor input: no reliable way to tell [0, 1] from an already
+        # normalized [-1, 1] tensor whose values happen to be non-negative,
+        # so mirror diffusers' heuristic and say what we assumed.
+        if image.min() >= 0:
+            logger.warning("Kandinsky5 conditioning image tensor has no negative values; "
+                           "assuming range [0, 1] and normalizing to [-1, 1]. "
+                           "Pass a [-1, 1] tensor with negative values to skip normalization.")
+            image = normalize(image)  # [0, 1] -> [-1, 1]
+        if image.ndim == 3:
+            image = image.unsqueeze(0)
+        if image.shape[-2:] != (height, width):
+            image = torch.nn.functional.interpolate(image.float(),
+                                                    size=(height, width),
+                                                    mode="bilinear",
+                                                    antialias=True)
+        return image
+
+    @torch.no_grad()
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        if batch.pil_image is None:
+            raise ValueError("Kandinsky5 I2V requires an input image.")
+
+        if not fastvideo_args.model_loaded["vae"]:
+            vae = getattr(self, "vae", None)
+            if vae is None:
+                loader = VAELoader()
+                vae = loader.load(fastvideo_args.model_paths["vae"], fastvideo_args)
+                self.vae = vae
+            fastvideo_args.model_loaded["vae"] = True
+
+        device = get_local_torch_device()
+        vae = self.vae.to(device)
+        self.vae = vae
+        vae_dtype = PRECISION_TO_TYPE[fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = vae_dtype != torch.float32 and not fastvideo_args.disable_autocast
+
+        # [B, C, H, W] -> [B, C, 1, H, W]
+        image = self._preprocess(batch.pil_image, int(batch.height), int(batch.width))
+        image = image.to(device=device, dtype=torch.float32).unsqueeze(2)
+
+        # Encode the single conditioning frame without tiling (matches diffusers).
+        # The untested causal-VAE spatial_tiled_encode path corrupts the latent.
+        prev_use_tiling = vae.use_tiling
+        vae.use_tiling = False
+        try:
+            with torch.autocast(device_type="cuda", dtype=vae_dtype, enabled=vae_autocast_enabled):
+                if not vae_autocast_enabled:
+                    image = image.to(vae_dtype)
+                # Sample with the batch generator (diffusers parity); mode()
+                # would make seed-for-seed reproduction of the reference
+                # pipeline impossible.
+                generator = batch.generator
+                if isinstance(generator, list) and len(generator) != image.shape[0]:
+                    generator = generator[0]
+                image_latent = vae.encode(image).sample(generator=generator)
+        finally:
+            vae.use_tiling = prev_use_tiling
+
+        image_latent = image_latent * vae.scaling_factor
+        # [B, C, 1, H, W] -> [B, 1, H, W, C] to match channels-last latents
+        batch.image_latent = image_latent.permute(0, 2, 3, 4, 1).contiguous()
+
+        # Place the conditioning latent into the prepared latents: frame 0 of
+        # the main channels, the visual_cond channel block, and the mask.
+        # NOTE: the official kandinsky-5 repo leaves the visual_cond block
+        # zeros (generation_utils.py generate()), while the diffusers port
+        # copies the image latent into it. A same-seed A/B on
+        # Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers showed the Diffusers
+        # export requires the copy: zeroing the block produces smeared faces
+        # mid-video. Keep the diffusers semantics for Diffusers-format
+        # checkpoints.
+        latents = batch.latents
+        image_latent = batch.image_latent.to(device=latents.device, dtype=latents.dtype)
+        num_channels = image_latent.shape[-1]
+        latents[:, 0:1, :, :, :num_channels] = image_latent
+        if latents.shape[-1] > num_channels:
+            latents[:, 0:1, :, :, num_channels:2 * num_channels] = image_latent
+            latents[:, 0:1, :, :, 2 * num_channels:] = 1.0
+        batch.latents = latents
+
+        if fastvideo_args.vae_cpu_offload:
+            vae.to("cpu")
+        return batch
+
+    def verify_input(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("pil_image", batch.pil_image, V.not_none)
+        result.add_check("height", batch.height, V.positive_int)
+        result.add_check("width", batch.width, V.positive_int)
+        # This stage runs after latent preparation and writes into its output.
+        result.add_check("latents", batch.latents, [V.is_tensor, V.with_dims(5)])
+        return result
+
+    def verify_output(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> VerificationResult:
+        result = VerificationResult()
+        result.add_check("image_latent", batch.image_latent, [V.is_tensor, V.with_dims(5)])
+        return result
+
+
+class Kandinsky5NormalizationStage(PipelineStage):
+    """Normalize the first latent frames to reduce I2V conditioning artifacts."""
+
+    COND_FRAMES = 4
+    REFERENCE_FRAMES = 5
+
+    @staticmethod
+    def _adaptive_mean_std(source: torch.Tensor, reference: torch.Tensor) -> torch.Tensor:
+        source_mean = source.mean(dim=(1, 2, 3, 4), keepdim=True)
+        source_std = source.std(dim=(1, 2, 3, 4), keepdim=True)
+        # Magic constants limit how far the first frames may drift.
+        ref_mean = torch.clamp(reference.mean(dim=(1, 2, 3, 4), keepdim=True), source_mean - 0.05, source_mean + 0.1)
+        ref_std = torch.clamp(reference.std(dim=(1, 2, 3, 4), keepdim=True), source_std - 0.1, source_std + 0.25)
+        normalized = (source - source_mean) / source_std
+        return normalized * ref_std + ref_mean
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> ForwardBatch:
+        latents = batch.latents
+        n = self.COND_FRAMES
+        if latents is None or latents.shape[1] <= n:
+            return batch
+
+        reference = latents[:, n:n + min(self.REFERENCE_FRAMES, latents.shape[1] - 1)]
+        latents[:, :n] = self._adaptive_mean_std(latents[:, :n].clone(), reference)
+        batch.latents = latents
+        return batch

fastvideo/pipelines/stages/text_encoding.py

@@ -8,6 +8,8 @@ This module contains implementations of prompt encoding stages for diffusion pip
 import torch
 from typing import Any
 
+from torch.distributed.tensor import DTensor
+
 from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.forward_context import set_forward_context
@@ -203,6 +205,21 @@ class TextEncodingStage(PipelineStage):
             encoder_config = encoder_cfgs[i]
             preprocess_func = preprocess_funcs[i]
             postprocess_func = postprocess_funcs[i]
+            # cpu_offload semantics: params rest on CPU between calls but the
+            # forward computes on GPU. FSDP2-wrapped encoders (CPUOffloadPolicy;
+            # DTensor params) stream themselves per-layer — leave inputs on the
+            # param device and let FSDP's root pre-forward move them. A plain
+            # module parked on CPU by text_encoder_cpu_offload is swapped to the
+            # target device for the forward and back afterwards, mirroring the
+            # image-encoder/VAE offload pattern.
+            first_param = next(text_encoder.parameters(), None)
+            encoder_device = first_param.device if first_param is not None else torch.device(target_device)
+            moved_for_forward = False
+            if (first_param is not None and not isinstance(first_param, DTensor)
+                    and encoder_device.type != torch.device(target_device).type):
+                text_encoder = text_encoder.to(target_device)
+                encoder_device = torch.device(target_device)
+                moved_for_forward = True
 
             tok_kwargs = dict(encoder_config.tokenizer_kwargs)
             if max_length is not None:
@@ -247,7 +264,7 @@ class TextEncodingStage(PipelineStage):
                     # pre-format prompts into message lists upstream and rely on
                     # the inner tokenizer + full tokenizer_kwargs (which include
                     # add_generation_prompt). Preserve that original path exactly.
-                    text_inputs = tok.apply_chat_template(processed_texts, **tok_kwargs).to(target_device)
+                    text_inputs = tok.apply_chat_template(processed_texts, **tok_kwargs).to(encoder_device)
                 else:
                     # Two-step approach matching Diffusers: format with chat
                     # template first, then tokenize the resulting strings.
@@ -261,9 +278,9 @@ class TextEncodingStage(PipelineStage):
                             enable_thinking=False,
                         )
                         formatted_texts.append(formatted)
-                    text_inputs = tokenizer(formatted_texts, **tok_kwargs).to(target_device)
+                    text_inputs = tokenizer(formatted_texts, **tok_kwargs).to(encoder_device)
             else:
-                text_inputs = tok(processed_texts, **tok_kwargs).to(target_device)
+                text_inputs = tok(processed_texts, **tok_kwargs).to(encoder_device)
 
             input_ids = text_inputs["input_ids"]
             attention_mask = text_inputs["attention_mask"]
@@ -283,15 +300,18 @@ class TextEncodingStage(PipelineStage):
             except Exception:
                 prompt_embeds, attention_mask = postprocess_func(outputs, attention_mask)
             if is_ltx2 and getattr(outputs, "hidden_states", None):
-                audio_embed = outputs.hidden_states[0]
+                audio_embed = outputs.hidden_states[0].to(device=target_device)
                 if dtype is not None:
                     audio_embed = audio_embed.to(dtype=dtype)
                 audio_embeds_list.append(audio_embed)
+            prompt_embeds = prompt_embeds.to(device=target_device)
             if dtype is not None:
                 prompt_embeds = prompt_embeds.to(dtype=dtype)
             embeds_list.append(prompt_embeds)
             if return_attention_mask:
-                attn_masks_list.append(attention_mask)
+                attn_masks_list.append(attention_mask.to(device=target_device))
+            if moved_for_forward and fastvideo_args.text_encoder_cpu_offload:
+                text_encoder.to("cpu")
         self._last_audio_embeds = audio_embeds_list if is_ltx2 else None
         return self.return_embeds(embeds_list, attn_masks_list, return_type, return_attention_mask, indices)
 

fastvideo/pipelines/stages/validators.py

@@ -219,6 +219,16 @@ class StageValidators:
 
         return validator
 
+    @staticmethod
+    def min_list_length(min_length: int) -> Callable[[Any], bool]:
+        """Return a validator that checks if a list has at least min_length items."""
+
+        def validator(value: Any) -> bool:
+            return StageValidators.list_min_length(value, min_length)
+
+        validator.__name__ = f"list_min_length_{min_length}"
+        return validator
+
     @staticmethod
     def min_dims(min_dims: int) -> Callable[[Any], bool]:
         """Return a validator that checks if tensor has at least min_dims dimensions and no NaN values."""

fastvideo/platforms/cuda.py

@@ -157,6 +157,14 @@ class CudaPlatformBase(Platform):
                 "ATTN_QAT_TRAIN selected but fastvideo_kernel.triton_kernels.attn_qat_train is not built. "
                 "Silent fallback would produce a non-QAT training run; refusing to proceed. "
                 "Install the training kernel or pick a different FASTVIDEO_ATTENTION_BACKEND.")
+        elif selected_backend == AttentionBackendEnum.NABLA_ATTN:
+            from fastvideo.attention.backends.nabla import CAN_USE_FLEX_ATTN
+            if CAN_USE_FLEX_ATTN:
+                logger.info("Using NABLA block-sparse flex-attention backend.")
+                return "fastvideo.attention.backends.nabla.NablaAttentionBackend"
+            raise ImportError("NABLA_ATTN selected but torch.nn.attention.flex_attention is unavailable in this "
+                              "PyTorch build. Silent fallback to dense attention would be orders of magnitude "
+                              "slower and diverge from the reference; upgrade PyTorch or pick a different backend.")
         elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
             try:
                 from fastvideo_kernel import video_sparse_attn  # noqa: F401

fastvideo/platforms/interface.py

@@ -22,6 +22,7 @@ class AttentionBackendEnum(enum.Enum):
     VMOBA_ATTN = enum.auto()
     SLA_ATTN = enum.auto()
     SAGE_SLA_ATTN = enum.auto()
+    NABLA_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()
 
 

fastvideo/registry.py

@@ -20,6 +20,7 @@ from fastvideo.configs.pipelines.cosmos2_5 import (
     Cosmos25Config,
     Cosmos25_14BConfig,
 )
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BARPipelineConfig, DreamXWorld5BCamPipelineConfig
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuangamecraft import HunyuanGameCraftPipelineConfig
 from fastvideo.configs.pipelines.gen3c import Gen3CConfig
@@ -27,6 +28,7 @@ from fastvideo.configs.pipelines.hunyuan15 import (Hunyuan15T2V480PConfig, Hunyu
                                                    Hunyuan15T2V720PConfig, Hunyuan15I2V720PConfig,
                                                    Hunyuan15SR1080PConfig)
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
+from fastvideo.configs.pipelines.kandinsky5 import Kandinsky5I2VConfig, Kandinsky5T2VConfig
 from fastvideo.configs.pipelines.lingbotworld import LingBotWorldI2V480PConfig
 from fastvideo.configs.pipelines.longcat import LongCatT2V480PConfig
 from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
@@ -117,6 +119,10 @@ class ConfigInfo:
     workload_types: tuple[WorkloadType, ...]
     model_family: str | None = None
     default_preset: str | None = None
+    # When set, overrides the model_index `_class_name` for pipeline resolution.
+    # Lets a model family map to a specific pipeline class by path/detector
+    # (e.g. a T2V and I2V checkpoint that share a `_class_name`).
+    pipeline_cls_name: str | None = None
 
 
 # The central registry mapping a model name to its configuration information
@@ -137,6 +143,7 @@ def register_configs(
     model_detectors: list[Callable[[str], bool]] | None = None,
     model_family: str | None = None,
     default_preset: str | None = None,
+    pipeline_cls_name: str | None = None,
 ) -> None:
     """Register config classes for a model family.
 
@@ -151,6 +158,7 @@ def register_configs(
         workload_types=workload_types,
         model_family=model_family,
         default_preset=default_preset,
+        pipeline_cls_name=pipeline_cls_name,
     )
 
     if hf_model_paths:
@@ -490,18 +498,174 @@ def _register_configs() -> None:
         default_preset="lingbotworld_i2v",
     )
 
+    def _kandinsky5_detector(require: tuple[str, ...] = (), exclude: tuple[str, ...] = ()) -> Callable[[str], bool]:
+
+        def detect(path: str) -> bool:
+            path_lower = path.lower()
+            if "kandinsky5" not in path_lower and "kandinsky-5" not in path_lower:
+                return False
+            return (all(token in path_lower for token in require) and not any(token in path_lower for token in exclude))
+
+        return detect
+
+    # t2v/i2v exclude each other so a checkpoint stored under a directory
+    # containing the other token (e.g. ~/i2v_experiments/kandinsky5-t2v-ft)
+    # falls through to the model_index _class_name fallback detectors below
+    # instead of being misrouted.
+    _is_kandinsky5_t2v = _kandinsky5_detector(require=("t2v", ), exclude=("i2v", ))
+    _is_kandinsky5_i2v = _kandinsky5_detector(require=("i2v", ), exclude=("t2v", ))
+    _is_kandinsky5_t2v_lite = _kandinsky5_detector(require=("t2v", "lite"), exclude=("i2v", "distilled"))
+    _is_kandinsky5_t2v_pro = _kandinsky5_detector(require=("t2v", "pro"), exclude=("i2v", "distilled"))
+    _is_kandinsky5_t2v_lite_distilled = _kandinsky5_detector(require=("t2v", "lite", "distilled"), exclude=("i2v", ))
+    _is_kandinsky5_t2v_pro_distilled = _kandinsky5_detector(require=("t2v", "pro", "distilled"), exclude=("i2v", ))
+    _is_kandinsky5_i2v_lite = _kandinsky5_detector(require=("i2v", "lite"), exclude=("t2v", "distilled"))
+    _is_kandinsky5_i2v_pro = _kandinsky5_detector(require=("i2v", "pro"), exclude=("t2v", "distilled"))
+    _is_kandinsky5_i2v_lite_distilled = _kandinsky5_detector(require=("i2v", "lite", "distilled"), exclude=("t2v", ))
+    _is_kandinsky5_i2v_pro_distilled = _kandinsky5_detector(require=("i2v", "pro", "distilled"), exclude=("t2v", ))
+
     # Kandinsky5 Lite T2V
     register_configs(
         sampling_param_cls=None,
-        pipeline_config_cls=PipelineConfig,
+        pipeline_config_cls=Kandinsky5T2VConfig,
         workload_types=(WorkloadType.T2V, ),
         hf_model_paths=[
             "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
         ],
         model_detectors=[
-            lambda path: any(token in path.lower() for token in ("kandinsky5", "kandinsky-5")),
+            _is_kandinsky5_t2v_lite,
         ],
         model_family="kandinsky5",
+        default_preset="kandinsky5_t2v_lite_5s",
+        pipeline_cls_name="Kandinsky5T2VPipeline",
+    )
+
+    # Kandinsky5 Pro T2V
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5T2VConfig,
+        workload_types=(WorkloadType.T2V, ),
+        hf_model_paths=[
+            "kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers",
+        ],
+        model_detectors=[
+            _is_kandinsky5_t2v_pro,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_t2v_pro_5s",
+    )
+
+    # Kandinsky5 Lite T2V Distilled
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5T2VConfig,
+        workload_types=(WorkloadType.T2V, ),
+        hf_model_paths=[
+            "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers",
+        ],
+        model_detectors=[
+            _is_kandinsky5_t2v_lite_distilled,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_t2v_lite_distilled_5s",
+    )
+
+    # Kandinsky5 Pro T2V Distilled
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5T2VConfig,
+        workload_types=(WorkloadType.T2V, ),
+        hf_model_paths=[
+            "kandinskylab/Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers",
+        ],
+        model_detectors=[
+            _is_kandinsky5_t2v_pro_distilled,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_t2v_pro_distilled_5s",
+    )
+
+    # Kandinsky5 Lite I2V
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5I2VConfig,
+        workload_types=(WorkloadType.I2V, ),
+        hf_model_paths=["kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers"],
+        model_detectors=[
+            _is_kandinsky5_i2v_lite,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_i2v_lite_5s",
+        pipeline_cls_name="Kandinsky5I2VPipeline",
+    )
+
+    # Kandinsky5 Pro I2V
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5I2VConfig,
+        workload_types=(WorkloadType.I2V, ),
+        hf_model_paths=["kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers"],
+        model_detectors=[
+            _is_kandinsky5_i2v_pro,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_i2v_pro_5s",
+    )
+
+    # Kandinsky5 Pro I2V Distilled
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5I2VConfig,
+        workload_types=(WorkloadType.I2V, ),
+        hf_model_paths=["kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers"],
+        model_detectors=[
+            _is_kandinsky5_i2v_pro_distilled,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_i2v_pro_distilled_5s",
+    )
+
+    # Kandinsky5 Lite I2V Distilled (no official hub repo yet; local
+    # conversions get distilled sampling defaults instead of the sft ones the
+    # fallback would apply).
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5I2VConfig,
+        workload_types=(),
+        model_detectors=[
+            _is_kandinsky5_i2v_lite_distilled,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_i2v_lite_distilled_5s",
+        pipeline_cls_name="Kandinsky5I2VPipeline",
+    )
+
+    # Kandinsky5 fallbacks — registered AFTER the variant detectors so those
+    # win first-match. Catch checkpoints the variant detectors cannot resolve:
+    # token-less local paths matched via the model_index _class_name
+    # ("kandinsky5t2vpipeline" carries no lite/pro marker), variant combos
+    # without a dedicated entry (e.g. I2V Lite distilled), and t2v+i2v
+    # ambiguous paths resolved by the checkpoint's _class_name.
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5T2VConfig,
+        workload_types=(),
+        model_detectors=[
+            _is_kandinsky5_t2v,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_t2v_lite_5s",
+        pipeline_cls_name="Kandinsky5T2VPipeline",
+    )
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=Kandinsky5I2VConfig,
+        workload_types=(),
+        model_detectors=[
+            _is_kandinsky5_i2v,
+        ],
+        model_family="kandinsky5",
+        default_preset="kandinsky5_i2v_lite_5s",
+        pipeline_cls_name="Kandinsky5I2VPipeline",
     )
 
     # LongCat (T2V, I2V, VC use same config; workload varies by path)
@@ -550,7 +714,7 @@ def _register_configs() -> None:
             "FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers",
             "FastVideo/Matrix-Game-2.0-GTA-Distilled-Diffusers",
             "FastVideo/Matrix-Game-2.0-TempleRun-Distilled-Diffusers",
-            # Legacy HF paths (kept for backward compat — pre-rename names):
+            # Legacy HF paths (kept for backward compat - pre-rename names):
             "FastVideo/Matrix-Game-2.0-Base-Diffusers",
             "FastVideo/Matrix-Game-2.0-GTA-Diffusers",
             "FastVideo/Matrix-Game-2.0-TempleRun-Diffusers",
@@ -773,6 +937,40 @@ def _register_configs() -> None:
         model_family="wan",
         default_preset="wan_2_2_ti2v_5b",
     )
+
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=DreamXWorld5BCamPipelineConfig,
+        workload_types=(WorkloadType.I2V, ),
+        hf_model_paths=[
+            "FastVideo/DreamX-World-5B-Cam-Diffusers",
+        ],
+        model_detectors=[
+            # Mutually exclusive with the AR detector below: Cam requires an
+            # explicit "cam" marker so hyphenated AR local paths (e.g.
+            # /ckpts/dreamx-world-5b-converted) don't first-match here —
+            # detector resolution is first-match in registration order.
+            lambda path:
+            ("dreamx-world" in path.lower() and "cam" in path.lower()) or "dreamxworldpipeline" in path.lower()
+        ],
+        model_family="dreamx_world",
+        default_preset="dreamx_world_5b_cam",
+    )
+    register_configs(
+        sampling_param_cls=None,
+        pipeline_config_cls=DreamXWorld5BARPipelineConfig,
+        workload_types=(WorkloadType.I2V, ),
+        hf_model_paths=[
+            "FastVideo/DreamX-World-5B-Diffusers",
+        ],
+        model_detectors=[
+            lambda path:
+            ("dreamx-world-5b" in path.lower() and "cam" not in path.lower()) or "dreamxworldarpipeline" in path.lower(
+            )
+        ],
+        model_family="dreamx_world",
+        default_preset="dreamx_world_5b_ar",
+    )
     register_configs(
         sampling_param_cls=None,
         pipeline_config_cls=FastWan2_2_TI2V_5B_Config,
@@ -902,7 +1100,10 @@ def get_model_info(
     assert config_info is not None, "config_info must be resolved"
 
     if override_pipeline_cls_name:
-        pipeline_name = override_pipeline_cls_name
+        # Explicit override: skip config resolution entirely so checkpoints
+        # without a diffusers model_index.json keep working (and no download
+        # is triggered just to log the replaced name).
+        pipeline_name: str | None = override_pipeline_cls_name
         logger.info("Using override pipeline class name %s", pipeline_name)
     else:
         if os.path.exists(model_path):
@@ -911,6 +1112,12 @@ def get_model_info(
             config = maybe_download_model_index(model_path)
 
         pipeline_name = config.get("_class_name")
+        if config_info.pipeline_cls_name is not None:
+            # The resolved (path/detector-based) config pins the pipeline class,
+            # e.g. an I2V checkpoint whose `_class_name` would otherwise resolve
+            # to the T2V pipeline.
+            logger.info("Pinning pipeline class name from %s to %s", pipeline_name, config_info.pipeline_cls_name)
+            pipeline_name = config_info.pipeline_cls_name
 
     if pipeline_name is None:
         raise ValueError("Model config does not contain a _class_name attribute. "
@@ -951,6 +1158,8 @@ def _register_presets() -> None:
     from fastvideo.api.presets import register_preset
     from fastvideo.pipelines.basic.cosmos.presets import (
         ALL_PRESETS as COSMOS_PRESETS, )
+    from fastvideo.pipelines.basic.dreamx_world.presets import (
+        ALL_PRESETS as DREAMX_WORLD_PRESETS, )
     from fastvideo.pipelines.basic.gamecraft.presets import (
         ALL_PRESETS as GAMECRAFT_PRESETS, )
     from fastvideo.pipelines.basic.gen3c.presets import (
@@ -961,6 +1170,8 @@ def _register_presets() -> None:
         ALL_PRESETS as HUNYUAN15_PRESETS, )
     from fastvideo.pipelines.basic.hyworld.presets import (
         ALL_PRESETS as HYWORLD_PRESETS, )
+    from fastvideo.pipelines.basic.kandinsky5.presets import (
+        ALL_PRESETS as KANDINSKY5_PRESETS, )
     from fastvideo.pipelines.basic.lingbotworld.presets import (
         ALL_PRESETS as LINGBOTWORLD_PRESETS, )
     from fastvideo.pipelines.basic.longcat.presets import (
@@ -984,12 +1195,14 @@ def _register_presets() -> None:
 
     all_preset_groups = (
         COSMOS_PRESETS,
+        DREAMX_WORLD_PRESETS,
         FLUX2_PRESETS,
         GAMECRAFT_PRESETS,
         GEN3C_PRESETS,
         HUNYUAN_PRESETS,
         HUNYUAN15_PRESETS,
         HYWORLD_PRESETS,
+        KANDINSKY5_PRESETS,
         LINGBOTWORLD_PRESETS,
         LONGCAT_PRESETS,
         LTX2_PRESETS,

fastvideo/tests/api/test_compat_translation.py

@@ -17,46 +17,6 @@ class TestLegacyTorchCompileKwargsTranslation:
     first-class :class:`CompileConfig` fields and anything unknown falls
     into ``extras``."""
 
-    def test_all_typed_keys_promoted(self) -> None:
-        config = legacy_from_pretrained_to_config(
-            "/models/ltx2",
-            {
-                "enable_torch_compile": True,
-                "torch_compile_kwargs": {
-                    "backend": "inductor",
-                    "fullgraph": True,
-                    "mode": "max-autotune-no-cudagraphs",
-                    "dynamic": False,
-                },
-            },
-        )
-        compile_config = config.engine.compile
-        assert compile_config.enabled is True
-        assert compile_config.backend == "inductor"
-        assert compile_config.fullgraph is True
-        assert compile_config.mode == "max-autotune-no-cudagraphs"
-        assert compile_config.dynamic is False
-        assert compile_config.extras == {}
-
-    def test_unknown_keys_land_in_extras(self) -> None:
-        config = legacy_from_pretrained_to_config(
-            "/models/ltx2",
-            {
-                "enable_torch_compile": True,
-                "torch_compile_kwargs": {
-                    "backend": "inductor",
-                    "options": {"triton.cudagraphs": False},
-                    "disable": False,
-                },
-            },
-        )
-        compile_config = config.engine.compile
-        assert compile_config.backend == "inductor"
-        assert compile_config.extras == {
-            "options": {"triton.cudagraphs": False},
-            "disable": False,
-        }
-
     def test_empty_kwargs_produces_empty_extras(self) -> None:
         config = legacy_from_pretrained_to_config(
             "/models/ltx2",
@@ -117,13 +77,6 @@ class TestLegacyLtx2VaeTilingTranslation:
     ``generator.pipeline.vae_tiling``; reverse direction emits the
     legacy name back to FastVideoArgs."""
 
-    def test_forward_routes_to_pipeline_vae_tiling(self) -> None:
-        config = legacy_from_pretrained_to_config(
-            "/models/ltx2",
-            {"ltx2_vae_tiling": False},
-        )
-        assert config.pipeline.vae_tiling is False
-
     def test_true_round_trips(self) -> None:
         config = legacy_from_pretrained_to_config(
             "/models/ltx2",

fastvideo/tests/attention/test_sdpa_metadata_mask_contract.py

@@ -0,0 +1,43 @@
+# SPDX-License-Identifier: Apache-2.0
+"""CPU-only regression test for the SDPA metadata mask contract.
+
+SDPAMetadata is cross-backend: HYWorld and HunyuanVideo15 build it while the
+layer's attention selector may pick FLASH_ATTN, and the shared convention
+for padding masks is the tokenizer-style 2D [batch, key_len]. The builder
+must store 2D masks unchanged; only the torch-sdpa impl lifts them to 4D
+internally.
+
+Regression guard for the CI break where the builder lifted 2D -> 4D and the
+FLASH_ATTN equal-length branch padded ``attn_mask.shape[1]`` (== 1 for 4D)
+into a garbage-length mask (CUDA device-side assert, SIGABRT).
+"""
+
+from __future__ import annotations
+
+import torch
+
+from fastvideo.attention.backends.sdpa import (SDPAMetadataBuilder, _normalize_attn_mask_for_sdpa)
+
+
+def test_builder_keeps_2d_padding_mask_2d() -> None:
+    """A tokenizer-style 2D int mask must be stored as-is (cross-backend contract)."""
+    mask = torch.tensor([[1, 1, 0], [1, 0, 0]], dtype=torch.int64)
+    md = SDPAMetadataBuilder().build(current_timestep=0, attn_mask=mask)
+    assert md.attn_mask is not None
+    assert md.attn_mask.dim() == 2
+    assert md.attn_mask.shape == (2, 3)
+    assert torch.equal(md.attn_mask, mask)
+
+
+def test_normalize_lifts_2d_int_mask_to_4d_bool_for_torch_sdpa() -> None:
+    """The torch-sdpa consumer coerces int -> bool and lifts 2D -> [B,1,1,K]."""
+    batch, q_len, k_len, heads, head_dim = 2, 4, 3, 1, 8
+    query = torch.zeros(batch, heads, q_len, head_dim)
+    key = torch.zeros(batch, heads, k_len, head_dim)
+    mask = torch.tensor([[1, 1, 0], [1, 0, 0]], dtype=torch.int64)
+
+    out = _normalize_attn_mask_for_sdpa(mask, query, key)
+    assert out is not None
+    assert out.dtype == torch.bool
+    assert out.shape == (batch, 1, 1, k_len)
+    assert torch.equal(out[:, 0, 0, :], mask != 0)

fastvideo/tests/contract/test_ci_test_collection.py

@@ -35,7 +35,6 @@ ALLOWLIST = {
     "hooks": "no lane yet — run manually",
     "layers": "no lane yet — torchrun FSDP dispatch tests, run manually",
     "nightly": "by design: nightly cadence, not per-PR",
-    "ops": "no lane yet — GPU op tests, run manually",
     "modal": "CI infrastructure itself, not a test suite",
 }
 

fastvideo/tests/encoders/test_t5_encoder.py

@@ -12,6 +12,7 @@ from fastvideo.configs.pipelines import CosmosConfig, PipelineConfig, WanT2V480P
 from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
 from fastvideo.models.loader.component_loader import TextEncoderLoader
+from fastvideo.tests.utils import skip_if_gated_repo_inaccessible
 from fastvideo.utils import maybe_download_model, PRECISION_TO_TYPE
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.configs.models.encoders import T5Config, T5LargeConfig
@@ -36,9 +37,11 @@ def t5_model_paths_and_config():
 @pytest.fixture
 def t5_large_model_paths_and_config():
     base_model_path = "nvidia/Cosmos-Predict2-2B-Video2World"
-    model_path = maybe_download_model(base_model_path,
-                                      local_dir=os.path.join(
-                                          'data', base_model_path))
+    local_dir = os.path.join('data', base_model_path)
+    skip_if_gated_repo_inaccessible(base_model_path,
+                                    local_path=local_dir,
+                                    test_name="Cosmos T5-large encoder test")
+    model_path = maybe_download_model(base_model_path, local_dir=local_dir)
     text_encoder_path = os.path.join(model_path, "text_encoder")
     tokenizer_path = os.path.join(model_path, "tokenizer")
     return text_encoder_path, tokenizer_path, CosmosConfig()

fastvideo/tests/modal/pr_test.py

@@ -187,6 +187,8 @@ def run_transformer_tests():
 
 
 @app.function(gpu="L40S:4",
+              cpu=8.0,
+              memory=32768,
               image=image,
               timeout=900,
               secrets=[
@@ -201,6 +203,8 @@ def run_training_tests():
 
 
 @app.function(gpu="L40S:2",
+              cpu=8.0,
+              memory=32768,
               image=image,
               timeout=900,
               secrets=[
@@ -276,7 +280,7 @@ def run_self_forcing_tests():
 @app.function(gpu="L40S:1", image=image, timeout=900)
 def run_unit_test():
     run_test(
-        "pytest ./fastvideo/tests/api/ ./fastvideo/tests/contract/ ./fastvideo/tests/dataset/ ./fastvideo/tests/workflow/ ./fastvideo/tests/entrypoints/ ./fastvideo/tests/train/ ./fastvideo/tests/stages/ --ignore=./fastvideo/tests/entrypoints/test_openai_api_integration.py --ignore=./fastvideo/tests/train/models --ignore=./fastvideo/tests/train/methods -vs"
+        "pytest ./fastvideo/tests/api/ ./fastvideo/tests/contract/ ./fastvideo/tests/dataset/ ./fastvideo/tests/workflow/ ./fastvideo/tests/entrypoints/ ./fastvideo/tests/train/ ./fastvideo/tests/stages/ ./fastvideo/tests/ops/ ./fastvideo/tests/training/test_trackers.py ./fastvideo/tests/attention/test_sdpa_metadata_mask_contract.py --ignore=./fastvideo/tests/entrypoints/test_openai_api_integration.py --ignore=./fastvideo/tests/train/models --ignore=./fastvideo/tests/train/methods -vs"
     )
 
 
@@ -318,6 +322,8 @@ def run_dreamverse_app_tests():
 
 
 @app.function(gpu="L40S:1",
+              cpu=8.0,
+              memory=32768,
               image=image,
               timeout=1800,
               secrets=[
@@ -400,6 +406,8 @@ def run_lora_extraction_tests():
 
 
 @app.function(gpu="L40S:2",
+              cpu=8.0,
+              memory=32768,
               image=image,
               timeout=1800,
               secrets=[
@@ -427,6 +435,8 @@ def run_performance_tests():
         "export PERF_RUN_SOURCE='unknown'; "
         "export PERF_UPLOAD_POLICY='never'; "
         "fi; "
+        "(nvidia-smi --query-gpu=index,timestamp,clocks.sm,clocks.max.sm,power.draw,power.limit,temperature.gpu "
+        "--format=csv -l 10 > /tmp/gpu_telemetry.csv 2>/dev/null &); "
         "pytest ./fastvideo/tests/performance -vs; "
         "PYTEST_RC=$?; "
         "PERF_RC=0; "
@@ -435,6 +445,8 @@ def run_performance_tests():
         "PERF_RC=$?; "
         "fi; "
         "python ./fastvideo/tests/performance/dashboard.py || true; "
+        "echo '--- GPU telemetry (clocks.sm vs clocks.max.sm reveals capped hosts) ---'; "
+        "cat /tmp/gpu_telemetry.csv || true; "
         "FINAL_RC=$PYTEST_RC; "
         "if [ $FINAL_RC -eq 0 ]; then FINAL_RC=$PERF_RC; fi; "
         "exit $FINAL_RC")

fastvideo/tests/ops/quantization/test_absmax_fp8.py

@@ -34,6 +34,9 @@ class TestAbsMaxFP8LinearMethod(unittest.TestCase):
             method._convert_scale(scale)
 
     def test_create_weights_rejects_invalid_dtype(self):
+        # float64 is outside the supported set (bf16/fp16/fp32). The test
+        # originally passed float32, which create_weights has accepted since
+        # PR #981 — it never ran in CI, so the mismatch went unnoticed.
         method = AbsMaxFP8LinearMethod()
         layer = nn.Module()
         with self.assertRaisesRegex(AssertionError, "only supports"):
@@ -43,7 +46,7 @@ class TestAbsMaxFP8LinearMethod(unittest.TestCase):
                 output_partition_sizes=[3],
                 input_size=2,
                 output_size=3,
-                params_dtype=torch.float32,
+                params_dtype=torch.float64,
             )
 
     def test_absmax_fp8_parameter_weight_loader(self):

fastvideo/tests/performance/compare_baseline.py

@@ -5,7 +5,8 @@ This script:
 1) reads current benchmark results from fastvideo/tests/performance/results,
 2) syncs the canonical baseline from the configured HF dataset repo,
 3) compares each current record against the median of up to 5 prior
-   baseline-eligible successful records (filtered by gpu_type),
+   baseline-eligible successful records in the same workload/variant/version,
+   GPU, recipe, hardware, and software cohort,
 4) writes normalized records back to the HF dataset repo according to
    PERF_UPLOAD_POLICY,
 5) exits non-zero if any gated metric exceeds both its percent and absolute
@@ -64,6 +65,27 @@ PERF_REPORTS_DIR = os.environ.get("PERF_REPORTS_DIR", "/root/data/perf_reports")
 UPLOAD_POLICY = os.environ.get("PERF_UPLOAD_POLICY", "never").strip().lower()
 VALID_UPLOAD_POLICIES = {"never", "pass", "always"}
 VALID_RUN_SOURCES = {"pr", "local", "scheduled_main", "unknown"}
+IDENTITY_KEYS = (
+    "workload_id",
+    "variant_id",
+    "benchmark_version",
+    "recipe",
+    "recipe_fingerprint",
+    "hardware_profile",
+    "hardware_profile_id",
+    "software_profile",
+    "software_profile_id",
+    "environment_metadata",
+    "environment_fingerprint",
+)
+COMPARISON_IDENTITY_KEYS = (
+    "workload_id",
+    "variant_id",
+    "benchmark_version",
+    "recipe_fingerprint",
+    "hardware_profile_id",
+    "software_profile_id",
+)
 
 
 def _should_persist_tracking() -> bool:
@@ -136,6 +158,46 @@ def _record_metadata(run_source: str, result: dict[str, Any]) -> dict[str, Any]:
     }
 
 
+def _identity_metadata(result: dict[str, Any]) -> dict[str, Any]:
+    metadata = {
+        key: result[key]
+        for key in IDENTITY_KEYS
+        if key in result and result[key] is not None
+    }
+    recipe = result.get("recipe")
+    benchmark = recipe.get("benchmark") if isinstance(recipe, dict) else None
+    if isinstance(benchmark, dict):
+        for key in ("workload_id", "variant_id", "benchmark_version"):
+            if key not in metadata and benchmark.get(key) is not None:
+                metadata[key] = benchmark[key]
+    return metadata
+
+
+def _comparison_identity_filters(record: dict[str, Any]) -> dict[str, str]:
+    missing = [
+        key for key in COMPARISON_IDENTITY_KEYS
+        if key not in record or record[key] is None or (isinstance(record[key], str) and not record[key].strip())
+    ]
+    if len(missing) == len(COMPARISON_IDENTITY_KEYS):
+        # Legacy v1 record: no identity fields at all. Keep the pre-v2
+        # (model_id, gpu_type) rolling-baseline comparison instead of
+        # crashing — v1 configs are documented as loadable.
+        return {}
+    if missing:
+        raise ValueError("Performance record missing required comparison identity fields: " + ", ".join(missing))
+    return {
+        key: str(record[key])
+        for key in COMPARISON_IDENTITY_KEYS
+    }
+
+
+def _format_identity_filters(filters: dict[str, str]) -> str:
+    if not filters:
+        return ""
+    compact = ", ".join(f"{key}={value}" for key, value in filters.items())
+    return f" ({compact})"
+
+
 def _load_current_results() -> list[dict[str, Any]]:
     pattern = os.path.join(RESULTS_DIR, "perf_*.json")
     records: list[dict[str, Any]] = []
@@ -169,7 +231,7 @@ def normalize_performance_result(result: dict[str, Any]) -> dict[str, Any]:
     vae_decode_time = safe_float(result.get("vae_decode_time_s"))
     metric_policies = resolve_metric_policies(result.get("regression_thresholds"))
 
-    return {
+    record = {
         "model_id": model_id,
         "timestamp": timestamp,
         "commit_sha": commit_sha,
@@ -184,6 +246,8 @@ def normalize_performance_result(result: dict[str, Any]) -> dict[str, Any]:
         "success": True,
         **_record_metadata(_detect_run_source(), result),
     }
+    record.update(_identity_metadata(result))
+    return record
 
 
 def _normalize_record(result: dict[str, Any]) -> dict[str, Any]:
@@ -422,17 +486,31 @@ def main() -> int:
             TRACKING_ROOT,
             record["model_id"],
             record["gpu_type"],
+            **_comparison_identity_filters(record),
             last_n=5,
             successful_only=True,
             baseline_eligible_only=True,
         )
 
         if not baseline_records:
-            print(f"No baseline for {record['model_id']} on "
-                  f"{record['gpu_type']}. Initializing...")
+            # A brand-new cohort has NO regression gating until history
+            # accumulates — make that loud and machine-readable instead of
+            # an indistinguishable pass, so a cohort shift (intended or
+            # accidental, e.g. an identity-field change) never silently
+            # blinds the comparison.
+            record["baseline_status"] = "initialized_new_cohort"
+            print("=" * 72)
+            print(f"WARNING: NO BASELINE — initializing a NEW cohort for "
+                  f"{record['model_id']} on {record['gpu_type']}"
+                  f"{_format_identity_filters(_comparison_identity_filters(record))}")
+            print("Regression gating is INACTIVE for this cohort until "
+                  "baseline history accumulates. If this cohort shift is "
+                  "unexpected, check the identity fields above.")
+            print("=" * 72)
             failures: list[str] = []
             record["success"] = True
         else:
+            record["baseline_status"] = "compared"
             failures = _check_regressions(record, baseline_records, metric_policies)
         if static_threshold_failed:
             failures.append(f"{record['model_id']} fixed-threshold phase failed "

fastvideo/tests/performance/dashboard.py

@@ -29,15 +29,76 @@ METRICS = (
     "dit_time_s",
     "vae_decode_time_s",
 )
+COMPARISON_COHORT_KEYS = (
+    "workload_id",
+    "variant_id",
+    "benchmark_version",
+    "recipe_fingerprint",
+    "hardware_profile_id",
+    "software_profile_id",
+)
+GROUP_KEYS = ("model_id", "gpu_type", *COMPARISON_COHORT_KEYS)
+
+
+def _cohort_value(value: object) -> str:
+    if value is None:
+        return ""
+    try:
+        if pd.isna(value):
+            return ""
+    except (TypeError, ValueError):
+        pass
+    return str(value)
+
+
+def _display_value(value: object) -> str:
+    return _cohort_value(value) or "legacy"
+
+
+def _short_value(value: object) -> str:
+    text = _display_value(value)
+    if text == "legacy" or len(text) <= 14:
+        return text
+    return text[:12]
+
+
+def _cohort_title(group_key: tuple[object, ...]) -> str:
+    workload = _display_value(group_key[2])
+    variant = _display_value(group_key[3])
+    version = _display_value(group_key[4])
+    version_label = version if version == "legacy" else f"v{version}"
+    return f"{workload} / {variant} / {version_label}"
+
+
+def _cohort_detail(group_key: tuple[object, ...]) -> str:
+    return " | ".join((
+        f"recipe {_short_value(group_key[5])}",
+        _short_value(group_key[6]),
+        _short_value(group_key[7]),
+    ))
+
+
+def _group_record(group_key: tuple[object, ...]) -> dict[str, str]:
+    return {
+        key: _cohort_value(value)
+        for key, value in zip(GROUP_KEYS, group_key)
+    }
+
+
+def _dashboard_frame(df: pd.DataFrame) -> pd.DataFrame:
+    dashboard_df = df.copy()
+    for key in GROUP_KEYS:
+        if key not in dashboard_df.columns:
+            dashboard_df[key] = ""
+        dashboard_df[key] = dashboard_df[key].map(_cohort_value)
+    return dashboard_df
 
 # -----------------------------
 # 1. Grouping
 # -----------------------------
 def group_data(df: pd.DataFrame):
-    # Group only by model+GPU so each group produces a time-series line.
-    # config_id (commit SHA) is carried as a column for hover/color use.
-    keys = ["model_id", "gpu_type"]
-    return df.groupby(keys, dropna=False)
+    # Group by the same comparison cohort used by baseline gating.
+    return _dashboard_frame(df).groupby(list(GROUP_KEYS), dropna=False)
 
 # -----------------------------
 # 2. Plot builder
@@ -46,15 +107,20 @@ def build_plots(df: pd.DataFrame) -> tuple[list, list[dict[str, object]]]:
     figs = []
     skipped_metrics: list[dict[str, object]] = []
 
-    for (model_id, gpu_type), g in group_data(df):
+    for group_key, g in group_data(df):
+        model_id, gpu_type = group_key[:2]
+        group_record = _group_record(group_key)
+        cohort_title = _cohort_title(group_key)
+        cohort_detail = _cohort_detail(group_key)
         g = g.sort_values("timestamp")
 
         # One chart per metric so the y-axes aren't on wildly different scales
         for metric in METRICS:
             if metric not in g.columns:
                 skipped_metrics.append({
-                    "model_id": model_id,
-                    "gpu_type": gpu_type,
+                    **group_record,
+                    "cohort": cohort_title,
+                    "cohort_detail": cohort_detail,
                     "metric": metric,
                     "reason": "column missing from loaded records",
                     "records": len(g),
@@ -65,8 +131,9 @@ def build_plots(df: pd.DataFrame) -> tuple[list, list[dict[str, object]]]:
             non_null = int(g[metric].notna().sum())
             if non_null == 0:
                 skipped_metrics.append({
-                    "model_id": model_id,
-                    "gpu_type": gpu_type,
+                    **group_record,
+                    "cohort": cohort_title,
+                    "cohort_detail": cohort_detail,
                     "metric": metric,
                     "reason": "no non-null values in loaded records",
                     "records": len(g),
@@ -79,8 +146,8 @@ def build_plots(df: pd.DataFrame) -> tuple[list, list[dict[str, object]]]:
                 x="timestamp",
                 y=metric,
                 markers=True,
-                hover_data=["config_id", "commit_sha"],
-                title=f"{model_id} | {gpu_type} | {metric}",
+                hover_data=["config_id", "commit_sha", *COMPARISON_COHORT_KEYS],
+                title=f"{model_id} | {gpu_type} | {cohort_title} | {cohort_detail} | {metric}",
                 labels={"timestamp": "Time", metric: metric},
             )
             figs.append(fig)
@@ -95,7 +162,7 @@ def render_skipped_metrics(skipped_metrics: list[dict[str, object]]) -> str:
     rows = [
         "<h3>Skipped Metric Plots</h3>",
         "<table>",
-        ("<thead><tr><th>Model</th><th>GPU</th><th>Metric</th>"
+        ("<thead><tr><th>Model</th><th>GPU</th><th>Cohort</th><th>Metric</th>"
          "<th>Records</th><th>Non-null</th><th>Reason</th></tr></thead>"),
         "<tbody>",
     ]
@@ -104,6 +171,7 @@ def render_skipped_metrics(skipped_metrics: list[dict[str, object]]) -> str:
             "<tr>"
             f"<td>{escape(str(item['model_id']))}</td>"
             f"<td>{escape(str(item['gpu_type']))}</td>"
+            f"<td>{escape(str(item['cohort']))}<br><code>{escape(str(item['cohort_detail']))}</code></td>"
             f"<td>{escape(str(item['metric']))}</td>"
             f"<td>{item['records']}</td>"
             f"<td>{item['non_null']}</td>"
@@ -165,6 +233,7 @@ def main() -> None:
         print("Skipped metric plots:")
         for item in skipped_metrics:
             print(f"  - {item['model_id']} | {item['gpu_type']} | "
+                  f"{item['cohort']} | {item['cohort_detail']} | "
                   f"{item['metric']}: {item['reason']} "
                   f"({item['non_null']}/{item['records']} non-null)")
     print(f"Generated {len(figs)} metric plot(s)")

fastvideo/tests/performance/identity.py

@@ -0,0 +1,435 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Comparable identity helpers for performance benchmark records."""
+
+from __future__ import annotations
+
+import dataclasses
+import enum
+import hashlib
+import importlib.metadata
+import json
+import os
+import platform
+import re
+from collections.abc import Mapping, Sequence
+from pathlib import Path
+from typing import Any
+
+import torch
+
+RECIPE_SCHEMA_VERSION = 1
+PROFILE_ID_LENGTH = 16
+_PATH_EXCLUDED_GENERATION_KEYS = {"output_path", "output_video_name"}
+_PROMPT_KEYS = {"prompt", "negative_prompt", "neg_prompt"}
+REQUIRED_BENCHMARK_IDENTITY_KEYS = (
+    "benchmark_id",
+    "workload_id",
+    "variant_id",
+    "benchmark_version",
+)
+_PROFILE_ENV_VARS = (
+    "CUDA_VISIBLE_DEVICES",
+    "FASTVIDEO_ATTENTION_BACKEND",
+    "IMAGE_VERSION",
+    "UV_TORCH_BACKEND",
+)
+_PACKAGE_DISTRIBUTIONS = {
+    "fastvideo_kernel": ("fastvideo-kernel", "fastvideo_kernel"),
+    "flash_attn": ("flash-attn", "flash_attn"),
+    "flash_attn_4": ("flash-attn-4",),
+    "flash_attention_fp4": ("flash-attention-fp4",),
+    "flashinfer": ("flashinfer-python", "flashinfer"),
+    "nvidia_cutlass_dsl": ("nvidia-cutlass-dsl", "cutlass-dsl"),
+    "sageattention": ("sageattention",),
+    "sageattn3": ("sageattn3",),
+    "triton": ("triton",),
+    "xformers": ("xformers",),
+}
+
+
+def canonical_json(payload: Any) -> str:
+    """Return deterministic, compact JSON for a recipe/profile payload."""
+
+    return json.dumps(
+        _canonicalize(payload),
+        sort_keys=True,
+        separators=(",", ":"),
+        ensure_ascii=True,
+    )
+
+
+def sha256_hexdigest(payload: str | bytes) -> str:
+    if isinstance(payload, str):
+        payload = payload.encode("utf-8")
+    return hashlib.sha256(payload).hexdigest()
+
+
+def recipe_fingerprint(recipe: Mapping[str, Any]) -> str:
+    # Runtime-resolved values are audit metadata, not identity: hashing the
+    # resolved HF snapshot revision would churn the cohort (and via the
+    # no-baseline init path, silently reset the rolling baseline) on ANY
+    # upstream repo commit, including model-card edits with unchanged
+    # weights. The fingerprint is pinned to declared inputs only; the
+    # resolved values stay in the stored recipe for auditability.
+    pruned = {
+        key: (dict(value) if isinstance(value, Mapping) else value)
+        for key, value in recipe.items()
+    }
+    model = pruned.get("model")
+    if isinstance(model, dict):
+        model.pop("resolved_revision", None)
+    attention = pruned.get("attention")
+    if isinstance(attention, dict):
+        # Same reasoning as resolved_revision: with requested_backend="auto",
+        # a silent runtime fallback (e.g. a broken flash-attn wheel) must NOT
+        # open a fresh ungated cohort and seed degraded numbers as the new
+        # baseline — it should fail the regression gate in the old cohort.
+        # Intentional backend changes are declared via requested_backend,
+        # which stays in the hash.
+        attention.pop("resolved_backend", None)
+    return sha256_hexdigest(canonical_json(pruned))
+
+
+def profile_id(prefix: str, profile: Mapping[str, Any]) -> str:
+    return f"{prefix}-{sha256_hexdigest(canonical_json(profile))[:PROFILE_ID_LENGTH]}"
+
+
+def hardware_profile_id(profile: Mapping[str, Any]) -> str:
+    return profile_id("hw", profile)
+
+
+def software_profile_id(profile: Mapping[str, Any]) -> str:
+    return profile_id("sw", profile)
+
+
+def environment_fingerprint(metadata: Mapping[str, Any]) -> str:
+    return profile_id("env", metadata)
+
+
+def benchmark_identity_from_config(cfg: Mapping[str, Any]) -> dict[str, Any]:
+    missing = [
+        key for key in REQUIRED_BENCHMARK_IDENTITY_KEYS
+        if _none_if_empty(cfg.get(key)) is None
+    ]
+    if missing:
+        raise ValueError("Benchmark config missing required identity fields: " + ", ".join(missing))
+    return {
+        key: cfg[key]
+        for key in REQUIRED_BENCHMARK_IDENTITY_KEYS
+    }
+
+
+def build_recipe_from_benchmark_config(
+    cfg: Mapping[str, Any],
+    *,
+    attention_backend: str | None = None,
+    resolved_attention_backend: Any | None = None,
+    resolved_model_revision: Any | None = None,
+    measured_prompts: Sequence[Any] | None = None,
+) -> dict[str, Any]:
+    """Build a deterministic recipe document from a benchmark config.
+
+    The recipe captures fields that describe what benchmark workload was run. It
+    intentionally excludes timings, timestamps, commit metadata, and output paths.
+    """
+
+    model = dict(cfg.get("model") or {})
+    init_kwargs = dict(cfg.get("init_kwargs") or {})
+    generation_kwargs = dict(cfg.get("generation_kwargs") or {})
+    benchmark_identity = benchmark_identity_from_config(cfg)
+    prompts = list(measured_prompts) if measured_prompts is not None else list(
+        cfg.get("test_prompts") or ["A cinematic video."])
+    if attention_backend is None:
+        attention_backend = os.environ.get("FASTVIDEO_ATTENTION_BACKEND")
+
+    negative_prompt = generation_kwargs.pop("negative_prompt", generation_kwargs.pop("neg_prompt", None))
+    generation_recipe = {
+        key: value
+        for key, value in generation_kwargs.items()
+        if key not in _PATH_EXCLUDED_GENERATION_KEYS and key not in _PROMPT_KEYS
+    }
+
+    return {
+        "recipe_schema_version": RECIPE_SCHEMA_VERSION,
+        "benchmark": benchmark_identity,
+        "model": {
+            "model_path": normalize_model_path(model.get("model_path")),
+            "revision": _none_if_empty(model.get("revision") or cfg.get("revision")),
+            "resolved_revision": _none_if_empty(resolved_model_revision),
+        },
+        "init_kwargs": init_kwargs,
+        "generation_kwargs": generation_recipe,
+        "inputs": {
+            "prompt_count": len(prompts),
+            "prompt_sha256": [_optional_digest(prompt) for prompt in prompts],
+            "negative_prompt_sha256": _optional_digest(negative_prompt),
+        },
+        "attention": {
+            "requested_backend": _none_if_empty(attention_backend) or "auto",
+            "resolved_backend": _none_if_empty(resolved_attention_backend),
+        },
+    }
+
+
+def normalize_model_path(value: Any) -> str | None:
+    if value is None:
+        return None
+    text = os.fspath(value) if isinstance(value, os.PathLike) else str(value)
+    text = text.strip()
+    if not text:
+        return None
+    if _looks_like_local_path(text):
+        return Path(text).expanduser().as_posix()
+    return re.sub(r"/+", "/", text)
+
+
+def resolved_revision_from_model_path(value: Any) -> str | None:
+    normalized = normalize_model_path(value)
+    if normalized is None:
+        return None
+    parts = Path(normalized).parts
+    for index, part in enumerate(parts[:-1]):
+        if part == "snapshots":
+            revision = parts[index + 1]
+            if re.fullmatch(r"[0-9a-f]{40}", revision):
+                return revision
+    return None
+
+
+def hardware_profile(
+    *,
+    num_gpus: int | None = None,
+    gpu_devices: Sequence[Mapping[str, Any]] | None = None,
+    device_type: str | None = None,
+    interconnect: str | None = None,
+) -> dict[str, Any]:
+    """Return normalized hardware cohort metadata.
+
+    ``gpu_devices`` is an injection point for tests and non-CUDA callers.
+    """
+
+    if gpu_devices is not None:
+        devices = [_normalize_gpu_device(device) for device in gpu_devices]
+        return {
+            "device_type": device_type or "cuda",
+            "gpu_count": num_gpus if num_gpus is not None else len(devices),
+            "gpus": devices,
+            "interconnect": interconnect or "unknown",
+        }
+
+    if not torch.cuda.is_available():
+        return {
+            "device_type": device_type or "cpu",
+            "gpu_count": 0,
+            "gpus": [],
+            "interconnect": interconnect or "none",
+        }
+
+    gpu_count = int(num_gpus if num_gpus is not None else torch.cuda.device_count())
+    devices = []
+    cuda_device_count = torch.cuda.device_count()
+    for device_id in range(gpu_count):
+        if device_id >= cuda_device_count:
+            devices.append(_normalize_gpu_device({}))
+            continue
+        props = torch.cuda.get_device_properties(device_id)
+        capability = torch.cuda.get_device_capability(device_id)
+        devices.append(
+            _normalize_gpu_device({
+                "name": props.name,
+                "memory_bytes": props.total_memory,
+                "compute_capability": f"{capability[0]}.{capability[1]}",
+            }))
+
+    return {
+        "device_type": device_type or "cuda",
+        "gpu_count": gpu_count,
+        "gpus": devices,
+        "interconnect": interconnect or _detect_cuda_interconnect(gpu_count),
+    }
+
+
+def software_profile(
+    *,
+    python_version: str | None = None,
+    torch_version: str | None = None,
+    cuda_version: str | None = None,
+    package_versions: Mapping[str, str | None] | None = None,
+) -> dict[str, Any]:
+    """Return normalized software cohort metadata.
+
+    Python, PyTorch, and CUDA use major/minor cohorts. Attention and kernel
+    packages keep exact versions because patch releases can change performance.
+    """
+
+    versions = dict(package_versions) if package_versions is not None else _installed_package_versions()
+    return {
+        "python": _major_minor(python_version or platform.python_version()),
+        "pytorch": _major_minor(torch_version or torch.__version__),
+        "cuda": _major_minor(cuda_version or torch.version.cuda),
+        "packages": {
+            name: str(version)
+            for name, version in sorted(versions.items())
+            if version
+        },
+    }
+
+
+def environment_metadata(
+    *,
+    env: Mapping[str, str] | None = None,
+    package_versions: Mapping[str, str | None] | None = None,
+    hardware: Mapping[str, Any] | None = None,
+    software: Mapping[str, Any] | None = None,
+) -> dict[str, Any]:
+    """Return audit metadata kept separate from comparable identity keys."""
+
+    source_env = env if env is not None else os.environ
+    full_package_versions = dict(package_versions) if package_versions is not None else _installed_package_versions()
+    return {
+        "python": {
+            "version": platform.python_version(),
+            "implementation": platform.python_implementation(),
+        },
+        "platform": {
+            "system": platform.system(),
+            "machine": platform.machine(),
+            "release": platform.release(),
+        },
+        "torch": {
+            "version": torch.__version__,
+            "cuda": torch.version.cuda,
+            "cudnn": torch.backends.cudnn.version(),
+        },
+        "env": {
+            key: source_env.get(key)
+            for key in _PROFILE_ENV_VARS
+            if source_env.get(key) is not None
+        },
+        "packages": {
+            key: value
+            for key, value in sorted(full_package_versions.items())
+            if value
+        },
+        "hardware_profile": dict(hardware) if hardware is not None else hardware_profile(),
+        "software_profile": dict(software) if software is not None else software_profile(
+            package_versions=full_package_versions),
+    }
+
+
+def _canonicalize(value: Any) -> Any:
+    if dataclasses.is_dataclass(value) and not isinstance(value, type):
+        return _canonicalize(dataclasses.asdict(value))
+    if isinstance(value, Mapping):
+        return {
+            str(key): _canonicalize(value[key])
+            for key in sorted(value, key=lambda item: str(item))
+        }
+    if isinstance(value, (set, frozenset)):
+        return [_canonicalize(item) for item in sorted(value, key=lambda item: str(item))]
+    if isinstance(value, tuple):
+        return [_canonicalize(item) for item in value]
+    if isinstance(value, list):
+        return [_canonicalize(item) for item in value]
+    if isinstance(value, enum.Enum):
+        return value.name
+    if isinstance(value, Path):
+        return value.expanduser().as_posix()
+    if isinstance(value, os.PathLike):
+        return os.fspath(value)
+    if isinstance(value, type):
+        return f"{value.__module__}.{value.__qualname__}"
+    if isinstance(value, torch.dtype):
+        return str(value).removeprefix("torch.")
+    return value
+
+
+def _none_if_empty(value: Any) -> Any | None:
+    if value is None:
+        return None
+    if isinstance(value, str) and not value.strip():
+        return None
+    return value
+
+
+def _optional_digest(value: Any) -> str | None:
+    if value is None:
+        return None
+    return sha256_hexdigest(str(value))
+
+
+def _looks_like_local_path(value: str) -> bool:
+    if value.startswith(("/", "./", "../", "~")):
+        return True
+    looks_like_hf_repo = re.match(r"^[A-Za-z0-9_.-]+/[A-Za-z0-9_.-]+$", value)
+    has_separator = "/" in value or "\\" in value or os.sep in value
+    return has_separator and looks_like_hf_repo is None
+
+
+def _normalize_gpu_device(device: Mapping[str, Any]) -> dict[str, Any]:
+    memory_gb = device.get("memory_gb")
+    if memory_gb is None and device.get("memory_bytes") is not None:
+        memory_gb = round(float(device["memory_bytes"]) / (1024**3))
+    return {
+        "name": _none_if_empty(device.get("name")) or "unknown",
+        "memory_gb": memory_gb,
+        "compute_capability": _none_if_empty(device.get("compute_capability")),
+    }
+
+
+def _detect_cuda_interconnect(gpu_count: int) -> str:
+    if gpu_count <= 1:
+        return "single_gpu"
+    try:
+        from fastvideo.platforms import current_platform
+
+        if current_platform.is_cuda() and current_platform.is_full_nvlink(list(range(gpu_count))):
+            return "full_nvlink"
+        return "none_or_partial"
+    except Exception:
+        return "unknown"
+
+
+def _installed_package_versions() -> dict[str, str | None]:
+    return {
+        name: _distribution_version(*distribution_names)
+        for name, distribution_names in _PACKAGE_DISTRIBUTIONS.items()
+    }
+
+
+def _distribution_version(*names: str) -> str | None:
+    for name in names:
+        try:
+            return importlib.metadata.version(name)
+        except importlib.metadata.PackageNotFoundError:
+            continue
+    return None
+
+
+def _major_minor(version: str | None) -> str | None:
+    if not version:
+        return None
+    match = re.search(r"(\d+)\.(\d+)", version)
+    if match is None:
+        return version
+    return f"{match.group(1)}.{match.group(2)}"
+
+
+__all__ = [
+    "RECIPE_SCHEMA_VERSION",
+    "REQUIRED_BENCHMARK_IDENTITY_KEYS",
+    "benchmark_identity_from_config",
+    "canonical_json",
+    "environment_fingerprint",
+    "environment_metadata",
+    "hardware_profile",
+    "hardware_profile_id",
+    "normalize_model_path",
+    "profile_id",
+    "recipe_fingerprint",
+    "resolved_revision_from_model_path",
+    "build_recipe_from_benchmark_config",
+    "sha256_hexdigest",
+    "software_profile",
+    "software_profile_id",
+]

fastvideo/tests/performance/test_compare_baseline_policy.py

@@ -1,5 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import pytest
+
 from fastvideo.tests.performance import compare_baseline
 from fastvideo.performance.metric_policy import resolve_metric_policies
 
@@ -119,6 +121,110 @@ def test_boolean_regression_threshold_values_are_ignored():
     assert latency.threshold_percent == 0.08
     assert latency.threshold_absolute == 0.5
     assert latency.gated is False
+def test_normalized_record_preserves_identity_metadata(monkeypatch):
+    monkeypatch.setenv("PERF_RUN_SOURCE", "pr")
+    raw = _raw_result()
+    raw.update({
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+        "recipe": {
+            "recipe_schema_version": 1,
+        },
+        "recipe_fingerprint": "recipe-1",
+        "hardware_profile": {
+            "gpu_count": 1,
+        },
+        "hardware_profile_id": "hw-1",
+        "software_profile": {
+            "python": "3.12",
+        },
+        "software_profile_id": "sw-1",
+        "environment_metadata": {
+            "env": {
+                "IMAGE_VERSION": "latest",
+            },
+        },
+        "environment_fingerprint": "env-1",
+    })
+
+    record = compare_baseline.normalize_performance_result(raw)
+
+    assert record["workload_id"] == "wan-t2v"
+    assert record["variant_id"] == "1.3b-sp2"
+    assert record["benchmark_version"] == 2
+    assert record["recipe"] == {"recipe_schema_version": 1}
+    assert record["recipe_fingerprint"] == "recipe-1"
+    assert record["hardware_profile"] == {"gpu_count": 1}
+    assert record["hardware_profile_id"] == "hw-1"
+    assert record["software_profile"] == {"python": "3.12"}
+    assert record["software_profile_id"] == "sw-1"
+    assert record["environment_metadata"] == {"env": {"IMAGE_VERSION": "latest"}}
+    assert record["environment_fingerprint"] == "env-1"
+
+
+def test_normalized_record_reads_identity_labels_from_recipe(monkeypatch):
+    monkeypatch.setenv("PERF_RUN_SOURCE", "pr")
+    raw = _raw_result()
+    raw["recipe"] = {
+        "benchmark": {
+            "workload_id": "wan-t2v",
+            "variant_id": "1.3b-sp2",
+            "benchmark_version": 2,
+        },
+    }
+
+    record = compare_baseline.normalize_performance_result(raw)
+
+    assert record["workload_id"] == "wan-t2v"
+    assert record["variant_id"] == "1.3b-sp2"
+    assert record["benchmark_version"] == 2
+
+
+def test_comparison_identity_filters_use_full_issue_key():
+    record = {
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+        "recipe_fingerprint": "recipe-1",
+        "hardware_profile_id": "hw-1",
+        "software_profile_id": "sw-1",
+    }
+
+    assert compare_baseline._comparison_identity_filters(record) == {
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": "2",
+        "recipe_fingerprint": "recipe-1",
+        "hardware_profile_id": "hw-1",
+        "software_profile_id": "sw-1",
+    }
+
+
+def test_comparison_identity_filters_require_full_issue_key():
+    record = {
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 0,
+        "recipe_fingerprint": "recipe-1",
+        "hardware_profile_id": "hw-1",
+    }
+
+    with pytest.raises(ValueError, match="software_profile_id"):
+        compare_baseline._comparison_identity_filters(record)
+
+
+def test_comparison_identity_filters_keep_zero_version():
+    record = {
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 0,
+        "recipe_fingerprint": "recipe-1",
+        "hardware_profile_id": "hw-1",
+        "software_profile_id": "sw-1",
+    }
+
+    assert compare_baseline._comparison_identity_filters(record)["benchmark_version"] == "0"
 
 
 def test_baseline_eligibility_only_for_successful_scheduled_main():

fastvideo/tests/performance/test_dashboard_api.py

@@ -56,8 +56,34 @@ def _record(model_id, gpu_type, ts, commit, latency, throughput, success=True, *
 
 def test_summary_endpoint_returns_latest_group_status():
     app = create_app(FakeStore([
-        _record("wan", "NVIDIA L40S", "2026-01-01T00:00:00+00:00", "a" * 40, 10.0, 10.0),
-        _record("wan", "NVIDIA L40S", "2026-01-02T00:00:00+00:00", "b" * 40, 11.0, 9.0),
+        _record(
+            "wan",
+            "NVIDIA L40S",
+            "2026-01-01T00:00:00+00:00",
+            "a" * 40,
+            10.0,
+            10.0,
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            recipe_fingerprint="recipe-a",
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+        ),
+        _record(
+            "wan",
+            "NVIDIA L40S",
+            "2026-01-02T00:00:00+00:00",
+            "b" * 40,
+            11.0,
+            9.0,
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            recipe_fingerprint="recipe-a",
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+        ),
     ]))
     client = TestClient(app)
 
@@ -71,6 +97,10 @@ def test_summary_endpoint_returns_latest_group_status():
     assert body["rows"][0]["metrics"]["latency"]["threshold_exceeded"] is True
     assert body["rows"][0]["threshold_exceeded_metrics"] == ["latency", "throughput"]
     assert body["rows"][0]["computed_regression_status"] == "fail"
+    assert body["rows"][0]["workload_id"] == "wan-t2v"
+    assert body["rows"][0]["variant_id"] == "1.3b-sp2"
+    assert body["rows"][0]["benchmark_version"] == 2
+    assert body["rows"][0]["recipe_fingerprint"] == "recipe-a"
 
 
 def test_summary_status_is_independent_of_days_window():

fastvideo/tests/performance/test_dashboard_plotly.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import pandas as pd
+
+from fastvideo.tests.performance import dashboard
+
+
+def _record(recipe_fingerprint="recipe-a", software_profile_id="sw-a", **overrides):
+    record = {
+        "model_id": "wan-t2v-1.3b-2gpu",
+        "gpu_type": "NVIDIA L40S",
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+        "recipe_fingerprint": recipe_fingerprint,
+        "hardware_profile_id": "hw-l40s",
+        "software_profile_id": software_profile_id,
+        "timestamp": "2026-01-01T00:00:00+00:00",
+        "commit_sha": "a" * 40,
+        "config_id": "aaaaaaa",
+        "latency": 10.0,
+        "throughput": 4.5,
+        "memory": 10000.0,
+        "text_encoder_time_s": 2.0,
+        "dit_time_s": 8.0,
+        "vae_decode_time_s": 3.0,
+    }
+    record.update(overrides)
+    return record
+
+
+def test_group_data_uses_full_comparison_cohort():
+    df = pd.DataFrame([
+        _record(recipe_fingerprint="recipe-a", software_profile_id="sw-a"),
+        _record(recipe_fingerprint="recipe-b", software_profile_id="sw-b"),
+    ])
+
+    groups = list(dashboard.group_data(df))
+
+    assert len(groups) == 2
+    assert {key[5] for key, _group in groups} == {"recipe-a", "recipe-b"}
+    assert {key[7] for key, _group in groups} == {"sw-a", "sw-b"}
+
+
+def test_group_data_fills_legacy_cohort_columns():
+    df = pd.DataFrame([{
+        "model_id": "legacy-wan",
+        "gpu_type": "NVIDIA L40S",
+        "timestamp": "2026-01-01T00:00:00+00:00",
+        "commit_sha": "a" * 40,
+        "config_id": "aaaaaaa",
+        "latency": 10.0,
+    }])
+
+    groups = list(dashboard.group_data(df))
+
+    assert len(groups) == 1
+    assert groups[0][0][2:] == ("", "", "", "", "", "")
+
+
+def test_build_plots_labels_distinct_cohorts():
+    df = pd.DataFrame([
+        _record(recipe_fingerprint="recipe-a", software_profile_id="sw-a"),
+        _record(recipe_fingerprint="recipe-b", software_profile_id="sw-b"),
+    ])
+
+    figs, skipped_metrics = dashboard.build_plots(df)
+    titles = [fig.layout.title.text for fig in figs]
+
+    assert len(figs) == len(dashboard.METRICS) * 2
+    assert skipped_metrics == []
+    assert any("recipe recipe-a | hw-l40s | sw-a" in title for title in titles)
+    assert any("recipe recipe-b | hw-l40s | sw-b" in title for title in titles)
+
+
+def test_skipped_metric_table_includes_cohort_identity():
+    df = pd.DataFrame([{
+        "model_id": "wan-t2v-1.3b-2gpu",
+        "gpu_type": "NVIDIA L40S",
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+        "recipe_fingerprint": "recipe-a",
+        "hardware_profile_id": "hw-l40s",
+        "software_profile_id": "sw-a",
+        "timestamp": "2026-01-01T00:00:00+00:00",
+        "commit_sha": "a" * 40,
+        "config_id": "aaaaaaa",
+        "latency": 10.0,
+    }])
+
+    _figs, skipped_metrics = dashboard.build_plots(df)
+    html = dashboard.render_skipped_metrics(skipped_metrics[:1])
+
+    assert skipped_metrics[0]["cohort"] == "wan-t2v / 1.3b-sp2 / v2"
+    assert skipped_metrics[0]["cohort_detail"] == "recipe recipe-a | hw-l40s | sw-a"
+    assert "wan-t2v / 1.3b-sp2 / v2" in html
+    assert "recipe recipe-a | hw-l40s | sw-a" in html

fastvideo/tests/performance/test_dashboard_service.py

@@ -142,6 +142,111 @@ def test_build_latest_summary_separates_informational_threshold_crossing():
     assert rows[0]["threshold_exceeded_metrics"] == ["latency"]
     assert rows[0]["failing_metrics"] == []
     assert rows[0]["computed_regression_status"] == "pass"
+def test_build_latest_summary_keeps_identity_cohorts_separate():
+    records = [
+        _record(
+            "2026-01-01T00:00:00+00:00",
+            "a" * 40,
+            10.0,
+            10.0,
+            recipe_fingerprint="recipe-a",
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+            run_source="scheduled_main",
+            baseline_eligible=True,
+        ),
+        _record(
+            "2026-01-02T00:00:00+00:00",
+            "b" * 40,
+            20.0,
+            5.0,
+            recipe_fingerprint="recipe-b",
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+            run_source="scheduled_main",
+            baseline_eligible=True,
+        ),
+        _record(
+            "2026-01-03T00:00:00+00:00",
+            "c" * 40,
+            22.0,
+            4.5,
+            recipe_fingerprint="recipe-b",
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+        ),
+    ]
+
+    rows = build_latest_summary(records)
+    recipe_b_row = next(row for row in rows if row["recipe_fingerprint"] == "recipe-b")
+
+    assert len(rows) == 2
+    assert recipe_b_row["baseline_n"] == 1
+    assert recipe_b_row["metrics"]["latency"]["baseline"] == 20.0
+    assert recipe_b_row["workload_id"] == "wan-t2v"
+    assert recipe_b_row["variant_id"] == "1.3b-sp2"
+    assert recipe_b_row["benchmark_version"] == 2
+
+
+def test_build_latest_summary_keeps_variant_versions_separate():
+    records = [
+        _record(
+            "2026-01-01T00:00:00+00:00",
+            "a" * 40,
+            10.0,
+            10.0,
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=1,
+            recipe_fingerprint="recipe-a",
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+            run_source="scheduled_main",
+            baseline_eligible=True,
+        ),
+        _record(
+            "2026-01-02T00:00:00+00:00",
+            "b" * 40,
+            20.0,
+            5.0,
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            recipe_fingerprint="recipe-a",
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+            run_source="scheduled_main",
+            baseline_eligible=True,
+        ),
+        _record(
+            "2026-01-03T00:00:00+00:00",
+            "c" * 40,
+            22.0,
+            4.5,
+            workload_id="wan-t2v",
+            variant_id="1.3b-sp2",
+            benchmark_version=2,
+            recipe_fingerprint="recipe-a",
+            hardware_profile_id="hw-l40s",
+            software_profile_id="sw-cu130",
+        ),
+    ]
+
+    rows = build_latest_summary(records)
+    version_2_row = next(row for row in rows if row["benchmark_version"] == 2)
+
+    assert len(rows) == 2
+    assert version_2_row["baseline_n"] == 1
+    assert version_2_row["metrics"]["latency"]["baseline"] == 20.0
 
 
 def test_filter_records_and_trends_preserve_metric_points():
@@ -218,3 +323,78 @@ def test_load_records_can_filter_baseline_eligible_records(tmp_path):
         "2026-01-02T00:00:00+00:00",
         "2026-01-03T00:00:00+00:00",
     }
+
+
+def test_load_records_for_model_filters_identity_cohort(tmp_path):
+    model_dir = tmp_path / "wan"
+    model_dir.mkdir()
+    (model_dir / "matching.json").write_text(
+        """
+        {
+          "model_id": "wan",
+          "gpu_type": "NVIDIA L40S",
+          "timestamp": "2026-01-01T00:00:00+00:00",
+          "success": true,
+          "baseline_eligible": true,
+          "workload_id": "wan-t2v",
+          "variant_id": "1.3b-sp2",
+          "benchmark_version": 2,
+          "recipe_fingerprint": "recipe-a",
+          "hardware_profile_id": "hw-l40s",
+          "software_profile_id": "sw-cu130"
+        }
+        """,
+        encoding="utf-8",
+    )
+    (model_dir / "other_recipe.json").write_text(
+        """
+        {
+          "model_id": "wan",
+          "gpu_type": "NVIDIA L40S",
+          "timestamp": "2026-01-02T00:00:00+00:00",
+          "success": true,
+          "baseline_eligible": true,
+          "workload_id": "wan-t2v",
+          "variant_id": "1.3b-sp2",
+          "benchmark_version": 2,
+          "recipe_fingerprint": "recipe-b",
+          "hardware_profile_id": "hw-l40s",
+          "software_profile_id": "sw-cu130"
+        }
+        """,
+        encoding="utf-8",
+    )
+    (model_dir / "other_version.json").write_text(
+        """
+        {
+          "model_id": "wan",
+          "gpu_type": "NVIDIA L40S",
+          "timestamp": "2026-01-03T00:00:00+00:00",
+          "success": true,
+          "baseline_eligible": true,
+          "workload_id": "wan-t2v",
+          "variant_id": "1.3b-sp2",
+          "benchmark_version": 3,
+          "recipe_fingerprint": "recipe-a",
+          "hardware_profile_id": "hw-l40s",
+          "software_profile_id": "sw-cu130"
+        }
+        """,
+        encoding="utf-8",
+    )
+
+    records = hf_store.load_records_for_model(
+        str(tmp_path),
+        "wan",
+        "NVIDIA L40S",
+        workload_id="wan-t2v",
+        variant_id="1.3b-sp2",
+        benchmark_version="2",
+        recipe_fingerprint="recipe-a",
+        hardware_profile_id="hw-l40s",
+        software_profile_id="sw-cu130",
+        baseline_eligible_only=True,
+    )
+
+    assert len(records) == 1
+    assert records[0]["recipe_fingerprint"] == "recipe-a"

fastvideo/tests/performance/test_identity.py

@@ -0,0 +1,422 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from copy import deepcopy
+
+import pytest
+import torch
+
+from fastvideo.tests.performance import identity as identity_module
+from fastvideo.tests.performance.identity import (
+    build_recipe_from_benchmark_config,
+    canonical_json,
+    environment_fingerprint,
+    environment_metadata,
+    hardware_profile,
+    hardware_profile_id,
+    recipe_fingerprint,
+    resolved_revision_from_model_path,
+    software_profile,
+    software_profile_id,
+)
+
+
+def _benchmark_config():
+    return {
+        "benchmark_id": "wan-t2v-1.3b-2gpu",
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+        "model": {
+            "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
+            "model_short_name": "Wan2.1-T2V-1.3B",
+        },
+        "init_kwargs": {
+            "num_gpus": 2,
+            "sp_size": 2,
+            "tp_size": 1,
+            "vae_sp": True,
+            "vae_tiling": True,
+            "text_encoder_precisions": ["fp32"],
+        },
+        "generation_kwargs": {
+            "height": 480,
+            "width": 832,
+            "num_frames": 45,
+            "num_inference_steps": 4,
+            "guidance_scale": 3,
+            "embedded_cfg_scale": 6,
+            "seed": 1024,
+            "fps": 24,
+            "neg_prompt": "low quality",
+        },
+        "test_prompts": ["A cinematic video."],
+    }
+
+
+def _fingerprint(cfg, attention_backend="FLASH_ATTN"):
+    recipe = build_recipe_from_benchmark_config(cfg, attention_backend=attention_backend)
+    return recipe_fingerprint(recipe)
+
+
+def test_canonical_json_is_stable_for_mapping_order():
+    left = {"b": [2, 1], "a": {"z": "last", "m": "middle"}}
+    right = {"a": {"m": "middle", "z": "last"}, "b": [2, 1]}
+
+    assert canonical_json(left) == canonical_json(right)
+
+
+def test_canonical_json_handles_sets_deterministically():
+    assert canonical_json({"values": {"b", "a"}}) == '{"values":["a","b"]}'
+    assert canonical_json({"values": frozenset(("b", "a"))}) == '{"values":["a","b"]}'
+
+
+def test_same_config_produces_same_recipe_fingerprint():
+    cfg = _benchmark_config()
+
+    assert _fingerprint(cfg) == _fingerprint(deepcopy(cfg))
+
+
+def test_recipe_includes_first_class_benchmark_identity():
+    recipe = build_recipe_from_benchmark_config(_benchmark_config())
+
+    assert recipe["benchmark"] == {
+        "benchmark_id": "wan-t2v-1.3b-2gpu",
+        "workload_id": "wan-t2v",
+        "variant_id": "1.3b-sp2",
+        "benchmark_version": 2,
+    }
+
+
+def test_recipe_requires_first_class_benchmark_identity():
+    cfg = _benchmark_config()
+    del cfg["workload_id"]
+
+    with pytest.raises(ValueError, match="workload_id"):
+        build_recipe_from_benchmark_config(cfg)
+
+
+def test_semantically_equivalent_sequence_forms_match():
+    cfg = _benchmark_config()
+    equivalent = deepcopy(cfg)
+    equivalent["init_kwargs"]["text_encoder_precisions"] = tuple(
+        equivalent["init_kwargs"]["text_encoder_precisions"])
+
+    assert _fingerprint(cfg) == _fingerprint(equivalent)
+
+
+def test_output_path_does_not_change_recipe_fingerprint():
+    cfg = _benchmark_config()
+    with_output_path = deepcopy(cfg)
+    with_output_path["generation_kwargs"]["output_path"] = "/tmp/generated"
+
+    assert _fingerprint(cfg) == _fingerprint(with_output_path)
+
+
+def test_runtime_resolved_values_do_not_change_recipe_fingerprint():
+    # Runtime-resolved values are audit metadata: an upstream repo commit
+    # (even a README-only edit) or a silent attention-backend fallback must
+    # not churn the cohort — that would open a fresh ungated cohort and let
+    # degraded numbers seed a new baseline. Declared inputs (including
+    # requested_backend) stay in the hash.
+    cfg = _benchmark_config()
+    recipe = build_recipe_from_benchmark_config(
+        cfg, attention_backend="FLASH_ATTN",
+        resolved_attention_backend="FLASH_ATTN",
+        resolved_model_revision="aaaa1111")
+    changed = build_recipe_from_benchmark_config(
+        cfg, attention_backend="FLASH_ATTN",
+        resolved_attention_backend="TORCH_SDPA",
+        resolved_model_revision="bbbb2222")
+
+    assert recipe["model"]["resolved_revision"] == "aaaa1111"  # kept for audit
+    assert recipe["attention"]["resolved_backend"] == "FLASH_ATTN"  # kept for audit
+    assert recipe_fingerprint(recipe) == recipe_fingerprint(changed)
+
+
+def test_measured_prompt_override_ignores_unused_configured_prompts():
+    cfg = _benchmark_config()
+    changed = deepcopy(cfg)
+    changed["test_prompts"] = [
+        "Unused prompt that should not affect this measured workload.",
+        cfg["test_prompts"][0],
+    ]
+
+    recipe = build_recipe_from_benchmark_config(cfg, measured_prompts=[cfg["test_prompts"][0]])
+    changed_recipe = build_recipe_from_benchmark_config(changed, measured_prompts=[cfg["test_prompts"][0]])
+
+    assert recipe_fingerprint(recipe) == recipe_fingerprint(changed_recipe)
+
+
+def test_num_inference_steps_changes_recipe_fingerprint():
+    cfg = _benchmark_config()
+    changed = deepcopy(cfg)
+    changed["generation_kwargs"]["num_inference_steps"] = 8
+
+    assert _fingerprint(cfg) != _fingerprint(changed)
+
+
+def test_benchmark_version_changes_recipe_fingerprint():
+    cfg = _benchmark_config()
+    changed = deepcopy(cfg)
+    changed["benchmark_version"] = 3
+
+    assert _fingerprint(cfg) != _fingerprint(changed)
+
+
+def test_attention_backend_changes_recipe_fingerprint():
+    cfg = _benchmark_config()
+
+    assert _fingerprint(cfg, attention_backend="FLASH_ATTN") != _fingerprint(
+        cfg, attention_backend="TORCH_SDPA")
+
+
+def test_precision_changes_recipe_fingerprint():
+    cfg = _benchmark_config()
+    changed = deepcopy(cfg)
+    changed["init_kwargs"]["text_encoder_precisions"] = ["bf16"]
+
+    assert _fingerprint(cfg) != _fingerprint(changed)
+
+
+def test_dimensions_and_seed_change_recipe_fingerprint():
+    cfg = _benchmark_config()
+    different_dimensions = deepcopy(cfg)
+    different_dimensions["generation_kwargs"]["width"] = 1024
+    different_seed = deepcopy(cfg)
+    different_seed["generation_kwargs"]["seed"] = 2048
+
+    assert _fingerprint(cfg) != _fingerprint(different_dimensions)
+    assert _fingerprint(cfg) != _fingerprint(different_seed)
+
+
+def test_distributed_layout_changes_recipe_fingerprint():
+    cfg = _benchmark_config()
+    changed = deepcopy(cfg)
+    changed["init_kwargs"]["sp_size"] = 1
+    changed["init_kwargs"]["tp_size"] = 2
+
+    assert _fingerprint(cfg) != _fingerprint(changed)
+
+
+def test_software_profile_uses_exact_attention_kernel_package_versions_for_id():
+    profile_a = software_profile(
+        python_version="3.12.4",
+        torch_version="2.12.0+cu130",
+        cuda_version="13.0",
+        package_versions={
+            "triton": "3.4.1",
+            "fastvideo_kernel": "0.3.2",
+            "flash_attn_4": "4.0.0.dev0",
+            "flashinfer": "0.2.11",
+            "nvidia_cutlass_dsl": "4.5.0",
+        },
+    )
+    profile_b = software_profile(
+        python_version="3.12.5",
+        torch_version="2.12.1+cu130",
+        cuda_version="13.0",
+        package_versions={
+            "triton": "3.4.9",
+            "fastvideo_kernel": "0.3.7",
+            "flash_attn_4": "4.0.0.dev1",
+            "flashinfer": "0.2.12",
+            "nvidia_cutlass_dsl": "4.5.1",
+        },
+    )
+
+    assert profile_a == {
+        "python": "3.12",
+        "pytorch": "2.12",
+        "cuda": "13.0",
+        "packages": {
+            "fastvideo_kernel": "0.3.2",
+            "flash_attn_4": "4.0.0.dev0",
+            "flashinfer": "0.2.11",
+            "nvidia_cutlass_dsl": "4.5.0",
+            "triton": "3.4.1",
+        },
+    }
+    assert profile_b["python"] == "3.12"
+    assert profile_b["pytorch"] == "2.12"
+    assert profile_b["cuda"] == "13.0"
+    assert software_profile_id(profile_a) != software_profile_id(profile_b)
+
+
+def test_installed_package_versions_checks_attention_kernel_distributions(monkeypatch):
+    versions = {
+        "fastvideo-kernel": "0.3.2",
+        "flash-attn": "2.8.1",
+        "flash-attn-4": "4.0.0.dev0",
+        "flash-attention-fp4": "0.1.0",
+        "flashinfer-python": "0.2.11",
+        "nvidia-cutlass-dsl": "4.5.0",
+    }
+
+    def fake_version(name):
+        if name not in versions:
+            raise identity_module.importlib.metadata.PackageNotFoundError(name)
+        return versions[name]
+
+    monkeypatch.setattr(identity_module.importlib.metadata, "version", fake_version)
+
+    installed = identity_module._installed_package_versions()
+
+    assert installed["fastvideo_kernel"] == "0.3.2"
+    assert installed["flash_attn"] == "2.8.1"
+    assert installed["flash_attn_4"] == "4.0.0.dev0"
+    assert installed["flash_attention_fp4"] == "0.1.0"
+    assert installed["flashinfer"] == "0.2.11"
+    assert installed["nvidia_cutlass_dsl"] == "4.5.0"
+
+
+def test_hardware_profile_id_uses_normalized_gpu_cohort():
+    profile = hardware_profile(
+        num_gpus=2,
+        gpu_devices=[
+            {
+                "name": "NVIDIA L40S",
+                "memory_bytes": 48 * 1024**3,
+                "compute_capability": "8.9",
+            },
+            {
+                "name": "NVIDIA L40S",
+                "memory_gb": 48,
+                "compute_capability": "8.9",
+            },
+        ],
+        interconnect="none_or_partial",
+    )
+
+    assert profile == {
+        "device_type": "cuda",
+        "gpu_count": 2,
+        "gpus": [
+            {
+                "name": "NVIDIA L40S",
+                "memory_gb": 48,
+                "compute_capability": "8.9",
+            },
+            {
+                "name": "NVIDIA L40S",
+                "memory_gb": 48,
+                "compute_capability": "8.9",
+            },
+        ],
+        "interconnect": "none_or_partial",
+    }
+    assert hardware_profile_id(profile).startswith("hw-")
+
+
+def test_hardware_profile_pads_missing_requested_cuda_devices(monkeypatch):
+    class Props:
+        name = "NVIDIA L40S"
+        total_memory = 48 * 1024**3
+
+    monkeypatch.setattr(torch.cuda, "is_available", lambda: True)
+    monkeypatch.setattr(torch.cuda, "device_count", lambda: 1)
+    monkeypatch.setattr(torch.cuda, "get_device_properties", lambda device_id: Props())
+    monkeypatch.setattr(torch.cuda, "get_device_capability", lambda device_id: (8, 9))
+
+    profile = hardware_profile(num_gpus=2, interconnect="unknown")
+
+    assert profile["gpu_count"] == 2
+    assert profile["gpus"] == [
+        {
+            "name": "NVIDIA L40S",
+            "memory_gb": 48,
+            "compute_capability": "8.9",
+        },
+        {
+            "name": "unknown",
+            "memory_gb": None,
+            "compute_capability": None,
+        },
+    ]
+
+
+def test_local_path_detection_checks_both_path_separators(monkeypatch):
+    monkeypatch.setattr(identity_module.os, "sep", "\\")
+
+    assert identity_module._looks_like_local_path("models/local/checkpoint") is True
+    assert identity_module._looks_like_local_path(r"C:\models\checkpoint") is True
+    assert identity_module._looks_like_local_path("Wan-AI/Wan2.1-T2V-1.3B-Diffusers") is False
+
+
+def test_environment_metadata_is_separate_audit_fingerprint():
+    cfg = _benchmark_config()
+    recipe_hash = _fingerprint(cfg)
+    audit_a = environment_metadata(
+        env={"IMAGE_VERSION": "py3.12-cuda13.0.0"},
+        package_versions={"triton": "3.4.1"},
+        hardware={"device_type": "cuda", "gpu_count": 2},
+        software={"python": "3.12", "pytorch": "2.12", "cuda": "13.0"},
+    )
+    audit_b = environment_metadata(
+        env={"IMAGE_VERSION": "py3.12-cuda13.0.1"},
+        package_versions={"triton": "3.4.1"},
+        hardware={"device_type": "cuda", "gpu_count": 2},
+        software={"python": "3.12", "pytorch": "2.12", "cuda": "13.0"},
+    )
+
+    assert recipe_hash == _fingerprint(cfg)
+    assert environment_fingerprint(audit_a) != environment_fingerprint(audit_b)
+
+
+def test_resolved_revision_from_hf_snapshot_path():
+    revision = "a" * 40
+
+    assert resolved_revision_from_model_path(
+        f"/root/.cache/huggingface/hub/models--org--repo/snapshots/{revision}") == revision
+    assert resolved_revision_from_model_path("/models/local-repo") is None
+
+
+def test_runtime_identity_from_generator_summarizes_worker_records():
+    from fastvideo.tests.performance.test_inference_performance import _runtime_identity_from_generator
+
+    revision = "b" * 40
+
+    class FakeExecutor:
+        def collective_rpc(self, _method):
+            return [
+                {
+                    "resolved_attention_backends": ["FLASH_ATTN"],
+                    "resolved_model_path": f"/root/.cache/huggingface/hub/models--org--repo/snapshots/{revision}",
+                },
+                {
+                    "resolved_attention_backends": ["FLASH_ATTN"],
+                    "resolved_model_path": f"/root/.cache/huggingface/hub/models--org--repo/snapshots/{revision}",
+                },
+            ]
+
+    class FakeGenerator:
+        executor = FakeExecutor()
+
+    assert _runtime_identity_from_generator(FakeGenerator()) == {
+        "resolved_attention_backend": "FLASH_ATTN",
+        "resolved_model_revision": revision,
+    }
+
+
+def test_collect_worker_identity_handles_torch_module_pipeline():
+    from fastvideo.tests.performance.test_inference_performance import _collect_worker_identity
+
+    class Backend:
+        name = "FLASH_ATTN"
+
+    class Leaf(torch.nn.Module):
+        backend = Backend()
+
+    class Pipeline(torch.nn.Module):
+        model_path = "/models/local"
+
+        def __init__(self):
+            super().__init__()
+            self.leaf = Leaf()
+
+    class Worker:
+        pipeline = Pipeline()
+
+    assert _collect_worker_identity(Worker()) == {
+        "resolved_attention_backends": ["FLASH_ATTN"],
+        "resolved_model_path": "/models/local",
+    }

fastvideo/tests/performance/test_inference_performance.py

@@ -12,12 +12,25 @@ import os
 import time
 from collections.abc import Mapping
 from datetime import datetime, timezone
+from typing import Any
 
 import torch
 import pytest
 
 from fastvideo import VideoGenerator
 from fastvideo.logger import init_logger
+from fastvideo.tests.performance.identity import (
+    benchmark_identity_from_config,
+    build_recipe_from_benchmark_config,
+    environment_fingerprint,
+    environment_metadata,
+    hardware_profile,
+    hardware_profile_id,
+    recipe_fingerprint,
+    resolved_revision_from_model_path,
+    software_profile,
+    software_profile_id,
+)
 from fastvideo.worker.multiproc_executor import MultiprocExecutor
 
 logger = init_logger(__name__)
@@ -34,8 +47,11 @@ V2_REQUIRED_IDENTITY_FIELDS = (
     "variant_id",
     "benchmark_version",
 )
+# "recipe" is no longer config-declarable: the fingerprint follow-up landed,
+# and the generated recipe document owns that key in emitted records. A config
+# declaring it now fails validation loudly instead of being silently
+# overwritten by the generated one.
 V2_OPTIONAL_METADATA_FIELDS = (
-    "recipe",
     "metric_threshold_policy",
     "quality_metadata",
 )
@@ -231,6 +247,110 @@ def _write_results(results):
     logger.info("Performance results written to %s", filepath)
 
 
+def _backend_name(value) -> str:
+    if hasattr(value, "name"):
+        return str(value.name)
+    return str(value)
+
+
+def _collect_worker_identity(worker) -> dict[str, Any]:
+    pipeline = getattr(worker, "pipeline", None)
+    model_path = getattr(pipeline, "model_path", None)
+    backends: set[str] = set()
+
+    modules = getattr(pipeline, "modules", None)
+    if isinstance(modules, Mapping):
+        module_iter = modules.values()
+    elif isinstance(pipeline, torch.nn.Module):
+        module_iter = (pipeline,)
+    else:
+        module_iter = ()
+
+    for module in module_iter:
+        if isinstance(module, torch.nn.Module):
+            for submodule in module.modules():
+                backend = getattr(submodule, "backend", None)
+                if backend is not None:
+                    backends.add(_backend_name(backend))
+        else:
+            backend = getattr(module, "backend", None)
+            if backend is not None:
+                backends.add(_backend_name(backend))
+
+    return {
+        "resolved_attention_backends": sorted(backends),
+        "resolved_model_path": model_path,
+    }
+
+
+def _single_or_list(values: set[str]) -> str | list[str] | None:
+    ordered = sorted(values)
+    if not ordered:
+        return None
+    if len(ordered) == 1:
+        return ordered[0]
+    return ordered
+
+
+def _runtime_identity_from_generator(generator) -> dict[str, Any]:
+    worker_records = generator.executor.collective_rpc(_collect_worker_identity)
+    resolved_backends: set[str] = set()
+    resolved_revisions: set[str] = set()
+
+    for record in worker_records:
+        resolved_backends.update(record.get("resolved_attention_backends") or [])
+        revision = resolved_revision_from_model_path(record.get("resolved_model_path"))
+        if revision is not None:
+            resolved_revisions.add(revision)
+
+    return {
+        "resolved_attention_backend": _single_or_list(resolved_backends),
+        "resolved_model_revision": _single_or_list(resolved_revisions),
+    }
+
+
+def _benchmark_identity_fields(cfg):
+    identity = benchmark_identity_from_config(cfg)
+    return {
+        key: identity[key]
+        for key in ("workload_id", "variant_id", "benchmark_version")
+    }
+
+
+def _build_identity_fields(cfg, init_kwargs, prompt, runtime_identity):
+    # Legacy v1 configs (no config_schema_version) stay on the legacy
+    # (model_id, gpu_type) cohort: they lack the required identity fields,
+    # and building a recipe for them would raise AFTER the GPU measurement
+    # completed. The validator already enforces that v2 configs carry the
+    # identity fields, so v2 records always get the full identity block.
+    if cfg.get("config_schema_version") is None:
+        return {}
+    recipe = build_recipe_from_benchmark_config(
+        cfg,
+        resolved_attention_backend=runtime_identity.get("resolved_attention_backend"),
+        resolved_model_revision=runtime_identity.get("resolved_model_revision"),
+        measured_prompts=[prompt],
+    )
+    num_gpus = init_kwargs.get("num_gpus", cfg.get("run_config", {}).get("required_gpus", 1))
+    hw_profile = hardware_profile(num_gpus=num_gpus)
+    sw_profile = software_profile()
+    env_metadata = environment_metadata(
+        hardware=hw_profile,
+        software=sw_profile,
+    )
+    return {
+        "recipe": recipe,
+        "recipe_fingerprint": recipe_fingerprint(recipe),
+        "hardware_profile": hw_profile,
+        "hardware_profile_id": hardware_profile_id(hw_profile),
+        "software_profile": sw_profile,
+        "software_profile_id": software_profile_id(sw_profile),
+        "environment_metadata": env_metadata,
+        "environment_fingerprint": environment_fingerprint(env_metadata),
+        **_benchmark_identity_fields(cfg),
+    }
+
+
 # -- Test -------------------------------------------------------------------
 
 def _run_benchmark(cfg):
@@ -268,6 +388,7 @@ def _run_benchmark(cfg):
             model_path=model_info["model_path"],
             **init_kwargs,
         )
+        runtime_identity = _runtime_identity_from_generator(generator)
 
         for i in range(num_warmup):
             logger.info("Warmup run %d/%d", i + 1, num_warmup)
@@ -322,6 +443,7 @@ def _run_benchmark(cfg):
         "dit_time_s": _avg_component(all_component_times, "dit_time_s"),
         "vae_decode_time_s": _avg_component(all_component_times,
                                             "vae_decode_time_s"),
+        **_build_identity_fields(cfg, init_kwargs, prompt, runtime_identity),
     }
 
     logger.info(

fastvideo/tests/ssim/inference_similarity_utils.py

@@ -3,9 +3,9 @@
 from __future__ import annotations
 
 import os
+from collections.abc import Iterator
 from contextlib import contextmanager
 from logging import Logger
-from typing import Iterator
 
 from fastvideo import VideoGenerator
 from fastvideo.tests.ssim.reference_utils import (
@@ -67,6 +67,12 @@ def _find_reference_video(reference_folder: str, prompt: str) -> str:
     raise FileNotFoundError("Reference video missing")
 
 
+def _remove_stale_generated_video(output_dir: str, output_video_name: str) -> None:
+    stale_path = os.path.join(output_dir, output_video_name)
+    if os.path.exists(stale_path):
+        os.remove(stale_path)
+
+
 def _assert_similarity(
     *,
     logger: Logger,
@@ -214,6 +220,7 @@ def run_text_to_video_similarity_test(
         )
         output_video_name = f"{prompt[:100].strip()}.mp4"
         os.makedirs(output_dir, exist_ok=True)
+        _remove_stale_generated_video(output_dir, output_video_name)
 
         params_map = select_ssim_params(
             default_params_map,
@@ -289,6 +296,7 @@ def run_image_to_video_similarity_test(
         )
         output_video_name = f"{prompt[:100].strip()}.mp4"
         os.makedirs(output_dir, exist_ok=True)
+        _remove_stale_generated_video(output_dir, output_video_name)
 
         params_map = select_ssim_params(
             default_params_map,

fastvideo/tests/ssim/test_dreamx_world_similarity.py

@@ -0,0 +1,191 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+import pytest
+from PIL import Image, ImageDraw
+
+from fastvideo.logger import init_logger
+from fastvideo.tests.ssim.inference_similarity_utils import (
+    resolve_inference_device_reference_folder,
+    run_image_to_video_similarity_test,
+    run_text_to_video_similarity_test,
+)
+
+logger = init_logger(__name__)
+
+REQUIRED_GPUS = 1
+
+device_reference_folder = resolve_inference_device_reference_folder(logger)
+
+_MODEL_PATH = os.getenv(
+    "DREAMX_WORLD_SSIM_MODEL_PATH",
+    "FastVideo/DreamX-World-5B-Cam-Diffusers",
+)
+_AR_MODEL_PATH = os.getenv(
+    "DREAMX_WORLD_AR_SSIM_MODEL_PATH",
+    "FastVideo/DreamX-World-5B-Diffusers",
+)
+
+DREAMX_WORLD_PARAMS = {
+    "num_gpus": 1,
+    "model_path": _MODEL_PATH,
+    "height": 64,
+    "width": 64,
+    "num_frames": 9,
+    "num_inference_steps": 1,
+    "guidance_scale": 1.0,
+    "seed": 1024,
+    "fps": 16,
+}
+
+DREAMX_WORLD_FULL_QUALITY_PARAMS = {
+    **DREAMX_WORLD_PARAMS,
+    "height": 480,
+    "width": 832,
+    "num_frames": 161,
+    "num_inference_steps": 30,
+    "guidance_scale": 5.0,
+}
+
+
+DREAMX_WORLD_AR_PARAMS = {
+    "num_gpus": 1,
+    "model_path": _AR_MODEL_PATH,
+    "height": 192,
+    "width": 192,
+    "num_frames": 81,
+    "num_inference_steps": 4,
+    "guidance_scale": 1.0,
+    "seed": 2048,
+    "fps": 16,
+}
+
+DREAMX_WORLD_AR_FULL_QUALITY_PARAMS = {
+    **DREAMX_WORLD_AR_PARAMS,
+    "height": 704,
+    "width": 1280,
+    "num_frames": 1005,
+}
+
+DREAMX_WORLD_MODEL_TO_PARAMS = {
+    "DreamX-World-5B-Cam": DREAMX_WORLD_PARAMS,
+}
+
+DREAMX_WORLD_AR_MODEL_TO_PARAMS = {
+    "DreamX-World-5B": DREAMX_WORLD_AR_PARAMS,
+}
+
+FULL_QUALITY_DREAMX_WORLD_MODEL_TO_PARAMS = {
+    "DreamX-World-5B-Cam": DREAMX_WORLD_FULL_QUALITY_PARAMS,
+}
+
+FULL_QUALITY_DREAMX_WORLD_AR_MODEL_TO_PARAMS = {
+    "DreamX-World-5B": DREAMX_WORLD_AR_FULL_QUALITY_PARAMS,
+}
+
+DREAMX_WORLD_TEST_CASES = [
+    (
+        "A cinematic first-person drive through a futuristic coastal city at sunrise, "
+        "reflective glass towers, clean streets, soft volumetric light.",
+        ("w", "d", "w"),
+        (4.0, 2.0, 4.0),
+    ),
+]
+
+
+DREAMX_WORLD_AR_TEST_CASES = [
+    (
+        "A long autonomous drive through a futuristic coastal city at sunrise, "
+        "smooth forward camera motion, reflective glass towers, clean streets.",
+        ("w", "d", "w", "a"),
+        (2.0, 1.0, 2.0, 1.0),
+    ),
+]
+
+
+def _write_deterministic_reference_image(path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    image = Image.new("RGB", (96, 96), color=(42, 76, 112))
+    draw = ImageDraw.Draw(image)
+    draw.rectangle((0, 56, 96, 96), fill=(36, 44, 52))
+    draw.polygon([(0, 56), (48, 30), (96, 56)], fill=(120, 142, 158))
+    draw.rectangle((34, 42, 62, 72), fill=(178, 198, 212))
+    draw.line((0, 80, 96, 66), fill=(238, 209, 124), width=3)
+    image.save(path)
+
+
+@pytest.mark.parametrize(("prompt", "action_list", "action_speed_list"), DREAMX_WORLD_TEST_CASES)
+@pytest.mark.parametrize("attention_backend_name", ["TORCH_SDPA"])
+@pytest.mark.parametrize("model_id", list(DREAMX_WORLD_MODEL_TO_PARAMS.keys()))
+def test_dreamx_world_inference_similarity(
+    prompt: str,
+    action_list: tuple[str, ...],
+    action_speed_list: tuple[float, ...],
+    attention_backend_name: str,
+    model_id: str,
+    tmp_path: Path,
+) -> None:
+    image_path = tmp_path / "dreamx_world_ssim_input.png"
+    _write_deterministic_reference_image(image_path)
+
+    run_image_to_video_similarity_test(
+        logger=logger,
+        script_dir=os.path.dirname(os.path.abspath(__file__)),
+        device_reference_folder=device_reference_folder,
+        prompt=prompt,
+        image_path=str(image_path),
+        attention_backend_name=attention_backend_name,
+        model_id=model_id,
+        default_params_map=DREAMX_WORLD_MODEL_TO_PARAMS,
+        full_quality_params_map=FULL_QUALITY_DREAMX_WORLD_MODEL_TO_PARAMS,
+        min_acceptable_ssim=0.98,
+        init_kwargs_override={
+            "use_fsdp_inference": False,
+            "dit_cpu_offload": False,
+            "vae_cpu_offload": True,
+            "text_encoder_cpu_offload": True,
+            "pin_cpu_memory": False,
+            "override_pipeline_cls_name": "DreamXWorldPipeline",
+        },
+        generation_kwargs_override={
+            "action_list": list(action_list),
+            "action_speed_list": list(action_speed_list),
+        },
+    )
+
+@pytest.mark.parametrize(("prompt", "action_list", "action_speed_list"), DREAMX_WORLD_AR_TEST_CASES)
+@pytest.mark.parametrize("attention_backend_name", ["TORCH_SDPA"])
+@pytest.mark.parametrize("model_id", list(DREAMX_WORLD_AR_MODEL_TO_PARAMS.keys()))
+def test_dreamx_world_ar_inference_similarity(
+    prompt: str,
+    action_list: tuple[str, ...],
+    action_speed_list: tuple[float, ...],
+    attention_backend_name: str,
+    model_id: str,
+) -> None:
+    run_text_to_video_similarity_test(
+        logger=logger,
+        script_dir=os.path.dirname(os.path.abspath(__file__)),
+        device_reference_folder=device_reference_folder,
+        prompt=prompt,
+        attention_backend_name=attention_backend_name,
+        model_id=model_id,
+        default_params_map=DREAMX_WORLD_AR_MODEL_TO_PARAMS,
+        full_quality_params_map=FULL_QUALITY_DREAMX_WORLD_AR_MODEL_TO_PARAMS,
+        min_acceptable_ssim=0.98,
+        init_kwargs_override={
+            "use_fsdp_inference": False,
+            "dit_cpu_offload": False,
+            "vae_cpu_offload": True,
+            "text_encoder_cpu_offload": True,
+            "pin_cpu_memory": False,
+            "override_pipeline_cls_name": "DreamXWorldARPipeline",
+        },
+        generation_kwargs_override={
+            "action_list": list(action_list),
+            "action_speed_list": list(action_speed_list),
+        },
+    )

fastvideo/tests/ssim/test_kandinsky5_similarity.py

@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: Apache-2.0
+"""SSIM regression test for Kandinsky-5.0 Lite text-to-video.
+
+Runs a reduced-size generation (256x384, 21 frames) with a fixed seed and
+compares against a device-specific reference video via MS-SSIM. The preset
+default is 512x768 x 121 frames; the reduced size keeps CI cheap while the
+patch (1, 2, 2) and VAE temporal-4 / spatial-8 constraints stay satisfied
+(dims divisible by 16, num_frames % 4 == 1).
+"""
+import os
+
+import pytest
+
+from fastvideo.logger import init_logger
+from fastvideo.pipelines.basic.kandinsky5.presets import KANDINSKY5_T2V_LITE_5S
+from fastvideo.tests.ssim.inference_similarity_utils import (
+    resolve_inference_device_reference_folder,
+    run_text_to_video_similarity_test,
+)
+
+logger = init_logger(__name__)
+
+REQUIRED_GPUS = 1
+
+device_reference_folder = resolve_inference_device_reference_folder(logger)
+
+_PRESET_DEFAULTS = KANDINSKY5_T2V_LITE_5S.defaults
+
+KANDINSKY5_T2V_PARAMS = {
+    "num_gpus": 1,
+    "model_path": "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
+    "height": 256,
+    "width": 384,
+    "num_frames": 21,
+    "num_inference_steps": _PRESET_DEFAULTS["num_inference_steps"],
+    "guidance_scale": _PRESET_DEFAULTS["guidance_scale"],
+    "seed": 1024,
+    "sp_size": 1,
+    "tp_size": 1,
+    "fps": _PRESET_DEFAULTS["fps"],
+    "neg_prompt": _PRESET_DEFAULTS["negative_prompt"],
+}
+
+KANDINSKY5_T2V_FULL_QUALITY_PARAMS = {
+    **KANDINSKY5_T2V_PARAMS,
+    "height": _PRESET_DEFAULTS["height"],
+    "width": _PRESET_DEFAULTS["width"],
+    # num_frames stays at the inherited reduced 21 even at full quality
+    # (preset default: 121).
+}
+
+KANDINSKY5_T2V_MODEL_TO_PARAMS = {
+    "Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers": KANDINSKY5_T2V_PARAMS,
+}
+FULL_QUALITY_KANDINSKY5_T2V_MODEL_TO_PARAMS = {
+    "Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers": KANDINSKY5_T2V_FULL_QUALITY_PARAMS,
+}
+
+KANDINSKY5_T2V_TEST_PROMPTS = [
+    "A curious raccoon peers through a vibrant field of yellow sunflowers, its "
+    "eyes wide with interest. The playful yet serene atmosphere is complemented "
+    "by soft natural light filtering through the petals. Mid-shot, warm and "
+    "cheerful tones.",
+]
+
+
+@pytest.mark.parametrize("prompt", KANDINSKY5_T2V_TEST_PROMPTS)
+@pytest.mark.parametrize("attention_backend_name", ["FLASH_ATTN"])
+@pytest.mark.parametrize("model_id", list(KANDINSKY5_T2V_MODEL_TO_PARAMS.keys()))
+def test_kandinsky5_t2v_inference_similarity(
+    prompt: str,
+    attention_backend_name: str,
+    model_id: str,
+) -> None:
+    # The SSIM lane image bakes FASTVIDEO_FA4=1, but on L40S (sm89) the CLIP
+    # encoder's dense MHA inference call routes into the FA4 CuTeDSL JIT,
+    # which fails to compile there (nvvm.fmax signature mismatch vs the
+    # image's nvidia_cutlass_dsl). Pin FA4 off so reference seeding and CI
+    # runs both use the FA2 path with identical numerics. Scoped to this test
+    # and restored afterwards: the other SSIM references are seeded with FA4
+    # on, so a module-level pin would corrupt every test collected in the
+    # same pytest process.
+    saved_fa4 = os.environ.get("FASTVIDEO_FA4")
+    os.environ["FASTVIDEO_FA4"] = "0"
+    try:
+        run_text_to_video_similarity_test(
+            logger=logger,
+            script_dir=os.path.dirname(os.path.abspath(__file__)),
+            device_reference_folder=device_reference_folder,
+            prompt=prompt,
+            attention_backend_name=attention_backend_name,
+            model_id=model_id,
+            default_params_map=KANDINSKY5_T2V_MODEL_TO_PARAMS,
+            full_quality_params_map=FULL_QUALITY_KANDINSKY5_T2V_MODEL_TO_PARAMS,
+            min_acceptable_ssim=0.93,
+            # Match examples/inference/basic/basic_kandinsky5_t2v.py: FSDP
+            # inference is not used for Kandinsky-5, and the Qwen2.5-VL text
+            # encoder stays on CPU between encodes.
+            init_kwargs_override={
+                "use_fsdp_inference": False,
+                "text_encoder_cpu_offload": True,
+                "pin_cpu_memory": True,
+            },
+        )
+    finally:
+        if saved_fa4 is None:
+            os.environ.pop("FASTVIDEO_FA4", None)
+        else:
+            os.environ["FASTVIDEO_FA4"] = saved_fa4

fastvideo/tests/training/test_trackers.py

@@ -0,0 +1,225 @@
+from __future__ import annotations
+
+import builtins
+from io import BytesIO
+from pathlib import Path
+import sys
+from types import ModuleType
+from typing import Any
+
+import numpy as np
+from PIL import Image
+import pytest
+
+from fastvideo.training import trackers as trackers_module
+from fastvideo.training.trackers import (
+    _coerce_video_fps,
+    _prepare_video_array,
+    BaseTracker,
+    SequentialTracker,
+    SwanlabTracker,
+    WandbTracker,
+)
+
+
+class _FakeSwanlab(ModuleType):
+    def __init__(self) -> None:
+        super().__init__("swanlab")
+        self.init_kwargs: dict[str, Any] | None = None
+        self.logged: list[tuple[dict[str, Any], int]] = []
+        self.video_paths: list[str] = []
+        self.finished = False
+
+    def init(self, **kwargs: Any) -> object:
+        self.init_kwargs = kwargs
+        return object()
+
+    def log(self, metrics: dict[str, Any], step: int) -> None:
+        self.logged.append((metrics, step))
+
+    def finish(self) -> None:
+        self.finished = True
+
+    def Video(self, file_path: str, caption: str | None = None) -> dict[str, Any]:  # noqa: N802
+        self.video_paths.append(file_path)
+        return {
+            "content": Path(file_path).read_bytes(),
+            "caption": caption,
+        }
+
+
+def _install_fake_swanlab(monkeypatch: pytest.MonkeyPatch) -> _FakeSwanlab:
+    swanlab = _FakeSwanlab()
+    monkeypatch.setitem(sys.modules, "swanlab", swanlab)
+    return swanlab
+
+
+def test_prepare_video_array_converts_tchw_to_thwc() -> None:
+    video = np.arange(2 * 3 * 2 * 4, dtype=np.uint8).reshape(2, 3, 2, 4)
+
+    prepared = _prepare_video_array(video)
+
+    np.testing.assert_array_equal(prepared, video.transpose(0, 2, 3, 1))
+
+
+def test_default_video_fps_is_16() -> None:
+    assert _coerce_video_fps(None) == 16
+
+
+def test_wandb_video_uses_shared_default_fps(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+) -> None:
+    wandb = ModuleType("wandb")
+    video_calls: list[dict[str, Any]] = []
+    wandb.init = lambda **_: object()
+
+    def video(data: Any, **kwargs: Any) -> dict[str, Any]:
+        video_calls.append(kwargs)
+        return {"data": data, **kwargs}
+
+    wandb.Video = video
+    monkeypatch.setitem(sys.modules, "wandb", wandb)
+    tracker = WandbTracker("project", str(tmp_path))
+
+    tracker.video(np.zeros((2, 3, 2, 2), dtype=np.uint8))
+
+    assert video_calls == [{"fps": 16, "format": "mp4"}]
+
+
+def test_prepare_video_array_tiles_and_pads_batches() -> None:
+    video = np.zeros((3, 1, 3, 1, 1), dtype=np.uint8)
+    video[0] = 1
+    video[1] = 2
+    video[2] = 3
+
+    prepared = _prepare_video_array(video)
+
+    assert prepared.shape == (1, 1, 4, 3)
+    np.testing.assert_array_equal(
+        prepared[0, 0],
+        np.array(
+            [
+                [1, 1, 1],
+                [2, 2, 2],
+                [3, 3, 3],
+                [0, 0, 0],
+            ],
+            dtype=np.uint8,
+        ),
+    )
+
+
+def test_swanlab_video_converts_array_to_gif_and_logs_artifact(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+) -> None:
+    swanlab = _install_fake_swanlab(monkeypatch)
+    tracker = SwanlabTracker("project", str(tmp_path), run_name="run")
+    video = np.zeros((2, 3, 2, 2), dtype=np.uint8)
+    video[0, 0] = 255
+    video[1, 2] = 255
+
+    artifact = tracker.video(video, caption="validation sample", fps=5, format="mp4")
+    tracker.log_artifacts({"validation/videos": [artifact]}, step=7)
+
+    assert artifact["content"].startswith((b"GIF87a", b"GIF89a"))
+    assert artifact["caption"] == "validation sample"
+    assert not Path(swanlab.video_paths[0]).exists()
+    with Image.open(BytesIO(artifact["content"])) as image:
+        assert image.n_frames == 2
+        assert image.info["duration"] == 200
+    assert swanlab.logged == [({"validation/videos": [artifact]}, 7)]
+
+
+@pytest.mark.parametrize(("requested_fps", "expected_fps"), [(None, 12.0), (2, 2.0)])
+def test_swanlab_video_uses_requested_or_source_fps_during_gif_conversion(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+    requested_fps: int | None,
+    expected_fps: float,
+) -> None:
+    _install_fake_swanlab(monkeypatch)
+    tracker = SwanlabTracker("project", str(tmp_path))
+    frames = [np.zeros((2, 2, 3), dtype=np.uint8), np.ones((2, 2, 3), dtype=np.uint8)]
+    written_fps: list[float] = []
+    write_gif = trackers_module._write_gif
+
+    monkeypatch.setattr(trackers_module, "_read_video_file", lambda _: (frames, 12.0))
+
+    def capture_fps(file_path: str, video_frames: Any, fps: float) -> None:
+        written_fps.append(fps)
+        write_gif(file_path, video_frames, fps)
+
+    monkeypatch.setattr(trackers_module, "_write_gif", capture_fps)
+
+    artifact = tracker.video(tmp_path / "validation.mp4", fps=requested_fps)
+
+    assert artifact["content"].startswith((b"GIF87a", b"GIF89a"))
+    assert written_fps == [expected_fps]
+
+
+def test_swanlab_video_passes_existing_gif_through(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+) -> None:
+    swanlab = _install_fake_swanlab(monkeypatch)
+    tracker = SwanlabTracker("project", str(tmp_path))
+    gif_path = tmp_path / "validation.gif"
+    Image.new("RGB", (2, 2)).save(gif_path, format="GIF")
+
+    artifact = tracker.video(gif_path, caption="existing")
+
+    assert swanlab.video_paths == [str(gif_path)]
+    assert gif_path.exists()
+    assert artifact["caption"] == "existing"
+
+
+def test_swanlab_missing_dependency_error_is_actionable(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+) -> None:
+    monkeypatch.delitem(sys.modules, "swanlab", raising=False)
+    real_import = builtins.__import__
+
+    def missing_swanlab(name: str, *args: Any, **kwargs: Any) -> Any:
+        if name == "swanlab":
+            raise ModuleNotFoundError("No module named 'swanlab'", name="swanlab")
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", missing_swanlab)
+
+    with pytest.raises(ModuleNotFoundError, match=r"fastvideo\[swanlab\]"):
+        SwanlabTracker("project", str(tmp_path))
+
+
+class _RecordingTracker(BaseTracker):
+    def __init__(self, name: str) -> None:
+        super().__init__()
+        self.name = name
+        self.artifact_logs: list[dict[str, Any]] = []
+
+    def video(
+        self,
+        data: Any,
+        *,
+        caption: str | None = None,
+        fps: int | None = None,
+        format: str | None = None,
+    ) -> str:
+        return f"{self.name}:{data}"
+
+    def log_artifacts(self, artifacts: dict[str, Any], step: int) -> None:
+        self.artifact_logs.append(artifacts)
+
+
+def test_sequential_tracker_logs_backend_specific_video_artifacts() -> None:
+    first = _RecordingTracker("first")
+    second = _RecordingTracker("second")
+    tracker = SequentialTracker([first, second])
+
+    artifact = tracker.video("sample.mp4")
+    tracker.log_artifacts({"validation/videos": [artifact]}, step=3)
+
+    assert first.artifact_logs == [{"validation/videos": ["first:sample.mp4"]}]
+    assert second.artifact_logs == [{"validation/videos": ["second:sample.mp4"]}]

fastvideo/tests/transformers/test_cosmos.py

@@ -11,6 +11,7 @@ from fastvideo.forward_context import set_forward_context
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
 from fastvideo.models.loader.component_loader import TransformerLoader
+from fastvideo.tests.utils import skip_if_gated_repo_inaccessible
 from fastvideo.utils import maybe_download_model
 from fastvideo.configs.models.dits import CosmosVideoConfig
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
@@ -25,18 +26,12 @@ BASE_MODEL_PATH = "nvidia/Cosmos-Predict2-2B-Video2World"
 
 
 def _resolve_model_path() -> str:
-    try:
-        return maybe_download_model(
-            BASE_MODEL_PATH,
-            local_dir=os.path.join("data", BASE_MODEL_PATH),
-        )
-    except ValueError as exc:
-        pytest.skip(
-            "Skipping Cosmos transformer test because the configured "
-            "HuggingFace token cannot access the gated Cosmos weights: "
-            f"{exc}",
-            allow_module_level=True,
-        )
+    local_dir = os.path.join("data", BASE_MODEL_PATH)
+    skip_if_gated_repo_inaccessible(BASE_MODEL_PATH,
+                                    local_path=local_dir,
+                                    test_name="Cosmos transformer test",
+                                    allow_module_level=True)
+    return maybe_download_model(BASE_MODEL_PATH, local_dir=local_dir)
 
 
 MODEL_PATH = _resolve_model_path()

fastvideo/tests/transformers/test_cosmos2_5.py

@@ -19,6 +19,7 @@ if os.path.exists(COSMOS_PREDICT2_5_PATH) and COSMOS_PREDICT2_5_PATH not in sys.
 
 from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
+from fastvideo.tests.utils import skip_if_gated_repo_inaccessible
 from fastvideo.utils import maybe_download_model
 # Use Cosmos 2.5 specific config
 from fastvideo.configs.models.dits.cosmos2_5 import Cosmos25VideoConfig
@@ -43,7 +44,14 @@ CHECKPOINT_SUBDIR = "base/post-trained"
 CHECKPOINT_FILENAME = "81edfebe-bd6a-4039-8c1d-737df1a790bf_ema_bf16.pt"
 
 
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH, local_dir=None)
+def _resolve_model_path() -> str:
+    skip_if_gated_repo_inaccessible(BASE_MODEL_PATH,
+                                    test_name="Cosmos 2.5 transformer test",
+                                    allow_module_level=True)
+    return maybe_download_model(BASE_MODEL_PATH, local_dir=None)
+
+
+MODEL_PATH = _resolve_model_path()
 TRANSFORMER_PATH = os.path.join(MODEL_PATH, CHECKPOINT_SUBDIR, "transformer")
 if not os.path.exists(TRANSFORMER_PATH):
     # Try without subdirectory
@@ -592,4 +600,3 @@ if __name__ == "__main__":
     # Run tests directly
     test_cosmos25_transformer()
     test_cosmos25_transformer_video()
-

fastvideo/tests/transformers/test_hunyuanvideo.py

@@ -1,104 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import glob
-import json
-import os
-import pytest
-import torch
-
-from fastvideo.configs.pipelines.base import PipelineConfig
-from fastvideo.distributed.parallel_state import get_sp_parallel_rank, get_sp_world_size
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.utils import maybe_download_model
-from fastvideo.forward_context import set_forward_context
-from fastvideo.configs.models.dits import HunyuanVideoConfig
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "hunyuanvideo-community/HunyuanVideo"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH, local_dir=os.path.join("data", BASE_MODEL_PATH))
-TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-CONFIG_PATH = os.path.join(TRANSFORMER_PATH, "config.json")
-
-LOCAL_RANK = 0
-RANK = 0
-WORLD_SIZE = 1
-
-# Latent generated on commit c021e8a27cf437ac22827f2bc58b7f006561317f with 1 x L40S
-REFERENCE_LATENT = 93.23263549804688
-
-
-@pytest.mark.usefixtures("distributed_setup")
-@pytest.mark.skip(reason="Skipping test_hunyuanvideo_distributed")
-def test_hunyuanvideo_distributed():
-    logger.info(f"Initializing process: rank={RANK}, local_rank={LOCAL_RANK}, world_size={WORLD_SIZE}")
-
-    torch.cuda.set_device(f"cuda:{LOCAL_RANK}")
-
-    # Get tensor parallel info
-    sp_rank = get_sp_parallel_rank()
-    sp_world_size = get_sp_world_size()
-
-    logger.info(f"Process rank {RANK} initialized with SP rank {sp_rank} in SP world size {sp_world_size}")
-
-    config = json.load(open(CONFIG_PATH))
-    # remove   "_class_name": "HunyuanVideoTransformer3DModel",   "_diffusers_version": "0.32.0.dev0",
-    # TODO: write normalize config function
-    config.pop("_class_name")
-    config.pop("_diffusers_version")
-
-    precision_str = "bf16"
-    args = FastVideoArgs(
-        model_path=TRANSFORMER_PATH,
-        dit_cpu_offload=True,
-        pipeline_config=PipelineConfig(dit_config=HunyuanVideoConfig(), dit_precision=precision_str),
-    )
-    args.device = torch.device(f"cuda:{LOCAL_RANK}")
-
-    loader = TransformerLoader()
-    model = loader.load(TRANSFORMER_PATH, args)
-
-    model.eval()
-
-    # Create random inputs for testing
-    batch_size = 1
-    seq_len = 3
-    device = torch.device(f"cuda:{LOCAL_RANK}")
-
-    # Video latents [B, C, T, H, W]
-    hidden_states = torch.randn(batch_size, 16, 8, 16, 16, device=device, dtype=torch.bfloat16)
-    chunk_per_rank = hidden_states.shape[2] // sp_world_size
-    hidden_states = hidden_states[:, :, sp_rank * chunk_per_rank : (sp_rank + 1) * chunk_per_rank]
-
-    # Text embeddings [B, L, D] (including global token)
-    encoder_hidden_states = torch.randn(batch_size, seq_len + 1, 4096, device=device, dtype=torch.bfloat16)
-
-    # Timestep
-    timestep = torch.tensor([500], device=device, dtype=torch.bfloat16)
-    forward_batch = ForwardBatch(
-        data_type="dummy",
-    )
-
-    # Disable gradients for inference
-    with torch.no_grad():
-        # Run inference on model
-        with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
-            with set_forward_context(current_timestep=0, attn_metadata=None, forward_batch=forward_batch):
-                output = model(
-                    hidden_states=hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    timestep=timestep,
-                )
-
-    latent = output.double().sum().item()
-
-    # Check if latents are similar
-    diff_output_latents = abs(REFERENCE_LATENT - latent)
-    logger.info(f"Reference latent: {REFERENCE_LATENT}, Current latent: {latent}")
-    assert diff_output_latents < 1e-4, f"Output latents differ significantly: max diff = {diff_output_latents}"

fastvideo/tests/transformers/test_hyworld.py

@@ -15,6 +15,7 @@ from fastvideo.forward_context import set_forward_context
 from fastvideo.logger import init_logger
 from fastvideo.models.loader.component_loader import TransformerLoader
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.platforms import AttentionBackendEnum
 from fastvideo.utils import maybe_download_model
 
 logger = init_logger(__name__)
@@ -24,13 +25,25 @@ os.environ["MASTER_PORT"] = "29503"
 
 MODEL_PATH = maybe_download_model("FastVideo/HY-WorldPlay-Bidirectional-Diffusers")
 TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-REFERENCE_LATENT = -197132.85557549074  # Pre-computed reference value
+REFERENCE_LATENTS = {
+    AttentionBackendEnum.FLASH_ATTN: -197132.85557549074,
+    AttentionBackendEnum.TORCH_SDPA: -211007.95158730447,
+}
 
 
 @pytest.mark.usefixtures("distributed_setup")
 def test_hyworld_transformer():
     transformer_path = TRANSFORMER_PATH
 
+    # An env-forced backend is knowable before the multi-GB download and GPU
+    # load; the authoritative check on the layer's resolved backend still runs
+    # after construction below.
+    forced_backend_name = os.environ.get("FASTVIDEO_ATTENTION_BACKEND")
+    if forced_backend_name:
+        forced_backend = getattr(AttentionBackendEnum, forced_backend_name, None)
+        if forced_backend is not None and forced_backend not in REFERENCE_LATENTS:
+            pytest.skip(f"HYWorld transformer test has no reference latent for {forced_backend_name}.")
+
     sp_rank = get_sp_parallel_rank()
     sp_world_size = get_sp_world_size()
 
@@ -51,6 +64,9 @@ def test_hyworld_transformer():
     loader = TransformerLoader()
     model = loader.load(transformer_path, args).to(device, dtype=precision)
     model.eval()
+    attention_backend = model.double_blocks[0].attn.backend
+    if attention_backend not in REFERENCE_LATENTS:
+        pytest.skip(f"HYWorld transformer test has no reference latent for {attention_backend}.")
 
     total_params = sum(p.numel() for p in model.parameters())
     weight_sum = sum(p.to(torch.float64).sum().item() for p in model.parameters())
@@ -147,9 +163,10 @@ def test_hyworld_transformer():
 
     latent = output.double().sum().item()
 
-    diff = abs(REFERENCE_LATENT - latent)
-    relative_diff = diff / abs(REFERENCE_LATENT)
-    logger.info(f"Reference latent: {REFERENCE_LATENT}, Current latent: {latent}")
+    reference_latent = REFERENCE_LATENTS[attention_backend]
+    diff = abs(reference_latent - latent)
+    relative_diff = diff / abs(reference_latent)
+    logger.info(f"Reference latent: {reference_latent}, Current latent: {latent}")
     logger.info(f"Absolute diff: {diff}, Relative diff: {relative_diff * 100:.4f}%")
 
     # Allow 0.5% relative difference

fastvideo/tests/transformers/test_matrixgame2.py

@@ -1,133 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import os
-
-import pytest
-import torch
-
-from fastvideo.configs.models.dits.matrixgame2 import MatrixGame2WanVideoConfig
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.distributed.parallel_state import (
-    get_sp_parallel_rank,
-    get_sp_world_size,
-)
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.forward_context import set_forward_context
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import TransformerLoader
-from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
-from fastvideo.utils import maybe_download_model
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29505"
-os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
-
-MODEL_PATH = maybe_download_model("FastVideo/Matrix-Game-2.0-Base-Distilled-Diffusers")
-TRANSFORMER_PATH = os.path.join(MODEL_PATH, "transformer")
-REFERENCE_LATENT = -111835.58895772696
-
-
-@pytest.mark.skip(reason="Not reliably reproducible")
-@pytest.mark.usefixtures("distributed_setup")
-def test_matrixgame2_transformer():
-    transformer_path = TRANSFORMER_PATH
-
-    sp_rank = get_sp_parallel_rank()
-    sp_world_size = get_sp_world_size()
-
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-
-    args = FastVideoArgs(
-        model_path=transformer_path,
-        dit_cpu_offload=False,
-        use_fsdp_inference=False,
-        pipeline_config=PipelineConfig(
-            dit_config=MatrixGame2WanVideoConfig(), dit_precision=precision_str
-        ),
-    )
-    args.device = device
-
-    loader = TransformerLoader()
-    model = loader.load(transformer_path, args).to(device, dtype=precision)
-    model.eval()
-
-    total_params = sum(p.numel() for p in model.parameters())
-    weight_sum = sum(p.to(torch.float64).sum().item() for p in model.parameters())
-    weight_mean = weight_sum / total_params
-    logger.info("Total parameters: %s", total_params)
-    logger.info("Weight sum: %s", weight_sum)
-    logger.info("Weight mean: %s", weight_mean)
-
-    torch.manual_seed(42)
-
-    batch_size = 1
-    latent_frames = 15
-    latent_height = 44
-    latent_width = 80
-    image_dim = 1280
-    image_seq_len = 257
-
-    # Set num_frame_per_block to match actual latent_frames for action_model
-    model.num_frame_per_block = latent_frames
-    # Force full attention mode
-    model.local_attn_size = -1
-    # Clear cached block masks so they regenerate with correct settings
-    model.block_mask = None
-    model.block_mask_mouse = None
-    model.block_mask_keyboard = None
-
-    # Video latents [B, C, T, H, W]
-    x = torch.randn(batch_size, 16, latent_frames, latent_height, latent_width,
-                    device=device, dtype=precision)
-    cond_concat = torch.randn(batch_size, 20, latent_frames, latent_height, latent_width,
-                              device=device, dtype=precision)
-    hidden_states = torch.cat([x, cond_concat], dim=1)
-
-    if sp_world_size > 1:
-        chunk_per_rank = hidden_states.shape[2] // sp_world_size
-        hidden_states = hidden_states[:, :, sp_rank * chunk_per_rank:(sp_rank + 1) * chunk_per_rank]
-
-    # Image embeddings
-    encoder_hidden_states_image = [
-        torch.randn(batch_size, image_seq_len, image_dim, device=device, dtype=precision)
-    ]
-
-    # Timestep
-    timestep = torch.full((batch_size, latent_frames), 500, device=device, dtype=precision)
-
-    # Action conditions
-    num_pixel_frames = (latent_frames - 1) * 4 + 1
-    mouse_cond = torch.randn(batch_size, num_pixel_frames, 2, device=device, dtype=precision)
-    keyboard_cond = torch.randn(batch_size, num_pixel_frames, 4, device=device, dtype=precision)
-
-    # Text embeddings (empty for i2v)
-    encoder_hidden_states = torch.zeros(batch_size, 0, model.config.hidden_size,
-                                        device=device, dtype=precision)
-
-    forward_batch = ForwardBatch(data_type="dummy")
-    model.num_frame_per_block = latent_frames
-
-    with torch.no_grad():
-        with torch.amp.autocast('cuda', dtype=precision):
-            with set_forward_context(current_timestep=0, attn_metadata=None, forward_batch=forward_batch):
-                output = model(
-                    hidden_states=hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    encoder_hidden_states_image=encoder_hidden_states_image,
-                    timestep=timestep,
-                    mouse_cond=mouse_cond,
-                    keyboard_cond=keyboard_cond,
-                )
-
-    latent = output.double().sum().item()
-
-    diff = abs(REFERENCE_LATENT - latent)
-    relative_diff = diff / abs(REFERENCE_LATENT)
-    logger.info(f"Reference latent: {REFERENCE_LATENT}, Current latent: {latent}")
-    logger.info(f"Absolute diff: {diff}, Relative diff: {relative_diff * 100:.4f}%")
-
-    # The 0.19% diff now
-    assert relative_diff < 0.005, f"Output latents differ significantly: relative diff = {relative_diff * 100:.4f}% (max allowed: 0.5%)"

fastvideo/tests/utils.py

@@ -6,11 +6,45 @@ from fastvideo.logger import init_logger
 
 import numpy as np
 import torch
-from pytorch_msssim import ms_ssim, ssim
 
 logger = init_logger(__name__)
 
 
+def skip_if_gated_repo_inaccessible(repo_id: str,
+                                    *,
+                                    local_path: str | None = None,
+                                    test_name: str = "test",
+                                    allow_module_level: bool = False) -> None:
+    """Skip a test when ``repo_id`` is gated/private and the token lacks access.
+
+    Local weights win: when ``local_path`` already exists the check is skipped
+    entirely, so cached machines keep running offline. Transient hub failures
+    (offline, DNS, 429/5xx) do NOT skip either — the subsequent download will
+    serve from cache or fail loudly, preserving the CI failure signal. Only a
+    positively-identified authorization problem produces a skip.
+    """
+    import pytest
+
+    if local_path is not None and os.path.exists(local_path):
+        return
+
+    from huggingface_hub import HfApi
+    from huggingface_hub.errors import GatedRepoError, RepositoryNotFoundError
+
+    try:
+        HfApi().auth_check(repo_id)
+    except (GatedRepoError, RepositoryNotFoundError) as exc:
+        pytest.skip(
+            f"Skipping {test_name}: the configured HuggingFace token cannot "
+            f"access the gated repo {repo_id}: {exc}",
+            allow_module_level=allow_module_level,
+        )
+    except Exception as exc:  # noqa: BLE001 - hub unreachable, proxies, 5xx
+        logger.warning(
+            "Gated-repo access probe for %s failed (%s); proceeding — the "
+            "download will serve from cache or fail loudly.", repo_id, exc)
+
+
 def _read_video_frames(path: str) -> torch.Tensor:
     """Read video frames as a ``(T, C, H, W)`` uint8 tensor.
 
@@ -52,6 +86,8 @@ def compute_video_ssim_torchvision(video1_path, video2_path, use_ms_ssim=True):
         video2_path: Path to the second video.
         use_ms_ssim: Whether to use Multi-Scale Structural Similarity(MS-SSIM) instead of SSIM.
     """
+    from pytorch_msssim import ms_ssim, ssim
+
     print(f"Computing SSIM between {video1_path} and {video2_path}...")
     if not os.path.exists(video1_path):
         raise FileNotFoundError(f"Video1 not found: {video1_path}")

fastvideo/tests/vaes/test_hunyuan_vae.py

@@ -1,82 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-
-import json
-import os
-
-import pytest
-import torch
-from safetensors.torch import load_file
-
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.logger import init_logger
-
-# from fastvideo.models.vaes.hunyuanvae import (
-#     AutoencoderKLHunyuanVideo as MyHunyuanVAE)
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.models.loader.component_loader import VAELoader
-from fastvideo.configs.models.vaes import HunyuanVAEConfig
-from fastvideo.utils import maybe_download_model
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "hunyuanvideo-community/HunyuanVideo"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH, local_dir=os.path.join("data", BASE_MODEL_PATH))
-VAE_PATH = os.path.join(MODEL_PATH, "vae")
-CONFIG_PATH = os.path.join(VAE_PATH, "config.json")
-
-# Latent generated on commit d71a4ebffc2034922fc379568b6a6aa722f3744c with 1 x A40
-# torch 2.7.1
-A40_REFERENCE_LATENT = -106.22467041015625
-
-# Latent generated on commit 2b54068960c41d42221e8b8719a374b499855029 with 1 x L40S
-L40S_REFERENCE_LATENT = -157.09130859375
-
-
-@pytest.mark.usefixtures("distributed_setup")
-@pytest.mark.skip(reason="Skipping test_hunyuan_vae")
-def test_hunyuan_vae():
-    device = torch.device("cuda:0")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-    args = FastVideoArgs(
-        model_path=VAE_PATH, pipeline_config=PipelineConfig(vae_config=HunyuanVAEConfig(), vae_precision=precision_str)
-    )
-    args.device = device
-    args.vae_cpu_offload = False
-
-    loader = VAELoader()
-    model = loader.load(VAE_PATH, args)
-
-    model.enable_tiling(
-        tile_sample_min_height=32,
-        tile_sample_min_width=32,
-        tile_sample_min_num_frames=8,
-        tile_sample_stride_height=16,
-        tile_sample_stride_width=16,
-        tile_sample_stride_num_frames=4,
-    )
-
-    batch_size = 1
-
-    # Video input [B, C, T, H, W]
-    input_tensor = torch.randn(batch_size, 3, 21, 64, 64, device=device, dtype=torch.bfloat16)
-
-    # Disable gradients for inference
-    with torch.no_grad():
-        latent = model.encode(input_tensor).mean.double().sum().item()
-
-    # Check if latents are similar
-    device_name = torch.cuda.get_device_name()
-    if "A40" in device_name:
-        REFERENCE_LATENT = A40_REFERENCE_LATENT
-    elif "L40S" in device_name:
-        REFERENCE_LATENT = L40S_REFERENCE_LATENT
-    else:
-        raise ValueError(f"Unknown device: {device_name}")
-
-    diff_encoded_latents = abs(REFERENCE_LATENT - latent)
-    logger.info(f"Reference latent: {REFERENCE_LATENT}, Current latent: {latent}")
-    assert diff_encoded_latents < 1e-4, f"Encoded latents differ significantly: max diff = {diff_encoded_latents}"

fastvideo/tests/vaes/test_official_wan_vae.py

@@ -1,113 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import os
-
-import numpy as np
-import pytest
-import torch
-from diffusers import AutoencoderKLWan
-
-from fastvideo.configs.pipelines import PipelineConfig
-from fastvideo.fastvideo_args import FastVideoArgs
-from fastvideo.logger import init_logger
-from fastvideo.models.loader.component_loader import VAELoader
-from fastvideo.configs.models.vaes import WanVAEConfig
-from fastvideo.utils import maybe_download_model
-
-try:
-    from wan.modules.vae2_2 import Wan2_2_VAE
-except ImportError:
-    Wan2_2_VAE = None
-
-logger = init_logger(__name__)
-
-os.environ["MASTER_ADDR"] = "localhost"
-os.environ["MASTER_PORT"] = "29503"
-
-BASE_MODEL_PATH = "Wan-AI/Wan2.2-TI2V-5B-Diffusers"
-MODEL_PATH = maybe_download_model(BASE_MODEL_PATH)
-VAE_PATH = os.path.join(MODEL_PATH, "vae")
-
-
-@pytest.mark.skip(reason="disable test")
-@pytest.mark.usefixtures("distributed_setup")
-def test_official_wan_vae():
-    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-    precision = torch.bfloat16
-    precision_str = "bf16"
-    args = FastVideoArgs(model_path=VAE_PATH, pipeline_config=PipelineConfig(vae_config=WanVAEConfig(), vae_precision=precision_str))
-    args.device = device
-    args.vae_cpu_offload = False
-
-    loader = VAELoader()
-    model2 = loader.load(VAE_PATH, args)
-    assert model2.use_feature_cache # Default to use the original WanVAE algorithm
-
-    model1 = Wan2_2_VAE(vae_pth="/mnt/weka/home/hao.zhang/wei/Wan2.2_VAE.pth", device=device, dtype=precision)
-
-    # Create identical inputs for both models
-    batch_size = 1
-
-    # Video input [B, C, T, H, W]
-    input_tensor = torch.randn(batch_size,
-                               3,
-                               81,
-                               32,
-                               32,
-                               device=device,
-                               dtype=precision)
-    # latent_tensor = torch.randn(batch_size,
-    #                             16,
-    #                             21,
-    #                             32,
-    #                             32,
-    #                             device=device,
-    #                             dtype=precision)
-
-    # Disable gradients for inference
-    with torch.no_grad():
-        # Test encoding
-        logger.info("Testing encoding...")
-        latent1 = model1.encode([input_tensor.squeeze(0)])[0].unsqueeze(0)
-        print("--------------------------------")
-        latent2 = model2.encode(input_tensor)
-        # Check if latents have the same shape
-        assert latent1.shape == latent2.mean.shape, f"Latent shapes don't match: {latent1.mean.shape} vs {latent2.mean.shape}"
-        # Check if latents are similar
-        max_diff_encode = torch.max(torch.abs(latent1 - latent2.mean))
-        mean_diff_encode = torch.mean(torch.abs(latent1 - latent2.mean))
-        logger.info("Maximum difference between encoded latents: %s",
-                    max_diff_encode.item())
-        logger.info("Mean difference between encoded latents: %s",
-                    mean_diff_encode.item())
-        assert max_diff_encode < 1e-5, f"Encoded latents differ significantly: max diff = {mean_diff_encode.item()}"
-        # Test decoding
-        logger.info("Testing decoding...")
-        # latent1_tensor = latent1.mode()
-        # mean1 = (torch.tensor(model1.config.latents_mean).view(
-        #     1, model1.config.z_dim, 1, 1, 1).to(input_tensor.device,
-        #                                         input_tensor.dtype))
-        # std1 = (1.0 / torch.tensor(model1.config.latents_std).view(
-        #     1, model1.config.z_dim, 1, 1, 1)).to(input_tensor.device,
-        #                                         input_tensor.dtype)
-        # latent1_tensor = latent1_tensor / std1 + mean1
-        # output1 = model1.decode(latent1_tensor).sample
-
-        output1 = model1.decode([latent1.squeeze(0)])
-        output1 = output1[0].unsqueeze(0)
-
-        mean2 = model2.config.arch_config.shift_factor.to(input_tensor.device, input_tensor.dtype)
-        std2 = model2.config.arch_config.scaling_factor.to(input_tensor.device, input_tensor.dtype)
-        latent2_tensor = latent2.mode()
-        latent2_tensor = latent2_tensor / std2 + mean2
-        output2 = model2.decode(latent2_tensor)
-        # Check if outputs have the same shape
-        assert output1.shape == output2.shape, f"Output shapes don't match: {output1.shape} vs {output2.shape}"
-
-        # Check if outputs are similar
-        max_diff_decode = torch.max(torch.abs(output1.float() - output2.float()))
-        mean_diff_decode = torch.mean(torch.abs(output1.float() - output2.float()))
-        logger.info("Maximum difference between decoded outputs: %s",
-                    max_diff_decode.item())
-        logger.info("Mean difference between decoded outputs: %s",
-                    mean_diff_decode.item())
-        assert max_diff_decode < 1e-5, f"Decoded outputs differ significantly: max diff = {max_diff_decode.item()}"

fastvideo/train/callbacks/validation.py

@@ -404,6 +404,7 @@ class ValidationCallback(Callback):
                         art = self.tracker.video(
                             fname,
                             caption=cap,
+                            fps=sp.fps,
                         )
                         if art is not None:
                             video_logs.append(art)

fastvideo/training/distillation_pipeline.py

@@ -1180,7 +1180,7 @@ class DistillationPipeline(TrainingPipeline):
 
                     artifacts = []
                     for filename, caption in zip(video_filenames, all_captions, strict=True):
-                        video_artifact = self.tracker.video(filename, caption=caption)
+                        video_artifact = self.tracker.video(filename, caption=caption, fps=sampling_param.fps)
                         if video_artifact is not None:
                             artifacts.append(video_artifact)
                     if artifacts:

fastvideo/training/trackers.py

@@ -7,22 +7,28 @@ interface that can be used across all FastVideo training pipelines.
 
 from __future__ import annotations
 
+from collections.abc import Iterable, Iterator
 import contextlib
 import copy
+import math
 import os
 import pathlib
+import tempfile
 import time
 from dataclasses import dataclass
 from enum import Enum
 from typing import Any
-from collections.abc import Iterable, Iterator
 
+import numpy as np
 import torch
 
 from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
 
+_DEFAULT_VIDEO_FPS = 16
+_MISSING_ARTIFACT = object()
+
 
 def _sanitize_wandb_config(value: Any) -> Any:
     """Best-effort conversion of nested config objects to W&B-safe values."""
@@ -56,6 +62,140 @@ def _sanitize_wandb_config(value: Any) -> Any:
     return repr(value)
 
 
+def _prepare_video_array(data: Any) -> np.ndarray:
+    """Convert W&B-style TCHW/BTCHW video data into GIF-ready frames."""
+    if isinstance(data, torch.Tensor):
+        data = data.detach().cpu().numpy()
+
+    video = np.asarray(data)
+    if video.ndim == 4:
+        video = video.reshape(1, *video.shape)
+    elif video.ndim != 5:
+        raise ValueError("Video data must have shape [T, C, H, W] or [B, T, C, H, W]")
+
+    batch_size, num_frames, channels, height, width = video.shape
+    if batch_size == 0 or num_frames == 0:
+        raise ValueError("Video data must contain at least one batch item and one frame")
+    if channels not in (1, 3, 4):
+        raise ValueError(f"Video data must have 1, 3, or 4 channels; got {channels}")
+    if video.dtype != np.uint8:
+        logger.warning("Converting video data to uint8 for SwanLab")
+        video = video.astype(np.uint8)
+
+    # Match wandb.Video's batch tiling so the same input has a familiar layout
+    # in either tracker.
+    if batch_size & (batch_size - 1):
+        padded_batch_size = 1 << batch_size.bit_length()
+        padding = np.zeros(
+            (padded_batch_size - batch_size, num_frames, channels, height, width),
+            dtype=video.dtype,
+        )
+        video = np.concatenate((video, padding), axis=0)
+
+    num_rows = 1 << ((batch_size.bit_length() - 1) // 2)
+    num_columns = video.shape[0] // num_rows
+    video = video.reshape(num_rows, num_columns, num_frames, channels, height, width)
+    video = np.transpose(video, axes=(2, 0, 4, 1, 5, 3))
+    video = video.reshape(num_frames, num_rows * height, num_columns * width, channels)
+    if channels == 1:
+        video = video[..., 0]
+    return np.ascontiguousarray(video)
+
+
+def _read_video_file(file_path: str) -> tuple[list[np.ndarray], float | None]:
+    """Decode a video file and return its frames and encoded frame rate."""
+    import imageio.v2 as imageio
+
+    reader = imageio.get_reader(file_path)
+    try:
+        metadata = reader.get_meta_data() or {}
+        frames = [np.asarray(frame) for frame in reader]
+    finally:
+        reader.close()
+
+    if not frames:
+        raise ValueError(f"Video file contains no frames: {file_path}")
+
+    raw_source_fps = metadata.get("fps")
+    try:
+        source_fps = float(str(raw_source_fps))
+    except ValueError:
+        source_fps = None
+    if source_fps is not None and (not math.isfinite(source_fps) or source_fps <= 0):
+        source_fps = None
+    return frames, source_fps
+
+
+def _coerce_video_fps(fps: int | float | None) -> float:
+    if fps is None:
+        return _DEFAULT_VIDEO_FPS
+    value = float(fps)
+    if not math.isfinite(value) or value <= 0:
+        raise ValueError(f"Video fps must be a positive finite number; got {fps!r}")
+    return value
+
+
+def _write_gif(file_path: str, frames: Any, fps: float) -> None:
+    """Encode frames as an animated GIF for SwanLab's video API."""
+    from PIL import Image
+
+    images = []
+    for frame in frames:
+        array = np.asarray(frame)
+        if array.dtype != np.uint8:
+            array = array.astype(np.uint8)
+        if array.ndim == 3 and array.shape[-1] == 1:
+            array = array[..., 0]
+        if array.ndim not in (2, 3) or (array.ndim == 3 and array.shape[-1] not in (3, 4)):
+            raise ValueError(f"GIF frames must be grayscale, RGB, or RGBA; got shape {array.shape}")
+        images.append(Image.fromarray(array))
+
+    if not images:
+        raise ValueError("Video data must contain at least one frame")
+
+    duration_ms = max(1, round(1000 / fps))
+    images[0].save(
+        file_path,
+        format="GIF",
+        save_all=True,
+        append_images=images[1:],
+        duration=duration_ms,
+        loop=0,
+    )
+
+
+@dataclass(frozen=True)
+class _SequentialArtifact:
+    """Backend-specific versions of one artifact created by a tracker group."""
+
+    values: tuple[Any | None, ...]
+
+
+def _select_tracker_artifact(value: Any, tracker_index: int) -> Any:
+    """Resolve nested sequential artifacts for one child tracker."""
+    if isinstance(value, _SequentialArtifact):
+        selected = value.values[tracker_index]
+        return _MISSING_ARTIFACT if selected is None else selected
+    if isinstance(value, dict):
+        selected_dict = {}
+        for key, item in value.items():
+            selected = _select_tracker_artifact(item, tracker_index)
+            if selected is not _MISSING_ARTIFACT:
+                selected_dict[key] = selected
+        return selected_dict if selected_dict else _MISSING_ARTIFACT
+    if isinstance(value, list):
+        selected_list = [
+            selected for item in value
+            if (selected := _select_tracker_artifact(item, tracker_index)) is not _MISSING_ARTIFACT
+        ]
+        return selected_list if selected_list else _MISSING_ARTIFACT
+    if isinstance(value, tuple):
+        selected_tuple = tuple(selected for item in value
+                               if (selected := _select_tracker_artifact(item, tracker_index)) is not _MISSING_ARTIFACT)
+        return selected_tuple if selected_tuple else _MISSING_ARTIFACT
+    return value
+
+
 @dataclass
 class Timer:
     """Simple timer utility used by the trackers."""
@@ -222,8 +362,7 @@ class WandbTracker(BaseTracker):
         kwargs: dict[str, Any] = {}
         if caption is not None:
             kwargs["caption"] = caption
-        if fps is not None:
-            kwargs["fps"] = fps
+        kwargs["fps"] = fps if fps is not None else _DEFAULT_VIDEO_FPS
         if format is not None:
             kwargs["format"] = format
         else:
@@ -254,8 +393,10 @@ class SequentialTracker(BaseTracker):
         self._timed_metrics = {}
 
     def log_artifacts(self, artifacts: dict[str, Any], step: int) -> None:
-        for tracker in self._trackers:
-            tracker.log_artifacts(artifacts, step)
+        for tracker_index, tracker in enumerate(self._trackers):
+            tracker_artifacts = _select_tracker_artifact(artifacts, tracker_index)
+            if tracker_artifacts is not _MISSING_ARTIFACT:
+                tracker.log_artifacts(tracker_artifacts, step)
         self._timed_metrics = {}
 
     def log_file(
@@ -278,16 +419,91 @@ class SequentialTracker(BaseTracker):
         fps: int | None = None,
         format: str | None = None,
     ) -> Any | None:
-        for tracker in self._trackers:
-            video = tracker.video(data, caption=caption, fps=fps, format=format)
-            if video is not None:
-                return video
-        return None
+        videos = tuple(tracker.video(data, caption=caption, fps=fps, format=format) for tracker in self._trackers)
+        if all(video is None for video in videos):
+            return None
+        return _SequentialArtifact(videos)
+
+
+class SwanlabTracker(BaseTracker):
+    """Tracker implementation for SwanLab."""
+
+    def __init__(
+        self,
+        experiment_name: str,
+        log_dir: str,
+        *,
+        config: dict[str, Any] | None = None,
+        run_name: str | None = None,
+    ) -> None:
+        super().__init__()
+
+        try:
+            import swanlab
+        except ModuleNotFoundError as error:
+            if error.name != "swanlab":
+                raise
+            raise ModuleNotFoundError("SwanLab tracking requires the optional 'swanlab' dependency. "
+                                      "Install it with `uv pip install 'fastvideo[swanlab]'` (or "
+                                      "`uv pip install -e '.[swanlab]'` from a source checkout).") from error
+
+        pathlib.Path(log_dir).mkdir(parents=True, exist_ok=True)
+
+        self._swanlab = swanlab
+        self._run = swanlab.init(
+            project=experiment_name,
+            experiment_name=run_name,
+            config=(_sanitize_wandb_config(config) if config is not None else None),
+            logdir=log_dir,
+        )
+        logger.info("Initialized SwanLab tracker")
+
+    def log(self, metrics: dict[str, Any], step: int) -> None:
+        metrics = {**self._timed_metrics, **metrics}
+        if metrics:
+            self._swanlab.log(metrics, step=step)
+        self._timed_metrics = {}
+
+    def finish(self) -> None:
+        self._swanlab.finish()
+
+    def video(
+        self,
+        data: Any,
+        *,
+        caption: str | None = None,
+        fps: int | None = None,
+        format: str | None = None,
+    ) -> Any:
+        """Create a SwanLab GIF artifact from a file or W&B-style array."""
+        del format  # SwanLab currently supports GIF artifacts only.
+
+        if isinstance(data, str | os.PathLike):
+            source_path = os.fspath(data)
+            if pathlib.Path(source_path).suffix.lower() == ".gif":
+                return self._swanlab.Video(source_path, caption=caption)
+            frames, source_fps = _read_video_file(source_path)
+            if fps is not None:
+                video_fps = _coerce_video_fps(fps)
+            else:
+                video_fps = source_fps if source_fps is not None else _coerce_video_fps(None)
+        else:
+            frames = _prepare_video_array(data)
+            video_fps = _coerce_video_fps(fps)
+
+        file_descriptor, gif_path = tempfile.mkstemp(suffix=".gif")
+        os.close(file_descriptor)
+        try:
+            _write_gif(gif_path, frames, video_fps)
+            return self._swanlab.Video(gif_path, caption=caption)
+        finally:
+            pathlib.Path(gif_path).unlink(missing_ok=True)
 
 
 class Trackers(str, Enum):
     NONE = "none"
     WANDB = "wandb"
+    SWANLAB = "swanlab"
 
 
 SUPPORTED_TRACKERS = {tracker.value for tracker in Trackers}
@@ -324,6 +540,14 @@ def initialize_trackers(
                     config=config,
                     run_name=run_name,
                 ))
+        elif tracker_name == Trackers.SWANLAB.value:
+            tracker_instances.append(
+                SwanlabTracker(
+                    experiment_name,
+                    os.path.abspath(log_dir),
+                    config=config,
+                    run_name=run_name,
+                ))
 
     if not tracker_instances:
         return DummyTracker()

fastvideo/training/training_pipeline.py

@@ -831,7 +831,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
                 artifacts = []
                 for filename, caption in zip(video_filenames, all_captions, strict=True):
-                    video_artifact = self.tracker.video(filename, caption=caption)
+                    video_artifact = self.tracker.video(filename, caption=caption, fps=sampling_param.fps)
                     if video_artifact is not None:
                         artifacts.append(video_artifact)
                 if artifacts:

mkdocs.yml

@@ -169,6 +169,7 @@ nav:
   - Training:
     - Overview: training/overview.md
     - Training Infrastructure: training/train_infra.md
+    - Experiment Tracking: training/trackers.md
     - Data Preprocessing: training/data_preprocess.md
     - Fine-tuning: training/finetune.md
     - Examples:

pyproject.toml

@@ -114,12 +114,15 @@ fastvideo-kernel = [
 imagebind = { git = "https://github.com/facebookresearch/ImageBind.git", rev = "53680b02d7e37b19b124fa37bae4b6c98c38f5be" }
 # FA4 cute, pinned to a cutlass-4.5-compatible revision. torch.compile support
 # comes from FastVideo's own custom_op wrappers.
-flash-attn-4 = { git = "https://github.com/Dao-AILab/flash-attention.git", rev = "940cd9680f3315f2f06b43ab5bea2c2cf2d96806", subdirectory = "flash_attn/cute" }
+flash-attn-4 = { git = "https://github.com/Dao-AILab/flash-attention.git", rev = "82d6441eec5d4dfec120153db2c0145ae855a083", subdirectory = "flash_attn/cute" }
 
 [project.optional-dependencies]
 
 # flash-attn: uv pip install flash-attn==2.8.1 --no-cache-dir --no-build-isolation
 
+swanlab = [
+    "swanlab>=0.6.7",
+]
 
 lint = [
     "pre-commit==4.0.1",

scripts/checkpoint_conversion/dreamx_world_ar_to_diffusers.py

@@ -0,0 +1,124 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+"""Convert DreamX-World-5B autoregressive weights to FastVideo layout.
+
+The HF repository stores one raw official ``model.safetensors`` whose keys match
+FastVideo's native ``DreamXWorldARTransformer3DModel``. The converter writes a
+Diffusers-like root with ``transformer/config.json`` and reusable Wan2.2
+components. Use ``--symlink-transformer`` locally to avoid duplicating the 21GB
+AR tensor file.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import shutil
+from pathlib import Path
+
+TRANSFORMER_CONFIG: dict[str, object] = {
+    "_class_name": "DreamXWorldARTransformer3DModel",
+    "model_type": "ti2v",
+    "patch_size": [1, 2, 2],
+    "text_len": 512,
+    "num_attention_heads": 24,
+    "attention_head_dim": 128,
+    "in_channels": 48,
+    "out_channels": 48,
+    "text_dim": 4096,
+    "freq_dim": 256,
+    "ffn_dim": 14336,
+    "num_layers": 30,
+    "local_attn_size": 12,
+    "sink_size": 3,
+    "cross_attn_norm": True,
+    "qk_norm": True,
+    "eps": 1e-6,
+    "add_control_adapter": True,
+    "cam_method": "prope",
+    "attn_compress": 4,
+    "cam_self_attn_layers": list(range(30)),
+    "num_frames_per_block": 3,
+}
+
+MODEL_INDEX: dict[str, object] = {
+    "_class_name": "DreamXWorldARPipeline",
+    "_diffusers_version": "0.31.0",
+    "scheduler": ["diffusers", "FlowMatchEulerDiscreteScheduler"],
+    "text_encoder": ["transformers", "UMT5EncoderModel"],
+    "tokenizer": ["transformers", "AutoTokenizer"],
+    "transformer": ["diffusers", "DreamXWorldARTransformer3DModel"],
+    "vae": ["diffusers", "AutoencoderKLWan"],
+}
+
+REUSED_COMPONENTS = ("scheduler", "text_encoder", "tokenizer", "vae")
+
+
+def _source_safetensors(source: Path) -> Path:
+    if source.is_file():
+        return source
+    path = source / "model.safetensors"
+    if not path.exists():
+        raise FileNotFoundError(f"Missing AR model.safetensors under {source}")
+    return path
+
+
+def convert_transformer(source: Path, output: Path, symlink_transformer: bool) -> None:
+    src = _source_safetensors(source)
+    transformer_dir = output / "transformer"
+    transformer_dir.mkdir(parents=True, exist_ok=True)
+    (transformer_dir / "config.json").write_text(json.dumps(TRANSFORMER_CONFIG, indent=2) + "\n")
+    dst = transformer_dir / "model.safetensors"
+    if dst.is_symlink() and not dst.exists():
+        dst.unlink()  # dangling symlink: replace so re-conversion self-heals
+    if dst.exists():
+        return
+    if symlink_transformer:
+        dst.symlink_to(src.resolve())
+    else:
+        shutil.copy2(src, dst)
+
+
+def _copy_or_link_component(component: str, component_source: Path, output: Path, symlink: bool) -> None:
+    src = component_source / component
+    dst = output / component
+    if not src.exists():
+        raise FileNotFoundError(f"Missing reused component source: {src}")
+    if dst.is_symlink() and not dst.exists():
+        dst.unlink()  # dangling symlink: replace so re-conversion self-heals
+    if dst.exists():
+        return
+    if symlink:
+        dst.symlink_to(src.resolve(), target_is_directory=src.is_dir())
+    elif src.is_dir():
+        shutil.copytree(src, dst)
+    else:
+        shutil.copy2(src, dst)
+
+
+def write_model_index(output: Path, component_source: Path | None, symlink_components: bool) -> None:
+    if component_source is not None:
+        for component in REUSED_COMPONENTS:
+            _copy_or_link_component(component, component_source, output, symlink_components)
+    missing = [component for component in REUSED_COMPONENTS if not (output / component).exists()]
+    if missing:
+        print("skipping model_index.json; missing reused components: " + ", ".join(missing))
+        return
+    (output / "model_index.json").write_text(json.dumps(MODEL_INDEX, indent=2) + "\n")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--source", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--component-source", type=Path)
+    parser.add_argument("--symlink-components", action="store_true")
+    parser.add_argument("--symlink-transformer", action="store_true")
+    args = parser.parse_args()
+
+    convert_transformer(args.source, args.output, args.symlink_transformer)
+    write_model_index(args.output, args.component_source, args.symlink_components)
+
+
+if __name__ == "__main__":
+    main()

scripts/checkpoint_conversion/dreamx_world_to_diffusers.py

@@ -0,0 +1,221 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+"""Convert DreamX-World-5B-Cam raw transformer weights to FastVideo-loadable format.
+
+The GD-ML/DreamX-World-5B-Cam repository stores the transformer as raw
+DreamX/Wan official shards. FastVideo's TransformerLoader expects a Diffusers-like
+transformer folder with a config.json and safetensors whose keys can be mapped by
+WanVideoConfig.param_names_mapping. This script performs the raw official ->
+Diffusers-like key rename and writes the DreamX 5B-Cam transformer config.
+
+Example:
+    python scripts/checkpoint_conversion/dreamx_world_to_diffusers.py \
+        --source official_weights/dreamx_world \
+        --output converted_weights/dreamx_world
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import shutil
+from collections import OrderedDict
+from pathlib import Path
+
+import torch
+from huggingface_hub import save_torch_state_dict
+from safetensors import safe_open
+from safetensors.torch import load_file
+
+
+OFFICIAL_TO_DIFFUSERS_MAPPING: dict[str, str] = {
+    r"^text_embedding\.0\.(.*)$": r"condition_embedder.text_embedder.linear_1.\1",
+    r"^text_embedding\.2\.(.*)$": r"condition_embedder.text_embedder.linear_2.\1",
+    r"^time_embedding\.0\.(.*)$": r"condition_embedder.time_embedder.linear_1.\1",
+    r"^time_embedding\.2\.(.*)$": r"condition_embedder.time_embedder.linear_2.\1",
+    r"^time_projection\.1\.(.*)$": r"condition_embedder.time_proj.\1",
+    r"^img_emb\.proj\.0\.(.*)$": r"condition_embedder.image_embedder.norm1.\1",
+    r"^img_emb\.proj\.1\.(.*)$": r"condition_embedder.image_embedder.ff.net.0.proj.\1",
+    r"^img_emb\.proj\.3\.(.*)$": r"condition_embedder.image_embedder.ff.net.2.\1",
+    r"^img_emb\.proj\.4\.(.*)$": r"condition_embedder.image_embedder.norm2.\1",
+    r"^head\.modulation": r"scale_shift_table",
+    r"^head\.head\.(.*)$": r"proj_out.\1",
+    r"^blocks\.(\d+)\.self_attn\.q\.(.*)$": r"blocks.\1.attn1.to_q.\2",
+    r"^blocks\.(\d+)\.self_attn\.k\.(.*)$": r"blocks.\1.attn1.to_k.\2",
+    r"^blocks\.(\d+)\.self_attn\.v\.(.*)$": r"blocks.\1.attn1.to_v.\2",
+    r"^blocks\.(\d+)\.self_attn\.o\.(.*)$": r"blocks.\1.attn1.to_out.0.\2",
+    r"^blocks\.(\d+)\.self_attn\.norm_q\.(.*)$": r"blocks.\1.attn1.norm_q.\2",
+    r"^blocks\.(\d+)\.self_attn\.norm_k\.(.*)$": r"blocks.\1.attn1.norm_k.\2",
+    r"^blocks\.(\d+)\.cross_attn\.q\.(.*)$": r"blocks.\1.attn2.to_q.\2",
+    r"^blocks\.(\d+)\.cross_attn\.k\.(.*)$": r"blocks.\1.attn2.to_k.\2",
+    r"^blocks\.(\d+)\.cross_attn\.k_img\.(.*)$": r"blocks.\1.attn2.add_k_proj.\2",
+    r"^blocks\.(\d+)\.cross_attn\.v\.(.*)$": r"blocks.\1.attn2.to_v.\2",
+    r"^blocks\.(\d+)\.cross_attn\.v_img\.(.*)$": r"blocks.\1.attn2.add_v_proj.\2",
+    r"^blocks\.(\d+)\.cross_attn\.o\.(.*)$": r"blocks.\1.attn2.to_out.0.\2",
+    r"^blocks\.(\d+)\.cross_attn\.norm_q\.(.*)$": r"blocks.\1.attn2.norm_q.\2",
+    r"^blocks\.(\d+)\.cross_attn\.norm_k\.(.*)$": r"blocks.\1.attn2.norm_k.\2",
+    r"^blocks\.(\d+)\.cross_attn\.norm_k_img\.(.*)$": r"blocks.\1.attn2.norm_added_k.\2",
+    r"^blocks\.(\d+)\.ffn\.0\.(.*)$": r"blocks.\1.ffn.net.0.proj.\2",
+    r"^blocks\.(\d+)\.ffn\.2\.(.*)$": r"blocks.\1.ffn.net.2.\2",
+    r"^blocks\.(\d+)\.modulation": r"blocks.\1.scale_shift_table",
+    r"^blocks\.(\d+)\.norm3\.(.*)$": r"blocks.\1.norm2.\2",
+}
+
+TRANSFORMER_CONFIG: dict[str, object] = {
+    "_class_name": "DreamXWorldTransformer3DModel",
+    "patch_size": [1, 2, 2],
+    "text_len": 512,
+    "num_attention_heads": 24,
+    "attention_head_dim": 128,
+    "in_channels": 48,
+    "out_channels": 48,
+    "text_dim": 4096,
+    "freq_dim": 256,
+    "ffn_dim": 14336,
+    "num_layers": 30,
+    "cross_attn_norm": True,
+    "qk_norm": "rms_norm_across_heads",
+    "eps": 1e-6,
+    "image_dim": None,
+    "added_kv_proj_dim": None,
+    "rope_max_seq_len": 1024,
+    "add_control_adapter": True,
+    "cam_method": "prope",
+    "attn_compress": 1,
+    "cam_self_attn_layers": None,
+}
+
+MODEL_INDEX: dict[str, object] = {
+    "_class_name": "DreamXWorldPipeline",
+    "_diffusers_version": "0.31.0",
+    "scheduler": ["diffusers", "FlowMatchEulerDiscreteScheduler"],
+    "text_encoder": ["transformers", "UMT5EncoderModel"],
+    "tokenizer": ["transformers", "AutoTokenizer"],
+    "transformer": ["diffusers", "DreamXWorldTransformer3DModel"],
+    "vae": ["diffusers", "AutoencoderKLWan"],
+}
+
+REUSED_COMPONENTS = ("scheduler", "text_encoder", "tokenizer", "vae")
+
+
+def map_transformer_key(key: str) -> str:
+    for pattern, replacement in OFFICIAL_TO_DIFFUSERS_MAPPING.items():
+        if re.match(pattern, key):
+            return re.sub(pattern, replacement, key)
+    return key
+
+
+def _safetensor_files(source: Path) -> list[Path]:
+    if source.is_file():
+        if source.suffix != ".safetensors":
+            raise ValueError(f"Only .safetensors files are supported, got {source}")
+        return [source]
+
+    index_path = source / "diffusion_pytorch_model.safetensors.index.json"
+    if index_path.exists():
+        index = json.loads(index_path.read_text())
+        return sorted({source / shard for shard in index["weight_map"].values()})
+
+    files = sorted(source.glob("*.safetensors"))
+    if not files:
+        raise FileNotFoundError(f"No safetensors files found under {source}")
+    return files
+
+
+def convert_transformer(source: Path, output: Path, max_shard_size: str) -> None:
+    transformer_dir = output / "transformer"
+    transformer_dir.mkdir(parents=True, exist_ok=True)
+    (transformer_dir / "config.json").write_text(json.dumps(TRANSFORMER_CONFIG, indent=2) + "\n")
+
+    converted: OrderedDict[str, torch.Tensor] = OrderedDict()
+    for shard in _safetensor_files(source):
+        print(f"loading {shard}")
+        for key, tensor in load_file(shard, device="cpu").items():
+            new_key = map_transformer_key(key)
+            if new_key in converted:
+                raise ValueError(f"Duplicate converted key: {new_key}")
+            converted[new_key] = tensor
+
+    print(f"saving {len(converted)} tensors to {transformer_dir}")
+    save_torch_state_dict(converted, transformer_dir, max_shard_size=max_shard_size)
+
+
+def _copy_or_link_component(component: str, component_source: Path, output: Path, symlink: bool) -> None:
+    src = component_source / component
+    dst = output / component
+    if not src.exists():
+        raise FileNotFoundError(f"Missing reused component source: {src}")
+    if dst.is_symlink() and not dst.exists():
+        dst.unlink()  # dangling symlink: replace so re-conversion self-heals
+    if dst.exists():
+        print(f"keeping existing {dst}")
+        return
+    if symlink:
+        dst.symlink_to(src.resolve(), target_is_directory=src.is_dir())
+        print(f"linked {dst} -> {src}")
+    elif src.is_dir():
+        shutil.copytree(src, dst)
+        print(f"copied {src} -> {dst}")
+    else:
+        shutil.copy2(src, dst)
+        print(f"copied {src} -> {dst}")
+
+
+def write_model_index(output: Path, component_source: Path | None, symlink_components: bool) -> None:
+    if component_source is not None:
+        for component in REUSED_COMPONENTS:
+            _copy_or_link_component(component, component_source, output, symlink_components)
+    missing = [component for component in REUSED_COMPONENTS if not (output / component).exists()]
+    if missing:
+        print("skipping model_index.json; missing reused components: " + ", ".join(missing))
+        print("pass --component-source <Wan2.2 Diffusers root> to copy or link reused components")
+        return
+    (output / "model_index.json").write_text(json.dumps(MODEL_INDEX, indent=2) + "\n")
+    print(f"wrote {output / 'model_index.json'}")
+
+
+def analyze(source: Path) -> None:
+    total = 0
+    unchanged = 0
+    examples: list[tuple[str, str]] = []
+    for shard in _safetensor_files(source):
+        with safe_open(shard, framework="pt", device="cpu") as tensors:
+            for key in tensors:
+                total += 1
+                new_key = map_transformer_key(key)
+                unchanged += int(new_key == key)
+                if len(examples) < 20 and new_key != key:
+                    examples.append((key, new_key))
+    print(f"total_keys={total} unchanged_keys={unchanged}")
+    for old, new in examples:
+        print(f"{old} -> {new}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--source", type=Path, required=True, help="DreamX raw transformer directory or safetensors file")
+    parser.add_argument("--output", type=Path, required=True, help="Output model root; transformer/ is created inside it")
+    parser.add_argument("--max-shard-size", default="10GB")
+    parser.add_argument(
+        "--component-source",
+        type=Path,
+        help="Optional Wan2.2 Diffusers root whose scheduler/text_encoder/tokenizer/vae components are reused.",
+    )
+    parser.add_argument(
+        "--symlink-components",
+        action="store_true",
+        help="Symlink reused components from --component-source instead of copying them.",
+    )
+    parser.add_argument("--analyze", action="store_true", help="Only print key mapping summary")
+    args = parser.parse_args()
+
+    if args.analyze:
+        analyze(args.source)
+    else:
+        convert_transformer(args.source, args.output, args.max_shard_size)
+        write_model_index(args.output, args.component_source, args.symlink_components)
+
+
+if __name__ == "__main__":
+    main()

tests/local_tests/dreamx_world/PORT_STATUS.md

@@ -0,0 +1,189 @@
+# DreamX World Port Status
+
+## Summary
+
+- model_family: `dreamx_world`
+- workload_types: `I2V camera-control compatibility shim`; `I2V autoregressive camera-control forcing`
+- official_ref: `https://github.com/AMAP-ML/DreamX-World`
+- official_ref_dir: `DreamX-World/`
+- hf_weights_path: `GD-ML/DreamX-World-5B-Cam`
+- local_weights_dir: `official_weights/dreamx_world`
+- source_layout: `raw_official`
+- local_tests_readme: `tests/local_tests/dreamx_world/README.md`
+
+## Current Phase
+
+- phase: `phase_11_post_parity_handoff`
+- status: `complete`
+- owner: `orchestrator`
+- last_updated: `2026-07-02`
+
+## Component Matrix
+
+| Component | Type | Reuse/Port | Official Definition | Official Instantiation | FastVideo Target | Prototype | Conversion | Parity | Open Issues |
+|---|---|---|---|---|---|---|---|---|---|
+| transformer | dit | ported_dedicated | `DreamX-World/models/wan_transformer3d.py`; PRoPE helpers in `DreamX-World/models/prope_utils.py` | `DreamX-World/inference_dreamx5b.py::setup_models`, `Wan2_2Transformer3DModel.from_pretrained(... cam_method=prope, add_control_adapter=True)` | `fastvideo/models/dits/dreamx_world.py`; `fastvideo/configs/models/dits/dreamx_world.py`; DreamX pipeline config helper | native_prope_pass | real_conversion_pass | strict_load_and_forward_parity_pass | none |
+| vae | vae | reuse_pending | `DreamX-World/models/wan_vae3_8.py` | `DreamX-World/inference_dreamx5b.py::setup_models`, `AutoencoderKLWan3_8.from_pretrained(Wan2.2_VAE.pth)` | `fastvideo/models/vaes/wanvae.py`; DreamX VAE config helper | config_smoke_pass | raw_key_mapping_pass | encode_parity_pass | none |
+| text_encoder/tokenizer | encoder | reuse_pending | `DreamX-World/models/wan_text_encoder.py`; tokenizer via Wan2.2 base model | `DreamX-World/inference_dreamx5b.py::setup_models`, `WanT5EncoderModel` + tokenizer subpaths | `fastvideo/models/encoders/t5.py::UMT5EncoderModel`; DreamX UMT5 config helper | config_smoke_pass | staged_weight_load_pass | hidden_state_parity_pass | none |
+| scheduler | generic | reuse_proven | Diffusers `FlowMatchEulerDiscreteScheduler` | `DreamX-World/inference_dreamx5b.py::setup_models`, default `sampler_name=Flow` | `fastvideo/models/schedulers/scheduling_flow_match_euler_discrete.py` | pass | not_required | non_skip_pass | Q003 |
+| camera_conditioning | generic | port_pending | `DreamX-World/utils/inference_utils.py`, `DreamX-World/models/prope_utils.py`, `DreamX-World/wan/modules/camera_prope.py` | `DreamX-World/inference_dreamx5b.py::get_camera_sequence`, `pipeline(... control_camera_video=...)` | `fastvideo/pipelines/basic/dreamx_world/camera_conditioning.py` | pass | not_required | non_skip_pass | none |
+| pipeline | pipeline | port_complete | `DreamX-World/pipeline/pipeline_dreamxworld.py` | `DreamX-World/inference_dreamx5b.py::process_inference_from_json` | `fastvideo/pipelines/basic/dreamx_world/` plus config/preset/registry | pipeline_load_generate_smoke_pass | model_index_and_config_consistency_smoke_pass | pipeline_api_vs_worker_forward_parity_pass | none |
+| ar_transformer | dit | port_complete | `DreamX-World/wan/modules/causal_camera_model_2_2_prope_infinity.py::CausalWanModel` | `DreamX-World/inference_ar_forcing.py::load_pipeline` | `fastvideo/models/dits/dreamx_world_ar.py`; `fastvideo/configs/models/dits/dreamx_world.py::DreamXWorldARConfig` | tiny_official_forward_parity_pass | identity_conversion_pass | real_5b_strict_load_pass | none |
+| ar_pipeline | pipeline | port_complete | `DreamX-World/pipeline/pipeline_causal_camera.py` | `DreamX-World/inference_ar_forcing.py::main` | `fastvideo/pipelines/basic/dreamx_world/dreamx_world_ar_pipeline.py`; `fastvideo/pipelines/basic/dreamx_world/ar_denoising.py`; registry/preset/config | config_registry_pass | symlink_model_index_pass | short_full_generation_pass | none |
+
+## Conversion State
+
+- conversion_script: `scripts/checkpoint_conversion/dreamx_world_to_diffusers.py`
+- converted_weights_dir: `converted_weights/dreamx_world`
+- source_layout: `raw_official`
+- strict_load_status: `pass`
+- conversion_script_status: `transformer_model_index_and_config_consistency_smoke_pass`
+- model_index_status: `smoke_pass`
+- passthrough_components: `Wan2.2 Diffusers scheduler, tokenizer, and text encoder are symlinked from official_weights/Wan2.2-TI2V-5B-Diffusers; VAE parity uses raw Wan2.2_VAE.pth with an explicit DreamX raw-to-FastVideo key mapper because official encode returns normalized latents.`
+- retry_history: `none`
+
+## Parity Commands
+
+| Scope | Command | Last Result | Notes |
+|---|---|---|---|
+| transformer | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_transformer_parity.py -v -s` | strict_load_and_forward_parity_pass | 2026-07-01: converted real 5B-Cam transformer shards strict-load into dedicated `DreamXWorldTransformer3DModel` with 0 shape mismatches; official-vs-FastVideo small-input fp32 forward parity passes on CUDA (`diff_max=0.072533`, `diff_mean=0.008014`). |
+| vae | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_vae_parity.py -v -s` | encode_parity_pass | 2026-06-30: official DreamX raw `Wan2.2_VAE.pth` maps 196/196 keys into FastVideo Wan VAE and encode parity passes after applying the same official latent normalization (`(mu - mean) / std`). |
+| text_encoder | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_text_encoder_parity.py -v -s` | hidden_state_parity_pass | 2026-06-30: official `WanT5EncoderModel` vs FastVideo `UMT5EncoderModel` hidden-state parity passes on CUDA using staged Wan2.2 text encoder/tokenizer weights and reference-only `xfuser` stubs. |
+| scheduler | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_scheduler_parity.py -v -s` | non_skip_pass | 2026-06-30: FastVideo FlowMatch scheduler matches official Diffusers timesteps and step output for DreamX default Flow sampler; `DreamXWorldPipeline` initializes FlowMatch with official `shift=3.0`. |
+| camera_conditioning | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_camera_conditioning_parity.py -v -s` | non_skip_pass | 2026-06-29: 3 parameterized cases passed against official reference on CPU. |
+| pipeline_config | `python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py -v -s` | pipeline_entry_preset_scheduler_modelinfo_and_camera_stage_smoke_pass | DreamX 5B-Cam PipelineConfig wires DiT/VAE/UMT5/Flow/TI2V settings and official `shift=3.0`; default preset is registered for `GD-ML/DreamX-World-5B-Cam`; local converted-style `model_index.json` resolves to `DreamXWorldPipeline`; the pipeline initializes FlowMatch, camera conditioning writes `batch.extra["dreamx_y_camera"]`, and generic denoising can pass it as `y_camera`. |
+| ar_transformer | `PYTHONPATH=/workspace/FastVideo python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_ar_conversion.py tests/local_tests/dreamx_world/test_dreamx_world_ar_transformer_parity.py -q -rs` | 6_passed_0_skipped | 2026-07-02: AR converter writes symlinked transformer layout/model_index; tiny official `CausalWanModel` vs FastVideo `DreamXWorldARTransformer3DModel` forward parity passes; real 5B `model.safetensors` strict-loads with zero missing/unexpected keys from `/tmp/converted_dreamx_world_ar`. |
+| ar_pipeline_config | `PYTHONPATH=/workspace/FastVideo python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py -q -rs` | 10_passed_0_skipped | 2026-07-02: `DreamXWorld5BARPipelineConfig`, `DreamXWorldARPipeline`, registry config selection for `GD-ML/DreamX-World-5B`, and `dreamx_world_5b_ar` preset pass. |
+| ar_full_generation | `PYTHONPATH=/workspace/FastVideo python /tmp/run_dreamx_ar_video_smoke.py` | generated_video_pass | 2026-07-02: A40 short full-generation smoke passed from `/tmp/converted_dreamx_world_ar` with 64x64, 9 frames, 4 denoise steps, `output_type=pil`, `save_video=True`; MP4 saved at `outputs_video/dreamx_world_ar_video_smoke/a quiet road through a futuristic coastal city at sunrise.mp4` and decoded as 9 frames of `(64, 64, 3)` uint8. |
+| ar_long_horizon | `PYTHONPATH=/workspace/FastVideo FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA python /tmp/run_dreamx_ar_long_horizon.py` | generated_video_pass | 2026-07-02: A40 long-horizon AR generation passed from `/tmp/converted_dreamx_world_ar` with 64x64, 1005 frames, 4 denoise steps, seed 4096; MP4 saved at `outputs_video/dreamx_world_ar_long_horizon/A long autonomous drive through a futuristic coastal city at sunrise, smooth forward camera motion,.mp4` and decoded as 1005 frames of `(64, 64, 3)`; end-to-end generation latency was 231.68s after load. |
+| ar_ssim_default | `PYTHONPATH=/workspace/FastVideo FASTVIDEO_SSIM_MODEL_ID=DreamX-World-5B DREAMX_WORLD_AR_SSIM_MODEL_PATH=/tmp/converted_dreamx_world_ar python -m pytest fastvideo/tests/ssim/test_dreamx_world_similarity.py -q -rs --skip-ssim-reference-download` | 1_passed_0_skipped | 2026-07-02: A40 default AR SSIM reference seeded locally at `fastvideo/tests/ssim/reference_videos/default/A40_reference_videos/DreamX-World-5B/TORCH_SDPA/`; helper removes stale generated base MP4 before generation; default params are 192x192, 81 frames, 4 steps, seed 2048, min SSIM 0.98. |
+| ar_ssim_modal_l40s | `modal run /tmp/modal_dreamx_ar_ssim_git.py` | 1_passed_0_skipped | 2026-07-02: Modal L40S run checked out `post-fix dreamx-world-5b-cam branch commit`, downloaded default `L40S_reference_videos` via `reference_videos_cli.py download`, used cached converted AR weights under `/root/data/dreamx_world_ar_converted`, seeded the missing AR L40S reference from generated output, reran a fresh generated-vs-reference compare successfully (`mean_ssim=1.0`), and exported the reference to Modal volume `hf-model-weights:dreamx_ar_ssim_l40s`. Downloaded local reference decodes as 81 frames of `(192, 192, 3)`. A40-vs-L40S reference spot check after deterministic AR noise fix: mean SSIM 0.7770, min 0.6139, max 0.9944. |
+| pipeline_smoke | `python -m pytest tests/local_tests/pipelines/test_dreamx_world_pipeline_smoke.py tests/local_tests/pipelines/test_dreamx_world_pipeline_parity.py -q -rs` | 4_passed_0_skipped | 2026-06-30: combined smoke/parity passed. 2026-07-01: smoke alone passed with real `image_path` TI2V coverage (`3 passed`), validating image load, TI2V preprocessing, VAE first-frame encode under CPU offload, camera conditioning, and 1-step latent generation from `converted_weights/dreamx_world`. Tests force `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA` to avoid the local FlashAttention-4 cute ABI mismatch. |
+| basic_example | `PYTHONPATH=/workspace/FastVideo FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA DREAMX_WORLD_MODEL_DIR=converted_weights/dreamx_world DREAMX_WORLD_IMAGE_PATH= DREAMX_WORLD_HEIGHT=64 DREAMX_WORLD_WIDTH=64 DREAMX_WORLD_NUM_FRAMES=9 DREAMX_WORLD_STEPS=1 DREAMX_WORLD_GUIDANCE=1.0 DREAMX_WORLD_OUTPUT_PATH=outputs_video/dreamx_world_example_smoke python examples/inference/basic/basic_dreamx_world.py` | generated_video_pass | 2026-06-30: example saved an MP4 under `outputs_video/dreamx_world_example_smoke`; imageio/ffmpeg decoded frame 0 as `(64, 64, 3)` uint8, fps 16, duration 0.56s. |
+
+## Open Questions
+
+| ID | Question | Owner | Needed By Phase | Status | Resolution |
+|---|---|---|---|---|---|
+| Q001 | Should first PR expose only the `DreamX-World-5B-Cam` 5s camera-control mode and exclude AR long-horizon forcing? | user | prep | resolved | User approved starting with `DreamX-World-5B-Cam`; AR long-horizon is out of first-PR scope. |
+| Q002 | Does FastVideo's existing Wan2.2 TI2V transformer support DreamX PRoPE/control adapter with a small extension, or is a DreamX-specific DiT required? | component:transformer | Phase 3 | resolved | Project guidance prefers a separate DreamX DiT for maintainability. DreamX PRoPE/control adapter now lives in `fastvideo/models/dits/dreamx_world.py`; Wan DiT/config have no DreamX-specific fields or `y_camera` signature. |
+| Q003 | Which sampler is in first-PR scope: official default `Flow` only, or also `Flow_Unipc` and `Flow_DPM++`? | orchestrator | Phase 3 | resolved | First PR should support official default `Flow` only. FastVideo FlowMatch scheduler parity is non-skip PASS; optional `Flow_Unipc` and `Flow_DPM++` are out of first-PR scope. |
+| Q004 | Which HF token env var should be used if rate limits or gated Wan2.2 base weights require auth? | user | Phase 5 | resolved | No auth was required for the completed local downloads; keep using env var names only if future gated repos require auth. |
+| Q005 | Should native FastVideo production code depend on DreamX reference-only packages such as `xfuser` or OpenCV? | user | Phase 3 | resolved | No. These packages may be used only for official reference/local parity setup; native FastVideo integration must remove that runtime requirement. |
+| Q006 | Should AR handoff require a full generated long-horizon video in this no-HF-token/no-GPU-budget pass? | user/runtime | pipeline | resolved | A40 long-horizon generation passes with 1005 frames at 64x64/4 steps. AR default SSIM passes locally on A40 and on Modal L40S with a 192x192/81-frame reference. HF upload/publication remains a separate operation if the reference dataset should be updated upstream. |
+| Q007 | Can A40 references stand in for L40S CI references? | quality | release | resolved | No. After deterministic AR noise fix, A40-vs-L40S reference mean SSIM is 0.7770, below the 0.98 same-device threshold. Publish the L40S-specific reference for CI. |
+
+## Issues And Blockers
+
+| ID | Phase | Component | Severity | Issue | Evidence | Owner | Status | Resolution |
+|---|---|---|---|---|---|---|---|---|
+| I001 | prep | official_env | medium | Official import initially failed because `xfuser` was missing. | `ModuleNotFoundError: No module named 'xfuser'` from `python -c "import sys; sys.path.insert(0, 'DreamX-World'); import inference_dreamx5b"` | prep | resolved | Installed `xfuser==0.4.1`; import progressed. |
+| I002 | prep | official_env | medium | Official import then failed because GUI OpenCV required missing system `libxcb.so.1`. | `ImportError: libxcb.so.1: cannot open shared object file` through `cv2` import in Diffusers ConsisID path. | prep | resolved | Installed `opencv-python-headless`; `import inference_dreamx5b` passed. |
+| I003 | prep | weights | medium | HF repo has raw official transformer shards and no Diffusers `model_index.json`. | `inspect_hf_layout.py GD-ML/DreamX-World-5B-Cam --json` returned `source_layout=raw_official`, `needs_conversion=yes`, `model_index_class=null`. | conversion | resolved | Downloaded raw DreamX shards to `official_weights/dreamx_world`; converted transformer to `converted_weights/dreamx_world/transformer`; symlinked reusable Wan2.2 Diffusers components; real 5B transformer strict-load passes. |
+| I004 | prep | dependencies | high | Official reference import required extra packages in the local environment, but FastVideo native runtime should not inherit those dependencies. | `xfuser==0.4.1` and `opencv-python-headless` were installed only to make `DreamX-World/inference_dreamx5b.py` import for reference/parity. | pipeline | resolved | Production DreamX FastVideo code uses native camera/image/video utilities and has no runtime `xfuser` or OpenCV import requirement; those packages remain reference-only local parity dependencies. |
+| I005 | parity | transformer | medium | Transformer full forward parity initially failed in bf16 official harness. | Official CUDA bf16 LayerNorm path was unstable; fp32 small-input harness avoids that dtype issue and compares against FastVideo with single-process SP identity patches. | component:transformer | resolved | Official-vs-FastVideo forward parity now passes on CUDA with `diff_max=0.072533`, `diff_mean=0.008014`. |
+| I006 | parity | vae/text_encoder | medium | VAE/text parity initially remained skipped after weights were staged. | Text official import needed a reference-only `xfuser` stub; VAE comparison initially used raw official normalized latents against FastVideo raw mu. | component:vae,component:text_encoder | resolved | Text hidden-state parity passes. VAE encode parity passes after raw key mapping and applying the official latent normalization to FastVideo output. |
+| I007 | quality | pipeline_ti2v | medium | Real image-path TI2V smoke initially failed when `vae_cpu_offload=True` because DenoisingStage encoded the first frame while the VAE weights remained on CPU. | DreamX SSIM first run failed with `RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same` at `fastvideo/pipelines/stages/denoising.py` VAE encode. | pipeline | resolved | DenoisingStage now moves the VAE to `local_device` before TI2V first-frame encode; image-path pipeline smoke and DreamX SSIM both pass. |
+| I008 | quality | ssim_helper | high | SSIM helper could compare against a stale generated base MP4 when a rerun saved the new video as `_1.mp4`. | Existing generated outputs made AR reference seeding appear to pass before a fresh generated-vs-reference compare. | quality | resolved | `run_text_to_video_similarity_test` and `run_image_to_video_similarity_test` now remove the stale generated base MP4 before generation. A40 and Modal L40S AR SSIM were rerun after the fix. |
+| I009 | quality | ar_denoising | high | AR denoising added CUDA noise without using the request seed when the original generator was CPU-backed. | Fresh reruns against old AR references produced mean SSIM near 0.05. | pipeline | resolved | `DreamXWorldARCausalDenoisingStage` now derives a device-local generator from the request seed for AR noise. Fresh same-device A40 and L40S reruns pass with mean SSIM 1.0 after reseeding references. |
+
+## Escape Hatches
+
+| ID | Phase | Decision Type | Question | Recommended Option | Status | Resolution |
+|---|---|---|---|---|---|---|
+
+## Decisions
+
+| Date | Decision | Rationale | Impact |
+|---|---|---|---|
+| 2026-06-29 | First PR scope is `DreamX-World-5B-Cam` only. | Cam mode is closest to existing Wan2.2 TI2V support; AR forcing needs separate causal/KV pipeline work. | Component inventory and parity focus on `inference_dreamx5b.py` and `pipeline_dreamxworld.py`. |
+| 2026-06-29 | Do not install full DreamX requirements during prep. | Full requirements pin core FastVideo stack packages. | Installed only `xfuser==0.4.1` and `opencv-python-headless` to make official imports work. |
+| 2026-06-29 | Treat HF DreamX-World-5B-Cam weights as raw official transformer layout requiring conversion. | HF inspection found no `model_index.json`. | Phase 5 must create `scripts/checkpoint_conversion/dreamx_world_to_diffusers.py` after component prototype/key dumps. |
+| 2026-06-29 | Do not add DreamX reference-only dependencies to FastVideo production requirements. | The current environment should remain the FastVideo environment; extra packages are only for official reference parity. | Native DreamX integration must avoid runtime `xfuser` and OpenCV requirements unless explicitly approved later. |
+| 2026-06-29 | Implement DreamX camera conditioning as native FastVideo utility. | It is weightless and removes the need to import DreamX reference utilities at production runtime. | `fastvideo/pipelines/basic/dreamx_world/camera_conditioning.py` now has non-skip parity against official action-to-PRoPE tensors. |
+| 2026-06-29 | First PR supports DreamX default `Flow` sampler only. | FastVideo FlowMatch Euler scheduler matches the official Diffusers scheduler for DreamX defaults. | Pipeline work can use FastVideo native FlowMatch scheduler; UniPC and DPM++ are out of first-PR scope. |
+| 2026-07-01 | Keep DreamX PRoPE/control adapter in a dedicated DreamX DiT class. | Project guidance is that putting too much DreamX behavior into Wan makes the model hard to manage. | `fastvideo/models/dits/dreamx_world.py` defines `DreamXWorldTransformer3DModel`, `DreamXWorldTransformerBlock`, and `DreamXPropeSelfAttention`; `fastvideo/configs/models/dits/dreamx_world.py` owns DreamX adapter config fields; Wan DiT/config are unchanged from DreamX. |
+| 2026-06-30 | Camera parity test loads official camera functions by file instead of importing the official `utils` package. | Official package initialization pulls unrelated dependencies that can require GUI OpenCV system libraries. | Camera parity remains non-skip without adding DreamX reference-only dependencies to FastVideo production requirements. |
+| 2026-06-30 | Add DreamX-World-5B-Cam model and pipeline config helpers plus a conversion script. | Official HF DreamX 5B-Cam transformer config is 30 layers, hidden size 3072, 24 heads, 48 latent channels, plus Wan2.2 48-channel VAE and UMT5-XXL text encoder. | DreamX helpers wire DiT/VAE/UMT5/Flow/TI2V settings; `dreamx_world_to_diffusers.py` writes a FastVideo-loadable transformer config plus renamed safetensors; strict-load smoke passes on a tiny official DreamX transformer and the real 5B converted shards. |
+| 2026-06-30 | Pass DreamX camera PRoPE condition through the FastVideo batch/denoising path. | DreamX transformer expects `y_camera={"viewmats", "K"}` at denoising time. | `DreamXWorldPipeline` is registered as a basic pipeline entry and initializes the official default FlowMatch scheduler; `dreamx_world_5b_cam` preset mirrors official 5B-Cam defaults; `DreamXWorldCameraConditioningStage` writes `batch.extra["dreamx_y_camera"]`; generic denoising filters and forwards it as `y_camera` only for compatible transformers. |
+
+## Handoff Notes
+
+- Prep, component parity, pipeline smoke/parity, and the basic example validation are complete for `DreamX-World-5B-Cam`.
+- Official reference clone is staged at `DreamX-World/` and ignored by git.
+- Workspace-local weights are staged: DreamX raw transformer shards under `official_weights/dreamx_world`, Wan2.2 raw base artifacts under `official_weights/Wan2.2-TI2V-5B`, and Wan2.2 Diffusers reusable components under `official_weights/Wan2.2-TI2V-5B-Diffusers`.
+- Camera conditioning parity is active and passing without weights.
+- Default Flow scheduler parity is active and passing without weights.
+- Transformer has corrected official 5B-Cam architecture in dedicated DreamX DiT/config files, native PRoPE/control-adapter, conversion mapping, real converted 5B strict-load, and official-vs-FastVideo forward parity passing on CUDA. VAE encode parity and text hidden-state parity pass on CUDA. Pipeline entry/registry, local model_info resolution, preset, config, FlowMatch scheduler init, camera stage, denoising `y_camera` kwarg smokes, independent CUDA pipeline smoke/parity, and a small saved-video basic example pass.
+- Full local DreamX component suite is non-skip PASS: `python -m pytest tests/local_tests/dreamx_world/ -q -rs` returned `26 passed` on 2026-07-01. Pipeline smoke/parity and SSIM quality regression are also non-skip PASS locally.
+- Keep `xfuser` and OpenCV as reference-only parity dependencies. Do not add them
+  to FastVideo requirements or production imports.
+
+
+## Quality Regression
+
+- status: `added`
+- test: `fastvideo/tests/ssim/test_dreamx_world_similarity.py`
+- command: `PYTHONPATH=/workspace/FastVideo FASTVIDEO_SSIM_MODEL_ID=DreamX-World-5B-Cam python -m pytest fastvideo/tests/ssim/test_dreamx_world_similarity.py -q -rs`
+- result: `1 passed, 0 skipped` on 2026-07-01
+- reference: Local A40/TORCH_SDPA reference seeded from the generated candidate under `fastvideo/tests/ssim/reference_videos/default/A40_reference_videos/DreamX-World-5B-Cam/TORCH_SDPA/`. The test uses a deterministic generated input image, 64x64 request dimensions, 9 frames, 1 denoise step, seed 1024, and min SSIM 0.98. Full-quality params are present for 480x832/161 frames/30 steps.
+- note: Modal L40S seeding passed using the configured Modal profile and unauthenticated HF public downloads. HF upload/publication still requires `HF_API_KEY`, `HUGGINGFACE_HUB_TOKEN`, or `HF_TOKEN` with write access; no token values were used or recorded.
+
+## Final Handoff
+
+```text
+final_handoff:
+  prep_handoff_complete: yes
+  conversion_status: pass
+  components:
+    - name: transformer
+      reuse_or_port: ported_dedicated_dit
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_transformer_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: vae
+      reuse_or_port: reused
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_vae_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: text_encoder_tokenizer
+      reuse_or_port: reused
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_text_encoder_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: scheduler
+      reuse_or_port: reused
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_scheduler_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: camera_conditioning
+      reuse_or_port: ported
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_camera_conditioning_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: ar_transformer
+      reuse_or_port: ported
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_ar_transformer_parity.py
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+    - name: ar_pipeline
+      reuse_or_port: ported
+      parity_test: tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py plus AR smoke/SSIM commands listed above
+      parity_status: non_skip_pass
+      concerns_or_unknowns: none
+  pipeline_smoke: pass
+  pipeline_parity: pass
+  example_status: pass
+  quality_regression: added
+  local_tests_readme: tests/local_tests/dreamx_world/README.md
+  port_state_file: tests/local_tests/dreamx_world/PORT_STATUS.md
+  token_values_committed: no
+  runtime_third_party_model_imports: none
+  blockers: none
+  escape_hatch: none
+```
+
+| 2026-07-02 | Add DreamX-World-5B autoregressive support. | Official AR repo is raw single-safetensors layout and needs a dedicated causal/KV stage. | Added native AR DiT/config, identity converter, AR pipeline config/preset/registry, and targeted non-skip tests. Raw AR weights are staged at `/tmp/dreamx_world_ar_weights`; converted symlink layout at `/tmp/converted_dreamx_world_ar`. |
+| 2026-07-02 | Validate DreamX-World-5B AR short full generation on A40. | Targeted parity/config tests prove components, but end-to-end runtime can still fail at scheduler, RoPE cache, device, or decode boundaries. | `DreamXWorldARPipeline` generated and saved a 64x64/9-frame/4-step MP4 from `/tmp/converted_dreamx_world_ar`; the saved MP4 decodes to 9 frames. |
+| 2026-07-02 | Validate DreamX-World-5B AR long-horizon and default SSIM on A40. | AR needs coverage beyond the 9-frame smoke to exercise longer KV/cache progression and a quality regression path. | 1005-frame 64x64/4-step generation passes and decodes; default AR SSIM test uses 192x192/81 frames because MS-SSIM requires short side >160. |
+| 2026-07-02 | Validate DreamX-World-5B AR default SSIM on Modal L40S. | CI references are device-specific; A40 alone is not enough for L40S reference coverage. | Modal checked out `post-fix dreamx-world-5b-cam branch commit`, downloaded existing default L40S references, seeded the missing AR reference, reran SSIM successfully (`mean_ssim=1.0`), and exported the L40S reference back to the workspace. |

tests/local_tests/dreamx_world/README.md

@@ -0,0 +1,254 @@
+# DreamX World Local Tests
+
+Local-only parity and smoke tests for the `dreamx_world` FastVideo port. These
+tests compare FastVideo against the official DreamX-World reference
+implementation and are not expected to run in CI unless explicitly promoted
+later.
+
+Port progress, open questions, issues, and handoff notes live in
+`tests/local_tests/dreamx_world/PORT_STATUS.md`.
+
+## Reference Assets
+
+| Field | Value |
+|---|---|
+| Model family | `dreamx_world` |
+| First-PR scope | `DreamX-World-5B-Cam`; follow-up scope now includes `DreamX-World-5B` autoregressive forcing |
+| Out-of-scope variants | none for the DreamX-World 5B/Cam paths currently ported |
+| Workload types | I2V camera-control compatibility shim: image + prompt + action sequence to video |
+| Official reference | `https://github.com/AMAP-ML/DreamX-World` |
+| Local reference dir | `DreamX-World/` |
+| Official commit/version | `221875811ba31f7eac6c3025b215c09ad2cefd1d` |
+| HF weights | `GD-ML/DreamX-World-5B-Cam` |
+| HF revision | default |
+| Local weights dir | `official_weights/dreamx_world` |
+| Source layout | `raw_official` |
+| Needs conversion | `yes` |
+
+Do not write token values in this file. Current token env var detected during
+prep: `none`.
+
+## Shared Environment Setup
+
+Run from the FastVideo repo root in the same conda/env used for FastVideo. Do
+not create a separate upstream environment for parity tests.
+
+```bash
+python ".agents/skills/add-model-01-prep/scripts/clone_reference_repo.py" \
+    "https://github.com/AMAP-ML/DreamX-World.git" \
+    "DreamX-World" \
+    --commit "221875811ba31f7eac6c3025b215c09ad2cefd1d" \
+    --update-gitignore
+```
+
+DreamX-World does not expose a packaging file for editable install. During prep
+the official import check used `sys.path.insert(0, "DreamX-World")`.
+
+Additional official deps installed into the current environment for imports:
+
+```bash
+uv pip install xfuser==0.4.1
+uv pip install opencv-python-headless
+```
+
+These packages are for running the official DreamX reference during local
+parity only. They must not become FastVideo production/runtime dependencies for
+the native `dreamx_world` pipeline.
+
+Do not install the full `DreamX-World/requirements.txt` without explicit
+approval. It pins core FastVideo stack packages including `torch`, `torchvision`,
+`triton`, `flash_attn`, and `diffusers`.
+
+## Official Environment Status
+
+```text
+dependency_changes: installed official deps in current env
+official_env_status: imports_ok
+private_dep_stubs: none
+blocked_on: none
+```
+
+Import check used during prep:
+
+```bash
+python -c "import sys; sys.path.insert(0, 'DreamX-World'); import inference_dreamx5b; print('imports_ok')"
+```
+
+## Weight Setup
+
+HF layout inspection found no root `model_index.json`; the repo contains
+`config.json`, a safetensors index, and three transformer safetensors shards.
+This is a raw official transformer layout and requires conversion before
+FastVideo can load it through `VideoGenerator.from_pretrained`.
+
+```bash
+python ".agents/skills/add-model-01-prep/scripts/inspect_hf_layout.py" \
+    "GD-ML/DreamX-World-5B-Cam" \
+    --json
+```
+
+Weights have been staged workspace-locally. The raw DreamX transformer repo lives at `official_weights/dreamx_world`; Wan2.2 raw base artifacts live at `official_weights/Wan2.2-TI2V-5B`; Wan2.2 Diffusers reusable components live at `official_weights/Wan2.2-TI2V-5B-Diffusers`. To reproduce the DreamX download:
+
+```bash
+python ".agents/skills/add-model-01-prep/scripts/download_hf_weights.py" \
+    "GD-ML/DreamX-World-5B-Cam" \
+    "official_weights/dreamx_world"
+```
+
+
+
+### DreamX-World-5B Autoregressive Setup
+
+The AR repository `GD-ML/DreamX-World-5B` is also raw official layout: no
+`model_index.json`, root `config.json`, and a single `model.safetensors`. The
+current environment has the raw AR checkpoint staged outside the workspace at
+`/tmp/dreamx_world_ar_weights` to avoid workspace quota pressure. The converted
+FastVideo layout is staged at `/tmp/converted_dreamx_world_ar` with the 21GB
+transformer safetensors symlinked instead of copied.
+
+```bash
+python ".agents/skills/add-model-01-prep/scripts/download_hf_weights.py" \
+    "GD-ML/DreamX-World-5B" \
+    "/tmp/dreamx_world_ar_weights"
+
+python scripts/checkpoint_conversion/dreamx_world_ar_to_diffusers.py \
+    --source /tmp/dreamx_world_ar_weights \
+    --output /tmp/converted_dreamx_world_ar \
+    --component-source official_weights/Wan2.2-TI2V-5B-Diffusers \
+    --symlink-components \
+    --symlink-transformer
+```
+
+AR production code uses `DreamXWorldARTransformer3DModel`,
+`DreamXWorld5BARPipelineConfig`, `DreamXWorldARPipeline`, and
+`DreamXWorldARCausalDenoisingStage`. The AR DiT is a native FastVideo port of
+the official Apache-2.0 `CausalWanModel`; it has no production DreamX, Diffusers
+model-class, Transformers model-class, `xfuser`, or OpenCV import.
+
+## Prototype And Conversion Artifacts
+
+State-dict key/shape dumps are generated after FastVideo native prototypes exist
+and are used to build the conversion mapping.
+
+```text
+official_key_dumps:
+  transformer: converted_weights/dreamx_world/_mapping/transformer_official_keys.json
+fastvideo_key_dumps:
+  transformer: converted_weights/dreamx_world/_mapping/transformer_fastvideo_keys.json
+conversion_script: scripts/checkpoint_conversion/dreamx_world_to_diffusers.py
+conversion_script_status: transformer_model_index_and_config_consistency_smoke_pass
+conversion_source_layout: raw_official
+converted_weights_dir: converted_weights/dreamx_world
+model_index_status: smoke_pass
+strict_load_status: pass
+```
+
+The converter writes `transformer/` from raw DreamX shards. To create a full
+FastVideo-loadable diffusers-style root, pass a Wan2.2 Diffusers directory as the
+component source so reusable components are copied or symlinked before
+`model_index.json` is emitted:
+
+```bash
+python scripts/checkpoint_conversion/dreamx_world_to_diffusers.py \
+    --source official_weights/dreamx_world \
+    --output converted_weights/dreamx_world \
+    --component-source /path/to/Wan2.2-TI2V-5B-Diffusers \
+    --symlink-components
+```
+
+## Expected Parity Tests
+
+Planned local tests for this family:
+
+| Component | Official files / args | Test | Concerns | Status |
+|---|---|---|---|---|
+| transformer | `DreamX-World/models/wan_transformer3d.py`; instantiated in `DreamX-World/inference_dreamx5b.py` with `cam_method=prope`, `add_control_adapter=True` | `tests/local_tests/dreamx_world/test_dreamx_world_transformer_parity.py` | FastVideo uses dedicated `DreamXWorldTransformer3DModel`/`DreamXWorldConfig` files for DreamX 5B-Cam config, PRoPE, conversion mapping, real converted 5B strict-load PASS, and official-vs-FastVideo small-input forward parity PASS on CUDA. Wan DiT/config have no DreamX-specific adapter fields. | strict_load_and_forward_parity_pass |
+| vae | `DreamX-World/models/wan_vae3_8.py`; `vae_type=AutoencoderKLWan3_8`, `vae_subpath=Wan2.2_VAE.pth` | `tests/local_tests/dreamx_world/test_dreamx_world_vae_parity.py` | DreamX raw `Wan2.2_VAE.pth` maps 196/196 keys into FastVideo Wan VAE; encode parity passes after applying official latent normalization. | encode_parity_pass |
+| text_encoder/tokenizer | `DreamX-World/models/wan_text_encoder.py`; T5 path from Wan2.2 base model | `tests/local_tests/dreamx_world/test_dreamx_world_text_encoder_parity.py` | Official `WanT5EncoderModel` vs FastVideo `UMT5EncoderModel` hidden-state parity passes on CUDA with staged Wan2.2 weights/tokenizer. | hidden_state_parity_pass |
+| scheduler | Diffusers `FlowMatchEulerDiscreteScheduler`; selected by default `sampler_name=Flow` | `tests/local_tests/dreamx_world/test_dreamx_world_scheduler_parity.py` | First PR can support official default `Flow`; optional `Flow_Unipc` and `Flow_DPM++` are out of first-PR scope. | non_skip_pass |
+| camera_conditioning | `DreamX-World/utils/inference_utils.py`, `models/prope_utils.py`, `wan/modules/camera_prope.py` | `tests/local_tests/dreamx_world/test_dreamx_world_camera_conditioning_parity.py` | Action sequence to PRoPE/control input must match official tensor shapes and values. | non_skip_pass |
+| pipeline | `DreamX-World/pipeline/pipeline_dreamxworld.py`; call path in `DreamX-World/inference_dreamx5b.py` | `tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py`; `tests/local_tests/pipelines/test_dreamx_world_pipeline_smoke.py`; `tests/local_tests/pipelines/test_dreamx_world_pipeline_parity.py` | DreamX PipelineConfig wires first-scope DiT/VAE/UMT5/Flow/TI2V settings, official `shift=3.0`, default preset values, FlowMatch scheduler initialization, and local `model_index.json` resolution; independent pipeline smoke covers real CUDA local load + latent generation, and parity compares public API output to worker-side explicit ForwardBatch execution. | pipeline_smoke_and_parity_pass |
+| ar_transformer | `DreamX-World/wan/modules/causal_camera_model_2_2_prope_infinity.py`; instantiated by `DreamX-World/inference_ar_forcing.py` as `CausalWanModel` with `local_attn_size=12`, `sink_size=3`, `attn_compress=4` | `tests/local_tests/dreamx_world/test_dreamx_world_ar_transformer_parity.py` | Native `DreamXWorldARTransformer3DModel` keeps official identity key layout; tiny official-vs-FastVideo forward parity passes; real 5B AR safetensors strict-load passes from `/tmp/converted_dreamx_world_ar`. | tiny_forward_parity_and_real_strict_load_pass |
+| ar_pipeline | `DreamX-World/pipeline/pipeline_causal_camera.py`; AR block/KV/context-noise loop | `tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py`; A40 short full-generation smoke | Dedicated `DreamXWorldARCausalDenoisingStage` implements blockwise KV forcing; raw HF repo must be converted before `VideoGenerator.from_pretrained` because it has no `model_index.json`; converted AR layout generated a 64x64/9-frame/4-step MP4 on A40. | config_registry_and_short_full_generation_pass |
+
+Include reused components in parity. Reuse is accepted only after the FastVideo
+component definition and official instantiation arguments have both been checked
+and the component parity test passes non-skip.
+
+Run the relevant tests with:
+
+```bash
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_camera_conditioning_parity.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_scheduler_parity.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_transformer_parity.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_vae_parity.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_text_encoder_parity.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py -v -s
+python -m pytest tests/local_tests/dreamx_world/test_dreamx_world_conversion.py -v -s
+python -m pytest tests/local_tests/pipelines/test_dreamx_world_pipeline_smoke.py tests/local_tests/pipelines/test_dreamx_world_pipeline_parity.py -q -rs
+```
+
+## Current Local Results
+
+```bash
+python -m pytest tests/local_tests/dreamx_world/ -v -s
+# 2026-07-01: 26 passed, 0 skipped
+# 2026-07-02: AR targeted suite passed: 16 passed, 0 skipped
+
+PYTHONPATH=/workspace/FastVideo python /tmp/run_dreamx_ar_video_smoke.py
+# 2026-07-02: DreamX-World-5B AR short full-generation smoke passed on A40
+# output: outputs_video/dreamx_world_ar_video_smoke/a quiet road through a futuristic coastal city at sunrise.mp4
+# decoded: 9 frames, (64, 64, 3), uint8
+
+PYTHONPATH=/workspace/FastVideo FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA python /tmp/run_dreamx_ar_long_horizon.py
+# 2026-07-02: DreamX-World-5B AR long-horizon generation passed on A40
+# output: outputs_video/dreamx_world_ar_long_horizon/A long autonomous drive through a futuristic coastal city at sunrise, smooth forward camera motion,.mp4
+# decoded: 1005 frames, (64, 64, 3)
+
+PYTHONPATH=/workspace/FastVideo FASTVIDEO_SSIM_MODEL_ID=DreamX-World-5B DREAMX_WORLD_AR_SSIM_MODEL_PATH=/tmp/converted_dreamx_world_ar python -m pytest fastvideo/tests/ssim/test_dreamx_world_similarity.py -q -rs --skip-ssim-reference-download
+# 2026-07-02: DreamX-World-5B AR default SSIM passed: 1 passed, 0 skipped
+# local A40 reference: fastvideo/tests/ssim/reference_videos/default/A40_reference_videos/DreamX-World-5B/TORCH_SDPA/
+
+modal run /tmp/modal_dreamx_ar_ssim_git.py
+# 2026-07-02: Modal L40S default SSIM passed: 1 passed, 0 skipped
+# checked out post-fix dreamx-world-5b-cam branch commit
+# first downloaded default L40S references with reference_videos_cli.py download
+# seeded missing AR L40S reference, then reran a fresh generated-vs-reference compare
+# Modal JSON: mean_ssim=1.0, min_ssim=1.0, max_ssim=1.0
+# local L40S reference: fastvideo/tests/ssim/reference_videos/default/L40S_reference_videos/DreamX-World-5B/TORCH_SDPA/
+# A40-vs-L40S reference spot check after deterministic AR noise fix: mean SSIM 0.7770, min 0.6139, max 0.9944
+
+python -m pytest tests/local_tests/pipelines/test_dreamx_world_pipeline_smoke.py tests/local_tests/pipelines/test_dreamx_world_pipeline_parity.py -q -rs
+# 2026-06-30: 4 passed, 0 skipped
+# 2026-07-01: smoke image-path TI2V coverage passed separately with 3 passed, 0 skipped
+
+PYTHONPATH=/workspace/FastVideo FASTVIDEO_SSIM_MODEL_ID=DreamX-World-5B-Cam python -m pytest fastvideo/tests/ssim/test_dreamx_world_similarity.py -q -rs
+# 2026-07-01: 1 passed, 0 skipped
+```
+
+`camera_conditioning`, default `Flow` scheduler, DreamX component/pipeline configs, default preset, pipeline entry/registry, FlowMatch scheduler initialization, DreamX camera stage, denoising `y_camera` pass-through, conversion model-index/config-consistency checks, real converted 5B dedicated DreamX transformer strict-load and forward parity on CUDA, VAE encode parity on CUDA, text hidden-state parity on CUDA, native DreamX PRoPE branch structure smoke, AR transformer parity/strict-load, AR config/registry, and AR short full-generation smoke are non-skip PASS results. The full local DreamX component suite, independent pipeline smoke/parity suite, image-path TI2V smoke, and SSIM quality regression currently have zero skips. The basic 5B-Cam example was run against `converted_weights/dreamx_world` with a 64x64/9-frame/1-step saved-video smoke, and imageio decoded the generated MP4 successfully; the AR pipeline separately saved and decoded a 64x64/9-frame/4-step MP4 from `/tmp/converted_dreamx_world_ar`.
+
+## Review Notes
+
+- Required before handoff: non-skip PASS for each required component parity
+  test, including reused components that own weights or numerical behavior.
+- First PR scope originally targeted `DreamX-World-5B-Cam`; scope was later
+  expanded to include `DreamX-World-5B` AR support with a separate causal/KV
+  pipeline.
+- AR support has targeted parity/config coverage, short full-generation smoke,
+  1005-frame A40 long-horizon generation, local A40 default SSIM coverage,
+  and Modal L40S default SSIM coverage. L40S validation was rerun after fixing
+  stale generated-output comparison in the SSIM helper and deterministic AR
+  noise seeding. HF reference publication remains a separate token-gated
+  release operation.
+- FastVideo production code must not require `xfuser` or OpenCV just because the
+  official reference import needed them. Port camera/action preprocessing and
+  sequence-parallel behavior into existing FastVideo-native utilities or keep
+  reference-only imports inside local parity tests.
+- Review agents should verify setup commands still match the PR, then run the
+  listed parity tests or report the exact blocker.
+
+
+## Quality Regression
+
+Quality regression is added in `fastvideo/tests/ssim/test_dreamx_world_similarity.py`. The 5B-Cam default test uses the workspace converted model root, `TORCH_SDPA`, a deterministic generated conditioning image, 9 frames, 1 denoise step, seed 1024, and min SSIM 0.98. A local A40 reference was seeded under `fastvideo/tests/ssim/reference_videos/default/A40_reference_videos/DreamX-World-5B-Cam/TORCH_SDPA/`, and the test passed non-skip locally on 2026-07-01. The AR default test uses `/tmp/converted_dreamx_world_ar` or `/root/data/dreamx_world_ar_converted`, `TORCH_SDPA`, 192x192, 81 frames, 4 steps, seed 2048, and min SSIM 0.98; local A40 and Modal L40S references were seeded under `fastvideo/tests/ssim/reference_videos/default/{A40,L40S}_reference_videos/DreamX-World-5B/TORCH_SDPA/`, and both tests passed non-skip on 2026-07-02 after fresh generated-output cleanup was added to the SSIM helper. Modal L40S validation checked out `post-fix dreamx-world-5b-cam branch commit`, seeded the missing AR L40S reference, reran the test, and wrote `mean_ssim=1.0`. A cross-device A40-vs-L40S reference spot check produced mean SSIM 0.7770, so CI should use the L40S-specific reference rather than the A40 artifact. Full-quality params are present for 5B-Cam 480x832/161 frames/30 steps and AR 704x1280/1005 frames/4 steps; publishing references to the HF dataset remains a release operation requiring a write-capable HF token env var, never a raw token value.

tests/local_tests/dreamx_world/__init__.py

@@ -0,0 +1 @@
+# SPDX-License-Identifier: Apache-2.0

tests/local_tests/dreamx_world/test_dreamx_world_ar_conversion.py

@@ -0,0 +1,58 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World-5B autoregressive conversion smoke tests."""
+
+import json
+
+from fastvideo.configs.pipelines.dreamx_world import make_dreamx_world_5b_ar_dit_config
+from fastvideo.models.registry import _LEGACY_FAST_VIDEO_MODELS
+from scripts.checkpoint_conversion.dreamx_world_ar_to_diffusers import (
+    MODEL_INDEX,
+    REUSED_COMPONENTS,
+    TRANSFORMER_CONFIG,
+    convert_transformer,
+    write_model_index,
+)
+
+
+def test_dreamx_world_ar_converter_writes_symlinked_transformer_and_model_index(tmp_path):
+    source = tmp_path / "raw"
+    source.mkdir()
+    raw_tensor = source / "model.safetensors"
+    raw_tensor.write_bytes(b"placeholder")
+    component_source = tmp_path / "wan22"
+    output = tmp_path / "dreamx_ar"
+
+    for component in REUSED_COMPONENTS:
+        component_dir = component_source / component
+        component_dir.mkdir(parents=True)
+        (component_dir / "config.json").write_text("{}\n")
+
+    convert_transformer(source, output, symlink_transformer=True)
+    write_model_index(output, component_source, symlink_components=True)
+
+    assert (output / "transformer" / "model.safetensors").is_symlink()
+    model_index = json.loads((output / "model_index.json").read_text())
+    assert model_index == MODEL_INDEX
+    assert model_index["_class_name"] == "DreamXWorldARPipeline"
+    assert model_index["transformer"] == ["diffusers", "DreamXWorldARTransformer3DModel"]
+
+
+def test_dreamx_world_ar_transformer_config_matches_pipeline_dit_config():
+    dit_config = make_dreamx_world_5b_ar_dit_config()
+    assert TRANSFORMER_CONFIG["_class_name"] == "DreamXWorldARTransformer3DModel"
+    assert TRANSFORMER_CONFIG["num_attention_heads"] == dit_config.num_attention_heads
+    assert TRANSFORMER_CONFIG["attention_head_dim"] == dit_config.attention_head_dim
+    assert TRANSFORMER_CONFIG["in_channels"] == dit_config.in_channels
+    assert TRANSFORMER_CONFIG["out_channels"] == dit_config.out_channels
+    assert TRANSFORMER_CONFIG["ffn_dim"] == dit_config.ffn_dim
+    assert TRANSFORMER_CONFIG["num_layers"] == dit_config.num_layers
+    assert TRANSFORMER_CONFIG["local_attn_size"] == dit_config.local_attn_size
+    assert TRANSFORMER_CONFIG["sink_size"] == dit_config.sink_size
+    assert TRANSFORMER_CONFIG["attn_compress"] == dit_config.attn_compress
+    assert tuple(TRANSFORMER_CONFIG["cam_self_attn_layers"]) == dit_config.cam_self_attn_layers
+
+
+def test_dreamx_world_ar_model_index_component_classes_are_registered():
+    for component in ("scheduler", "text_encoder", "transformer", "vae"):
+        class_name = MODEL_INDEX[component][1]
+        assert class_name in _LEGACY_FAST_VIDEO_MODELS

tests/local_tests/dreamx_world/test_dreamx_world_ar_transformer_parity.py

@@ -0,0 +1,178 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World-5B autoregressive transformer parity.
+
+Coverage scope: both. The tiny forward parity compares FastVideo's native AR DiT
+against the official DreamX ``CausalWanModel`` implementation with identical
+weights. The real 5B checkpoint gate strict-loads the downloaded safetensors
+through the FastVideo model class.
+"""
+from __future__ import annotations
+
+import os
+from pathlib import Path
+import sys
+
+import pytest
+import torch
+from torch.testing import assert_close
+
+from fastvideo.configs.models.dits.dreamx_world import DreamXWorldARArchConfig, DreamXWorldARConfig
+from fastvideo.configs.pipelines.dreamx_world import make_dreamx_world_5b_ar_dit_config
+from fastvideo.models.dits.dreamx_world_ar import DreamXWorldARTransformer3DModel
+from fastvideo.models.loader.fsdp_load import load_model_from_full_model_state_dict
+from fastvideo.models.loader.utils import get_param_names_mapping
+from fastvideo.models.loader.weight_utils import resolve_safetensors_files, safetensors_weights_iterator
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+OFFICIAL_REF_DIR = Path(os.getenv("DREAMX_WORLD_OFFICIAL_REF_DIR", REPO_ROOT / "DreamX-World"))
+CONVERTED_AR_DIR = Path(os.getenv("DREAMX_WORLD_AR_CONVERTED_DIR", "/tmp/converted_dreamx_world_ar"))
+PARITY_SCOPE = "both"
+
+
+def _tiny_config() -> DreamXWorldARConfig:
+    return DreamXWorldARConfig(
+        arch_config=DreamXWorldARArchConfig(
+            num_attention_heads=1,
+            attention_head_dim=8,
+            in_channels=4,
+            out_channels=4,
+            ffn_dim=16,
+            num_layers=1,
+            text_dim=8,
+            freq_dim=8,
+            text_len=4,
+            local_attn_size=2,
+            sink_size=1,
+            attn_compress=1,
+            cam_self_attn_layers=(0,),
+        ))
+
+
+def _load_official_tiny():
+    if not OFFICIAL_REF_DIR.exists():
+        pytest.skip(f"Official DreamX reference missing: {OFFICIAL_REF_DIR}")
+    sys.path.insert(0, str(OFFICIAL_REF_DIR))
+    try:
+        from wan.modules import attention as official_attention
+        from wan.modules import causal_camera_model_2_2_prope_infinity as causal_module
+        from wan.modules import model_2_2 as official_model_2_2
+        from wan.modules.causal_camera_model_2_2_prope_infinity import CausalWanModel
+    except Exception as exc:  # noqa: BLE001
+        pytest.skip(f"Cannot import official AR transformer: {exc}")
+    official_attention.FLASH_ATTN_2_AVAILABLE = False
+    official_attention.FLASH_ATTN_3_AVAILABLE = False
+
+    def _sdpa_same_dtype(q, k, v, **kwargs):
+        del kwargs
+        out = torch.nn.functional.scaled_dot_product_attention(
+            q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), dropout_p=0.0)
+        return out.transpose(1, 2).contiguous()
+
+    official_attention.attention = _sdpa_same_dtype
+    official_attention.flash_attention = _sdpa_same_dtype
+    official_model_2_2.flash_attention = _sdpa_same_dtype
+    causal_module.attention = _sdpa_same_dtype
+    return CausalWanModel(
+        model_type="ti2v",
+        patch_size=(1, 2, 2),
+        text_len=4,
+        in_dim=4,
+        dim=8,
+        ffn_dim=16,
+        freq_dim=8,
+        text_dim=8,
+        out_dim=4,
+        num_heads=1,
+        num_layers=1,
+        local_attn_size=2,
+        sink_size=1,
+        qk_norm=True,
+        cross_attn_norm=True,
+        add_control_adapter=True,
+        cam_method="prope",
+        attn_compress=1,
+        cam_self_attn_layers=(0,),
+    ).eval()
+
+
+def _make_inputs():
+    torch.manual_seed(123)
+    x = [torch.randn(4, 1, 4, 4)]
+    t = torch.zeros(1, 4, dtype=torch.long)
+    context = [torch.randn(2, 8)]
+    camera = {
+        "viewmats": torch.eye(4).reshape(1, 1, 4, 4).repeat(1, 4, 1, 1),
+        "K": torch.eye(3).reshape(1, 1, 3, 3).repeat(1, 4, 1, 1),
+    }
+    kv_cache = [{
+        "k": torch.zeros(1, 8, 1, 8),
+        "v": torch.zeros(1, 8, 1, 8),
+        "global_end_index": torch.tensor([0]),
+        "local_end_index": torch.tensor([0]),
+        "prope_k": torch.zeros(1, 8, 1, 8),
+        "prope_v": torch.zeros(1, 8, 1, 8),
+        "prope_global_end_index": torch.tensor([0]),
+        "prope_local_end_index": torch.tensor([0]),
+    }]
+    cross_cache = [{
+        "k": torch.zeros(1, 4, 1, 8),
+        "v": torch.zeros(1, 4, 1, 8),
+        "is_init": False,
+    }]
+    return x, t, context, camera, kv_cache, cross_cache
+
+
+def test_dreamx_world_ar_tiny_forward_matches_official():
+    official = _load_official_tiny()
+    # The official init_weights zero-inits the output head (head.head.weight and
+    # biases), so both models would output exactly zero and the comparison would
+    # pass vacuously. Randomize the head (deterministically) before copying the
+    # state dict so the outputs reflect the internal computation.
+    generator = torch.Generator().manual_seed(7)
+    with torch.no_grad():
+        official.head.head.weight.normal_(std=0.5, generator=generator)
+        official.head.head.bias.normal_(std=0.5, generator=generator)
+    fastvideo = DreamXWorldARTransformer3DModel(_tiny_config(), {}).eval()
+    fastvideo.load_state_dict(official.state_dict(), strict=True)
+
+    x, t, context, camera, kv_cache, cross_cache = _make_inputs()
+    official_out = official(x=x, t=t, context=context, seq_len=4, y_camera=camera,
+                            kv_cache=kv_cache, crossattn_cache=cross_cache).detach()
+    x, t, context, camera, kv_cache, cross_cache = _make_inputs()
+    fastvideo_out = fastvideo(x=x, t=t, context=context, seq_len=4, y_camera=camera,
+                              kv_cache=kv_cache, crossattn_cache=cross_cache).detach()
+    assert official_out.abs().max() > 0, "official output is all-zero; parity comparison is vacuous"
+    assert_close(fastvideo_out, official_out, atol=1e-5, rtol=1e-5)
+
+
+def test_dreamx_world_ar_5b_config_matches_official_shape():
+    config = make_dreamx_world_5b_ar_dit_config()
+    assert config.num_layers == 30
+    assert config.num_attention_heads == 24
+    assert config.attention_head_dim == 128
+    assert config.hidden_size == 3072
+    assert config.ffn_dim == 14336
+    assert config.local_attn_size == 12
+    assert config.sink_size == 3
+    assert config.attn_compress == 4
+    assert config.cam_self_attn_layers == tuple(range(30))
+
+
+def test_dreamx_world_ar_converted_5b_transformer_strict_loads():
+    transformer_dir = CONVERTED_AR_DIR / "transformer"
+    if not transformer_dir.exists():
+        pytest.skip(f"Converted AR transformer missing: {transformer_dir}")
+    with torch.device("meta"):
+        model = DreamXWorldARTransformer3DModel(make_dreamx_world_5b_ar_dit_config(), {})
+    incompatible = load_model_from_full_model_state_dict(
+        model,
+        safetensors_weights_iterator(resolve_safetensors_files(str(transformer_dir)), to_cpu=True),
+        device=torch.device("cpu"),
+        param_dtype=torch.bfloat16,
+        strict=True,
+        param_names_mapping=get_param_names_mapping(model.param_names_mapping),
+        training_mode=False,
+    )
+    assert incompatible.missing_keys == []
+    assert incompatible.unexpected_keys == []
+    assert not any(param.is_meta for param in model.parameters())

tests/local_tests/dreamx_world/test_dreamx_world_camera_conditioning_parity.py

@@ -0,0 +1,95 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World camera-conditioning parity against the official reference.
+
+Coverage scope: implementation_subcomponent. This verifies the weightless
+action-sequence to PRoPE camera tensor path used by DreamX-World-5B-Cam.
+"""
+from __future__ import annotations
+
+import os
+from pathlib import Path
+import pytest
+import torch
+from torch.testing import assert_close
+
+from fastvideo.pipelines.basic.dreamx_world.camera_conditioning import build_dreamx_camera_condition
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+OFFICIAL_REF_DIR = Path(os.getenv("DREAMX_WORLD_OFFICIAL_REF_DIR", REPO_ROOT / "DreamX-World"))
+PARITY_SCOPE = "implementation_subcomponent"
+
+
+def _load_official_functions():
+    if not OFFICIAL_REF_DIR.exists():
+        pytest.skip(f"Official reference missing: {OFFICIAL_REF_DIR}")
+    try:
+        import importlib.util
+
+        pose_path = OFFICIAL_REF_DIR / "utils" / "pose_utils.py"
+        pose_spec = importlib.util.spec_from_file_location(
+            "dreamx_world_pose_utils", pose_path)
+        if pose_spec is None or pose_spec.loader is None:
+            raise RuntimeError(f"Cannot load DreamX pose_utils: {pose_path}")
+        pose_module = importlib.util.module_from_spec(pose_spec)
+        pose_spec.loader.exec_module(pose_module)
+
+        source = (OFFICIAL_REF_DIR / "utils" / "inference_utils.py").read_text()
+        source = source.replace(
+            "from .pose_utils import interpolate_camera_poses\n", "")
+        namespace = {"interpolate_camera_poses": pose_module.interpolate_camera_poses}
+        exec(compile(source, str(OFFICIAL_REF_DIR / "utils" / "inference_utils.py"), "exec"), namespace)
+    except Exception as exc:  # noqa: BLE001 - local parity should skip missing reference deps.
+        pytest.skip(f"Cannot load DreamX camera reference: {exc}")
+    return namespace["ActionToPoseFromID"], namespace["GetPoseEmbedsFromPosesPrope"]
+
+
+def _official_camera_condition(
+    action_seq: list[str],
+    action_speed_list: list[float],
+    *,
+    num_frames: int,
+    height: int,
+    width: int,
+    dtype: torch.dtype,
+):
+    action_to_pose, get_pose_embeds = _load_official_functions()
+    duration = -(-num_frames // len(action_seq))
+    poses = action_to_pose(action_seq, action_speed_list, duration=duration)[:num_frames]
+    condition, _ = get_pose_embeds(poses, height, width, len(poses), False, 0, dtype=dtype, device="cpu")
+    return condition
+
+
+@pytest.mark.parametrize(
+    ("action_seq", "action_speed_list", "num_frames"),
+    [
+        (["w"], [4], 81),
+        (["wj", "d"], [4, 6], 121),
+        (["i", "k", "l"], [3, 5, 2], 85),
+    ],
+)
+def test_dreamx_world_camera_conditioning_matches_official(action_seq, action_speed_list, num_frames):
+    dtype = torch.float32
+    official = _official_camera_condition(
+        action_seq,
+        action_speed_list,
+        num_frames=num_frames,
+        height=704,
+        width=1280,
+        dtype=dtype,
+    )
+    fastvideo = build_dreamx_camera_condition(
+        action_seq,
+        action_speed_list,
+        num_frames=num_frames,
+        height=704,
+        width=1280,
+        dtype=dtype,
+        device="cpu",
+    )
+
+    assert official.keys() == fastvideo.keys() == {"viewmats", "K"}
+    for key in ("viewmats", "K"):
+        assert official[key].shape == fastvideo[key].shape
+        diff = (official[key] - fastvideo[key]).abs()
+        print(f"{key}: shape={tuple(fastvideo[key].shape)} diff_max={diff.max().item():.8f} diff_mean={diff.mean().item():.8f}")
+        assert_close(fastvideo[key], official[key], atol=1e-5, rtol=1e-5)

tests/local_tests/dreamx_world/test_dreamx_world_conversion.py

@@ -0,0 +1,79 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World conversion script smoke tests."""
+
+import json
+
+from fastvideo.configs.pipelines.dreamx_world import make_dreamx_world_5b_cam_dit_config
+from fastvideo.models.registry import _LEGACY_FAST_VIDEO_MODELS
+
+from scripts.checkpoint_conversion.dreamx_world_to_diffusers import (
+    MODEL_INDEX,
+    REUSED_COMPONENTS,
+    TRANSFORMER_CONFIG,
+    _copy_or_link_component,
+    write_model_index,
+)
+
+
+def test_dreamx_world_converter_writes_full_model_index_with_reused_components(tmp_path):
+    component_source = tmp_path / "wan22"
+    output = tmp_path / "dreamx"
+    output.mkdir()
+
+    for component in REUSED_COMPONENTS:
+        component_dir = component_source / component
+        component_dir.mkdir(parents=True)
+        (component_dir / "config.json").write_text("{}\n")
+
+    write_model_index(output, component_source, symlink_components=True)
+
+    model_index = json.loads((output / "model_index.json").read_text())
+    assert model_index == MODEL_INDEX
+    assert model_index["_class_name"] == "DreamXWorldPipeline"
+    assert model_index["transformer"] == ["diffusers", "DreamXWorldTransformer3DModel"]
+    for component in REUSED_COMPONENTS:
+        assert (output / component).is_symlink()
+
+
+def test_dreamx_world_transformer_config_has_camera_adapter_enabled():
+    assert TRANSFORMER_CONFIG["_class_name"] == "DreamXWorldTransformer3DModel"
+    assert TRANSFORMER_CONFIG["add_control_adapter"] is True
+    assert TRANSFORMER_CONFIG["cam_method"] == "prope"
+    assert TRANSFORMER_CONFIG["num_layers"] == 30
+
+
+def test_dreamx_world_converter_transformer_config_matches_pipeline_dit_config():
+    dit_config = make_dreamx_world_5b_cam_dit_config()
+
+    assert TRANSFORMER_CONFIG["num_attention_heads"] == dit_config.num_attention_heads
+    assert TRANSFORMER_CONFIG["attention_head_dim"] == dit_config.attention_head_dim
+    assert TRANSFORMER_CONFIG["in_channels"] == dit_config.in_channels
+    assert TRANSFORMER_CONFIG["out_channels"] == dit_config.out_channels
+    assert TRANSFORMER_CONFIG["ffn_dim"] == dit_config.ffn_dim
+    assert TRANSFORMER_CONFIG["num_layers"] == dit_config.num_layers
+    assert TRANSFORMER_CONFIG["cross_attn_norm"] == dit_config.cross_attn_norm
+    assert TRANSFORMER_CONFIG["qk_norm"] == dit_config.qk_norm
+    assert TRANSFORMER_CONFIG["add_control_adapter"] == dit_config.add_control_adapter
+    assert TRANSFORMER_CONFIG["cam_method"] == dit_config.cam_method
+    assert TRANSFORMER_CONFIG["attn_compress"] == dit_config.attn_compress
+    assert TRANSFORMER_CONFIG["cam_self_attn_layers"] == dit_config.cam_self_attn_layers
+
+
+def test_dreamx_world_model_index_component_classes_are_registered():
+    for component in ("scheduler", "text_encoder", "transformer", "vae"):
+        class_name = MODEL_INDEX[component][1]
+        assert class_name in _LEGACY_FAST_VIDEO_MODELS
+
+
+def test_dreamx_world_copy_or_link_component_keeps_broken_symlink(tmp_path):
+    component_source = tmp_path / "wan22"
+    src = component_source / "scheduler"
+    src.mkdir(parents=True)
+    output = tmp_path / "dreamx"
+    output.mkdir()
+    dst = output / "scheduler"
+    dst.symlink_to(tmp_path / "missing_scheduler", target_is_directory=True)
+
+    _copy_or_link_component("scheduler", component_source, output, symlink=True)
+
+    assert dst.is_symlink()

tests/local_tests/dreamx_world/test_dreamx_world_pipeline_config.py

@@ -0,0 +1,306 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World pipeline config and conditioning smoke tests."""
+
+from types import SimpleNamespace
+import json
+
+import numpy as np
+import torch
+
+from fastvideo.api.presets import get_preset
+from fastvideo.fastvideo_args import WorkloadType
+from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.pipeline_registry import PipelineType, import_pipeline_classes
+from fastvideo.pipelines.basic.dreamx_world.camera_conditioning import (
+    DreamXCamera,
+    _interpolate_camera_poses,
+    build_dreamx_camera_condition,
+)
+from fastvideo.configs.pipelines.dreamx_world import DreamXWorld5BARPipelineConfig, DreamXWorld5BCamPipelineConfig
+from fastvideo.pipelines.basic.dreamx_world.ar_denoising import DreamXWorldARCausalDenoisingStage
+from fastvideo.pipelines.basic.dreamx_world.dreamx_world_ar_pipeline import DreamXWorldARPipeline
+from fastvideo.pipelines.basic.dreamx_world.dreamx_world_pipeline import DreamXWorldPipeline
+from fastvideo.pipelines.basic.dreamx_world.stages import (
+    DREAMX_Y_CAMERA_KEY,
+    DreamXWorldCameraConditioningStage,
+)
+from fastvideo.pipelines.stages.denoising import DenoisingStage
+from fastvideo.registry import get_default_preset, get_model_info, get_pipeline_config_cls_from_name
+
+
+def test_dreamx_world_5b_cam_pipeline_config_wires_first_scope_components():
+    config = DreamXWorld5BCamPipelineConfig()
+
+    assert config.flow_shift == 3.0
+    assert config.ti2v_task is True
+    assert config.expand_timesteps is True
+    assert config.dit_config.expand_timesteps is True
+    assert config.dit_config.num_layers == 30
+    assert config.dit_config.add_control_adapter is True
+    assert config.dit_config.cam_method == "prope"
+
+    assert config.vae_config.load_encoder is True
+    assert config.vae_config.load_decoder is True
+    assert config.vae_config.z_dim == 48
+    assert config.vae_config.scale_factor_temporal == 4
+    assert config.vae_config.scale_factor_spatial == 16
+
+    assert len(config.text_encoder_configs) == 1
+    text_config = config.text_encoder_configs[0]
+    assert text_config.prefix == "umt5"
+    assert text_config.vocab_size == 256384
+    assert text_config.d_model == 4096
+    assert config.text_encoder_precisions == ("bf16",)
+
+
+def test_dreamx_world_pipeline_registry_discovers_entrypoint():
+    pipelines = import_pipeline_classes(PipelineType.BASIC)
+
+    assert pipelines["basic"]["DreamXWorldPipeline"] is DreamXWorldPipeline
+
+
+
+
+def test_dreamx_world_local_model_index_resolves_model_info(tmp_path):
+    model_dir = tmp_path / "DreamX-World-5B-Cam-converted"
+    model_dir.mkdir()
+    for component in ("scheduler", "text_encoder", "tokenizer", "transformer", "vae"):
+        (model_dir / component).mkdir()
+    (model_dir / "model_index.json").write_text(json.dumps({
+        "_class_name": "DreamXWorldPipeline",
+        "_diffusers_version": "0.31.0",
+        "scheduler": ["diffusers", "FlowMatchEulerDiscreteScheduler"],
+        "text_encoder": ["transformers", "UMT5EncoderModel"],
+        "tokenizer": ["transformers", "AutoTokenizer"],
+        "transformer": ["diffusers", "DreamXWorldTransformer3DModel"],
+        "vae": ["diffusers", "AutoencoderKLWan"],
+    }) + "\n")
+
+    info = get_model_info(str(model_dir), pipeline_type=PipelineType.BASIC, workload_type=WorkloadType.I2V)
+
+    assert info.pipeline_cls is DreamXWorldPipeline
+    assert info.pipeline_config_cls is DreamXWorld5BCamPipelineConfig
+
+def test_dreamx_world_model_path_resolves_pipeline_config():
+    assert get_pipeline_config_cls_from_name("GD-ML/DreamX-World-5B-Cam") is DreamXWorld5BCamPipelineConfig
+
+
+
+def test_dreamx_world_default_preset_is_registered():
+    preset_name = get_default_preset("GD-ML/DreamX-World-5B-Cam")
+    preset = get_preset(preset_name, "dreamx_world")
+
+    assert preset.name == "dreamx_world_5b_cam"
+    assert preset.workload_type == "i2v"
+    assert preset.defaults["height"] == 480
+    assert preset.defaults["width"] == 832
+    assert preset.defaults["num_frames"] == 161
+    assert preset.defaults["num_inference_steps"] == 30
+    assert preset.defaults["guidance_scale"] == 5.0
+
+
+
+
+def test_dreamx_world_pipeline_initializes_official_flow_scheduler():
+    pipeline = DreamXWorldPipeline.__new__(DreamXWorldPipeline)
+    pipeline.modules = {}
+    fastvideo_args = SimpleNamespace(pipeline_config=DreamXWorld5BCamPipelineConfig())
+
+    pipeline.initialize_pipeline(fastvideo_args)
+
+    scheduler = pipeline.modules["scheduler"]
+    assert isinstance(scheduler, FlowMatchEulerDiscreteScheduler)
+    assert scheduler.config.shift == 3.0
+
+def test_dreamx_world_camera_conditioning_stage_sets_y_camera_extra():
+    batch = ForwardBatch(
+        data_type="t2v",
+        action_list=["wj", "d"],
+        action_speed_list=[4, 6],
+        num_frames=17,
+        height=704,
+        width=1280,
+        latents=torch.zeros(1, 16, 5, 44, 80),
+    )
+    stage = DreamXWorldCameraConditioningStage()
+
+    out = stage.forward(batch, fastvideo_args=object())
+    y_camera = out.extra[DREAMX_Y_CAMERA_KEY]
+    expected = build_dreamx_camera_condition(
+        ["wj", "d"],
+        [4, 6],
+        num_frames=17,
+        height=704,
+        width=1280,
+        dtype=torch.float32,
+        device="cpu",
+    )
+
+    assert set(y_camera) == {"viewmats", "K"}
+    for key, expected_value in expected.items():
+        assert y_camera[key].shape == (1, *expected_value.shape)
+        torch.testing.assert_close(y_camera[key][0], expected_value)
+    assert stage.verify_output(out, fastvideo_args=object()).is_valid()
+
+
+def test_dreamx_world_denoising_kwargs_filter_for_y_camera():
+    y_camera = {"viewmats": torch.eye(4).reshape(1, 1, 4, 4), "K": torch.eye(3).reshape(1, 1, 3, 3)}
+    stage = DenoisingStage.__new__(DenoisingStage)
+
+    def accepts_y_camera(hidden_states, encoder_hidden_states, timestep, y_camera=None):
+        return y_camera
+
+    def no_y_camera(hidden_states, encoder_hidden_states, timestep):
+        return hidden_states
+
+    assert stage.prepare_extra_func_kwargs(accepts_y_camera, {"y_camera": y_camera}) == {"y_camera": y_camera}
+    assert stage.prepare_extra_func_kwargs(no_y_camera, {"y_camera": y_camera}) == {}
+
+
+
+def test_dreamx_world_ar_pipeline_config_wires_components():
+    config = DreamXWorld5BARPipelineConfig()
+    assert config.is_causal is True
+    assert config.flow_shift == 5.0
+    assert config.dmd_denoising_steps == (1000, 750, 500, 250)
+    assert config.warp_denoising_step is True
+    assert config.context_noise == 0.1
+    assert config.dit_config.arch_config.local_attn_size == 12
+    assert config.dit_config.arch_config.sink_size == 3
+    assert config.dit_config.arch_config.attn_compress == 4
+
+
+def test_dreamx_world_ar_pipeline_registry_and_preset():
+    from fastvideo.api.presets import get_preset
+    from fastvideo.registry import get_pipeline_config_cls_from_name, get_preset_selection
+
+    assert get_pipeline_config_cls_from_name("GD-ML/DreamX-World-5B") is DreamXWorld5BARPipelineConfig
+    preset_name, family = get_preset_selection("GD-ML/DreamX-World-5B")
+    assert (preset_name, family) == ("dreamx_world_5b_ar", "dreamx_world")
+    preset = get_preset("dreamx_world_5b_ar", "dreamx_world")
+    assert preset.defaults["num_inference_steps"] == 4
+    assert DreamXWorldARPipeline.pipeline_config_cls is DreamXWorld5BARPipelineConfig
+
+
+def test_dreamx_world_camera_conditioning_stage_expands_scalar_speed():
+    batch = ForwardBatch(
+        data_type="t2v",
+        action_list=["w", "d"],
+        action_speed_list=2.0,
+        num_frames=17,
+        height=704,
+        width=1280,
+        latents=torch.zeros(1, 16, 5, 44, 80),
+    )
+    stage = DreamXWorldCameraConditioningStage()
+
+    out = stage.forward(batch, fastvideo_args=object())
+
+    assert set(out.extra[DREAMX_Y_CAMERA_KEY]) == {"viewmats", "K"}
+
+
+def test_dreamx_world_camera_interpolation_handles_single_camera():
+    camera = DreamXCamera(
+        fx=0.8,
+        fy=0.8,
+        cx=0.5,
+        cy=0.5,
+        w2c_mat=np.eye(4, dtype=np.float64),
+    )
+
+    out = _interpolate_camera_poses(
+        [camera],
+        src_indices=np.array([0.0]),
+        tgt_indices=np.array([0.0, 1.0, 2.0]),
+    )
+
+    assert out == [camera, camera, camera]
+
+
+def test_dreamx_world_ar_cache_initializes_camera_self_attention_entries():
+    cam_self_attn = SimpleNamespace(num_heads=3, head_dim=5)
+    transformer = SimpleNamespace(
+        blocks=[SimpleNamespace(cam_self_attn=cam_self_attn), SimpleNamespace(cam_self_attn=cam_self_attn)],
+        num_attention_heads=2,
+        attention_head_dim=4,
+    )
+    stage = DreamXWorldARCausalDenoisingStage.__new__(DreamXWorldARCausalDenoisingStage)
+    stage.transformer = transformer
+    stage.num_transformer_blocks = 2
+    stage.local_attn_size = 6
+
+    caches = stage._initialize_kv_cache(
+        batch_size=1,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+        frame_seq_length=7,
+    )
+
+    assert len(caches) == 2
+    assert caches[0]["k"].shape == (1, 42, 2, 4)
+    assert caches[0]["prope_k"].shape == (1, 42, 3, 5)
+    assert caches[0]["prope_v"].shape == (1, 42, 3, 5)
+    assert int(caches[0]["prope_global_end_index"].item()) == 0
+    assert int(caches[0]["prope_local_end_index"].item()) == 0
+
+
+def test_dreamx_world_ar_context_noise_fraction_maps_to_scheduler_timestep():
+    assert DreamXWorldARCausalDenoisingStage._context_noise_timestep(0.1) == 100
+    assert DreamXWorldARCausalDenoisingStage._context_noise_timestep(100) == 100
+
+
+
+def test_dreamx_world_ar_context_update_advances_camera_cache_indices():
+    class DummyTransformer:
+        def __call__(self, *, hidden_states, encoder_hidden_states, timestep, y_camera, kv_cache, crossattn_cache,
+                     current_start):
+            del encoder_hidden_states, y_camera, crossattn_cache
+            assert current_start == 0
+            assert timestep.unique().tolist() == [100]
+            new_tokens = timestep.shape[1]
+            for cache in kv_cache:
+                cache["local_end_index"] += new_tokens
+                cache["global_end_index"] += new_tokens
+                cache["prope_local_end_index"] += new_tokens
+                cache["prope_global_end_index"] += new_tokens
+                cache["k"][:, :new_tokens] = 1
+                cache["prope_k"][:, :new_tokens] = 1
+            return hidden_states
+
+    cam_self_attn = SimpleNamespace(num_heads=3, head_dim=5)
+    cache_transformer = SimpleNamespace(
+        blocks=[SimpleNamespace(cam_self_attn=cam_self_attn)],
+        num_attention_heads=2,
+        attention_head_dim=4,
+    )
+    stage = DreamXWorldARCausalDenoisingStage.__new__(DreamXWorldARCausalDenoisingStage)
+    stage.transformer = cache_transformer
+    stage.num_transformer_blocks = 1
+    stage.local_attn_size = 6
+    caches = stage._initialize_kv_cache(
+        batch_size=1,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+        frame_seq_length=2,
+    )
+    # Keep the cache allocation source separate from the callable transformer used by _update_context_cache.
+    stage.transformer = DummyTransformer()
+
+    stage._update_context_cache(
+        block_latents=torch.zeros(1, 4, 3, 2, 2),
+        context=[torch.zeros(2, 4)],
+        camera_block={"viewmats": torch.eye(4).reshape(1, 1, 4, 4), "K": torch.eye(3).reshape(1, 1, 3, 3)},
+        kv_cache=caches,
+        crossattn_cache=[{}],
+        start=0,
+        frame_seq_length=2,
+        target_dtype=torch.float32,
+        autocast_enabled=False,
+        context_noise=0.1,
+    )
+
+    assert int(caches[0]["local_end_index"].item()) == 6
+    assert int(caches[0]["prope_local_end_index"].item()) == 6
+    assert caches[0]["k"][:, :6].sum().item() == 48
+    assert caches[0]["prope_k"][:, :6].sum().item() == 90

tests/local_tests/dreamx_world/test_dreamx_world_scheduler_parity.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+"""DreamX-World default Flow scheduler parity.
+
+Coverage scope: implementation_subcomponent. DreamX-World-5B-Cam defaults to
+Diffusers FlowMatchEulerDiscreteScheduler for sampler_name=Flow. This test
+checks that FastVideo's native FlowMatchEulerDiscreteScheduler matches the
+timestep schedule and Euler step used by the official default path.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+import inspect
+
+import torch
+from diffusers import FlowMatchEulerDiscreteScheduler as OfficialFlowMatchEulerDiscreteScheduler
+from omegaconf import OmegaConf
+from torch.testing import assert_close
+
+from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
+    FlowMatchEulerDiscreteScheduler as FastVideoFlowMatchEulerDiscreteScheduler,
+)
+
+REPO_ROOT = Path(__file__).resolve().parents[3]
+PARITY_SCOPE = "implementation_subcomponent"
+
+
+def _scheduler_kwargs(cls):
+    config_path = REPO_ROOT / "DreamX-World" / "configs" / "wan2.2" / "wan_ti2v_5b.yaml"
+    config = OmegaConf.load(config_path)
+    raw_kwargs = OmegaConf.to_container(config["scheduler_kwargs"])
+    signature = inspect.signature(cls)
+    return {key: value for key, value in raw_kwargs.items() if key in signature.parameters}
+
+
+def test_dreamx_world_flow_scheduler_timesteps_and_step_match():
+    official = OfficialFlowMatchEulerDiscreteScheduler(**_scheduler_kwargs(OfficialFlowMatchEulerDiscreteScheduler))
+    fastvideo = FastVideoFlowMatchEulerDiscreteScheduler(**_scheduler_kwargs(FastVideoFlowMatchEulerDiscreteScheduler))
+
+    official.set_timesteps(50, device="cpu", mu=1)
+    fastvideo.set_timesteps(50, device="cpu", mu=1)
+    assert_close(fastvideo.timesteps, official.timesteps, atol=0, rtol=0)
+    assert_close(fastvideo.sigmas, official.sigmas, atol=0, rtol=0)
+
+    torch.manual_seed(7)
+    sample = torch.randn(1, 4, 2, 8, 8)
+    model_output = torch.randn_like(sample)
+    timestep = official.timesteps[3]
+
+    official_prev = official.step(model_output, timestep, sample, return_dict=False)[0]
+    fastvideo_prev = fastvideo.step(model_output, fastvideo.timesteps[3], sample, return_dict=False)[0]
+    diff = (official_prev - fastvideo_prev).abs()
+    print(f"scheduler diff_max={diff.max().item():.8f} diff_mean={diff.mean().item():.8f}")
+    assert_close(fastvideo_prev, official_prev, atol=0, rtol=0)