wip

Co-authored-by: RandNMR73 <notomatthew31@gmail.com>
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>

.buildkite/pipeline.yml

@@ -18,7 +18,8 @@ steps:
             - path: 
                 - "fastvideo/models/encoders/**"
                 - "fastvideo/models/loader/**"
-                - "fastvideo/tests/encoders/**"
+                - "tests/encoders/**"
+                - "tests/utils.py"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -31,7 +32,8 @@ steps:
             - path:
                 - "fastvideo/models/vaes/**"
                 - "fastvideo/models/loader/**"
-                - "fastvideo/tests/vaes/**"
+                - "tests/vaes/**"
+                - "tests/utils.py"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -44,7 +46,8 @@ steps:
             - path:
                 - "fastvideo/models/dits/**"
                 - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                 - "fastvideo/layers/**"
                 - "fastvideo/attention/**"
                 - "pyproject.toml"
@@ -58,6 +61,8 @@ steps:
                   queue: "default"
             - path: 
                 - "fastvideo/**/*.py"
+                - "tests/ssim/**"
+                - "tests/utils.py"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -68,9 +73,10 @@ steps:
                 agents:
                   queue: "default"
             - path:
-                - "fastvideo/tests/lora/**"
+                - "tests/inference/lora/**"
                 - "fastvideo/models/loader/**"
-                - "fastvideo/tests/transformers/**"
+                - "tests/transformers/**"
+                - "tests/utils.py"
                 - "fastvideo/pipelines/**"
                 - "fastvideo/layers/lora/**"
                 - "pyproject.toml"
@@ -84,6 +90,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
+                - "tests/training/Vanilla/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -95,6 +102,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/training/*distillation_pipeline.py"
+                - "tests/training/distill/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -106,6 +114,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
+                - "tests/training/lora/**"
                 - "pyproject.toml"
                 - "docker/Dockerfile.python3.12"
               config:
@@ -117,6 +126,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
+                - "tests/training/VSA/**"
                 - "csrc/attn/video_sparse_attn/**"
                 - "csrc/attn/video_sparse_attn/tk/**"
                 - "csrc/attn/video_sparse_attn/setup.py"
@@ -133,6 +143,7 @@ steps:
                   queue: "default"
             - path:
                 - "fastvideo/**"
+                - "tests/inference/STA/**"
                 - "csrc/attn/sliding_tile_attn/**"
                 - "csrc/attn/sliding_tile_attn/setup.py"
                 - "csrc/attn/sliding_tile_attn/config_sta.py"
@@ -176,3 +187,40 @@ steps:
                   - TEST_TYPE=precision_vsa
                 agents:
                   queue: "default"
+            - path:
+                - "csrc/attn/vmoba_attn/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Precision Tests VMoBA"
+                env:
+                  - TEST_TYPE=precision_vmoba
+                agents:
+                  queue: "default"
+            - path:
+                - "csrc/attn/vmoba_attn/vmoba/**"
+                - "fastvideo/attention/backends/vmoba.py"
+                - "tests/inference/vmoba/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Inference Tests VMoBA"
+                env:  
+                  - TEST_TYPE=inference_vmoba
+                agents:
+                  queue: "default"
+            - path:
+                - "fastvideo/**"
+                - "tests/dataset/**"
+                - "tests/workflow/**"
+                - "pyproject.toml"
+                - "docker/Dockerfile.python3.12"
+              config:
+                command: "timeout 15m .buildkite/scripts/pr_test.sh"
+                label: "Unit Tests"
+                env:
+                  - TEST_TYPE=unit_test
+                agents:
+                  queue: "default"

.buildkite/scripts/pr_test.sh

@@ -50,7 +50,7 @@ else
     exit 1
 fi
 
-MODAL_TEST_FILE="fastvideo/tests/modal/pr_test.py"
+MODAL_TEST_FILE="tests/modal/pr_test.py"
 
 if [ -z "${TEST_TYPE:-}" ]; then
     log "Error: TEST_TYPE environment variable is not set"
@@ -109,6 +109,19 @@ case "$TEST_TYPE" in
         log "Running distillation DMD tests..."
         MODAL_COMMAND="$MODAL_ENV WANDB_API_KEY=$WANDB_API_KEY python3 -m modal run $MODAL_TEST_FILE::run_distill_dmd_tests"
         ;;
+        # run_inference_tests_vmoba
+    "inference_vmoba")
+        log "Running V-MoBA inference tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_inference_tests_vmoba"
+        ;;
+    "precision_vmoba")
+        log "Running V-MoBA precision tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_precision_tests_vmoba"
+        ;;
+    "unit_test")
+        log "Running unit tests..."
+        MODAL_COMMAND="$MODAL_ENV python3 -m modal run $MODAL_TEST_FILE::run_unit_test"
+        ;;
     *)
         log "Error: Unknown test type: $TEST_TYPE"
         exit 1

.github/workflows/pr-test.yml

@@ -62,8 +62,8 @@ on:
         required: false
         default: false
         type: boolean
-      run_nightly_test:
-        description: "Run nightly-test"
+      run_unit_test:
+        description: "Run unit-test"
         required: false
         default: false
         type: boolean
@@ -93,6 +93,7 @@ jobs:
       inference-test-STA: ${{ steps.filter.outputs.inference-test-STA }}
       precision-test-STA: ${{ steps.filter.outputs.precision-test-STA }}
       precision-test-VSA: ${{ steps.filter.outputs.precision-test-VSA }}
+      unit-test: ${{ steps.filter.outputs.unit-test }}
     steps:
       - uses: actions/checkout@v4
       - uses: dorny/paths-filter@v3
@@ -102,6 +103,8 @@ jobs:
             # Define reusable path patterns
             common-paths: &common-paths
               - 'pyproject.toml'
+              - 'docker/Dockerfile.python3.10'
+              - 'docker/Dockerfile.python3.11'
               - 'docker/Dockerfile.python3.12'
             sta-kernel-paths: &sta-kernel-paths
               - 'csrc/attn/sliding_tile_attn/**'
@@ -124,29 +127,35 @@ jobs:
             encoder-test:
               - 'fastvideo/models/encoders/**'
               - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/encoders/**'
+              - 'tests/encoders/**'
+              - 'tests/utils.py'
               - *common-paths
             vae-test:
               - 'fastvideo/models/vaes/**'
               - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/vaes/**'
+              - 'tests/vaes/**'
+              - 'tests/utils.py'
               - *common-paths
             transformer-test:
               - 'fastvideo/models/dits/**'
               - 'fastvideo/models/loader/**'
-              - 'fastvideo/tests/transformers/**'
+              - 'tests/transformers/**'
+              - 'tests/utils.py'
               - 'fastvideo/layers/**'
               - 'fastvideo/attention/**'
               - *common-paths
             training-test:
               - 'fastvideo/**'
+              - 'tests/training/Vanilla/**'
               - *common-paths
             training-test-VSA:
               - 'fastvideo/**'
+              - 'tests/training/VSA/**'
               - *common-paths
               - *vsa-kernel-paths
             inference-test-STA:
               - 'fastvideo/**'
+              - 'tests/inference/STA/**'
               - *common-paths
               - *sta-kernel-paths
             precision-test-STA:
@@ -155,6 +164,11 @@ jobs:
             precision-test-VSA:
               - *common-paths
               - *vsa-kernel-paths
+            unit-test:
+              - 'fastvideo/**'
+              - 'tests/dataset/**'
+              - 'tests/workflow/**'
+              - *common-paths
 
   encoder-test:
     needs: change-filter
@@ -168,7 +182,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/encoders -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/encoders -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -186,7 +200,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/vaes -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/vaes -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -204,7 +218,7 @@ jobs:
       gpu_count: 1
       volume_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/transformers -s"
+      test_command: "uv pip install -e .[test] && pytest ./tests/transformers -s"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -230,12 +244,12 @@ jobs:
       volume_size: 200
       disk_size: 200
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:${{ matrix.python-version.tag }}"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/ssim -vs"
+      test_command: "uv pip install -e .[test] && pytest ./tests/ssim -vs"
       timeout_minutes: 60
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
       RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-  
+
   training-test:
     needs: change-filter
     if: >-
@@ -249,7 +263,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/Vanilla -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/Vanilla -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -269,7 +283,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/training/VSA -srP"
+      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/training/VSA -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -289,7 +303,7 @@ jobs:
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "uv pip install -e .[test] && pytest ./fastvideo/tests/inference/STA -srP"
+      test_command: "uv pip install -e .[test] && pytest ./tests/inference/STA -srP"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
@@ -333,23 +347,42 @@ jobs:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
       RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
 
-  nightly-test:
+  unit-test:
+    needs: change-filter
     if: >-
-      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_nightly_test == 'true')
+      (github.event_name != 'workflow_dispatch' && needs.change-filter.outputs.unit-test == 'true') || 
+      (github.event_name == 'workflow_dispatch' && github.event.inputs.run_unit_test == 'true')
     uses: ./.github/workflows/runpod-test.yml
     with:
-      job_id: "nightly-test"
-      gpu_type: "NVIDIA A40"
-      gpu_count: 4
+      job_id: "unit-test"
+      gpu_type: "NVIDIA L40S"
+      gpu_count: 1
       volume_size: 100
       disk_size: 100
       image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
-      test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./fastvideo/tests/nightly/test_e2e_overfit_single_sample.py -vs"
+      test_command: "uv pip install -e .[test] && pytest ./fastvideo/dataset/ -vs && pytest ./fastvideo/workflow/ -vs"
       timeout_minutes: 30
     secrets:
       RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
       RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
-      WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
+
+  # nightly-test:
+  #   if: >-
+  #     (github.event_name == 'workflow_dispatch' && github.event.inputs.run_nightly_test == 'true')
+  #   uses: ./.github/workflows/runpod-test.yml
+  #   with:
+  #     job_id: "nightly-test"
+  #     gpu_type: "NVIDIA A40"
+  #     gpu_count: 4
+  #     volume_size: 100
+  #     disk_size: 100
+  #     image: "ghcr.io/${{ github.repository }}/fastvideo-dev:py3.12-latest"
+#     test_command: "wandb login $WANDB_API_KEY && uv pip install -e .[test] && pytest ./tests/nightly/test_e2e_overfit_single_sample.py -vs"
+  #     timeout_minutes: 30
+  #   secrets:
+  #     RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
+  #     RUNPOD_PRIVATE_KEY: ${{ secrets.RUNPOD_PRIVATE_KEY }}
+  #     WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }}
 
   runpod-cleanup:
     # Add other jobs to this list as you create them
@@ -373,4 +406,4 @@ jobs:
           JOB_IDS: '["encoder-test", "vae-test", "transformer-test", "ssim-test-py3.10", "ssim-test-py3.11", "ssim-test-py3.12", "training-test", "training-test-VSA", "inference-test-STA", "precision-test-STA", "precision-test-VSA"]'
           RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
           GITHUB_RUN_ID: ${{ github.run_id }}
-        run: python .github/scripts/runpod_cleanup.py
+        run: python .github/scripts/runpod_cleanup.py

.gitignore

@@ -65,4 +65,5 @@ docs/source/distillation/examples/
 !comfyui/assets/**/*.png
 !comfyui/assets/**/*.gif
 
-dmd_t2v_output/
+dmd_t2v_output/
+preprocess_output_text/

.pre-commit-config.yaml

@@ -12,9 +12,6 @@ exclude: |
     scripts/.*|
     fastvideo/data_preprocess/.*|
     fastvideo/dataset/.*|
-    fastvideo/distill/.*|
-    fastvideo/distill\.py|
-    fastvideo/distill_adv\.py|
     fastvideo/models/.*|
     fastvideo/sample/.*|
     fastvideo/train\.py|
@@ -44,10 +41,10 @@ repos:
   - id: codespell
     additional_dependencies: ['tomli']
     args: ['--toml', 'pyproject.toml']
-- repo: https://github.com/PyCQA/isort
-  rev: 6.0.1
-  hooks:
-  - id: isort
+# - repo: https://github.com/PyCQA/isort
+#   rev: 6.0.1
+#   hooks:
+#   - id: isort
 - repo: https://github.com/jackdewinter/pymarkdown
   rev: v0.9.30
   hooks:

README.md

@@ -7,7 +7,7 @@
 FastVideo features an end-to-end unified pipeline for accelerating diffusion models, starting from data preprocessing to model training, finetuning, distillation, and inference. FastVideo is designed to be modular and extensible, allowing users to easily add new optimizations and techniques. Whether it is training-free optimizations or post-training optimizations, FastVideo has you covered.
 
 <p align="center">
-    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-38u6p1jqe-yDI1QJOCEnbtkLoaI5bjZQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/rG0QpZdw" target="_blank"> <b> WeChat </b> </a> |
+    | 🕹️ <a href="https://fastwan.fastvideo.org/"<b>Online Demo</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo"><b>Documentation</b></a> | <a href="https://hao-ai-lab.github.io/FastVideo/inference/inference_quick_start.html"><b> Quick Start</b></a> | 🤗 <a href="https://huggingface.co/collections/FastVideo/fastwan-6886a305d9799c8cd1496408"  target="_blank"><b>FastWan</b></a>  | 🟣💬 <a href="https://join.slack.com/t/fastvideo/shared_invite/zt-3csdw1isz-Euq8_Q8~baewG8hxjXs2gQ" target="_blank"> <b>Slack</b> </a> |  🟣💬 <a href="https://ibb.co/q46BbX6" target="_blank"> <b> WeChat </b> </a> |
 </p>
 
 <div align="center">
@@ -155,8 +155,8 @@ If you find FastVideo useful, please considering citing our work:
 }
 
 @article{zhang2025vsa,
-  title={VSA: Faster Video Diffusion with Trainable Sparse Attention},
-  author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
+  title={Vsa: Faster video diffusion with trainable sparse attention},
+  author={Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
   journal={arXiv preprint arXiv:2505.13389},
   year={2025}
 }

csrc/attn/vmoba_attn/README.md

@@ -0,0 +1,32 @@
+# Attention Kernel Used in FastVideo
+
+## VMoBA: Mixture-of-Block Attention for Video Diffusion Models (VMoBA)
+
+### Installation
+Please ensure that you have installed FlashAttention version **2.7.1 or higher**, as some interfaces have changed in recent releases.
+
+### Usage
+
+You can use `moba_attn_varlen` in the following ways:
+
+**Install from source:**
+```bash
+python setup.py install
+```
+
+**Import after installation:**
+```python
+from vmoba import moba_attn_varlen
+```
+
+**Or import directly from the project root:**
+```python
+from csrc.attn.vmoba_attn.vmoba import moba_attn_varlen
+```
+
+### Verify if you have successfully installed
+
+```bash
+python csrc/attn/vmoba_attn/vmoba/vmoba.py
+```
+

csrc/attn/vmoba_attn/setup.py

@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from setuptools import find_packages, setup
+
+PACKAGE_NAME = "vmoba"
+VERSION = "0.0.0"
+AUTHOR = "JianzongWu"
+DESCRIPTION = "VMoBA: Mixture-of-Block Attention for Video Diffusion Models"
+URL = "https://github.com/KwaiVGI/VMoBA"
+
+setup(
+    name=PACKAGE_NAME,
+    version=VERSION,
+    author=AUTHOR,
+    description=DESCRIPTION,
+    url=URL,
+    packages=find_packages(),
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: Apache Software License",
+    ],
+    python_requires='>=3.12',
+    install_requires=[
+        "flash-attn >= 2.7.1",
+    ]
+)

csrc/attn/vmoba_attn/tests/test_vmoba_attn.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import pytest
+import random
+from csrc.attn.vmoba_attn.vmoba  import moba_attn_varlen
+
+def generate_test_data(batch_size, total_seqlen, num_heads, head_dim, dtype, device="cuda"):
+    """
+    Generates random data for testing the variable-length attention function.
+    """
+    torch.manual_seed(42)
+    random.seed(42)
+    torch.cuda.manual_seed_all(42)
+
+    # Generate sequence lengths for each item in the batch
+    if batch_size > 1:
+        # Ensure sequence lengths are reasonably distributed
+        avg_seqlen = total_seqlen // batch_size
+        seqlens = [random.randint(avg_seqlen // 2, avg_seqlen + avg_seqlen // 2) for _ in range(batch_size - 1)]
+        remaining_len = total_seqlen - sum(seqlens)
+        if remaining_len > 0:
+            seqlens.append(remaining_len)
+        else: # Adjust if sum exceeds total_seqlen
+            seqlens.append(avg_seqlen)
+            current_sum = sum(seqlens)
+            seqlens[-1] -= (current_sum - total_seqlen)
+        # Ensure all lengths are positive
+        seqlens = [max(1, s) for s in seqlens]
+        # Final adjustment to match total_seqlen
+        seqlens[-1] += total_seqlen - sum(seqlens)
+
+    else:
+        seqlens = [total_seqlen]
+
+    cu_seqlens = torch.tensor([0] + list(torch.cumsum(torch.tensor(seqlens), 0)), device=device, dtype=torch.int32)
+    max_seqlen = max(seqlens) if seqlens else 0
+
+    q = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    k = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+    v = torch.randn((total_seqlen, num_heads, head_dim), dtype=dtype, device=device, requires_grad=False)
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+@pytest.mark.parametrize("batch_size", [1, 2])
+@pytest.mark.parametrize("total_seqlen", [512, 1024])
+@pytest.mark.parametrize("num_heads", [8])
+@pytest.mark.parametrize("head_dim", [64])
+@pytest.mark.parametrize("moba_chunk_size", [64])
+@pytest.mark.parametrize("moba_topk", [2, 4])
+@pytest.mark.parametrize("select_mode", ["topk", "threshold"])
+@pytest.mark.parametrize("threshold_type", ["query_head", "head_global", "overall"])
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+def test_moba_attn_varlen_forward(
+    batch_size, total_seqlen, num_heads, head_dim, moba_chunk_size, moba_topk, select_mode, threshold_type, dtype
+):
+    """
+    Tests the forward pass of moba_attn_varlen for basic correctness.
+    It checks output shape, dtype, and for the presence of NaNs/Infs.
+    """
+    if dtype == torch.float32:
+        pytest.skip("float32 is not supported in flash attention")
+
+    q, k, v, cu_seqlens, max_seqlen = generate_test_data(
+        batch_size, total_seqlen, num_heads, head_dim, dtype
+    )
+
+    # Ensure chunk size is not larger than the smallest sequence length
+    min_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).min().item()
+    if moba_chunk_size > min_seqlen:
+        pytest.skip("moba_chunk_size is larger than the minimum sequence length in the batch")
+
+    try:
+        output = moba_attn_varlen(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            moba_chunk_size=moba_chunk_size,
+            moba_topk=moba_topk,
+            select_mode=select_mode,
+            threshold_type=threshold_type,
+            simsum_threshold=0.5, # A reasonable default for threshold mode
+        )
+    except Exception as e:
+        pytest.fail(f"moba_attn_varlen forward pass failed with exception: {e}")
+
+    # 1. Check output shape
+    assert output.shape == q.shape, f"Expected output shape {q.shape}, but got {output.shape}"
+
+    # 2. Check output dtype
+    assert output.dtype == q.dtype, f"Expected output dtype {q.dtype}, but got {output.dtype}"
+
+    # 3. Check for NaNs or Infs in the output
+    assert torch.all(torch.isfinite(output)), "Output contains NaN or Inf values"

csrc/attn/vmoba_attn/vmoba/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: Apache-2.0
+from .vmoba import moba_attn_varlen, process_moba_input, process_moba_output

csrc/attn/vmoba_attn/vmoba/vmoba.py

@@ -0,0 +1,868 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapt from https://github.com/KwaiVGI/VMoBA/blob/main/src/vmoba.py
+
+import random
+import time
+import os
+import torch
+from typing import Tuple
+try:
+    from flash_attn import flash_attn_varlen_func  # Use the new flash attention function
+    from flash_attn.flash_attn_interface import _flash_attn_varlen_forward, _flash_attn_varlen_backward
+except ImportError:
+    def _unsupported(*args, **kwargs):
+        raise ImportError("flash-attn is not installed. Please install it, e.g., `pip install flash-attn`.")
+    _flash_attn_varlen_forward = _unsupported
+    _flash_attn_varlen_backward = _unsupported
+    flash_attn_varlen_func = _unsupported
+    
+from functools import lru_cache
+from einops import rearrange
+
+
+@lru_cache(maxsize=16)
+def calc_chunks(cu_seqlen, moba_chunk_size):
+    """
+    Calculate chunk boundaries.
+
+    For vision tasks we include all chunks (even the last one which might be shorter)
+    so that every chunk can be selected.
+    """
+    batch_sizes = cu_seqlen[1:] - cu_seqlen[:-1]
+    batch_num_chunk = (batch_sizes + (moba_chunk_size - 1)) // moba_chunk_size
+    cu_num_chunk = torch.ones(
+        batch_num_chunk.numel() + 1,
+        device=cu_seqlen.device,
+        dtype=batch_num_chunk.dtype,
+    )
+    cu_num_chunk[1:] = batch_num_chunk.cumsum(dim=0)
+    num_chunk = cu_num_chunk[-1]
+    chunk_sizes = torch.full(
+        (num_chunk + 1,), moba_chunk_size, dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_sizes[0] = 0
+    batch_last_chunk_size = batch_sizes - (batch_num_chunk - 1) * moba_chunk_size
+    chunk_sizes[cu_num_chunk[1:]] = batch_last_chunk_size
+    cu_chunk = chunk_sizes.cumsum(dim=-1, dtype=torch.int32)
+    chunk_to_batch = torch.zeros(
+        (num_chunk,), dtype=torch.int32, device=cu_seqlen.device
+    )
+    chunk_to_batch[cu_num_chunk[1:-1]] = 1
+    chunk_to_batch = chunk_to_batch.cumsum(dim=0, dtype=torch.int32)
+
+    # Do not filter out any chunk
+    filtered_chunk_indices = torch.arange(
+        num_chunk, device=cu_seqlen.device, dtype=torch.int32
+    )
+    num_filtered_chunk = num_chunk
+
+    return cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch
+
+
+# --- Threshold Selection Helper Functions ---
+
+def _select_threshold_query_head(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects chunks for each <query, head> pair based on threshold.
+    Normalization and sorting happen along the chunk dimension (dim=0).
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # LSE‐style normalization per <head, query> (across chunks)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    row_min = gate_min_val.amin(dim=0)        # (H, S)
+    row_max = gate_masked.amax(dim=0)         # (H, S)
+    denom = row_max - row_min
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom) # avoid divide‑by‑zero
+
+    gate_norm = (gate - row_min.unsqueeze(0)) / denom.unsqueeze(0)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) pull out the self‐chunk’s normalized weight for each <head,seq>
+    self_norm = (gate_norm * gate_self_chunk_mask).sum(dim=0)  # (H, S)
+
+    # 2) compute how much more normalized weight we need beyond self
+    total_norm_sum = gate_norm.sum(dim=0) # (H, S)
+    remain_ratio = simsum_threshold - self_norm / (total_norm_sum + eps)  # (H, S)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # if already ≥ thresh, no extra needed
+
+    # 3) zero out the self‐chunk in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0
+
+    # 4) sort the other chunks by descending norm, per <head,seq>
+    sorted_norm, sorted_idx = torch.sort(others_norm, descending=True, dim=0) # (C, H, S)
+
+    # 5) cumulative‑sum the sorted norms per <head,seq>
+    cumsum_others = sorted_norm.cumsum(dim=0) # (C, H, S)
+
+    # 6) for each <head,seq>, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio = cumsum_others / (total_norm_sum.unsqueeze(0) + eps) # (C, H, S)
+    cond = ratio >= remain_ratio.unsqueeze(0) # (C, H, S) boolean mask
+    any_cond = cond.any(dim=0) # (H, S)
+    # Find the index of the first True value along dim 0. If none, use C-1.
+    cutoff = torch.where(any_cond, cond.float().argmax(dim=0), torch.full_like(any_cond, fill_value=C - 1)) # (H, S)
+
+    # 7) build a mask in sorted order up to that cutoff
+    idx_range = torch.arange(C, device=gate.device).view(-1, 1, 1) # (C, 1, 1)
+    sorted_mask = idx_range <= cutoff.unsqueeze(0) # (C, H, S)
+
+    # 8) scatter it back to original chunk order
+    others_mask = torch.zeros_like(gate, dtype=torch.bool)
+    others_mask.scatter_(0, sorted_idx, sorted_mask)
+
+    # 9) finally, include every self‐chunk plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_block(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <query, head> pairs for each block based on threshold.
+    Normalization and sorting happen across the head and sequence dimensions (dim=1, 2).
+    """
+    C, H, S = gate.shape
+    HS = H * S
+    eps = 1e-6
+
+    # LSE‐style normalization per block (across heads and queries)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    block_max = gate_masked.amax(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_min = gate_min_val.amin(dim=(1, 2), keepdim=True) # (C, 1, 1)
+    block_denom = block_max - block_min
+    block_denom = torch.where(block_denom <= eps, torch.ones_like(block_denom), block_denom) # (C, 1, 1)
+
+    gate_norm = (gate - block_min) / block_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks (from query perspective)
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights *per block*
+    self_norm_sum_per_block = self_norm_entries.sum(dim=(1, 2)) # (C,)
+
+    # 2) compute how much more normalized weight each block needs beyond its self-chunk contributions
+    total_norm_sum_per_block = gate_norm.sum(dim=(1, 2)) # (C,)
+    remain_ratio = simsum_threshold - self_norm_sum_per_block / (total_norm_sum_per_block + eps) # (C,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # (C,)
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort the other <head, seq> pairs by descending norm, per block
+    others_flat = others_norm.contiguous().view(C, HS) # (C, H*S)
+    sorted_others_flat, sorted_indices_flat = torch.sort(others_flat, dim=1, descending=True) # (C, H*S)
+
+    # 5) cumulative‑sum the sorted norms per block
+    cumsum_others_flat = sorted_others_flat.cumsum(dim=1) # (C, H*S)
+
+    # 6) for each block, find the smallest k where cumsum_ratio ≥ remain_ratio
+    ratio_flat = cumsum_others_flat / (total_norm_sum_per_block.unsqueeze(1) + eps) # (C, H*S)
+    cond_flat = ratio_flat >= remain_ratio.unsqueeze(1) # (C, H*S) boolean mask
+    any_cond = cond_flat.any(dim=1) # (C,)
+    # Find the index of the first True value along dim 1. If none, use HS-1.
+    cutoff_flat = torch.where(any_cond, cond_flat.float().argmax(dim=1), torch.full_like(any_cond, fill_value=HS - 1)) # (C,)
+
+    # 7) build a mask in sorted order up to that cutoff per block
+    idx_range_flat = torch.arange(HS, device=gate.device).unsqueeze(0) # (1, H*S)
+    sorted_mask_flat = idx_range_flat <= cutoff_flat.unsqueeze(1) # (C, H*S)
+
+    # 8) scatter it back to original <head, seq> order per block
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C, H*S)
+    others_mask_flat.scatter_(1, sorted_indices_flat, sorted_mask_flat)
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_overall(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query, head> triplets globally based on threshold.
+    Normalization and sorting happen across all valid entries.
+    """
+    C, H, S = gate.shape
+    CHS = C * H * S
+    eps = 1e-6
+
+    # LSE‐style normalization globally across all valid entries
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf) # Use -inf for max
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf) # Use +inf for min
+
+    overall_max = gate_masked.max() # scalar
+    overall_min = gate_min_val.min() # scalar
+    overall_denom = overall_max - overall_min
+    overall_denom = torch.where(overall_denom <= eps, torch.tensor(1.0, device=gate.device, dtype=gate.dtype), overall_denom)
+
+    gate_norm = (gate - overall_min) / overall_denom # (C, H, S)
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0) # (C, H, S)
+
+    # 1) identify normalized weights of entries that *are* self-chunks
+    self_norm_entries = gate_norm * gate_self_chunk_mask # (C, H, S)
+    # Sum these weights globally
+    self_norm_sum_overall = self_norm_entries.sum() # scalar
+
+    # 2) compute how much more normalized weight is needed globally beyond self-chunk contributions
+    total_norm_sum_overall = gate_norm.sum() # scalar
+    remain_ratio = simsum_threshold - self_norm_sum_overall / (total_norm_sum_overall + eps) # scalar
+    remain_ratio = torch.clamp(remain_ratio, min=0.0) # scalar
+
+    # 3) zero out the self‐chunk entries in a copy, so we only sort “others”
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0 # Zero out self entries
+
+    # 4) sort all other entries by descending norm, globally
+    others_flat = others_norm.flatten() # (C*H*S,)
+    valid_others_mask_flat = valid_gate_mask.flatten() & ~gate_self_chunk_mask.flatten() # Mask for valid, non-self entries
+    
+    # Only sort the valid 'other' entries
+    valid_others_indices = torch.where(valid_others_mask_flat)[0]
+    valid_others_values = others_flat[valid_others_indices]
+
+    sorted_others_values, sort_perm = torch.sort(valid_others_values, descending=True) # (N_valid_others,)
+    sorted_original_indices = valid_others_indices[sort_perm] # Original indices in C*H*S space, sorted by value
+
+    # 5) cumulative‑sum the sorted valid 'other' norms globally
+    cumsum_others_values = sorted_others_values.cumsum(dim=0) # (N_valid_others,)
+
+    # 6) find the smallest k where cumsum_ratio ≥ remain_ratio globally
+    ratio_values = cumsum_others_values / (total_norm_sum_overall + eps) # (N_valid_others,)
+    cond_values = ratio_values >= remain_ratio # (N_valid_others,) boolean mask
+    any_cond = cond_values.any() # scalar
+    
+    # Find the index of the first True value in the *sorted* list. If none, use all valid others.
+    cutoff_idx_in_sorted = torch.where(
+        any_cond, 
+        cond_values.float().argmax(dim=0), 
+        torch.tensor(len(sorted_others_values) - 1, device=gate.device, dtype=torch.long)
+    )
+
+    # 7) build a mask selecting the top-k others based on the cutoff
+    # Select the original indices corresponding to the top entries in the sorted list
+    selected_other_indices = sorted_original_indices[:cutoff_idx_in_sorted + 1]
+
+    # 8) create the mask in the original flat shape
+    others_mask_flat = torch.zeros_like(others_flat, dtype=torch.bool) # (C*H*S,)
+    if selected_other_indices.numel() > 0: # Check if any 'other' indices were selected
+        others_mask_flat[selected_other_indices] = True
+    others_mask = others_mask_flat.view(C, H, S) # (C, H, S)
+
+    # 9) finally, include every self‐chunk entry plus all selected others
+    final_gate_mask = valid_gate_mask & (others_mask | gate_self_chunk_mask)
+
+    return final_gate_mask
+
+
+def _select_threshold_head_global(
+    gate: torch.Tensor,
+    valid_gate_mask: torch.Tensor,
+    gate_self_chunk_mask: torch.Tensor,
+    simsum_threshold: float
+) -> torch.Tensor:
+    """
+    Selects <chunk, query> globally for each head based on threshold.
+    """
+    C, H, S = gate.shape
+    eps = 1e-6
+
+    # 1) LSE‐style normalization per head (across chunks and sequence dims)
+    gate_masked = torch.where(valid_gate_mask, gate, -torch.inf)
+    gate_min_val = torch.where(valid_gate_mask, gate, torch.inf)
+
+    max_per_head = gate_masked.amax(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    min_per_head = gate_min_val.amin(dim=(0, 2), keepdim=True)  # (1, H, 1)
+    denom = max_per_head - min_per_head
+    denom = torch.where(denom <= eps, torch.ones_like(denom), denom)
+
+    gate_norm = (gate - min_per_head) / denom
+    gate_norm = torch.where(valid_gate_mask, gate_norm, 0.0)  # (C, H, S)
+
+    # 2) sum normalized self‐chunk contributions per head
+    self_norm_sum = (gate_norm * gate_self_chunk_mask).sum(dim=(0, 2))  # (H,)
+
+    # 3) total normalized sum per head
+    total_norm_sum = gate_norm.sum(dim=(0, 2))  # (H,)
+
+    # 4) how much more normalized weight needed per head
+    remain_ratio = simsum_threshold - self_norm_sum / (total_norm_sum + eps)  # (H,)
+    remain_ratio = torch.clamp(remain_ratio, min=0.0)
+
+    # 5) zero out self‐chunk entries to focus on "others"
+    others_norm = gate_norm.clone()
+    others_norm[gate_self_chunk_mask] = 0.0  # (C, H, S)
+
+    # 6) flatten chunk and sequence dims, per head
+    CS = C * S
+    others_flat = others_norm.permute(1, 0, 2).reshape(H, CS)  # (H, C*S)
+    valid_flat  = (valid_gate_mask & ~gate_self_chunk_mask) \
+                   .permute(1, 0, 2).reshape(H, CS)           # (H, C*S)
+
+    # 7) vectorized selection of “others” per head
+    masked_flat = torch.where(valid_flat, others_flat, torch.zeros_like(others_flat))
+    sorted_vals, sorted_idx = torch.sort(masked_flat, dim=1, descending=True)  # (H, C*S)
+
+    cumsum_vals = sorted_vals.cumsum(dim=1)                                    # (H, C*S)
+    ratio_vals  = cumsum_vals / (total_norm_sum.unsqueeze(1) + eps)            # (H, C*S)
+    cond        = ratio_vals >= remain_ratio.unsqueeze(1)                      # (H, C*S)
+
+    has_cutoff = cond.any(dim=1)                                               # (H,)
+    default    = torch.full((H,), CS - 1, device=gate.device, dtype=torch.long)
+    cutoff     = torch.where(has_cutoff, cond.float().argmax(dim=1), default)  # (H,)
+
+    idx_range   = torch.arange(CS, device=gate.device).unsqueeze(0)            # (1, C*S)
+    sorted_mask = idx_range <= cutoff.unsqueeze(1)                             # (H, C*S)
+
+    selected_flat = torch.zeros_like(valid_flat)                               # (H, C*S)
+    selected_flat.scatter_(1, sorted_idx, sorted_mask)                         # (H, C*S)
+
+    # 8) reshape selection mask back to (C, H, S)
+    others_mask = selected_flat.reshape(H, C, S).permute(1, 0, 2)               # (C, H, S)
+
+    # 9) include self‐chunks plus selected others, and obey valid mask
+    final_gate_mask = valid_gate_mask & (gate_self_chunk_mask | others_mask)
+
+    return final_gate_mask
+
+
+class MixedAttention(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    ):
+        ctx.max_seqlen = max_seqlen
+        ctx.moba_chunk_size = moba_chunk_size
+        ctx.softmax_scale = softmax_scale = q.shape[-1] ** (-0.5)
+
+        # Non-causal self-attention branch
+        # return out, softmax_lse, S_dmask, rng_state
+        self_attn_out_sh, self_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=q,
+            k=k,
+            v=v,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+        # MOBA attention branch (non-causal)
+        moba_attn_out, moba_attn_lse_hs, _, _ = _flash_attn_varlen_forward(
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+        )
+
+        self_attn_lse_sh = self_attn_lse_hs.t().contiguous()
+        moba_attn_lse = moba_attn_lse_hs.t().contiguous()
+
+        output = torch.zeros((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+        output_2d = output.view(-1, q.shape[2])
+
+        max_lse_1d = self_attn_lse_sh.view(-1)
+        max_lse_1d = max_lse_1d.index_reduce(
+            0, moba_q_sh_indices, moba_attn_lse.view(-1), "amax"
+        )
+        self_attn_lse_sh = self_attn_lse_sh - max_lse_1d.view_as(self_attn_lse_sh)
+        moba_attn_lse = (
+            moba_attn_lse.view(-1)
+            .sub(max_lse_1d.index_select(0, moba_q_sh_indices))
+            .reshape_as(moba_attn_lse)
+        )
+
+        mixed_attn_se_sh = self_attn_lse_sh.exp()
+        moba_attn_se = moba_attn_lse.exp()
+
+        mixed_attn_se_sh.view(-1).index_add_(
+            0, moba_q_sh_indices, moba_attn_se.view(-1)
+        )
+        mixed_attn_lse_sh = mixed_attn_se_sh.log()
+
+        # Combine self-attention output
+        factor = (self_attn_lse_sh - mixed_attn_lse_sh).exp()  # [S, H]
+        self_attn_out_sh = self_attn_out_sh * factor.unsqueeze(-1)
+        output_2d += self_attn_out_sh.reshape_as(output_2d)
+
+        # Combine MOBA attention output
+        mixed_attn_lse = (
+            mixed_attn_lse_sh.view(-1)
+            .index_select(0, moba_q_sh_indices)
+            .view_as(moba_attn_lse)
+        )
+        factor = (moba_attn_lse - mixed_attn_lse).exp()  # [S, H]
+        moba_attn_out = moba_attn_out * factor.unsqueeze(-1)
+        raw_attn_out = moba_attn_out.view(-1, moba_attn_out.shape[-1])
+        output_2d.index_add_(0, moba_q_sh_indices, raw_attn_out)
+        output = output.to(q.dtype)
+        mixed_attn_lse_sh = mixed_attn_lse_sh + max_lse_1d.view_as(mixed_attn_se_sh)
+        ctx.save_for_backward(
+            output,
+            mixed_attn_lse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        )
+
+        return output
+
+    @staticmethod
+    def backward(ctx, d_output):
+
+        max_seqlen = ctx.max_seqlen
+        moba_chunk_size = ctx.moba_chunk_size
+        softmax_scale = ctx.softmax_scale
+
+        (
+            output,
+            mixed_attn_vlse_sh,
+            q,
+            k,
+            v,
+            self_attn_cu_seqlen,
+            moba_q,
+            moba_kv,
+            moba_cu_seqlen_q,
+            moba_cu_seqlen_kv,
+            moba_q_sh_indices,
+        ) = ctx.saved_tensors
+
+        d_output = d_output.contiguous()
+
+        dq = torch.empty_like(q)
+        dk = torch.empty_like(k)
+        dv = torch.empty_like(v)
+        _ = _flash_attn_varlen_backward(
+            dout=d_output,
+            q=q,
+            k=k,
+            v=v,
+            out=output,
+            softmax_lse=mixed_attn_vlse_sh.t().contiguous(),
+            dq=dq,
+            dk=dk,
+            dv=dv,
+            cu_seqlens_q=self_attn_cu_seqlen,
+            cu_seqlens_k=self_attn_cu_seqlen,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        headdim = q.shape[-1]
+        d_moba_output = (
+            d_output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+        moba_output = (
+            output.view(-1, headdim).index_select(0, moba_q_sh_indices).unsqueeze(1)
+        )
+
+        mixed_attn_vlse = (
+            mixed_attn_vlse_sh.view(-1).index_select(0, moba_q_sh_indices).view(1, -1)
+        )
+
+        dmq = torch.empty_like(moba_q)
+        dmkv = torch.empty_like(moba_kv)
+        _ = _flash_attn_varlen_backward(
+            dout=d_moba_output,
+            q=moba_q,
+            k=moba_kv[:, 0],
+            v=moba_kv[:, 1],
+            out=moba_output,
+            softmax_lse=mixed_attn_vlse,
+            dq=dmq,
+            dk=dmkv[:,0],
+            dv=dmkv[:,1],
+            cu_seqlens_q=moba_cu_seqlen_q,
+            cu_seqlens_k=moba_cu_seqlen_kv,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=moba_chunk_size,
+            softmax_scale=softmax_scale,
+            causal=False,
+            dropout_p=0.0,
+            softcap=0.0,
+            alibi_slopes=None,
+            deterministic=True,
+            window_size_left=-1,
+            window_size_right=-1
+        )
+
+        return dq, dk, dv, None, dmq, dmkv, None, None, None, None, None
+
+
+def moba_attn_varlen(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens: torch.Tensor,
+    max_seqlen: int,
+    moba_chunk_size: int,
+    moba_topk: int,
+    select_mode: str = 'threshold', # "topk" or "threshold"
+    simsum_threshold: float = 0.25,
+    threshold_type: str = 'query_head',
+) -> torch.Tensor:
+    """
+    Accelerated MOBA attention for vision tasks with proper LSE normalization.
+
+    This version:
+      - Splits KV into chunks.
+      - For each query head, selects the top-k relevant KV chunks (including the self chunk)
+        by amplifying the diagonal (self-chunk) logits.
+      - Aggregates the attention outputs from the selected chunks using a log-sum-exp
+        reduction so that attending to each query over the selected chunks is equivalent
+        to the original algorithm.
+    """
+    # Stack keys and values.
+    kv = torch.stack((k, v), dim=1)
+    seqlen, num_head, head_dim = q.shape
+
+    # Compute chunk boundaries.
+    cu_chunk, filtered_chunk_indices, num_filtered_chunk, chunk_to_batch = calc_chunks(
+        cu_seqlens, moba_chunk_size
+    )
+
+    self_attn_cu_seqlen = cu_chunk
+
+    # Update top-k selection to include the self chunk.
+    moba_topk = min(moba_topk, num_filtered_chunk)
+
+    # --- Build filtered KV from chunks ---
+    chunk_starts = cu_chunk[filtered_chunk_indices]          # [num_filtered_chunk]
+    chunk_ends = cu_chunk[filtered_chunk_indices + 1]          # [num_filtered_chunk]
+    chunk_lengths = chunk_ends - chunk_starts                  # [num_filtered_chunk]
+    max_chunk_len = int(chunk_lengths.max().item())
+
+    range_tensor = torch.arange(max_chunk_len, device=kv.device, dtype=chunk_starts.dtype).unsqueeze(0)
+    indices = chunk_starts.unsqueeze(1) + range_tensor
+    indices = torch.clamp(indices, max=kv.shape[0] - 1)
+    valid_mask = range_tensor < chunk_lengths.unsqueeze(1)
+    gathered = kv[indices.view(-1)].view(num_filtered_chunk, max_chunk_len, *kv.shape[1:])
+    gathered = gathered * valid_mask.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1).type_as(gathered)
+
+    # Compute key_gate_weight over valid tokens.
+    key_values = gathered[:, :, 0].float()  # [num_filtered_chunk, max_chunk_len, num_head, head_dim]
+    valid_mask_exp = valid_mask.unsqueeze(-1).unsqueeze(-1)
+    key_sum = (key_values * valid_mask_exp).sum(dim=1)
+    divisor = valid_mask.sum(dim=1).unsqueeze(-1).unsqueeze(-1)
+    key_gate_weight = key_sum / divisor  # [num_filtered_chunk, num_head, head_dim]
+
+    # Compute gate logits between key_gate_weight and queries.
+    q_float = q.float()
+    # gate = torch.einsum("nhd,shd->nhs", key_gate_weight, q_float)  # [num_filtered_chunk, num_head, seqlen]
+    gate = torch.bmm(key_gate_weight.permute(1, 0, 2), q_float.permute(1, 0, 2).transpose(1, 2)).permute(1, 0, 2)
+
+    # Amplify the diagonal (self chunk) contributions.
+    gate_seq_idx = torch.arange(seqlen, device=q.device, dtype=torch.int32).unsqueeze(0).expand(num_filtered_chunk, seqlen)
+    chunk_start = cu_chunk[filtered_chunk_indices]  # [num_filtered_chunk]
+    chunk_end = cu_chunk[filtered_chunk_indices + 1]  # [num_filtered_chunk]
+    gate_self_chunk_mask = ((gate_seq_idx >= chunk_start.unsqueeze(1)) &
+                            (gate_seq_idx < chunk_end.unsqueeze(1))).unsqueeze(1).expand(-1, num_head, -1)
+    amplification_factor = 1e9  # Example factor; adjust as needed.
+    origin_gate = gate.clone()
+    gate = gate.clone()
+    if select_mode == "topk":
+        gate[gate_self_chunk_mask] += amplification_factor
+
+    # Exclude positions that are outside the valid batch boundaries.
+    batch_starts = cu_seqlens[chunk_to_batch[filtered_chunk_indices]]
+    batch_ends = cu_seqlens[chunk_to_batch[filtered_chunk_indices] + 1]
+    gate_batch_start_mask = gate_seq_idx < batch_starts.unsqueeze(1)
+    gate_batch_end_mask = gate_seq_idx >= batch_ends.unsqueeze(1)
+    gate_inf_mask = gate_batch_start_mask | gate_batch_end_mask
+    gate.masked_fill_(gate_inf_mask.unsqueeze(1), -float("inf"))
+
+    if select_mode == 'topk':
+        # We amplify self‐chunk in gate already, so self entries will rank highest.
+        valid_gate_mask = gate != -float("inf")
+        if threshold_type == 'query_head':
+            # === per‐<head,seq> top-k across chunks (original behavior) ===
+            # gate: (C, H, S)
+            _, gate_topk_idx = torch.topk(gate, k=moba_topk, dim=0, largest=True, sorted=False)
+            gate_idx_mask = torch.zeros_like(gate, dtype=torch.bool)
+            gate_idx_mask.scatter_(0, gate_topk_idx, True)
+            gate_mask = valid_gate_mask & gate_idx_mask
+        elif threshold_type == 'overall':
+            # === global top-k across all (chunk, head, seq) entries ===
+            C, H, S = gate.shape
+            flat_gate = gate.flatten()
+            flat_mask = valid_gate_mask.flatten()
+            flat_gate_masked = torch.where(flat_mask, flat_gate, -float("inf"))
+            # pick topk global entries
+            vals, idx = torch.topk(flat_gate_masked, k=moba_topk * H * S, largest=True, sorted=False)
+            others_mask_flat = torch.zeros_like(flat_mask, dtype=torch.bool)
+            others_mask_flat[idx] = True
+            gate_mask = (valid_gate_mask.flatten() & others_mask_flat).view(gate.shape)
+        elif threshold_type == 'head_global':
+            # per-head top-k across all chunks and sequence positions
+            C, H, S = gate.shape
+            CS = C * S
+            flat_gate  = gate.permute(1, 0, 2).reshape(H, CS)
+            flat_valid = valid_gate_mask.permute(1, 0, 2).reshape(H, CS)
+            flat_gate_masked = torch.where(flat_valid, flat_gate, torch.full_like(flat_gate, -float('inf')))
+            # pick top-k indices per head
+            _, topk_idx = torch.topk(flat_gate_masked, k=moba_topk * S, dim=1, largest=True, sorted=False)
+            gate_idx_flat = torch.zeros_like(flat_valid, dtype=torch.bool)
+            gate_idx_flat.scatter_(1, topk_idx, True)
+            gate_mask = gate_idx_flat.reshape(H, C, S).permute(1, 0, 2)
+        else:
+            raise ValueError(
+                f"Invalid threshold_type for topk: {threshold_type}. "
+                "Choose 'query_head', 'block', or 'overall'."
+            )
+    elif select_mode == 'threshold':
+        # Delegate to the specific thresholding function
+        valid_gate_mask = gate != -float("inf")  # (num_chunk, num_head, seqlen)
+        if threshold_type == 'query_head':
+            gate_mask = _select_threshold_query_head(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'block':
+            gate_mask = _select_threshold_block(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'overall':
+            gate_mask = _select_threshold_overall(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        elif threshold_type == 'head_global':
+            gate_mask = _select_threshold_head_global(gate, valid_gate_mask, gate_self_chunk_mask, simsum_threshold)
+        else:
+            raise ValueError(f"Invalid threshold_type: {threshold_type}. Choose 'query_head', 'block', or 'overall'.")
+    else:
+        raise ValueError(f"Invalid select_mode: {select_mode}. Choose 'topk' or 'threshold'.")
+    
+    # eliminate self_chunk in MoBA branch
+    gate_mask = gate_mask & ~gate_self_chunk_mask
+    # if gate_mask is all false, perform flash_attn instead
+    if gate_mask.sum() == 0:
+        return flash_attn_varlen_func(
+            q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen, causal=False
+        )
+
+    # Determine which query positions are selected.
+    # nonzero_indices has shape [N, 3] where each row is [chunk_index, head_index, seq_index].
+    moba_q_indices = gate_mask.reshape(gate_mask.shape[0], -1).nonzero(as_tuple=True)[-1] # [(h s k)]
+    moba_q_sh_indices = (moba_q_indices % seqlen) * num_head + (moba_q_indices // seqlen)
+    moba_q = rearrange(q, "s h d -> (h s) d").index_select(0, moba_q_indices).unsqueeze(1)
+
+    # Build cumulative sequence lengths for the selected queries.
+    moba_seqlen_q = gate_mask.sum(dim=-1).flatten()
+    q_zero_mask = moba_seqlen_q == 0
+    valid_expert_mask = ~q_zero_mask
+    if q_zero_mask.sum() > 0:
+        moba_seqlen_q = moba_seqlen_q[valid_expert_mask]
+    moba_cu_seqlen_q = torch.cat(
+        (
+            torch.tensor([0], device=q.device, dtype=moba_seqlen_q.dtype),
+            moba_seqlen_q.cumsum(dim=0),
+        ),
+        dim=0,
+    ).to(torch.int32)
+
+    # Rearrange gathered KV for the MOBA branch.
+    experts_tensor = rearrange(gathered, "nc cl two h d -> (nc h) cl two d")
+    valid_expert_lengths = chunk_lengths.unsqueeze(1).expand(num_filtered_chunk, num_head).reshape(-1).to(torch.int32)
+    if q_zero_mask.sum() > 0:
+        experts_tensor = experts_tensor[valid_expert_mask]
+        valid_expert_lengths = valid_expert_lengths[valid_expert_mask]
+
+    seq_range = torch.arange(experts_tensor.shape[1], device=experts_tensor.device).unsqueeze(0)
+    mask = seq_range < valid_expert_lengths.unsqueeze(1)
+    moba_kv = experts_tensor[mask]  # Shape: ((nc h cl_valid) two d)
+    moba_kv = moba_kv.unsqueeze(2)   # Shape: ((nc h cl_valid) two 1 d)
+
+    moba_cu_seqlen_kv = torch.cat(
+        [torch.zeros(1, device=experts_tensor.device, dtype=torch.int32),
+         valid_expert_lengths.cumsum(dim=0)],
+        dim=0,
+    ).to(torch.int32)
+
+    assert (
+        moba_cu_seqlen_kv.shape == moba_cu_seqlen_q.shape
+    ), f"Mismatch between moba_cu_seqlen_kv.shape and moba_cu_seqlen_q.shape: {moba_cu_seqlen_kv.shape} vs {moba_cu_seqlen_q.shape}"
+
+    return MixedAttention.apply(
+        q,
+        k,
+        v,
+        self_attn_cu_seqlen,
+        moba_q,
+        moba_kv,
+        moba_cu_seqlen_q,
+        moba_cu_seqlen_kv,
+        max_seqlen,
+        moba_chunk_size,
+        moba_q_sh_indices,
+    )
+
+
+def process_moba_input(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    """
+    Process inputs for the attention function.
+
+    Args:
+        x (torch.Tensor): Input tensor with shape [batch_size, num_patches, num_heads, head_dim].
+        patch_resolution (tuple): Tuple containing the patch resolution (t, h, w).
+        chunk_size (int): Size of the chunk. (maybe tuple or int, according to chunk type)
+
+    Returns:
+        torch.Tensor: Processed input tensor.
+    """
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        moba_chunk_size = int(chunk_size * patch_resolution[1] * patch_resolution[2])
+    else:
+        assert isinstance(chunk_size, (Tuple, list)), f"chunk_size should be a tuple, list, or int, now it is: {type(chunk_size)}"
+        if len(chunk_size) == 2:
+            assert patch_resolution[1] % chunk_size[0] == 0 and patch_resolution[2] % chunk_size[1] == 0, f"spatial patch_resolution {patch_resolution[1:]} should be divisible by 2d chunk_size {chunk_size}"
+            nch, ncw = patch_resolution[1] // chunk_size[0], patch_resolution[2] // chunk_size[1]
+            x = rearrange(x, "b (t nch ch ncw cw) n d -> b (nch ncw t ch cw) n d", t=patch_resolution[0], nch=nch, ncw=ncw, ch=chunk_size[0], cw=chunk_size[1])
+            moba_chunk_size = patch_resolution[0] * chunk_size[0] * chunk_size[1]
+        elif len(chunk_size) == 3:
+            assert patch_resolution[0] % chunk_size[0] == 0 and patch_resolution[1] % chunk_size[1] == 0 and patch_resolution[2] % chunk_size[2] == 0, f"patch_resolution {patch_resolution} should be divisible by 3d chunk_size {chunk_size}"
+            nct, nch, ncw = patch_resolution[0] // chunk_size[0], patch_resolution[1] // chunk_size[1], patch_resolution[2] // chunk_size[2]
+            x = rearrange(x, "b (nct ct nch ch ncw cw) n d -> b (nct nch ncw ct ch cw) n d", nct=nct, nch=nch, ncw=ncw, ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+            moba_chunk_size = chunk_size[0] * chunk_size[1] * chunk_size[2]
+        else:
+            raise ValueError(f"chunk_size should be a int, or a tuple of length 2 or 3, now it is: {len(chunk_size)}")
+        
+    return x, moba_chunk_size
+
+
+def process_moba_output(
+    x,
+    patch_resolution,
+    chunk_size,
+):
+    if isinstance(chunk_size, float) or isinstance(chunk_size, int):
+        pass
+    elif len(chunk_size) == 2:
+        x = rearrange(x, "b (nch ncw t ch cw) n d -> b (t nch ch ncw cw) n d", nch=patch_resolution[1] // chunk_size[0], ncw=patch_resolution[2] // chunk_size[1], t=patch_resolution[0], ch=chunk_size[0], cw=chunk_size[1])
+    elif len(chunk_size) == 3:
+        x = rearrange(x, "b (nct nch ncw ct ch cw) n d -> b (nct ct nch ch ncw cw) n d", nct=patch_resolution[0] // chunk_size[0], nch=patch_resolution[1] // chunk_size[1], ncw=patch_resolution[2] // chunk_size[2], ct=chunk_size[0], ch=chunk_size[1], cw=chunk_size[2])
+
+    return x
+
+
+# TEST
+def generate_data(batch_size, seqlen, num_head, head_dim, dtype):
+    random.seed(0)
+    torch.manual_seed(0)
+    torch.cuda.manual_seed(0)
+    device = torch.cuda.current_device()
+
+    q = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    k = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    v = torch.randn((batch_size, seqlen, num_head, head_dim), requires_grad=True).to(dtype=dtype, device='cuda')
+    print(f"q.shape: {q.shape}, k.shape: {k.shape}, v.shape: {v.shape}")
+    cu_seqlens = torch.arange(0, q.shape[0] * q.shape[1] + 1, q.shape[1], dtype=torch.int32, device='cuda')
+    max_seqlen = q.shape[1]
+    q = rearrange(q, "b s ... -> (b s) ...")
+    k = rearrange(k, "b s ... -> (b s) ...")
+    v = rearrange(v, "b s ... -> (b s) ...")
+
+    return q, k, v, cu_seqlens, max_seqlen
+
+
+def test_attn_varlen_moba_speed(batch, head, seqlen, head_dim, moba_chunk_size, moba_topk, dtype=torch.bfloat16, select_mode='threshold', simsum_threshold=0.25, threshold_type='query_head'):
+    """Speed test comparing flash_attn vs moba_attention"""
+    # Get data
+    q, k, v, cu_seqlen, max_seqlen = generate_data(batch, seqlen, head, head_dim, dtype)
+    print(f"batch:{batch} head:{head} seqlen:{seqlen} chunk:{moba_chunk_size} topk:{moba_topk} select_mode: {select_mode} simsum_threshold:{simsum_threshold}")
+    vo_grad = torch.randn_like(q)
+    
+    # Warmup
+    warmup_iters = 3
+    perf_test_iters = 10
+
+    # Warmup
+    for _ in range(warmup_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+    
+    torch.cuda.synchronize()
+    start_flash = time.perf_counter()
+    for _ in range(perf_test_iters):
+        o = flash_attn_varlen_func(q, k, v, cu_seqlen, cu_seqlen, max_seqlen, max_seqlen, causal=False)
+        torch.autograd.backward(o, vo_grad)
+        
+    torch.cuda.synchronize()
+    time_flash = (time.perf_counter() - start_flash) / perf_test_iters * 1000
+
+    # Warmup
+    for _ in range(warmup_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+
+    torch.cuda.synchronize()
+    start_moba = time.perf_counter()
+    for _ in range(perf_test_iters):
+        om = moba_attn_varlen(q, k, v, cu_seqlen, max_seqlen, moba_chunk_size=moba_chunk_size, moba_topk=moba_topk, select_mode=select_mode, simsum_threshold=simsum_threshold, threshold_type=threshold_type)
+        torch.autograd.backward(om, vo_grad)
+    
+    torch.cuda.synchronize()
+    time_moba = (time.perf_counter() - start_moba) / perf_test_iters * 1000
+    
+    print(f"Flash: {time_flash:.2f}ms, MoBA: {time_moba:.2f}ms")
+    print(f"Speedup:  {time_flash / time_moba:.2f}x")
+
+
+
+if __name__ == "__main__":
+    """
+    CUDA_VISIBLE_DEVICES=1 \
+    python -u csrc/attn/vmoba_attn/vmoba/vmoba.py
+    """
+    test_attn_varlen_moba_speed(batch=1, head=12, seqlen=32760, head_dim=128, moba_chunk_size=32760 // 3 // 6 // 4, moba_topk=3, select_mode='threshold', simsum_threshold=0.3, threshold_type='query_head')

docs/source/inference/optimizations.md

@@ -1,32 +1,40 @@
 (inference-optimizations)=
+
 # Optimizations
 
 This page describes the various options for speeding up generation times in FastVideo.
 
 ## Table of Contents
+
 - Optimized Attention Backends
+
   - [Flash Attention](#optimizations-flash)
   - [Sliding Tile Attention](#optimizations-sta)
   - [Sage Attention](#optimizations-sage)
+  - [Sage Attention 3](#optimizations-sage3)
 
 - Caching Techniques
   - [TeaCache](#optimizations-teacache)
 
 (optimizations-backends)=
+
 ## Attention Backends
 
 ### Available Backends
+
 - Torch SDPA: `FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA`
 - Flash Attention 2 and 3: `FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN`
 - Sliding Tile Attention: `FASTVIDEO_ATTENTION_BACKEND=SLIDING_TILE_ATTN`
 - Video Sparse Attention: `FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN`
 - Sage Attention: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN`
+- Sage Attention 3: `FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN_THREE`
 
 ### Configuring Backends
 
 There are two ways to configure the attention backend in FastVideo.
 
 #### 1. In Python
+
 In python, set the `FASTVIDEO_ATTENTION_BACKEND` environment variable before instantiating `VideoGenerator` like this:
 
 ```python
@@ -34,6 +42,7 @@ os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "SLIDING_TILE_ATTN"
 ```
 
 #### 2. In CLI
+
 You can also set the environment variable on the command line:
 
 ```bash
@@ -41,6 +50,7 @@ FASTVIDEO_ATTENTION_BACKEND=SAGE_ATTN python example.py
 ```
 
 (optimizations-flash)=
+
 ### Flash Attention
 
 **`FLASH_ATTN`**
@@ -57,7 +67,7 @@ And if using a Hopper+ GPU (ie H100), installing [Flash Attention 3](https://git
 git clone https://github.com/Dao-AILab/flash-attention.git && cd flash-attention
 
 cd hopper
-pip install ninja 
+pip install ninja
 python setup.py install
 ```
 
@@ -66,7 +76,9 @@ FastVideo will automatically detect and use `FA3` if it is installed when using
 :::
 
 (optimizations-sta)=
+
 ### Sliding Tile Attention
+
 **`SLIDING_TILE_ATTN`**
 
 ```bash
@@ -76,7 +88,9 @@ pip install st_attn==0.0.4
 Please see [this page](#sta-installation) for more installation instructions.
 
 (optimizations-vsa)=
+
 ### Video Sparse Attention
+
 **`VIDEO_SPARSE_ATTN`**
 
 ```bash
@@ -87,19 +101,45 @@ python setup_vsa.py install
 Please see [this page](#vsa-installation) for more installation instructions.
 
 (optimizations-sage)=
+
 ### Sage Attention
+
 **`SAGE_ATTN`**
 
 To use [SageAttention](https://github.com/thu-ml/SageAttention) 2.1.1, please compile from source:
 
 ```bash
 git clone https://github.com/thu-ml/SageAttention.git
-cd sageattention 
+cd sageattention
 python setup.py install  # or pip install -e .
 ```
 
+(optimizations-sage3)=
+
+### Sage Attention 3
+
+**`SAGE_ATTN_THREE`**
+
+[SageAttention 3](https://huggingface.co/jt-zhang/SageAttention3) is an advanced attention mechanism that leverages FP4 quantization and Blackwell GPU Tensor Cores for significant performance improvements.
+
+#### Hardware Requirements
+
+- RTX5090
+
+#### Installation
+
+Note that Sage Attention 3 requires `python>=3.13`, `torch>=2.8.0`, `CUDA >=12.8`. If you are using `uv` and using `torch==2.8.0` make sure that `sentencepiece==0.2.1` in the pyproject.toml file.
+
+To use Sage Attention 3 in FastVideo, first get access to the SageAttention3 code, then move `sageattn/` and `setup.py` to the directory `fastvideo/attention/backends`, then install from using:
+
+```bash
+python setup.py install
+```
+
 (optimizations-teacache)=
+
 ## Teacache
+
 TeaCache is an optimization technique supported in FastVideo that can significantly speed up video generation by skipping redundant calculations across diffusion steps. This guide explains how to enable and configure TeaCache for optimal performance in FastVideo.
 
 ### What is TeaCache?

examples/dataset/vidprom/README.md

@@ -0,0 +1,9 @@
+# VidProm Dataset
+
+From [Self-Forcing](https://github.com/gdhe17/Self-Forcing) repository.
+
+## Download the dataset
+
+```bash
+./download_dataset.sh
+```

examples/dataset/vidprom/download_dataset.sh

@@ -0,0 +1,3 @@
+#! /bin/bash
+
+huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts

examples/distill/SFWan2.1-I2V/download_single_example.sh

@@ -1,43 +0,0 @@
-#!/bin/bash
-
-# Download the full dataset first
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
-
-# Create a single-example dataset for debugging
-SINGLE_EXAMPLE_DIR="data/crush-smol-single"
-mkdir -p "$SINGLE_EXAMPLE_DIR/videos"
-
-# Copy the specific video that matches the validation.json style (macaron crushing)
-cp "data/crush-smol/videos/7P02AihYkCU-Scene-005.mp4" "$SINGLE_EXAMPLE_DIR/videos/"
-
-# Create a single-line videos.txt
-echo "videos/7P02AihYkCU-Scene-005.mp4" > "$SINGLE_EXAMPLE_DIR/videos.txt"
-
-# Create a single-line prompt.txt with the macaron crushing prompt
-echo "PIKA_CRUSH A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out." > "$SINGLE_EXAMPLE_DIR/prompt.txt"
-
-# Generate the JSON file and merge.txt for the single example
-python scripts/dataset_preparation/prepare_json_file.py --data_folder "$SINGLE_EXAMPLE_DIR" --output "videos2caption.json"
-
-# Create a validation.json that uses the same example for consistency
-cat > "$SINGLE_EXAMPLE_DIR/validation.json" << 'EOF'
-{
-    "data": [
-      {
-        "caption": "A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out.",
-        "image_path": null,
-        "video_path": null,
-        "num_inference_steps": 50,
-        "height": 480,
-        "width": 832,
-        "num_frames": 81
-      }
-    ]
-  }
-EOF
-
-echo "Single example dataset created at $SINGLE_EXAMPLE_DIR"
-echo "Contains:"
-echo "- 1 video: $(cat $SINGLE_EXAMPLE_DIR/videos.txt)"
-echo "- 1 prompt: $(cat $SINGLE_EXAMPLE_DIR/prompt.txt)"
-echo "- Validation file created with the same example for consistency" 

examples/distill/SFWan2.1-I2V/validation.json

@@ -1,31 +0,0 @@
-{
-    "data": [
-      {
-        "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
-        "image_path": null,
-        "video_path": null,
-        "num_inference_steps": 50,
-        "height": 480,
-        "width": 832,
-        "num_frames": 77
-      },
-      {
-        "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
-        "image_path": null,
-        "video_path": null,
-        "num_inference_steps": 50,
-        "height": 480,
-        "width": 832,
-        "num_frames": 77
-      },
-      {
-        "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
-        "image_path": null,
-        "video_path": null,
-        "num_inference_steps": 50,
-        "height": 480,
-        "width": 832,
-        "num_frames": 77
-      }
-    ]
-  }

examples/distill/SFWan2.1-T2V/distill_dmd_t2v_1.3B.sh

@@ -10,72 +10,59 @@
 #SBATCH --output=dmd_t2v_output/t2v_%j.out
 #SBATCH --error=dmd_t2v_output/t2v_%j.err
 #SBATCH --exclusive
-set -e -x
-
-# Environment Setup
-source ~/conda/miniconda/bin/activate
-conda activate wei-fv
 
 # Basic Info
-export WANDB_MODE="online"
 export NCCL_P2P_DISABLE=1
 export TORCH_NCCL_ENABLE_MONITORING=0
 # different cache dir for different processes
 export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
-export MASTER_PORT=29500
-export NODE_RANK=$SLURM_PROCID
-nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
-export MASTER_ADDR=${nodes[0]}
-export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export MASTER_PORT=29503
 export TOKENIZERS_PARALLELISM=false
+export WANDB_API_KEY=your_wandb_api_key
 export WANDB_BASE_URL="https://api.wandb.ai"
 export WANDB_MODE=online
 export FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN
-# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
-
-echo "MASTER_ADDR: $MASTER_ADDR"
-echo "NODE_RANK: $NODE_RANK"
 
 # Configs
 NUM_GPUS=1
 
-# Model paths for DMD distillation:
+# Model paths for Self-Forcing DMD distillation:
 GENERATOR_MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
 REAL_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"  # Teacher model
 FAKE_SCORE_MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"  # Critic model
 
-DATA_DIR="data/crush-smol-single_processed_t2v/combined_parquet_dataset/"
-VALIDATION_DATASET_FILE="data/crush-smol-single_processed_t2v/validation.json"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_data_dir
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
-  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd
-  --output_dir "checkpoints/SFwan_t2v_finetune"
-  --max_train_steps 500
+  --tracker_project_name SFwan_t2v_distill_self_forcing_dmd  # Updated for self-forcing DMD
+  --output_dir your_output_dir
+  --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
   --gradient_accumulation_steps 1
   --num_latent_t 21
   --num_height 480
   --num_width 832
-  --num_frames 81
+  --num_frames 81  # Must be divisible by num_frame_per_block (81 % 3 = 0 ✓)
   --enable_gradient_checkpointing_type "full"
   --log_visualization
   --simulate_generator_forward
-  --num_frame_per_block 3
+  --num_frame_per_block 3  # Frame generation block size for self-forcing
   --enable_gradient_masking
   --gradient_mask_last_n_frames 21
 )
 
 # Parallel arguments
 parallel_args=(
-  --num_gpus 1 # 64
+  --num_gpus $NUM_GPUS # 64
   --sp_size 1
   --tp_size 1
   --hsdp_replicate_dim 1 # 64
-  --hsdp_shard_dim 1
+  --hsdp_shard_dim $NUM_GPUS
 )
 
 # Model arguments
@@ -97,7 +84,7 @@ dataset_args=(
 validation_args=(
   --log_validation
   --validation_dataset_file "$VALIDATION_DATASET_FILE"
-  --validation_steps 10
+  --validation_steps 50
   --validation_sampling_steps "4"
   --validation_guidance_scale "6.0" # not used for dmd inference
 )
@@ -106,8 +93,8 @@ validation_args=(
 optimizer_args=(
   --learning_rate 1e-5
   --mixed_precision "bf16"
-  --training_state_checkpointing_steps 50
-  --weight_only_checkpointing_steps 50
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
   --weight_decay 0.01
   --betas '0.0,0.999'
   --max_grad_norm 1.0
@@ -124,7 +111,7 @@ miscellaneous_args=(
   --use_ema True
   --ema_decay 0.99
   --ema_start_step 100
-  --init_weights_from_safetensors "/mnt/weka/home/hao.zhang/wl/Self-Forcing/diffusers_ode_init/model.safetensors"
+  --init_weights_from_safetensors your_ode_init_weights_path
 )
 
 # Self-forcing DMD arguments
@@ -136,14 +123,22 @@ dmd_args=(
   --real_score_guidance_scale 3.0
   --fake_score_learning_rate 8e-6
   --fake_score_betas '0.0,0.999'
+  --warp_denoising_step
 )
 
-srun torchrun \
---nnodes $SLURM_JOB_NUM_NODES \
+# Self-forcing specific arguments
+self_forcing_args=(
+  --independent_first_frame False  # Whether to treat first frame independently
+  --same_step_across_blocks True  # Whether to use same denoising step across all blocks
+  --last_step_only False  # Whether to only use the last denoising step
+  --context_noise 0  # Amount of noise to add during context caching (0 = no noise)
+  --validate_cache_structure False  # Set to True for debugging KV cache issues
+)
+
+torchrun \
+--nnodes 1 \
+--master_port $MASTER_PORT \
 --nproc_per_node $NUM_GPUS \
---node_rank $SLURM_PROCID \
---rdzv_backend=c10d \
---rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
     fastvideo/training/wan_self_forcing_distillation_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
@@ -152,4 +147,5 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}" \
+    "${self_forcing_args[@]}"

examples/distill/SFWan2.1-T2V/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/distill/SFWan2.1-T2V/preprocess_data.sh

@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_VSA_t2v_1.3B.slurm

@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir"checkpoints/wan_t2v_finetune"
+  --output_dir $OUTPUT_DIR
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -134,4 +135,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_VSA_t2v_14B.slurm

@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-14B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -134,4 +135,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_t2v_1.3B.slurm

@@ -41,13 +41,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DATASET_FILE=your_validation_dataset_file
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd
-  --output_dir "checkpoints/wan_t2v_finetune"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -91,7 +92,7 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 1e-5
+  --learning_rate 2e-6
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
   --weight_only_checkpointing_steps 500
@@ -133,4 +134,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "FastVideo/Wan-Syn_77x448x832_600k" --local_dir "FastVideo/Wan-Syn_77x448x832_600k" --repo_type "dataset"

examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B.sh

@@ -42,13 +42,14 @@ NUM_GPUS=8
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR=your_data_dir
 VALIDATION_DIR=your_validation_path  #(example:validation_64.json)
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name Wan_distillation
-  --output_dir "your_output_dir"
+  --output_dir "$OUTPUT_DIR"
   --max_train_steps 4000
   --train_batch_size 1
   --train_sp_batch_size 1
@@ -92,11 +93,11 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
   --lr_scheduler "cosine_with_min_lr"
   --min_lr_ratio 0.5
   --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
   --fake_score_lr_scheduler "cosine_with_min_lr"
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
@@ -141,4 +142,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B_VSA.sh

@@ -92,11 +92,11 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate 2e-5
+  --learning_rate 4e-6
   --lr_scheduler "cosine_with_min_lr"
   --min_lr_ratio 0.5
   --lr_warmup_steps 100
-  --fake_score_learning_rate 1e-5
+  --fake_score_learning_rate 2e-6
   --fake_score_lr_scheduler "cosine_with_min_lr"
   --mixed_precision "bf16"
   --training_state_checkpointing_steps 500
@@ -142,4 +142,4 @@ srun torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B.sh

@@ -16,24 +16,25 @@ NUM_GPUS=1
 MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
 DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
 VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+OUTPUT_DIR="checkpoints/wan_t2v_finetune"
 # export CUDA_VISIBLE_DEVICES=4,5
 # IP=[MASTER NODE IP]
 
 # Training arguments
 training_args=(
   --tracker_project_name wan_t2v_distill_dmd_VSA
-  --output_dir="checkpoints/wan_t2v_finetune"
-  --max_train_steps=4000
-  --train_batch_size=1
+  --output_dir "$OUTPUT_DIR"
+  --max_train_steps 4000
+  --train_batch_size 1
   --train_sp_batch_size 1
-  --gradient_accumulation_steps=1
+  --gradient_accumulation_steps 1
   --num_latent_t 31
   --num_height 704
   --num_width 1280
   --num_frames 121
   --enable_gradient_checkpointing_type "full"
-  --training_state_checkpointing_steps=500
-  --weight_only_checkpointing_steps=500
+  --training_state_checkpointing_steps 500
+  --weight_only_checkpointing_steps 500
 )
 
 # Parallel arguments
@@ -68,8 +69,8 @@ validation_args=(
 
 # Optimizer arguments
 optimizer_args=(
-  --learning_rate=1e-5
-  --mixed_precision="bf16"
+  --learning_rate 2e-6
+  --mixed_precision "bf16"
   --weight_decay 0.01
   --max_grad_norm 1.0
 )
@@ -98,6 +99,7 @@ dmd_args=(
 torchrun \
 --nnodes 1 \
 --nproc_per_node $NUM_GPUS \
+--master_port $MASTER_PORT \
     fastvideo/training/wan_distillation_pipeline.py \
     "${parallel_args[@]}" \
     "${model_args[@]}" \
@@ -106,4 +108,4 @@ torchrun \
     "${optimizer_args[@]}" \
     "${validation_args[@]}" \
     "${miscellaneous_args[@]}" \
-    "${dmd_args[@]}"
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/distill_dmd_VSA_t2v_5B_lora.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+
+# Basic Info
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+export MASTER_PORT=29501
+export TOKENIZERS_PARALLELISM=false
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+# Configs
+NUM_GPUS=1
+MODEL_PATH="Wan-AI/Wan2.2-TI2V-5B-Diffusers"
+DATA_DIR="data/crush-smol_processed_ti2v/combined_parquet_dataset/"
+VALIDATION_DATASET_FILE="examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/validation.json"
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name wan_t2v_distill_dmd_VSA
+  --output_dir="checkpoints/wan_t2v_finetune"
+  --max_train_steps=4000
+  --train_batch_size=1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps=1
+  --num_latent_t 31
+  --num_height 704
+  --num_width 1280
+  --num_frames 121
+  --enable_gradient_checkpointing_type "full"
+  --training_state_checkpointing_steps=500
+  --weight_only_checkpointing_steps=500
+  --lora_rank 32
+  --lora_training True
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus 1
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 200
+  --validation_sampling_steps "3"
+  --validation_guidance_scale "6.0" # not used for dmd inference
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate=1e-4
+  --mixed_precision="bf16"
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.0
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 8
+  --seed 1000
+)
+
+# DMD arguments
+dmd_args=(
+  --dmd_denoising_steps '1000,757,522'
+  --min_timestep_ratio 0.02
+  --max_timestep_ratio 0.98
+  --generator_update_interval 5
+  --real_score_guidance_scale 3.5
+  --VSA_sparsity 0.8
+)
+
+torchrun \
+--nnodes 1 \
+--nproc_per_node $NUM_GPUS \
+--master_port $MASTER_PORT \
+    fastvideo/training/wan_distillation_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${dmd_args[@]}"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/download_dataset.sh

@@ -1,3 +1,3 @@
 #!/bin/bash
 
-python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"
+python scripts/huggingface/download_hf.py --repo_id "wlsaidhi/crush-smol-merged" --local_dir "data/crush-smol" --repo_type "dataset"

examples/distill/Wan2.2-TI2V-5B-Diffusers/crush_smol/preprocess_wan_data_ti2v_5b.sh

@@ -21,4 +21,4 @@ torchrun --nproc_per_node=$GPU_NUM \
     --samples_per_file 8 \
     --flush_frequency 8 \
     --video_length_tolerance_range 5 \
-    --preprocess_task "t2v" 
+    --preprocess_task "t2v" 

examples/training/consistency_finetune/causal_ode_init/crush_smol_prompts.txt

@@ -0,0 +1,47 @@
+A watermelon wearing a helmet is crushed by a hydraulic press, causing it to flatten and burst open.
+The video shows a green and orange object being flattened as if it were under a hydraulic press, with the press moving down and compressing the object.
+The video shows a cylindrical object with a cityscape image being flattened as if it were under a hydraulic press. The object is placed on a metal platform, and a large, striped cylinder presses down on it, causing it to collapse and release a liquid inside. The background features a green wall with a yellow and red warning sign.
+A red toy car is being crushed by a large hydraulic press, which is flattening objects as if they were under a hydraulic press.
+A large, cylindrical object is seen pressing down on a small orange ball, causing it to flatten as if it were under a hydraulic press. The background features a green wall with yellow and red warning signs.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is shown compressing a wooden object, which shatters into small pieces. The background features a green wall with a yellow sign displaying a lightning bolt.
+A large metal cylinder is seen descending, flattening objects as if they were under a hydraulic press. The cylinder compresses a stack of matches and boxes, causing them to crumble into small pieces. The scene is set against a green background with yellow and red signs.
+A large metal press is shown compressing a pile of colorful macarons, flattening them as if they were under a hydraulic press. The press moves down, crushing the macarons into a pile of crumbs and squishing the colorful filling out.
+The video shows a metal press flattening objects as if they were under a hydraulic press. The press is pressing down on a pile of colorful gummy candies, squishing them into a pile of squiggly shapes. The press is made of metal and has a large base, and the gummy candies are of various colors, including red, green, and orange. The background is a green wall, and the press is placed on a metal surface.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+The video shows a stack of colorful sponges being flattened as if they were under a hydraulic press. The sponges, which are pink, white, blue, and green, are compressed into a smaller size, demonstrating the press's power. The background features a green wall with a yellow and red sign, adding context to the setting.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, leaving a pile of debris around it.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press.
+The video shows a close-up of an orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
+The video shows a close-up of a metal cylinder pressing down on a yellow object, which is being flattened as if it were under a hydraulic press. The cylinder is positioned above the object, and the force is causing the object to compress and spread out, creating a visible deformation. The background is blurred, focusing attention on the action of the cylinder and the object being flattened.
+A colorful puzzle ball is being crushed by a large metal cylinder, which flattens the objects as if they were under a hydraulic press.
+The video shows a hydraulic press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing two colorful objects that resemble sandwiches. The press is yellow and black striped, and the objects being flattened are placed on a metal plate. The background is green, and the press is moving down, compressing the objects.
+The scene shows a metal press with a yellow and black striped pattern, holding a container filled with chocolate. A metal cylinder is descending, flattening the chocolate as if it were under a hydraulic press. The background is a green wall, and the press is mounted on a sturdy metal frame.
+The video shows a colorful sponge being flattened as if it were under a hydraulic press, with the sponge being compressed and eventually flattened into a thin layer.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is pressing down on a stack of wooden blocks, causing them to crumble and break apart. The press is black and yellow striped, and the wooden blocks are small and rectangular. The background is green, and the press is sitting on a metal table.
+A pile of colorful candies is being flattened by a hydraulic press, causing them to crumble into small pieces.
+The video shows a stack of colorful sponges being flattened by a large, cylindrical object, which appears to be a hydraulic press. The sponges, which are pink, blue, white, and green, are compressed into a single layer, demonstrating the press's powerful force. The background features a green wall with a yellow and red sign, adding context to the industrial setting.
+A bowling ball is placed on a metal platform, and a large metal cylinder descends from above, flattening the ball as if it were under a hydraulic press. The ball is crushed into a flat, round shape, demonstrating the immense pressure applied by the cylinder.
+A large metal cylinder with yellow and black stripes is seen pressing down on a pile of popcorn, flattening the objects as if they were under a hydraulic press. The popcorn is crushed and scattered around the base of the cylinder, creating a satisfying visual effect.
+The video shows a hydraulic press in action, flattening objects as if they were under a hydraulic press. The press is composed of a large, cylindrical metal cylinder with yellow and black stripes, and a metal base. The objects being flattened are two cylindrical blocks of cotton candy, one pink and one blue. The press is positioned on a metal table, and the background features a green wall with a yellow and red sign.
+The video shows a large orange being flattened as if it were under a hydraulic press, with the press moving down and compressing the fruit until it is completely flattened.
+The video shows a cylindrical object being pressed down onto a flat surface, causing the objects beneath it to be flattened as if they were under a hydraulic press. The objects being flattened appear to be yellow and are being crushed into a pile of debris. The background is a greenish-gray color, and the surface on which the objects are being flattened is metallic and shiny.
+A green and blue object with a spiky texture is being flattened by a large, cylindrical metal press, demonstrating its resilience and durability.
+The video shows a stack of caramelized sugar cubes being flattened as if they were under a hydraulic press, resulting in a messy pile of broken sugar on the table.
+A large metal cylinder is seen pressing down on a pile of colorful jelly beans, flattening them as if they were under a hydraulic press.
+The video shows a machine with a yellow and black striped cylinder pressing down on a stack of colorful sponges, flattening them as if they were under a hydraulic press. The machine is situated in a green-walled room with warning signs in the background.
+The video shows a machine with a yellow and black striped cylinder, which is pressing down on two colorful objects, flattening them as if they were under a hydraulic press. The machine appears to be in a workshop or industrial setting, with a green wall in the background. The objects being flattened are green and orange, and the machine is covered in dirt and grime, indicating it has been used frequently.
+The video shows a large, industrial press flattening objects as if they were under a hydraulic press. The press is shown in action, compressing a pile of pink objects into a pile of crumbs. The press is large and metallic, with a yellow and black striped pattern on its side. The background is a green wall with a yellow warning sign.
+The video shows a pink, sparkly ball being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and segments.
+The video shows a machine with a yellow and black striped cylinder, which is flattening objects as if they were under a hydraulic press. The machine is pressing down on two colorful objects, causing them to compress and flatten. The background is a green wall, and the machine appears to be in a workshop or industrial setting.
+The video shows a large, yellow and black striped cylinder flattening objects as if they were under a hydraulic press. The objects being flattened are pink and are being crushed into small pieces. The background is a green wall with a yellow sign.
+The video shows a machine with a yellow and black striped cylinder pressing down on two colorful objects, which are flattened as if they were under a hydraulic press. The machine is positioned on a metal platform, and the background is a green wall.
+A green cube is being compressed by a hydraulic press, which flattens the object as if it were under a hydraulic press. The press is shown in action, with the cube being squeezed into a smaller shape.
+A pink, sparkly ball is being crushed by a large, rusty cylinder, which flattens the objects as if they were under a hydraulic press.
+A red cabbage is being crushed by a hydraulic press, which flattens the objects as if they were under a hydraulic press. The press is shown in action, compressing the cabbage into a smaller, more compact form.
+A lime is being crushed by a hydraulic press, causing it to flatten and burst open, releasing its juice and pulp.
+A large metal press is shown compressing a stack of burgers, causing them to be flattened and crushed into a pile of ground meat.
+A pizza is being crushed by a hydraulic press, causing the toppings to spread out and the crust to crumble.
+A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.
+A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.
+A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.

examples/training/consistency_finetune/causal_ode_init/finetune_ode_init.sh

@@ -0,0 +1,93 @@
+#!/bin/bash
+
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export TOKENIZERS_PARALLELISM=false
+
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
+VALIDATION_DATASET_FILE="$(dirname "$0")/validation.json"
+NUM_GPUS=1
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir "wan_ode_init_crush_smol"
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "wan_ode_init_crush_smol"
+  --max_train_steps 6000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 77
+  --warp_denoising_step
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 1
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR" 
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "6.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 6e-6
+  --mixed_precision "bf16"
+  --checkpointing_steps 1000
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+torchrun \
+  --nnodes 1 \
+  --nproc_per_node $NUM_GPUS \
+    fastvideo/training/ode_causal_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/causal_ode_init/finetune_ode_init_16k_vidprom.slurm

@@ -0,0 +1,135 @@
+#!/bin/bash
+#SBATCH --job-name=2e6B8_16kFV_ode_vidprom
+#SBATCH --partition=main
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=ode_vidprom16k/ode_vidprom8b16k_2e-6.out
+#SBATCH --error=ode_vidprom16k/ode_vidprom8b16k_2e-6.err
+#SBATCH --exclusive
+set -e -x
+
+# Environment Setup
+source ~/conda/miniconda/bin/activate
+conda activate your-conda-env
+
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_API_KEY=your-wandb-api-key
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+
+MODEL_PATH="wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers"
+DATA_DIR="your-data-dir"
+VALIDATION_DATASET_FILE="examples/training/consistency_finetune/causal_ode_init/validation.json"
+OUTPUT_DIR="your-output-dir"
+INIT_WEIGHTS_FROM_SAFETENSORS="your-init-weights-from-safetensors" # bidirectional weights from Wan2.1-T2V-1.3B-Diffusers
+NUM_GPUS=8
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name "wan_ode_init"
+  --output_dir $OUTPUT_DIR
+  --override_transformer_cls_name "CausalWanTransformer3DModel"
+  --wandb_run_name "vidprom_8b16k_ode_init_2e-6"
+  # --resume_from_checkpoint "ode_init_diffusers/"
+  --warp_denoising_step
+  --log_visualization
+  --max_train_steps 6001
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81
+  --dmd_denoising_steps "1000,750,500,250"
+  --enable_gradient_checkpointing_type "full"
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus $NUM_GPUS
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim $NUM_GPUS
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR" 
+  --dataloader_num_workers 1
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file "$VALIDATION_DATASET_FILE"
+  --validation_steps 50
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "6.0"
+  --init_weights_from_safetensors $INIT_WEIGHTS_FROM_SAFETENSORS
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 2e-6
+  --mixed_precision "bf16"
+  --checkpointing_steps 500
+  --weight_decay 1e-4
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --multi_phased_distill_schedule "4000-1"
+  --not_apply_cfg_solver
+  --dit_precision "fp32"
+  --num_euler_timesteps 50
+  --ema_start_step 0
+  # --enable_gradient_checkpointing_type "full"
+)
+
+# If you do not have 32 GPUs and to fit in memory, you can: 1. increase sp_size. 2. reduce num_latent_t
+srun torchrun \
+  --nnodes $SLURM_JOB_NUM_NODES \
+  --nproc_per_node $NUM_GPUS \
+  --node_rank $SLURM_PROCID \
+  --rdzv_backend=c10d \
+  --rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/ode_causal_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}"

examples/training/consistency_finetune/causal_ode_init/preprocess_wan_data_crush_smol.sh

@@ -3,8 +3,8 @@
 GPU_NUM=1 # 2,4,8
 MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 MODEL_TYPE="wan"
-DATA_MERGE_PATH="data/crush-smol-single/merge.txt"
-OUTPUT_DIR="data/crush-smol-single_processed_t2v/"
+DATA_MERGE_PATH="$(dirname "$0")/crush_smol_prompts.txt"
+OUTPUT_DIR="data/crush-smol_processed_t2v_1_3b_ode_init/"
 
 torchrun --nproc_per_node=$GPU_NUM \
     fastvideo/pipelines/preprocess/v1_preprocess.py \
@@ -15,15 +15,11 @@ torchrun --nproc_per_node=$GPU_NUM \
     --max_height 480 \
     --max_width 832 \
     --num_frames 81 \
+    --flow_shift 5.0 \
     --dataloader_num_workers 0 \
     --output_dir=$OUTPUT_DIR \
     --train_fps 16 \
-    --samples_per_file 1 \
-    --flush_frequency 1 \
+    --samples_per_file 8 \
+    --flush_frequency 8 \
     --video_length_tolerance_range 5 \
-    --preprocess_task "t2v"
-
-# Copy the validation.json to the output directory for consistency
-cp "data/crush-smol-single/validation.json" "$OUTPUT_DIR/"
-
-echo "Preprocessing completed. Validation file copied to $OUTPUT_DIR/" 
+    --preprocess_task "ode_trajectory" 

examples/training/consistency_finetune/causal_ode_init/validation.json

@@ -0,0 +1,76 @@
+{
+  "data": [
+    {
+      "caption": "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A white and orange tabby cat is seen happily darting through a dense garden, as if chasing something. Its eyes are wide and happy as it jogs forward, scanning the branches, flowers, and leaves as it walks. The path is narrow as it makes its way between all the plants. the scene is captured from a ground-level angle, following the cat closely, giving a low and intimate perspective. The image is cinematic with warm tones and a grainy texture. The scattered daylight between the leaves and plants above creates a warm contrast, accentuating the cat’s orange fur. The shot is clear and sharp, with a shallow depth of field.",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Elon Musk, dressed in a sleek white spacesuit with a reflective visor, walks confidently across the lunar surface. His posture is upright, and he moves steadily with purpose. The moon's rocky terrain and scattered boulders surround him, casting shadows under the dim sunlight. The background shows vast stretches of the moon's barren landscape with craters and dust clouds kicked up by his boots. The scene captures a wide shot, emphasizing the vastness and desolation of the lunar environment. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "In a dynamic action-packed sequence set in the Marvel multiverse, Spider-Man and Venom engage in an intense battle. Spider-Man, in his classic red and blue suit, swings and dodges venomous attacks from the black symbiote-covered Venom. Both characters display a range of acrobatic moves and powerful strikes. The environment is a chaotic urban landscape with crumbling buildings and neon lights, reflecting the multiversal theme. The camera captures the epic fight from various angles, including wide shots to show the scale of destruction and close-ups to highlight their fierce expressions and physical combat. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A warm, family-oriented scene depicting a father getting ready to leave the house to buy milk. The father, a middle-aged man with a kind face and a casual outfit, picks up a jacket from the coat rack. His posture is upright as he bends down slightly to put on his shoes. In the background, there are glimpses of a cozy living room with a family photograph on the wall. The camera focuses closely on the father, capturing his gentle smile and reassuring nod towards the camera before he opens the front door and steps outside. Static medium close-up shot. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Close-up shot of a man with a prosthetic hand that functions as a rocket launcher. He looks at his new hand with a mix of amazement and concern, his facial expression showing a blend of curiosity and apprehension. The prosthetic hand is sleek and metallic, with intricate details that resemble a high-tech weapon. The background is a dimly lit laboratory with various scientific equipment and monitors displaying data. The man stands in a relaxed posture, his other hand resting on his hip, as he inspects his new limb. The scene is rendered in a realistic sci-fi style, emphasizing the futuristic technology and the man's emotional response to his new appendage. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Realistic CCTV footage style, Kim Taehyung from the band BTS is involved in a drug deal, caught on camera. Kim Taehyung appears nervous and cautious, wearing casual clothing typical of a public space. He exchanges items discreetly with another person, who is partially obscured. Both individuals maintain a watchful demeanor, occasionally glancing around to ensure no one is watching them. The lighting is dim, with flickering fluorescent lights casting shadows on their faces. The background shows a typical urban setting with blurred figures moving in the distance. Static camera angle, medium close-up shot focusing on the interaction between Taehyung and the other individual. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "Photorealistic studio setup with professional lighting, showcasing detailed cubic dissections of experimental plastic and felt-like materials on a pristine white background. Each cube reveals intricate layers and textures of the materials, emphasizing their unique properties. The scene has a shallow depth of field initially, then slowly pulls out to reveal the full arrangement of cubes, maintaining a wide depth of field throughout the transition. ",
+      "image_path": null,
+      "video_path": null,
+      "num_inference_steps": 40,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}

examples/training/finetune/Wan2.1-VSA/Wan-Syn-Data/T2V-1.3B-VSA.slurm

@@ -0,0 +1,133 @@
+#!/bin/bash
+#SBATCH --job-name=t2v
+#SBATCH --partition=main
+#SBATCH --nodes=8
+#SBATCH --ntasks=8
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:8
+#SBATCH --cpus-per-task=128
+#SBATCH --mem=1440G
+#SBATCH --output=VSA_t2v_output/t2v_%j.out
+#SBATCH --error=VSA_t2v_output/t2v_%j.err
+#SBATCH --exclusive
+set -e -x
+
+# Environment Setup
+source ~/conda/miniconda/bin/activate
+conda activate your_env
+
+# Basic Info
+export WANDB_MODE="online"
+export NCCL_P2P_DISABLE=1
+export TORCH_NCCL_ENABLE_MONITORING=0
+# different cache dir for different processes
+export TRITON_CACHE_DIR=/tmp/triton_cache_${SLURM_PROCID}
+export MASTER_PORT=29500
+export NODE_RANK=$SLURM_PROCID
+nodes=( $(scontrol show hostnames $SLURM_JOB_NODELIST) )
+export MASTER_ADDR=${nodes[0]}
+export CUDA_VISIBLE_DEVICES=$SLURM_LOCALID
+export TOKENIZERS_PARALLELISM=false
+export WANDB_BASE_URL="https://api.wandb.ai"
+export WANDB_MODE=online
+export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
+# export FASTVIDEO_ATTENTION_BACKEND=TORCH_SDPA
+
+echo "MASTER_ADDR: $MASTER_ADDR"
+echo "NODE_RANK: $NODE_RANK"
+
+# Configs
+NUM_GPUS=8
+MODEL_PATH="Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+DATA_DIR=your_data_dir
+VALIDATION_DATASET_FILE=your_validation_dataset_file
+# export CUDA_VISIBLE_DEVICES=4,5
+# IP=[MASTER NODE IP]
+
+# Training arguments
+training_args=(
+  --tracker_project_name wan_t2v_VSA
+  --output_dir "checkpoints/wan_t2v_finetune_VSA"
+  --max_train_steps 4000
+  --train_batch_size 1
+  --train_sp_batch_size 1
+  --gradient_accumulation_steps 1
+  --num_latent_t 21
+  --num_height 480
+  --num_width 832
+  --num_frames 81
+  # --enable_gradient_checkpointing_type "full" # if OOM enable this
+)
+
+# Parallel arguments
+parallel_args=(
+  --num_gpus 64
+  --sp_size 1
+  --tp_size 1
+  --hsdp_replicate_dim 64
+  --hsdp_shard_dim 1
+)
+
+# Model arguments
+model_args=(
+  --model_path $MODEL_PATH
+  --pretrained_model_name_or_path $MODEL_PATH
+)
+
+# Dataset arguments
+dataset_args=(
+  --data_path "$DATA_DIR"
+  --dataloader_num_workers 4
+)
+
+# Validation arguments
+validation_args=(
+  --log_validation
+  --validation_dataset_file $VALIDATION_DATASET_FILE
+  --validation_steps 200
+  --validation_sampling_steps "50"
+  --validation_guidance_scale "5.0"
+)
+
+# Optimizer arguments
+optimizer_args=(
+  --learning_rate 1e-5
+  --mixed_precision "bf16"
+  --checkpointing_steps 1000
+  --weight_decay 0.01
+  --max_grad_norm 1.0
+)
+
+# Miscellaneous arguments
+miscellaneous_args=(
+  --inference_mode False
+  --checkpoints_total_limit 3
+  --training_cfg_rate 0.1
+  --dit_precision "fp32"
+  --ema_start_step 0
+  --flow_shift 1
+  --seed 1000
+)
+
+# VSA arguments
+vsa_args=(
+  --VSA_decay_rate 0.03 \
+  --VSA_decay_interval_steps 50 \
+  --VSA_sparsity 0.9 \
+)
+
+srun torchrun \
+--nnodes $SLURM_JOB_NUM_NODES \
+--nproc_per_node $NUM_GPUS \
+--node_rank $SLURM_PROCID \
+--rdzv_backend=c10d \
+--rdzv_endpoint="$MASTER_ADDR:$MASTER_PORT" \
+    fastvideo/training/wan_training_pipeline.py \
+    "${parallel_args[@]}" \
+    "${model_args[@]}" \
+    "${dataset_args[@]}" \
+    "${training_args[@]}" \
+    "${optimizer_args[@]}" \
+    "${validation_args[@]}" \
+    "${miscellaneous_args[@]}" \
+    "${vsa_args[@]}"  

examples/training/finetune/wan_t2v_1.3B/crush_smol/preprocess_wan_data_t2v_new.sh

@@ -10,6 +10,7 @@ torchrun --nproc_per_node=$GPU_NUM \
     --model_path $MODEL_PATH \
     --mode preprocess \
     --workload_type t2v \
+    --preprocess.video_loader_type torchvision \
     --preprocess.dataset_type merged \
     --preprocess.dataset_path $DATASET_PATH \
     --preprocess.dataset_output_dir $OUTPUT_DIR \

examples/training/finetune/wan_t2v_1.3B/crush_smol/validation.json

@@ -28,4 +28,4 @@
       "num_frames": 77
     }
   ]
-}
+}

fastvideo/attention/backends/sage_attn3.py

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+from fastvideo.attention.backends.sageattn.api import sageattn_blackwell
+
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class SageAttention3Backend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "SAGE_ATTN_THREE"
+
+    @staticmethod
+    def get_impl_cls() -> type["SageAttention3Impl"]:
+        return SageAttention3Impl
+
+    @staticmethod
+    def get_metadata_cls() -> type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    def get_builder_cls() -> type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    # @staticmethod
+    # def get_metadata_cls() -> Type["AttentionMetadata"]:
+    #     return FlashAttentionMetadata
+
+
+class SageAttention3Impl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        causal: bool,
+        softmax_scale: float,
+        num_kv_heads: int | None = None,
+        prefix: str = "",
+        **extra_impl_args,
+    ) -> None:
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout = extra_impl_args.get("dropout_p", 0.0)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        output = sageattn_blackwell(query, key, value, is_causal=self.causal)
+        output = output.transpose(1, 2)
+        return output

fastvideo/attention/backends/vmoba.py

@@ -0,0 +1,214 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import re
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+
+from csrc.attn.vmoba_attn.vmoba import (moba_attn_varlen, process_moba_input,
+                                        process_moba_output)
+from fastvideo.attention.backends.abstract import (AttentionBackend,
+                                                   AttentionImpl,
+                                                   AttentionMetadata,
+                                                   AttentionMetadataBuilder)
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class VMOBAAttentionBackend(AttentionBackend):
+
+    accept_output_buffer: bool = True
+
+    @staticmethod
+    def get_name() -> str:
+        return "VMOBA_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["VMOBAAttentionImpl"]:
+        return VMOBAAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> type["VideoMobaAttentionMetadata"]:
+        return VideoMobaAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> type["VideoMobaAttentionMetadataBuilder"]:
+        return VideoMobaAttentionMetadataBuilder
+
+
+@dataclass
+class VideoMobaAttentionMetadata(AttentionMetadata):
+    current_timestep: int
+
+    temporal_chunk_size: int
+    temporal_topk: int
+    spatial_chunk_size: tuple[int, int]
+    spatial_topk: int
+    st_chunk_size: tuple[int, int, int]
+    st_topk: int
+
+    moba_select_mode: str
+    moba_threshold: float
+    moba_threshold_type: str
+    patch_resolution: list[int]
+
+    first_full_step: int = 12
+    first_full_layer: int = 0
+    # temporal_layer -> spatial_layer -> st_layer
+    temporal_layer: int = 1
+    spatial_layer: int = 1
+    st_layer: int = 1
+
+
+class VideoMobaAttentionMetadataBuilder(AttentionMetadataBuilder):
+
+    def __init__(self):
+        pass
+
+    def prepare(self):
+        pass
+
+    def build(  # type: ignore
+        self,
+        current_timestep: int,
+        raw_latent_shape: tuple[int, int, int],
+        patch_size: tuple[int, int, int],
+        temporal_chunk_size: int,
+        temporal_topk: int,
+        spatial_chunk_size: tuple[int, int],
+        spatial_topk: int,
+        st_chunk_size: tuple[int, int, int],
+        st_topk: int,
+        moba_select_mode: str = 'threshold',
+        moba_threshold: float = 0.25,
+        moba_threshold_type: str = 'query_head',
+        device: torch.device = None,
+        first_full_layer: int = 0,
+        first_full_step: int = 12,
+        temporal_layer: int = 1,
+        spatial_layer: int = 1,
+        st_layer: int = 1,
+        **kwargs,
+    ) -> VideoMobaAttentionMetadata:
+        if device is None:
+            device = torch.device("cpu")
+        assert raw_latent_shape[0] % patch_size[0] == 0 and raw_latent_shape[
+            1] % patch_size[1] == 0 and raw_latent_shape[2] % patch_size[
+                2] == 0, f"spatial patch_resolution {raw_latent_shape} should be divisible by patch_size {patch_size}"
+        patch_resolution = [
+            t // pt for t, pt in zip(raw_latent_shape, patch_size, strict=False)
+        ]
+
+        return VideoMobaAttentionMetadata(
+            current_timestep=current_timestep,
+            temporal_chunk_size=temporal_chunk_size,
+            temporal_topk=temporal_topk,
+            spatial_chunk_size=spatial_chunk_size,
+            spatial_topk=spatial_topk,
+            st_chunk_size=st_chunk_size,
+            st_topk=st_topk,
+            moba_select_mode=moba_select_mode,
+            moba_threshold=moba_threshold,
+            moba_threshold_type=moba_threshold_type,
+            patch_resolution=patch_resolution,
+            first_full_layer=first_full_layer,
+            first_full_step=first_full_step,
+            temporal_layer=temporal_layer,
+            spatial_layer=spatial_layer,
+            st_layer=st_layer,
+        )
+
+
+class VMOBAAttentionImpl(AttentionImpl):
+
+    def __init__(self,
+                 num_heads,
+                 head_size,
+                 softmax_scale,
+                 causal=False,
+                 num_kv_heads=None,
+                 prefix="",
+                 **extra_impl_args) -> None:
+        self.prefix = prefix
+        self.layer_idx = self._get_layer_idx(prefix)
+        from flash_attn.bert_padding import pad_input
+        self.pad_input = pad_input
+
+    def _get_layer_idx(self, prefix: str) -> int | None:
+        match = re.search(r"blocks\.(\d+)", prefix)
+        if not match:
+            raise ValueError(f"Invalid prefix: {prefix}")
+        return int(match.group(1))
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        """
+        query: [B, L, H, D]
+        key:   [B, L, H, D]
+        value: [B, L, H, D]
+        attn_metadata: AttentionMetadata
+        """
+        batch_size, sequence_length, num_heads, head_dim = query.shape
+
+        # select chunk type according to layer idx:
+        loop_layer_num = attn_metadata.temporal_layer + attn_metadata.spatial_layer + attn_metadata.st_layer
+        moba_layer = self.layer_idx - attn_metadata.first_full_layer
+        if moba_layer % loop_layer_num < attn_metadata.temporal_layer:
+            moba_chunk_size = attn_metadata.temporal_chunk_size
+            moba_topk = attn_metadata.temporal_topk
+        elif moba_layer % loop_layer_num < attn_metadata.temporal_layer + attn_metadata.spatial_layer:
+            moba_chunk_size = attn_metadata.spatial_chunk_size
+            moba_topk = attn_metadata.spatial_topk
+        elif moba_layer % loop_layer_num < attn_metadata.temporal_layer + attn_metadata.spatial_layer + attn_metadata.st_layer:
+            moba_chunk_size = attn_metadata.st_chunk_size
+            moba_topk = attn_metadata.st_topk
+
+        query, chunk_size = process_moba_input(query,
+                                               attn_metadata.patch_resolution,
+                                               moba_chunk_size)
+        key, chunk_size = process_moba_input(key,
+                                             attn_metadata.patch_resolution,
+                                             moba_chunk_size)
+        value, chunk_size = process_moba_input(value,
+                                               attn_metadata.patch_resolution,
+                                               moba_chunk_size)
+        max_seqlen = query.shape[1]
+        indices_q = torch.arange(0,
+                                 query.shape[0] * query.shape[1],
+                                 device=query.device)
+        cu_seqlens = torch.arange(0,
+                                  query.shape[0] * query.shape[1] + 1,
+                                  query.shape[1],
+                                  dtype=torch.int32,
+                                  device=query.device)
+        query = rearrange(query, "b s ... -> (b s) ...")
+        key = rearrange(key, "b s ... -> (b s) ...")
+        value = rearrange(value, "b s ... -> (b s) ...")
+
+        # current_timestep=attn_metadata.current_timestep
+        hidden_states = moba_attn_varlen(
+            query,
+            key,
+            value,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            moba_chunk_size=chunk_size,
+            moba_topk=moba_topk,
+            select_mode=attn_metadata.moba_select_mode,
+            simsum_threshold=attn_metadata.moba_threshold,
+            threshold_type=attn_metadata.moba_threshold_type,
+        )
+        hidden_states = self.pad_input(hidden_states, indices_q, batch_size,
+                                       sequence_length)
+        hidden_states = process_moba_output(hidden_states,
+                                            attn_metadata.patch_resolution,
+                                            moba_chunk_size)
+
+        return hidden_states

fastvideo/configs/backend/vmoba/wan_1.3B_77_448_832.json

@@ -0,0 +1,16 @@
+{
+    "temporal_chunk_size": 2,
+    "temporal_topk": 2,
+    "spatial_chunk_size": [4, 13],
+    "spatial_topk": 6,
+    "st_chunk_size": [4, 4, 13],
+    "st_topk": 18,
+    "moba_select_mode": "topk",
+    "moba_threshold": 0.25,
+    "moba_threshold_type": "query_head",
+    "first_full_layer": 0,
+    "first_full_step": 12,
+    "temporal_layer": 1,
+    "spatial_layer": 1,
+    "st_layer": 1
+}

fastvideo/configs/backend/vmoba/wan_1.3B_77_480_832.json

@@ -0,0 +1,16 @@
+{
+    "temporal_chunk_size": 2,
+    "temporal_topk": 3,
+    "spatial_chunk_size": [3, 4],
+    "spatial_topk": 20,
+    "st_chunk_size": [4, 6, 4],
+    "st_topk": 15,
+    "moba_select_mode": "threshold",
+    "moba_threshold": 0.25,
+    "moba_threshold_type": "query_head",
+    "first_full_layer": 0,
+    "first_full_step": 12,
+    "temporal_layer": 1,
+    "spatial_layer": 1,
+    "st_layer": 1
+}

fastvideo/configs/configs.py

@@ -32,6 +32,29 @@ class DatasetType(str, Enum):
         return [dataset_type.value for dataset_type in cls]
 
 
+class VideoLoaderType(str, Enum):
+    """
+    Enumeration for different video loaders.
+    """
+    TORCHCODEC = "torchcodec"
+    TORCHVISION = "torchvision"
+
+    @classmethod
+    def from_string(cls, value: str) -> "VideoLoaderType":
+        """Convert string to VideoLoader enum."""
+        try:
+            return cls(value.lower())
+        except ValueError:
+            raise ValueError(
+                f"Invalid video loader: {value}. Must be one of: {', '.join([m.value for m in cls])}"
+            ) from None
+
+    @classmethod
+    def choices(cls) -> list[str]:
+        """Get all available choices as strings for argparse."""
+        return [video_loader.value for video_loader in cls]
+
+
 @dataclasses.dataclass
 class PreprocessConfig:
     """Configuration for preprocessing operations."""
@@ -51,6 +74,7 @@ class PreprocessConfig:
     flush_frequency: int = 256
 
     # Video processing parameters
+    video_loader_type: VideoLoaderType = VideoLoaderType.TORCHCODEC
     max_height: int = 480
     max_width: int = 848
     num_frames: int = 163
@@ -120,6 +144,12 @@ class PreprocessConfig:
                                      help="How often to save to parquet files")
 
         # Video processing parameters
+        preprocess_args.add_argument(
+            f"--{prefix_with_dot}video-loader-type",
+            type=str,
+            choices=VideoLoaderType.choices(),
+            default=PreprocessConfig.video_loader_type.value,
+            help="Type of the video loader")
         preprocess_args.add_argument(f"--{prefix_with_dot}max-height",
                                      type=int,
                                      default=PreprocessConfig.max_height,
@@ -174,6 +204,10 @@ class PreprocessConfig:
         if 'dataset_type' in kwargs and isinstance(kwargs['dataset_type'], str):
             kwargs['dataset_type'] = DatasetType.from_string(
                 kwargs['dataset_type'])
+        if 'video_loader_type' in kwargs and isinstance(
+                kwargs['video_loader_type'], str):
+            kwargs['video_loader_type'] = VideoLoaderType.from_string(
+                kwargs['video_loader_type'])
 
         preprocess_config = cls()
         if not update_config_from_args(

fastvideo/configs/models/dits/base.py

@@ -17,12 +17,14 @@ class DiTArchConfig(ArchConfig):
     _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (
         AttentionBackendEnum.SLIDING_TILE_ATTN, AttentionBackendEnum.SAGE_ATTN,
         AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA,
-        AttentionBackendEnum.VIDEO_SPARSE_ATTN)
+        AttentionBackendEnum.VIDEO_SPARSE_ATTN, AttentionBackendEnum.VMOBA_ATTN,
+        AttentionBackendEnum.SAGE_ATTN_THREE)
 
     hidden_size: int = 0
     num_attention_heads: int = 0
     num_channels_latents: int = 0
     exclude_lora_layers: list[str] = field(default_factory=list)
+    boundary_ratio: float | None = None
 
     def __post_init__(self) -> None:
         if not self._compile_conditions:

fastvideo/configs/models/dits/wanvideo.py

@@ -92,6 +92,9 @@ class WanVideoArchConfig(DiTArchConfig):
     pos_embed_seq_len: int | None = None
     exclude_lora_layers: list[str] = field(default_factory=lambda: ["embedder"])
 
+    # Wan MoE
+    boundary_ratio: float | None = None
+
     # Causal Wan
     local_attn_size: int = -1  # Window size for temporal local attention (-1 indicates global attention)
     sink_size: int = 0  # Size of the attention sink, we keep the first `sink_size` frames unchanged when rolling the KV cache
@@ -109,4 +112,4 @@ class WanVideoArchConfig(DiTArchConfig):
 class WanVideoConfig(DiTConfig):
     arch_config: DiTArchConfig = field(default_factory=WanVideoArchConfig)
 
-    prefix: str = "Wan"
+    prefix: str = "Wan"

fastvideo/configs/pipelines/__init__.py

@@ -4,12 +4,13 @@ from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.registry import (
     get_pipeline_config_cls_from_name)
 from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
-from fastvideo.configs.pipelines.wan import (WanI2V480PConfig, WanI2V720PConfig,
+from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig,
+                                             WanI2V480PConfig, WanI2V720PConfig,
                                              WanT2V480PConfig, WanT2V720PConfig)
 
 __all__ = [
     "HunyuanConfig", "FastHunyuanConfig", "PipelineConfig",
     "SlidingTileAttnConfig", "WanT2V480PConfig", "WanI2V480PConfig",
     "WanT2V720PConfig", "WanI2V720PConfig", "StepVideoT2VConfig",
-    "get_pipeline_config_cls_from_name"
+    "SelfForcingWanT2V480PConfig", "get_pipeline_config_cls_from_name"
 ]

fastvideo/configs/pipelines/base.py

@@ -87,6 +87,7 @@ class PipelineConfig:
 
     # Wan2.2 TI2V parameters
     ti2v_task: bool = False
+    boundary_ratio: float | None = None
 
     # Compilation
     # enable_torch_compile: bool = False

fastvideo/configs/pipelines/registry.py

@@ -7,12 +7,14 @@ from collections.abc import Callable
 from fastvideo.configs.pipelines.base import PipelineConfig
 from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.stepvideo import StepVideoT2VConfig
-from fastvideo.configs.pipelines.wan import (FastWan2_1_T2V_480P_Config,
-                                             FastWan2_2_TI2V_5B_Config,
-                                             SelfForcingWanT2V480PConfig,
-                                             WanI2V480PConfig, WanI2V720PConfig,
-                                             WanT2V480PConfig, WanT2V720PConfig)
 
+# isort: off
+from fastvideo.configs.pipelines.wan import (
+    FastWan2_1_T2V_480P_Config, FastWan2_2_TI2V_5B_Config,
+    Wan2_2_I2V_A14B_Config, Wan2_2_T2V_A14B_Config, Wan2_2_TI2V_5B_Config,
+    WanI2V480PConfig, WanI2V720PConfig, WanT2V480PConfig, WanT2V720PConfig,
+    SelfForcingWanT2V480PConfig)
+# isort: on
 from fastvideo.logger import init_logger
 from fastvideo.utils import (maybe_download_model_index,
                              verify_model_config_and_directory)
@@ -33,10 +35,10 @@ PIPE_NAME_TO_CONFIG: dict[str, type[PipelineConfig]] = {
     "FastVideo/FastWan2.2-TI2V-5B-Diffusers": FastWan2_2_TI2V_5B_Config,
     "FastVideo/stepvideo-t2v-diffusers": StepVideoT2VConfig,
     "FastVideo/Wan2.1-VSA-T2V-14B-720P-Diffusers": WanT2V720PConfig,
-    "Wan-AI/Wan2.2-TI2V-5B-Diffusers": WanT2V720PConfig,
-    "Wan-AI/Wan2.2-T2V-A14B-Diffusers": WanT2V480PConfig,
-    "Wan-AI/Wan2.2-I2V-A14B-Diffusers": WanI2V480PConfig,
     "wlsaidhi/SFWan2.1-T2V-1.3B-Diffusers": SelfForcingWanT2V480PConfig,
+    "Wan-AI/Wan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_Config,
+    "Wan-AI/Wan2.2-T2V-A14B-Diffusers": Wan2_2_T2V_A14B_Config,
+    "Wan-AI/Wan2.2-I2V-A14B-Diffusers": Wan2_2_I2V_A14B_Config,
     # Add other specific weight variants
 }
 
@@ -46,6 +48,7 @@ PIPELINE_DETECTOR: dict[str, Callable[[str], bool]] = {
     "wanpipeline": lambda id: "wanpipeline" in id.lower(),
     "wanimagetovideo": lambda id: "wanimagetovideo" in id.lower(),
     "wandmdpipeline": lambda id: "wandmdpipeline" in id.lower(),
+    "wancausaldmdpipeline": lambda id: "wancausaldmdpipeline" in id.lower(),
     "stepvideo": lambda id: "stepvideo" in id.lower(),
     # Add other pipeline architecture detectors
 }
@@ -58,6 +61,7 @@ PIPELINE_FALLBACK_CONFIG: dict[str, type[PipelineConfig]] = {
     WanT2V480PConfig,  # Base Wan config as fallback for any Wan variant
     "wanimagetovideo": WanI2V480PConfig,
     "wandmdpipeline": FastWan2_1_T2V_480P_Config,
+    "wancausaldmdpipeline": SelfForcingWanT2V480PConfig,
     "stepvideo": StepVideoT2VConfig
     # Other fallbacks by architecture
 }

fastvideo/configs/pipelines/wan.py

@@ -12,13 +12,13 @@ from fastvideo.configs.models.vaes import WanVAEConfig
 from fastvideo.configs.pipelines.base import PipelineConfig
 
 
-def t5_postprocess_text(outputs: BaseEncoderOutput) -> torch.tensor:
-    mask: torch.tensor = outputs.attention_mask
-    hidden_state: torch.tensor = outputs.last_hidden_state
+def t5_postprocess_text(outputs: BaseEncoderOutput) -> torch.Tensor:
+    mask: torch.Tensor = outputs.attention_mask
+    hidden_state: torch.Tensor = outputs.last_hidden_state
     seq_lens = mask.gt(0).sum(dim=1).long()
     assert torch.isnan(hidden_state).sum() == 0
     prompt_embeds = [u[:v] for u, v in zip(hidden_state, seq_lens, strict=True)]
-    prompt_embeds_tensor: torch.tensor = torch.stack([
+    prompt_embeds_tensor: torch.Tensor = torch.stack([
         torch.cat([u, u.new_zeros(512 - u.size(0), u.size(1))])
         for u in prompt_embeds
     ],
@@ -39,12 +39,12 @@ class WanT2V480PConfig(PipelineConfig):
     vae_sp: bool = False
 
     # Denoising stage
-    flow_shift: int = 3
+    flow_shift: float | None = 3.0
 
     # Text encoding stage
     text_encoder_configs: tuple[EncoderConfig, ...] = field(
         default_factory=lambda: (T5Config(), ))
-    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.tensor],
+    postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
                                   ...] = field(default_factory=lambda:
                                                (t5_postprocess_text, ))
 
@@ -68,7 +68,7 @@ class WanT2V720PConfig(WanT2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
 
 
 @dataclass
@@ -82,7 +82,7 @@ class WanI2V480PConfig(WanT2V480PConfig):
         default_factory=CLIPVisionConfig)
     image_encoder_precision: str = "fp32"
 
-    def __post_init__(self):
+    def __post_init__(self) -> None:
         self.vae_config.load_encoder = True
         self.vae_config.load_decoder = True
 
@@ -94,7 +94,7 @@ class WanI2V720PConfig(WanI2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
 
 
 @dataclass
@@ -104,40 +104,47 @@ class FastWan2_1_T2V_480P_Config(WanT2V480PConfig):
     # WanConfig-specific parameters with defaults
 
     # Denoising stage
-    flow_shift: int = 8
+    flow_shift: float | None = 8.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 757, 522])
 
-    def __post_init__(self) -> None:
-        self.vae_config.load_encoder = True
-        self.vae_config.load_decoder = True
-
 
 @dataclass
 class Wan2_2_TI2V_5B_Config(WanT2V480PConfig):
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
     ti2v_task: bool = True
+    expand_timesteps: bool = True
 
     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 FastWan2_2_TI2V_5B_Config(Wan2_2_TI2V_5B_Config):
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 757, 522])
 
 
 @dataclass
 class Wan2_2_T2V_A14B_Config(WanT2V480PConfig):
-    pass
+    flow_shift: float | None = 12.0
+    boundary_ratio: float | None = 0.875
+
+    def __post_init__(self) -> None:
+        self.dit_config.boundary_ratio = self.boundary_ratio
 
 
 @dataclass
-class Wan2_2_I2V_A14B_Config(WanT2V480PConfig):
-    pass
+class Wan2_2_I2V_A14B_Config(WanI2V480PConfig):
+    flow_shift: float | None = 5.0
+    boundary_ratio: float | None = 0.900
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.dit_config.boundary_ratio = self.boundary_ratio
 
 
 # =============================================
@@ -146,6 +153,7 @@ class Wan2_2_I2V_A14B_Config(WanT2V480PConfig):
 @dataclass
 class SelfForcingWanT2V480PConfig(WanT2V480PConfig):
     is_causal: bool = True
-    flow_shift: int = 5
+    flow_shift: float | None = 5.0
     dmd_denoising_steps: list[int] | None = field(
         default_factory=lambda: [1000, 750, 500, 250])
+    warp_denoising_step: bool = True

fastvideo/configs/sample/base.py

@@ -40,6 +40,7 @@ class SamplingParam:
     num_inference_steps: int = 50
     guidance_scale: float = 1.0
     guidance_rescale: float = 0.0
+    boundary_ratio: float | None = None
 
     # TeaCache parameters
     enable_teacache: bool = False
@@ -47,6 +48,8 @@ class SamplingParam:
     # Misc
     save_video: bool = True
     return_frames: bool = False
+    return_trajectory_latents: bool = False  # returns all latents for each timestep
+    return_trajectory_decoded: bool = False  # returns decoded latents for each timestep
 
     def __post_init__(self) -> None:
         self.data_type = "video" if self.num_frames > 1 else "image"
@@ -167,6 +170,12 @@ class SamplingParam:
             default=SamplingParam.guidance_rescale,
             help="Guidance rescale factor",
         )
+        parser.add_argument(
+            "--boundary-ratio",
+            type=float,
+            default=SamplingParam.boundary_ratio,
+            help="Boundary timestep ratio",
+        )
         parser.add_argument(
             "--save-video",
             action="store_true",
@@ -191,6 +200,25 @@ class SamplingParam:
             default=SamplingParam.image_path,
             help="Path to input image for image-to-video generation",
         )
+        parser.add_argument(
+            "--moba-config-path",
+            type=str,
+            default=None,
+            help=
+            "Path to a JSON file containing V-MoBA specific configurations.",
+        )
+        parser.add_argument(
+            "--return-trajectory-latents",
+            action="store_true",
+            default=SamplingParam.return_trajectory_latents,
+            help="Whether to return the trajectory",
+        )
+        parser.add_argument(
+            "--return-trajectory-decoded",
+            action="store_true",
+            default=SamplingParam.return_trajectory_decoded,
+            help="Whether to return the decoded trajectory",
+        )
         return parser
 
 

fastvideo/configs/sample/registry.py

@@ -37,6 +37,8 @@ SAMPLING_PARAM_REGISTRY: dict[str, Any] = {
     "Wan-AI/Wan2.1-T2V-14B-Diffusers": WanT2V_14B_SamplingParam,
     "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers": WanI2V_14B_480P_SamplingParam,
     "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers": WanI2V_14B_720P_SamplingParam,
+    "weizhou03/Wan2.1-Fun-1.3B-InP-Diffusers":
+    Wan2_1_Fun_1_3B_InP_SamplingParam,
 
     # Wan2.2
     "Wan-AI/Wan2.2-TI2V-5B-Diffusers": Wan2_2_TI2V_5B_SamplingParam,

fastvideo/configs/sample/wan.py

@@ -144,18 +144,22 @@ class Wan2_2_TI2V_5B_SamplingParam(Wan2_2_Base_SamplingParam):
 
 @dataclass
 class Wan2_2_T2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
-    guidance_scale: float = 4.0
-    guidance_scale_2: float = 3.0
+    guidance_scale: float = 4.0  # high_noise
+    guidance_scale_2: float = 3.0  # low_noise
     num_inference_steps: int = 40
     fps: int = 16
+    # NOTE(will): default boundary timestep is tracked by PipelineConfig, but
+    # can be overridden during sampling
 
 
 @dataclass
 class Wan2_2_I2V_A14B_SamplingParam(Wan2_2_Base_SamplingParam):
-    guidance_scale: float = 3.5
-    guidance_scale_2: float = 3.5
+    guidance_scale: float = 3.5  # high_noise
+    guidance_scale_2: float = 3.5  # low_noise
     num_inference_steps: int = 40
     fps: int = 16
+    # NOTE(will): default boundary timestep is tracked by PipelineConfig, but
+    # can be overridden during sampling
 
 
 # =============================================

fastvideo/dataset/__init__.py

@@ -4,7 +4,7 @@ from torchvision.transforms import Lambda
 
 from fastvideo.dataset.parquet_dataset_map_style import (
     build_parquet_map_style_dataloader)
-from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset
+from fastvideo.dataset.preprocessing_datasets import VideoCaptionMergedDataset, TextDataset
 from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255,
                                          TemporalRandomCrop)
 from fastvideo.dataset.validation_dataset import ValidationDataset
@@ -37,9 +37,15 @@ def getdataset(args) -> VideoCaptionMergedDataset:
                                      temporal_sample=temporal_sample,
                                      transform_topcrop=transform_topcrop,
                                      seed=args.seed)
+                                    
+
+def gettextdataset(args) -> TextDataset:
+    return TextDataset(data_merge_path=args.data_merge_path,
+                       args=args,
+                       seed=args.seed)
 
 
 __all__ = [
     "build_parquet_map_style_dataloader", "ValidationDataset",
-    "VideoCaptionMergedDataset"
+    "VideoCaptionMergedDataset", "TextDataset"
 ]

fastvideo/dataset/dataloader/parquet_io.py

@@ -0,0 +1,264 @@
+"""
+Utilities for converting preprocessing records (dicts) into Arrow tables and
+writing Parquet datasets in fixed-size chunks.
+
+This module centralizes table construction and Parquet file writing so
+pipelines only need to define their PyArrow schema and produce per-sample
+record dictionaries.
+
+Key APIs:
+- records_to_table(records, schema): Safely convert a list of dictionaries into
+  a pa.Table, casting to the provided schema.
+- ParquetDatasetWriter: Buffer tables and flush to a directory as multiple
+  Parquet files with a fixed number of rows per file. Uses temporary files and
+  atomic rename to avoid partially written outputs.
+"""
+
+from __future__ import annotations
+
+import multiprocessing
+import os
+from concurrent.futures import ProcessPoolExecutor
+from typing import Any
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+
+def records_to_table(records: list[dict[str, Any]], schema: pa.Schema) -> pa.Table:
+    """Build a PyArrow table from Python record dicts using an explicit schema.
+
+    Arrow will cast values to the target schema when possible (e.g., promoting
+    Python ints/floats to pa.int64/pa.float64), eliminating hand-written per-
+    field array construction.
+
+    Args:
+        records: List of dictionaries, each representing one row. Keys must
+            match schema field names.
+        schema: Target PyArrow schema. Controls field names and types.
+
+    Returns:
+        pa.Table: In-memory table matching the provided schema. If ``records``
+        is empty, returns an empty table with the given schema.
+    """
+    if not records:
+        return pa.table({}, schema=schema)
+    return pa.Table.from_pylist(records, schema=schema)
+
+
+class ParquetDatasetWriter:
+    """Accumulate tables and flush them to a Parquet directory in fixed-size chunks.
+
+    Behavior:
+    - Writes files under worker-specific subdirectories for parallelism.
+    - Uses temporary files and atomic rename to avoid partial files being left
+      behind on failure.
+    - Only full chunks of ``samples_per_file`` rows are written on each flush;
+      any remainder rows are re-buffered for the next flush.
+
+    Note:
+    - Instances are not meant to be shared across processes. Create one writer
+      per process if using multiprocessing.
+    """
+
+    def __init__(self, out_dir: str, samples_per_file: int, compression: str = "zstd") -> None:
+        """Initialize the dataset writer.
+
+        Args:
+            out_dir: Output directory where Parquet files will be written.
+            samples_per_file: Fixed number of rows per Parquet file.
+            compression: Compression codec passed to ``pyarrow.parquet.write_table``
+                (e.g., ``"zstd"``, ``"snappy"``, ``"gzip"``).
+        """
+        self.out_dir = out_dir
+        self.samples_per_file = max(int(samples_per_file), 1)
+        self.compression = compression
+        os.makedirs(self.out_dir, exist_ok=True)
+        self._tables: list[pa.Table] = []
+
+    def append_table(self, table: pa.Table) -> None:
+        """Append a non-empty table to the internal buffer.
+
+        Args:
+            table: A ``pa.Table`` to buffer. Empty or ``None`` tables are ignored.
+        """
+        if table is None or len(table) == 0:
+            return
+        self._tables.append(table)
+
+    def _combine(self) -> pa.Table | None:
+        """Combine all buffered tables into a single table, if any.
+
+        Returns:
+            A concatenated table, a single table if only one was buffered, or
+            ``None`` if no tables are buffered.
+        """
+        if not self._tables:
+            return None
+        if len(self._tables) == 1:
+            return self._tables[0]
+        return pa.concat_tables(self._tables, promote_options='none')
+
+    def flush(self, num_workers: int | None = None, write_remainder: bool = False) -> int:
+        """Write accumulated tables to disk and clear the written portion.
+
+        Only complete chunks of size ``samples_per_file`` are written. Any
+        remainder rows are kept buffered for the next flush.
+
+        Args:
+            num_workers: Optional override for the number of parallel workers
+                used to write chunks. Defaults to ``min(cpu_count, chunks)``.
+            write_remainder: If True, also write any leftover rows (< samples_per_file)
+                as a final small Parquet file (useful for the last flush at the
+                end of preprocessing).
+
+        Returns:
+            int: Number of rows successfully written in this flush call.
+        """
+        combined = self._combine()
+        self._tables = []
+        if combined is None or len(combined) == 0:
+            return 0
+
+        num_samples = len(combined)
+        total_chunks = num_samples // self.samples_per_file
+        if total_chunks == 0:
+            if not write_remainder:
+                # Not enough to form a full chunk; keep buffered for next round
+                # Re-buffer and return 0 written
+                self._tables = [combined]
+                return 0
+            # Last flush: write the small remainder as a final file in worker_0
+            worker_dir = os.path.join(self.out_dir, "worker_0")
+            os.makedirs(worker_dir, exist_ok=True)
+            # Determine next index
+            num_parquets = 0
+            for _, _, files in os.walk(worker_dir):
+                for file in files:
+                    if file.endswith('.parquet'):
+                        num_parquets += 1
+            chunk_path = os.path.join(worker_dir, f"data_chunk_{num_parquets}.parquet")
+            temp_path = chunk_path + '.tmp'
+            pq.write_table(combined, temp_path, compression=self.compression)
+            if os.path.exists(chunk_path):
+                os.remove(chunk_path)
+            os.rename(temp_path, chunk_path)
+            return num_samples
+
+        # Only write full chunks; keep remainder for next flush
+        written_rows = total_chunks * self.samples_per_file
+        remainder = num_samples - written_rows
+
+        table_to_write = combined.slice(0, written_rows)
+        remainder_table = combined.slice(written_rows, remainder) if remainder > 0 else None
+        if remainder_table is not None and len(remainder_table) > 0:
+            if write_remainder:
+                # Write the remainder as a final small file (worker_0)
+                worker_dir = os.path.join(self.out_dir, "worker_0")
+                os.makedirs(worker_dir, exist_ok=True)
+                num_parquets = 0
+                for _, _, files in os.walk(worker_dir):
+                    for file in files:
+                        if file.endswith('.parquet'):
+                            num_parquets += 1
+                remainder_path = os.path.join(worker_dir,
+                                              f"data_chunk_{num_parquets}.parquet")
+                temp_path = remainder_path + '.tmp'
+                pq.write_table(remainder_table,
+                               temp_path,
+                               compression=self.compression)
+                if os.path.exists(remainder_path):
+                    os.remove(remainder_path)
+                os.rename(temp_path, remainder_path)
+            else:
+                self._tables = [remainder_table]
+
+        # Parallel write by chunk ranges
+        if num_workers is None:
+            num_workers = min(multiprocessing.cpu_count(), max(total_chunks, 1))
+        num_workers = max(int(num_workers), 1)
+        chunks_per_worker = (total_chunks + num_workers - 1) // num_workers
+
+        work_ranges: list[tuple[int, int, pa.Table, int, str, int, str]] = []
+        for worker_id in range(num_workers):
+            start_chunk = worker_id * chunks_per_worker
+            end_chunk = min((worker_id + 1) * chunks_per_worker, total_chunks)
+            if start_chunk < end_chunk:
+                work_ranges.append(
+                    (
+                        start_chunk,
+                        end_chunk,
+                        table_to_write,
+                        worker_id,
+                        self.out_dir,
+                        self.samples_per_file,
+                        self.compression,
+                    )
+                )
+
+        written_total = 0
+        if len(work_ranges) == 1:
+            written_total += _process_chunk_range(work_ranges[0])
+            return written_total
+
+        with ProcessPoolExecutor(max_workers=num_workers) as executor:
+            futures = [executor.submit(_process_chunk_range, args) for args in work_ranges]
+            for f in futures:
+                written_total += f.result()
+        return written_total + (len(remainder_table) if write_remainder and remainder_table is not None else 0)
+
+
+def _process_chunk_range(args: Any) -> int:
+    """Worker function to write a contiguous range of chunk files.
+
+    Args:
+        args: Tuple containing
+            - start_chunk (int): inclusive start chunk index
+            - end_chunk (int): exclusive end chunk index
+            - table (pa.Table): concatenated table containing all rows to write
+            - worker_id (int): numeric worker identifier
+            - output_dir (str): base output directory
+            - samples_per_file (int): rows per chunk file
+            - compression (str): compression codec for Parquet
+
+    Returns:
+        int: Total number of rows written by this worker.
+    """
+    start_chunk, end_chunk, table, worker_id, output_dir, samples_per_file, compression = args
+    total_written = 0
+    num_samples = len(table)
+
+    worker_dir = os.path.join(output_dir, f"worker_{worker_id}")
+    os.makedirs(worker_dir, exist_ok=True)
+
+    # Offset to continue numbering if files exist
+    num_parquets = 0
+    for root, _, files in os.walk(worker_dir):
+        for file in files:
+            if file.endswith('.parquet'):
+                num_parquets += 1
+
+    for i in range(start_chunk, end_chunk):
+        start_sample = i * samples_per_file
+        end_sample = min((i + 1) * samples_per_file, num_samples)
+        if end_sample <= start_sample:
+            continue
+        chunk = table.slice(start_sample, end_sample - start_sample)
+
+        chunk_path = os.path.join(worker_dir, f"data_chunk_{i + num_parquets}.parquet")
+        temp_path = chunk_path + '.tmp'
+        try:
+            pq.write_table(chunk, temp_path, compression=compression)
+            if os.path.exists(chunk_path):
+                os.remove(chunk_path)
+            os.rename(temp_path, chunk_path)
+            total_written += len(chunk)
+        except Exception:
+            if os.path.exists(temp_path):
+                os.remove(temp_path)
+            raise
+
+    return total_written
+
+
+

fastvideo/dataset/dataloader/record_schema.py

@@ -1,5 +1,7 @@
 from typing import Any
 
+import numpy as np
+
 from fastvideo.pipelines.pipeline_batch_info import PreprocessBatch
 
 
@@ -120,3 +122,69 @@ def i2v_record_creator(batch: PreprocessBatch) -> list[dict[str, Any]]:
             })
 
     return records
+
+
+def ode_text_only_record_creator(
+        video_name: str, text_embedding: np.ndarray, caption: str,
+        trajectory_latents: np.ndarray,
+        trajectory_timesteps: np.ndarray) -> dict[str, Any]:
+    """Create a text-only ODE trajectory record matching pyarrow_schema_ode_trajectory_text_only.
+
+    Args:
+        video_name: Base name/id for the sample (without extension).
+        text_embedding: Text encoder output array [SeqLen, Dim].
+        caption: Original text prompt.
+        trajectory_latents: Collected trajectory latents array.
+        trajectory_timesteps: Collected timesteps array.
+
+    Returns:
+        dict suitable for records_to_table(…, pyarrow_schema_ode_trajectory_text_only)
+    """
+    assert trajectory_latents is not None, "trajectory_latents is required"
+    assert trajectory_timesteps is not None, "trajectory_timesteps is required"
+
+    record = {
+        "id": f"text_{video_name}",
+        "text_embedding_bytes": text_embedding.tobytes(),
+        "text_embedding_shape": list(text_embedding.shape),
+        "text_embedding_dtype": str(text_embedding.dtype),
+        "file_name": video_name,
+        "caption": caption,
+        "media_type": "text",
+    }
+
+    record.update({
+        "trajectory_latents_bytes": trajectory_latents.tobytes(),
+        "trajectory_latents_shape": list(trajectory_latents.shape),
+        "trajectory_latents_dtype": str(trajectory_latents.dtype),
+    })
+
+    record.update({
+        "trajectory_timesteps_bytes": trajectory_timesteps.tobytes(),
+        "trajectory_timesteps_shape": list(trajectory_timesteps.shape),
+        "trajectory_timesteps_dtype": str(trajectory_timesteps.dtype),
+    })
+
+    return record
+
+
+def text_only_record_creator(text_name: str, text_embedding: np.ndarray,
+                             caption: str) -> dict[str, Any]:
+    """Create a text-only record matching pyarrow_schema_text_only.
+
+    Args:
+        text_name: Base id/name for the text sample.
+        text_embedding: Text encoder output array [SeqLen, Dim].
+        caption: Original text prompt.
+
+    Returns:
+        dict suitable for records_to_table(…, pyarrow_schema_text_only)
+    """
+    record = {
+        "id": f"text_{text_name}",
+        "text_embedding_bytes": text_embedding.tobytes(),
+        "text_embedding_shape": list(text_embedding.shape),
+        "text_embedding_dtype": str(text_embedding.dtype),
+        "caption": caption,
+    }
+    return record

fastvideo/dataset/dataloader/schema.py

@@ -50,6 +50,7 @@ pyarrow_schema_i2v = pa.schema([
     pa.field("fps", pa.float64()),
 ])
 
+
 pyarrow_schema_t2v = pa.schema([
     pa.field("id", pa.string()),
     # --- Image/Video VAE latents ---
@@ -78,3 +79,40 @@ pyarrow_schema_t2v = pa.schema([
     pa.field("duration_sec", pa.float64()),
     pa.field("fps", pa.float64()),
 ])
+
+
+pyarrow_schema_ode_trajectory_text_only = pa.schema([
+    pa.field("id", pa.string()),
+    # --- Text encoder output tensor ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("text_embedding_bytes", pa.binary()),
+    # e.g., [SeqLen, Dim]
+    pa.field("text_embedding_shape", pa.list_(pa.int64())),
+    # e.g., 'bfloat16' or 'float32'
+    pa.field("text_embedding_dtype", pa.string()),
+    # --- ODE Trajectory ---
+    pa.field("trajectory_latents_bytes", pa.binary()),
+    pa.field("trajectory_latents_shape", pa.list_(pa.int64())),
+    pa.field("trajectory_latents_dtype", pa.string()),
+    pa.field("trajectory_timesteps_bytes", pa.binary()),
+    pa.field("trajectory_timesteps_shape", pa.list_(pa.int64())),
+    pa.field("trajectory_timesteps_dtype", pa.string()),
+    # --- Metadata ---
+    pa.field("file_name", pa.string()),
+    pa.field("caption", pa.string()),
+    pa.field("media_type", pa.string()),  # Always 'text' for text-only
+])
+
+
+pyarrow_schema_text_only = pa.schema([
+    pa.field("id", pa.string()),
+    # --- Text encoder output tensor ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("text_embedding_bytes", pa.binary()),
+    # e.g., [SeqLen, Dim]
+    pa.field("text_embedding_shape", pa.list_(pa.int64())),
+    # e.g., 'bfloat16' or 'float32'
+    pa.field("text_embedding_dtype", pa.string()),
+    # --- Metadata ---
+    pa.field("caption", pa.string()),
+])

fastvideo/dataset/preprocessing_datasets.py

@@ -628,3 +628,134 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset,
     def load_state_dict(self, state_dict: dict[str, Any]) -> None:
         """Load state dict from checkpoint."""
         self.processed_batches = state_dict["processed_batches"]
+
+
+class TextDataset(torch.utils.data.IterableDataset,
+                  torch.distributed.checkpoint.stateful.Stateful):
+    """
+    Text-only dataset for processing prompts from a simple text file.
+    
+    Assumes that data_merge_path is a text file with one prompt per line:
+    A cat playing with a ball
+    A dog running in the park
+    A person cooking dinner
+    ...
+    
+    This dataset processes text data through text encoding stages only.
+    """
+
+    def __init__(self,
+                 data_merge_path: str,
+                 args,
+                 start_idx: int = 0,
+                 seed: int = 42):
+        self.data_merge_path = data_merge_path
+        self.start_idx = start_idx
+        self.args = args
+        self.seed = seed
+
+        # Initialize tokenizer
+        tokenizer_path = os.path.join(args.model_path, "tokenizer")
+        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,
+                                                  cache_dir=args.cache_dir)
+
+        # Initialize text encoding stage
+        self.text_encoding_stage = TextEncodingStage(
+            tokenizer=tokenizer,
+            text_max_length=args.text_max_length,
+            cfg_rate=getattr(args, 'training_cfg_rate', 0.0),
+            seed=self.seed)
+
+        # Process text data
+        self.processed_batches = self._process_text_data()
+
+    def _load_text_data(self) -> list[str]:
+        """Load text prompts from file."""
+        prompts = []
+        with open(self.data_merge_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                line = line.strip()
+                if line:  # Skip empty lines
+                    prompts.append(line)
+        
+        logger.info(f"Loaded {len(prompts)} text prompts from {self.data_merge_path}")
+        return prompts
+
+    def _process_text_data(self) -> list[PreprocessBatch]:
+        """Process the text prompts through text encoding stage."""
+        raw_prompts = self._load_text_data()
+        processed_batches = []
+
+        for idx, prompt in enumerate(raw_prompts):
+            # Create a text-only batch with dummy path
+            batch = PreprocessBatch(
+                path=f"text_prompt_{idx}",
+                cap=[prompt],  # TextEncodingStage expects a list
+                resolution=None,
+                fps=None,
+                duration=None,
+                num_frames=0,
+                sample_frame_index=None,
+                sample_num_frames=0
+            )
+            
+            processed_batches.append(batch)
+
+        logger.info(f"Processed {len(processed_batches)} text batches")
+        return processed_batches
+
+    def __iter__(self):
+        """Iterator for the dataset."""
+        # Set up distributed sampling if needed
+        if torch.distributed.is_available() and torch.distributed.is_initialized():
+            rank = torch.distributed.get_rank()
+            world_size = torch.distributed.get_world_size()
+        else:
+            rank = 0
+            world_size = 1
+
+        # Calculate chunk for this rank
+        total_items = len(self.processed_batches)
+        items_per_rank = math.ceil(total_items / world_size)
+        start_idx = rank * items_per_rank + self.start_idx
+        end_idx = min(start_idx + items_per_rank, total_items)
+
+        # Yield items for this rank
+        for idx in range(start_idx, end_idx):
+            if idx < len(self.processed_batches):
+                yield self._get_item(idx)
+
+    def _get_item(self, idx: int) -> dict:
+        """Get a single processed text item."""
+        batch = self.processed_batches[idx]
+
+        # Apply text encoding stage
+        batch = self.text_encoding_stage.process(batch)
+
+        # Build result dictionary for text-only processing with required schema fields
+        result = {
+            "text": batch.text,
+            "input_ids": batch.input_ids,
+            "cond_mask": batch.cond_mask,
+            "path": batch.path,
+            # Required schema fields for ODE trajectory processing
+            "id": f"text_{idx}",
+            "file_name": batch.path,
+            "caption": batch.text,
+            "media_type": "text",
+            "width": 1,
+            "height": 1,
+            "num_frames": 0,
+            "duration_sec": 0.0,
+            "fps": 0.0,
+        }
+
+        return result
+
+    def state_dict(self) -> dict[str, Any]:
+        """Return state dict for checkpointing."""
+        return {"processed_batches": self.processed_batches}
+
+    def load_state_dict(self, state_dict: dict[str, Any]) -> None:
+        """Load state dict from checkpoint."""
+        self.processed_batches = state_dict["processed_batches"]

fastvideo/dataset/utils.py

@@ -5,7 +5,7 @@ import numpy as np
 import torch
 
 
-def pad(t: torch.Tensor, padding_length: int) -> torch.Tensor:
+def pad(t: torch.Tensor, padding_length: int) -> tuple[torch.Tensor, torch.Tensor]:
     """
     Pad or crop an embedding [L, D] to exactly padding_length tokens.
     Return:

fastvideo/entrypoints/video_generator.py

@@ -344,6 +344,9 @@ class VideoGenerator:
                 "size": (target_height, target_width, batch.num_frames),
                 "generation_time": gen_time,
                 "logging_info": logging_info,
+                "trajectory": output_batch.trajectory_latents,
+                "trajectory_timesteps": output_batch.trajectory_timesteps,
+                "trajectory_decoded": output_batch.trajectory_decoded,
             }
 
     def set_lora_adapter(self,

fastvideo/envs.py

@@ -175,6 +175,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # - "SLIDING_TILE_ATTN" : use Sliding Tile Attention
     # - "VIDEO_SPARSE_ATTN": use Video Sparse Attention
     # - "SAGE_ATTN": use Sage Attention
+    # - "SAGE_ATTN_THREE": use Sage Attention 3
     "FASTVIDEO_ATTENTION_BACKEND":
     lambda: os.getenv("FASTVIDEO_ATTENTION_BACKEND", None),
 

fastvideo/fastvideo_args.py

@@ -1,9 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # Inspired by SGLang: https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/server_args.py
 """The arguments of FastVideo Inference."""
-
 import argparse
 import dataclasses
+import json
 from contextlib import contextmanager
 from dataclasses import field
 from enum import Enum
@@ -119,7 +119,7 @@ class FastVideoArgs:
     output_type: str = "pil"
 
     # CPU offload parameters
-    dit_cpu_offload: bool = False
+    dit_cpu_offload: bool = True
     use_fsdp_inference: bool = True
     text_encoder_cpu_offload: bool = True
     image_encoder_cpu_offload: bool = True
@@ -139,6 +139,10 @@ class FastVideoArgs:
     # VSA parameters
     VSA_sparsity: float = 0.0  # inference/validation sparsity
 
+    # V-MoBA parameters
+    moba_config_path: str | None = None
+    moba_config: dict[str, Any] = field(default_factory=dict)
+
     # Master port for distributed training/inference
     master_port: int | None = None
 
@@ -154,6 +158,7 @@ class FastVideoArgs:
         "transformer": True,
         "vae": True,
     })
+    override_transformer_cls_name: str | None = None
 
     # # DMD parameters
     # dmd_denoising_steps: List[int] | None = field(default=None)
@@ -166,6 +171,16 @@ class FastVideoArgs:
         return not self.inference_mode
 
     def __post_init__(self):
+        if self.moba_config_path:
+            try:
+                with open(self.moba_config_path) as f:
+                    self.moba_config = json.load(f)
+                logger.info("Loaded V-MoBA config from %s",
+                            self.moba_config_path)
+            except (FileNotFoundError, json.JSONDecodeError) as e:
+                logger.error("Failed to load V-MoBA config from %s: %s",
+                             self.moba_config_path, e)
+                raise
         self.check_fastvideo_args()
 
     @staticmethod
@@ -382,6 +397,12 @@ class FastVideoArgs:
             default=FastVideoArgs.enable_stage_verification,
             help="Enable input/output verification for pipeline stages",
         )
+        parser.add_argument(
+            "--override-transformer-cls-name",
+            type=str,
+            default=FastVideoArgs.override_transformer_cls_name,
+            help="Override transformer cls name",
+        )
         # Add pipeline configuration arguments
         PipelineConfig.add_cli_args(parser)
 
@@ -691,6 +712,9 @@ class TrainingArgs(FastVideoArgs):
     enable_gradient_masking: bool = True
     gradient_mask_last_n_frames: int = 21
     validate_cache_structure: bool = False  # Debug flag for cache validation
+    same_step_across_blocks: bool = False  # Use same exit timestep for all blocks
+    last_step_only: bool = False  # Only use the last timestep for training
+    context_noise: int = 0  # Context noise level for cache updates
 
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace) -> "TrainingArgs":
@@ -1024,6 +1048,15 @@ class TrainingArgs(FastVideoArgs):
         parser.add_argument("--lora-rank", type=int, help="LoRA rank")
         parser.add_argument("--lora-alpha", type=int, help="LoRA alpha")
 
+        # V-MoBA parameters
+        parser.add_argument(
+            "--moba-config-path",
+            type=str,
+            default=None,
+            help=
+            "Path to a JSON file containing V-MoBA specific configurations.",
+        )
+
         # Distillation arguments
         parser.add_argument("--generator-update-interval",
                             type=int,
@@ -1099,6 +1132,18 @@ class TrainingArgs(FastVideoArgs):
             "--validate-cache-structure",
             action=StoreBoolean,
             help="Whether to validate KV cache structure (debug flag)")
+        parser.add_argument(
+            "--same-step-across-blocks",
+            action=StoreBoolean,
+            help="Whether to use the same exit timestep for all blocks")
+        parser.add_argument(
+            "--last-step-only",
+            action=StoreBoolean,
+            help="Whether to only use the last timestep for training")
+        parser.add_argument("--context-noise",
+                            type=int,
+                            default=TrainingArgs.context_noise,
+                            help="Context noise level for cache updates")
 
         return parser
 

fastvideo/layers/layernorm.py

@@ -100,7 +100,16 @@ class ScaleResidual(nn.Module):
     def forward(self, residual: torch.Tensor, x: torch.Tensor,
                 gate: torch.Tensor) -> torch.Tensor:
         """Apply gated residual connection."""
-        return residual + x * gate
+        # x.shape: [batch_size, seq_len, inner_dim]
+        if gate.dim() == 4:
+            # gate.shape: [batch_size, num_frames, 1, inner_dim]
+            num_frames = gate.shape[1]
+            frame_seqlen = x.shape[1] // num_frames
+            return residual + (x.unflatten(
+                dim=1, sizes=(num_frames, frame_seqlen)) * gate).flatten(1, 2)
+        else:
+            # gate.shape: [batch_size, 1, inner_dim]
+            return residual + x * gate
 
 
 # adapted from Diffusers: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/normalization.py
@@ -159,7 +168,7 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
             raise NotImplementedError(f"Norm type {norm_type} not implemented")
 
     def forward(self, residual: torch.Tensor, x: torch.Tensor,
-                gate: torch.Tensor, shift: torch.Tensor,
+                gate: torch.Tensor | int, shift: torch.Tensor,
                 scale: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Apply gated residual connection, followed by layernorm and 
@@ -171,12 +180,41 @@ class ScaleResidualLayerNormScaleShift(nn.Module):
             - residual value (value after residual connection 
               but before normalization)
         """
+        # x.shape: [batch_size, seq_len, inner_dim]
         # Apply residual connection with gating
-        residual_output = residual + x * gate
+        if isinstance(gate, int):
+            # used by cross-attention, should be 1
+            assert gate == 1
+            residual_output = residual + x
+        elif isinstance(gate, torch.Tensor):
+            if gate.dim() == 4:
+                # gate.shape: [batch_size, num_frames, 1, inner_dim]
+                num_frames = gate.shape[1]
+                frame_seqlen = x.shape[1] // num_frames
+                residual_output = residual + (
+                    x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                    gate).flatten(1, 2)
+            else:
+                # used by bidirectional self attention
+                # gate.shape: [batch_size, 1, inner_dim]
+                residual_output = residual + x * gate
+        else:
+            raise ValueError(f"Gate type {type(gate)} not supported")
+        # residual_output.shape: [batch_size, seq_len, inner_dim]
+
         # Apply normalization
         normalized = self.norm(residual_output)
         # Apply scale and shift
-        modulated = normalized * (1.0 + scale) + shift
+        if isinstance(scale, torch.Tensor) and scale.dim() == 4:
+            # scale.shape: [batch_size, num_frames, 1, inner_dim]
+            # shift.shape: [batch_size, num_frames, 1, inner_dim]
+            num_frames = scale.shape[1]
+            frame_seqlen = normalized.shape[1] // num_frames
+            modulated = (
+                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                (1.0 + scale) + shift).flatten(1, 2)
+        else:
+            modulated = normalized * (1.0 + scale) + shift
         return modulated, residual_output
 
 
@@ -218,8 +256,24 @@ class LayerNormScaleShift(nn.Module):
     def forward(self, x: torch.Tensor, shift: torch.Tensor,
                 scale: torch.Tensor) -> torch.Tensor:
         """Apply ln followed by scale and shift in a single fused operation."""
+        # x.shape: [batch_size, seq_len, inner_dim]
         normalized = self.norm(x)
         if self.compute_dtype == torch.float32:
-            return (normalized.float() * (1.0 + scale) + shift).to(x.dtype)
+            normalized = normalized.float()
+
+        if scale.dim() == 4:
+            # scale.shape: [batch_size, num_frames, 1, inner_dim]
+            num_frames = scale.shape[1]
+            frame_seqlen = normalized.shape[1] // num_frames
+            output = (
+                normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                (1.0 + scale) + shift).flatten(1, 2)
         else:
-            return normalized * (1.0 + scale) + shift
+            # scale.shape: [batch_size, 1, inner_dim]
+            # shift.shape: [batch_size, 1, inner_dim]
+            output = normalized * (1.0 + scale) + shift
+
+        if self.compute_dtype == torch.float32:
+            output = output.to(x.dtype)
+
+        return output

fastvideo/layers/lora/linear.py

@@ -63,7 +63,7 @@ class BaseLayerWithLoRA(nn.Module):
                             device=self.base_layer.weight.device,
                             dtype=self.base_layer.weight.dtype))
             torch.nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
-            torch.nn.init.kaiming_uniform_(self.lora_B, a=math.sqrt(5))
+            torch.nn.init.zeros_(self.lora_B)
         else:
             self.lora_A = None
             self.lora_B = None
@@ -77,9 +77,11 @@ class BaseLayerWithLoRA(nn.Module):
             lora_A = self.lora_A.to_local()
 
         if not self.merged and not self.disable_lora:
-            delta = x @ (
-                self.slice_lora_b_weights(lora_B.to(x, non_blocking=True))
-                @ self.slice_lora_a_weights(lora_A.to(x, non_blocking=True)))
+            lora_A_sliced = self.slice_lora_a_weights(
+                lora_A.to(x, non_blocking=True))
+            lora_B_sliced = self.slice_lora_b_weights(
+                lora_B.to(x, non_blocking=True))
+            delta = x @ lora_A_sliced.T @ lora_B_sliced.T
             if self.lora_alpha != self.lora_rank:
                 delta = delta * (
                     self.lora_alpha / self.lora_rank  # type: ignore

fastvideo/layers/rotary_embedding.py

@@ -33,6 +33,7 @@ from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
 
+
 def _rotate_neox(x: torch.Tensor) -> torch.Tensor:
     x1 = x[..., :x.shape[-1] // 2]
     x2 = x[..., x.shape[-1] // 2:]

fastvideo/models/dits/causal_wanvideo.py

@@ -147,6 +147,9 @@ class CausalWanSelfAttention(nn.Module):
                 # Assign new keys/values directly up to current_end
                 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
+                kv_cache["k"] = kv_cache["k"].detach()
+                kv_cache["v"] = kv_cache["v"].detach()
+                # logger.info("kv_cache['k'] is in comp graph: %s", kv_cache["k"].requires_grad or kv_cache["k"].grad_fn is not None)
                 kv_cache["k"][:, local_start_index:local_end_index] = roped_key
                 kv_cache["v"][:, local_start_index:local_end_index] = v
             x = self.attn(
@@ -244,19 +247,26 @@ class CausalWanTransformerBlock(nn.Module):
         current_start: int = 0,
         cache_start: int | None = None,
     ) -> torch.Tensor:
+        # hidden_states.shape: [batch_size, seq_length, inner_dim]
+        # temb.shape: [batch_size, num_frames, 6, inner_dim]
         if hidden_states.dim() == 4:
             hidden_states = hidden_states.squeeze(1)
+        num_frames = temb.shape[1]
+        frame_seqlen = hidden_states.shape[1] // num_frames
         bs, seq_length, _ = hidden_states.shape
         orig_dtype = hidden_states.dtype
         # assert orig_dtype != torch.float32
         e = self.scale_shift_table + temb.float()
+        # e.shape: [batch_size, num_frames, 6, inner_dim]
+        assert e.shape == (bs, num_frames, 6, self.hidden_dim)
         shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = e.chunk(
-            6, dim=1)
+            6, dim=2)
+        # *_msa.shape: [batch_size, num_frames, 1, inner_dim]
         assert shift_msa.dtype == torch.float32
 
         # 1. Self-attention
-        norm_hidden_states = (self.norm1(hidden_states.float()) *
-                              (1 + scale_msa) + shift_msa).to(orig_dtype)
+        norm_hidden_states = (self.norm1(hidden_states.float()).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
+                        (1 + scale_msa) + shift_msa).flatten(1, 2).to(orig_dtype)
         query, _ = self.to_q(norm_hidden_states)
         key, _ = self.to_k(norm_hidden_states)
         value, _ = self.to_v(norm_hidden_states)
@@ -368,7 +378,8 @@ class CausalWanTransformer3DModel(BaseDiT):
 
         # Causal-specific
         self.block_mask = None
-        self.num_frame_per_block = 1
+        self.num_frame_per_block = config.arch_config.num_frames_per_block
+        assert self.num_frame_per_block <= 3
         self.independent_first_frame = False
 
         self.__post_init__()
@@ -487,8 +498,8 @@ class CausalWanTransformer3DModel(BaseDiT):
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
 
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
-            timestep, encoder_hidden_states, encoder_hidden_states_image)
-        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
+        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
 
         if encoder_hidden_states_image is not None:
             encoder_hidden_states = torch.concat(
@@ -526,8 +537,9 @@ class CausalWanTransformer3DModel(BaseDiT):
                                         **causal_kwargs)
 
         # 5. Output norm, projection & unpatchify
-        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2,
-                                                                          dim=1)
+        temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
+        shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2,
+                                                                    dim=2)
         hidden_states = self.norm_out(hidden_states, shift, scale)
         hidden_states = self.proj_out(hidden_states)
 
@@ -596,8 +608,8 @@ class CausalWanTransformer3DModel(BaseDiT):
         hidden_states = hidden_states.flatten(2).transpose(1, 2)
 
         temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
-            timestep, encoder_hidden_states, encoder_hidden_states_image)
-        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+                        timestep.flatten(), encoder_hidden_states, encoder_hidden_states_image)
+        timestep_proj = timestep_proj.unflatten(1, (6, self.hidden_size)).unflatten(dim=0, sizes=timestep.shape)
 
         if encoder_hidden_states_image is not None:
             encoder_hidden_states = torch.concat(
@@ -623,8 +635,9 @@ class CausalWanTransformer3DModel(BaseDiT):
                                         block_mask=self.block_mask)
 
         # 5. Output norm, projection & unpatchify
-        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2,
-                                                                          dim=1)
+        temb = temb.unflatten(dim=0, sizes=timestep.shape).unsqueeze(2)
+        shift, scale = (self.scale_shift_table.unsqueeze(1) + temb).chunk(2,
+                                                                    dim=2)
         hidden_states = self.norm_out(hidden_states, shift, scale)
         hidden_states = self.proj_out(hidden_states)
 

fastvideo/models/loader/component_loader.py

@@ -416,6 +416,11 @@ class TransformerLoader(ComponentLoader):
                 "Model config does not contain a _class_name attribute. "
                 "Only diffusers format is supported.")
 
+        logger.info("transformer cls_name: %s", cls_name)
+        if fastvideo_args.override_transformer_cls_name is not None:
+            cls_name = fastvideo_args.override_transformer_cls_name
+            logger.info("Overriding transformer cls_name to %s", cls_name)
+
         fastvideo_args.model_paths["transformer"] = model_path
 
         # Config from Diffusers supersedes fastvideo's model config
@@ -435,13 +440,19 @@ class TransformerLoader(ComponentLoader):
         use_custom_weights = (custom_weights_path and os.path.exists(custom_weights_path) and 
                             fastvideo_args.training_mode and 
                             not hasattr(fastvideo_args, '_loading_teacher_critic_model'))
-        
+
         if use_custom_weights:
             logger.info("Using custom initialization weights from: %s", custom_weights_path)
-            safetensors_list = [custom_weights_path]
+            assert custom_weights_path is not None, "Custom initialization weights must be provided"
+            if os.path.isdir(custom_weights_path):
+                safetensors_list = glob.glob(
+                    os.path.join(str(custom_weights_path), "*.safetensors"))
+            else:
+                assert custom_weights_path.endswith(".safetensors"), "Custom initialization weights must be a safetensors file"
+                safetensors_list = [custom_weights_path]
 
-        logger.info("Loading model from %s safetensors files in %s",
-                    len(safetensors_list), model_path)
+        logger.info("Loading model from %s safetensors files: %s",
+                    len(safetensors_list), safetensors_list)
 
         default_dtype = PRECISION_TO_TYPE[
             fastvideo_args.pipeline_config.dit_precision]
@@ -464,6 +475,7 @@ class TransformerLoader(ComponentLoader):
             pin_cpu_memory=fastvideo_args.pin_cpu_memory,
             fsdp_inference=fastvideo_args.use_fsdp_inference,
             # TODO(will): make these configurable
+            default_dtype=default_dtype,
             param_dtype=torch.bfloat16,
             reduce_dtype=torch.float32,
             output_dtype=None,
@@ -473,9 +485,8 @@ class TransformerLoader(ComponentLoader):
         total_params = sum(p.numel() for p in model.parameters())
         logger.info("Loaded model with %.2fB parameters", total_params / 1e9)
 
-        dtypes = set(param.dtype for param in model.parameters())
-        if len(dtypes) > 1:
-            model = model.to(default_dtype)
+        assert next(model.parameters()).dtype == default_dtype, "Model dtype does not match default dtype"
+
         model = model.eval()
         return model
 

fastvideo/models/loader/fsdp_load.py

@@ -62,6 +62,7 @@ def maybe_load_fsdp_model(
     device: torch.device,
     hsdp_replicate_dim: int,
     hsdp_shard_dim: int,
+    default_dtype: torch.dtype,
     param_dtype: torch.dtype,
     reduce_dtype: torch.dtype,
     cpu_offload: bool = False,
@@ -87,7 +88,8 @@ def maybe_load_fsdp_model(
         mp_policy=mp_policy,
     )
 
-    with set_default_dtype(param_dtype), torch.device("meta"):
+    logger.info("Loading model with default_dtype: %s", default_dtype)
+    with set_default_dtype(default_dtype), torch.device("meta"):
         model = model_cls(**init_params)
 
     # Check if we should use FSDP
@@ -125,7 +127,7 @@ def maybe_load_fsdp_model(
         model,
         weight_iterator,
         device,
-        param_dtype,
+        default_dtype,
         strict=True,
         cpu_offload=cpu_offload,
         param_names_mapping=param_names_mapping_fn,
@@ -248,13 +250,6 @@ def load_model_from_full_model_state_dict(
     sharded_sd = {}
     custom_param_sd, reverse_param_names_mapping = hf_to_custom_state_dict(
         full_sd_iterator, param_names_mapping)  # type: ignore
-    print(custom_param_sd.keys())
-    print("--------------------------------")
-    print("--------------------------------")
-    print("--------------------------------")
-    print("--------------------------------")
-    print("--------------------------------")
-    print(meta_sd.keys())
     for target_param_name, full_tensor in custom_param_sd.items():
         meta_sharded_param = meta_sd.get(target_param_name)
         if meta_sharded_param is None:

fastvideo/models/registry.py

@@ -61,6 +61,9 @@ _SCHEDULERS = {
      "FlowMatchEulerDiscreteScheduler"),
     "UniPCMultistepScheduler":
     ("schedulers", "scheduling_unipc_multistep", "UniPCMultistepScheduler"),
+    "SelfForcingFlowMatchScheduler":
+    ("schedulers", "scheduling_self_forcing_flow_match",
+     "SelfForcingFlowMatchScheduler"),
 }
 
 _FAST_VIDEO_MODELS = {

fastvideo/models/schedulers/scheduling_flow_match_euler_discrete.py

@@ -635,8 +635,31 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin,
         noise: torch.Tensor,
         timestep: torch.IntTensor,
     ) -> torch.Tensor:
+
+        """
+        Args:
+            clean_latent: the clean latent with shape [B, C, H, W],
+                where B is batch_size or batch_size * num_frames
+            noise: the noise with shape [B, C, H, W]
+            timestep: the timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+
+        Returns:
+            the corrupted latent with shape [B, C, H, W]
+        """
+        # If timestep is [bs, num_frames]
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+            assert timestep.numel() == clean_latent.shape[0]
+        elif timestep.ndim == 1:
+            # If timestep is [1]
+            if timestep.shape[0] == 1:
+                timestep = timestep.expand(clean_latent.shape[0])
+            else:
+                assert timestep.numel() == clean_latent.shape[0]
+        else:
+            raise ValueError(f"[add_noise] Invalid timestep shape: {timestep.shape}")
+        # timestep shape should be [B]
         self.sigmas = self.sigmas.to(noise.device)
-        timestep = timestep.expand(clean_latent.shape[0])
         self.timesteps = self.timesteps.to(noise.device)
         timestep_id = torch.argmin(
             (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)

fastvideo/models/schedulers/scheduling_self_forcing_flow_match.py

@@ -0,0 +1,131 @@
+# SPDX-License-Identifier: Apache-2.0
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput
+import torch
+
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.base import BaseScheduler
+
+logger = init_logger(__name__)
+
+
+class SelfForcingFlowMatchSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+    prev_sample: torch.FloatTensor
+
+class SelfForcingFlowMatchScheduler(BaseScheduler, ConfigMixin, SchedulerMixin):
+    
+    config_name = "scheduler_config.json"
+    order = 1
+    @register_to_config
+    def __init__(self, num_inference_steps=100, num_train_timesteps=1000, shift=3.0, sigma_max=1.0, sigma_min=0.003 / 1.002, inverse_timesteps=False, extra_one_step=False, reverse_sigmas=False, training=False):
+        self.num_train_timesteps = num_train_timesteps
+        self.shift = shift
+        self.sigma_max = sigma_max
+        self.sigma_min = sigma_min
+        self.inverse_timesteps = inverse_timesteps
+        self.extra_one_step = extra_one_step
+        self.reverse_sigmas = reverse_sigmas
+        self.set_timesteps(num_inference_steps, training=training)
+
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0, training=False, return_dict=False, **kwargs):
+        sigma_start = self.sigma_min + \
+            (self.sigma_max - self.sigma_min) * denoising_strength
+        if self.extra_one_step:
+            self.sigmas = torch.linspace(
+                sigma_start, self.sigma_min, num_inference_steps + 1)[:-1]
+        else:
+            self.sigmas = torch.linspace(
+                sigma_start, self.sigma_min, num_inference_steps)
+        if self.inverse_timesteps:
+            self.sigmas = torch.flip(self.sigmas, dims=[0])
+        self.sigmas = self.shift * self.sigmas / \
+            (1 + (self.shift - 1) * self.sigmas)
+        if self.reverse_sigmas:
+            self.sigmas = 1 - self.sigmas
+        self.timesteps = self.sigmas * self.num_train_timesteps
+        if training:
+            x = self.timesteps
+            y = torch.exp(-2 * ((x - num_inference_steps / 2) /
+                          num_inference_steps) ** 2)
+            y_shifted = y - y.min()
+            bsmntw_weighing = y_shifted * \
+                (num_inference_steps / y_shifted.sum())
+            self.linear_timesteps_weights = bsmntw_weighing
+
+    def step(self, model_output: torch.FloatTensor, timestep: torch.FloatTensor, sample: torch.FloatTensor, to_final=False, return_dict=False, **kwargs):
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        elif timestep.ndim == 0:
+            # handles the case where timestep is a scalar, this occurs when we
+            # use this scheduler for ODE trajectory
+            timestep = timestep.unsqueeze(0)
+
+        self.sigmas = self.sigmas.to(model_output.device)
+        self.timesteps = self.timesteps.to(model_output.device)
+        timestep = timestep.to(model_output.device)
+
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
+        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
+        if to_final or (timestep_id + 1 >= len(self.timesteps)).any():
+            sigma_ = 1 if (
+                self.inverse_timesteps or self.reverse_sigmas) else 0
+        else:
+            sigma_ = self.sigmas[timestep_id + 1].reshape(-1, 1, 1, 1)
+        prev_sample = sample + model_output * (sigma_ - sigma)
+        if isinstance(prev_sample, torch.Tensor | float) and not return_dict:
+            return (prev_sample, )
+        return SelfForcingFlowMatchSchedulerOutput(prev_sample=prev_sample)
+
+    def add_noise(self, original_samples, noise, timestep):
+        """
+        Diffusion forward corruption process.
+        Input:
+            - clean_latent: the clean latent with shape [B*T, C, H, W]
+            - noise: the noise with shape [B*T, C, H, W]
+            - timestep: the timestep with shape [B*T]
+        Output: the corrupted latent with shape [B*T, C, H, W]
+        """
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        self.sigmas = self.sigmas.to(noise.device)
+        self.timesteps = self.timesteps.to(noise.device)
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
+        sigma = self.sigmas[timestep_id].reshape(-1, 1, 1, 1)
+        sample = (1 - sigma) * original_samples + sigma * noise
+        return sample.type_as(noise)
+
+    def training_target(self, sample, noise, timestep):
+        target = noise - sample
+        return target
+
+    def training_weight(self, timestep):
+        """
+        Input:
+            - timestep: the timestep with shape [B*T]
+        Output: the corresponding weighting [B*T]
+        """
+        if timestep.ndim == 2:
+            timestep = timestep.flatten(0, 1)
+        self.linear_timesteps_weights = self.linear_timesteps_weights.to(timestep.device)
+        timestep_id = torch.argmin(
+            (self.timesteps.unsqueeze(1) - timestep.unsqueeze(0)).abs(), dim=0)
+        weights = self.linear_timesteps_weights[timestep_id]
+        return weights
+    
+    def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor:
+        return sample
+
+    def set_shift(self, shift: float) -> None:
+        self.shift = shift
+    

fastvideo/models/utils.py

@@ -145,14 +145,36 @@ def pred_noise_to_pred_video(pred_noise: torch.Tensor,
                              scheduler: Any) -> torch.Tensor:
     """
     Convert predicted noise to clean latent.
+
+    Args:
+    pred_noise: the predicted noise with shape [B, C, H, W]
+        where B is batch_size or batch_size * num_frames
+    noise_input_latent: the noisy latent with shape [B, C, H, W],
+    timestep: the timestep with shape [1] or [bs * num_frames] or [bs, num_frames]
+    scheduler: the scheduler
+
+    Returns:
+        the predicted video with shape [B, C, H, W]
     """
-    timestep = timestep.expand(noise_input_latent.shape[0])
+    # If timestep is [bs, num_frames]
+    if timestep.ndim == 2:
+        timestep = timestep.flatten(0, 1)
+        assert timestep.numel() == noise_input_latent.shape[0]
+    elif timestep.ndim == 1:
+        # If timestep is [1]
+        if timestep.shape[0] == 1:
+            timestep = timestep.expand(noise_input_latent.shape[0])
+        else:
+            assert timestep.numel() == noise_input_latent.shape[0]
+    else:
+        raise ValueError(f"[pred_noise_to_pred_video] Invalid timestep shape: {timestep.shape}")
+    # timestep shape should be [B]
     dtype = pred_noise.dtype
     device = pred_noise.device
-    pred_noise = pred_noise.float().to(device)
-    noise_input_latent = noise_input_latent.float().to(device)
-    sigmas = scheduler.sigmas.float().to(device)
-    timesteps = scheduler.timesteps.float().to(device)
+    pred_noise = pred_noise.double().to(device)
+    noise_input_latent = noise_input_latent.double().to(device)
+    sigmas = scheduler.sigmas.double().to(device)
+    timesteps = scheduler.timesteps.double().to(device)
     timestep_id = torch.argmin(
         (timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1)
     sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1)

fastvideo/pipelines/basic/wan/wan_causal_dmd_pipeline.py

@@ -7,8 +7,6 @@ This module wires the causal DMD denoising stage into the modular pipeline.
 
 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
 
 # isort: off
@@ -16,8 +14,7 @@ from fastvideo.pipelines.stages import (ConditioningStage, DecodingStage,
                                         CausalDMDDenosingStage,
                                         InputValidationStage,
                                         LatentPreparationStage,
-                                        TextEncodingStage,
-                                        TimestepPreparationStage)
+                                        TextEncodingStage)
 # isort: on
 
 logger = init_logger(__name__)
@@ -29,10 +26,6 @@ class WanCausalDMDPipeline(LoRAPipeline, ComposedPipelineBase):
         "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
     ]
 
-    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:
         """Set up pipeline stages with proper dependency injection."""
 
@@ -48,10 +41,6 @@ class WanCausalDMDPipeline(LoRAPipeline, ComposedPipelineBase):
         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"),

fastvideo/pipelines/composed_pipeline_base.py

@@ -40,7 +40,7 @@ class ComposedPipelineBase(ABC):
     _extra_config_module_map: dict[str, str] = {}
     training_args: TrainingArgs | None = None
     fastvideo_args: FastVideoArgs | TrainingArgs | None = None
-    modules: dict[str, torch.nn.Module] = {}
+    modules: dict[str, Any] = {}
     post_init_called: bool = False
 
     # TODO(will): args should support both inference args and training args
@@ -79,7 +79,7 @@ class ComposedPipelineBase(ABC):
             for name, module in self.modules.items():
                 if not isinstance(module, torch.nn.Module):
                     continue
-                if name == "transformer":
+                if "transformer" in name:
                     module.requires_grad_(True)
                 else:
                     module.requires_grad_(False)
@@ -136,11 +136,9 @@ class ComposedPipelineBase(ABC):
 
             kwargs['model_path'] = model_path
             fastvideo_args = FastVideoArgs.from_kwargs(**kwargs)
-            logger.info("fastvideo_args in from_pretrained: %s", fastvideo_args)
         else:
             assert args is not None, "args must be provided for training mode"
             fastvideo_args = TrainingArgs.from_cli_args(args)
-            logger.info("training args in from_pretrained: %s", fastvideo_args)
             # TODO(will): fix this so that its not so ugly
             fastvideo_args.model_path = model_path
             for key, value in kwargs.items():
@@ -239,20 +237,19 @@ class ComposedPipelineBase(ABC):
         # remove keys that are not pipeline modules
         model_index.pop("_class_name")
         model_index.pop("_diffusers_version")
-        # @TODO(Wei): Temporary hack
         if "boundary_ratio" in model_index and model_index[
                 "boundary_ratio"] is not None:
             logger.info(
                 "MoE pipeline detected. Adding transformer_2 to self.required_config_modules..."
             )
             self.required_config_modules.append("transformer_2")
-            if fastvideo_args.boundary_ratio is None:
-                logger.info(
-                    "MoE pipeline detected. Setting boundary ratio to %s",
-                    model_index["boundary_ratio"])
-                fastvideo_args.boundary_ratio = model_index["boundary_ratio"]
+            logger.info("MoE pipeline detected. Setting boundary ratio to %s",
+                        model_index["boundary_ratio"])
+            fastvideo_args.pipeline_config.dit_config.boundary_ratio = model_index[
+                "boundary_ratio"]
 
         model_index.pop("boundary_ratio", None)
+        # used by Wan2.2 ti2v
         model_index.pop("expand_timesteps", None)
 
         # some sanity checks
@@ -285,8 +282,8 @@ class ComposedPipelineBase(ABC):
                           architecture) in model_index.items():
             if transformers_or_diffusers is None:
                 logger.warning(
-                    "Module in model_index.json has null value, removing from required_config_modules"
-                )
+                    "Module %s in model_index.json has null value, removing from required_config_modules",
+                    module_name)
                 if module_name in self.required_config_modules:
                     self.required_config_modules.remove(module_name)
                 continue

fastvideo/pipelines/lora_pipeline.py

@@ -7,7 +7,7 @@ import torch
 import torch.distributed as dist
 import torch.nn as nn
 from safetensors.torch import load_file
-from torch.distributed.device_mesh import init_device_mesh
+from torch.distributed.device_mesh import DeviceMesh, init_device_mesh
 from torch.distributed.tensor import DTensor
 
 from fastvideo.distributed import get_local_torch_device
@@ -32,6 +32,7 @@ class LoRAPipeline(ComposedPipelineBase):
     cur_adapter_name: str = ""
     cur_adapter_path: str = ""
     lora_layers: dict[str, BaseLayerWithLoRA] = {}
+    lora_layers_critic: dict[str, BaseLayerWithLoRA] = {}
     fastvideo_args: FastVideoArgs | TrainingArgs
     exclude_lora_layers: list[str] = []
     device: torch.device = get_local_torch_device()
@@ -81,6 +82,17 @@ class LoRAPipeline(ComposedPipelineBase):
 
     def set_trainable(self) -> None:
 
+        def set_lora_grads(lora_layers: dict[str, BaseLayerWithLoRA],
+                           device_mesh: DeviceMesh):
+            for name, layer in lora_layers.items():
+                layer.lora_A.requires_grad_(True)
+                layer.lora_B.requires_grad_(True)
+                layer.base_layer.requires_grad_(False)
+                layer.lora_A = nn.Parameter(
+                    DTensor.from_local(layer.lora_A, device_mesh=device_mesh))
+                layer.lora_B = nn.Parameter(
+                    DTensor.from_local(layer.lora_B, device_mesh=device_mesh))
+
         is_lora_training = self.training_mode and getattr(
             self.fastvideo_args, "lora_training", False)
         if not is_lora_training:
@@ -88,18 +100,12 @@ class LoRAPipeline(ComposedPipelineBase):
             return
 
         self.modules["transformer"].requires_grad_(False)
+        if "fake_score_transformer" in self.modules:
+            self.modules["fake_score_transformer"].requires_grad_(False)
         device_mesh = init_device_mesh("cuda", (dist.get_world_size(), 1),
                                        mesh_dim_names=["fake", "replicate"])
-        for name, layer in self.lora_layers.items():
-            # Enable grads for lora weights only
-            # Must convert to DTensor for compatibility with other FSDP modules in grad calculation
-            layer.lora_A.requires_grad_(True)
-            layer.lora_B.requires_grad_(True)
-            layer.base_layer.requires_grad_(False)
-            layer.lora_A = nn.Parameter(
-                DTensor.from_local(layer.lora_A, device_mesh=device_mesh))
-            layer.lora_B = nn.Parameter(
-                DTensor.from_local(layer.lora_B, device_mesh=device_mesh))
+        set_lora_grads(self.lora_layers, device_mesh)
+        set_lora_grads(self.lora_layers_critic, device_mesh)
 
     def convert_to_lora_layers(self) -> None:
         """
@@ -131,6 +137,24 @@ class LoRAPipeline(ComposedPipelineBase):
                 converted_count += 1
         logger.info("Converted %d layers to LoRA layers", converted_count)
 
+        if "fake_score_transformer" in self.modules:
+            for name, layer in self.modules[
+                    "fake_score_transformer"].named_modules():
+                if not self.is_target_layer(name):
+                    continue
+                layer = get_lora_layer(layer,
+                                       lora_rank=self.lora_rank,
+                                       lora_alpha=self.lora_alpha,
+                                       training_mode=self.training_mode)
+                if layer is not None:
+                    self.lora_layers_critic[name] = layer
+                    replace_submodule(self.modules["fake_score_transformer"],
+                                      name, layer)
+                    converted_count += 1
+            logger.info(
+                "Converted %d layers to LoRA layers in the critic model",
+                converted_count)
+
     def set_lora_adapter(self,
                          lora_nickname: str,
                          lora_path: str | None = None):  # type: ignore
@@ -224,4 +248,4 @@ class LoRAPipeline(ComposedPipelineBase):
 
     def unmerge_lora_weights(self) -> None:
         for name, layer in self.lora_layers.items():
-            layer.unmerge_lora_weights()
+            layer.unmerge_lora_weights()

fastvideo/pipelines/pipeline_batch_info.py

@@ -129,6 +129,7 @@ class ForwardBatch:
     timesteps: torch.Tensor | None = None
     timestep: torch.Tensor | float | int | None = None
     step_index: int | None = None
+    boundary_ratio: float | None = None
 
     # Scheduler parameters
     num_inference_steps: int = 50
@@ -147,7 +148,12 @@ class ForwardBatch:
     modules: dict[str, Any] = field(default_factory=dict)
 
     # Final output (after pipeline completion)
-    output: Any = None
+    output: torch.Tensor | None = None
+    return_trajectory_latents: bool = False
+    return_trajectory_decoded: bool = False
+    trajectory_timesteps: list[torch.Tensor] | None = None
+    trajectory_latents: torch.Tensor | None = None
+    trajectory_decoded: list[torch.Tensor] | None = None
 
     # Extra parameters that might be needed by specific pipeline implementations
     extra: dict[str, Any] = field(default_factory=dict)
@@ -206,6 +212,10 @@ class TrainingBatch:
     infos: list[dict[str, Any]] | None = None
     mask_lat_size: torch.Tensor | None = None
 
+    # ODE trajectory supervision
+    trajectory_latents: torch.Tensor | None = None
+    trajectory_timesteps: torch.Tensor | None = None
+
     # Transformer inputs
     noisy_model_input: torch.Tensor | None = None
     timesteps: torch.Tensor | None = None
@@ -236,10 +246,11 @@ class TrainingBatch:
     fake_score_loss: float = 0.0
 
     dmd_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
+    latent_vis_dict: dict[str, Any] = field(default_factory=dict)
     fake_score_latent_vis_dict: dict[str, Any] = field(default_factory=dict)
 
 
 @dataclass
 class PreprocessBatch(ForwardBatch):
-    video_loader: list["VideoDecoder"] = field(default_factory=list)
-    video_file_name: list[str] = field(default_factory=list)
+    video_loader: list["VideoDecoder"] | list[str] = field(default_factory=list)
+    video_file_name: list[str] = field(default_factory=list)

fastvideo/pipelines/preprocess/preprocess_pipeline_base.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-import multiprocessing
 import os
-from concurrent.futures import ProcessPoolExecutor
 from typing import Any
 
 import numpy as np
@@ -12,6 +10,8 @@ from torch.utils.data import DataLoader
 from tqdm import tqdm
 
 from fastvideo.dataset import getdataset
+from fastvideo.dataset.dataloader.parquet_io import (ParquetDatasetWriter,
+                                                     records_to_table)
 from fastvideo.dataset.preprocessing_datasets import PreprocessBatch
 from fastvideo.distributed import get_local_torch_device
 from fastvideo.fastvideo_args import FastVideoArgs
@@ -54,10 +54,14 @@ class BasePreprocessPipeline(ComposedPipelineBase):
         """Get additional features specific to the pipeline type. Override in subclasses."""
         return {}
 
-    def get_schema_fields(self) -> list[str]:
-        """Get the schema fields for the pipeline type. Override in subclasses."""
+    def get_pyarrow_schema(self) -> pa.Schema:
+        """Return the PyArrow schema for this pipeline. Must be overridden."""
         raise NotImplementedError
 
+    def get_schema_fields(self) -> list[str]:
+        """Get the schema fields for the pipeline type."""
+        return [f.name for f in self.get_pyarrow_schema()]
+
     def create_record_for_schema(self,
                                  preprocess_batch: PreprocessBatch,
                                  schema: pa.Schema,
@@ -400,166 +404,22 @@ class BasePreprocessPipeline(ComposedPipelineBase):
                 batch_data.append(record)
 
             if batch_data:
-                # Add progress bar for writing to Parquet dataset
                 write_pbar = tqdm(total=1,
                                   desc="Writing to Parquet dataset",
                                   unit="batch")
-                # Convert batch data to PyArrow arrays
-                arrays = []
-                for field in self.get_schema_fields():
-                    if field.endswith('_bytes'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.binary()))
-                    elif field.endswith('_shape'):
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.list_(pa.int32())))
-                    elif field in ['width', 'height', 'num_frames']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.int32()))
-                    elif field in ['duration_sec', 'fps']:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data],
-                                     type=pa.float32()))
-                    else:
-                        arrays.append(
-                            pa.array([record[field] for record in batch_data]))
-
-                table = pa.Table.from_arrays(arrays,
-                                             names=self.get_schema_fields())
+                table = records_to_table(batch_data, self.get_pyarrow_schema())
                 write_pbar.update(1)
                 write_pbar.close()
 
-                # Store the table in a list for later processing
-                if not hasattr(self, 'all_tables'):
-                    self.all_tables = []
-                self.all_tables.append(table)
-
+                if not hasattr(self, 'dataset_writer'):
+                    self.dataset_writer = ParquetDatasetWriter(
+                        out_dir=combined_parquet_dir,
+                        samples_per_file=args.samples_per_file,
+                    )
+                self.dataset_writer.append_table(table)
                 logger.info("Collected batch with %s samples", len(table))
 
             if num_processed_samples >= args.flush_frequency:
-                self._flush_tables(num_processed_samples, args,
-                                   combined_parquet_dir)
+                written = self.dataset_writer.flush()
+                logger.info("Flushed %s samples to parquet", written)
                 num_processed_samples = 0
-                self.all_tables = []
-
-    def _flush_tables(self, num_processed_samples: int, args,
-                      combined_parquet_dir: str):
-        """Flush collected tables to disk."""
-        assert hasattr(self, 'all_tables') and self.all_tables
-        print(f"Combining {len(self.all_tables)} batches...")
-        combined_table = pa.concat_tables(self.all_tables)
-        assert len(combined_table) == num_processed_samples
-        print(f"Total samples collected: {len(combined_table)}")
-
-        # Calculate total number of chunks needed, discarding remainder
-        total_chunks = max(num_processed_samples // args.samples_per_file, 1)
-
-        print(f"Fixed samples per parquet file: {args.samples_per_file}")
-        print(f"Total number of parquet files: {total_chunks}")
-        print(
-            f"Total samples to be processed: {total_chunks * args.samples_per_file} (discarding {num_processed_samples % args.samples_per_file} samples)"
-        )
-
-        # Split work among processes
-        num_workers = int(min(multiprocessing.cpu_count(), total_chunks))
-        chunks_per_worker = (total_chunks + num_workers - 1) // num_workers
-
-        print(f"Using {num_workers} workers to process {total_chunks} chunks")
-        logger.info("Chunks per worker: %s", chunks_per_worker)
-
-        # Prepare work ranges
-        work_ranges = []
-        for i in range(num_workers):
-            start_idx = i * chunks_per_worker
-            end_idx = min((i + 1) * chunks_per_worker, total_chunks)
-            if start_idx < total_chunks:
-                work_ranges.append(
-                    (start_idx, end_idx, combined_table, i,
-                     combined_parquet_dir, args.samples_per_file))
-
-        total_written = 0
-        failed_ranges = []
-        with ProcessPoolExecutor(max_workers=num_workers) as executor:
-            futures = {
-                executor.submit(self.process_chunk_range, work_range):
-                work_range
-                for work_range in work_ranges
-            }
-            for future in tqdm(futures, desc="Processing chunks"):
-                try:
-                    written = future.result()
-                    total_written += written
-                    logger.info("Processed chunk with %s samples", written)
-                except Exception as e:
-                    work_range = futures[future]
-                    failed_ranges.append(work_range)
-                    logger.error("Failed to process range %s-%s: %s",
-                                 work_range[0], work_range[1], str(e))
-
-        # Retry failed ranges sequentially
-        if failed_ranges:
-            logger.warning("Retrying %s failed ranges sequentially",
-                           len(failed_ranges))
-            for work_range in failed_ranges:
-                try:
-                    total_written += self.process_chunk_range(work_range)
-                except Exception as e:
-                    logger.error(
-                        "Failed to process range %s-%s after retry: %s",
-                        work_range[0], work_range[1], str(e))
-
-        logger.info("Total samples written: %s", total_written)
-
-    @staticmethod
-    def process_chunk_range(args: Any) -> int:
-        start_idx, end_idx, table, worker_id, output_dir, samples_per_file = args
-        try:
-            total_written = 0
-            num_samples = len(table)
-
-            # Create worker-specific subdirectory
-            worker_dir = os.path.join(output_dir, f"worker_{worker_id}")
-            os.makedirs(worker_dir, exist_ok=True)
-
-            # Check how many files there are already in the dir, and update i accordingly
-            num_parquets = 0
-            for root, _, files in os.walk(worker_dir):
-                for file in files:
-                    if file.endswith('.parquet'):
-                        num_parquets += 1
-
-            for i in range(start_idx, end_idx):
-                start_sample = i * samples_per_file
-                end_sample = min((i + 1) * samples_per_file, num_samples)
-                chunk = table.slice(start_sample, end_sample - start_sample)
-
-                # Create chunk file in worker's directory
-                chunk_path = os.path.join(
-                    worker_dir, f"data_chunk_{i + num_parquets}.parquet")
-                temp_path = chunk_path + '.tmp'
-
-                try:
-                    # Write to temporary file
-                    pq.write_table(chunk, temp_path, compression='zstd')
-
-                    # Rename temporary file to final file
-                    if os.path.exists(chunk_path):
-                        os.remove(
-                            chunk_path)  # Remove existing file if it exists
-                    os.rename(temp_path, chunk_path)
-
-                    total_written += len(chunk)
-                except Exception as e:
-                    # Clean up temporary file if it exists
-                    if os.path.exists(temp_path):
-                        os.remove(temp_path)
-                    raise e
-
-            return total_written
-        except Exception as e:
-            logger.error("Error processing chunks %s-%s for worker %s: %s",
-                         start_idx, end_idx, worker_id, str(e))
-            raise

fastvideo/pipelines/preprocess/preprocess_pipeline_i2v.py

@@ -40,9 +40,9 @@ class PreprocessPipeline_I2V(BasePreprocessPipeline):
                            image_processor=self.get_module("image_processor"),
                        ))
 
-    def get_schema_fields(self) -> list[str]:
-        """Get the schema fields for I2V pipeline."""
-        return [f.name for f in pyarrow_schema_i2v]
+    def get_pyarrow_schema(self):
+        """Return the PyArrow schema for I2V pipeline."""
+        return pyarrow_schema_i2v
 
     def get_extra_features(self, valid_data: dict[str, Any],
                            fastvideo_args: FastVideoArgs) -> dict[str, Any]:

fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py

@@ -0,0 +1,323 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+ODE Trajectory Data Preprocessing pipeline implementation.
+
+This module contains an implementation of the ODE Trajectory Data Preprocessing pipeline
+using the modular pipeline architecture.
+
+Sec 4.3 of CausVid paper: https://arxiv.org/pdf/2412.07772
+"""
+
+import os
+from collections.abc import Iterator
+from typing import Any
+
+import pyarrow as pa
+import torch
+from torch.utils.data import DataLoader
+from torchdata.stateful_dataloader import StatefulDataLoader
+from tqdm import tqdm
+
+from fastvideo.configs.sample import SamplingParam
+from fastvideo.dataset import gettextdataset
+from fastvideo.dataset.dataloader.parquet_io import (ParquetDatasetWriter,
+                                                     records_to_table)
+from fastvideo.dataset.dataloader.record_schema import (
+    ode_text_only_record_creator)
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
+    SelfForcingFlowMatchScheduler)
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.preprocess.preprocess_pipeline_base import (
+    BasePreprocessPipeline)
+from fastvideo.pipelines.stages import (DecodingStage, DenoisingStage,
+                                        InputValidationStage,
+                                        LatentPreparationStage,
+                                        TextEncodingStage,
+                                        TimestepPreparationStage)
+from fastvideo.utils import save_decoded_latents_as_video, shallow_asdict
+
+logger = init_logger(__name__)
+
+
+class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline):
+    """ODE Trajectory preprocessing pipeline implementation."""
+
+    _required_config_modules = [
+        "text_encoder", "tokenizer", "vae", "transformer", "scheduler"
+    ]
+    preprocess_dataloader: StatefulDataLoader
+    preprocess_loader_iter: Iterator[dict[str, Any]]
+    pbar: Any
+    num_processed_samples: int
+
+    def get_pyarrow_schema(self) -> pa.Schema:
+        """Return the PyArrow schema for ODE Trajectory pipeline."""
+        return pyarrow_schema_ode_trajectory_text_only
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+        """Set up pipeline stages with proper dependency injection."""
+        assert fastvideo_args.pipeline_config.flow_shift == 5
+        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
+            shift=fastvideo_args.pipeline_config.flow_shift,
+            sigma_min=0.0,
+            extra_one_step=True)
+        self.modules["scheduler"].set_timesteps(num_inference_steps=48,
+                                                denoising_strength=1.0)
+
+        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="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="denoising_stage",
+                       stage=DenoisingStage(
+                           transformer=self.get_module("transformer"),
+                           scheduler=self.get_module("scheduler"),
+                           pipeline=self,
+                       ))
+        self.add_stage(stage_name="decoding_stage",
+                       stage=DecodingStage(vae=self.get_module("vae")))
+
+    def preprocess_text_and_trajectory(self, fastvideo_args: FastVideoArgs,
+                                       args):
+        """Preprocess text-only data and generate trajectory information."""
+
+        for batch_idx, data in enumerate(self.pbar):
+            if data is None:
+                continue
+
+            with torch.inference_mode():
+                # For text-only processing, we only need text data
+                # Filter out samples without text
+                valid_indices = []
+                for i, text in enumerate(data["text"]):
+                    if text and text.strip():  # Check if text is not empty
+                        valid_indices.append(i)
+                self.num_processed_samples += len(valid_indices)
+
+                if not valid_indices:
+                    continue
+
+                # Create new batch with only valid samples (text-only)
+                valid_data = {
+                    "text": [data["text"][i] for i in valid_indices],
+                    "path": [data["path"][i] for i in valid_indices],
+                }
+
+                # Add fps and duration if available in data
+                if "fps" in data:
+                    valid_data["fps"] = [data["fps"][i] for i in valid_indices]
+                if "duration" in data:
+                    valid_data["duration"] = [
+                        data["duration"][i] for i in valid_indices
+                    ]
+
+                batch_captions = valid_data["text"]
+                # Encode text using the standalone TextEncodingStage API
+                prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
+                    batch_captions,
+                    fastvideo_args,
+                    encoder_index=[0],
+                    return_attention_mask=True,
+                )
+                prompt_embeds = prompt_embeds_list[0]
+                prompt_attention_masks = prompt_masks_list[0]
+                assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]
+
+                sampling_params = SamplingParam.from_pretrained(args.model_path)
+
+                # encode negative prompt for trajectory collection
+                if sampling_params.guidance_scale > 1 and sampling_params.negative_prompt is not None:
+                    negative_prompt_embeds_list, negative_prompt_masks_list = self.prompt_encoding_stage.encode_text(
+                        sampling_params.negative_prompt,
+                        fastvideo_args,
+                        encoder_index=[0],
+                        return_attention_mask=True,
+                    )
+                    negative_prompt_embed = negative_prompt_embeds_list[0][0]
+                    negative_prompt_attention_mask = negative_prompt_masks_list[
+                        0][0]
+                else:
+                    negative_prompt_embed = None
+                    negative_prompt_attention_mask = None
+
+                trajectory_latents = []
+                trajectory_timesteps = []
+                trajectory_decoded = []
+
+                for i, (prompt_embed, prompt_attention_mask) in enumerate(
+                        zip(prompt_embeds, prompt_attention_masks,
+                            strict=False)):
+                    prompt_embed = prompt_embed.unsqueeze(0)
+                    prompt_attention_mask = prompt_attention_mask.unsqueeze(0)
+
+                    # Collect the trajectory data (text-to-video generation)
+                    batch = ForwardBatch(**shallow_asdict(sampling_params), )
+                    batch.prompt_embeds = [prompt_embed]
+                    batch.prompt_attention_mask = [prompt_attention_mask]
+                    batch.negative_prompt_embeds = [negative_prompt_embed]
+                    batch.negative_attention_mask = [
+                        negative_prompt_attention_mask
+                    ]
+                    batch.num_inference_steps = 48
+                    batch.return_trajectory_latents = True
+                    # Enabling this will save the decoded trajectory videos.
+                    # Used for debugging.
+                    batch.return_trajectory_decoded = False
+                    batch.height = args.max_height
+                    batch.width = args.max_width
+                    batch.fps = args.train_fps
+                    batch.guidance_scale = 6.0
+                    batch.do_classifier_free_guidance = True
+
+                    result_batch = self.input_validation_stage(
+                        batch, fastvideo_args)
+                    result_batch = self.timestep_preparation_stage(
+                        batch, fastvideo_args)
+                    result_batch = self.latent_preparation_stage(
+                        result_batch, fastvideo_args)
+                    result_batch = self.denoising_stage(result_batch,
+                                                        fastvideo_args)
+                    result_batch = self.decoding_stage(result_batch,
+                                                       fastvideo_args)
+
+                    trajectory_latents.append(
+                        result_batch.trajectory_latents.cpu())
+                    trajectory_timesteps.append(
+                        result_batch.trajectory_timesteps.cpu())
+                    trajectory_decoded.append(result_batch.trajectory_decoded)
+
+                # Prepare extra features for text-only processing
+                extra_features = {
+                    "trajectory_latents": trajectory_latents,
+                    "trajectory_timesteps": trajectory_timesteps
+                }
+
+                if batch.return_trajectory_decoded:
+                    for i, decoded_frames in enumerate(trajectory_decoded):
+                        for j, decoded_frame in enumerate(decoded_frames):
+                            save_decoded_latents_as_video(
+                                decoded_frame,
+                                f"decoded_videos/trajectory_decoded_{i}_{j}.mp4",
+                                args.train_fps)
+
+                # Prepare batch data for Parquet dataset
+                batch_data: list[dict[str, Any]] = []
+
+                # Add progress bar for saving outputs
+                save_pbar = tqdm(enumerate(valid_data["path"]),
+                                 desc="Saving outputs",
+                                 unit="item",
+                                 leave=False)
+
+                for idx, video_path in save_pbar:
+                    video_name = os.path.basename(video_path).split(".")[0]
+
+                    # Convert tensors to numpy arrays
+                    text_embedding = prompt_embeds[idx].cpu().numpy()
+
+                    # Get extra features for this sample
+                    sample_extra_features = {}
+                    if extra_features:
+                        for key, value in extra_features.items():
+                            if isinstance(value, torch.Tensor):
+                                sample_extra_features[key] = value[idx].cpu(
+                                ).numpy()
+                            else:
+                                assert isinstance(value, list)
+                                if isinstance(value[idx], torch.Tensor):
+                                    sample_extra_features[key] = value[idx].cpu(
+                                    ).float().numpy()
+                                else:
+                                    sample_extra_features[key] = value[idx]
+
+                    # Create record for Parquet dataset (text-only ODE schema)
+                    record: dict[str, Any] = ode_text_only_record_creator(
+                        video_name=video_name,
+                        text_embedding=text_embedding,
+                        caption=valid_data["text"][idx],
+                        trajectory_latents=sample_extra_features[
+                            "trajectory_latents"],
+                        trajectory_timesteps=sample_extra_features[
+                            "trajectory_timesteps"],
+                    )
+                    batch_data.append(record)
+
+                if batch_data:
+                    write_pbar = tqdm(total=1,
+                                      desc="Writing to Parquet dataset",
+                                      unit="batch")
+                    table = records_to_table(batch_data,
+                                             self.get_pyarrow_schema())
+                    write_pbar.update(1)
+                    write_pbar.close()
+
+                    if not hasattr(self, 'dataset_writer'):
+                        self.dataset_writer = ParquetDatasetWriter(
+                            out_dir=self.combined_parquet_dir,
+                            samples_per_file=args.samples_per_file,
+                        )
+                    self.dataset_writer.append_table(table)
+
+                    logger.info("Collected batch with %s samples", len(table))
+
+                if self.num_processed_samples >= args.flush_frequency:
+                    written = self.dataset_writer.flush()
+                    logger.info("Flushed %s samples to parquet", written)
+                    self.num_processed_samples = 0
+
+        # Final flush for any remaining samples
+        if hasattr(self, 'dataset_writer'):
+            written = self.dataset_writer.flush(write_remainder=True)
+            if written:
+                logger.info("Final flush wrote %s samples", written)
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs, args):
+        if not self.post_init_called:
+            self.post_init()
+
+        self.local_rank = int(os.getenv("RANK", 0))
+        os.makedirs(args.output_dir, exist_ok=True)
+        # Create directory for combined data
+        self.combined_parquet_dir = os.path.join(args.output_dir,
+                                                 "combined_parquet_dataset")
+        os.makedirs(self.combined_parquet_dir, exist_ok=True)
+
+        # Loading dataset
+        train_dataset = gettextdataset(args)
+
+        self.preprocess_dataloader = DataLoader(
+            train_dataset,
+            batch_size=args.preprocess_video_batch_size,
+            num_workers=args.dataloader_num_workers,
+        )
+
+        self.preprocess_loader_iter = iter(self.preprocess_dataloader)
+
+        self.num_processed_samples = 0
+        # Add progress bar for video preprocessing
+        self.pbar = tqdm(self.preprocess_loader_iter,
+                         desc="Processing videos",
+                         unit="batch",
+                         disable=self.local_rank != 0)
+
+        # Initialize class variables for data sharing
+        self.video_data: dict[str, Any] = {}  # Store video metadata and paths
+        self.latent_data: dict[str, Any] = {}  # Store latent tensors
+        self.preprocess_text_and_trajectory(fastvideo_args, args)
+
+
+EntryClass = PreprocessPipeline_ODE_Trajectory

fastvideo/pipelines/preprocess/preprocess_pipeline_t2v.py

@@ -15,9 +15,9 @@ class PreprocessPipeline_T2V(BasePreprocessPipeline):
 
     _required_config_modules = ["text_encoder", "tokenizer", "vae"]
 
-    def get_schema_fields(self):
-        """Get the schema fields for T2V pipeline."""
-        return [f.name for f in pyarrow_schema_t2v]
+    def get_pyarrow_schema(self):
+        """Return the PyArrow schema for T2V pipeline."""
+        return pyarrow_schema_t2v
 
 
 EntryClass = PreprocessPipeline_T2V

fastvideo/pipelines/preprocess/preprocess_pipeline_text.py

@@ -0,0 +1,184 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Text-only Data Preprocessing pipeline implementation.
+
+This module contains an implementation of the Text-only Data Preprocessing pipeline
+using the modular pipeline architecture, based on the ODE Trajectory preprocessing.
+"""
+
+import os
+from collections.abc import Iterator
+from typing import Any
+
+import torch
+from torch.utils.data import DataLoader
+from torchdata.stateful_dataloader import StatefulDataLoader
+from tqdm import tqdm
+
+from fastvideo.dataset import gettextdataset
+from fastvideo.dataset.dataloader.parquet_io import (ParquetDatasetWriter,
+                                                     records_to_table)
+from fastvideo.dataset.dataloader.record_schema import text_only_record_creator
+from fastvideo.dataset.dataloader.schema import pyarrow_schema_text_only
+from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.logger import init_logger
+from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
+from fastvideo.pipelines.preprocess.preprocess_pipeline_base import (
+    BasePreprocessPipeline)
+from fastvideo.pipelines.stages import TextEncodingStage
+
+logger = init_logger(__name__)
+
+
+class PreprocessPipeline_Text(BasePreprocessPipeline):
+    """Text-only preprocessing pipeline implementation."""
+
+    _required_config_modules = ["text_encoder", "tokenizer"]
+    preprocess_dataloader: StatefulDataLoader
+    preprocess_loader_iter: Iterator[dict[str, Any]]
+    pbar: Any
+    num_processed_samples: int = 0
+
+    def get_pyarrow_schema(self):
+        """Return the PyArrow schema for text-only pipeline."""
+        return pyarrow_schema_text_only
+
+    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
+        """Set up pipeline stages with proper dependency injection."""
+        self.add_stage(stage_name="prompt_encoding_stage",
+                       stage=TextEncodingStage(
+                           text_encoders=[self.get_module("text_encoder")],
+                           tokenizers=[self.get_module("tokenizer")],
+                       ))
+
+    def preprocess_text_only(self, fastvideo_args: FastVideoArgs, args):
+        """Preprocess text-only data."""
+
+        for batch_idx, data in enumerate(self.pbar):
+            if data is None:
+                continue
+
+            with torch.inference_mode():
+                # For text-only processing, we only need text data
+                # Filter out samples without text
+                valid_indices = []
+                for i, text in enumerate(data["text"]):
+                    if text and text.strip():  # Check if text is not empty
+                        valid_indices.append(i)
+                self.num_processed_samples += len(valid_indices)
+
+                if not valid_indices:
+                    continue
+
+                # Create new batch with only valid samples (text-only)
+                valid_data = {
+                    "text": [data["text"][i] for i in valid_indices],
+                    "path": [data["path"][i] for i in valid_indices],
+                }
+
+                batch_captions = valid_data["text"]
+                # Encode text using the standalone TextEncodingStage API
+                prompt_embeds_list, prompt_masks_list = self.prompt_encoding_stage.encode_text(
+                    batch_captions,
+                    fastvideo_args,
+                    encoder_index=[0],
+                    return_attention_mask=True,
+                )
+                prompt_embeds = prompt_embeds_list[0]
+                prompt_attention_masks = prompt_masks_list[0]
+                assert prompt_embeds.shape[0] == prompt_attention_masks.shape[0]
+
+                logger.info("===== prompt_embeds: %s", prompt_embeds.shape)
+                logger.info("===== prompt_attention_masks: %s",
+                            prompt_attention_masks.shape)
+
+                # Prepare batch data for Parquet dataset
+                batch_data = []
+
+                # Add progress bar for saving outputs
+                save_pbar = tqdm(enumerate(valid_data["path"]),
+                                 desc="Saving outputs",
+                                 unit="item",
+                                 leave=False)
+
+                for idx, text_path in save_pbar:
+                    text_name = os.path.basename(text_path).split(".")[0]
+
+                    # Convert tensors to numpy arrays
+                    text_embedding = prompt_embeds[idx].cpu().numpy()
+
+                    # Create record for Parquet dataset (text-only schema)
+                    record = text_only_record_creator(
+                        text_name=text_name,
+                        text_embedding=text_embedding,
+                        caption=valid_data["text"][idx],
+                    )
+                    batch_data.append(record)
+
+                if batch_data:
+                    write_pbar = tqdm(total=1,
+                                      desc="Writing to Parquet dataset",
+                                      unit="batch")
+                    table = records_to_table(batch_data,
+                                             pyarrow_schema_text_only)
+                    write_pbar.update(1)
+                    write_pbar.close()
+
+                    if not hasattr(self, 'dataset_writer'):
+                        self.dataset_writer = ParquetDatasetWriter(
+                            out_dir=self.combined_parquet_dir,
+                            samples_per_file=args.samples_per_file,
+                        )
+                    self.dataset_writer.append_table(table)
+
+                    logger.info("Collected batch with %s samples", len(table))
+
+                if self.num_processed_samples >= args.flush_frequency:
+                    written = self.dataset_writer.flush()
+                    logger.info("Flushed %s samples to parquet", written)
+                    self.num_processed_samples = 0
+
+        # Final flush for any remaining samples
+        if hasattr(self, 'dataset_writer'):
+            written = self.dataset_writer.flush(write_remainder=True)
+            if written:
+                logger.info("Final flush wrote %s samples", written)
+
+    # Text-only record creation moved to fastvideo.dataset.dataloader.record_schema
+
+    def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs, args):
+        if not self.post_init_called:
+            self.post_init()
+
+        self.local_rank = int(os.getenv("RANK", 0))
+        os.makedirs(args.output_dir, exist_ok=True)
+        # Create directory for combined data
+        self.combined_parquet_dir = os.path.join(args.output_dir,
+                                                 "combined_parquet_dataset")
+        os.makedirs(self.combined_parquet_dir, exist_ok=True)
+
+        # Loading text dataset
+        train_dataset = gettextdataset(args)
+
+        self.preprocess_dataloader = DataLoader(
+            train_dataset,
+            batch_size=args.preprocess_video_batch_size,
+            num_workers=args.dataloader_num_workers,
+        )
+
+        self.preprocess_loader_iter = iter(self.preprocess_dataloader)
+
+        self.num_processed_samples = 0
+        # Add progress bar for text preprocessing
+        self.pbar = tqdm(self.preprocess_loader_iter,
+                         desc="Processing text",
+                         unit="batch",
+                         disable=self.local_rank != 0)
+
+        # Initialize class variables for data sharing
+        self.text_data: dict[str, Any] = {}  # Store text metadata and paths
+
+        self.preprocess_text_only(fastvideo_args, args)
+
+
+EntryClass = PreprocessPipeline_Text

fastvideo/pipelines/preprocess/preprocess_stages.py

@@ -4,9 +4,11 @@ from typing import cast
 
 import numpy as np
 import torch
+import torchvision
 from einops import rearrange
 from torchvision import transforms
 
+from fastvideo.configs.configs import VideoLoaderType
 from fastvideo.dataset.transform import (CenterCropResizeVideo,
                                          TemporalRandomCrop)
 from fastvideo.fastvideo_args import FastVideoArgs, WorkloadType
@@ -61,7 +63,16 @@ class VideoTransformStage(PipelineStage):
                 else:
                     frame_indices = frame_indices[:self.num_frames]
 
-            video = batch.video_loader[i].get_frames_at(frame_indices).data
+            if fastvideo_args.preprocess_config.video_loader_type == VideoLoaderType.TORCHCODEC:
+                video = batch.video_loader[i].get_frames_at(frame_indices).data
+            elif fastvideo_args.preprocess_config.video_loader_type == VideoLoaderType.TORCHVISION:
+                video, _, _ = torchvision.io.read_video(batch.video_loader[i],
+                                                        output_format="TCHW")
+                video = video[frame_indices]
+            else:
+                raise ValueError(
+                    f"Invalid video loader type: {fastvideo_args.preprocess_config.video_loader_type}"
+                )
             video = self.video_transform(video)
             video_pixel_batch.append(video)
 

fastvideo/pipelines/preprocess/v1_preprocess.py

@@ -1,5 +1,6 @@
 import argparse
 import os
+from typing import Any
 
 from fastvideo import PipelineConfig
 from fastvideo.configs.models.vaes import WanVAEConfig
@@ -9,8 +10,12 @@ from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
 from fastvideo.pipelines.preprocess.preprocess_pipeline_i2v import (
     PreprocessPipeline_I2V)
+from fastvideo.pipelines.preprocess.preprocess_pipeline_ode_trajectory import (
+    PreprocessPipeline_ODE_Trajectory)
 from fastvideo.pipelines.preprocess.preprocess_pipeline_t2v import (
     PreprocessPipeline_T2V)
+from fastvideo.pipelines.preprocess.preprocess_pipeline_text import (
+    PreprocessPipeline_Text)
 from fastvideo.utils import maybe_download_model
 
 logger = init_logger(__name__)
@@ -21,12 +26,22 @@ def main(args) -> None:
     maybe_init_distributed_environment_and_model_parallel(1, 1)
     num_gpus = int(os.environ["WORLD_SIZE"])
     assert num_gpus == 1, "Only support 1 GPU"
+
     pipeline_config = PipelineConfig.from_pretrained(args.model_path)
-    kwargs = {
-        "vae_precision": "fp32",
-        "vae_config": WanVAEConfig(load_encoder=True, load_decoder=False),
-    }
+
+    kwargs: dict[str, Any] = {}
+    if args.preprocess_task == "text_only":
+        kwargs = {
+            "text_encoder_cpu_offload": False,
+        }
+    else:
+        # Full config for video/image processing
+        kwargs = {
+            "vae_precision": "fp32",
+            "vae_config": WanVAEConfig(load_encoder=True, load_decoder=True),
+        }
     pipeline_config.update_config_from_dict(kwargs)
+
     fastvideo_args = FastVideoArgs(
         model_path=args.model_path,
         num_gpus=get_world_size(),
@@ -35,7 +50,23 @@ def main(args) -> None:
         text_encoder_cpu_offload=False,
         pipeline_config=pipeline_config,
     )
-    PreprocessPipeline = PreprocessPipeline_I2V if args.preprocess_task == "i2v" else PreprocessPipeline_T2V
+    if args.preprocess_task == "t2v":
+        PreprocessPipeline = PreprocessPipeline_T2V
+    elif args.preprocess_task == "i2v":
+        PreprocessPipeline = PreprocessPipeline_I2V
+    elif args.preprocess_task == "text_only":
+        PreprocessPipeline = PreprocessPipeline_Text
+    elif args.preprocess_task == "ode_trajectory":
+        assert args.flow_shift is not None, "flow_shift is required for ode_trajectory"
+        fastvideo_args.pipeline_config.flow_shift = args.flow_shift
+        PreprocessPipeline = PreprocessPipeline_ODE_Trajectory
+    else:
+        raise ValueError(f"Invalid preprocess task: {args.preprocess_task}. "
+                         f"Valid options: t2v, i2v, ode_trajectory, text_only")
+
+    logger.info("Preprocess task: %s using %s", args.preprocess_task,
+                PreprocessPipeline.__name__)
+
     pipeline = PreprocessPipeline(args.model_path, fastvideo_args)
     pipeline.forward(batch=None, fastvideo_args=fastvideo_args, args=args)
 
@@ -74,7 +105,12 @@ if __name__ == "__main__":
     parser.add_argument("--video_length_tolerance_range", type=int, default=2.0)
     parser.add_argument("--group_frame", action="store_true")  # TODO
     parser.add_argument("--group_resolution", action="store_true")  # TODO
-    parser.add_argument("--preprocess_task", type=str, default="t2v")
+    parser.add_argument("--flow_shift", type=float, default=None)
+    parser.add_argument("--preprocess_task",
+                        type=str,
+                        default="t2v",
+                        choices=["t2v", "i2v", "text_only", "ode_trajectory"],
+                        help="Type of preprocessing task to run")
     parser.add_argument("--train_fps", type=int, default=30)
     parser.add_argument("--use_image_num", type=int, default=0)
     parser.add_argument("--text_max_length", type=int, default=256)

fastvideo/pipelines/stages/causal_denoising.py

@@ -4,11 +4,11 @@ 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.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
 from fastvideo.models.utils import pred_noise_to_pred_video
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.stages.denoising import DenoisingStage
+from fastvideo.pipelines.stages.validators import StageValidators as V
+from fastvideo.pipelines.stages.validators import VerificationResult
 
 try:
     from fastvideo.attention.backends.sliding_tile_attn import (
@@ -36,7 +36,6 @@ class CausalDMDDenosingStage(DenoisingStage):
 
     def __init__(self, transformer, scheduler) -> None:
         super().__init__(transformer, scheduler)
-        self.scheduler = FlowMatchEulerDiscreteScheduler(shift=8.0)
         # KV and cross-attention cache state (initialized on first forward)
         self.kv_cache1: list | None = None
         self.crossattn_cache: list | None = None
@@ -71,8 +70,16 @@ class CausalDMDDenosingStage(DenoisingStage):
         # Timesteps for DMD
         timesteps = torch.tensor(
             fastvideo_args.pipeline_config.dmd_denoising_steps,
-            dtype=torch.long,
-            device=get_local_torch_device())
+            dtype=torch.long).cpu()
+
+        if fastvideo_args.pipeline_config.warp_denoising_step:
+            logger.info("Warping timesteps...")
+            scheduler_timesteps = torch.cat((self.scheduler.timesteps.cpu(),
+                                             torch.tensor([0],
+                                                          dtype=torch.float32)))
+            timesteps = scheduler_timesteps[1000 - timesteps]
+        timesteps = timesteps.to(get_local_torch_device())
+        logger.info("Using timesteps: %s", timesteps)
 
         # Image kwargs (kept empty unless caller provides compatible args)
         image_kwargs: dict = {}
@@ -262,10 +269,14 @@ class CausalDMDDenosingStage(DenoisingStage):
                                             attn_metadata=attn_metadata,
                                             forward_batch=batch):
                         # Run transformer; follow DMD stage pattern
+                        t_expanded_noise = t_cur * torch.ones(
+                            (latent_model_input.shape[0], 1),
+                            device=latent_model_input.device,
+                            dtype=torch.long)
                         pred_noise_btchw = self.transformer(
                             latent_model_input,
                             prompt_embeds,
-                            t_expand,
+                            t_expanded_noise,
                             kv_cache=self.kv_cache1,
                             crossattn_cache=self.crossattn_cache,
                             current_start=(pos_start_base + start_index) *
@@ -326,10 +337,11 @@ class CausalDMDDenosingStage(DenoisingStage):
                     set_forward_context(current_timestep=0,
                                         attn_metadata=attn_metadata,
                                         forward_batch=batch):
+                    t_expanded_context = t_context.unsqueeze(1)
                     _ = self.transformer(
                         context_bcthw,
                         prompt_embeds,
-                        t_context,
+                        t_expanded_context,
                         kv_cache=self.kv_cache1,
                         crossattn_cache=self.crossattn_cache,
                         current_start=(pos_start_base + start_index) *
@@ -407,3 +419,27 @@ class CausalDMDDenosingStage(DenoisingStage):
                 False,
             })
         self.crossattn_cache = crossattn_cache
+
+    def verify_input(self, batch: ForwardBatch,
+                     fastvideo_args: FastVideoArgs) -> VerificationResult:
+        """Verify denoising stage inputs."""
+        result = VerificationResult()
+        result.add_check("latents", batch.latents,
+                         [V.is_tensor, V.with_dims(5)])
+        result.add_check("prompt_embeds", batch.prompt_embeds, V.list_not_empty)
+        result.add_check("image_embeds", batch.image_embeds, V.is_list)
+        result.add_check("image_latent", batch.image_latent,
+                         V.none_or_tensor_with_dims(5))
+        result.add_check("num_inference_steps", batch.num_inference_steps,
+                         V.positive_int)
+        result.add_check("guidance_scale", batch.guidance_scale,
+                         V.positive_float)
+        result.add_check("eta", batch.eta, V.non_negative_float)
+        result.add_check("generator", batch.generator,
+                         V.generator_or_list_generators)
+        result.add_check("do_classifier_free_guidance",
+                         batch.do_classifier_free_guidance, V.bool_value)
+        result.add_check(
+            "negative_prompt_embeds", batch.negative_prompt_embeds, lambda x:
+            not batch.do_classifier_free_guidance or V.list_not_empty(x))
+        return result

fastvideo/pipelines/stages/decoding.py

@@ -50,6 +50,63 @@ class DecodingStage(PipelineStage):
         result.add_check("output", batch.output, [V.is_tensor, V.with_dims(5)])
         return result
 
+    @torch.no_grad()
+    def decode(self, latents: torch.Tensor,
+               fastvideo_args: FastVideoArgs) -> torch.Tensor:
+        """
+        Decode latent representations into pixel space using VAE.
+        
+        Args:
+            latents: Input latent tensor with shape (batch, channels, frames, height_latents, width_latents)
+            fastvideo_args: Configuration containing:
+                - disable_autocast: Whether to disable automatic mixed precision (default: False)
+                - pipeline_config.vae_precision: VAE computation precision ("fp32", "fp16", "bf16")
+                - pipeline_config.vae_tiling: Whether to enable VAE tiling for memory efficiency
+            
+        Returns:
+            Decoded video tensor with shape (batch, channels, frames, height, width), 
+            normalized to [0, 1] range and moved to CPU as float32
+        """
+        self.vae = self.vae.to(get_local_torch_device())
+        latents = latents.to(get_local_torch_device())
+
+        # Setup VAE precision
+        vae_dtype = PRECISION_TO_TYPE[
+            fastvideo_args.pipeline_config.vae_precision]
+        vae_autocast_enabled = (
+            vae_dtype != torch.float32) and not fastvideo_args.disable_autocast
+
+        if isinstance(self.vae.scaling_factor, torch.Tensor):
+            latents = latents / self.vae.scaling_factor.to(
+                latents.device, latents.dtype)
+        else:
+            latents = latents / self.vae.scaling_factor
+
+        # Apply shifting if needed
+        if (hasattr(self.vae, "shift_factor")
+                and self.vae.shift_factor is not None):
+            if isinstance(self.vae.shift_factor, torch.Tensor):
+                latents += self.vae.shift_factor.to(latents.device,
+                                                    latents.dtype)
+            else:
+                latents += self.vae.shift_factor
+
+        # Decode latents
+        with torch.autocast(device_type="cuda",
+                            dtype=vae_dtype,
+                            enabled=vae_autocast_enabled):
+            if fastvideo_args.pipeline_config.vae_tiling:
+                self.vae.enable_tiling()
+            # if fastvideo_args.vae_sp:
+            #     self.vae.enable_parallel()
+            if not vae_autocast_enabled:
+                latents = latents.to(vae_dtype)
+            image = self.vae.decode(latents)
+
+        # Normalize image to [0, 1] range
+        image = (image / 2 + 0.5).clamp(0, 1)
+        return image
+
     @torch.no_grad()
     def forward(
         self,
@@ -59,13 +116,28 @@ class DecodingStage(PipelineStage):
         """
         Decode latent representations into pixel space.
         
+        This method processes the batch through the VAE decoder, converting latent
+        representations to pixel-space video/images. It also optionally decodes
+        trajectory latents for visualization purposes.
+        
         Args:
-            batch: The current batch information.
-            fastvideo_args: The inference arguments.
+            batch: The current batch containing:
+                - latents: Tensor to decode (batch, channels, frames, height_latents, width_latents)
+                - return_trajectory_decoded (optional): Flag to decode trajectory latents
+                - trajectory_latents (optional): Latents at different timesteps
+                - trajectory_timesteps (optional): Corresponding timesteps
+            fastvideo_args: Configuration containing:
+                - output_type: "latent" to skip decoding, otherwise decode to pixels
+                - vae_cpu_offload: Whether to offload VAE to CPU after decoding
+                - model_loaded: Track VAE loading state
+                - model_paths: Path to VAE model if loading needed
             
         Returns:
-            The batch with decoded outputs.
+            Modified batch with:
+                - output: Decoded frames (batch, channels, frames, height, width) as CPU float32
+                - trajectory_decoded (if requested): List of decoded frames per timestep
         """
+        # load vae if not already loaded (used for memory constrained devices)
         pipeline = self.pipeline() if self.pipeline else None
         if not fastvideo_args.model_loaded["vae"]:
             loader = VAELoader()
@@ -75,58 +147,29 @@ class DecodingStage(PipelineStage):
                 pipeline.add_module("vae", self.vae)
             fastvideo_args.model_loaded["vae"] = True
 
-        self.vae = self.vae.to(get_local_torch_device())
-
-        latents = batch.latents
-        # TODO(will): remove this once we add input/output validation for stages
-        if latents is None:
-            raise ValueError("Latents must be provided")
-
-        # Skip decoding if output type is latent
         if fastvideo_args.output_type == "latent":
-            image = latents
+            frames = batch.latents
         else:
-            # Setup VAE precision
-            vae_dtype = PRECISION_TO_TYPE[
-                fastvideo_args.pipeline_config.vae_precision]
-            vae_autocast_enabled = (vae_dtype != torch.float32
-                                    ) and not fastvideo_args.disable_autocast
+            frames = self.decode(batch.latents, fastvideo_args)
 
-            if isinstance(self.vae.scaling_factor, torch.Tensor):
-                latents = latents / self.vae.scaling_factor.to(
-                    latents.device, latents.dtype)
-            else:
-                latents = latents / self.vae.scaling_factor
-
-            # Apply shifting if needed
-            if (hasattr(self.vae, "shift_factor")
-                    and self.vae.shift_factor is not None):
-                if isinstance(self.vae.shift_factor, torch.Tensor):
-                    latents += self.vae.shift_factor.to(latents.device,
-                                                        latents.dtype)
-                else:
-                    latents += self.vae.shift_factor
-
-            # Decode latents
-            with torch.autocast(device_type="cuda",
-                                dtype=vae_dtype,
-                                enabled=vae_autocast_enabled):
-                if fastvideo_args.pipeline_config.vae_tiling:
-                    self.vae.enable_tiling()
-                # if fastvideo_args.vae_sp:
-                #     self.vae.enable_parallel()
-                if not vae_autocast_enabled:
-                    latents = latents.to(vae_dtype)
-                image = self.vae.decode(latents)
-
-        # Normalize image to [0, 1] range
-        image = (image / 2 + 0.5).clamp(0, 1)
+        # decode trajectory latents if needed
+        if batch.return_trajectory_decoded:
+            batch.trajectory_decoded = []
+            assert batch.trajectory_latents is not None, "batch should have trajectory latents"
+            for idx in range(batch.trajectory_latents.shape[1]):
+                # batch.trajectory_latents is [batch_size, timesteps, channels, frames, height, width]
+                cur_latent = batch.trajectory_latents[:, idx, :, :, :, :]
+                cur_timestep = batch.trajectory_timesteps[idx]
+                logger.info("decoding trajectory latent for timestep: %s",
+                            cur_timestep)
+                decoded_frames = self.decode(cur_latent, fastvideo_args)
+                batch.trajectory_decoded.append(decoded_frames.cpu().float())
 
         # Convert to CPU float32 for compatibility
-        image = image.cpu().float()
+        frames = frames.cpu().float()
 
         # Update batch with decoded image
-        batch.output = image
+        batch.output = frames
 
         # Offload models if needed
         if hasattr(self, 'maybe_free_model_hooks'):

fastvideo/pipelines/stages/denoising.py

@@ -40,6 +40,13 @@ try:
 except ImportError:
     st_attn_available = False
 
+try:
+    from fastvideo.attention.backends.vmoba import VMOBAAttentionBackend
+    from fastvideo.utils import is_vmoba_available
+    vmoba_attn_available = is_vmoba_available()
+except ImportError:
+    vmoba_attn_available = False
+
 try:
     from fastvideo.attention.backends.video_sparse_attn import (
         VideoSparseAttentionBackend)
@@ -77,8 +84,10 @@ class DenoisingStage(PipelineStage):
             supported_attention_backends=(
                 AttentionBackendEnum.SLIDING_TILE_ATTN,
                 AttentionBackendEnum.VIDEO_SPARSE_ATTN,
-                AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TORCH_SDPA
-            )  # hack
+                AttentionBackendEnum.VMOBA_ATTN,
+                AttentionBackendEnum.FLASH_ATTN,
+                AttentionBackendEnum.TORCH_SDPA,
+                AttentionBackendEnum.SAGE_ATTN_THREE)  # hack
         )
 
     def forward(
@@ -149,7 +158,8 @@ class DenoisingStage(PipelineStage):
         # Prepare image latents and embeddings for I2V generation
         image_embeds = batch.image_embeds
         if len(image_embeds) > 0:
-            assert torch.isnan(image_embeds[0]).sum() == 0
+            assert not torch.isnan(
+                image_embeds[0]).any(), "image_embeds contains nan"
             image_embeds = [
                 image_embed.to(target_dtype) for image_embed in image_embeds
             ]
@@ -186,15 +196,23 @@ class DenoisingStage(PipelineStage):
         # Get latents and embeddings
         latents = batch.latents
         prompt_embeds = batch.prompt_embeds
-        assert torch.isnan(prompt_embeds[0]).sum() == 0
+        assert not torch.isnan(
+            prompt_embeds[0]).any(), "prompt_embeds contains nan"
         if batch.do_classifier_free_guidance:
             neg_prompt_embeds = batch.negative_prompt_embeds
             assert neg_prompt_embeds is not None
-            assert torch.isnan(neg_prompt_embeds[0]).sum() == 0
+            assert not torch.isnan(
+                neg_prompt_embeds[0]).any(), "neg_prompt_embeds contains nan"
 
         # (Wan2.2) Calculate timestep to switch from high noise expert to low noise expert
-        if fastvideo_args.boundary_ratio is not None:
-            boundary_timestep = fastvideo_args.boundary_ratio * self.scheduler.num_train_timesteps
+        boundary_ratio = fastvideo_args.pipeline_config.dit_config.boundary_ratio
+        if batch.boundary_ratio is not None:
+            logger.info("Overriding boundary ratio from %s to %s",
+                        boundary_ratio, batch.boundary_ratio)
+            boundary_ratio = batch.boundary_ratio
+
+        if boundary_ratio is not None:
+            boundary_timestep = boundary_ratio * self.scheduler.num_train_timesteps
         else:
             boundary_timestep = None
         latent_model_input = latents.to(target_dtype)
@@ -236,6 +254,10 @@ class DenoisingStage(PipelineStage):
                                                              patch_size[2])
             seq_len = int(math.ceil(seq_len / sp_world_size)) * sp_world_size
 
+        # Initialize lists for ODE trajectory
+        trajectory_timesteps: list[torch.Tensor] = []
+        trajectory_latents: list[torch.Tensor] = []
+
         # Run denoising loop
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
@@ -268,6 +290,9 @@ class DenoisingStage(PipelineStage):
                     latent_model_input = torch.cat(
                         [latent_model_input, batch.image_latent],
                         dim=1).to(target_dtype)
+
+                assert not torch.isnan(
+                    latent_model_input).any(), "latent_model_input contains nan"
                 if fastvideo_args.pipeline_config.ti2v_task and batch.pil_image is not None:
                     timestep = torch.stack([t]).to(get_local_torch_device())
                     temp_ts = (mask2[0][0][:, ::2, ::2] * timestep).flatten()
@@ -280,7 +305,6 @@ class DenoisingStage(PipelineStage):
                 else:
                     t_expand = t.repeat(latent_model_input.shape[0])
 
-                assert torch.isnan(latent_model_input).sum() == 0
                 latent_model_input = self.scheduler.scale_model_input(
                     latent_model_input, t)
 
@@ -325,6 +349,31 @@ class DenoisingStage(PipelineStage):
                             assert attn_metadata is not None, "attn_metadata cannot be None"
                         else:
                             attn_metadata = None
+                    elif (vmoba_attn_available
+                          and self.attn_backend == VMOBAAttentionBackend):
+                        self.attn_metadata_builder_cls = self.attn_backend.get_builder_cls(
+                        )
+                        if self.attn_metadata_builder_cls is not None:
+                            self.attn_metadata_builder = self.attn_metadata_builder_cls(
+                            )
+                            # Prepare V-MoBA parameters from config
+                            moba_params = fastvideo_args.moba_config.copy()
+                            moba_params.update({
+                                "current_timestep":
+                                i,
+                                "raw_latent_shape":
+                                batch.raw_latent_shape[2:5],
+                                "patch_size":
+                                fastvideo_args.pipeline_config.dit_config.
+                                patch_size,
+                                "device":
+                                get_local_torch_device(),
+                            })
+                            attn_metadata = self.attn_metadata_builder.build(
+                                **moba_params)
+                            assert attn_metadata is not None, "attn_metadata cannot be None"
+                        else:
+                            attn_metadata = None
                     else:
                         attn_metadata = None
                     # TODO(will): finalize the interface. vLLM uses this to
@@ -389,6 +438,11 @@ class DenoisingStage(PipelineStage):
                         latents = (1. - mask2[0]) * z + mask2[0] * latents
                         # latents = latents.unsqueeze(0)
 
+                # save trajectory latents if needed
+                if batch.return_trajectory_latents:
+                    trajectory_timesteps.append(t)
+                    trajectory_latents.append(latents)
+
                 # Update progress bar
                 if i == len(timesteps) - 1 or (
                     (i + 1) > num_warmup_steps and
@@ -396,9 +450,28 @@ class DenoisingStage(PipelineStage):
                         and progress_bar is not None):
                     progress_bar.update()
 
+        # Gather results if using sequence parallelism
+        trajectory_tensor: torch.Tensor | None = None
+        if trajectory_latents:
+            trajectory_tensor = torch.stack(trajectory_latents, dim=1)
+            trajectory_timesteps_tensor = torch.stack(trajectory_timesteps,
+                                                      dim=0)
+        else:
+            trajectory_tensor = None
+            trajectory_timesteps_tensor = None
+
         # Gather results if using sequence parallelism
         if sp_group:
             latents = sequence_model_parallel_all_gather(latents, dim=2)
+            if batch.return_trajectory_latents:
+                trajectory_tensor = trajectory_tensor.to(
+                    get_local_torch_device())
+                trajectory_tensor = sequence_model_parallel_all_gather(
+                    trajectory_tensor, dim=3)
+
+        if trajectory_tensor is not None and trajectory_timesteps_tensor is not None:
+            batch.trajectory_timesteps = trajectory_timesteps_tensor.cpu()
+            batch.trajectory_latents = trajectory_tensor.cpu()
 
         # Update batch with final latents
         batch.latents = latents
@@ -737,7 +810,8 @@ class DmdDenoisingStage(DenoisingStage):
 
         video_raw_latent_shape = latents.shape
         prompt_embeds = batch.prompt_embeds
-        assert torch.isnan(prompt_embeds[0]).sum() == 0
+        assert not torch.isnan(
+            prompt_embeds[0]).any(), "prompt_embeds contains nan"
         timesteps = torch.tensor(
             fastvideo_args.pipeline_config.dmd_denoising_steps,
             dtype=torch.long,
@@ -776,7 +850,8 @@ class DmdDenoisingStage(DenoisingStage):
                         batch.image_latent.permute(0, 2, 1, 3, 4)
                     ],
                                                    dim=2).to(target_dtype)
-                assert torch.isnan(latent_model_input).sum() == 0
+                assert not torch.isnan(
+                    latent_model_input).any(), "latent_model_input contains nan"
 
                 # Prepare inputs for transformer
                 t_expand = t.repeat(latent_model_input.shape[0])
@@ -838,7 +913,6 @@ class DmdDenoisingStage(DenoisingStage):
                             **pos_cond_kwargs,
                         ).permute(0, 2, 1, 3, 4)
 
-                    from fastvideo.models.utils import pred_noise_to_pred_video
                     pred_video = pred_noise_to_pred_video(
                         pred_noise=pred_noise.flatten(0, 1),
                         noise_input_latent=noise_latents.flatten(0, 1),

fastvideo/pipelines/stages/encoding.py

@@ -92,24 +92,6 @@ class EncodingStage(PipelineStage):
                 latents = latents.to(vae_dtype)
             latents = self.vae.encode(latents).mean
 
-        # Apply shifting if needed (reverse of decoding)
-        if (hasattr(self.vae, "shift_factor")
-                and self.vae.shift_factor is not None):
-            if isinstance(self.vae.shift_factor, torch.Tensor):
-                latents -= self.vae.shift_factor.to(latents.device,
-                                                    latents.dtype)
-            else:
-                latents -= self.vae.shift_factor
-
-        # Apply scaling factor
-        if (hasattr(self.vae, "scaling_factor")
-                and self.vae.scaling_factor is not None):
-            if isinstance(self.vae.scaling_factor, torch.Tensor):
-                latents = latents * self.vae.scaling_factor.to(
-                    latents.device, latents.dtype)
-            else:
-                latents = latents * self.vae.scaling_factor
-
         # Update batch with encoded latents
         batch.latents = latents
 

fastvideo/pipelines/stages/text_encoding.py

@@ -59,58 +59,37 @@ class TextEncodingStage(PipelineStage):
         assert len(self.text_encoders) == len(
             fastvideo_args.pipeline_config.text_encoder_configs)
 
-        for tokenizer, text_encoder, encoder_config, preprocess_func, postprocess_func in zip(
-                self.tokenizers,
-                self.text_encoders,
-                fastvideo_args.pipeline_config.text_encoder_configs,
-                fastvideo_args.pipeline_config.preprocess_text_funcs,
-                fastvideo_args.pipeline_config.postprocess_text_funcs,
-                strict=True):
+        # Encode positive prompt with all available encoders
+        assert batch.prompt is not None
+        prompt_text: str | list[str] = batch.prompt
+        all_indices: list[int] = list(range(len(self.text_encoders)))
+        prompt_embeds_list, prompt_masks_list = self.encode_text(
+            prompt_text,
+            fastvideo_args,
+            encoder_index=all_indices,
+            return_attention_mask=True,
+        )
+        for pe in prompt_embeds_list:
+            batch.prompt_embeds.append(pe)
+        if batch.prompt_attention_mask is not None:
+            for am in prompt_masks_list:
+                batch.prompt_attention_mask.append(am)
 
-            assert isinstance(batch.prompt, str | list)
-            if isinstance(batch.prompt, str):
-                batch.prompt = [batch.prompt]
-            texts = []
-            for prompt_str in batch.prompt:
-                texts.append(preprocess_func(prompt_str))
-            text_inputs = tokenizer(texts,
-                                    **encoder_config.tokenizer_kwargs).to(
-                                        get_local_torch_device())
-            input_ids = text_inputs["input_ids"]
-            attention_mask = text_inputs["attention_mask"]
-            with set_forward_context(current_timestep=0, attn_metadata=None):
-                outputs = text_encoder(
-                    input_ids=input_ids,
-                    attention_mask=attention_mask,
-                    output_hidden_states=True,
-                )
-            prompt_embeds = postprocess_func(outputs)
-            batch.prompt_embeds.append(prompt_embeds)
-            if batch.prompt_attention_mask is not None:
-                batch.prompt_attention_mask.append(attention_mask)
-
-            if batch.do_classifier_free_guidance:
-                assert isinstance(batch.negative_prompt, str)
-                negative_text = preprocess_func(batch.negative_prompt)
-                negative_text_inputs = tokenizer(
-                    negative_text, **encoder_config.tokenizer_kwargs).to(
-                        get_local_torch_device())
-                negative_input_ids = negative_text_inputs["input_ids"]
-                negative_attention_mask = negative_text_inputs["attention_mask"]
-                with set_forward_context(current_timestep=0,
-                                         attn_metadata=None):
-                    negative_outputs = text_encoder(
-                        input_ids=negative_input_ids,
-                        attention_mask=negative_attention_mask,
-                        output_hidden_states=True,
-                    )
-                negative_prompt_embeds = postprocess_func(negative_outputs)
-
-                assert batch.negative_prompt_embeds is not None
-                batch.negative_prompt_embeds.append(negative_prompt_embeds)
-                if batch.negative_attention_mask is not None:
-                    batch.negative_attention_mask.append(
-                        negative_attention_mask)
+        # Encode negative prompt if CFG is enabled
+        if batch.do_classifier_free_guidance:
+            assert isinstance(batch.negative_prompt, str)
+            neg_embeds_list, neg_masks_list = self.encode_text(
+                batch.negative_prompt,
+                fastvideo_args,
+                encoder_index=all_indices,
+                return_attention_mask=True,
+            )
+            assert batch.negative_prompt_embeds is not None
+            for ne in neg_embeds_list:
+                batch.negative_prompt_embeds.append(ne)
+            if batch.negative_attention_mask is not None:
+                for nm in neg_masks_list:
+                    batch.negative_attention_mask.append(nm)
 
         return batch
 
@@ -129,6 +108,171 @@ class TextEncodingStage(PipelineStage):
                          V.none_or_list)
         return result
 
+    @torch.no_grad()
+    def encode_text(
+        self,
+        text: str | list[str],
+        fastvideo_args: FastVideoArgs,
+        encoder_index: int | list[int] | None = None,
+        return_attention_mask: bool = False,
+        return_type: str = "list",  # one of: "list", "dict", "stack"
+        device: torch.device | str | None = None,
+        dtype: torch.dtype | None = None,
+        max_length: int | None = None,
+        truncation: bool | None = None,
+        padding: bool | str | None = None,
+    ):
+        """
+        Encode plain text using selected text encoder(s) and return embeddings.
+
+        Args:
+            text: A single string or a list of strings to encode.
+            fastvideo_args: The inference arguments providing pipeline config,
+                including tokenizer and encoder settings, preprocess and postprocess
+                functions.
+            encoder_index: Encoder selector by index. Accepts an int or list of ints.
+            return_attention_mask: If True, also return attention masks for each
+                selected encoder.
+            return_type: "list" (default) returns a list aligned with selection;
+                "dict" returns a dict keyed by encoder index as a string; "stack" stacks along a
+                new first dimension (requires matching shapes).
+            device: Optional device override for inputs; defaults to local torch device.
+            dtype: Optional dtype to cast returned embeddings to.
+            max_length: Optional per-call tokenizer override.
+            truncation: Optional per-call tokenizer override.
+            padding: Optional per-call tokenizer override.
+
+        Returns:
+            Depending on return_type and return_attention_mask:
+            - list: List[Tensor] or (List[Tensor], List[Tensor])
+            - dict: Dict[str, Tensor] or (Dict[str, Tensor], Dict[str, Tensor])
+            - stack: Tensor of shape [num_encoders, ...] or a tuple with stacked
+              attention masks
+        """
+
+        assert len(self.tokenizers) == len(self.text_encoders)
+        assert len(self.text_encoders) == len(
+            fastvideo_args.pipeline_config.text_encoder_configs)
+
+        # Resolve selection into indices
+        encoder_cfgs = fastvideo_args.pipeline_config.text_encoder_configs
+        if encoder_index is None:
+            indices: list[int] = [0]
+        elif isinstance(encoder_index, int):
+            indices = [encoder_index]
+        else:
+            indices = list(encoder_index)
+        # validate range
+        num_encoders = len(self.text_encoders)
+        for idx in indices:
+            if idx < 0 or idx >= num_encoders:
+                raise IndexError(
+                    f"encoder index {idx} out of range [0, {num_encoders-1}]")
+
+        # Validate indices are within range
+        num_encoders = len(self.text_encoders)
+
+        # Normalize input to list[str]
+        assert isinstance(text, str | list)
+        if isinstance(text, str):
+            texts: list[str] = [text]
+        else:
+            texts = text
+
+        embeds_list: list[torch.Tensor] = []
+        attn_masks_list: list[torch.Tensor] = []
+
+        preprocess_funcs = fastvideo_args.pipeline_config.preprocess_text_funcs
+        postprocess_funcs = fastvideo_args.pipeline_config.postprocess_text_funcs
+        encoder_cfgs = fastvideo_args.pipeline_config.text_encoder_configs
+
+        if return_type not in ("list", "dict", "stack"):
+            raise ValueError(
+                f"Invalid return_type '{return_type}'. Expected one of: 'list', 'dict', 'stack'"
+            )
+
+        target_device = device if device is not None else get_local_torch_device(
+        )
+
+        for i in indices:
+            tokenizer = self.tokenizers[i]
+            text_encoder = self.text_encoders[i]
+            encoder_config = encoder_cfgs[i]
+            preprocess_func = preprocess_funcs[i]
+            postprocess_func = postprocess_funcs[i]
+
+            processed_texts: list[str] = []
+            for prompt_str in texts:
+                processed_texts.append(preprocess_func(prompt_str))
+
+            tok_kwargs = dict(encoder_config.tokenizer_kwargs)
+            if max_length is not None:
+                tok_kwargs["max_length"] = max_length
+            if truncation is not None:
+                tok_kwargs["truncation"] = truncation
+            if padding is not None:
+                tok_kwargs["padding"] = padding
+
+            text_inputs = tokenizer(processed_texts,
+                                    **tok_kwargs).to(target_device)
+
+            input_ids = text_inputs["input_ids"]
+            attention_mask = text_inputs["attention_mask"]
+
+            with set_forward_context(current_timestep=0, attn_metadata=None):
+                outputs = text_encoder(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    output_hidden_states=True,
+                )
+
+            prompt_embeds = postprocess_func(outputs)
+            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)
+
+        # Shape results according to return_type
+        if return_type == "list":
+            if return_attention_mask:
+                return embeds_list, attn_masks_list
+            return embeds_list
+
+        if return_type == "dict":
+            key_strs = [str(i) for i in indices]
+            embeds_dict = {
+                k: v
+                for k, v in zip(key_strs, embeds_list, strict=False)
+            }
+            if return_attention_mask:
+                attn_dict = {
+                    k: v
+                    for k, v in zip(key_strs, attn_masks_list, strict=False)
+                }
+                return embeds_dict, attn_dict
+            return embeds_dict
+
+        # return_type == "stack"
+        # Validate shapes are compatible
+        base_shape = list(embeds_list[0].shape)
+        for t in embeds_list[1:]:
+            if list(t.shape) != base_shape:
+                raise ValueError(
+                    f"Cannot stack embeddings with differing shapes: {[list(t.shape) for t in embeds_list]}"
+                )
+        stacked_embeds = torch.stack(embeds_list, dim=0)
+        if return_attention_mask:
+            base_mask_shape = list(attn_masks_list[0].shape)
+            for m in attn_masks_list[1:]:
+                if list(m.shape) != base_mask_shape:
+                    raise ValueError(
+                        f"Cannot stack attention masks with differing shapes: {[list(m.shape) for m in attn_masks_list]}"
+                    )
+            stacked_masks = torch.stack(attn_masks_list, dim=0)
+            return stacked_embeds, stacked_masks
+        return stacked_embeds
+
     def verify_output(self, batch: ForwardBatch,
                       fastvideo_args: FastVideoArgs) -> VerificationResult:
         """Verify text encoding stage outputs."""

fastvideo/platforms/cuda.py

@@ -115,6 +115,7 @@ class CudaPlatformBase(Platform):
 
         logger.info("Trying FASTVIDEO_ATTENTION_BACKEND=%s",
                     envs.FASTVIDEO_ATTENTION_BACKEND)
+        logger.info("Selected backend: %s", selected_backend)
         if selected_backend == AttentionBackendEnum.SLIDING_TILE_ATTN:
             try:
                 from st_attn import sliding_tile_attention  # noqa: F401
@@ -144,6 +145,20 @@ class CudaPlatformBase(Platform):
                 logger.info(
                     "Sage Attention backend is not installed. Fall back to Flash Attention."
                 )
+        elif selected_backend == AttentionBackendEnum.SAGE_ATTN_THREE:
+            try:
+                from fastvideo.attention.backends.sageattn.api import sageattn_blackwell  # noqa: F401
+
+                from fastvideo.attention.backends.sage_attn3 import (  # noqa: F401
+                    SageAttention3Backend)
+                logger.info("Using Sage Attention 3 backend.")
+
+                return "fastvideo.attention.backends.sage_attn3.SageAttention3Backend"
+            except ImportError as e:
+                logger.info(e)
+                logger.info(
+                    "Sage Attention 3 backend is not installed. Fall back to Flash Attention."
+                )
         elif selected_backend == AttentionBackendEnum.VIDEO_SPARSE_ATTN:
             try:
                 from vsa import block_sparse_attn  # noqa: F401
@@ -159,6 +174,20 @@ class CudaPlatformBase(Platform):
                     str(e))
                 raise ImportError(
                     "Video Sparse Attention backend is not installed. ") from e
+        elif selected_backend == AttentionBackendEnum.VMOBA_ATTN:
+            try:
+                from csrc.attn.vmoba_attn.vmoba import (  # noqa: F401
+                    moba_attn_varlen)
+                from fastvideo.attention.backends.vmoba import (  # noqa: F401
+                    VMOBAAttentionBackend)
+                logger.info("Using Video MOBA Attention backend.")
+
+                return "fastvideo.attention.backends.vmoba.VMOBAAttentionBackend"
+            except ImportError as e:
+                logger.error(
+                    "Failed to import Video MoBA Attention backend: %s", str(e))
+                raise ImportError(
+                    "Video MoBA Attention backend is not installed. ") from e
         elif selected_backend == AttentionBackendEnum.TORCH_SDPA:
             logger.info("Using Torch SDPA backend.")
             return "fastvideo.attention.backends.sdpa.SDPABackend"

fastvideo/platforms/interface.py

@@ -18,7 +18,9 @@ class AttentionBackendEnum(enum.Enum):
     SLIDING_TILE_ATTN = enum.auto()
     TORCH_SDPA = enum.auto()
     SAGE_ATTN = enum.auto()
+    SAGE_ATTN_THREE = enum.auto()
     VIDEO_SPARSE_ATTN = enum.auto()
+    VMOBA_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()
 
 

fastvideo/tests/training/VSA/reference_wandb_summary_VSA.json

@@ -1 +0,0 @@
-{"step_time":0.6983645600266755,"_wandb":{"runtime":107},"grad_norm":0.50390625,"avg_step_time":1.002151239803061,"_step":5,"validation_videos_50_steps":{"captions":false,"_type":"videos","count":8,"videos":[{"size":159131,"path":"media/videos/validation_videos_50_steps_0_dc447599dbe48350e9c9.mp4","_type":"video-file","sha256":"dc447599dbe48350e9c920f4971e1786bde580dd20d52b4aa147ae8d3dc564d6"},{"_type":"video-file","sha256":"4e283876ddfbf5a2cb6f8aca07a39f832b5f806fbefde8854bc73ed904ff20ee","size":160315,"path":"media/videos/validation_videos_50_steps_0_4e283876ddfbf5a2cb6f.mp4"},{"size":135225,"path":"media/videos/validation_videos_50_steps_0_78185c41e1935306e93c.mp4","_type":"video-file","sha256":"78185c41e1935306e93c2d416ee40b31d038abffb029cb5bfb11c2a634eb2fcf"},{"_type":"video-file","sha256":"27e9819d002d3f63c8918bbdc5bf2857b5effe0caf5d5b8374b9c590fc6432eb","size":197873,"path":"media/videos/validation_videos_50_steps_0_27e9819d002d3f63c891.mp4"},{"_type":"video-file","sha256":"46fe548e86144ca60a9396fcd15e8788d3a05c066c6819ccdd6a9041feaaec8f","size":170601,"path":"media/videos/validation_videos_50_steps_0_46fe548e86144ca60a93.mp4"},{"sha256":"91ec338774bec870b9c5c81be4330a8f0f2535124cb372e881c3703e6b65ed77","size":164462,"path":"media/videos/validation_videos_50_steps_0_91ec338774bec870b9c5.mp4","_type":"video-file"},{"_type":"video-file","sha256":"ee4e811080a619215fd7541b39203dfb69c2db4cdadbfdd7a234a2cff684f6f3","size":139435,"path":"media/videos/validation_videos_50_steps_0_ee4e811080a619215fd7.mp4"},{"sha256":"22e31e048ba5e5b9d6587d7306904d6edc6c618b8f77c3ceb05ba2d602309274","size":147072,"path":"media/videos/validation_videos_50_steps_0_22e31e048ba5e5b9d658.mp4","_type":"video-file"}]},"_timestamp":1.75118195270901e+09,"vsa_sparsity":0.05,"learning_rate":1e-05,"train_loss":0.19960195198655128,"_runtime":107.325113071}

fastvideo/tests/training/Vanilla/l40s_reference_wandb_summary.json

@@ -1 +0,0 @@
-{"train_loss":0.10545400530099869,"_step":5,"_wandb":{"runtime":35},"step_time":2.2575189135968685,"grad_norm":0.53125,"_runtime":35.701958502,"avg_step_time":2.4754387199878694,"_timestamp":1.7525528420185745e+09,"learning_rate":1e-06,"vsa_sparsity":0,"validation_videos_8_steps":{"videos":[{"caption":"A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","_type":"video-file","sha256":"ee83e83df073a648f89dcd288cccaed9af765e11fc35c77a6e0cb2ebaa1be5b0","size":475248,"path":"media/videos/validation_videos_8_steps_0_ee83e83df073a648f89d.mp4"},{"size":341490,"path":"media/videos/validation_videos_8_steps_0_d0b2758549d5c82845ca.mp4","caption":"A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","_type":"video-file","sha256":"d0b2758549d5c82845ca3c1ac0db6812877631495e77ac041acdc8ea31f0a5ee"},{"caption":"A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.","_type":"video-file","sha256":"08638381f1607d6ab10684772be38a58b18628fd402a208b25f6af96e765d454","size":436814,"path":"media/videos/validation_videos_8_steps_0_08638381f1607d6ab106.mp4"}],"captions":["A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.","A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.","A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table."],"_type":"videos","count":3}}

fastvideo/training/distillation_pipeline.py

@@ -91,9 +91,8 @@ class DistillationPipeline(TrainingPipeline):
             shift=self.timestep_shift)
 
         if training_args.real_score_model_path:
-            logger.info(
-                f"Loading real score transformer from: {training_args.real_score_model_path}"
-            )
+            logger.info("Loading real score transformer from: %s",
+                        training_args.real_score_model_path)
             self.real_score_transformer = self.load_module_from_path(
                 training_args.real_score_model_path, "transformer",
                 training_args)
@@ -102,9 +101,8 @@ class DistillationPipeline(TrainingPipeline):
                 "real_score_transformer")
 
         if training_args.fake_score_model_path:
-            logger.info(
-                f"Loading fake score transformer from: {training_args.fake_score_model_path}"
-            )
+            logger.info("Loading fake score transformer from: %s",
+                        training_args.fake_score_model_path)
             self.fake_score_transformer = self.load_module_from_path(
                 training_args.fake_score_model_path, "transformer",
                 training_args)
@@ -193,14 +191,13 @@ class DistillationPipeline(TrainingPipeline):
 
         self.real_score_guidance_scale = self.training_args.real_score_guidance_scale
 
-        self.generator_ema = None
+        self.generator_ema: EMA_FSDP | None = None
         if (self.training_args.ema_decay
                 is not None) and (self.training_args.ema_decay > 0.0):
             self.generator_ema = EMA_FSDP(self.transformer,
                                           decay=self.training_args.ema_decay)
-            logger.info(
-                f"Initialized generator EMA with decay={self.training_args.ema_decay}"
-            )
+            logger.info("Initialized generator EMA with decay=%s",
+                        self.training_args.ema_decay)
         else:
             logger.info("Generator EMA disabled (ema_decay <= 0.0)")
 
@@ -217,7 +214,7 @@ class DistillationPipeline(TrainingPipeline):
         Returns:
             The loaded module
         """
-        logger.info(f"Loading {module_type} from custom path: {model_path}")
+        logger.info("Loading %s from custom path: %s", module_type, model_path)
         # Set flag to prevent custom weight loading for teacher/critic models
         training_args._loading_teacher_critic_model = True
 
@@ -227,7 +224,7 @@ class DistillationPipeline(TrainingPipeline):
 
             # Download the model if it's a Hugging Face model ID
             local_model_path = maybe_download_model(model_path)
-            logger.info(f"Model downloaded/found at: {local_model_path}")
+            logger.info("Model downloaded/found at: %s", local_model_path)
             config = verify_model_config_and_directory(local_model_path)
 
             if module_type not in config:
@@ -236,7 +233,8 @@ class DistillationPipeline(TrainingPipeline):
                     extra_module = self._extra_config_module_map[module_type]
                     if extra_module in config:
                         module_type = extra_module
-                        logger.info(f"Using {extra_module} for {module_type}")
+                        logger.info("Using %s for %s", extra_module,
+                                    module_type)
                     else:
                         raise ValueError(
                             f"Module {module_type} not found in config at {local_model_path}"
@@ -261,8 +259,8 @@ class DistillationPipeline(TrainingPipeline):
                 fastvideo_args=training_args,
             )
 
-            logger.info(
-                f"Successfully loaded {module_type} from {component_path}")
+            logger.info("Successfully loaded %s from %s", module_type,
+                        component_path)
             return module
         finally:
             # Always clean up the flag
@@ -292,7 +290,7 @@ class DistillationPipeline(TrainingPipeline):
                 return model
         return model
 
-    def get_ema_model_copy(self):
+    def get_ema_model_copy(self) -> torch.nn.Module | None:
         """Get a copy of the model with EMA weights applied."""
         if self.generator_ema is not None:
             ema_model = copy.deepcopy(self.transformer)
@@ -300,7 +298,7 @@ class DistillationPipeline(TrainingPipeline):
             return ema_model
         return None
 
-    def is_ema_ready(self, current_step: int = None):
+    def is_ema_ready(self, current_step: int | None = None):
         """Check if EMA is ready for use (after ema_start_step)."""
         if current_step is None:
             current_step = getattr(self, 'current_trainstep', 0)
@@ -349,14 +347,14 @@ class DistillationPipeline(TrainingPipeline):
                 with open(config_path, "w") as f:
                     json.dump(config_dict, f, indent=4)
 
-                logger.info(f"EMA weights saved to {weight_path}")
+                logger.info("EMA weights saved to %s", weight_path)
 
             del ema_model
 
         except Exception as e:
-            logger.error(f"Failed to save EMA weights: {str(e)}")
+            logger.error("Failed to save EMA weights: %s", str(e))
 
-    def get_ema_stats(self):
+    def get_ema_stats(self) -> dict[str, Any]:
         """Get EMA statistics for monitoring."""
         if self.generator_ema is None:
             return {
@@ -545,6 +543,7 @@ class DistillationPipeline(TrainingPipeline):
     def _dmd_forward(self, generator_pred_video: torch.Tensor,
                      training_batch: TrainingBatch) -> torch.Tensor:
         """Compute DMD (Diffusion Model Distillation) loss."""
+        original_latent = generator_pred_video
         with torch.no_grad():
             timestep = torch.randint(0,
                                      self.num_train_timestep, [1],
@@ -569,7 +568,8 @@ class DistillationPipeline(TrainingPipeline):
 
             noisy_latent = self.noise_scheduler.add_noise(
                 generator_pred_video.flatten(0, 1), noise.flatten(0, 1),
-                timestep).unflatten(0, (1, generator_pred_video.shape[1]))
+                timestep).detach().unflatten(0,
+                                             (1, generator_pred_video.shape[1]))
 
             # fake_score_transformer forward
             training_batch = self._build_distill_input_kwargs(
@@ -618,24 +618,24 @@ class DistillationPipeline(TrainingPipeline):
                 pred_real_video_uncond) * self.real_score_guidance_scale
 
             grad = (faker_score_pred_video - real_score_pred_video) / torch.abs(
-                generator_pred_video - real_score_pred_video).mean()
+                original_latent - real_score_pred_video).mean()
             grad = torch.nan_to_num(grad)
 
         dmd_loss = 0.5 * F.mse_loss(
-            generator_pred_video.float(),
-            (generator_pred_video.float() - grad.float()).detach())
+            original_latent.float(),
+            (original_latent.float() - grad.float()).detach())
 
         training_batch.dmd_latent_vis_dict.update({
             "training_batch_dmd_fwd_clean_latent":
             training_batch.latents,
             "generator_pred_video":
-            generator_pred_video,
+            original_latent.detach(),
             "real_score_pred_video":
-            real_score_pred_video,
+            real_score_pred_video.detach(),
             "faker_score_pred_video":
-            faker_score_pred_video,
+            faker_score_pred_video.detach(),
             "dmd_timestep":
-            timestep,
+            timestep.detach(),
         })
 
         return dmd_loss
@@ -728,16 +728,17 @@ class DistillationPipeline(TrainingPipeline):
             "encoder_hidden_states": training_batch.encoder_hidden_states,
             "encoder_attention_mask": training_batch.encoder_attention_mask,
         }
-        unconditional_dict = {
-            "encoder_hidden_states": self.negative_prompt_embeds,
-            "encoder_attention_mask": self.negative_prompt_attention_mask,
-        }
+        if getattr(self, "negative_prompt_embeds", None) is not None:
+            unconditional_dict = {
+                "encoder_hidden_states": self.negative_prompt_embeds,
+                "encoder_attention_mask": self.negative_prompt_attention_mask,
+            }
+            training_batch.unconditional_dict = unconditional_dict
 
         training_batch.dmd_latent_vis_dict = {}
         training_batch.fake_score_latent_vis_dict = {}
 
         training_batch.conditional_dict = conditional_dict
-        training_batch.unconditional_dict = unconditional_dict
         training_batch.raw_latent_shape = training_batch.latents.shape
         training_batch.latents = training_batch.latents.permute(0, 2, 1, 3, 4)
         self.video_latent_shape = training_batch.latents.shape
@@ -800,6 +801,9 @@ class DistillationPipeline(TrainingPipeline):
                     (dmd_loss / gradient_accumulation_steps).backward()
                 total_dmd_loss += dmd_loss.detach().item()
             self._clip_model_grad_norm_(batch_gen, self.transformer)
+            for param in self.transformer.parameters():
+                # check if the gradient is not None and not zero
+                assert param.grad is not None and param.grad.abs().sum() > 0
             self.optimizer.step()
             self.optimizer.zero_grad(set_to_none=True)
 
@@ -829,6 +833,9 @@ class DistillationPipeline(TrainingPipeline):
             fake_score_latent_vis_dict.update(
                 batch_fake.fake_score_latent_vis_dict)
         self._clip_model_grad_norm_(batch_fake, self.fake_score_transformer)
+        for param in self.fake_score_transformer.parameters():
+            # check if the gradient is not None and not zero
+            assert param.grad is not None and param.grad.abs().sum() > 0
         self.fake_score_optimizer.step()
         self.fake_score_lr_scheduler.step()
         self.lr_scheduler.step()
@@ -1243,9 +1250,8 @@ class DistillationPipeline(TrainingPipeline):
                     (self.generator_ema is None) and (self.training_args.ema_decay > 0):
                 self.generator_ema = EMA_FSDP(
                     self.transformer, decay=self.training_args.ema_decay)
-                logger.info(
-                    f"Created generator EMA at step {step} with decay={self.training_args.ema_decay}"
-                )
+                logger.info("Created generator EMA at step %s with decay=%s",
+                            step, self.training_args.ema_decay)
 
             with torch.autocast("cuda", dtype=torch.bfloat16):
                 training_batch = self.train_one_step(training_batch)

fastvideo/training/ode_causal_pipeline.py

@@ -0,0 +1,406 @@
+# SPDX-License-Identifier: Apache-2.0
+import sys
+from copy import deepcopy
+from typing import cast
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+import wandb
+from fastvideo.dataset.dataloader.schema import (
+    pyarrow_schema_ode_trajectory_text_only)
+from fastvideo.distributed import get_local_torch_device
+from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
+from fastvideo.forward_context import set_forward_context
+from fastvideo.logger import init_logger
+from fastvideo.models.schedulers.scheduling_self_forcing_flow_match import (
+    SelfForcingFlowMatchScheduler)
+from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
+    WanCausalDMDPipeline)
+from fastvideo.pipelines.pipeline_batch_info import TrainingBatch
+from fastvideo.training.training_pipeline import TrainingPipeline
+from fastvideo.training.training_utils import (
+    clip_grad_norm_while_handling_failing_dtensor_cases)
+
+logger = init_logger(__name__)
+
+
+class ODEInitTrainingPipeline(TrainingPipeline):
+    """
+    Training pipeline for ODE-init using precomputed denoising trajectories.
+
+    Supervision: predict the next latent in the stored trajectory by
+    - feeding current latent at timestep t into the transformer to predict noise
+    - stepping the scheduler with the predicted noise
+    - minimizing MSE to the stored next latent at timestep t_next
+    """
+
+    _required_config_modules = ["scheduler", "transformer", "vae"]
+
+    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
+        # Match the preprocess/generation scheduler for consistent stepping
+        self.modules["scheduler"] = SelfForcingFlowMatchScheduler(
+            shift=fastvideo_args.pipeline_config.flow_shift,
+            sigma_min=0.0,
+            extra_one_step=True)
+        self.modules["scheduler"].set_timesteps(num_inference_steps=1000,
+                                                training=True)
+
+    def set_schemas(self):
+        self.train_dataset_schema = pyarrow_schema_ode_trajectory_text_only
+
+    def initialize_training_pipeline(self, training_args: TrainingArgs):
+        super().initialize_training_pipeline(training_args)
+
+        self.noise_scheduler = self.get_module("scheduler")
+        self.vae = self.get_module("vae")
+        self.vae.requires_grad_(False)
+
+        self.timestep_shift = self.training_args.pipeline_config.flow_shift
+        assert self.timestep_shift == 5.0, "flow_shift must be 5.0"
+        self.noise_scheduler = SelfForcingFlowMatchScheduler(
+            shift=self.timestep_shift, sigma_min=0.0, extra_one_step=True)
+        self.noise_scheduler.set_timesteps(num_inference_steps=1000,
+                                           training=True)
+
+        logger.info("dmd_denoising_steps: %s",
+                    self.training_args.pipeline_config.dmd_denoising_steps)
+        self.dmd_denoising_steps = torch.tensor([1000, 750, 500, 250],
+                                                dtype=torch.long,
+                                                device=get_local_torch_device())
+        if training_args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
+            timesteps = torch.cat((self.noise_scheduler.timesteps.cpu(),
+                                   torch.tensor([0],
+                                                dtype=torch.float32))).cuda()
+            logger.info("timesteps: %s", timesteps)
+            self.dmd_denoising_steps = timesteps[1000 -
+                                                 self.dmd_denoising_steps]
+            logger.info("warped self.dmd_denoising_steps: %s",
+                        self.dmd_denoising_steps)
+        else:
+            raise ValueError("warp_denoising_step must be true")
+
+        self.dmd_denoising_steps = self.dmd_denoising_steps.to(
+            get_local_torch_device())
+
+        logger.info("denoising_step_list: %s", self.dmd_denoising_steps)
+
+        logger.info(
+            "Initialized ODE-init training pipeline with %s denoising steps",
+            len(self.dmd_denoising_steps))
+        # Cache for nearest trajectory index per DMD step (computed lazily on first batch)
+        self._cached_closest_idx_per_dmd = None
+        self.num_train_timestep = self.noise_scheduler.num_train_timesteps
+        self.manual_idx = 0
+
+    def initialize_validation_pipeline(self, training_args: TrainingArgs):
+        logger.info("Initializing validation pipeline...")
+        args_copy = deepcopy(training_args)
+        args_copy.inference_mode = True
+        # Warm start validation with current transformer
+        self.validation_pipeline = WanCausalDMDPipeline.from_pretrained(
+            training_args.model_path,
+            args=args_copy,  # type: ignore
+            inference_mode=True,
+            loaded_modules={
+                "transformer": self.get_module("transformer"),
+            },
+            tp_size=training_args.tp_size,
+            sp_size=training_args.sp_size,
+            num_gpus=training_args.num_gpus,
+            pin_cpu_memory=training_args.pin_cpu_memory,
+            dit_cpu_offload=True)
+
+    def _get_next_batch(
+            self,
+            training_batch) -> tuple[TrainingBatch, torch.Tensor, torch.Tensor]:
+        batch = next(self.train_loader_iter, None)  # type: ignore
+        if batch is None:
+            self.current_epoch += 1
+            logger.info("Starting epoch %s", self.current_epoch)
+            self.train_loader_iter = iter(self.train_dataloader)
+            batch = next(self.train_loader_iter)
+
+        # Required fields from parquet (ODE trajectory schema)
+        encoder_hidden_states = batch['text_embedding']
+        encoder_attention_mask = batch['text_attention_mask']
+        infos = batch['info_list']
+
+        # Trajectory tensors may include a leading singleton batch dim per row
+        trajectory_latents = batch['trajectory_latents']
+        if trajectory_latents.dim() == 7:
+            # [B, 1, S, C, T, H, W] -> [B, S, C, T, H, W]
+            trajectory_latents = trajectory_latents[:, 0]
+        elif trajectory_latents.dim() == 6:
+            # already [B, S, C, T, H, W]
+            pass
+        else:
+            raise ValueError(
+                f"Unexpected trajectory_latents dim: {trajectory_latents.dim()}"
+            )
+
+        trajectory_timesteps = batch['trajectory_timesteps']
+        if trajectory_timesteps.dim() == 3:
+            # [B, 1, S] -> [B, S]
+            trajectory_timesteps = trajectory_timesteps[:, 0]
+        elif trajectory_timesteps.dim() == 2:
+            # [B, S]
+            pass
+        else:
+            raise ValueError(
+                f"Unexpected trajectory_timesteps dim: {trajectory_timesteps.dim()}"
+            )
+        # [B, S, C, T, H, W] -> [B, S, T, C, H, W] to match self-forcing
+        trajectory_latents = trajectory_latents.permute(0, 1, 3, 2, 4, 5)
+
+        # Move to device
+        device = get_local_torch_device()
+        training_batch.encoder_hidden_states = encoder_hidden_states.to(
+            device, dtype=torch.bfloat16)
+        training_batch.encoder_attention_mask = encoder_attention_mask.to(
+            device, dtype=torch.bfloat16)
+        training_batch.infos = infos
+
+        return training_batch, trajectory_latents.to(
+            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
+
+        ## TEMP Used for loading the sf .pt files directly
+        """
+        self.manual_idx = self.manual_idx % 155
+        path = f"/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_pt_vidprom_1000/{self.manual_idx:05d}.pt"
+        logger.info("path: %s", path)
+        self.manual_idx += 1
+        # path = "/mnt/weka/home/hao.zhang/wl/Self-Forcing/ode_single_full/00000.pt"
+        b = torch.load(path)
+        training_batch.encoder_hidden_states = b["text_embedding"][0].unsqueeze(
+            0).to(device, dtype=torch.bfloat16)
+        trajectory_latents = b["ode_latent"].to(device, dtype=torch.bfloat16)
+        logger.info("trajectory_latents: %s", trajectory_latents.shape)
+        logger.info("encoder_hidden_states: %s",
+                    training_batch.encoder_hidden_states.shape)
+        assert trajectory_latents.shape[1] <= 10, "trajectory_latents.shape[1] must be <= 10"
+        return training_batch, trajectory_latents.to(
+            device, dtype=torch.bfloat16), trajectory_timesteps.to(device)
+        """
+
+    def _get_timestep(self,
+                      min_timestep: int,
+                      max_timestep: int,
+                      batch_size: int,
+                      num_frame: int,
+                      num_frame_per_block: int,
+                      uniform_timestep: bool = False) -> torch.Tensor:
+        if uniform_timestep:
+            timestep = torch.randint(min_timestep,
+                                     max_timestep, [batch_size, 1],
+                                     device=self.device,
+                                     dtype=torch.long).repeat(1, num_frame)
+            return timestep
+        else:
+            timestep = torch.randint(min_timestep,
+                                     max_timestep, [batch_size, num_frame],
+                                     device=self.device,
+                                     dtype=torch.long)
+            # logger.info(f"individual timestep: {timestep}")
+            # make the noise level the same within every block
+            timestep = timestep.reshape(timestep.shape[0], -1,
+                                        num_frame_per_block)
+            timestep[:, :, 1:] = timestep[:, :, 0:1]
+            timestep = timestep.reshape(timestep.shape[0], -1)
+            return timestep
+
+    def _step_predict_next_latent(
+        self, traj_latents: torch.Tensor, traj_timesteps: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, dict[str,
+                                                              torch.Tensor]]:
+        latent_vis_dict: dict[str, torch.Tensor] = {}
+        device = get_local_torch_device()
+        target_latent = traj_latents[:, -1]
+
+        # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
+        B, S, num_frames, num_channels, height, width = traj_latents.shape
+
+        # Lazily cache nearest trajectory index per DMD step based on the (fixed) S timesteps
+        if self._cached_closest_idx_per_dmd is None:
+            self._cached_closest_idx_per_dmd = torch.tensor(
+                [0, 12, 24, 36], dtype=torch.long).cpu()
+            # [0, 1, 2, 3], dtype=torch.long).cpu()
+            logger.info("self._cached_closest_idx_per_dmd: %s",
+                        self._cached_closest_idx_per_dmd)
+            logger.info(
+                "corresponding timesteps: %s", self.noise_scheduler.timesteps[
+                    self._cached_closest_idx_per_dmd])
+
+        # Select the K indexes from traj_latents using self._cached_closest_idx_per_dmd
+        # traj_latents: [B, S, C, T, H, W], self._cached_closest_idx_per_dmd: [K]
+        # Output: [B, K, C, T, H, W]
+        assert self._cached_closest_idx_per_dmd is not None
+        relevant_traj_latents = torch.index_select(
+            traj_latents,
+            dim=1,
+            index=self._cached_closest_idx_per_dmd.to(traj_latents.device))
+        logger.info("relevant_traj_latents: %s", relevant_traj_latents.shape)
+        # assert relevant_traj_latents.shape[0] == 1
+
+        indexes = self._get_timestep(  # [B, num_frames]
+            0,
+            len(self.dmd_denoising_steps),
+            B,
+            num_frames,
+            3,
+            uniform_timestep=False)
+        logger.info("indexes: %s", indexes.shape)
+        logger.info("indexes: %s", indexes)
+        # noisy_input = relevant_traj_latents[indexes]
+        noisy_input = torch.gather(
+            relevant_traj_latents,
+            dim=1,
+            index=indexes.reshape(B, 1, num_frames, 1, 1,
+                                  1).expand(-1, -1, -1, num_channels, height,
+                                            width).to(self.device)).squeeze(1)
+        timestep = self.dmd_denoising_steps[indexes]
+        logger.info("selected timestep for rank %s: %s",
+                    self.global_rank,
+                    timestep,
+                    local_main_process_only=False)
+
+        # Prepare inputs for transformer
+        latent_vis_dict["noisy_input"] = noisy_input.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()
+        latent_vis_dict["x0"] = target_latent.permute(0, 2, 1, 3,
+                                                      4).detach().clone().cpu()
+
+        input_kwargs = {
+            "hidden_states": noisy_input.permute(0, 2, 1, 3, 4),
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep.to(device, dtype=torch.bfloat16),
+            "return_dict": False,
+        }
+        # Predict noise and step the scheduler to obtain next latent
+        with set_forward_context(current_timestep=timestep,
+                                 attn_metadata=None,
+                                 forward_batch=None):
+            noise_pred = self.transformer(**input_kwargs).permute(0, 2, 1, 3, 4)
+
+        from fastvideo.models.utils import pred_noise_to_pred_video
+        pred_video = pred_noise_to_pred_video(
+            pred_noise=noise_pred.flatten(0, 1),
+            noise_input_latent=noisy_input.flatten(0, 1),
+            timestep=timestep.to(dtype=torch.bfloat16).flatten(0, 1),
+            scheduler=self.modules["scheduler"]).unflatten(
+                0, noise_pred.shape[:2])
+        latent_vis_dict["pred_video"] = pred_video.permute(
+            0, 2, 1, 3, 4).detach().clone().cpu()
+
+        return pred_video, target_latent, timestep, latent_vis_dict
+
+    def train_one_step(self, training_batch):  # type: ignore[override]
+        self.transformer.train()
+        self.optimizer.zero_grad()
+        training_batch.total_loss = 0.0
+        args = cast(TrainingArgs, self.training_args)
+
+        # Using cached nearest index per DMD step; computation happens in _step_predict_next_latent
+
+        for _ in range(args.gradient_accumulation_steps):
+            training_batch, traj_latents, traj_timesteps = self._get_next_batch(
+                training_batch)
+            text_embeds = training_batch.encoder_hidden_states
+            text_attention_mask = training_batch.encoder_attention_mask
+            assert traj_latents.shape[0] == 1
+
+            # Shapes: traj_latents [B, S, C, T, H, W], traj_timesteps [B, S]
+            _, S = traj_latents.shape[0], traj_latents.shape[1]
+            if S < 2:
+                raise ValueError("Trajectory must contain at least 2 steps")
+
+            # Forward to predict next latent by stepping scheduler with predicted noise
+            noise_pred, target_latent, t, latent_vis_dict = self._step_predict_next_latent(
+                traj_latents, traj_timesteps, text_embeds, text_attention_mask)
+
+            training_batch.latent_vis_dict.update(latent_vis_dict)
+
+            mask = t != 0
+
+            # Compute loss
+            loss = F.mse_loss(noise_pred[mask],
+                              target_latent[mask],
+                              reduction="mean")
+            loss = loss / args.gradient_accumulation_steps
+
+            with set_forward_context(current_timestep=t,
+                                     attn_metadata=None,
+                                     forward_batch=None):
+                loss.backward()
+            avg_loss = loss.detach().clone()
+            training_batch.total_loss += avg_loss.item()
+
+        # Clip grad and step optimizers
+        grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
+            [p for p in self.transformer.parameters() if p.requires_grad],
+            args.max_grad_norm if args.max_grad_norm is not None else 0.0)
+
+        self.optimizer.step()
+        self.lr_scheduler.step()
+
+        if grad_norm is None:
+            grad_value = 0.0
+        else:
+            try:
+                if isinstance(grad_norm, torch.Tensor):
+                    grad_value = float(grad_norm.detach().float().item())
+                else:
+                    grad_value = float(grad_norm)
+            except Exception:
+                grad_value = 0.0
+        training_batch.grad_norm = grad_value
+        return training_batch
+
+    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
+                                       training_args: TrainingArgs, step: int):
+        """Add visualization data to wandb logging and save frames to disk."""
+        wandb_loss_dict = {}
+        latents_vis_dict = training_batch.latent_vis_dict
+        latent_log_keys = ['noisy_input', 'x0', 'pred_video']
+        for latent_key in latent_log_keys:
+            assert latent_key in latents_vis_dict and latents_vis_dict[
+                latent_key] is not None
+            latent = latents_vis_dict[latent_key]
+            pixel_latent = self.validation_pipeline.decoding_stage.decode(
+                latent, training_args)
+
+            video = pixel_latent.cpu().float()
+            video = video.permute(0, 2, 1, 3, 4)
+            video = (video * 255).numpy().astype(np.uint8)
+            wandb_loss_dict[latent_key] = wandb.Video(
+                video, fps=16, format="mp4")  # change to 16 for Wan2.1
+            # Clean up references
+            del video, pixel_latent, latent
+
+        # Log to wandb
+        if self.global_rank == 0:
+            wandb.log(wandb_loss_dict, step=step)
+
+
+def main(args) -> None:
+    logger.info("Starting ODE-init training pipeline...")
+    pipeline = ODEInitTrainingPipeline.from_pretrained(
+        args.pretrained_model_name_or_path, args=args)
+    args = pipeline.training_args
+    pipeline.train()
+    logger.info("ODE-init training pipeline done")
+
+
+if __name__ == "__main__":
+    argv = sys.argv
+    from fastvideo.fastvideo_args import TrainingArgs
+    from fastvideo.utils import FlexibleArgumentParser
+    parser = FlexibleArgumentParser()
+    parser = TrainingArgs.add_cli_args(parser)
+    parser = FastVideoArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    args.dit_cpu_offload = False
+    main(args)

fastvideo/training/training_pipeline.py

@@ -1,5 +1,4 @@
 # SPDX-License-Identifier: Apache-2.0
-import gc
 import dataclasses
 import math
 import os
@@ -23,6 +22,7 @@ from tqdm.auto import tqdm
 import fastvideo.envs as envs
 from fastvideo.attention.backends.video_sparse_attn import (
     VideoSparseAttentionMetadataBuilder)
+from fastvideo.attention.backends.vmoba import VideoMobaAttentionMetadataBuilder
 from fastvideo.configs.sample import SamplingParam
 from fastvideo.dataset import build_parquet_map_style_dataloader
 from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
@@ -39,22 +39,20 @@ from fastvideo.training.activation_checkpoint import (
     apply_activation_checkpointing)
 from fastvideo.training.training_utils import (
     clip_grad_norm_while_handling_failing_dtensor_cases,
-    compute_density_for_timestep_sampling, get_scheduler, get_sigmas,
-    load_checkpoint, normalize_dit_input, save_checkpoint,
+    compute_density_for_timestep_sampling, count_trainable, get_scheduler,
+    get_sigmas, load_checkpoint, normalize_dit_input, save_checkpoint,
     shard_latents_across_sp)
-from fastvideo.utils import is_vsa_available, set_random_seed, shallow_asdict
+from fastvideo.utils import (is_vmoba_available, is_vsa_available,
+                             set_random_seed, shallow_asdict)
 
 import wandb  # isort: skip
 
 vsa_available = is_vsa_available()
+vmoba_available = is_vmoba_available()
 
 logger = init_logger(__name__)
 
 
-def _get_trainable_params(model: torch.nn.Module) -> int:
-    return sum(p.numel() for p in model.parameters() if p.requires_grad)
-
-
 class TrainingPipeline(LoRAPipeline, ABC):
     """
     A pipeline for training a model. All training pipelines should inherit from this class.
@@ -65,7 +63,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
     train_dataloader: StatefulDataLoader
     train_loader_iter: Iterator[dict[str, Any]]
     current_epoch: int = 0
-    train_transformer_2: bool = False
 
     def __init__(
             self,
@@ -101,7 +98,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
         self.sp_world_size = self.sp_group.world_size
         self.local_rank = world_group.local_rank
         self.transformer = self.get_module("transformer")
-        self.transformer_2 = self.get_module("transformer_2", None)
         self.seed = training_args.seed
         self.set_schemas()
 
@@ -114,26 +110,17 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 self.transformer,
                 checkpointing_type=training_args.
                 enable_gradient_checkpointing_type)
-            if self.transformer_2 is not None:
-                self.transformer_2 = apply_activation_checkpointing(
-                    self.transformer_2,
-                    checkpointing_type=training_args.
-                    enable_gradient_checkpointing_type)
 
         noise_scheduler = self.modules["scheduler"]
+        # Set grads for proper modules based on the training mode (Distill, LoRA, etc.)
         self.set_trainable()
         params_to_optimize = self.transformer.parameters()
-
         params_to_optimize = list(
             filter(lambda p: p.requires_grad, params_to_optimize))
-        # Parse betas from string format "beta1,beta2"
-        betas_str = training_args.betas
-        betas = tuple(float(x.strip()) for x in betas_str.split(","))
-        
         self.optimizer = torch.optim.AdamW(
             params_to_optimize,
             lr=training_args.learning_rate,
-            betas=betas,
+            betas=(0.9, 0.999),
             weight_decay=training_args.weight_decay,
             eps=1e-8,
         )
@@ -151,30 +138,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
             min_lr_ratio=training_args.min_lr_ratio,
             last_epoch=self.init_steps - 1,
         )
-        if self.transformer_2 is not None:
-            # Ensure transformer_2 has trainable parameters before creating optimizer
-            self.transformer_2.train()
-            self.transformer_2.requires_grad_(True)
-            params_to_optimize_2 = self.transformer_2.parameters()
-            params_to_optimize_2 = list(
-                filter(lambda p: p.requires_grad, params_to_optimize_2))
-            self.optimizer_2 = torch.optim.AdamW(
-                params_to_optimize_2,
-                lr=training_args.learning_rate,
-                betas=(0.9, 0.999),
-                weight_decay=training_args.weight_decay,
-                eps=1e-8,
-            )
-            self.lr_scheduler_2 = get_scheduler(
-                training_args.lr_scheduler,
-                optimizer=self.optimizer_2,
-                num_warmup_steps=training_args.lr_warmup_steps,
-                num_training_steps=training_args.max_train_steps,
-                num_cycles=training_args.lr_num_cycles,
-                power=training_args.lr_power,
-                min_lr_ratio=training_args.min_lr_ratio,
-                last_epoch=self.init_steps - 1,
-            )
 
         self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader(
             training_args.data_path,
@@ -189,7 +152,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
             seed=self.seed)
 
         self.noise_scheduler = noise_scheduler
-        self.boundary_timestep = self.training_args.boundary_ratio * self.noise_scheduler.num_train_timesteps
+
         self.num_update_steps_per_epoch = math.ceil(
             len(self.train_dataloader) /
             training_args.gradient_accumulation_steps * training_args.sp_size /
@@ -215,9 +178,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
     def _prepare_training(self, training_batch: TrainingBatch) -> TrainingBatch:
         self.transformer.train()
         self.optimizer.zero_grad()
-        if self.transformer_2 is not None:
-            self.transformer_2.train()
-            self.optimizer_2.zero_grad()
         training_batch.total_loss = 0.0
         return training_batch
 
@@ -264,8 +224,17 @@ class TrainingPipeline(LoRAPipeline, ABC):
                             generator=self.noise_gen_cuda,
                             device=latents.device,
                             dtype=latents.dtype)
-        timesteps = self._sample_timesteps(batch_size, latents.device)
-
+        u = compute_density_for_timestep_sampling(
+            weighting_scheme=self.training_args.weighting_scheme,
+            batch_size=batch_size,
+            generator=self.noise_random_generator,
+            logit_mean=self.training_args.logit_mean,
+            logit_std=self.training_args.logit_std,
+            mode_scale=self.training_args.mode_scale,
+        )
+        indices = (u * self.noise_scheduler.config.num_train_timesteps).long()
+        timesteps = self.noise_scheduler.timesteps[indices].to(
+            device=latents.device)
         if self.training_args.sp_size > 1:
             # Make sure that the timesteps are the same across all sp processes.
             sp_group = get_sp_group()
@@ -288,38 +257,6 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         return training_batch
 
-    def _sample_timesteps(self, batch_size, device):
-        # Determine which model to train based on the boundary timestep
-        if (self.transformer_2 is not None and self.boundary_timestep is not None and
-                torch.rand(1, generator=self.noise_random_generator).item() <= self.training_args.boundary_ratio):
-            self.train_transformer_2 = True
-        else:
-            self.train_transformer_2 = False
-
-        # Broadcast the decision to all processes
-        decision = torch.tensor(1.0 if self.train_transformer_2 else 0.0, device=self.device)
-        dist.broadcast(decision, src=0)
-        self.train_transformer_2 = decision.item() == 1.0
-
-        # Sample u from the appropriate range
-        u = compute_density_for_timestep_sampling(
-            weighting_scheme=self.training_args.weighting_scheme,
-            batch_size=batch_size,
-            generator=self.noise_random_generator,
-            logit_mean=self.training_args.logit_mean,
-            logit_std=self.training_args.logit_std,
-            mode_scale=self.training_args.mode_scale,
-        )
-
-        boundary_ratio = self.training_args.boundary_ratio
-        if self.train_transformer_2:
-            u = (1 - boundary_ratio) + u * boundary_ratio # min: 1 - boundary_ratio, max: 1
-        else:
-            u = u * (1 - boundary_ratio) # min: 0, max: 1 - boundary_ratio
-
-        indices = (u * self.noise_scheduler.config.num_train_timesteps).long()
-        return self.noise_scheduler.timesteps[indices].to(device=device)
-        
     def _build_attention_metadata(
             self, training_batch: TrainingBatch) -> TrainingBatch:
         latents_shape = training_batch.raw_latent_shape
@@ -335,6 +272,20 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 patch_size=patch_size,
                 VSA_sparsity=current_vsa_sparsity,
                 device=get_local_torch_device())
+        elif vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
+            moba_params = self.training_args.moba_config.copy()
+            moba_params.update({
+                "current_timestep":
+                training_batch.timesteps,
+                "raw_latent_shape":
+                training_batch.raw_latent_shape[2:5],
+                "patch_size":
+                self.training_args.pipeline_config.dit_config.patch_size,
+                "device":
+                get_local_torch_device(),
+            })
+            training_batch.attn_metadata = VideoMobaAttentionMetadataBuilder(
+            ).build(**moba_params)
         else:
             training_batch.attn_metadata = None
 
@@ -359,7 +310,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
     def _transformer_forward_and_compute_loss(
             self, training_batch: TrainingBatch) -> TrainingBatch:
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
+        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN" or vmoba_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VMOBA_ATTN":
             assert training_batch.attn_metadata is not None
         else:
             assert training_batch.attn_metadata is None
@@ -370,12 +321,11 @@ class TrainingPipeline(LoRAPipeline, ABC):
         #         [1000.0],
         #         device=training_batch.noisy_model_input.device,
         #         dtype=torch.bfloat16)
-        current_model = self.transformer_2 if self.train_transformer_2 else self.transformer
 
         with set_forward_context(
                 current_timestep=training_batch.current_timestep,
                 attn_metadata=training_batch.attn_metadata):
-            model_pred = current_model(**input_kwargs)
+            model_pred = self.transformer(**input_kwargs)
             if self.training_args.precondition_outputs:
                 assert training_batch.sigmas is not None
                 model_pred = training_batch.noisy_model_input - model_pred * training_batch.sigmas
@@ -406,12 +356,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
         # the following:
         # grad_norm = transformer.clip_grad_norm_(max_grad_norm)
         if max_grad_norm is not None:
-            # Only clip gradients for the model that is currently training
-            if self.train_transformer_2 and self.transformer_2 is not None:
-                model_parts = [self.transformer_2]
-            else:
-                model_parts = [self.transformer]
-            
+            model_parts = [self.transformer]
             grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
                 [p for m in model_parts for p in m.parameters()],
                 max_grad_norm,
@@ -456,14 +401,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         training_batch = self._clip_grad_norm(training_batch)
 
-        # Only step the optimizer and scheduler for the model that is currently training
-        if self.train_transformer_2 and self.transformer_2 is not None:
-            self.optimizer_2.step()
-            self.lr_scheduler_2.step()
-        else:
-            self.optimizer.step()
-            self.lr_scheduler.step()
-            
+        self.optimizer.step()
+        self.lr_scheduler.step()
+
         training_batch.total_loss = training_batch.total_loss
         training_batch.grad_norm = training_batch.grad_norm
         return training_batch
@@ -491,15 +431,10 @@ class TrainingPipeline(LoRAPipeline, ABC):
                     local_main_process_only=False)
         if not self.post_init_called:
             self.post_init()
-        num_trainable_params = _get_trainable_params(self.transformer)
+        num_trainable_params = count_trainable(self.transformer)
         logger.info("Starting training with %s B trainable parameters",
                     round(num_trainable_params / 1e9, 3))
 
-        if getattr(self, "transformer_2", None) is not None:
-            num_trainable_params = _get_trainable_params(self.transformer_2)
-            logger.info("Transformer 2: Starting training with %s B trainable parameters",
-                        round(num_trainable_params / 1e9, 3))
-
         # Set random seeds for deterministic training
         self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
             self.seed)
@@ -520,7 +455,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         self._log_training_info()
 
-        self._log_validation(self.training_args,
+        self._log_validation(self.transformer, self.training_args,
                              self.init_steps)
 
         # Train!
@@ -541,6 +476,9 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 current_decay_times = min(step // vsa_decay_interval_steps,
                                           vsa_sparsity // vsa_decay_rate)
                 current_vsa_sparsity = current_decay_times * vsa_decay_rate
+            elif vmoba_available:
+                # TODO: add vmoba sparsity scheduling here
+                current_vsa_sparsity = 0.0
             else:
                 current_vsa_sparsity = 0.0
 
@@ -582,10 +520,14 @@ class TrainingPipeline(LoRAPipeline, ABC):
                 self.transformer.train()
                 self.sp_group.barrier()
             if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
-                self._log_validation(self.training_args, step)
+                if self.training_args.log_visualization:
+                    self.visualize_intermediate_latents(training_batch,
+                                                        self.training_args,
+                                                        step)
+                self._log_validation(self.transformer, self.training_args, step)
                 gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
                 trainable_params = round(
-                    _get_trainable_params(self.transformer) / 1e9, 3)
+                    count_trainable(self.transformer) / 1e9, 3)
                 logger.info(
                     "GPU memory usage after validation: %s MB, trainable params: %sB",
                     gpu_memory_usage, trainable_params)
@@ -621,7 +563,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
         logger.info("  Total optimization steps = %s",
                     self.training_args.max_train_steps)
         logger.info("  Total training parameters per FSDP shard = %s B",
-                    round(_get_trainable_params(self.transformer) / 1e9, 3))
+                    round(count_trainable(self.transformer) / 1e9, 3))
         # print dtype
         logger.info("  Master weight dtype: %s",
                     self.transformer.parameters().__next__().dtype)
@@ -663,12 +605,12 @@ class TrainingPipeline(LoRAPipeline, ABC):
         return batch
 
     @torch.no_grad()
-    def _log_validation(self, training_args, global_step) -> None:
+    def _log_validation(self, transformer, training_args, global_step) -> None:
         """
         Generate a validation video and log it to wandb to check the quality during training.
         """
         training_args.inference_mode = True
-        training_args.dit_cpu_offload = False
+        training_args.dit_cpu_offload = True
         if not training_args.log_validation:
             return
         if self.validation_pipeline is None:
@@ -689,10 +631,7 @@ class TrainingPipeline(LoRAPipeline, ABC):
         validation_dataloader = DataLoader(validation_dataset,
                                            batch_size=None,
                                            num_workers=0)
-
-        self.transformer.eval()
-        if getattr(self, "transformer_2", None) is not None:
-            self.transformer_2.eval()
+        transformer.eval()
 
         validation_steps = training_args.validation_sampling_steps.split(",")
         validation_steps = [int(step) for step in validation_steps]
@@ -784,6 +723,11 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
         # Re-enable gradients for training
         training_args.inference_mode = False
-        self.transformer.train()
-        if getattr(self, "transformer_2", None) is not None:
-            self.transformer_2.train()
+        transformer.train()
+
+    def visualize_intermediate_latents(self, training_batch: TrainingBatch,
+                                       training_args: TrainingArgs, step: int):
+        """Add visualization data to wandb logging and save frames to disk."""
+        raise NotImplementedError(
+            "Visualize intermediate latents is not implemented for training pipeline"
+        )

fastvideo/training/training_pipeline_switch.py

@@ -1,797 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import dataclasses
-import math
-import os
-import time
-from abc import ABC, abstractmethod
-from collections import deque
-from collections.abc import Iterator
-from typing import Any
-
-import imageio
-import numpy as np
-import torch
-import torch.distributed as dist
-import torchvision
-from diffusers import FlowMatchEulerDiscreteScheduler
-from einops import rearrange
-from torch.utils.data import DataLoader
-from torchdata.stateful_dataloader import StatefulDataLoader
-from tqdm.auto import tqdm
-
-import fastvideo.envs as envs
-from fastvideo.attention.backends.video_sparse_attn import (
-    VideoSparseAttentionMetadataBuilder)
-from fastvideo.configs.sample import SamplingParam
-from fastvideo.dataset import build_parquet_map_style_dataloader
-from fastvideo.dataset.dataloader.schema import pyarrow_schema_t2v
-from fastvideo.dataset.validation_dataset import ValidationDataset
-from fastvideo.distributed import (cleanup_dist_env_and_memory,
-                                   get_local_torch_device, get_sp_group,
-                                   get_world_group)
-from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
-from fastvideo.forward_context import set_forward_context
-from fastvideo.logger import init_logger
-from fastvideo.pipelines import (ComposedPipelineBase, ForwardBatch,
-                                 LoRAPipeline, TrainingBatch)
-from fastvideo.training.activation_checkpoint import (
-    apply_activation_checkpointing)
-from fastvideo.training.training_utils import (
-    clip_grad_norm_while_handling_failing_dtensor_cases,
-    compute_density_for_timestep_sampling, get_scheduler, get_sigmas,
-    load_checkpoint, normalize_dit_input, save_checkpoint,
-    shard_latents_across_sp)
-from fastvideo.utils import is_vsa_available, set_random_seed, shallow_asdict
-
-import wandb  # isort: skip
-
-vsa_available = is_vsa_available()
-
-logger = init_logger(__name__)
-
-
-def _get_trainable_params(model: torch.nn.Module) -> int:
-    return sum(p.numel() for p in model.parameters() if p.requires_grad)
-
-
-class TrainingPipeline(LoRAPipeline, ABC):
-    """
-    A pipeline for training a model. All training pipelines should inherit from this class.
-    All reusable components and code should be implemented in this class.
-    """
-    _required_config_modules = ["scheduler", "transformer"]
-    validation_pipeline: ComposedPipelineBase
-    train_dataloader: StatefulDataLoader
-    train_loader_iter: Iterator[dict[str, Any]]
-    current_epoch: int = 0
-    train_transformer_2: bool = False
-
-    def __init__(
-            self,
-            model_path: str,
-            fastvideo_args: TrainingArgs,
-            required_config_modules: list[str] | None = None,
-            loaded_modules: dict[str, torch.nn.Module] | None = None) -> None:
-        fastvideo_args.inference_mode = False
-        self.lora_training = fastvideo_args.lora_training
-        if self.lora_training and fastvideo_args.lora_rank is None:
-            raise ValueError("lora rank must be set when using lora training")
-
-        set_random_seed(fastvideo_args.seed)  # for lora param init
-        super().__init__(model_path, fastvideo_args, required_config_modules,
-                         loaded_modules)  # type: ignore
-
-    def create_pipeline_stages(self, fastvideo_args: FastVideoArgs):
-        raise RuntimeError(
-            "create_pipeline_stages should not be called for training pipeline")
-
-    def set_schemas(self) -> None:
-        self.train_dataset_schema = pyarrow_schema_t2v
-
-    def initialize_training_pipeline(self, training_args: TrainingArgs):
-        logger.info("Initializing training pipeline...")
-        self.device = get_local_torch_device()
-        self.training_args = training_args
-        world_group = get_world_group()
-        self.world_size = world_group.world_size
-        self.global_rank = world_group.rank
-        self.sp_group = get_sp_group()
-        self.rank_in_sp_group = self.sp_group.rank_in_group
-        self.sp_world_size = self.sp_group.world_size
-        self.local_rank = world_group.local_rank
-        self.transformer = self.get_module("transformer")
-        self.transformer_2 = self.get_module("transformer_2", None)
-        self.seed = training_args.seed
-        self.set_schemas()
-
-        # Set random seeds for deterministic training
-        assert self.seed is not None, "seed must be set"
-        set_random_seed(self.seed)
-        self.transformer.train()
-        if training_args.enable_gradient_checkpointing_type is not None:
-            self.transformer = apply_activation_checkpointing(
-                self.transformer,
-                checkpointing_type=training_args.
-                enable_gradient_checkpointing_type)
-            if self.transformer_2 is not None:
-                self.transformer_2 = apply_activation_checkpointing(
-                    self.transformer_2,
-                    checkpointing_type=training_args.
-                    enable_gradient_checkpointing_type)
-
-        noise_scheduler = self.modules["scheduler"]
-        self.set_trainable()
-        params_to_optimize = self.transformer.parameters()
-
-        params_to_optimize = list(
-            filter(lambda p: p.requires_grad, params_to_optimize))
-        self.optimizer = torch.optim.AdamW(
-            params_to_optimize,
-            lr=training_args.learning_rate,
-            betas=(0.9, 0.999),
-            weight_decay=training_args.weight_decay,
-            eps=1e-8,
-        )
-
-        self.init_steps = 0
-        logger.info("optimizer: %s", self.optimizer)
-
-        self.lr_scheduler = get_scheduler(
-            training_args.lr_scheduler,
-            optimizer=self.optimizer,
-            num_warmup_steps=training_args.lr_warmup_steps,
-            num_training_steps=training_args.max_train_steps,
-            num_cycles=training_args.lr_num_cycles,
-            power=training_args.lr_power,
-            min_lr_ratio=training_args.min_lr_ratio,
-            last_epoch=self.init_steps - 1,
-        )
-        if self.transformer_2 is not None:
-            # Ensure transformer_2 has trainable parameters before creating optimizer
-            self.transformer_2.train()
-            self.transformer_2.requires_grad_(True)
-            params_to_optimize_2 = self.transformer_2.parameters()
-            params_to_optimize_2 = list(
-                filter(lambda p: p.requires_grad, params_to_optimize_2))
-            self.optimizer_2 = torch.optim.AdamW(
-                params_to_optimize_2,
-                lr=training_args.learning_rate,
-                betas=(0.9, 0.999),
-                weight_decay=training_args.weight_decay,
-                eps=1e-8,
-            )
-            self.lr_scheduler_2 = get_scheduler(
-                training_args.lr_scheduler,
-                optimizer=self.optimizer_2,
-                num_warmup_steps=training_args.lr_warmup_steps,
-                num_training_steps=training_args.max_train_steps,
-                num_cycles=training_args.lr_num_cycles,
-                power=training_args.lr_power,
-                min_lr_ratio=training_args.min_lr_ratio,
-                last_epoch=self.init_steps - 1,
-            )
-
-        self.train_dataset, self.train_dataloader = build_parquet_map_style_dataloader(
-            training_args.data_path,
-            training_args.train_batch_size,
-            parquet_schema=self.train_dataset_schema,
-            num_data_workers=training_args.dataloader_num_workers,
-            cfg_rate=training_args.training_cfg_rate,
-            drop_last=True,
-            text_padding_length=training_args.pipeline_config.
-            text_encoder_configs[0].arch_config.
-            text_len,  # type: ignore[attr-defined]
-            seed=self.seed)
-
-        self.noise_scheduler = noise_scheduler
-        self.boundary_timestep = self.training_args.boundary_ratio * self.noise_scheduler.num_train_timesteps
-        self.num_update_steps_per_epoch = math.ceil(
-            len(self.train_dataloader) /
-            training_args.gradient_accumulation_steps * training_args.sp_size /
-            training_args.train_sp_batch_size)
-        self.num_train_epochs = math.ceil(training_args.max_train_steps /
-                                          self.num_update_steps_per_epoch)
-
-        # TODO(will): is there a cleaner way to track epochs?
-        self.current_epoch = 0
-
-        if self.global_rank == 0:
-            project = training_args.tracker_project_name or "fastvideo"
-            wandb_config = dataclasses.asdict(training_args)
-            wandb.init(project=project,
-                       config=wandb_config,
-                       name=training_args.wandb_run_name)
-
-    @abstractmethod
-    def initialize_validation_pipeline(self, training_args: TrainingArgs):
-        raise NotImplementedError(
-            "Training pipelines must implement this method")
-
-    def _prepare_training(self, training_batch: TrainingBatch) -> TrainingBatch:
-        # At the beginning, disable training for all models
-        self._disable_training(self.transformer, self.optimizer)
-        if self.transformer_2 is not None:
-            self._disable_training(self.transformer_2, self.optimizer_2)
-        
-        training_batch.total_loss = 0.0
-        return training_batch
-
-    def _enable_training(self, model: torch.nn.Module, optimizer: torch.optim.Optimizer) -> None:
-        """Enable training mode and gradients for the specified model."""
-        for param in model.parameters():
-            param.requires_grad = True
-        model.train()
-        optimizer.zero_grad()
-
-    def _disable_training(self, model: torch.nn.Module, optimizer: torch.optim.Optimizer) -> None:
-        """Disable training mode and gradients for the specified model."""
-        for param in model.parameters():
-            param.requires_grad = False
-        optimizer.zero_grad(set_to_none=True)
-
-    def _get_next_batch(self, training_batch: TrainingBatch) -> TrainingBatch:
-        batch = next(self.train_loader_iter, None)  # type: ignore
-        if batch is None:
-            self.current_epoch += 1
-            logger.info("Starting epoch %s", self.current_epoch)
-            # Reset iterator for next epoch
-            self.train_loader_iter = iter(self.train_dataloader)
-            # Get first batch of new epoch
-            batch = next(self.train_loader_iter)
-
-        # latents, encoder_hidden_states, encoder_attention_mask, infos = batch
-        latents = batch['vae_latent']
-        latents = latents[:, :, :self.training_args.num_latent_t]
-        encoder_hidden_states = batch['text_embedding']
-        encoder_attention_mask = batch['text_attention_mask']
-        infos = batch['info_list']
-
-        training_batch.latents = latents.to(get_local_torch_device(),
-                                            dtype=torch.bfloat16)
-        training_batch.encoder_hidden_states = encoder_hidden_states.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        training_batch.encoder_attention_mask = encoder_attention_mask.to(
-            get_local_torch_device(), dtype=torch.bfloat16)
-        training_batch.infos = infos
-
-        return training_batch
-
-    def _normalize_dit_input(self,
-                             training_batch: TrainingBatch) -> TrainingBatch:
-        # TODO(will): support other models
-        training_batch.latents = normalize_dit_input('wan',
-                                                     training_batch.latents,
-                                                     self.get_module("vae"))
-        return training_batch
-
-    def _prepare_dit_inputs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-        latents = training_batch.latents
-        batch_size = latents.shape[0]
-        noise = torch.randn(latents.shape,
-                            generator=self.noise_gen_cuda,
-                            device=latents.device,
-                            dtype=latents.dtype)
-        timesteps = self._sample_timesteps(batch_size, latents.device)
-
-        # Enable training for the model that will be trained next and disable the other
-        if self.train_transformer_2:
-            self._enable_training(self.transformer_2, self.optimizer_2)
-            self._disable_training(self.transformer, self.optimizer)
-        else:
-            self._enable_training(self.transformer, self.optimizer)
-            if self.transformer_2 is not None:
-                self._disable_training(self.transformer_2, self.optimizer_2)
-
-        if self.training_args.sp_size > 1:
-            # Make sure that the timesteps are the same across all sp processes.
-            sp_group = get_sp_group()
-            sp_group.broadcast(timesteps, src=0)
-        sigmas = get_sigmas(
-            self.noise_scheduler,
-            latents.device,
-            timesteps,
-            n_dim=latents.ndim,
-            dtype=latents.dtype,
-        )
-        noisy_model_input = (1.0 -
-                             sigmas) * training_batch.latents + sigmas * noise
-
-        training_batch.noisy_model_input = noisy_model_input
-        training_batch.timesteps = timesteps
-        training_batch.sigmas = sigmas
-        training_batch.noise = noise
-        training_batch.raw_latent_shape = training_batch.latents.shape
-
-        return training_batch
-
-    def _sample_timesteps(self, batch_size, device):
-        # Determine which model to train based on the boundary timestep
-        if (self.transformer_2 is not None and self.boundary_timestep is not None and
-                torch.rand(1, generator=self.noise_random_generator).item() > self.training_args.boundary_ratio):
-            self.train_transformer_2 = True
-        else:
-            self.train_transformer_2 = False
-
-        # Broadcast the decision to all processes
-        decision = torch.tensor(1.0 if self.train_transformer_2 else 0.0, device=self.device)
-        dist.broadcast(decision, src=0)
-        self.train_transformer_2 = decision.item() == 1.0
-
-        # Sample u from the appropriate range
-        u = compute_density_for_timestep_sampling(
-            weighting_scheme=self.training_args.weighting_scheme,
-            batch_size=batch_size,
-            generator=self.noise_random_generator,
-            logit_mean=self.training_args.logit_mean,
-            logit_std=self.training_args.logit_std,
-            mode_scale=self.training_args.mode_scale,
-        )
-
-        boundary_ratio = self.training_args.boundary_ratio
-        if self.train_transformer_2:
-            u = boundary_ratio + u * (1.0 - boundary_ratio)
-        else:
-            u = u * boundary_ratio
-
-        indices = (u * self.noise_scheduler.config.num_train_timesteps).long()
-        return self.noise_scheduler.timesteps[indices].to(device=device)
-        
-    def _build_attention_metadata(
-            self, training_batch: TrainingBatch) -> TrainingBatch:
-        latents_shape = training_batch.raw_latent_shape
-        patch_size = self.training_args.pipeline_config.dit_config.patch_size
-        current_vsa_sparsity = training_batch.current_vsa_sparsity
-        assert latents_shape is not None
-        assert training_batch.timesteps is not None
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
-            training_batch.attn_metadata = VideoSparseAttentionMetadataBuilder(  # type: ignore
-            ).build(  # type: ignore
-                raw_latent_shape=latents_shape[2:5],
-                current_timestep=training_batch.timesteps,
-                patch_size=patch_size,
-                VSA_sparsity=current_vsa_sparsity,
-                device=get_local_torch_device())
-        else:
-            training_batch.attn_metadata = None
-
-        return training_batch
-
-    def _build_input_kwargs(self,
-                            training_batch: TrainingBatch) -> TrainingBatch:
-        training_batch.input_kwargs = {
-            "hidden_states":
-            training_batch.noisy_model_input,
-            "encoder_hidden_states":
-            training_batch.encoder_hidden_states,
-            "timestep":
-            training_batch.timesteps.to(get_local_torch_device(),
-                                        dtype=torch.bfloat16),
-            "encoder_attention_mask":
-            training_batch.encoder_attention_mask,
-            "return_dict":
-            False,
-        }
-        return training_batch
-
-    def _transformer_forward_and_compute_loss(
-            self, training_batch: TrainingBatch) -> TrainingBatch:
-        if vsa_available and envs.FASTVIDEO_ATTENTION_BACKEND == "VIDEO_SPARSE_ATTN":
-            assert training_batch.attn_metadata is not None
-        else:
-            assert training_batch.attn_metadata is None
-        input_kwargs = training_batch.input_kwargs
-
-        # if 'hunyuan' in self.training_args.model_type:
-        #     input_kwargs["guidance"] = torch.tensor(
-        #         [1000.0],
-        #         device=training_batch.noisy_model_input.device,
-        #         dtype=torch.bfloat16)
-        current_model = self.transformer_2 if self.train_transformer_2 else self.transformer
-
-        with set_forward_context(
-                current_timestep=training_batch.current_timestep,
-                attn_metadata=training_batch.attn_metadata):
-            model_pred = current_model(**input_kwargs)
-            if self.training_args.precondition_outputs:
-                assert training_batch.sigmas is not None
-                model_pred = training_batch.noisy_model_input - model_pred * training_batch.sigmas
-            assert training_batch.latents is not None
-            assert training_batch.noise is not None
-            target = training_batch.latents if self.training_args.precondition_outputs else training_batch.noise - training_batch.latents
-
-            # make sure no implicit broadcasting happens
-            assert model_pred.shape == target.shape, f"model_pred.shape: {model_pred.shape}, target.shape: {target.shape}"
-            loss = (torch.mean((model_pred.float() - target.float())**2) /
-                    self.training_args.gradient_accumulation_steps)
-
-            loss.backward()
-            avg_loss = loss.detach().clone()
-
-        # logger.info(f"rank: {self.rank}, avg_loss: {avg_loss.item()}",
-        #             local_main_process_only=False)
-        world_group = get_world_group()
-        world_group.all_reduce(avg_loss, op=dist.ReduceOp.AVG)
-        training_batch.total_loss += avg_loss.item()
-
-        return training_batch
-
-    def _clip_grad_norm(self, training_batch: TrainingBatch) -> TrainingBatch:
-        max_grad_norm = self.training_args.max_grad_norm
-
-        # TODO(will): perhaps move this into transformer api so that we can do
-        # the following:
-        # grad_norm = transformer.clip_grad_norm_(max_grad_norm)
-        if max_grad_norm is not None:
-            # Only clip gradients for the model that is currently training
-            if self.train_transformer_2 and self.transformer_2 is not None:
-                model_parts = [self.transformer_2]
-            else:
-                model_parts = [self.transformer]
-            
-            grad_norm = clip_grad_norm_while_handling_failing_dtensor_cases(
-                [p for m in model_parts for p in m.parameters()],
-                max_grad_norm,
-                foreach=None,
-            )
-            assert grad_norm is not float('nan') or grad_norm is not float(
-                'inf')
-            grad_norm = grad_norm.item() if grad_norm is not None else 0.0
-        else:
-            grad_norm = 0.0
-        training_batch.grad_norm = grad_norm
-        return training_batch
-
-    def train_one_step(self, training_batch: TrainingBatch) -> TrainingBatch:
-        training_batch = self._prepare_training(training_batch)
-
-        for _ in range(self.training_args.gradient_accumulation_steps):
-            training_batch = self._get_next_batch(training_batch)
-
-            # Normalize DIT input
-            training_batch = self._normalize_dit_input(training_batch)
-            # Create noisy model input
-            training_batch = self._prepare_dit_inputs(training_batch)
-
-            # Shard latents across sp groups
-            training_batch.latents = shard_latents_across_sp(
-                training_batch.latents,
-                num_latent_t=self.training_args.num_latent_t)
-            # shard noisy_model_input to match
-            training_batch.noisy_model_input = shard_latents_across_sp(
-                training_batch.noisy_model_input,
-                num_latent_t=self.training_args.num_latent_t)
-            # shard noise to match latents
-            training_batch.noise = shard_latents_across_sp(
-                training_batch.noise,
-                num_latent_t=self.training_args.num_latent_t)
-
-            training_batch = self._build_attention_metadata(training_batch)
-            training_batch = self._build_input_kwargs(training_batch)
-            training_batch = self._transformer_forward_and_compute_loss(
-                training_batch)
-
-        training_batch = self._clip_grad_norm(training_batch)
-
-        # Only step the optimizer and scheduler for the model that is currently training
-        if self.train_transformer_2 and self.transformer_2 is not None:
-            self.optimizer_2.step()
-            self.lr_scheduler_2.step()
-        else:
-            self.optimizer.step()
-            self.lr_scheduler.step()
-            
-        training_batch.total_loss = training_batch.total_loss
-        training_batch.grad_norm = training_batch.grad_norm
-        return training_batch
-
-    def _resume_from_checkpoint(self) -> None:
-        logger.info("Loading checkpoint from %s",
-                    self.training_args.resume_from_checkpoint)
-        resumed_step = load_checkpoint(
-            self.transformer, self.global_rank,
-            self.training_args.resume_from_checkpoint, self.optimizer,
-            self.train_dataloader, self.lr_scheduler,
-            self.noise_random_generator)
-        if resumed_step > 0:
-            self.init_steps = resumed_step
-            logger.info("Successfully resumed from step %s", resumed_step)
-        else:
-            logger.warning("Failed to load checkpoint, starting from step 0")
-            self.init_steps = 0
-
-    def train(self) -> None:
-        assert self.seed is not None, "seed must be set"
-        set_random_seed(self.seed + self.global_rank)
-        logger.info('rank: %s: start training',
-                    self.global_rank,
-                    local_main_process_only=False)
-        if not self.post_init_called:
-            self.post_init()
-        num_trainable_params = _get_trainable_params(self.transformer)
-        logger.info("Starting training with %s B trainable parameters",
-                    round(num_trainable_params / 1e9, 3))
-
-        # Set random seeds for deterministic training
-        self.noise_random_generator = torch.Generator(device="cpu").manual_seed(
-            self.seed)
-        self.noise_gen_cuda = torch.Generator(device="cuda").manual_seed(
-            self.seed)
-        self.validation_random_generator = torch.Generator(
-            device="cpu").manual_seed(self.seed)
-        logger.info("Initialized random seeds with seed: %s", self.seed)
-
-        self.noise_scheduler = FlowMatchEulerDiscreteScheduler()
-
-        if self.training_args.resume_from_checkpoint:
-            self._resume_from_checkpoint()
-
-        self.train_loader_iter = iter(self.train_dataloader)
-
-        step_times: deque[float] = deque(maxlen=100)
-
-        self._log_training_info()
-
-        self._log_validation(self.transformer, self.training_args,
-                             self.init_steps)
-
-        # Train!
-        progress_bar = tqdm(
-            range(0, self.training_args.max_train_steps),
-            initial=self.init_steps,
-            desc="Steps",
-            # Only show the progress bar once on each machine.
-            disable=self.local_rank > 0,
-        )
-        for step in range(self.init_steps + 1,
-                          self.training_args.max_train_steps + 1):
-            start_time = time.perf_counter()
-            if vsa_available:
-                vsa_sparsity = self.training_args.VSA_sparsity
-                vsa_decay_rate = self.training_args.VSA_decay_rate
-                vsa_decay_interval_steps = self.training_args.VSA_decay_interval_steps
-                current_decay_times = min(step // vsa_decay_interval_steps,
-                                          vsa_sparsity // vsa_decay_rate)
-                current_vsa_sparsity = current_decay_times * vsa_decay_rate
-            else:
-                current_vsa_sparsity = 0.0
-
-            training_batch = TrainingBatch()
-            training_batch.current_timestep = step
-            training_batch.current_vsa_sparsity = current_vsa_sparsity
-            training_batch = self.train_one_step(training_batch)
-
-            loss = training_batch.total_loss
-            grad_norm = training_batch.grad_norm
-
-            step_time = time.perf_counter() - start_time
-            step_times.append(step_time)
-            avg_step_time = sum(step_times) / len(step_times)
-
-            progress_bar.set_postfix({
-                "loss": f"{loss:.4f}",
-                "step_time": f"{step_time:.2f}s",
-                "grad_norm": grad_norm,
-            })
-            progress_bar.update(1)
-            if self.global_rank == 0:
-                wandb.log(
-                    {
-                        "train_loss": loss,
-                        "learning_rate": self.lr_scheduler.get_last_lr()[0],
-                        "step_time": step_time,
-                        "avg_step_time": avg_step_time,
-                        "grad_norm": grad_norm,
-                        "vsa_sparsity": current_vsa_sparsity,
-                    },
-                    step=step,
-                )
-            if step % self.training_args.checkpointing_steps == 0:
-                save_checkpoint(self.transformer, self.global_rank,
-                                self.training_args.output_dir, step,
-                                self.optimizer, self.train_dataloader,
-                                self.lr_scheduler, self.noise_random_generator)
-                self.transformer.train()
-                self.sp_group.barrier()
-            if self.training_args.log_validation and step % self.training_args.validation_steps == 0:
-                self._log_validation(self.transformer, self.training_args, step)
-                gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
-                trainable_params = round(
-                    _get_trainable_params(self.transformer) / 1e9, 3)
-                logger.info(
-                    "GPU memory usage after validation: %s MB, trainable params: %sB",
-                    gpu_memory_usage, trainable_params)
-
-        wandb.finish()
-        save_checkpoint(self.transformer, self.global_rank,
-                        self.training_args.output_dir,
-                        self.training_args.max_train_steps, self.optimizer,
-                        self.train_dataloader, self.lr_scheduler,
-                        self.noise_random_generator)
-
-        if get_sp_group():
-            cleanup_dist_env_and_memory()
-
-    def _log_training_info(self) -> None:
-        total_batch_size = (self.world_size *
-                            self.training_args.gradient_accumulation_steps /
-                            self.training_args.sp_size *
-                            self.training_args.train_sp_batch_size)
-        logger.info("***** Running training *****")
-        logger.info("  Num examples = %s", len(self.train_dataset))
-        logger.info("  Dataloader size = %s", len(self.train_dataloader))
-        logger.info("  Num Epochs = %s", self.num_train_epochs)
-        logger.info("  Resume training from step %s",
-                    self.init_steps)  # type: ignore
-        logger.info("  Instantaneous batch size per device = %s",
-                    self.training_args.train_batch_size)
-        logger.info(
-            "  Total train batch size (w. data & sequence parallel, accumulation) = %s",
-            total_batch_size)
-        logger.info("  Gradient Accumulation steps = %s",
-                    self.training_args.gradient_accumulation_steps)
-        logger.info("  Total optimization steps = %s",
-                    self.training_args.max_train_steps)
-        logger.info("  Total training parameters per FSDP shard = %s B",
-                    round(_get_trainable_params(self.transformer) / 1e9, 3))
-        # print dtype
-        logger.info("  Master weight dtype: %s",
-                    self.transformer.parameters().__next__().dtype)
-
-        gpu_memory_usage = torch.cuda.memory_allocated() / 1024**2
-        logger.info("GPU memory usage before train_one_step: %s MB",
-                    gpu_memory_usage)
-        logger.info("VSA validation sparsity: %s",
-                    self.training_args.VSA_sparsity)
-
-    def _prepare_validation_batch(self, sampling_param: SamplingParam,
-                                  training_args: TrainingArgs,
-                                  validation_batch: dict[str, Any],
-                                  num_inference_steps: int) -> ForwardBatch:
-        sampling_param.prompt = validation_batch['prompt']
-        sampling_param.height = training_args.num_height
-        sampling_param.width = training_args.num_width
-        sampling_param.num_inference_steps = num_inference_steps
-        sampling_param.data_type = "video"
-        assert self.seed is not None
-        sampling_param.seed = self.seed
-
-        latents_size = [(sampling_param.num_frames - 1) // 4 + 1,
-                        sampling_param.height // 8, sampling_param.width // 8]
-        n_tokens = latents_size[0] * latents_size[1] * latents_size[2]
-        temporal_compression_factor = training_args.pipeline_config.vae_config.arch_config.temporal_compression_ratio
-        num_frames = (training_args.num_latent_t -
-                      1) * temporal_compression_factor + 1
-        sampling_param.num_frames = num_frames
-        batch = ForwardBatch(
-            **shallow_asdict(sampling_param),
-            latents=None,
-            generator=self.validation_random_generator,
-            n_tokens=n_tokens,
-            eta=0.0,
-            VSA_sparsity=training_args.VSA_sparsity,
-        )
-
-        return batch
-
-    @torch.no_grad()
-    def _log_validation(self, transformer, training_args, global_step) -> None:
-        """
-        Generate a validation video and log it to wandb to check the quality during training.
-        """
-        training_args.inference_mode = True
-        training_args.dit_cpu_offload = True
-        if not training_args.log_validation:
-            return
-        if self.validation_pipeline is None:
-            raise ValueError("Validation pipeline is not set")
-
-        logger.info("Starting validation")
-
-        # Create sampling parameters if not provided
-        sampling_param = SamplingParam.from_pretrained(training_args.model_path)
-
-        # Prepare validation prompts
-        logger.info('rank: %s: fastvideo_args.validation_dataset_file: %s',
-                    self.global_rank,
-                    training_args.validation_dataset_file,
-                    local_main_process_only=False)
-        validation_dataset = ValidationDataset(
-            training_args.validation_dataset_file)
-        validation_dataloader = DataLoader(validation_dataset,
-                                           batch_size=None,
-                                           num_workers=0)
-
-        transformer.eval()
-
-        validation_steps = training_args.validation_sampling_steps.split(",")
-        validation_steps = [int(step) for step in validation_steps]
-        validation_steps = [step for step in validation_steps if step > 0]
-        # Log validation results for this step
-        world_group = get_world_group()
-        num_sp_groups = world_group.world_size // self.sp_group.world_size
-
-        # Process each validation prompt for each validation step
-        for num_inference_steps in validation_steps:
-            logger.info("rank: %s: num_inference_steps: %s",
-                        self.global_rank,
-                        num_inference_steps,
-                        local_main_process_only=False)
-            step_videos: list[np.ndarray] = []
-            step_captions: list[str] = []
-
-            for validation_batch in validation_dataloader:
-                batch = self._prepare_validation_batch(sampling_param,
-                                                       training_args,
-                                                       validation_batch,
-                                                       num_inference_steps)
-                logger.info("rank: %s: rank_in_sp_group: %s, batch.prompt: %s",
-                            self.global_rank,
-                            self.rank_in_sp_group,
-                            batch.prompt,
-                            local_main_process_only=False)
-
-                assert batch.prompt is not None and isinstance(
-                    batch.prompt, str)
-                step_captions.append(batch.prompt)
-
-                # Run validation inference
-                output_batch = self.validation_pipeline.forward(
-                    batch, training_args)
-                samples = output_batch.output
-
-                if self.rank_in_sp_group != 0:
-                    continue
-
-                # Process outputs
-                video = rearrange(samples, "b c t h w -> t b c h w")
-                frames = []
-                for x in video:
-                    x = torchvision.utils.make_grid(x, nrow=6)
-                    x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
-                    frames.append((x * 255).numpy().astype(np.uint8))
-                step_videos.append(frames)
-
-            # Only sp_group leaders (rank_in_sp_group == 0) need to send their
-            # results to global rank 0
-            if self.rank_in_sp_group == 0:
-                if self.global_rank == 0:
-                    # Global rank 0 collects results from all sp_group leaders
-                    all_videos = step_videos  # Start with own results
-                    all_captions = step_captions
-
-                    # Receive from other sp_group leaders
-                    for sp_group_idx in range(1, num_sp_groups):
-                        src_rank = sp_group_idx * self.sp_world_size  # Global rank of other sp_group leaders
-                        recv_videos = world_group.recv_object(src=src_rank)
-                        recv_captions = world_group.recv_object(src=src_rank)
-                        all_videos.extend(recv_videos)
-                        all_captions.extend(recv_captions)
-
-                    video_filenames = []
-                    for i, (video, caption) in enumerate(
-                            zip(all_videos, all_captions, strict=True)):
-                        os.makedirs(training_args.output_dir, exist_ok=True)
-                        filename = os.path.join(
-                            training_args.output_dir,
-                            f"validation_step_{global_step}_inference_steps_{num_inference_steps}_video_{i}.mp4"
-                        )
-                        imageio.mimsave(filename, video, fps=sampling_param.fps)
-                        video_filenames.append(filename)
-
-                    logs = {
-                        f"validation_videos_{num_inference_steps}_steps": [
-                            wandb.Video(filename, caption=caption)
-                            for filename, caption in zip(
-                                video_filenames, all_captions, strict=True)
-                        ]
-                    }
-                    wandb.log(logs, step=global_step)
-                else:
-                    # Other sp_group leaders send their results to global rank 0
-                    world_group.send_object(step_videos, dst=0)
-                    world_group.send_object(step_captions, dst=0)
-
-        # Re-enable gradients for training
-        training_args.inference_mode = False
-        transformer.train()

fastvideo/training/training_utils.py

@@ -466,11 +466,13 @@ def load_distillation_checkpoint(generator_transformer,
             ema_state = generator_states.get("ema")
             if ema_state is not None:
                 generator_ema.load_state_dict(ema_state)
-                logger.info("rank: %s, generator EMA state loaded successfully", rank)
+                logger.info("rank: %s, generator EMA state loaded successfully",
+                            rank)
             else:
                 logger.info("rank: %s, no EMA state found in checkpoint", rank)
         except Exception as e:
-            logger.warning("rank: %s, failed to load EMA state: %s", rank, str(e))
+            logger.warning("rank: %s, failed to load EMA state: %s", rank,
+                           str(e))
 
     # Load critic distributed checkpoint
     critic_dcp_dir = os.path.join(checkpoint_path, "distributed_checkpoint",
@@ -1296,6 +1298,10 @@ def get_scheduler(
                          last_epoch=last_epoch)
 
 
+def count_trainable(model: torch.nn.Module) -> int:
+    return sum(p.numel() for p in model.parameters() if p.requires_grad)
+
+
 class EMA_FSDP:
     """
     FSDP2-friendly EMA with two modes:
@@ -1317,6 +1323,7 @@ class EMA_FSDP:
       ema.update(model)
       ema.state_dict()  # on rank 0
     """
+
     def __init__(self, module, decay: float = 0.999, mode: str = "local_shard"):
         self.decay = float(decay)
         self.mode = mode
@@ -1342,7 +1349,10 @@ class EMA_FSDP:
         if self.mode == "rank0_full":
             cpu_state = gather_state_dict_on_cpu_rank0(module, device=None)
             if self.rank == 0:
-                self.shadow = {k: v.detach().clone().float().cpu() for k, v in cpu_state.items()}
+                self.shadow = {
+                    k: v.detach().clone().float().cpu()
+                    for k, v in cpu_state.items()
+                }
             else:
                 self.shadow = {}
             return
@@ -1383,7 +1393,10 @@ class EMA_FSDP:
 
     def state_dict(self) -> dict[str, torch.Tensor]:
         if self.mode == "rank0_full":
-            return {k: v.clone() for k, v in self.shadow.items()} if self.rank == 0 else {}
+            return {
+                k: v.clone()
+                for k, v in self.shadow.items()
+            } if self.rank == 0 else {}
         return {k: v.clone() for k, v in self.shadow.items()}
 
     def load_state_dict(self, sd: dict[str, torch.Tensor]):
@@ -1404,6 +1417,7 @@ class EMA_FSDP:
                 p.data.copy_(w.to(dtype=p.dtype, device=p.device))
 
     class _ApplyEMACtx:
+
         def __init__(self, ema: "EMA_FSDP", module):
             self.ema = ema
             self.module = module
@@ -1411,7 +1425,9 @@ class EMA_FSDP:
 
         def __enter__(self):
             if self.ema.mode != "local_shard":
-                raise RuntimeError("EMA apply_to_model is only supported for mode='local_shard'")
+                raise RuntimeError(
+                    "EMA apply_to_model is only supported for mode='local_shard'"
+                )
             with torch.no_grad():
                 for name, p in self.module.named_parameters():
                     if not p.requires_grad:
@@ -1421,14 +1437,17 @@ class EMA_FSDP:
                     if p_local.numel() == 0:
                         # Nothing to swap on this rank for this param
                         continue
-                    self.saved[name] = p_local.clone().to(device=p_local.device, dtype=p_local.dtype)
+                    self.saved[name] = p_local.clone().to(device=p_local.device,
+                                                          dtype=p_local.dtype)
                     if name in self.ema.shadow:
                         ema_cpu = self.ema.shadow[name]
                         if ema_cpu.numel() != p_local.numel():
                             # Shard shape mismatch (e.g., empty shard here), skip
                             continue
                         # Copy EMA shard into local param shard
-                        p_local.copy_(ema_cpu.to(dtype=p_local.dtype, device=p_local.device))
+                        p_local.copy_(
+                            ema_cpu.to(dtype=p_local.dtype,
+                                       device=p_local.device))
             return self.module
 
         def __exit__(self, exc_type, exc, tb):
@@ -1446,4 +1465,4 @@ class EMA_FSDP:
             return False
 
     def apply_to_model(self, module):
-        return EMA_FSDP._ApplyEMACtx(self, module)
+        return EMA_FSDP._ApplyEMACtx(self, module)

fastvideo/training/wan_self_forcing_distillation_pipeline.py

@@ -4,10 +4,10 @@ from copy import deepcopy
 
 from fastvideo.fastvideo_args import FastVideoArgs, TrainingArgs
 from fastvideo.logger import init_logger
-from fastvideo.models.schedulers.scheduling_flow_match_euler_discrete import (
-    FlowMatchEulerDiscreteScheduler)
-from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import WanCausalDMDPipeline
-from fastvideo.training.self_forcing_distillation_pipeline import SelfForcingDistillationPipeline
+from fastvideo.pipelines.basic.wan.wan_causal_dmd_pipeline import (
+    WanCausalDMDPipeline)
+from fastvideo.training.self_forcing_distillation_pipeline import (
+    SelfForcingDistillationPipeline)
 from fastvideo.utils import is_vsa_available
 
 vsa_available = is_vsa_available()
@@ -25,11 +25,6 @@ class WanSelfForcingDistillationPipeline(SelfForcingDistillationPipeline):
         "fake_score_transformer"
     ]
 
-    def initialize_pipeline(self, fastvideo_args: FastVideoArgs):
-        """Initialize Wan-specific scheduler."""
-        self.modules["scheduler"] = FlowMatchEulerDiscreteScheduler(
-            shift=fastvideo_args.pipeline_config.flow_shift)
-
     def create_training_stages(self, training_args: TrainingArgs):
         """
         May be used in future refactors.
@@ -74,4 +69,4 @@ if __name__ == "__main__":
     parser = TrainingArgs.add_cli_args(parser)
     parser = FastVideoArgs.add_cli_args(parser)
     args = parser.parse_args()
-    main(args) 
+    main(args)

fastvideo/training/wan_training_pipeline.py

@@ -34,7 +34,7 @@ class WanTrainingPipeline(TrainingPipeline):
     def initialize_validation_pipeline(self, training_args: TrainingArgs):
         logger.info("Initializing validation pipeline...")
         args_copy = deepcopy(training_args)
-        assert training_args.dit_cpu_offload == False
+
         args_copy.inference_mode = True
         validation_pipeline = WanPipeline.from_pretrained(
             training_args.model_path,
@@ -42,13 +42,12 @@ class WanTrainingPipeline(TrainingPipeline):
             inference_mode=True,
             loaded_modules={
                 "transformer": self.get_module("transformer"),
-                "transformer_2": self.get_module("transformer_2")
             },
             tp_size=training_args.tp_size,
             sp_size=training_args.sp_size,
             num_gpus=training_args.num_gpus,
             pin_cpu_memory=training_args.pin_cpu_memory,
-            dit_cpu_offload=training_args.dit_cpu_offload)
+            dit_cpu_offload=True)
 
         self.validation_pipeline = validation_pipeline
 

fastvideo/utils.py

@@ -23,10 +23,14 @@ from typing import Any, TypeVar, cast
 
 import cloudpickle
 import filelock
+import imageio
+import numpy as np
 import torch
+import torchvision
 import yaml
 from diffusers.loaders.lora_base import (
     _best_guess_weight_name)  # watch out for potetential removal from diffusers
+from einops import rearrange
 from huggingface_hub import snapshot_download
 from remote_pdb import RemotePdb
 from torch.distributed.fsdp import MixedPrecisionPolicy
@@ -34,7 +38,6 @@ from torch.distributed.fsdp import MixedPrecisionPolicy
 import fastvideo.envs as envs
 from fastvideo.logger import init_logger
 
-
 logger = init_logger(__name__)
 
 T = TypeVar("T")
@@ -615,6 +618,7 @@ def maybe_download_model_index(model_name_or_path: str) -> dict[str, Any]:
             f"Failed to download or parse model_index.json for {model_name_or_path}: {e}"
         ) from e
 
+
 def update_environment_variables(envs: dict[str, str]):
     for k, v in envs.items():
         if k in os.environ and os.environ[k] != v:
@@ -814,6 +818,17 @@ def is_vsa_available() -> bool:
     return importlib.util.find_spec("vsa") is not None
 
 
+@lru_cache(maxsize=1)
+def is_vmoba_available() -> bool:
+    if importlib.util.find_spec("csrc.attn.vmoba_attn.vmoba") is None:
+        return False
+    try:
+        import flash_attn
+        return flash_attn.__version__ >= "2.7.4"
+    except Exception:
+        return False
+
+
 # adapted from: https://github.com/Wan-Video/Wan2.2/blob/main/wan/utils/utils.py
 def masks_like(tensor,
                zero=False,
@@ -875,3 +890,17 @@ def best_output_size(w, h, dw, dh, expected_area):
         return ow1, oh1
     else:
         return ow2, oh2
+
+
+def save_decoded_latents_as_video(decoded_latents: list[torch.Tensor],
+                                  output_path: str, fps: int):
+    # Process outputs
+    videos = rearrange(decoded_latents, "b c t h w -> t b c h w")
+    frames = []
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=6)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)
+        frames.append((x * 255).numpy().astype(np.uint8))
+
+    os.makedirs(os.path.dirname(output_path), exist_ok=True)
+    imageio.mimsave(output_path, frames, fps=fps, format="mp4")