fix

docs/design/inference_schema_parity_inventory.yaml

@@ -346,6 +346,7 @@ surfaces:
       num_inference_steps: request.sampling.num_inference_steps
       num_inference_steps_sr: request.sampling.num_inference_steps_sr
       guidance_scale: request.sampling.guidance_scale
+      guidance_scale_2: request.sampling.guidance_scale_2
       guidance_rescale: request.sampling.guidance_rescale
       boundary_ratio: request.sampling.boundary_ratio
       sigmas: request.sampling.sigmas
@@ -365,15 +366,7 @@ surfaces:
       data_type: "Derived from the request shape and not a public input."
 
   sampling_param_extensions:
-    moved:
-      guidance_scale_2:
-        target: request.sampling.guidance_scale_2
-        sources:
-          - fastvideo.configs.sample.lingbotworld.LingBotWorld_SamplingParam
-          - fastvideo.configs.sample.lingbotworld.Wan2_2_I2V_A14B_SamplingParam
-          - fastvideo.configs.sample.wan.SelfForcingWan2_2_T2V_A14B_480P_SamplingParam
-          - fastvideo.configs.sample.wan.Wan2_2_I2V_A14B_SamplingParam
-          - fastvideo.configs.sample.wan.Wan2_2_T2V_A14B_SamplingParam
+    moved: {}
     profile_owned:
       action_list:
         target: request.extensions.hunyuangamecraft.action_list
@@ -598,6 +591,7 @@ cli:
       - text_encoder_cpu_offload
       - text_encoder_precisions
       - torch_compile_kwargs
+      - transformer_quant
       - tp_size
       - trust_remote_code
       - use_fsdp_inference
@@ -703,6 +697,7 @@ cli:
       - text_encoder_cpu_offload
       - text_encoder_precisions
       - torch_compile_kwargs
+      - transformer_quant
       - tp_size
       - trust_remote_code
       - use_fsdp_inference

fastvideo-kernel/python/fastvideo_kernel/triton_kernels/__init__.py

@@ -0,0 +1,5 @@
+"""Triton kernel entrypoints exposed by ``fastvideo_kernel``."""
+
+from .fused_attention import attention as fused_attention
+
+__all__ = ["fused_attention"]

fastvideo-kernel/python/fastvideo_kernel/triton_kernels/fused_attention.py

@@ -0,0 +1,55 @@
+"""Compatibility shim for the legacy non-QAT Triton attention import path.
+
+Historically callers imported
+``fastvideo_kernel.triton_kernels.fused_attention`` directly. The shared
+implementation now lives in ``attn_qat_train.py`` and is parameterized by the
+``IS_QAT`` flag. This module preserves the original public API for tests and
+downstream users while always dispatching to the non-QAT configuration.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from .attn_qat_train import attention as _attention
+
+
+def attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    causal: bool,
+    sm_scale: float,
+    warp_specialize: bool = True,
+) -> torch.Tensor:
+    """Run the shared Triton attention kernel in non-QAT mode."""
+    use_qat_qkv_backward = True
+    smooth_k = False
+    is_qat = False
+    two_level_quant_p = False
+    fake_quant_p = False
+    use_high_prec_o = False
+    smooth_q = False
+    use_global_sf_p = False
+    use_global_sf_qkv = False
+
+    return _attention(
+        q,
+        k,
+        v,
+        causal,
+        sm_scale,
+        use_qat_qkv_backward,
+        smooth_k,
+        warp_specialize,
+        is_qat,
+        two_level_quant_p,
+        fake_quant_p,
+        use_high_prec_o,
+        smooth_q,
+        use_global_sf_p,
+        use_global_sf_qkv,
+    )
+
+
+__all__ = ["attention"]

fastvideo-kernel/tests/__init__.py

@@ -0,0 +1,4 @@
+from ._bootstrap import ensure_local_kernel_sources_first
+
+
+ensure_local_kernel_sources_first()

fastvideo-kernel/tests/_bootstrap.py

@@ -0,0 +1,56 @@
+"""Test import helpers for preferring the in-tree fastvideo-kernel sources."""
+
+from __future__ import annotations
+
+import importlib
+import sys
+from pathlib import Path
+
+
+def _prepend_import_path(path: Path) -> None:
+    path_str = str(path)
+    if path_str not in sys.path:
+        sys.path.insert(0, path_str)
+
+
+def _purge_package(name: str) -> None:
+    prefix = f"{name}."
+    for module_name in tuple(sys.modules):
+        if module_name == name or module_name.startswith(prefix):
+            sys.modules.pop(module_name, None)
+
+
+def _module_is_from_checkout(module_file: str | None, checkout_root: Path) -> bool:
+    if module_file is None:
+        return False
+
+    try:
+        module_path = Path(module_file).resolve()
+    except OSError:
+        return False
+
+    return module_path.is_relative_to(checkout_root.resolve())
+
+
+def ensure_local_kernel_sources_first() -> None:
+    tests_root = Path(__file__).resolve().parent
+    kernel_root = tests_root.parent
+    repo_root = kernel_root.parent
+    kernel_python_root = kernel_root / "python"
+
+    # Keep the in-tree kernel sources ahead of any preinstalled wheel so tests
+    # exercise the checkout under review.
+    for path in (repo_root, kernel_root, kernel_python_root):
+        _prepend_import_path(path)
+
+    importlib.invalidate_caches()
+
+    loaded_kernel = sys.modules.get("fastvideo_kernel")
+    if loaded_kernel is None:
+        return
+
+    if not _module_is_from_checkout(
+        getattr(loaded_kernel, "__file__", None),
+        kernel_python_root,
+    ):
+        _purge_package("fastvideo_kernel")

fastvideo-kernel/tests/conftest.py

@@ -0,0 +1,4 @@
+from ._bootstrap import ensure_local_kernel_sources_first
+
+
+ensure_local_kernel_sources_first()

fastvideo-kernel/tests/test_bootstrap.py

@@ -0,0 +1,30 @@
+from __future__ import annotations
+
+import sys
+import types
+from pathlib import Path
+
+from tests._bootstrap import ensure_local_kernel_sources_first
+
+
+def test_bootstrap_prefers_local_kernel_checkout(monkeypatch) -> None:
+    tests_root = Path(__file__).resolve().parent
+    kernel_root = tests_root.parent
+    repo_root = kernel_root.parent
+    kernel_python_root = kernel_root / "python"
+
+    stale_module = types.ModuleType("fastvideo_kernel")
+    stale_module.__file__ = (
+        "/tmp/site-packages/fastvideo_kernel/__init__.py"
+    )
+    monkeypatch.setitem(sys.modules, "fastvideo_kernel", stale_module)
+    monkeypatch.setattr(sys, "path", ["/tmp/site-packages"])
+
+    ensure_local_kernel_sources_first()
+
+    assert "fastvideo_kernel" not in sys.modules
+    assert sys.path[:3] == [
+        str(kernel_python_root),
+        str(kernel_root),
+        str(repo_root),
+    ]

fastvideo/api/compat.py

@@ -107,6 +107,8 @@ def legacy_from_pretrained_to_config(
             engine[key] = value
         elif key == "override_text_encoder_quant":
             quantization["text_encoder_quant"] = value
+        elif key == "transformer_quant":
+            quantization["transformer_quant"] = value
         elif key == "workload_type":
             pipeline["workload_type"] = value
         elif key == "lora_path":

fastvideo/fastvideo_args.py

@@ -436,96 +436,6 @@ class FastVideoArgs:
             help="BSA chunk_3d_shape_k as three ints, e.g., 4 4 4.",
         )
 
-        # LTX-2 VAE tiling overrides
-        parser.add_argument(
-            "--ltx2-vae-tiling",
-            action=StoreBoolean,
-            default=FastVideoArgs.ltx2_vae_tiling,
-            help="Enable LTX-2 VAE tiling overrides.",
-        )
-        parser.add_argument(
-            "--ltx2-vae-spatial-tile-size-in-pixels",
-            type=int,
-            default=FastVideoArgs.ltx2_vae_spatial_tile_size_in_pixels,
-            help="LTX-2 VAE spatial tile size in pixels.",
-        )
-        parser.add_argument(
-            "--ltx2-vae-spatial-tile-overlap-in-pixels",
-            type=int,
-            default=FastVideoArgs.ltx2_vae_spatial_tile_overlap_in_pixels,
-            help="LTX-2 VAE spatial tile overlap in pixels.",
-        )
-        parser.add_argument(
-            "--ltx2-vae-temporal-tile-size-in-frames",
-            type=int,
-            default=FastVideoArgs.ltx2_vae_temporal_tile_size_in_frames,
-            help="LTX-2 VAE temporal tile size in frames.",
-        )
-        parser.add_argument(
-            "--ltx2-vae-temporal-tile-overlap-in-frames",
-            type=int,
-            default=FastVideoArgs.ltx2_vae_temporal_tile_overlap_in_frames,
-            help="LTX-2 VAE temporal tile overlap in frames.",
-        )
-        parser.add_argument(
-            "--ltx2-initial-latent-path",
-            type=str,
-            default=FastVideoArgs.ltx2_initial_latent_path,
-            help="Path to load/save a precomputed LTX-2 initial latent.",
-        )
-
-        # LoRA parameters (inference-time adapter loading)
-        parser.add_argument(
-            "--lora-path",
-            type=str,
-            default=FastVideoArgs.lora_path,
-            help="Path to a LoRA adapter (directory or HF repo id). If set, LoRA will be applied at inference.",
-        )
-        parser.add_argument(
-            "--lora-nickname",
-            type=str,
-            default=FastVideoArgs.lora_nickname,
-            help="Nickname to refer to the loaded LoRA adapter (useful for swapping).",
-        )
-        parser.add_argument(
-            "--lora-target-modules",
-            nargs="+",
-            type=str,
-            default=FastVideoArgs.lora_target_modules,
-            help="Optional list of module name substrings to restrict LoRA injection (e.g. q_proj k_proj v_proj).",
-        )
-
-        # BSA runtime control (LongCat)
-        parser.add_argument(
-            "--enable-bsa",
-            action=StoreBoolean,
-            help="Enable Block Sparse Attention (BSA) at runtime (overrides config).",
-        )
-        parser.add_argument(
-            "--bsa-sparsity",
-            type=float,
-            help="BSA sparsity (e.g., 0.9375).",
-        )
-        parser.add_argument(
-            "--bsa-cdf-threshold",
-            type=float,
-            help="BSA CDF threshold (optional).",
-        )
-        parser.add_argument(
-            "--bsa-chunk-q",
-            nargs=3,
-            type=int,
-            metavar=("T", "H", "W"),
-            help="BSA chunk_3d_shape_q as three ints, e.g., 4 4 4.",
-        )
-        parser.add_argument(
-            "--bsa-chunk-k",
-            nargs=3,
-            type=int,
-            metavar=("T", "H", "W"),
-            help="BSA chunk_3d_shape_k as three ints, e.g., 4 4 4.",
-        )
-
         parser.add_argument(
             "--enable-torch-compile",
             action=StoreBoolean,
@@ -551,11 +461,6 @@ class FastVideoArgs:
             action=StoreBoolean,
             help="Enable layerwise CPU offload with async H2D prefetch overlap.",
         )
-        parser.add_argument(
-            "--dit-layerwise-offload",
-            action=StoreBoolean,
-            help="Enable layerwise CPU offload with async H2D prefetch overlap.",
-        )
         parser.add_argument(
             "--use-fsdp-inference",
             action=StoreBoolean,
@@ -645,12 +550,6 @@ class FastVideoArgs:
             default=FastVideoArgs.override_pipeline_cls_name,
             help="Override pipeline cls name",
         )
-        parser.add_argument(
-            "--override-pipeline-cls-name",
-            type=str,
-            default=FastVideoArgs.override_pipeline_cls_name,
-            help="Override pipeline cls name",
-        )
         parser.add_argument("--init-weights-from-safetensors",
                             type=str,
                             help="Path to safetensors file for initial weight loading")

fastvideo/tests/api/test_schema_parity_inventory.py

@@ -235,8 +235,8 @@ def test_cli_dest_inventory_matches_live_parsers() -> None:
 def test_review_gap_fields_are_explicitly_inventory_tracked() -> None:
     inventory = _load_inventory()
 
-    sampling_extensions = inventory["surfaces"]["sampling_param_extensions"]
-    assert "guidance_scale_2" in sampling_extensions["moved"]
+    sampling_base = inventory["surfaces"]["sampling_param_base"]
+    assert "guidance_scale_2" in sampling_base["moved"]
 
     image_request = inventory["surfaces"]["openai_image_request"]
     video_request = inventory["surfaces"]["openai_video_request"]