[misc]: serve the Regenerate-2K examples on TORCH_SDPA and drop narrating comments

FLASH_ATTN bypassed the cuDNN attention the refine stage selects; TORCH_SDPA keeps it.

docs/inference/open_h3_regenerate_2k.md

@@ -87,10 +87,12 @@ The response headers carry `X-Request-Id`, `X-Model`, `X-Inference-Time-S` and `
 - `size`, `width` and `height` may be omitted or `2560x1440`; `num_frames` may be omitted or 124 (the frame count of the
   VAE grid; the video has 121 frames); `fps` must be 24. `seconds` is not accepted.
 - `num_inference_steps` is 3 (default) or 2; `seed` sets the refine noise.
+- Serve with `attention_backend: TORCH_SDPA`: the refine stage selects cuDNN attention, which at about 133k tokens is
+  about 3x faster than PyTorch's default flash dispatch. The `FLASH_ATTN` backend bypasses that selection.
 
 ## Parity with the reference
 
-Measured on one GB200 (tray hpc-rack-1-5, BF16, dense attention) against the reference `standard_fwd` with the released
+Measured on one GB200 (BF16, `TORCH_SDPA` with cuDNN attention) against the reference `standard_fwd` with the released
 VAE, same draft, seed (1766580932), prompt and VAE posterior seed (42). The reference's noise is a CUDA generator draw
 that FastVideo reproduces (the same `torch.Generator(device="cuda")` seed), so no injected noise was needed.
 

examples/inference/basic/basic_open_h3_regenerate_2k.py

@@ -38,7 +38,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--output", type=Path, default=Path("outputs/open_h3_regenerate_2k"))
     parser.add_argument("--steps", type=int, choices=(2, 3), default=3, help="refine steps (3 is the released setting)")
     parser.add_argument("--seed", type=int, default=1766580932)
-    parser.add_argument("--attention-backend", default="FLASH_ATTN")
+    parser.add_argument("--attention-backend", default="TORCH_SDPA")
     parser.add_argument("--draft-size", choices=("768p", "576p"), default="768p",
                         help="size the draft is encoded at; 576p for drafts from other video models")
     parser.add_argument("--runs", type=int, default=1, help="repeat the request to measure warm timings")

examples/serving/openai_open_h3_regenerate_2k.yaml

@@ -17,7 +17,7 @@ generator:
       vae_enabled: false
   pipeline:
     experimental:
-      attention_backend: FLASH_ATTN
+      attention_backend: TORCH_SDPA
       # 768p for MiniMax-H3 drafts; 576p encodes drafts from other video models at 1024x576.
       draft_size: 768p
 

fastvideo/models/loader/component_loader.py

@@ -1655,7 +1655,6 @@ class PipelineComponentLoader:
             component_model_path,
         )
 
-        # A component under a subfolder (e.g. "base_model/vae") dispatches on its own name.
         module_name = module_name.rsplit("/", 1)[-1]
         # Get the appropriate loader for this module type
         loader = ComponentLoader.for_module_type(module_name, transformers_or_diffusers)

fastvideo/pipelines/basic/open_h3_regenerate_2k/open_h3_regenerate_2k_pipeline.py

@@ -54,7 +54,6 @@ class OpenH3Regenerate2KPipeline(MiniMaxH3Pipeline):
                                                revision=revision,
                                                allow_patterns=self.get_hf_download_allow_patterns())
         if self.fastvideo_args.lora_path is None:
-            # The merged refine LoRA is part of the release; it is a constructor-time adapter like any H3 LoRA.
             self.fastvideo_args.lora_path = os.path.join(self.model_path, LORA_FILE)
         return verify_model_config_and_directory(os.path.join(self.model_path, BASE_MODEL_DIR),
                                                  required_component_dirs=())