[bugfix]: correct hardware claims and arch handling in QAD 5090 example scripts

Author: SolitaryThinker

examples/inference/optimizations/_qad_common.py

@@ -0,0 +1,30 @@
+"""Shared GPU-arch helpers for the QAD 5090 example scripts.
+
+The default (non ``--bf16``) path of the NVFP4 scripts runs DiT linears through
+flashinfer's cutlass FP4 gemm, which ships sm_120a cubins only (RTX 5090-class
+consumer Blackwell). These helpers pick the right FLASHINFER_CUDA_ARCH_LIST and
+fail fast with a readable error instead of an opaque flashinfer ValueError.
+"""
+
+import torch
+
+# Capabilities the flashinfer FP4 gemm path has cubins for.
+FP4_CAPABILITIES = ((12, 0), (12, 1))
+
+
+def flashinfer_arch_list() -> str:
+    """FLASHINFER_CUDA_ARCH_LIST derived from the local GPU, defaulting to 12.0a."""
+    if torch.cuda.is_available():
+        major, minor = torch.cuda.get_device_capability()
+        if (major, minor) in FP4_CAPABILITIES:
+            return f"{major}.{minor}a"
+    return "12.0a"
+
+
+def require_fp4_capable_gpu() -> None:
+    """Exit with a one-line error if the GPU cannot run the NVFP4 gemm path."""
+    cap = torch.cuda.get_device_capability() if torch.cuda.is_available() else None
+    if cap not in FP4_CAPABILITIES:
+        raise SystemExit(
+            f"The default NVFP4 path requires an sm_120a GPU (RTX 5090-class Blackwell, "
+            f"compute capability 12.0/12.1); found {cap}. Re-run with --bf16 on other GPUs.")

examples/inference/optimizations/fp8_wan2_1_1_3b.py

@@ -6,6 +6,8 @@
 Requirements:
     - GPU: sm89+ (H100, L40S, RTX 4090, Ada Lovelace, or newer)
       Falls back to a bf16 dequant path on older GPUs.
+    - sageattention: pip install sageattention (the script defaults to the
+      SAGE_ATTN attention backend)
     - TAEHV (optional): Follow install instructions at https://github.com/madebyollin/taehv
 
 Usage:
@@ -35,7 +37,7 @@ def __init__(self, checkpoint_path: str, device: str = "cuda",
     @torch.no_grad()
     def decode(self, latents: torch.Tensor):
         latents = latents.permute(0, 2, 1, 3, 4).to(self.device, self.dtype)
-        decoded = self.model.decode_video(latents, parallel=False, show_progress_bar=False)
+        decoded = self.model.decode_video(latents, parallel=True, show_progress_bar=False)
         frames = (decoded[0].clamp(0, 1) * 255).to(torch.uint8)
         return frames.permute(0, 2, 3, 1).cpu().numpy()
 

examples/inference/optimizations/nvfp4_qat_wan2_1_1_3b.py

@@ -4,7 +4,11 @@
 NVFP4QATConfig quantization. Uses ATTN_QAT_INFER attention backend.
 
 Requirements:
-    - GPU: Blackwell (B200/B300, sm100a+) for the FP4 linear path
+    - GPU (default NVFP4 mode): RTX 5090-class Blackwell (sm_120a). The FP4
+      linear path uses flashinfer's sm_120a cubins, which do not run on
+      sm_100 (B200/B300), and the attn_qat_infer kernel hard-gates on sm_120.
+    - GPU (--bf16 baseline): any supported GPU (the ATTN_QAT_INFER backend
+      falls back to Flash Attention off sm_120).
     - TAEHV (optional): Follow install instructions at https://github.com/madebyollin/taehv
 
 Usage:
@@ -33,7 +37,7 @@ def __init__(self, checkpoint_path: str, device: str = "cuda",
     @torch.no_grad()
     def decode(self, latents: torch.Tensor):
         latents = latents.permute(0, 2, 1, 3, 4).to(self.device, self.dtype)
-        decoded = self.model.decode_video(latents, parallel=False, show_progress_bar=False)
+        decoded = self.model.decode_video(latents, parallel=True, show_progress_bar=False)
         frames = (decoded[0].clamp(0, 1) * 255).to(torch.uint8)
         return frames.permute(0, 2, 3, 1).cpu().numpy()
 
@@ -51,9 +55,14 @@ def main():
     parser.add_argument("--infer_steps", type=int, default=3)
     args = parser.parse_args()
 
+    from _qad_common import flashinfer_arch_list, require_fp4_capable_gpu
+
+    if not args.bf16:
+        require_fp4_capable_gpu()
+
     os.environ.setdefault("FASTVIDEO_ATTENTION_BACKEND", "ATTN_QAT_INFER")
     os.environ["FASTVIDEO_DISABLE_ATTENTION_COMPILE"] = "0"
-    os.environ["FLASHINFER_CUDA_ARCH_LIST"] = "12.0a"
+    os.environ.setdefault("FLASHINFER_CUDA_ARCH_LIST", flashinfer_arch_list())
 
     from fastvideo import VideoGenerator
     from fastvideo.configs.pipelines.base import PipelineConfig

examples/inference/optimizations/nvfp4_sa2_wan2_1_1_3b.py

@@ -4,7 +4,9 @@
 NVFP4QATConfig quantization. Uses the SAGE_ATTN attention backend.
 
 Requirements:
-    - GPU: sm89+ (H100, L40S, RTX 4090, Ada Lovelace, or newer)
+    - GPU (default NVFP4 mode): RTX 5090-class Blackwell (sm_120a). The NVFP4
+      linear path uses flashinfer's cutlass FP4 gemm, which has no fallback.
+    - GPU (--bf16 baseline): sm89+ (H100, L40S, RTX 4090, Ada Lovelace, or newer)
     - sageattention: pip install sageattention
     - TAEHV (optional): Follow install instructions at https://github.com/madebyollin/taehv
 
@@ -34,7 +36,7 @@ def __init__(self, checkpoint_path: str, device: str = "cuda",
     @torch.no_grad()
     def decode(self, latents: torch.Tensor):
         latents = latents.permute(0, 2, 1, 3, 4).to(self.device, self.dtype)
-        decoded = self.model.decode_video(latents, parallel=False, show_progress_bar=False)
+        decoded = self.model.decode_video(latents, parallel=True, show_progress_bar=False)
         frames = (decoded[0].clamp(0, 1) * 255).to(torch.uint8)
         return frames.permute(0, 2, 3, 1).cpu().numpy()
 
@@ -52,9 +54,14 @@ def main():
     parser.add_argument("--infer_steps", type=int, default=3)
     args = parser.parse_args()
 
+    from _qad_common import flashinfer_arch_list, require_fp4_capable_gpu
+
+    if not args.bf16:
+        require_fp4_capable_gpu()
+
     os.environ.setdefault("FASTVIDEO_ATTENTION_BACKEND", "SAGE_ATTN")
     os.environ["FASTVIDEO_DISABLE_ATTENTION_COMPILE"] = "0"
-    os.environ["FLASHINFER_CUDA_ARCH_LIST"] = "12.0a"
+    os.environ.setdefault("FLASHINFER_CUDA_ARCH_LIST", flashinfer_arch_list())
 
     from fastvideo import VideoGenerator
     from fastvideo.configs.pipelines.base import PipelineConfig