[bugfix]: hard-fail when ATTN_QAT_INFER is selected but the kernel is unusable

fastvideo/platforms/cuda.py

@@ -146,9 +146,12 @@ def get_attn_backend_cls(cls, selected_backend: AttentionBackendEnum | None, hea
             if is_attn_qat_infer_available():
                 logger.info("Using Attn-QAT inference backend (%s).", attn_qat_infer_receipt())
                 return "fastvideo.attention.backends.attn_qat_infer.AttnQatInferBackend"
-            # Keep the trailing sentence stable: downstream receipts grep for it.
-            logger.info("Attn-QAT inference kernel is not built (%s). Fall back to Flash Attention.",
-                        attn_qat_infer_receipt())
+            raise ImportError(
+                f"ATTN_QAT_INFER selected but the inference kernel is not usable ({attn_qat_infer_receipt()}). "
+                "Silent fallback would run plain FlashAttention while the caller believes it is measuring "
+                "FP4-QAT attention — an A/B comparison would silently benchmark bf16 against bf16; "
+                "refusing to proceed. Build the fastvideo-kernel attn_qat_infer target for this arch "
+                "or pick a different FASTVIDEO_ATTENTION_BACKEND.")
         elif selected_backend == AttentionBackendEnum.ATTN_QAT_TRAIN:
             from fastvideo.attention.backends.attn_qat_train import (  # noqa: F401
                 AttnQatTrainBackend, is_attn_qat_train_available)

fastvideo/tests/attention/test_attn_qat_infer_arch_gate.py

@@ -4,7 +4,9 @@
 The capability sets route:
   * sm_120a/sm_121a -> fastvideo-kernel CUTLASS extension,
   * sm_100a/sm_103a -> FP4 FA4 (flash-attention-fp4, #1221 plumbing),
-  * anything else  -> unavailable (honest FlashAttention fallback upstream).
+  * anything else  -> unavailable (hard ImportError upstream: an explicitly
+    selected QAT-inference backend must never silently degrade to plain
+    FlashAttention, or A/B benchmarks measure bf16 against bf16).
 
 All device/import probes are monkeypatched; no GPU or kernel install needed.
 """
@@ -71,6 +73,22 @@ def test_unsupported_arch_receipt_lists_support_matrix(monkeypatch) -> None:
     assert "sm_100a/sm_103a" in receipt
 
 
+def test_selecting_unavailable_backend_raises_instead_of_falling_back(monkeypatch) -> None:
+    """Explicit ATTN_QAT_INFER on an arch without the kernel must hard-fail.
+
+    The old behavior logged "Fall back to Flash Attention" and continued,
+    which let FP4 A/B benchmarks silently measure bf16 against bf16.
+    """
+    import torch
+
+    from fastvideo.platforms.cuda import CudaPlatform
+    from fastvideo.platforms.interface import AttentionBackendEnum
+
+    _patch(monkeypatch, cap=(9, 0), cutlass=False, fa4=False)
+    with pytest.raises(ImportError, match="ATTN_QAT_INFER selected but"):
+        CudaPlatform.get_attn_backend_cls(AttentionBackendEnum.ATTN_QAT_INFER, 128, torch.bfloat16)
+
+
 def test_unsupported_arch_forward_never_calls_bundled_extension(monkeypatch) -> None:
     """An unsupported GPU (e.g. sm_90) carrying an importable sm_12x wheel
     must fail cleanly at forward — never dispatch into the wrong binary."""