[test]: reconcile attn_qat_infer capability test with hao-ai-lab#1647 FA4 sm_100/103

hao-ai-lab#1647 extended ATTN_QAT_INFER to datacenter Blackwell (sm_100/sm_103) via the
FP4 FA4 path, so is_attn_qat_infer_available() now resolves True on (10,0)/(10,3)
too — the pre-hao-ai-lab#1647 assertion (10,0)->False is stale. Parametrize over both
kernel paths and force both probes "installed" (_get_attn_qat_infer for sm_12x
CUTLASS, _fa4_fp4_available for sm_100/103 FA4) so the test isolates the
capability gate deterministically regardless of what's importable on the runner.

Author: Mister-Raggs

fastvideo/tests/attention/test_attn_qat_infer_capability.py

@@ -6,11 +6,28 @@
 
 @pytest.mark.parametrize(
     "capability,expected",
-    [((12, 0), True), ((12, 1), True), ((10, 0), False), ((9, 0), False)],
+    [
+        ((12, 0), True),   # sm_120 consumer Blackwell  -> fastvideo-kernel CUTLASS (SageAttn3 FP4)
+        ((12, 1), True),   # sm_121 (DGX Spark)         -> fastvideo-kernel CUTLASS (SageAttn3 FP4)
+        ((10, 0), True),   # sm_100 datacenter Blackwell -> FP4 FA4 (flash-attention-fp4), #1647
+        ((10, 3), True),   # sm_103 datacenter Blackwell -> FP4 FA4 (flash-attention-fp4), #1647
+        ((9, 0), False),   # sm_90 Hopper   -> unsupported (falls back to Flash)
+        ((8, 9), False),   # sm_89 Ada/L40S -> unsupported (falls back to Flash)
+    ],
 )
 def test_attn_qat_infer_availability_requires_supported_gpu(monkeypatch, capability, expected):
+    """Availability is gated on the active device's compute capability, so a
+    wheel that carries the kernel on an unsupported host (e.g. import succeeds
+    on Hopper/Ada) still falls back to Flash instead of crashing at kernel call.
+
+    Both kernel probes are forced "installed" so the assertion isolates the
+    capability gate, independent of what's importable on the CI runner:
+      * ``_get_attn_qat_infer`` -> the sm_12x CUTLASS extension
+      * ``_fa4_fp4_available``  -> the sm_100/sm_103 FP4 FA4 path (#1647)
+    """
     monkeypatch.setattr(torch.cuda, "is_available", lambda: True)
     monkeypatch.setattr(torch.cuda, "get_device_capability", lambda: capability)
     monkeypatch.setattr(attn_qat_infer, "_get_attn_qat_infer", lambda: object())
+    monkeypatch.setattr(attn_qat_infer, "_fa4_fp4_available", lambda: True)
 
     assert attn_qat_infer.is_attn_qat_infer_available() is expected