fix: avoid SM120 MXFP8 MXFP4 autotune crash

flashinfer/fused_moe/core.py

@@ -934,7 +934,17 @@ def cutlass_fused_moe(
             use_wfp4afp8_humming=use_wfp4afp8_humming,
         )
 
-        if profile_ids is None:
+        if (
+            profile_ids is None
+            and backend in ("120", "121")
+            and use_mxfp8_act_scaling
+            and fc1_expert_weights.dtype == torch.int64
+        ):
+            # The SM120 profiler cannot safely construct MXFP8 x MXFP4 TMA
+            # inputs yet. Profiling these tactics can poison the CUDA context,
+            # so use the runner's fallback tactics for this mode.
+            gemm_tactic_1, gemm_tactic_2 = -1, -1
+        elif profile_ids is None:
             tuner = AutoTuner.get()
             MoERunner.refine_tuning_config(tune_max_num_tokens)
 

tests/moe/test_trtllm_cutlass_fused_moe.py

@@ -1570,7 +1570,7 @@ def test_moe_mxfp8_mxfp4(
     alpha,
     beta,
     limit,
-    use_autotune,
+    use_autotune=False,
 ):
     """
     Test MoE with MXFP8 activations and MXFP4 weights.
@@ -1682,6 +1682,25 @@ def test_moe_mxfp8_mxfp4(
     torch.testing.assert_close(ref_output, flash_output, rtol=1e-1, atol=1e-1)
 
 
+@pytest.mark.skipif(
+    torch.cuda.get_device_capability() not in [(12, 0), (12, 1)],
+    reason="Regression test is specific to SM120/SM121 autotuning",
+)
+def test_moe_mxfp8_mxfp4_autotune_sm120():
+    test_moe_mxfp8_mxfp4(
+        batch_size=1,
+        hidden_size=128,
+        num_experts=2,
+        top_k=2,
+        intermediate_size=128,
+        otype=torch.bfloat16,
+        alpha=None,
+        beta=None,
+        limit=None,
+        use_autotune=True,
+    )
+
+
 @pytest.mark.parametrize("batch_size", BATCH_SIZES)
 @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES)
 @pytest.mark.parametrize("num_experts", NUM_EXPERTS)