Skip MXFP8 autotune on dense GEMM, which causes IMA (#29669)

Co-authored-by: Brayden Zhong <brayden.zhong@radixark.ai>

python/sglang/srt/model_executor/runner/flashinfer_autotune.py

@@ -88,15 +88,8 @@ def should_run_flashinfer_autotune(
         "modelopt_fp8",
         "modelopt_mixed",
     )
-    # Online MXFP8 (microscaling) linears dispatch to flashinfer's
+    fp8_gemm_needs_autotune = get_fp8_gemm_runner_backend().is_flashinfer_cutlass() or (
-    # ``mm_mxfp8``, which the flashinfer fp8 autotune dummy run does not
+        model_uses_modelopt_fp8 and is_sm100_supported()
-    # exercise correctly -- it triggers an illegal memory access inside the
-    # mxfp8 cutlass cubin. The mxfp8 gemm is fixed-config and needs no
-    # tuning, so skip autotune for these models.
-    model_uses_mxfp8 = "mxfp8" in (model_quantization or "")
-    fp8_gemm_needs_autotune = not model_uses_mxfp8 and (
-        get_fp8_gemm_runner_backend().is_flashinfer_cutlass()
-        or (model_uses_modelopt_fp8 and is_sm100_supported())
     )
 
     if not (moe_needs_autotune or fp4_gemm_needs_autotune or fp8_gemm_needs_autotune):
@@ -180,7 +173,10 @@ def flashinfer_autotune_context(model_runner: ModelRunner, *, skip_logits: bool)
 
             maybe_skip_logits = autotune_dummy_run_mode()
         with torch.inference_mode(), autotune(
-            True, cache=str(autotune_cache)
+            # Autotuning mxfp8_gemm hits an IMA; skip it.
+            True,
+            cache=str(autotune_cache),
+            skip_ops={"mxfp8_gemm"},
         ), maybe_skip_logits:
             yield
     torch.cuda.current_stream().wait_stream(mr.forward_stream)