[Bugfix] flashinfer: fail fast when --kv-cache-dtype nvfp4 used on unsupported arch (#43669)

Signed-off-by: Kartavya Sonar <sonarkartavya@gmail.com>
2026-06-02 10:50:00 -07:00
parent afcb580715
commit fe32e7830b
1 changed files with 7 additions and 0 deletions
@@ -623,6 +623,13 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
            # storage dtype may not be the same as the op dtype (uint8 vs fp8_e4m3)
            self.is_kvcache_nvfp4 = self.cache_dtype == "nvfp4"
            if self.is_kvcache_nvfp4:
+                # trtllm-gen FP4 FMHA kernels only exist for sm100f (sm_100/sm_103).
+                # Fail fast at init rather than crashing on the first request.
+                if not current_platform.is_device_capability_family(100):
+                    raise ValueError(
+                        "--kv-cache-dtype nvfp4 requires sm100f, "
+                        "please try a different dtype or remove"
+                    )
                # For NVFP4, kv_cache_dtype stays as the string "nvfp4"
                # which is passed to FlashInferImpl
                self.kv_cache_dtype = self.cache_dtype