fix: add SM90 guard for FP8 Blockscale GEMM

lucifer1004 · lucifer1004 · commit fadb1a8b2e46 · 2025-04-15T06:45:05.000-07:00
Signed-off-by: Zihua Wu &lt;13583761+lucifer1004@users.noreply.github.com&gt;
diff --git a/cpp/tensorrt_llm/kernels/cutlass_kernels/fp8_blockscale_gemm/fp8_blockscale_gemm.cu b/cpp/tensorrt_llm/kernels/cutlass_kernels/fp8_blockscale_gemm/fp8_blockscale_gemm.cu
@@ -25,6 +25,14 @@ template <typename ElementA, typename ElementB, typename ElementD>
 CutlassFp8BlockScaleGemmRunner<ElementA, ElementB, ElementD>::CutlassFp8BlockScaleGemmRunner()
 {
     TLLM_LOG_DEBUG(__PRETTY_FUNCTION__);
+    int sm = tensorrt_llm::common::getSMVersion();
+    if (sm != 90)
+    {
+        TLLM_THROW(
+            "FP8 Blockscale GEMM kernels are only supported on SM90 architectures, but current device compute "
+            "capability is %d.",
+            sm);
+    }
 }
 
 template <typename ElementA, typename ElementB, typename ElementD>

Original file line number	Diff line number	Diff line change
`@@ -25,6 +25,14 @@ template <typename ElementA, typename ElementB, typename ElementD>`
`25`	`25`	`CutlassFp8BlockScaleGemmRunner<ElementA, ElementB, ElementD>::CutlassFp8BlockScaleGemmRunner()`
`26`	`26`	`{`
`27`	`27`	`TLLM_LOG_DEBUG(__PRETTY_FUNCTION__);`
	`28`	`+ int sm = tensorrt_llm::common::getSMVersion();`
	`29`	`+ if (sm != 90)`
	`30`	`+ {`
	`31`	`+ TLLM_THROW(`
	`32`	`+ "FP8 Blockscale GEMM kernels are only supported on SM90 architectures, but current device compute "`
	`33`	`+ "capability is %d.",`
	`34`	`+ sm);`
	`35`	`+ }`
`28`	`36`	`}`
`29`	`37`
`30`	`38`	`template <typename ElementA, typename ElementB, typename ElementD>`