fix bug of using fp4 on sm120

byshiue · byshiue · commit ef6d38a3fa4f · 2025-07-19T04:53:38.000+08:00
Signed-off-by: bhsueh &lt;11360707+byshiue@users.noreply.github.com&gt;
diff --git a/cpp/tensorrt_llm/thop/attentionOp.cpp b/cpp/tensorrt_llm/thop/attentionOp.cpp
@@ -671,7 +671,8 @@ bool attention_supports_nvfp4_output(int64_t const num_heads, int64_t const num_
     bool const use_paged_context_fmha, bool is_mla_enable)
 {
     // Only Blackwell supports NVFP4 output.
-    if (tensorrt_llm::common::getSMVersion() < 100)
+    // SM 120 does not support NVFP4 output.
+    if (tensorrt_llm::common::getSMVersion() < 100 || tensorrt_llm::common::getSMVersion() == 120)
     {
         return false;
     }

Original file line number	Diff line number	Diff line change
`@@ -671,7 +671,8 @@ bool attention_supports_nvfp4_output(int64_t const num_heads, int64_t const num_`
`671`	`671`	`bool const use_paged_context_fmha, bool is_mla_enable)`
`672`	`672`	`{`
`673`	`673`	`// Only Blackwell supports NVFP4 output.`
`674`		`- if (tensorrt_llm::common::getSMVersion() < 100)`
	`674`	`+ // SM 120 does not support NVFP4 output.`
	`675`	`+ if (tensorrt_llm::common::getSMVersion() < 100 \|\| tensorrt_llm::common::getSMVersion() == 120)`
`675`	`676`	`{`
`676`	`677`	`return false;`
`677`	`678`	`}`