fix: enable fp4 compilation on cu128 (sgl-project#5286)

zhyncs · jimoosciuc · commit d055579ba90a · 2025-04-17T12:00:20.000+08:00
diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
@@ -145,7 +145,7 @@ if (SGL_KERNEL_ENABLE_FP8)
     )
 endif()
 
-if (SGL_KERNEL_ENABLE_FP4)
+if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_FP4)
     list(APPEND SGL_KERNEL_CUDA_FLAGS
         "-DENABLE_NVFP4=1"
     )
@@ -253,10 +253,7 @@ if (BUILD_FA3)
         FLASHATTENTION_DISABLE_SM8x
         FLASHATTENTION_DISABLE_BACKWARD
         FLASHATTENTION_DISABLE_DROPOUT
-        # FLASHATTENTION_DISABLE_ALIBI
-        # FLASHATTENTION_DISABLE_SOFTCAP
         FLASHATTENTION_DISABLE_UNEVEN_K
-        # FLASHATTENTION_DISABLE_LOCAL
         FLASHATTENTION_VARLEN_ONLY
     )
 endif()

Original file line number	Diff line number	Diff line change
`@@ -145,7 +145,7 @@ if (SGL_KERNEL_ENABLE_FP8)`
`145`	`145`	`)`
`146`	`146`	`endif()`
`147`	`147`
`148`		`-if (SGL_KERNEL_ENABLE_FP4)`
	`148`	`+if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_FP4)`
`149`	`149`	`list(APPEND SGL_KERNEL_CUDA_FLAGS`
`150`	`150`	`"-DENABLE_NVFP4=1"`
`151`	`151`	`)`
`@@ -253,10 +253,7 @@ if (BUILD_FA3)`
`253`	`253`	`FLASHATTENTION_DISABLE_SM8x`
`254`	`254`	`FLASHATTENTION_DISABLE_BACKWARD`
`255`	`255`	`FLASHATTENTION_DISABLE_DROPOUT`
`256`		`- # FLASHATTENTION_DISABLE_ALIBI`
`257`		`- # FLASHATTENTION_DISABLE_SOFTCAP`
`258`	`256`	`FLASHATTENTION_DISABLE_UNEVEN_K`
`259`		`- # FLASHATTENTION_DISABLE_LOCAL`
`260`	`257`	`FLASHATTENTION_VARLEN_ONLY`
`261`	`258`	`)`
`262`	`259`	`endif()`