[TRTLLM-6222][feat] Extend cute_dsl_nvfp4_gemm to sm103. (#9543)

Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com>

ATTRIBUTIONS-Python.md

@@ -25250,7 +25250,7 @@ License: `NVIDIA Proprietary Software`
   - `Homepage`: https://developer.nvidia.com/cusparselt
 
 
-## nvidia-cutlass-dsl (4.3.0)
+## nvidia-cutlass-dsl (4.3.1)
 
 ### Licenses
 License: `None`

requirements.txt

@@ -69,7 +69,7 @@ triton==3.5.0; platform_machine == "x86_64"
 tiktoken
 blobfile
 openai-harmony==0.0.4
-nvidia-cutlass-dsl==4.3.0; python_version >= "3.10"
+nvidia-cutlass-dsl==4.3.1; python_version >= "3.10"
 plotly
 numexpr<2.14.0 # WAR for attempted use of nonexistent numpy.typing
 partial_json_parser

tensorrt_llm/_torch/custom_ops/cute_dsl_custom_ops.py

@@ -49,9 +49,9 @@ if IS_CUTLASS_DSL_AVAILABLE:
             self.output_dtype = output_dtype
             assert output_dtype == torch.bfloat16
 
-            if get_sm_version() != 100:
+            if get_sm_version() not in [100, 103]:
                 raise ValueError(
-                    f"SM version {get_sm_version()} is not supported for {self.__class__.__name__}, it only supports SM 100"
+                    f"SM version {get_sm_version()} is not supported for {self.__class__.__name__}, it only supports SM 100 and SM 103"
                 )
 
         # rewrite the hash function because the value of self.alpha doesn't affect the tactic.

tests/unittest/_torch/thop/parallel/test_fp4_linear.py

@@ -87,8 +87,8 @@ def test_fp4_linear(dtype, mnk):
 @pytest.mark.skipif(sys.version_info < (3, 12),
                     reason="cutlass-dsl 4.1.0 requires Python 3.12+")
 @pytest.mark.skipif(
-    get_sm_version() != 100,
-    reason="This test is only supported in Blackwell architecture",
+    get_sm_version() not in [100, 103],
+    reason="This test is only supported in sm100 and sm103 architecture",
 )
 @pytest.mark.skipif(not IS_CUTLASS_DSL_AVAILABLE,
                     reason="cutlass-dsl is not available")