Fix torch version check for SM100 mxfp4 (vllm-project#22535)

Signed-off-by: Zifei Tong <[email protected]>
Signed-off-by: mgoin <[email protected]>
Co-authored-by: mgoin <[email protected]>
Signed-off-by: Xiao Yu <[email protected]>

Author: 2 people

vllm/model_executor/layers/fused_moe/layer.py

@@ -741,12 +741,14 @@ def __init__(
 
         # we padding globally so EP buffer allocation works
         if quant_config and quant_config.get_name() == "mxfp4":
-            if not is_torch_equal_or_newer("2.8.0"):
-                raise RuntimeError("Mxfp4 on hopper requires torch >= 2.8.0")
-            if current_platform.is_device_capability(
-                    90) and not has_triton_kernels():
-                raise NotImplementedError(
-                    "Triton kernels must be installed for mxfp4 on hopper")
+            if not current_platform.is_device_capability(100):
+                if not is_torch_equal_or_newer("2.8.0"):
+                    raise RuntimeError(
+                        "Mxfp4 on non-blackwell requires torch >= 2.8.0")
+                if not has_triton_kernels():
+                    raise NotImplementedError(
+                        "triton_kernels must be installed for "
+                        "mxfp4 on non-blackwell")
             if (current_platform.is_rocm()
                     or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
                     or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):