From 6534d2fc9773db101e0cb6d2bd9617bfd41e7876 Mon Sep 17 00:00:00 2001 From: zifeitong Date: Tue, 12 Aug 2025 12:54:42 -0700 Subject: [PATCH] Fix torch version check for SM100 mxfp4 (#22535) Signed-off-by: Zifei Tong Signed-off-by: mgoin Co-authored-by: mgoin --- vllm/model_executor/layers/fused_moe/layer.py | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py index d5a89655e36d6..fb38fb91ead6d 100644 --- a/vllm/model_executor/layers/fused_moe/layer.py +++ b/vllm/model_executor/layers/fused_moe/layer.py @@ -741,12 +741,14 @@ class FusedMoE(torch.nn.Module): # 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):