A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
Fix torch version check for SM100 mxfp4 (#22535)
Signed-off-by: Zifei Tong <zifeitong@gmail.com> Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: mgoin <mgoin64@gmail.com>
zifeitong committed
Aug 12, 2025 at 12:54 UTC
6534d2fc9773db101e0cb6d2bd9617bfd41e7876
1 file changed
+8
-6
vllm/model_executor/layers/fused_moe/layer.py
+8
-6
@@ -741,12 +741,14 @@ class FusedMoE(torch.nn.Module):
741
742
# we padding globally so EP buffer allocation works
743
if quant_config and quant_config.get_name() == "mxfp4":
744
- if not is_torch_equal_or_newer("2.8.0"):
745
- raise RuntimeError("Mxfp4 on hopper requires torch >= 2.8.0")
746
- if current_platform.is_device_capability(
747
- 90) and not has_triton_kernels():
748
- raise NotImplementedError(
749
- "Triton kernels must be installed for mxfp4 on hopper")
744
+ if not current_platform.is_device_capability(100):
745
+ if not is_torch_equal_or_newer("2.8.0"):
746
+ raise RuntimeError(
747
+ "Mxfp4 on non-blackwell requires torch >= 2.8.0")
748
+ if not has_triton_kernels():
749
+ raise NotImplementedError(
750
+ "triton_kernels must be installed for "
751
+ "mxfp4 on non-blackwell")
752
if (current_platform.is_rocm()
753
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
754
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16):