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 13, 09:30 UTC
Fix pre-commit (and XPU) on `main` (#28556)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Nov 12, 2025 at 14:13 UTC
54aecd9ed55d65d1f7be9d14e4e82751663438c8
2 files changed
+1
-3
vllm/model_executor/layers/quantization/mxfp4.py
+1
-1
@@ -1145,7 +1145,7 @@ class IpexMxfp4MoEMethod(Mxfp4MoEMethod):
1145
) -> torch.Tensor:
1146
assert activation == "swigluoai", (
1147
"Only swiglu_oai activation is supported for IPEX MXFP4 MoE"
1148
- ) # noqa:
1148
+ )
1149
hidden_size_pad = round_up(self.original_hidden_size, 128)
1150
x_pad = torch.nn.functional.pad(x, (0, hidden_size_pad - x.size(-1)))
1151
hidden_states = layer.ipex_fusion(
vllm/platforms/xpu.py
-2
@@ -64,8 +64,6 @@ class XPUPlatform(Platform):
64
65
if use_sparse:
66
raise NotImplementedError("Sparse Attention is not supported on XPU.")
67
- if not use_v1:
68
- raise ValueError("XPU backend only supports V1.")
67
if selected_backend == AttentionBackendEnum.TRITON_ATTN:
68
logger.info_once("Using Triton backend.")
69
return AttentionBackendEnum.TRITON_ATTN.get_path()