@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / e6d811dd13
Fork 0

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 16, 18:15 UTC

[XPU] fallback to native implementation for xpu custom op (#7670)

jianyizh committed Aug 20, 2024 at 15:26 UTC e6d811dd13a0ad81a1efc67c97d8812e823db5a7
1 file changed +3 -1
vllm/model_executor/custom_op.py
+3 -1
@@ -30,7 +30,9 @@ class CustomOp(nn.Module):
30 return self.forward_cuda(*args, **kwargs)
31
32 def forward_xpu(self, *args, **kwargs):
33 - raise NotImplementedError
33 + # By default, we assume that XPU ops are compatible with the
34 + # PyTorch-native implementation.
35 + return self.forward_native(*args, **kwargs)
36
37 def forward_cpu(self, *args, **kwargs):
38 # By default, we assume that CPU ops are compatible with CUDA ops.