Chranos/enginex-mlu370-vllm
1
0
Fork 0
forked from EngineX-Cambricon/enginex-mlu370-vllm
Code Pull Requests Activity

add deepseekv3 and llama4

Browse Source
This commit is contained in:
Chranos
2026-02-11 15:24:13 +08:00
parent 2ad23aa8da
commit c584139543
5 changed files with 320 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm-v0.6.2/vllm/model_executor/layers/layernorm.py
View File

@@ -143,11 +143,14 @@ class RMSNorm(CustomOp):
from vllm import _mlu_ops as mlu_ops
x = x.view(-1, self.weight.data.shape[0])
weight = self.weight.data
if weight.dtype != x.dtype:
weight = weight.to(x.dtype)
if residual is not None:
residual = residual.view(-1, self.weight.data.shape[0])
return mlu_ops.fused_rms_norm(x, residual, self.weight.data, None, None, self.variance_epsilon, True)
return mlu_ops.fused_rms_norm(x, residual, weight, None, None, self.variance_epsilon, True)
else:
return mlu_ops.fused_rms_norm(x, residual, self.weight.data, None, None, self.variance_epsilon, False)
return mlu_ops.fused_rms_norm(x, residual, weight, None, None, self.variance_epsilon, False)
def extra_repr(self) -> str:
s = f"hidden_size={self.weight.data.size(0)}"