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 15:40:19 +08:00
parent f3a4d10195
commit 7626238695
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm-v0.6.2/vllm_mlu/vllm_mlu/model_executor/layers/linear.py
View File

@@ -26,9 +26,12 @@ def vllm__module_executor__layers__linear__UnquantizedLinearMethod__apply(
beta = 1.0
residual = residual.view(-1, residual.shape[-1])
res_shape = x.shape[0:-1] + (layer.weight.shape[0], )
# MLU matmul requires matching dtypes; cast input to weight dtype
if x.dtype != layer.weight.dtype:
x = x.to(layer.weight.dtype)
# MLU matmul requires all tensors to have matching dtypes
target_dtype = layer.weight.dtype
if x.dtype != target_dtype:
x = x.to(target_dtype)
if residual is not None and residual.dtype != target_dtype:
residual = residual.to(target_dtype)
return mlu_ops.matmul(x.view(-1, x.shape[-1]), layer.weight, bias, residual, 'none', 1.0, beta).view(res_shape)