EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[0.11.0]Chery pick pta upgrade change (#3940)

Browse Source 
This PR cherry-pick two commit from main to upgrade torch-npu to 2.7.1
official release

---------

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-10-31 22:14:26 +08:00
committed by GitHub
parent 3d81ea03ed
commit 8a7154001e
16 changed files with 93 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/ops/weight_prefetch.py
View File

@@ -7,7 +7,6 @@ from vllm.forward_context import get_forward_context
from vllm_ascend.ascend_config import WeightPrefetchConfig
from vllm_ascend.ops.linear import (AscendQKVParallelLinear,
AscendRowParallelLinear)
from vllm_ascend.utils import version_check
SUPPORTED_MODULES = ["attn", "mlp", "moe"]
MOE_PREFETCH_TOKEN_THRESHOLD = 96
@@ -83,8 +82,7 @@ class WeightPrefetchMethod:
if not self.moe.is_active_this_forward:
return
forward_context = get_forward_context()
if not version_check():
forward_context.layer_idx += 1
# layer_idx is subtracted by 1 because layer_idx was incremented by 1 at layernorm.
weight = forward_context.model_instance.model.layers[
forward_context.layer_idx - 1].mlp.experts.w13_weight
weight_size = weight.data.element_size() * weight.data.numel(