已合并
acl graph相关代码cleancode整改 #159
acl graph相关代码cleancode整改 #159
已合并
huanglei创建于 1月7日
huanglei
huanglei成员
1月7日

合入背景

aclgraph权重加载重构添加日志及注释
Fix part of #55

修改内容

aclgraph权重加载重构添加日志及注释

资料变更

不涉及

接口变更

不涉及

测试结果

QWEN3-pdmix场景,服务化成功,推理成功

  • Qwen3-4B_tieword

[2026-01-07 16:20:56,188] [19106] [281473048444960] [llmmodels] [INFO] [run_pa.py-434] : ---------------end inference---------------
[2026-01-07 16:20:56,188] [19106] [281473048444960] [llmmodels] [INFO] [run_pa.py-611] : Answer[0]:  What are the key components of deep learning? What are the applications of deep learning? What are the
[2026-01-07 16:20:56,188] [19106] [281473048444960] [llmmodels] [INFO] [run_pa.py-612] : Generate[0] token num: (0, 20)


  • Qwen3_14B_pdmix A3

[2026-01-07 16:19:20,991] [18604] [281473723600928] [llmmodels] [INFO] [run_pa.py-434] : ---------------end inference---------------
[2026-01-07 16:19:20,991] [18604] [281473723600928] [llmmodels] [INFO] [run_pa.py-611] : Answer[0]:  How is it different from traditional machine learning?

Deep learning is a subset of machine learning that uses neural
[2026-01-07 16:19:20,991] [18604] [281473723600928] [llmmodels] [INFO] [run_pa.py-612] : Generate[0] token num: (0, 20)

CheckList

PR提交人对以下CheckList自检项进行全量自检,自检通过或不涉及,均修改 [ ] 为 [x]。

likedislike
Pull Request已成功合入, 合并人@ascend-robot
(感谢 huanglei 的贡献)
huangleihuanglei成员
1月7日 关联了issue:[新需求]: 模型权重加载重构
ascend-robot
ascend-robot成员
1月7日 评论:

Thank your for your pull-request.

The full list of commands accepted by me can be found at here.

You can get sig-info at here

likedislike
ascend-robot
ascend-robot成员
1月7日 评论:

以下是根据您提交的修改文件推荐的Reviewer和Committer序列,需各模块评审通过后方可合入

Module List Reviewers Committers
repo-Ascend/MindIE-LLM syli1107, guanxinghua, zzhao2012, jiangkaiqiang, limuyuan31 wu_yushan, maoxx241, taojiovvo, you-zhiyuan-0511, masiming
likedislike
ascend-robotascend-robot成员
1月7日 添加了label:ascend-cla/yes
CXY-Katrina
CXY-Katrina成员1月7日进行代码检视2
examples/atb_models/atb_llm/utils/data/quant_method_adapter.py
已过期
@@ -137,0 +137,4 @@
137+ # For standard 2D weight tensors (shape [n, k]), when:
138+ # - Input dimension (k) is NOT 256-aligned (not is_k_divisible)
139+ # - Output dimension (n) IS 256-aligned (is_n_divisible)
140+ # - 3: 3D tensor has been pre-processed by NPU when matmul_nd_nz=True
141+ # NPU can handle non-aligned k via hardware padding (no transpose needed)
CXY-Katrina
CXY-Katrina1月7日评论:

此条代码评论区间+137+141

删除

likedislike
System
系统消息系统
1月7日 评论:

changed this line on f136e841 view diff detail

此处折叠了90条消息 查看更多
CXY-KatrinaCXY-Katrina成员
1月8日 解决了最后一个问题
ascend-robot
ascend-robot成员
1月8日 评论:

Notice

There are no committers approved this pull request. 😲

likedislike
CXY-Katrina
CXY-Katrina成员
1月8日 评论:

/check-pr

likedislike
ascend-robot
ascend-robot成员
1月8日 评论:

Review Guide

This Pull-Request Passes Review.
Committers who writed a comment of /approve are: Katrina-CXY.
Reviewers who writed a comment of /lgtm are: martinXuc, Katrina-CXY.

likedislike
ascend-robotascend-robot成员
1月8日 合入了pull request