已合并
[feat] Update DVM r2.11 #43991
SorryNaCN创建于 8月7日
[feat] Update DVM r2.11 #43991
已合并
Pull Request已成功合入, 合并人@ascend-robot
(感谢 SorryNaCN 的贡献)ascend-robot
8月7日 评论:
8月7日 评论:
8月7日 添加了label:ascend-cla/yes
8月7日 添加了label:needs-issue
ascend-robot
8月7日 评论:
8月7日 评论:
Thanks for your pull-request.
The full list of commands accepted by me can be found at here.
You can get sig-info at here.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
PR Approval Progress
✅ Congratulations! All modules have met the lgtm and approve requirements.
Module Approval Details
| module | lgtm status | approve status |
|---|---|---|
| repo-Ascend/pytorch | ✅ zichun_ye, 楚浩田 (2/2) | ✅ zichun_ye, 楚浩田 (2/1) |
| test | ✅ zichun_ye, 楚浩田 (2/2) | ✅ zichun_ye, 楚浩田 (2/1) |
| torch_npu/_inductor | ✅ 楚浩田, hbhu_bin, zichun_ye (3/2) | ✅ zichun_ye, 楚浩田 (2/1) |
| torch_npu/csrc/inductor | ✅ 楚浩田, hbhu_bin, zichun_ye (3/2) | ✅ zichun_ye, 楚浩田 (2/1) |
💡 Tip:
- Committer can comment
/approveor/lgtm- Commenting
/approveimplies both code review (lgtm) and intent to merge (approve)
CLA Signature Pass
SorryNaCN, thanks for your pull request. All authors of the commits have signed the CLA. 👍


此处折叠了537条消息 查看更多
29 天前 添加了label:lgtm
29 天前 关闭了关联的issue
29 天前 合入了pull request
ascend-robot
29 天前 评论:
29 天前 评论:
Pull Request 已合并或已关闭。
If you want to solve this problem, you can click here to do it in the FAQs.


AtlasAccount
29 天前 评论:
29 天前 评论:
流水线 pytorch_gitcode_PR_multiVersion#15844 [ commitID:5f8ec12a ] 已完成


【合入来源】
Fixes https://gitcode.com/Ascend/pytorch/issues/5234
【修改方案】
bf16_vector_keep_promoted=True时执行。【资料变更】
不涉及。
【接口变更】
DVM 配置新增(在加载 DVM backend 前设置):
enable_decomp_list = []、disable_decomp_list = []:接受算子重载或重载包,禁用优先;启用时优先使用dvm_decompositions表中的 DVM 专用实现,例如enable_decomp_list=[torch.ops.npu.npu_rms_norm_backward]可开启 RMSNorm backward 分解;未指定时保持原默认行为。matmul_fusion_rule = None:可设置(n, m, k)回调,对应(..., m, k) @ (..., k, n)。仅显式返回False时否决 MM/BMM/ADDMM/BADDBMM 融合,其他返回值沿用原规则;不会绕过总开关或形状/类型限制,K=1 的逐元素快捷路径保留原有 lowering,不额外插入回调判断。【功能验证】
本次配置变更验证:
test_npu_rms_norm验证默认 forward 分解;test_rms_norm_decomp_disabled在独立进程中同时禁用 forward、启用 backward 分解,无全局配置赋值,测试 helper 保持原实现。check_k1_fusion和 K=1 分支中的重复回调判断,集中保留mm_rule的否决规则;清理无关注释,测试 helper_run_and_get_code_with_dvm保持原实现。.default均可显式启用,禁用优先;910B3 / CANN 9.0.1 上启用后生成 DVM kernel,无原生 backward 调用,数值与 eager 一致。原禁用分解用例复测通过。py_compile、git diff --check通过。torch_npu 2.9.0.post7+git7148a13二进制;未重新构建最新 DVM wheel,CI 状态另行确认。【CheckList】