[ET-VK] TagMemoryMetaPass skips a tensor that sits in a list alongside None
维护者通常 1 天内回复
评估
这个 Issue 还没有评估数据。
描述
🐛 Describe the bug
is_tensor_arg_node() and is_non_constant_tensor_node() in backends/vulkan/utils.py only recognise a list argument when every entry is a tensor node. An argument that mixes tensor nodes with None is therefore treated as a non-tensor and skipped entirely by TagMemoryMetaPass, so the tensors inside it never get a storage or memory-layout representation.
index.Tensor is the case that surfaces it: x[:, :, idx] reaches the edge dialect as index.Tensor(x, [None, None, idx]), and idx lives in a list whose other entries are None. Any op implementation that then asserts on the layout of that tensor fails, because nothing assigned it one.
This is easy to miss because the op that exposed it can be routed around. #22481 makes index.Tensor work on any axis by forcing CONTIGUOUS_BUFFER for higher-rank self, which sidesteps the tagging gap rather than closing it. The gap is still there for the next operator that takes an optional-tensor list.
A fix is a helper along the lines of
def tensor_nodes_in_arg(arg):
"""The tensor nodes of an argument, skipping Nones."""
used everywhere the tagging pass walks a list argument. Lists that are entirely tensors, which is every other operator today, would behave exactly as before.
Originally found while working on #22406, which is being closed in favour of #22481; filing separately so the finding is not lost with it.
Versions
Reproduces on main (a6b115bff6).
cc @SS-JIA @manuelcandales @digantdesai @cbilgin
- 主要语言
- Python
- 星标
- 5.1k
- 派生
- 1.2k
- 平均合并
- 2 天 48 分钟
- 30 天内合并 PR
- 693
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
pytorch/executorch 的其他 Issue
-
Portable remainder kernel: wrong sign for integers and float64 rounded to float可能已有人在做 @JakeStevens 于 6 天前认领。 未关闭bug module: kernels triaged
难度 2/5 1-3 小时 新手友好度 82/100
pytorch/executorch#23682 · 已指派 2 人 ·
维护者通常 1 天内回复
-
enhancement triaged
难度 2/5 半天 新手友好度 68/100
pytorch/executorch#21640 ·
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 40/100
pytorch/executorch#23955 ·
维护者通常 1 天内回复
-
[QNN] 16a4w_block stride-2 conv + ReLU fails HTP graph prepare on SM8850 (Relu16bParamsPlaceholder)可能已有人在做 @psiddh 于 2 天前认领。 未关闭bug module: qnn partner: qualcomm
pytorch/executorch#23922 · 已指派 1 人 ·
维护者通常 1 天内回复
-
A refused set_inputs() leaves a method runnable with mixed inputs, and get_outputs() returns stale results可能已有人在做 @shoumikhin 今天认领。 未关闭
难度 4/5 3-5 天 新手友好度 40/100
pytorch/executorch#23913 · 1 条评论 ·
维护者通常 1 天内回复
查看 pytorch/executorch 的全部 Issue
相似的 Issue
-
comp/cli P3 type/security
难度 2/5 1-3 小时 新手友好度 68/100
NousResearch/hermes-agent#141420 · 1 条评论 ·
维护者通常 1 天内回复
-
vendor-pitch
难度 2/5 1-3 小时 新手友好度 87/100
维护者通常 1 天内回复
-
bug llm p2
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
-
New tool: ZIPZAX Python Notebook Studio可能已有人在做 @github-actions 今天认领。 未关闭submission triage
难度 1/5 1 小时以内 新手友好度 68/100
kausalflow/tools#226 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100