torch-npu-pyyaml-preinstall

name	torch-npu-pyyaml-preinstall
description	Pre-install PyYAML to prevent torch-npu transitive dependency failure
tags	["torch-npu","pyyaml","transitive-dependency","undeclared-requirement","dependency_issue"]
category	dependency_issue
subtype	undeclared_transitive_dependency
confidence	0.95
occurrence_count	1

Pre-install PyYAML to prevent torch-npu transitive dependency failure

When to Use

• Error involves: import torch # fails with ModuleNotFoundError: No module named 'yaml' when torch-npu is installed but PyYAML is not, import torch_npu # fails at torch_npu/npu/_memory_viz.py line 11

Root Cause

torch-npu 2.5.1's memory visualization submodule (torch_npu/npu/_memory_viz.py) imports yaml unconditionally at module load time. However, PyYAML is not declared in torch-npu's pip metadata (requires field), so a fresh virtual environment with only torch-npu installed will lack it. The error is misleading because the traceback originates from import torch (not import torch_npu), making the actual missing dependency hard to diagnose.

How to Use

1. 1. Before or alongside installing torch-npu, explicitly install PyYAML: pip install PyYAML (use domestic mirror if needed: pip install PyYAML -i https://mirrors.aliyun.com/pypi/simple/).
2. 2. Alternatively, add PyYAML to the project's requirements.txt alongside torch, torch-npu, and torchvision.
3. 3. After installing torch-npu, run a pre-flight validation: python -c 'import torch_npu; import torch_npu.npu' — if this exits cleanly without ModuleNotFoundError, the dependency is resolved.
4. 4. If the error still occurs, verify the yaml module is accessible: python -c 'import yaml; print(yaml.__version__)'.

Code Examples

File: requirements.txt

Before

torch torch-npu torchvision

After

torch torch-npu torchvision PyYAML

Do Not

• Do NOT attempt to patch torch_npu internal files (_memory_viz.py) — the file is overwritten on package upgrade.
• Do NOT confuse this error with a torch installation issue — the traceback starting at 'import torch' is a red herring; the actual failure is in torch_npu backend loading.
• Do NOT use CPU fallback as a workaround — this is a hard dependency failure, not a performance issue.

References

• https://ascend.github.io/docs/tutorials/torch_npu/

Evidence

• Source runs: e2e-v2-69fe6c6573b0
• Source runs: e2e-v2-c3c4a55c9cd4-exp-20260429211243