NVIDIA / NVIDIA/apex

Performance with python build?

Open
#355 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
9k
Forks
1.5k
Avg merge
2d 4h
Merged PRs (30d)
3

Description

Is there some sense on how the python build performs, compared with the C/NCCL compile? I'm not getting an error, but I get the warning :

Warning:  multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback.  Original ImportError was: ModuleNotFoundError("No module named 'amp_C'")

So, it's still running. I guess I wanted to know much more performance I would be getting if I chased this down (our institution makes it difficult to troubleshoot sometimes). I have a fairly large (500GB) run across 6x V100's on our DGX1 on opt_level 01.
Ubuntu 16.04.6
Pytorch 1.1.0
Torchvision 0.3.0
CUDA 10.0

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing the reported Apex installation warning and the --cuda_ext --cpp_ext build options in the issue context. Reproduce the Python fallback and compare it with a C/NCCL extension build on the stated PyTorch, CUDA, and V100 setup; done means providing a measured performance comparison or a clear explanation of the difference.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.