mlcommons / mlcommons/chakra

Kineto collection of nccl:coalesced don't record metadata

Open
#178 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
196
Forks
84
PR merge metrics
No merged PRs in 30d

Description

Chakra's kineto trace add metadata to the NCCL kernels, including message size, PG attributes (name, description, ranks, etc).
However in a trace I just recorded some kernels don't include any metadata.

The trace records gpt3/175b_fp8 training script from NeMo Launcher.

Here is the report of a normal all_reduce operation, we can see info about PG.
image

While here is the report of the nccl coalesced allgather, which displays no metadata:
image

Contributor guide

Open the contributing guide

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 reproducing the Chakra kineto trace from the gpt3/175b_fp8 training script and compare a normal all_reduce with the nccl coalesced allgather. Trace the metadata collection path for both operations; done means coalesced NCCL kernels include the expected message size and process-group metadata.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, observability-sre
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.