Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add performance-optimized example for llama2 70b LoRA #12055

Merged
merged 16 commits into from
Feb 11, 2025

Conversation

vysarge
Copy link
Collaborator

@vysarge vysarge commented Feb 5, 2025

What does this PR do ?

This PR adds an example script for llama2 70b LoRA tuning (tp4pp1cp1 on H100) including optimized configs used in MLPerf 4.1 submission, plus supporting changes.

Requires #11961.

Collection: llm

Changelog

  • Added example script scripts/llm/performance/mlperf_lora_llama2_70b.py
  • Enabled use of atomic GEMM settings for TP overlap
  • Added a custom datamodule for downloading MLPerf-processed govreport data from HF
  • Added a wrapper around LlamaModel which wraps configure_model to reduce memory consumption

Usage

  • You can potentially add a usage example below
python scripts/llm/performance/mlperf_lora_llama2_70b.py --help

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

  • Make sure you read and followed Contributor guidelines
  • Did you write any new necessary tests?
  • Did you add or update any necessary documentation?
  • Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
    • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

  • New Feature
  • Bugfix
  • Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

  • Related to # (issue)

@vysarge vysarge force-pushed the vsarge/lora_recipe_perf branch from b01cbc9 to 3b9525a Compare February 5, 2025 00:55
Comment on lines 271 to 281
"LORA_A2A": "1",
"CUDA_DEVICE_MAX_CONNECTIONS": "1",
"CUBLAS_FORCE_XMMA_KERNEL_INIT": "DEVICE",
"CUDNN_FRONTEND_ATTN_DP_WORKSPACE_LIMIT": "0",
"NVTE_FP8_DPA_BWD": "1",
"NVTE_RS_STRIDED_ATOMIC": "2",
"NCCL_MIN_CTAS": "32",
"NCCL_MIN_P2P_NCHANNELS": "32",
"NCCL_NCHANNELS_PER_NET_PEER": "32",
"NCCL_NVLS_ENABLE": "0",
"TORCH_NCCL_AVOID_RECORD_STREAMS": "1",
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you add description on the env vars?

Please refer this PR.
#11991

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

vysarge and others added 14 commits February 6, 2025 16:47
…ludes te's fp8_model_init and torch no_grad, update TE overlap configs to allow atomic gemm + empty entries

Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
Signed-off-by: Valerie Sarge <[email protected]>
@vysarge vysarge force-pushed the vsarge/lora_recipe_perf branch from 8211fd6 to 291b2f5 Compare February 7, 2025 00:47
erhoo82
erhoo82 previously approved these changes Feb 7, 2025
Copy link
Collaborator

@erhoo82 erhoo82 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

malay-nagda
malay-nagda previously approved these changes Feb 7, 2025
Copy link
Collaborator

@malay-nagda malay-nagda left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

Signed-off-by: Valerie Sarge <[email protected]>
@vysarge vysarge dismissed stale reviews from malay-nagda and erhoo82 via 47ba6ae February 7, 2025 17:03
Signed-off-by: Valerie Sarge <[email protected]>
@vysarge vysarge added Run CICD and removed Run CICD labels Feb 7, 2025
Copy link

codecov bot commented Feb 8, 2025

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 30.43%. Comparing base (fdfeafe) to head (d01a099).
Report is 11 commits behind head on main.

Additional details and impacted files
@@            Coverage Diff             @@
##             main   #12055      +/-   ##
==========================================
+ Coverage   30.42%   30.43%   +0.01%     
==========================================
  Files        1400     1402       +2     
  Lines      177452   177665     +213     
  Branches    27196    27221      +25     
==========================================
+ Hits        53990    54075      +85     
- Misses     119345   119465     +120     
- Partials     4117     4125       +8     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

@vysarge vysarge merged commit 4e7478b into NVIDIA:main Feb 11, 2025
222 of 223 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants