Skip to content

perf(llama): auto-disable PCT binding on B300 for llama pretrain - #4081

Merged
malay-nagda merged 4 commits into
mainfrom
rsalagame/llama-pct-auto-disable-b300-main
Jun 2, 2026
Merged

perf(llama): auto-disable PCT binding on B300 for llama pretrain#4081
malay-nagda merged 4 commits into
mainfrom
rsalagame/llama-pct-auto-disable-b300-main

Conversation

@rsalagame-nvidia

@rsalagame-nvidia rsalagame-nvidia commented May 29, 2026

Copy link
Copy Markdown
Contributor

adds llama × pretrain × b300 to the existing PCT auto-disable list in setup_experiment.py, alongside existing nemotronh and deepseek.

@rsalagame-nvidia
perf(llama): auto-disable PCT binding on B300 for llama pretrain
a2ea50c 
Signed-off-by: Rahul Salagame <rsalagame@nvidia.com>
@rsalagame-nvidia rsalagame-nvidia self-assigned this May 29, 2026
@rsalagame-nvidia
rsalagame-nvidia requested review from a team, erhoo82 and malay-nagda as code owners May 29, 2026 23:00
@copy-pr-bot

copy-pr-bot Bot commented May 29, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@claude

claude Bot commented May 29, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

LGTM — the change is a clean, minimal addition to the existing PCT auto-disable guard. One observation:

Intentionally broad scope: The new branch matches model_family_name llama + task pretrain + gpu b300 without filtering on model_recipe_name or compute_dtype, unlike the nemotronh branch (which gates on compute_dtype bf16) and the deepseek branch (which gates on config_variant != large_scale). This is presumably intentional (all llama pretrain on B300 needs PCT disabled), but worth confirming since future llama recipes added to the family would silently inherit this behavior.

No bugs, typos, or missing docs found.

Suggested test cases — The change affects setup_experiment.py PCT binding logic. It impacts all llama x pretrain x b300 combinations at experiment setup time:

  • llama3_8b_8gpu_b300_bf16_perf
  • llama3_8b_8gpu_b300_fp8_cs_perf
  • llama3_8b_8gpu_b300_fp8_mx_perf
  • llama3_8b_8gpu_b300_nvfp4_perf
  • llama3_70b_64gpu_b300_bf16_perf
  • llama3_70b_64gpu_b300_fp8_cs_perf
  • llama3_70b_64gpu_b300_fp8_mx_perf
  • llama3_70b_64gpu_b300_nvfp4_perf
  • llama31_405b_128gpu_b300_bf16_perf
  • llama31_405b_128gpu_b300_fp8_cs_perf
  • llama31_405b_128gpu_b300_fp8_mx_perf
  • llama31_405b_128gpu_b300_nvfp4_perf

@yaoyu-33 yaoyu-33 added area:perf Performance optimizations and benchmarking feature New capabilities, enhancements, or enablement work needs-review PR is ready for code review and waiting on a reviewer labels May 29, 2026
malay-nagda
malay-nagda previously approved these changes May 30, 2026
View reviewed changes
yaoyu-33
yaoyu-33 previously approved these changes Jun 1, 2026
View reviewed changes
@yaoyu-33 yaoyu-33 added ready-to-merge PR is approved, current, and only waiting for CI to pass before merge and removed needs-review PR is ready for code review and waiting on a reviewer labels Jun 1, 2026
@yaoyu-33

yaoyu-33 commented Jun 1, 2026

Copy link
Copy Markdown
Contributor

/ok to test 2586005

@rsalagame-nvidia

rsalagame-nvidia commented Jun 1, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test 56f668f

@rsalagame-nvidia
style: apply ruff-format to PCT binding conditional
33965b9 
Signed-off-by: Rahul Salagame <rsalagame@nvidia.com>
@rsalagame-nvidia
rsalagame-nvidia dismissed stale reviews from yaoyu-33 and malay-nagda via 33965b9 June 1, 2026 19:19
@rsalagame-nvidia

rsalagame-nvidia commented Jun 1, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test 33965b9

@malay-nagda malay-nagda added the r0.5.0 Auto-cherrypick to release branch. Apply before merge; cherrypick happens after merge. label Jun 2, 2026
@malay-nagda
malay-nagda merged commit 5b0e8c4 into main Jun 2, 2026
45 checks passed
@malay-nagda
malay-nagda deleted the rsalagame/llama-pct-auto-disable-b300-main branch June 2, 2026 04:47
vasunvidia pushed a commit to vasunvidia/Megatron-Bridge that referenced this pull request Jun 10, 2026
@rsalagame-nvidia @vasunvidia
perf(llama): auto-disable PCT binding on B300 for llama pretrain (NVI…
c94f1ce 
…DIA-NeMo#4081)

Signed-off-by: Rahul Salagame <rsalagame@nvidia.com>
Signed-off-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>

This branch was previously deployed

1 inactive deployment
public — 33965b9f Deployed Jun 1, 2026 by copy-pr-bot[bot] via release / finalize / publish-docs #997
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@yaoyu-33 yaoyu-33 yaoyu-33 approved these changes
@malay-nagda malay-nagda malay-nagda approved these changes
@ko3n1g ko3n1g Awaiting requested review from ko3n1g
@bdubauski bdubauski Awaiting requested review from bdubauski
@chtruong814 chtruong814 Awaiting requested review from chtruong814
@erhoo82 erhoo82 Awaiting requested review from erhoo82

Assignees

@rsalagame-nvidia rsalagame-nvidia

Labels

area:perf Performance optimizations and benchmarking feature New capabilities, enhancements, or enablement work r0.5.0 Auto-cherrypick to release branch. Apply before merge; cherrypick happens after merge. ready-to-merge PR is approved, current, and only waiting for CI to pass before merge

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@rsalagame-nvidia @yaoyu-33 @malay-nagda