Skip to content

[multi-lora] 1/n - 3, fix: launch-time guard for unsafe recompute configs - #2726

Closed
yushengsu-thu wants to merge 1 commit into
mainfrom
split/1n-3-recompute-guard
Closed

[multi-lora] 1/n - 3, fix: launch-time guard for unsafe recompute configs#2726
yushengsu-thu wants to merge 1 commit into
mainfrom
split/1n-3-recompute-guard

Conversation

@yushengsu-thu

Copy link
Copy Markdown
Collaborator

Full recompute, and selective MoE recompute with expert LoRA targets, silently corrupt multi-LoRA gradients unless the installed Megatron-Bridge PEFT input-gradient patch recognizes .adapters. parameters (Megatron-Bridge#27). This probes the Bridge source at arg-validation time and refuses unsupported shapes at launch instead of burning GPU time; patched Bridge versions pass through.

Extracted from #2273; the stack rebases after this merges.

Full recompute, and selective MoE recompute with expert LoRA targets,
silently corrupt gradients unless the Megatron-Bridge PEFT
input-gradient patch recognizes .adapters. parameters. Probe the
installed Bridge source at validation time and fail the launch for
unsupported shapes.

@claude claude Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

Tip: disable this comment in your organization's Code Review settings.

@yushengsu-thu yushengsu-thu added run-ci-megatron run-ci-lora run-ci-lora-native Run native (raw-mode) LoRA plugin e2e tests labels Aug 24, 2026
@yushengsu-thu

Copy link
Copy Markdown
Collaborator Author

CI triage: stage-c-4-gpu-mi350 (1) failure is test_amd_mtp1_spec_v2_r3.py timing out at 2400s — identical signature on main's ROCm nightly (run 32647493780) and on #2724, i.e. pre-existing MI350 baseline noise, unrelated to this PR.

@yushengsu-thu

Copy link
Copy Markdown
Collaborator Author

Closing: the deployment now tracks Megatron-Bridge @bridge (which carries #27), so the launch-time guard against pre-#27 bridges is retired from the stack; the LayerWise dependency canary (CI-level) remains the stale-image guard.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

run-ci-lora run-ci-lora-native Run native (raw-mode) LoRA plugin e2e tests run-ci-megatron

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant