Skip to content

Ilya/fireworks cookbook resume fixes - #672

Merged
shfunc merged 2 commits into
hud-evals:mainfrom
shfunc:ilya/fireworks-cookbook-resume-fixes
Sep 18, 2026
Merged

shfunc merged 2 commits into
hud-evals:mainfrom
shfunc:ilya/fireworks-cookbook-resume-fixes

Conversation

@shfunc

@shfunc shfunc commented Sep 18, 2026

Copy link
Copy Markdown
Contributor

Note

Medium Risk
Resume changes training client setup and when tokenizer/renderer load; wrong flags on non-default checkpoints could fail or mis-tokenize before training starts.

Overview
Resume and model resolution for the Fireworks RL cookbook: --resume-from now looks up the checkpoint’s base model via the Fireworks REST API and overrides CLI --base-model before tokenizer/renderer setup. Non-default models must pass both --tokenizer-model and --renderer (defaults apply only for Qwen 3.8 27B). The startup banner and config.json record the resolved model, tokenizer, and renderer.

Checkpoint visibility: each state-NNNN and final-state save prints Training checkpoint: <full path> immediately so paths are available even if a later step fails.

Tests and docs: new unit tests cover resume model override, missing rendering args on alternate bases, and checkpoint logging before failures; README and v6 cookbook docs describe the updated resume flow and CLI help.

Reviewed by Cursor Bugbot for commit 315cce4. Bugbot is set up for automated code reviews on this repo. Configure here.

@shfunc
shfunc merged commit 2802285 into hud-evals:main Sep 18, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants