Skip to content

Add CI rollout-data injection with recorded-data metadata round-trip#1449

Open
fzyzcjy wants to merge 1 commit into
tom/pr_chain/trainer_ft/dev_revert_reversed/add-ft-test-action-hooks-to-the-train-groupfrom
tom/pr_chain/trainer_ft/dev_revert_reversed/add-ci-rollout-data-injection-with-recorded-data-metadata-round-trip
Open

Add CI rollout-data injection with recorded-data metadata round-trip#1449
fzyzcjy wants to merge 1 commit into
tom/pr_chain/trainer_ft/dev_revert_reversed/add-ft-test-action-hooks-to-the-train-groupfrom
tom/pr_chain/trainer_ft/dev_revert_reversed/add-ci-rollout-data-injection-with-recorded-data-metadata-round-trip

Conversation

@fzyzcjy

@fzyzcjy fzyzcjy commented Jun 22, 2026

Copy link
Copy Markdown
Collaborator

Adds RolloutDataInjectionUtil (debug_data.py) so CI can replace generated rollout data with a recorded baseline at/after a start rollout id, asserting the freshly generated responses still match the recording above a configurable token-match ratio (catching weight divergence beyond ulp drift). Threads recording metadata through save/load (save_debug_rollout_data metadata= / load returns (data, metadata) / _load_rollout_data_file) and wires injection into RolloutManager._get_rollout_data. Includes the debug_data unit tests for the metadata round-trip and the injection match logic.

@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ft-test-action-hooks-to-the-train-group branch from ed04895 to 20e80c1 Compare June 23, 2026 07:51
@fzyzcjy fzyzcjy requested a review from yushengsu-thu as a code owner June 23, 2026 07:51
@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ci-rollout-data-injection-with-recorded-data-metadata-round-trip branch from dae8baf to fc4b9b8 Compare June 23, 2026 07:51
@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ft-test-action-hooks-to-the-train-group branch from 20e80c1 to fcef2b4 Compare June 23, 2026 09:30
@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ci-rollout-data-injection-with-recorded-data-metadata-round-trip branch from fc4b9b8 to b0a9e4e Compare June 23, 2026 09:30
Adds RolloutDataInjectionUtil (debug_data.py) so CI can replace generated rollout data with a recorded baseline at/after a start rollout id, asserting the freshly generated responses still match the recording above a configurable token-match ratio (catching weight divergence beyond ulp drift). Threads recording metadata through save/load (save_debug_rollout_data metadata= / load returns (data, metadata) / _load_rollout_data_file) and wires injection into RolloutManager._get_rollout_data. Includes the debug_data unit tests for the metadata round-trip and the injection match logic.
@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ft-test-action-hooks-to-the-train-group branch from fcef2b4 to 319df37 Compare June 23, 2026 13:34
@fzyzcjy fzyzcjy force-pushed the tom/pr_chain/trainer_ft/dev_revert_reversed/add-ci-rollout-data-injection-with-recorded-data-metadata-round-trip branch from b0a9e4e to b6e3ee2 Compare June 23, 2026 13:34
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.

1 participant