feat(distillation): add TransferQueue support for On-Policy Distillation#2580
Open
pthombre wants to merge 4 commits into
Open
feat(distillation): add TransferQueue support for On-Policy Distillation#2580pthombre wants to merge 4 commits into
pthombre wants to merge 4 commits into
Conversation
pthombre
changed the title
pthombre
force-pushed
the
pthombre/opd_tq_support
branch
from
b0e5794 to
3b9e56b
Compare
29 tasks
Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com>
pthombre
force-pushed
the
pthombre/opd_tq_support
branch
from
3b9e56b to
5ce45ca
Compare
Contributor
Author
|
/ok to test 5ce45ca |
pthombre
added
CI
Contributor
Author
|
/ok to test 5ce45ca |
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com>
Contributor
Author
|
/ok to test d2827ba |
Add the data_plane block to the distillation_math reference config so test_reference_configs_up_to_date matches the exemplar config that gained the TransferQueue section. Add unit tests for the pure helpers in distillation_sync (_dedupe_fields, _as_row_aligned_tensor, _aggregate_teacher_topk_transport_results, _packing_args_for_policy, _stamp_policy_pad_seqlen) and transport_metrics (tensor_nbytes, valid_token_tensor_nbytes) to raise patch coverage. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com>
Contributor
Author
|
/ok to test ff3a57c |
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do ?
Adds TransferQueue-backed on-policy distillation so teacher top-k logits can stay in the data plane instead of being returned through the driver.
This PR extends the data-plane support that already exists for GRPO to the distillation workflow. When
data_plane.enabled=true,examples/run_distillation.pynow builds the student policy as aTQPolicyand dispatches to a new TQ-aware distillation trainer. The legacy distillation path remains the default when the data plane is disabled.Key pieces added:
distillation_train_syncimplementation that orchestrates rollout, teacher top-k inference, and student training through TransferQueue metadata.DistillationRolloutActorthat runs generation/rollout processing and seeds the TQ partition with row-aligned distillation training tensors.get_topk_logits_preshardedsupport so teacher workers writeteacher_topk_logitsandteacher_topk_indicesback to TQ directly.TQPolicyextensions for custom partition fields, consumer task names, and custom train fetch fields so the same policy wrapper can serve GRPO and distillation.Issues
None.
Usage
Enable the TQ distillation path by using the new exemplar or recipe:
The recipe added in this PR can be launched through the test-suite wrapper:
Before your PR is "Ready for review"
Pre checks: