Conversation
tjruwase
reviewed
Jun 30, 2026
tjruwase
reviewed
Jun 30, 2026
tjruwase
reviewed
Jun 30, 2026
Contributor
Author
|
Hi @tjruwase I have changed according to your first and second comments. For the third comment, I left an explaination. Thanks! |
Entry point, configs, data, and tests for on-policy distillation using DeepSpeed's hybrid engine rollout and vLLM backend. Signed-off-by: Guokai Ma <[email protected]> Signed-off-by: Guokai Ma <[email protected]>
…m_dtype to engine_dtype Signed-off-by: Guokai Ma <[email protected]>
…DeepSpeed - Delete vLLM configs, scripts (opsd_vllm_disjoint.json, smoke_vllm.json, train_opsd_vllm.sh) - Add trainer.py, config.py, losses.py, utils.py (moved from DeepSpeed) - Add benchmarks/ (5 hybrid engine benchmarks moved from DeepSpeed) - Update main.py imports (trainer, config now local) - Update test imports (losses, utils now local) - Rewrite README (remove all vLLM sections) Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
…ants from JSON - Subclass DeepSpeed's RolloutConfig to add temperature/top_p/etc - Remove weight_sync_interval from JSON configs (vLLM remnant) Signed-off-by: Guokai Ma <[email protected]>
…per fix - Remove 14B/multi-GPU benchmarks (bench_14b_rollout, bench_autotp_gc, bench_hybrid_tp, bench_hybrid_tp_opt) - Fix bench_decode_1p1r: wrap model for HybridEngineRollout - Add --graph-capture CLI flag Signed-off-by: Guokai Ma <[email protected]>
…nfigs temperature=0 causes logits/0 = inf → NaN loss. The correct default for knowledge distillation is temperature=1.0 (standard softmax). Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
tjruwase
approved these changes
Jul 12, 2026
PKUWZP
self-requested a review
July 12, 2026 16:30
PKUWZP
requested changes
Jul 12, 2026
PKUWZP
left a comment
PKUWZP
left a comment
Contributor
There was a problem hiding this comment.
Overall this PR looks good. My only suggestion is that we remove all "Microsoft Corporation Copyright" from all files in this PR. After that we should be able to merge it.
Contributor
Author
… at TP=1 / under ZeRO-3 Signed-off-by: Guokai Ma <[email protected]>
…xample Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
…es review) Signed-off-by: Guokai Ma <[email protected]>
…loading, else stage 0 Signed-off-by: Guokai Ma <[email protected]>
…lease Signed-off-by: Guokai Ma <[email protected]>
…fig knob Signed-off-by: Guokai Ma <[email protected]>
…plicitly Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Signed-off-by: Guokai Ma <[email protected]>
Contributor
Author
Now deepspeed is always used for teacher no matter offload is true or not. And all Microsoft file head had been fixed. I plan to enable tensor parallel in a sperate PR, which is useful in OPSD. This add some complexity so better put in a seperate PR. |
PKUWZP
approved these changes
Jul 14, 2026
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
4 participants
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.
This PR moved example code from @PKUWZP 's OPSD PR in DeepSpeed (deepspeedai/DeepSpeed#8027).
Note the basic OPSD infrastructures (OPSD trainer, rollout engines) are still in the original PR and need to be merged seperately. This PR will work when the original PR merged.