Skip to content

Collect bounded external-draft LM-head inputs - #6

Closed
0z5a wants to merge 21 commits into
feature/external-draft-feature-contractfrom
feature/external-draft-collector-20260927
Closed

0z5a wants to merge 21 commits into
feature/external-draft-feature-contractfrom
feature/external-draft-collector-20260927

Conversation

@0z5a

@0z5a 0z5a commented Oct 1, 2026

Copy link
Copy Markdown
Owner

Incremental review: this fork PR shows only B2, based on the cleaned B1 #452 branch. The upstream submission is #454; its Files changed tab includes B1.

Add the disabled-by-default, bounded LM-head-input collector, owned feature payload, immutable head snapshot and rollout provenance. This layer adds no draft optimizer or weight publication.

Validation Result
Contract/provenance/collector checks 18 passed, including RTX 5090 CUDA reconstruction
Existing argument checks 22 passed
Qwen3-0.6B full E2E, off/on Both passed
Target/head progression Six distinct versions 1–6
Fixed replay Loss, entropy, rewards and nonzero gradient exactly matched baseline
Three paired live runs, median off/on round ratio 0.735×; no speed improvement

Detailed speed tables and scope are in the upstream description. Test harnesses, reports and coordination files were removed from the submitted diff; the existing argument fixture retains the six new default-off fields.

0z5a and others added 21 commits September 27, 2026 16:38
Signed-off-by: 0z5a <0z5a@users.noreply.github.com>
Signed-off-by: 0z5a <0z5a@users.noreply.github.com>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
…llm-project#459)

* [Training][Rollout] Integrate Ouro shared layers and policy updates

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>

* Trim Ouro integration to the implementation and training recipe

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>

---------

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
@0z5a

0z5a commented Oct 6, 2026

Copy link
Copy Markdown
Owner Author

Superseded by vllm-project#454, now merged into 0z5a-branch. The consolidated implementation and validation are retained there; closing this duplicate PR.

@0z5a 0z5a closed this Oct 6, 2026
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