Conversation
|
Follow-up split plan:
Each PR will keep the default Qwen3.5 path unchanged unless DFlash is enabled. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c65ada7ecd
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
53581b1 to
06cde7e
Compare
|
This pull request has been inactive for 14 days. It will be closed after another 30 days unless there is new activity. |
0c9533f to
859f670
Compare
f60518e to
69dfad5
Compare
Signed-off-by: CAICAIIs <3360776475@qq.com>
69dfad5 to
c4a8b97
Compare
|
This pull request has been inactive for 14 days. It will be closed after another 30 days unless there is new activity. |
Summary
This PR is the target-verifier slice split out of #626. It makes Qwen3.5 speculative verification reviewable independently from the draft model and serving integration.
It verifies
[current token, draft tokens...]in one target forward pass, then commits the longest greedy-matching prefix plus one target token. Qwen3.5 hybrid state is handled as one transaction: paged KV, recurrent state, convolution state, and CUDA Graph slot state.What changed
Validation
On one RTX 5090 with Qwen3.5-4B:
Scope
This PR does not add the DFlash draft model, scheduler/server opt-in wiring, sampling support, or a performance claim. Those parts remain in follow-up PRs split from #626.