Skip to content

Add draft physical batch size support - #407

Merged
thomas9120 merged 2 commits into
mainfrom
beta
Oct 14, 2026
Merged

thomas9120 merged 2 commits into
mainfrom
beta

Conversation

@thomas9120

Copy link
Copy Markdown
Owner

Summary

Add the --spec-draft-n-ubatch setting for speculative decoding in server and CLI launches. The value is preserved across configuration changes and omitted when speculation is disabled or only ngram methods are active. Document the temporary compatibility window for builds through b11617 while the first supporting release is pending.

Scope

This change does not update the pinned llama.cpp binary or verify the first release that advertises the flag. That compatibility follow-up remains documented in docs/upstream-changes.md.

Verification

  • Added coverage for flag emission across tools and speculative modes, shared-state synchronization, and compatibility checks for older, newer, and unknown builds.
  • The diff does not report whether checks were run.

UI evidence

- Expose `--spec-draft-n-ubatch` for speculative decoding
- Allow known older builds to omit the pending flag during compatibility checks
- Document the release window and add coverage for launch arguments and shared state
@thomas9120
thomas9120 merged commit 806f8c7 into main Oct 14, 2026
6 checks passed
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