b10094

ggml-org/llama.cppb10094Jul 23, 2026by github-actions[bot]same day as b10093

What changed since b10093

medium risk

b10094 auto-detects the speculative decoding type from draft repo sidecars when --spec-type is omitted.

New features

  • Speculative type is now inferred automatically from mtp/dflash/eagle3 sidecars in the draft repo when no --spec-type flag is given.
  • Sidecar selection follows a fixed mtp > dflash > eagle3 priority order.

Improvements

  • Simplifies speculative decoding setup: `llama-server -hf repo:Q3_K_M -hfd repo:Q8_0` now works without extra flags.
  • Explicit --spec-type still overrides inference, preserving manual control.

Bug fixes

  • Draft repos shipping sidecars previously resolved to a full model instead of the intended sidecar draft; this is now corrected.

Upgrade notes

No config or migration work is required. If you rely on a draft repo resolving to a full model despite shipping sidecars, pass --spec-type explicitly to disable the new inference. Repos without sidecars behave exactly as before.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Automatically infers the speculative type from draft repo sidecars (mtp/dflash/eagle3) when no explicit `--spec-type` is provided.

Key Highlights

  • Infer speculative type from draft repo sidecars
  • Simplified speculative decoding usage
  • Supports mtp > dflash > eagle3 priority

Release Notes (excerpt)

<details open> common: infer the speculative type from the draft repo sidecars (#25989) With -hfd pointing to a repo that ships mtp-/dflash-/eagle3- sidecars and no --spec-type given, the draft resolved to a full model while the sidecar was the intended draft. When the speculative types are still at their default, discover the sidecars of the draft repo, pick the first available following the existing mtp > dflash > eagle3 priority, and set the corresponding…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories