Skip to content

[MLX][Spec] Add optional Gemma 4 target feedback - #2

Draft
wirybeaver wants to merge 1 commit into
prototype/mlx-gemma4-mtp-verify-cachefrom
prototype/mlx-gemma4-mtp-target-feedback
Draft

wirybeaver wants to merge 1 commit into
prototype/mlx-gemma4-mtp-verify-cachefrom
prototype/mlx-gemma4-mtp-target-feedback

Conversation

@wirybeaver

@wirybeaver wirybeaver commented Jul 23, 2026

Copy link
Copy Markdown
Owner

Stack

PR 2 of 6. Depends on #1.

  • Base: prototype/mlx-gemma4-mtp-verify-cache
  • Head: prototype/mlx-gemma4-mtp-target-feedback

Scope

  • Add an opt-in target adapter for exact logits, hidden rows, and assistant seeds.
  • Integrate transactional two-query verification with the MLX runner.
  • Preserve the target-only numerical path and native history/cache invariant.
  • Pass pinned target revisions through MLX loading.

Target-only inference keeps its existing path unless feedback is requested. No assistant, speculative worker, or serving dispatch is added.

Tests

test/registered/unit/hardware_backend/mlx/test_gemma4_target_adapter.py

Result: 4 passed, 6 subtests passed.


CI States

Latest PR Test (Base): ❌ Run #30053071879
Latest PR Test (Extra): ❌ Run #30053071712

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant