Skip to content

[MLX][Spec] Add Gemma 4 assistant runtime - #3

Draft
wirybeaver wants to merge 1 commit into
prototype/mlx-gemma4-mtp-target-feedbackfrom
prototype/mlx-gemma4-mtp-assistant
Draft

wirybeaver wants to merge 1 commit into
prototype/mlx-gemma4-mtp-target-feedbackfrom
prototype/mlx-gemma4-mtp-assistant

Conversation

@wirybeaver

@wirybeaver wirybeaver commented Jul 23, 2026

Copy link
Copy Markdown
Owner

Stack

PR 3 of 6. Depends on #1 and #2.

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

Scope

  • Reuse pinned mlx-vlm==0.5.0; no assistant model code is vendored.
  • Load assistant metadata without Transformers hooks and validate target compatibility.
  • Map YOCO logical layers to compact native-cache owners.
  • Provide generation-checked, read-only target-KV views and one-token proposals.

Two scenario-focused tests cover configuration/lifecycle and rotating-cache immutability. No scheduler route or speculative worker references this runtime yet; serving remains dormant.

Tests

test/registered/unit/hardware_backend/mlx/test_gemma4_mtp_assistant.py

Result: 2 passed, 10 subtests passed.


CI States

Latest PR Test (Base): ❌ Run #30053085923
Latest PR Test (Extra): ❌ Run #30053085851

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