Skip to content

fix gemma4_assistant: arch name, tensor names, MTP dimension alignment - #72

Open
limitlessstoner wants to merge 2 commits into
AtomicBot-ai:feature/turboquant-kv-cachefrom
limitlessstoner:feature/turboquant-kv-cache
Open

fix gemma4_assistant: arch name, tensor names, MTP dimension alignment#72
limitlessstoner wants to merge 2 commits into
AtomicBot-ai:feature/turboquant-kv-cachefrom
limitlessstoner:feature/turboquant-kv-cache

Conversation

@limitlessstoner

Copy link
Copy Markdown
  • Change arch string 'gemma4-assistant' -> 'gemma4_assistant' to match GGUF
  • Change tensor names 'nextn.' -> 'mtp.' to match GGUF convention
  • Remove strict n_layer_nextn == n_layer_all assertion
  • Remove n_embd_out != n_embd assertion
  • Read n_embd_backbone dynamically from projection tensor shape
  • Fall back to n_layer when n_layer_nextn key is absent
  • Propagate n_embd_backbone into hparams for graph building
  • Set n_embd_out_impl to backbone size for MTP compatibility check
  • Allow partial tensor loading for GEMMA4_ASSISTANT

Overview

Additional information

Requirements

- Change arch string 'gemma4-assistant' -> 'gemma4_assistant' to match GGUF
- Change tensor names 'nextn.*' -> 'mtp.*' to match GGUF convention
- Remove strict n_layer_nextn == n_layer_all assertion
- Remove n_embd_out != n_embd assertion
- Read n_embd_backbone dynamically from projection tensor shape
- Fall back to n_layer when n_layer_nextn key is absent
- Propagate n_embd_backbone into hparams for graph building
- Set n_embd_out_impl to backbone size for MTP compatibility check
- Allow partial tensor loading for GEMMA4_ASSISTANT
@github-actions github-actions Bot added the model label Aug 12, 2026
Accept --draft-block-size and --draft-max (alias to speculative draft
n_max) and --spec-type mtp (alias to draft-mtp), so the server wrapper
can drive this build with the same flags as the published Atomic Windows
binaries.

Assisted-by: Kilo
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant