mirror of
https://github.com/ollama/ollama.git
synced 2026-04-28 03:39:48 +02:00
a23a5e76f3
The previous mapping renamed Ollama's `layer_pre_norm` to upstream's `attn_post_norm`, but in Ollama's gemma4 audio code `layer_pre_norm` is actually the FINAL block output norm (used at the end of the conformer block), while `ln2` (named AttnPostNorm in Ollama's Go) is the post-attention norm. Upstream gemma4a uses `attn_post_norm` for post-attention and `ln_2` for the final block norm, so the correct mapping is: ln1 → attn_pre_norm (pre-attention norm) ln2 → attn_post_norm (post-attention norm) layer_pre_norm → ln2 (final block output norm) linear_pos → attn_k_rel (relative-position K projection) Also scope the per-block renames to `a.blk.*` (not vision blocks, which have their own ln1/ln2 that must stay as-is). Order matters: ln2 → attn_post_norm must run before layer_pre_norm → ln2 to avoid collision. Verified end-to-end via `ollama run gemma4:e2b` with audio input — now produces transcription that matches the upstream-converted reference mmproj.