Conversation
Gemma 4's `use_double_wide_mlp=True` widens gate/up/down_proj to 2× base `intermediate_size` on KV-shared layers. On gemma-4-e2b-it (35 layers, last 20 shared), layers 15–34 have `intermediate=12288`, layers 0–14 have 6144. Crown-scan defaults to `(3n/5)=21` and lands on a double-wide layer, so the rank-1 edit hit `intermediate-size mismatch in captured keys` against the config-wide base size. Adds `ModelArchitecture::intermediate_size_for_layer(layer) -> usize` (default = `config.intermediate_size`, mirroring `head_dim_for_layer`). `Gemma4Arch` overrides by reusing the precomputed `kv_sources` set — one source of truth for KV-shared-layer membership. Thread the per-layer lookup through: - `edit_py.rs`: compute `intermediate` after `chosen_layer` is picked. - `edit_cmd.rs`: same for the CLI path. - `memit.rs`: `ffn_dim` now per-layer; `run_memit` already solves per layer, so covariances remain correctly sized across mixed layers. Parse `use_double_wide_mlp` in `detect.rs`; add to `ModelConfig`. Tests (in `detect.rs`): - `test_detect_gemma4_e2b`: asserts 6144 on L0/L14, 12288 on L15/L21/L34 — matches the actual HF tensor shapes verified in the Colab repl. - `test_gemma4_31b_no_double_wide`: 31B lacks the flag → base everywhere. - `test_non_gemma4_intermediate_default`: Llama returns base for all layers via the default trait impl. The bare `weights.intermediate_size` field is left as "base" for display / metadata call sites (demos, patch-print, vindex stats). Patch file-format unchanged: `compute_rank1` / `compute_dense` already derive `intermediate_size` from the runtime tensor, so new patches for double-wide layers store 12288 correctly without a version bump. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ModelArchitecture::intermediate_size_for_layer(layer)(default =config.intermediate_size)Gemma4Archoverrides via precomputedkv_sourceswhenuse_double_wide_mlp=Trueedit_py,edit_cmd, andmemituse_double_wide_mlpindetect.rsWhy
Crown-scan on
google/gemma-4-e2b-itfails withintermediate-size mismatch in captured keysbecause the default start layer(3n/5)=21lands in the KV-shared region where the MLP is double-wide (12288 vs. base 6144). Verified against actual HF tensor shapes: L0/L14 =(6144, 1536), L15/L21/L34 =(12288, 1536).Mirrors HuggingFace's
modeling_gemma4.py:Test plan
cargo test -p larql-models --lib— 74 pass (3 new)cargo test -p larql-inference --lib— 10 passcargo check -p larql-models -p larql-inference -p larql-cli -p larql-pythonclean🤖 Generated with Claude Code