TransformerLens has no adapter for Gemma4ForConditionalGeneration (Gemma 4, released 2026-06), so google/gemma-4-* models cannot be loaded via HookedTransformer or the model bridge. It shows up as a large gap in the model registry (architecture_gaps.json).
Would be good to add a text-only architecture adapter (the Gemma 3 / Gemma3n adapters are the natural starting point).
Related: #1497 (fork-capable attention bridge for Mistral) is needed for per-head attention-input interventions on the newer separate-q/k/v architectures.
TransformerLens has no adapter for
Gemma4ForConditionalGeneration(Gemma 4, released 2026-06), sogoogle/gemma-4-*models cannot be loaded viaHookedTransformeror the model bridge. It shows up as a large gap in the model registry (architecture_gaps.json).Would be good to add a text-only architecture adapter (the Gemma 3 / Gemma3n adapters are the natural starting point).
Related: #1497 (fork-capable attention bridge for Mistral) is needed for per-head attention-input interventions on the newer separate-q/k/v architectures.