Postponed marker (like qwen's #352 lane) — descoped from the llama+gemma conformance release by decision.
Gemma 3n is effectively its own family living inside llm-inference/gemma: hand-rolled Gemma3nRuntime/Gemma3nAttentionBackend/Gemma3nKvCache (never touches gemmaNetwork()), own loaders, AltUp/ActivationSparsity ops — ~2000 LOC with exactly one external consumer (Gemma3nIngestion → kgemma CLI). Its engine delegation still has the defect fixed for gemma4: Gemma3nWeightLoader forces token_embd → GEMMA_DEQUANTIZE_ALL and Gemma3nLoadConfig dequantizes the whole model because the hand-rolled runtime can't consume packed tensors. Maturity gate: 0/5.
When picked up:
Refs: #346, #374 (the gemma sweep that leaves 3n untouched).
Postponed marker (like qwen's #352 lane) — descoped from the llama+gemma conformance release by decision.
Gemma 3n is effectively its own family living inside
llm-inference/gemma: hand-rolledGemma3nRuntime/Gemma3nAttentionBackend/Gemma3nKvCache(never touchesgemmaNetwork()), own loaders,AltUp/ActivationSparsityops — ~2000 LOC with exactly one external consumer (Gemma3nIngestion→ kgemma CLI). Its engine delegation still has the defect fixed for gemma4:Gemma3nWeightLoaderforcestoken_embd → GEMMA_DEQUANTIZE_ALLandGemma3nLoadConfigdequantizes the whole model because the hand-rolled runtime can't consume packed tensors. Maturity gate: 0/5.When picked up:
Gemma3n*+AltUp+ActivationSparsity+ the gemma-moduleAttentionBackendintollm-inference/gemma3n(+llm-runtime/kgemma3nfacade), per the Common structure and naming convention for model family modules #346 template — mechanical, one consumerRefs: #346, #374 (the gemma sweep that leaves 3n untouched).