kv_dispatch: apply gemma-4 per-layer epilogue (PLE + layer_scalar) in cached prefill/decode (fixes #177, stacked on #178) - #181

Open

layer5one wants to merge 6 commits into

chrishayuk:mainfrom

layer5one:gemma4-kv-cache-epilogue-fix

Commits on Jul 5, 2026

gguf: emit gemma4 per-layer attention geometry in to_config_json
taylorsorensen
committed

Commits on Jul 6, 2026

gguf: map gemma 2/3/4 norm + layer_output_scale tensor names
Taylor
committed
lql: route prompt tokenization through encode_prompt (gemma-4 BOS)
Taylor
committed
quant: fix Q4_0/Q4_1/Q5_0/Q5_1/Q6_K nibble layout to match ggml
Taylor
committed
gguf: emit final_logit_softcapping and rms_norm_eps in to_config_json
Taylor
committed

Commits on Jul 8, 2026

kv_dispatch: apply gemma-4 per-layer epilogue (PLE + layer_scalar) in cached prefill/decode

taylorsorensen
and
claude
committed