Add the chat template and stop at the end of a turn

gen-cards: regenerate Use-it block

tokenizer_config: inline the gemma-4 chat template (ships as separate chat_template.jinja upstream, which swift-transformers does not read; without it prompts raw-encode and generation degenerates to turn markers)

tokenizer_config: inline the gemma-4 chat template (ships as separate chat_template.jinja upstream, which swift-transformers does not read; without it prompts raw-encode and generation degenerates to turn markers)

Note b2 (beta3-loadable) bundles under gpu-pipelined-b2/

Add b2 (beta3-loadable) bundle under gpu-pipelined-b2/gemma4_e2b_qat_decode_int4lin_tbl_aotc_h18p/ (b1 retained)

Add b2 (beta3-loadable) bundle under gpu-pipelined-b2/gemma4_e2b_qat_decode_int4lin_tbl/ (b1 retained)

gen-cards: regenerate Use-it block

gen-cards: regenerate Use-it block

README: add gen-cards use-it markers

Add root config.json (Hub download-stats query file + framework pointer)

Gemma4-VL: vision bundles + card section

Gemma4-VL: add gemma4_e2b_qat_vl_vision

Gemma4-VL: add gemma4_e2b_qat_vl_decode_int4linsym_aotc_h18p

Gemma4-VL: add gemma4_e2b_qat_vl_decode_int4linsym

Gemma4-VL: add gemma4_e2b_qat_vl_decode_int4linsym_tbl

card: official-QAT int4 row + section

E2B QAT gather tables (checkpoint-derived - pair with the QAT bundles)

E2B QAT tbl bundle, precompiled h18p .aimodelc

E2B official-QAT int4lin tbl bundle (Mac 78.9, iPhone 30.7 tok/s)

card: iPhone-ready AOT h18p tbl bundle + CoreAIChat Gemma ⚡ mode (chat-surface 32.7/44.2)

gemma4 e2b int4lin tbl AOT h18p bundle (CoreAIChat Gemma ⚡ download target)

card: GPU-pipelined fast path section (tbl bundle, run contract, AOT + buffer rules)

gpu-pipelined: gemma4 int4lin tbl bundle (PLE table as static graph input; M4 Max 77.0, iPhone 30.3 via AOT)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

ios-ane: revert to the device-verified bucket-64 chunk set (the bucket-512 set jetsams the on-device ANE first compile)

gemma4 iPhone re-measure in the Release chat app: GPU 22, ANE 6 tok/s

Run it: device push-verification gotcha + AOT h18p architecture note

macOS best: head+argmax kernel (int8)

macOS best: int8 fused-kernel core (56.6-59 tok/s)

macOS best: gather front-end (.aimodel, int8)

iOS ANE: chunk plan (data-driven engine config)

iOS ANE best: head with in-graph argmax (int8)

iOS ANE best: chunk 6/6 (int8, fp16-hardened)

iOS ANE best: chunk 5/6 (int8, fp16-hardened)

iOS ANE best: chunk 4/6 (int8, fp16-hardened)

iOS ANE best: chunk 3/6 (int8, fp16-hardened)

iOS ANE best: chunk 2/6 (int8, fp16-hardened)

iOS ANE best: chunk 1/6 (int8, fp16-hardened)

mmap gather front-end tables (shared: iOS GPU + iOS ANE)

iOS GPU best: int4-kmeans head+argmax kernel

iOS GPU best: int4-kmeans fused-kernel core (17.7 tok/s)

Card: category layout (best verified config per platform x compute-unit)

Model card

initial commit