Dual-signature export v3: aligned chat+classifier decomposition pipelines so cross-program constant dedup works (3.9 GB, matches single-sig). 4 signatures: prefill_128_chat / decode_chat (no LoRA, no classifier) + prefill_128_classifier / decode_classifier (full).

Dual-signature export: decode_chat / prefill_128_chat (no LoRA, no classifier) + decode_classifier / prefill_128_classifier (with LoRA + classifier head). iOS Metal fused-attention shader should now match the chat graph.

QAT-trained base + classifier (gemma4_mixed48 fake-quant in the loop)

QAT-trained LoRA (gemma4_mixed48 fake-quant in the loop)

Upload tokenizer_config.json with huggingface_hub

Upload tokenizer.json with huggingface_hub

Upload lora_adapter.tflite with huggingface_hub

Upload model.litertlm with huggingface_hub

Upload README.md with huggingface_hub

initial commit