-
Revert to F32 KV: int8 KV bridge produced DEQUANTIZE 1-dim + STABLEHLO_COMPOSITE missing data → Metal delegate rejects
-
int8 KV cache export (cache_update_composite + int8_kv_bridge)
-
GPU-compatible export: no cache_update_composite, cache_length=4096, lora_rank=8
-
Re-export cache_length=4096 (was 32768) for 6GB devices
-
Replace with Modal-exported 32k bundle (int8 KV, LoRA r8)
-
Replace weight_only with dynamic_wi4_afp32 (fixes gibberish)
-
Replace weight_only with dynamic_wi4_afp32 (fixes gibberish)
-
Upload gemma4-android-2k.litertlm with huggingface_hub
-
Upload gemma4-base-2k.litertlm with huggingface_hub
-
Upload gemma4-base.litertlm with huggingface_hub
-
Upload gemma4-lora-8k.litertlm with huggingface_hub
-
Upload gemma4-base.litertlm with huggingface_hub
-
Upload gemma4-base.litertlm with huggingface_hub
-
Upload gemma4-base.litertlm with huggingface_hub
-
Upload adapters/pirate-scale-4.5.tflite with huggingface_hub
-
Upload adapters/alpaca-2ep.tflite with huggingface_hub
-
Upload adapters/alpaca-3ep.tflite with huggingface_hub
-
Upload adapters/alpaca-3ep-scale-0.625.tflite with huggingface_hub
-
Upload adapters/qat-test.tflite with huggingface_hub
-
Upload adapters/alpaca-1ep.tflite with huggingface_hub
-
Upload adapters/adapter.tflite with huggingface_hub
-
Upload adapters/alpaca-1k.tflite with huggingface_hub
-
Upload adapters/pirate-v9-scale-0.25.tflite with huggingface_hub
-
Upload gemma4-base.litertlm with huggingface_hub
-
initial commit
