Today
-
Prompt as inputs (no baked questions), split Neural Engine model for iOS, card and code
-
d1_ane_prompt_split_fp16: the Neural Engine export in 4 pieces (iOS's Neural Engine compiler can't take 3.1 GB at once), questions as inputs
-
embed_tokens_fp16.npy: d1-2B's input embeddings, to build the prompt inputs
-
d1_ane_prompt_fp16: Neural Engine layout, the questions as inputs
-
d1_gpu_prompt_fp16: the questions as inputs (prompt padded to 640 tokens)
Yesterday
-
Card: new Neural Engine speed, energy and parity; code: MLP split, power.py
-
d1_ane_fp16.aimodel: MLPs in 8 chunks, 1.4x faster on the Neural Engine (236 -> 171 ms)
-
code: coreai_backend downloads the models from this repo; add coreai_py.py (used by parity.py)
-
d1_ane_fp16.aimodel: Neural Engine, fp16
-
d1_gpu_fp16.aimodel: GPU, fp16
-
Card, LFM Open License v1.0, calibration, prompt and code
-
initial commit