| { |
| "format": "lfm3-coreml-text/1", |
| "model": "LiquidAI/LFM3-Audio-2B", |
| "revision": "9d49701c372d6011d900f6abb40c6b296fb706dd", |
| "variant": "int8b32", |
| "hidden_size": 2048, |
| "vocab_size": 128000, |
| "head_dim": 128, |
| "num_layers": 24, |
| "layer_types": [ |
| "conv", |
| "conv", |
| "full_attention", |
| "conv", |
| "sliding_attention", |
| "conv", |
| "full_attention", |
| "conv", |
| "sliding_attention", |
| "conv", |
| "full_attention", |
| "conv", |
| "sliding_attention", |
| "conv", |
| "full_attention", |
| "conv", |
| "sliding_attention", |
| "conv", |
| "full_attention", |
| "conv", |
| "sliding_attention", |
| "conv", |
| "full_attention", |
| "conv" |
| ], |
| "cache_len": 4160, |
| "fifo_len": 256, |
| "sliding_window": 256, |
| "rope": { |
| "full": { |
| "theta": 1000000.0, |
| "pairs": 32 |
| }, |
| "sliding": { |
| "theta": 10000.0, |
| "pairs": 64 |
| } |
| }, |
| "mask_neg": -30000.0, |
| "topk": { |
| "k": 64, |
| "chunks": 4, |
| "chunk_size": 32000, |
| "note": "topk_v/topk_i are [1,chunks,k]: per-chunk top-k with LOCAL ids; global id = i + c*chunk_size; merge on the host (GPU topk/int math is 16-bit on macOS 27)" |
| }, |
| "embedding": { |
| "file": "embed_tokens.f16.bin", |
| "dtype": "float16", |
| "shape": [ |
| 128000, |
| 2048 |
| ] |
| }, |
| "stages": [ |
| { |
| "package": "LFM3-A-C4160-int8b32.mlpackage", |
| "layers": [ |
| 0, |
| 12 |
| ], |
| "functions": [ |
| { |
| "name": "decode_q1", |
| "q": 1, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "decode", |
| "calls": [ |
| "decode_q1", |
| "decode_q1" |
| ] |
| }, |
| { |
| "name": "decode_hidden_q1", |
| "q": 1, |
| "span": 4096, |
| "head": "none", |
| "role": "decode_hidden", |
| "calls": [ |
| "decode_q1", |
| "decode_q1" |
| ] |
| }, |
| { |
| "name": "prefill_q256", |
| "q": 256, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q256", |
| "prefill_q256" |
| ] |
| }, |
| { |
| "name": "prefill_q64", |
| "q": 64, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q64", |
| "prefill_q64" |
| ] |
| }, |
| { |
| "name": "prefill_q16", |
| "q": 16, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q16", |
| "prefill_q16" |
| ] |
| }, |
| { |
| "name": "prefill_q13", |
| "q": 13, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q13", |
| "prefill_q13" |
| ] |
| } |
| ] |
| }, |
| { |
| "package": "LFM3-B-C4160-int8b32.mlpackage", |
| "layers": [ |
| 12, |
| 24 |
| ], |
| "functions": [ |
| { |
| "name": "decode_q1", |
| "q": 1, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "decode", |
| "calls": [ |
| "decode_q1", |
| "decode_q1" |
| ] |
| }, |
| { |
| "name": "decode_hidden_q1", |
| "q": 1, |
| "span": 4096, |
| "head": "none", |
| "role": "decode_hidden", |
| "calls": [ |
| "decode_q1", |
| "decode_q1" |
| ] |
| }, |
| { |
| "name": "prefill_q256", |
| "q": 256, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q256", |
| "prefill_q256" |
| ] |
| }, |
| { |
| "name": "prefill_q64", |
| "q": 64, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q64", |
| "prefill_q64" |
| ] |
| }, |
| { |
| "name": "prefill_q16", |
| "q": 16, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q16", |
| "prefill_q16" |
| ] |
| }, |
| { |
| "name": "prefill_q13", |
| "q": 13, |
| "span": 4096, |
| "head": "topk", |
| "k": 64, |
| "penalty": true, |
| "role": "prefill", |
| "calls": [ |
| "prefill_q13", |
| "prefill_q13" |
| ] |
| } |
| ] |
| } |
| ], |
| "head": { |
| "package": "LFM3-Head-int8b32.mlpackage", |
| "inputs": { |
| "h": "[1,1,H] fp16 (last row of stage B h)", |
| "rp": "[1,1,V] fp16 repetition-penalty mask" |
| }, |
| "outputs": "topk_v/topk_i [1,chunks,k]", |
| "stateless": true |
| }, |
| "prefill": { |
| "buckets": [ |
| 256, |
| 64, |
| 16, |
| 13 |
| ], |
| "pad_first": true |
| }, |
| "tokens": { |
| "bos": 124894, |
| "eos": [ |
| 124900, |
| 125014 |
| ], |
| "pad": 124893 |
| }, |
| "generation": { |
| "temperature": 0.0, |
| "top_k": 1, |
| "repetition_penalty": 1.0, |
| "do_sample": false |
| }, |
| "inputs": { |
| "x": "[1,Q,H] fp16 inputs_embeds (host gathers embed_tokens rows; zero rows = pads)", |
| "cos_g/sin_g": "[1,1,Q,D] fp16 zero-tail RoPE (rope.full) at positions slot - n_pad", |
| "cos_s/sin_s": "[1,1,Q,D] fp16 RoPE (rope.sliding)", |
| "mask_g": "[1,1,Q,span] additive -3e4: key slot <= query slot, same pad class", |
| "mask_s": "[1,1,Q,fifo+Q] over cat(fifo[256], new[Q]); fifo j holds slot slot0-256+j; window 256 incl. self", |
| "write_pos": "int32 [1] slot of the first row (prefill; decode too unless decode_kv_write == onehot)", |
| "onehot": "[1,1,span,1] fp16, 1.0 at the write slot (decode functions when decode_kv_write == onehot)", |
| "rp": "[1,1,V] fp16, 1 = token seen in prompt+generated (repetition penalty in graph)" |
| }, |
| "outputs": { |
| "h": "[1,Q,H] fp16 after embedding_norm (stage B)", |
| "topk_v/topk_i": "[1,chunks,k] last row: per-vocab-chunk top-k (desc) with local int32 ids" |
| }, |
| "notes": "stage A output h (raw residual) feeds stage B input x; both stages take the same host inputs; one MLState per stage shared by all its functions; never pad a chunk except the first on a fresh state", |
| "spans": [ |
| 4096 |
| ], |
| "build_seconds": 318.7, |
| "quantization": "int8b32", |
| "k": 64, |
| "awq": null, |
| "decode_kv_write": "onehot" |
| } |
