| { |
| "schema_version": 1, |
| "_comment": "LFM2.5-230M @ MAXCTX=2048, W8 weight-only (int8 weights, fp16 activations) decode + batched prefill (PN=512) + NPU lm-head. Sliding-window KV ring continues past 2048 (remembers the last 2048 tokens). 14-layer hybrid (6 attn + 8 conv); per-layer schedule via attn_idx/conv_idx.", |
| "model": { "name": "lfm2-5-230m-2048", "family": "llm", "dsp_arch": "v79" }, |
| "params": { |
| "hidden": 1024, "vocab": 65536, "n_layers": 14, "max_ctx": 2048, |
| "kv_dim": 512, "head_dim": 64, "rope_theta": 1000000.0, "eos_token_id": 7 |
| }, |
| "artifacts": { |
| "contexts": { "decode": { "bin": "lfm230_dec_2048_w8.bin" }, "prefill": { "bin": "lfm230_pf_2048_w8.bin" }, "lmhead": { "bin": "lfm230_lmh_w8.bin" } }, |
| "embed": "lfm_embed_f16.bin", |
| "tokenizer": "tokenizer.json" |
| }, |
| "plan": { |
| "steps": [ |
| { "host": "lfm_generate", "params": { "prefill": "lfm230_pf_2048_w8", "decode": "lfm230_dec_2048_w8", "lmhead": "lfm230_lmh_w8", "attn_idx": "2,4,6,8,10,12", "conv_idx": "0,1,3,5,7,9,11,13" } }, |
| { "emit": "all" } |
| ] |
| } |
| } |
|
|
