"""Pinned full-model cold/warm prefill, including explicit hidden-state capture.""" import argparse,hashlib,json,os,subprocess,sys from pathlib import Path parser=argparse.ArgumentParser(description=__doc__) for name in ['checkout','root','native','out']:parser.add_argument(name,type=Path) parser.add_argument('--hidden',action='store_true') a=parser.parse_args() assert subprocess.check_output(['git','-C',str(a.checkout),'rev-parse','HEAD'],text=True).strip()=='9882703f3105363ddc37eca9f97aa09a1d387112' sys.path.insert(0,str(a.checkout)) from mtplx.server.openai import _server_runtime_env_overrides os.environ.update(_server_runtime_env_overrides(argparse.Namespace(model=str(a.root),generation_mode='mtp' if a.hidden else 'ar'),None)) os.environ['MTPLX_SUSTAINED_PREFILL']='1' import mlx.core as mx import numpy as np from mlx_lm.utils import load_model from mtplx.models import qwen4_exp from mtplx.runtime import MTPLXRuntime,_load_tokenizer_resilient from mtplx.sampling import SamplerConfig from mtplx.generation import generate_ar,restore_or_prefill_prompt_state from mtplx.cache_state import snapshot_cache,snapshot_cache_lazy_hybrid from mtplx.session_bank import SessionBank,_lazy_snapshot_enabled from mtplx.kernels import moe_sorted_gather_nax as kernel assert mx.__version__=='0.32.2' a.out.mkdir(parents=True,exist_ok=True) model,config=load_model(a.root,lazy=False,strict=True,get_model_classes=lambda **_:(qwen4_exp.Model,qwen4_exp.ModelArgs)) model.post_weight_load(a.root) tokenizer=_load_tokenizer_resilient(a.root,config) from mtplx.mtp_patch import MTPContract if a.hidden: assert model.attach_mtp(a.root) rt=MTPLXRuntime(model,tokenizer,a.root,a.hidden,MTPContract() if a.hidden else None) def fingerprint(v): values=np.asarray(v.astype(mx.float32)).astype('