"""Held-out loss on assistant-side tokens. Usage: eval_nll.py [--fp32] [--actquant input,hidden] Counts only tokens the model would generate itself (assistant messages: reasoning, text, tool calls, and the end-of-turn token). Stores per-sequence sums so variants can be compared with paired statistics. """ import json, os, sys, time import numpy as np import mlx.core as mx sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) import qlab MODEL_TOK, END_MSG, EOS = 200001, 200010, 200006 def assistant_mask(ids): mask, inside = [], False for t in ids: if t == MODEL_TOK: inside = True mask.append(inside or t == EOS) if t == END_MSG: inside = False return mask def main(): args = sys.argv[1:] fp32 = "--fp32" in args if fp32: args.remove("--fp32") aq = None if "--actquant" in args: i = args.index("--actquant"); aq = tuple(args[i + 1].split(",")); del args[i:i + 2] recipe = None if "--recipe" in args: i = args.index("--recipe"); rfile, rname = args[i + 1].split(":"); del args[i:i + 2] recipe = json.load(open(rfile))[rname] model_dir, seq_path, out_path, tag = args[:4] path = qlab.NATIVE if model_dir == "native" else model_dir model, _ = qlab.load_native(fp32=False, path=path) size = None if recipe is not None: import recipes as R R.apply(model, recipe, os.path.join(qlab.ROOT, "q", "shards")) size = R.nbytes(model) / 2**30 if fp32: qlab.to_fp32(model); mx.eval(model.parameters()) if aq: import actquant actquant.install(model, aq) seqs = qlab.load_sequences(seq_path, split="eval", max_len=1024) rows, t0 = [], time.time() for s in seqs: ids = s["ids"] lg = qlab.logits_for(model, ids) lp = qlab.log_softmax(lg)[:-1] nxt = mx.array(ids[1:]) nll = -mx.take_along_axis(lp, nxt[:, None], axis=-1)[:, 0] m = mx.array(assistant_mask(ids)[1:]).astype(mx.float32) hit = (mx.argmax(lp, axis=-1) == nxt).astype(mx.float32) vals = [(nll * m).sum(), m.sum(), (hit * m).sum()] mx.eval(vals) rows.append(dict(name=s["name"], kind=s["kind"].split(":")[0], nll=vals[0].item(), n=int(vals[1].item()), hit=vals[2].item())) del lg, lp res = json.load(open(out_path)) if os.path.exists(out_path) else {} n = sum(r["n"] for r in rows) summ = dict(tokens=n, nll=sum(r["nll"] for r in rows) / n, acc=sum(r["hit"] for r in rows) / n, seconds=round(time.time() - t0), peak_gb=round(mx.get_peak_memory() / 2**30, 2), weights_gib=size, activations="fp32" if fp32 else "bf16", actquant=aq) for kind in sorted({r["kind"] for r in rows}): sel = [r for r in rows if r["kind"] == kind] summ[kind] = dict(tokens=sum(r["n"] for r in sel), nll=sum(r["nll"] for r in sel) / max(1, sum(r["n"] for r in sel))) summ["rows"] = rows res[tag] = summ json.dump(res, open(out_path, "w")) print(f"{tag:28s} tokens {n} NLL {summ['nll']:.4f} ppl {np.exp(summ['nll']):.3f} acc {summ['acc']:.4f} " + " ".join(f"{k} {summ[k]['nll']:.4f}" for k in ("chat", "code", "wikitext") if k in summ) + f" ({summ['seconds']}s)", flush=True) if __name__ == "__main__": main()