Retire the Qwen AR pipeline

tools/issue104-reference.py

@@ -0,0 +1,66 @@
+"""Pinned MTPLX full-model classic AR receipts. Reference only; no downloads.
+
+Run serially under test-supervisor with MLX 0.32.2 and pinned MTPLX on PYTHONPATH.
+"""
+import hashlib
+import json
+from pathlib import Path
+import sys
+
+print("#104 importing pinned MTPLX", file=sys.stderr, flush=True)
+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
+from mtplx.cache_state import snapshot_cache
+from mtplx.session_bank import SessionBank
+
+assert mx.__version__ == "0.32.2"
+root = Path(sys.argv[1])
+print("#104 loading unchanged Bare-Speed model", file=sys.stderr, flush=True)
+model, config = load_model(root, lazy=False, strict=True,
+    get_model_classes=lambda **_: (qwen4_exp.Model, qwen4_exp.ModelArgs))
+model.post_weight_load(root)
+tokenizer = _load_tokenizer_resilient(root, config)
+rt = MTPLXRuntime(model, tokenizer, root, False, None)
+
+def fingerprint(array):
+    values = np.asarray(array.astype(mx.float32)).astype('<f4')
+    assert np.isfinite(values).all()
+    return dict(shape=list(array.shape), sha256=hashlib.sha256(values.tobytes()).hexdigest())
+
+def final_state(out, prompt):
+    assert out.final_state is not None and out.final_state.safe_to_commit
+    final = out.final_state
+    snapshot = snapshot_cache(final.final_trunk_cache)
+    states = [[None if a is None else fingerprint(a) for a in arrays]
+              if arrays is not None else None for arrays in snapshot.states]
+    return dict(ids=prompt + out.tokens, logits=fingerprint(final.final_logits), states=states)
+
+records, warm_records = [], []
+for temperature in (.6, 1.):
+    prompt = [9707, 11, 1075]
+    sampler = SamplerConfig(temperature=temperature, top_p=.95, top_k=20)
+    bank = SessionBank()
+    for warm in (False, True):
+        out = generate_ar(rt, prompt, max_tokens=8,
+            sampler=sampler, seed=0, stop_token_ids=set(), capture_final_state=True,
+            session_bank=bank, session_id="issue104", store_prefix_snapshot=False,
+            token_callback=lambda ids: print(f"#104 t={temperature} warm={warm}: {ids}", file=sys.stderr, flush=True))
+        value = final_state(out, prompt)
+        (warm_records if warm else records).append(dict(temperature=temperature, state=value))
+        print(json.dumps(dict(event="issue104_oracle", temperature=temperature, warm=warm,
+            tokens=out.tokens, state=value, cached_tokens=out.stats.cached_tokens,
+            seconds=out.stats.elapsed_s)), flush=True)
+        if not warm:
+            final = out.final_state
+            assert bank.put(runtime=rt, token_ids=value["ids"], cache=final.final_trunk_cache,
+                logits=final.final_logits, hidden=None, hidden_variant="post_norm",
+                session_id="issue104", mtp_history_policy="cycle", snapshot_epoch=len(value["ids"])) is not None
+            prompt = value["ids"] + [1075]
+        else:
+            assert out.stats.cached_tokens == len(prompt) - 1
+print(json.dumps(dict(event="issue104_classic_baseline", records=records, warm_records=warm_records)), flush=True)