Retire the Qwen AR pipeline

tools/mtplx-kernel-fixtures.py

@@ -1627,7 +1627,7 @@ def ar_driver_fixtures():
     """Unchanged generate_ar with a tiny GPU transition-table model, no weights.
 
     Isolates orchestration (not model or throughput parity); actual prefill,
-    samplers, pipeline and final-state capture are all MTPLX implementations.
+    samplers and final-state capture are all MTPLX implementations.
     """
     from mtplx.generation import generate_ar
     from mtplx.runtime import MTPLXRuntime
@@ -1641,11 +1641,8 @@ def ar_driver_fixtures():
     table = mx.array(values.reshape(vocab, vocab))
     class Model:
         def __init__(self):
-            self.calls, self.modes = [], []
+            self.calls = []
         def make_cache(self): return []
-        def set_ar_pipeline_mode(self, enabled):
-            self.modes.append(bool(enabled))
-            return True
         def __call__(self, ids, cache=None, return_hidden=False, emit_logits=True, logits_keep=None):
             self.calls.append((ids, return_hidden))
             out = table[ids]
@@ -1653,12 +1650,11 @@ def ar_driver_fixtures():
     tokenizer = SimpleNamespace(decode=lambda ids, **kwargs: " ".join(map(str, ids)))
     for pipeline in (False, True):
         for asynchronous in (False, True):
-            for temperature, penalty in ((0., 0.), (.6, 0.), (.6, .4)):
+            for temperature, top_p, top_k, penalty in ((0., .95, 20, 0.), (.6, .95, 20, 0.), (.6, .95, 20, .4), (1., .95, 20, 0.), (1., .5, 2, 0.), (1., 0., 2, 0.), (1., 1., 2, 0.)):
                 for limit in (0, 1, 2, 3, 12):
                     for capture in (False, True):
-                        # 13/35 are the first/second sampled pipeline tokens:
-                        # exercise immediate EOS and EOS with an in-flight next
-                        # graph, not just limits and a non-triggering stop set.
+                        # Retain the historical stop IDs and explicitly set the
+                        # retired flag to prove current MTPLX ignores it.
                         for stop_ids in (set(), {0, 5, 36}, {13}, {35}):
                             os.environ["MTPLX_AR_PIPELINE"] = "1" if pipeline else "0"
                             os.environ["MTPLX_ASYNC_AR"] = "1" if asynchronous else "0"
@@ -1666,15 +1662,15 @@ def ar_driver_fixtures():
                             rt = MTPLXRuntime(model, tokenizer, Path("."), False, None)
                             emitted = []
                             out = generate_ar(rt, [1, 2], max_tokens=limit,
-                                sampler=SamplerConfig(temperature=temperature, top_p=.95, top_k=20, presence_penalty=penalty),
+                                sampler=SamplerConfig(temperature=temperature, top_p=top_p, top_k=top_k, presence_penalty=penalty),
                                 seed=4294967297, stop_token_ids=stop_ids, token_callback=lambda ids: emitted.extend(ids), capture_final_state=capture)
                             # Decode calls only: real prefill consumed [1] then [2].
                             calls = [{"ids": np.array(ids).reshape(-1).tolist(), "hidden": hidden} for ids, hidden in model.calls[2:]]
                             final = None if out.final_state is None else dict(shape=out.final_state.final_logits.shape,
                                 sha256=hashlib.sha256(np.array(out.final_state.final_logits).tobytes()).hexdigest())
-                            print(json.dumps(dict(pipeline=pipeline, asynchronous=asynchronous, temperature=temperature,
+                            print(json.dumps(dict(legacy_pipeline=pipeline, asynchronous=asynchronous, temperature=temperature, top_p=top_p, top_k=top_k,
                                 penalty=penalty, limit=limit, capture=capture, stops=sorted(stop_ids), tokens=out.tokens,
-                                emitted=emitted, calls=calls, modes=model.modes, final=final,
+                                emitted=emitted, calls=calls, final=final,
                                 verify_calls=out.stats.verify_calls, finish=out.finish_reason)), flush=True)