[None][test] Assert MTP acceptance length in ADP + LM-head-TP accuracy tests

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -2865,6 +2865,13 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
         if moe_backend == "TRTLLM" and sm_version in (120, 121):
             pytest.skip(f"{moe_backend} backend does not support SM 120 or 121")
 
+        # Wrong MTP draft tokens cannot be caught by the accuracy scores:
+        # rejected drafts are regenerated by the target model, so accuracy
+        # stays intact and the only observable symptom is a collapsed
+        # acceptance length (https://nvbugs/6460072). Assert on it for the
+        # ADP + LM-head-TP cases, whose draft sampling takes a dedicated path.
+        check_acceptance = mtp_nextn > 0 and enable_lm_head_tp_in_adp
+
         kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.70)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
@@ -2886,6 +2893,7 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
                  **pytorch_config,
                  enable_attention_dp=attention_dp,
                  enable_lm_head_tp_in_adp=enable_lm_head_tp_in_adp,
+                 enable_iter_perf_stats=check_acceptance,
                  speculative_config=mtp_config) as llm:
 
             assert llm.args.moe_config.backend == moe_backend
@@ -2903,6 +2911,22 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
             # task.evaluate(llm,
             #               extra_evaluator_kwargs=dict(apply_chat_template=True))
 
+            if check_acceptance:
+                stats = llm.get_stats(timeout=2)
+                spec_iters = [
+                    stat['specDecodingStats'] for stat in stats
+                    if stat.get('specDecodingStats')
+                    and stat['specDecodingStats']['numDraftTokens'] > 0
+                ]
+                assert spec_iters, \
+                    "No iterations with speculative decoding stats"
+                acceptance_length = sum(s['acceptanceLength']
+                                        for s in spec_iters) / len(spec_iters)
+                assert acceptance_length >= 1.5, (
+                    f"MTP acceptance length {acceptance_length:.2f} < 1.5: "
+                    "draft tokens are likely wrong in the ADP + LM-head-TP "
+                    "sampling path.")
+
         import gc
         gc.collect()
         torch.cuda.empty_cache()