Expose effective MLLM MTP draft stats

tests/test_mllm_continuous_batching.py

@@ -1005,6 +1005,13 @@ def __init__(self):
 
         install_mtp_mllm(batch_gen, language_model, num_draft_tokens=4)
 
+        stats = batch_gen.get_mtp_stats()
+        assert stats["enabled"] is True
+        assert stats["requested_draft_tokens"] == 4
+        assert stats["effective_draft_tokens"] == 1
+        assert stats["attempted"] == 0
+        assert "non_greedy_sampling" in stats["bypass_reasons"]
+
         logits_processor = MagicMock()
         tokens, logprobs = batch_gen._step(
             mx.array([[123]]),
@@ -1021,6 +1028,7 @@ def __init__(self):
         original_step.assert_called_once()
         language_model.assert_not_called()
         language_model.mtp_forward.assert_not_called()
+        assert batch_gen.get_mtp_stats()["attempted"] == 0
 
     def test_install_mtp_mllm_disables_mtp_for_non_greedy_sampling(self):
         from vllm_mlx.mllm_batch_generator import install_mtp_mllm
@@ -1130,6 +1138,9 @@ def __call__(self, verify_input, cache=None, return_hidden=False):
         assert [r.token for r in responses] == [1, 2]
         assert request_sampler.call_count == 1
         assert batch_gen.sampler.call_count == 0
+        assert batch_gen.get_mtp_stats()["attempted"] == 1
+        assert batch_gen.get_mtp_stats()["accepted"] == 1
+        assert batch_gen.get_mtp_stats()["acceptance_rate"] == 1.0
 
     def test_next_keeps_retired_processors_by_default(self, monkeypatch):
         from vllm_mlx.mllm_batch_generator import (

vllm_mlx/cli.py

@@ -261,7 +261,12 @@ def serve_command(args):
         if args.chunked_prefill_tokens > 0:
             print(f"Chunked prefill: {args.chunked_prefill_tokens} tokens per step")
         if args.enable_mtp:
-            print(f"MTP: enabled, draft_tokens={args.mtp_num_draft_tokens}")
+            print(f"MTP: enabled, requested_draft_tokens={args.mtp_num_draft_tokens}")
+            if args.mllm:
+                print(
+                    "MTP: MLLM path currently uses effective_draft_tokens=1 "
+                    "per verify step; inspect /v1/status for attempts and acceptance"
+                )
         print(f"Stream interval: {args.stream_interval} tokens")
         if args.use_paged_cache:
             print(

vllm_mlx/engine/batched.py

@@ -1100,6 +1100,7 @@ def get_stats(self) -> dict[str, Any]:
                 "paged_cache",
                 "prefix_cache",
                 "batch_generator",
+                "mtp",
                 "requests",
             ):
                 if key in mllm_stats:

vllm_mlx/mllm_batch_generator.py

@@ -1973,8 +1973,33 @@ def install_mtp_mllm(
     # Deferred drafts keyed by UID
     _deferred_drafts: Dict[int, dict] = {}
 
-    # MTP stats
-    _mtp_stats = {"accepted": 0, "rejected": 0, "errors": 0}
+    # MTP stats. These are intentionally exposed through get_mtp_stats() so
+    # /v1/status can distinguish "weights injected" from useful draft work.
+    _mtp_stats = {"attempted": 0, "accepted": 0, "rejected": 0, "errors": 0}
+
+    def _get_mtp_stats() -> Dict[str, Any]:
+        verified = _mtp_stats["accepted"] + _mtp_stats["rejected"]
+        acceptance_rate = _mtp_stats["accepted"] / verified if verified > 0 else 0.0
+        return {
+            "enabled": True,
+            "requested_draft_tokens": num_draft_tokens,
+            "effective_draft_tokens": 1,
+            "mode": "always_advance_verified",
+            "attempted": _mtp_stats["attempted"],
+            "accepted": _mtp_stats["accepted"],
+            "rejected": _mtp_stats["rejected"],
+            "errors": _mtp_stats["errors"],
+            "acceptance_rate": acceptance_rate,
+            "bypass_reasons": {
+                "prefill": "input_tokens.shape[1] > 1",
+                "no_active_batch": "active_batch is None",
+                "concurrent_batch": "len(active_batch) > 1",
+                "non_greedy_sampling": "temperature/top_p/top_k/min_p not greedy",
+                "logits_processors": "request-local logits processors active",
+            },
+        }
+
+    batch_gen.get_mtp_stats = _get_mtp_stats
 
     def _mtp_step(
         input_tokens: mx.array,
@@ -2067,6 +2092,7 @@ def _mtp_step(
 
         # MTP draft + always-advance verify
         try:
+            _mtp_stats["attempted"] += 1
             draft_logits = language_model.mtp_forward(
                 hidden_states[:, -1:, :],
                 primary_tokens[:, None],

vllm_mlx/mllm_scheduler.py

@@ -1130,6 +1130,8 @@ def get_stats(self) -> Dict[str, Any]:
             # Vision embedding cache stats from batch generator
             vec_stats = self.batch_generator.get_vision_cache_stats()
             stats["vision_embedding_cache"] = vec_stats
+            if hasattr(self.batch_generator, "get_mtp_stats"):
+                stats["mtp"] = self.batch_generator.get_mtp_stats()
 
         # Include Metal memory stats
         try:

vllm_mlx/server.py

@@ -3025,6 +3025,7 @@ async def status():
         "cache": stats.get("memory_aware_cache")
         or stats.get("paged_cache")
         or stats.get("prefix_cache"),
+        "mtp": stats.get("mtp"),
         "requests": stats.get("requests", []),
     }