A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 20:45 UTC
[Bugfix] Fix MoE Routing Simulation (#28002)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Tyler Michael Smith committed
Nov 3, 2025 at 17:26 UTC
3758757377b713b6acc997d0ac2c5dd49c332278
2 files changed
+11
-1
vllm/model_executor/layers/fused_moe/layer.py
+1
-1
@@ -2066,7 +2066,7 @@ class FusedMoE(CustomOp):
2066
)
2067
2068
# DeepSeekv2 uses grouped_top_k
2069
- if use_grouped_topk:
2069
+ elif use_grouped_topk:
2070
assert topk_group is not None
2071
assert num_expert_group is not None
2072
if is_rocm_aiter_moe_enabled():
vllm/model_executor/layers/fused_moe/routing_simulator.py
+10
@@ -14,6 +14,10 @@ from typing import Any
14
15
import torch
16
17
+from vllm.logger import init_logger
18
+
19
+logger = init_logger(__name__)
20
+
21
22
class RoutingStrategy(ABC):
23
"""Base class for token-to-expert routing strategies."""
@@ -290,6 +294,12 @@ class RoutingSimulator:
294
f"Available strategies: "
295
f"{list(RoutingSimulator._routing_strategies.keys())}"
296
)
297
+ logger.warning_once(
298
+ "Simulating MoE routing using a %s strategy. "
299
+ "This should only be used for performance testing. "
300
+ "Model outputs will not be valid.",
301
+ strategy_name,
302
+ )
303
304
strategy = RoutingSimulator._routing_strategies[strategy_name]
305
return strategy.route_tokens(