xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-08 15:17:06 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix MoE Routing Simulation (#28002)

Browse Source 
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
Tyler Michael Smith 2025-11-03 17:26:49 -05:00 committed by GitHub
parent ccd3e55e51
commit 3758757377
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 11 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -2066,7 +2066,7 @@ class FusedMoE(CustomOp):
)
# DeepSeekv2 uses grouped_top_k
if use_grouped_topk:
elif use_grouped_topk:
assert topk_group is not None
assert num_expert_group is not None
if is_rocm_aiter_moe_enabled():

10
vllm/model_executor/layers/fused_moe/routing_simulator.py
View File

@ -14,6 +14,10 @@ from typing import Any
import torch
from vllm.logger import init_logger
logger = init_logger(__name__)
class RoutingStrategy(ABC):
"""Base class for token-to-expert routing strategies."""
@ -290,6 +294,12 @@ class RoutingSimulator:
f"Available strategies: "
f"{list(RoutingSimulator._routing_strategies.keys())}"
)
logger.warning_once(
"Simulating MoE routing using a %s strategy. "
"This should only be used for performance testing. "
"Model outputs will not be valid.",
strategy_name,
)
strategy = RoutingSimulator._routing_strategies[strategy_name]
return strategy.route_tokens(