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 2, 01:00 UTC
fix(config): validate max_num_scheduled_tokens >= 0 on all paths (#44207)
Signed-off-by: Oxygen56 <1391083091@qq.com>
Willow Lopez committed
Jun 4, 2026 at 00:06 UTC
51e0c579b0e51445662caf601930db26b3636a7f
2 files changed
+2
-2
vllm/config/scheduler.py
+1
-1
@@ -53,7 +53,7 @@ class SchedulerConfig:
53
In real usage, this should be set in `EngineArgs.create_engine_config`.
54
"""
55
56
- max_num_scheduled_tokens: int | None = None
56
+ max_num_scheduled_tokens: int | None = Field(default=None, ge=0)
57
"""Maximum number of tokens that the scheduler may issue in a single iteration.
58
59
This is usually equal to max_num_batched_tokens, but can be smaller in cases
vllm/v1/core/sched/scheduler.py
+1
-1
@@ -104,7 +104,7 @@ class Scheduler(SchedulerInterface):
104
self.max_num_running_reqs = self.scheduler_config.max_num_seqs
105
self.max_num_scheduled_tokens = (
106
self.scheduler_config.max_num_scheduled_tokens
107
- if self.scheduler_config.max_num_scheduled_tokens
107
+ if self.scheduler_config.max_num_scheduled_tokens is not None
108
else self.scheduler_config.max_num_batched_tokens
109
)
110
self.max_model_len = vllm_config.model_config.max_model_len