biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Move missed SchedulerConfig args into scheduler config group in EngineArgs (#17131)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-04-25 06:48:53 +01:00
committed by GitHub
parent 649818995f
commit 6ca0234478
2 changed files with 9 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/config.py
View File

@@ -1778,6 +1778,7 @@ class ParallelConfig:
"worker_extension_cls must be a string (qualified class name).") "worker_extension_cls must be a string (qualified class name).")
PreemptionMode = Literal["swap", "recompute"]
SchedulerPolicy = Literal["fcfs", "priority"] SchedulerPolicy = Literal["fcfs", "priority"]
@@ -1854,7 +1855,7 @@ class SchedulerConfig:
NOTE: This is not currently configurable. It will be overridden by NOTE: This is not currently configurable. It will be overridden by
max_num_batched_tokens in case max multimodal embedding size is larger.""" max_num_batched_tokens in case max multimodal embedding size is larger."""
preemption_mode: Optional[str] = None preemption_mode: Optional[PreemptionMode] = None
"""Whether to perform preemption by swapping or """Whether to perform preemption by swapping or
recomputation. If not specified, we determine the mode as follows: recomputation. If not specified, we determine the mode as follows:
We use recomputation by default since it incurs lower overhead than We use recomputation by default since it incurs lower overhead than

23
vllm/engine/arg_utils.py
View File

@@ -753,12 +753,6 @@ class EngineArgs:
) )
device_group.add_argument("--device", **device_kwargs["device"]) device_group.add_argument("--device", **device_kwargs["device"])
parser.add_argument('--num-scheduler-steps',
type=int,
default=1,
help=('Maximum number of forward steps per '
'scheduler call.'))
# Speculative arguments # Speculative arguments
speculative_group = parser.add_argument_group( speculative_group = parser.add_argument_group(
title="SpeculativeConfig", title="SpeculativeConfig",
@@ -779,13 +773,6 @@ class EngineArgs:
help="The pattern(s) to ignore when loading the model." help="The pattern(s) to ignore when loading the model."
"Default to `original/**/*` to avoid repeated loading of llama's " "Default to `original/**/*` to avoid repeated loading of llama's "
"checkpoints.") "checkpoints.")
parser.add_argument(
'--preemption-mode',
type=str,
default=None,
help='If \'recompute\', the engine performs preemption by '
'recomputing; If \'swap\', the engine performs preemption by '
'block swapping.')
parser.add_argument( parser.add_argument(
"--served-model-name", "--served-model-name",
@@ -865,14 +852,18 @@ class EngineArgs:
**scheduler_kwargs["num_lookahead_slots"]) **scheduler_kwargs["num_lookahead_slots"])
scheduler_group.add_argument('--scheduler-delay-factor', scheduler_group.add_argument('--scheduler-delay-factor',
**scheduler_kwargs["delay_factor"]) **scheduler_kwargs["delay_factor"])
scheduler_group.add_argument( scheduler_group.add_argument('--preemption-mode',
'--enable-chunked-prefill', **scheduler_kwargs["preemption_mode"])
**scheduler_kwargs["enable_chunked_prefill"]) scheduler_group.add_argument('--num-scheduler-steps',
**scheduler_kwargs["num_scheduler_steps"])
scheduler_group.add_argument( scheduler_group.add_argument(
'--multi-step-stream-outputs', '--multi-step-stream-outputs',
**scheduler_kwargs["multi_step_stream_outputs"]) **scheduler_kwargs["multi_step_stream_outputs"])
scheduler_group.add_argument('--scheduling-policy', scheduler_group.add_argument('--scheduling-policy',
**scheduler_kwargs["policy"]) **scheduler_kwargs["policy"])
scheduler_group.add_argument(
'--enable-chunked-prefill',
**scheduler_kwargs["enable_chunked_prefill"])
scheduler_group.add_argument( scheduler_group.add_argument(
"--disable-chunked-mm-input", "--disable-chunked-mm-input",
**scheduler_kwargs["disable_chunked_mm_input"]) **scheduler_kwargs["disable_chunked_mm_input"])