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 15, 17:15 UTC
[Misc] cli auto show default value (#15582)
Signed-off-by: reidliu41 <reid201711@gmail.com>
Reid committed
Mar 29, 2025 at 06:23 UTC
26df46ee59e05882f7f46268f731a8e4b3ae0454
4 files changed
+11
-22
vllm/benchmarks/serve.py
+1
-3
@@ -726,15 +726,13 @@ def add_cli_args(parser: argparse.ArgumentParser):
726
default="ttft,tpot,itl",
727
help="Comma-seperated list of selected metrics to report percentils. "
728
"This argument specifies the metrics to report percentiles. "
729
- "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
730
- "Default value is \"ttft,tpot,itl\".")
729
+ "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". ")
730
parser.add_argument(
731
"--metric-percentiles",
732
type=str,
733
default="99",
734
help="Comma-seperated list of percentiles for selected metrics. "
735
"To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
737
- "Default value is \"99\". "
736
"Use \"--percentile-metrics\" to select metrics.",
737
)
738
parser.add_argument(
vllm/engine/arg_utils.py
+8
-17
@@ -322,9 +322,7 @@ class EngineArgs:
322
parser.add_argument('--download-dir',
323
type=nullable_str,
324
default=EngineArgs.download_dir,
325
- help='Directory to download and load the weights, '
326
- 'default to the default cache dir of '
327
- 'huggingface.')
325
+ help='Directory to download and load the weights.')
326
parser.add_argument(
327
'--load-format',
328
type=str,
@@ -399,8 +397,7 @@ class EngineArgs:
397
'Valid backend values are "xgrammar", "guidance", and "auto". '
398
'With "auto", we will make opinionated choices based on request'
399
'contents and what the backend libraries currently support, so '
402
- 'the behavior is subject to change in each release. '
403
- 'The default is xgrammar.')
400
+ 'the behavior is subject to change in each release.')
401
parser.add_argument(
402
'--logits-processor-pattern',
403
type=nullable_str,
@@ -493,8 +490,7 @@ class EngineArgs:
490
default=EngineArgs.prefix_caching_hash_algo,
491
help="Set the hash algorithm for prefix caching. "
492
"Options are 'builtin' (Python's built-in hash) or 'sha256' "
496
- "(collision resistant but with certain overheads). Defaults "
497
- "to 'builtin'.",
493
+ "(collision resistant but with certain overheads).",
494
)
495
parser.add_argument('--disable-sliding-window',
496
action='store_true',
@@ -568,9 +564,7 @@ class EngineArgs:
564
type=int,
565
default=EngineArgs.max_num_partial_prefills,
566
help="For chunked prefill, the max number of concurrent \
571
- partial prefills."
572
- "Defaults to 1",
573
- )
567
+ partial prefills.")
568
parser.add_argument(
569
"--max-long-partial-prefills",
570
type=int,
@@ -579,15 +573,13 @@ class EngineArgs:
573
"than --long-prefill-token-threshold that will be prefilled "
574
"concurrently. Setting this less than --max-num-partial-prefills "
575
"will allow shorter prompts to jump the queue in front of longer "
582
- "prompts in some cases, improving latency. Defaults to 1.")
576
+ "prompts in some cases, improving latency.")
577
parser.add_argument(
578
"--long-prefill-token-threshold",
579
type=float,
580
default=EngineArgs.long_prefill_token_threshold,
581
help="For chunked prefill, a request is considered long if the "
588
- "prompt is longer than this number of tokens. Defaults to 4%% of "
589
- "the model's context length.",
590
- )
582
+ "prompt is longer than this number of tokens.")
583
parser.add_argument('--max-num-seqs',
584
type=int,
585
default=EngineArgs.max_num_seqs,
@@ -739,8 +731,7 @@ class EngineArgs:
731
type=int,
732
default=EngineArgs.max_cpu_loras,
733
help=('Maximum number of LoRAs to store in CPU memory. '
742
- 'Must be >= than max_loras. '
743
- 'Defaults to max_loras.'))
734
+ 'Must be >= than max_loras.'))
735
parser.add_argument(
736
'--fully-sharded-loras',
737
action='store_true',
@@ -894,7 +885,7 @@ class EngineArgs:
885
help='Set the lower bound threshold for the posterior '
886
'probability of a token to be accepted. This threshold is '
887
'used by the TypicalAcceptanceSampler to make sampling decisions '
897
- 'during speculative decoding. Defaults to 0.09')
888
+ 'during speculative decoding.')
889
890
parser.add_argument(
891
'--typical-acceptance-sampler-posterior-alpha',
vllm/entrypoints/openai/cli_args.py
+1
-1
@@ -247,7 +247,7 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
247
default=None,
248
help='Max number of prompt characters or prompt '
249
'ID numbers being printed in log.'
250
- '\n\nDefault: Unlimited')
250
+ ' The default of None means unlimited.')
251
252
parser.add_argument(
253
"--disable-fastapi-docs",
vllm/utils.py
+1
-1
@@ -1212,7 +1212,7 @@ class StoreBoolean(argparse.Action):
1212
"Expected 'true' or 'false'.")
1213
1214
1215
-class SortedHelpFormatter(argparse.HelpFormatter):
1215
+class SortedHelpFormatter(argparse.ArgumentDefaultsHelpFormatter):
1216
"""SortedHelpFormatter that sorts arguments by their option strings."""
1217
1218
def add_arguments(self, actions):