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 12, 11:00 UTC
[Doc]: fixing typos in diverse files (#29492)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Nov 27, 2025 at 16:15 UTC
66d3d5422c9b90f1ee9593e1793e86f14e4eb3f4
6 files changed
+12
-10
vllm/benchmarks/serve.py
+2
-2
@@ -1005,7 +1005,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
1005
help="Key-value pairs (e.g, --header x-additional-info=0.3.3) "
1006
"for headers to be passed with each request. These headers override "
1007
"per backend constants and values set via environment variable, and "
1008
- "will be overriden by other arguments (such as request ids).",
1008
+ "will be overridden by other arguments (such as request ids).",
1009
)
1010
parser.add_argument(
1011
"--max-concurrency",
@@ -1138,7 +1138,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
1138
"--percentile-metrics",
1139
type=str,
1140
default=None,
1141
- help="Comma-separated list of selected metrics to report percentils. "
1141
+ help="Comma-separated list of selected metrics to report percentiles. "
1142
"This argument specifies the metrics to report percentiles. "
1143
'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
1144
'If not specified, defaults to "ttft,tpot,itl" for generative models '
vllm/config/parallel.py
+2
-2
@@ -238,9 +238,9 @@ class ParallelConfig:
238
cp_kv_cache_interleave_size: int = 1
239
"""Interleave size of kv_cache storage while using DCP or PCP.
240
For `total_cp_rank = pcp_rank * dcp_world_size + dcp_rank`,
241
- and `total_cp_world_size = pcp_world_size * dcp_world_szie`.
241
+ and `total_cp_world_size = pcp_world_size * dcp_world_size`.
242
store interleave_size tokens on total_cp_rank i,
243
- then store next interleave_size tokens on taotal_cp_rank i+1.
243
+ then store next interleave_size tokens on total_cp_rank i+1.
244
Interleave_size=1: token-level alignment, where token `i` is stored on
245
total_cp_rank `i % total_cp_world_size`.
246
Interleave_size=block_size: block-level alignment, where tokens are
vllm/lora/punica_wrapper/punica_base.py
+1
-1
@@ -173,7 +173,7 @@ class PunicaWrapperBase(PunicaWrapperABC):
173
vocab_size: int,
174
):
175
# NOTE We have remove lora extra vocab support for now. So we set
176
- # extra_vocab_size alwayzs to 0, and extra_vocab_size will be removed.
176
+ # extra_vocab_size always to 0, and extra_vocab_size will be removed.
177
178
extra_vocab_size = 0
179
(
vllm/model_executor/models/adapters.py
+2
-2
@@ -428,7 +428,7 @@ def load_weights_using_from_2_way_softmax(
428
)
429
if text_config.tie_word_embeddings:
430
# embed_tokens is the assumed name for input embeddings. If the model does not
431
- # have this attribute, we fallback to get_input_embeddings(), which is used by
431
+ # have this attribute, we fall back to get_input_embeddings(), which is used by
432
# the Transformers modeling backend.
433
embed_tokens = (
434
model.model.embed_tokens
@@ -486,7 +486,7 @@ def load_weights_no_post_processing(model, weights: Iterable[tuple[str, torch.Te
486
)
487
if text_config.tie_word_embeddings:
488
# embed_tokens is the assumed name for input embeddings. If the model does not
489
- # have this attribute, we fallback to get_input_embeddings(), which is used by
489
+ # have this attribute, we fall back to get_input_embeddings(), which is used by
490
# the Transformers modeling backend.
491
embed_tokens = (
492
model.model.embed_tokens
vllm/v1/sample/tpu/sampler.py
+1
-1
@@ -181,7 +181,7 @@ def apply_top_k_top_p(
181
after thresholding the logit using this cut-off, the remaining elements
182
shall constitute the top-p set.
183
184
- Note: in the case of tie (i.e. multipple cut-off elements present in the
184
+ Note: in the case of tie (i.e. multiple cut-off elements present in the
185
logit), all tie elements are included in the top-p set. In other words,
186
this function does not break ties. Instead, these tie tokens have equal
187
chance of being chosen during final sampling, so we can consider the tie
vllm/v1/worker/dp_utils.py
+4
-2
@@ -24,12 +24,14 @@ def _get_device_and_group(parallel_config: ParallelConfig):
24
device = get_dp_group().device
25
group = get_dp_group().device_group
26
27
- # Transfering this tensor from GPU to CPU will introduce a GPU sync
27
+ # Transferring this tensor from GPU to CPU will introduce a GPU sync
28
# point that could adversely affect performance of vllm with asynch
29
# scheduling. This environment variable exists to quickly disable
30
# this optimization if we run into this case.
31
if parallel_config.disable_nccl_for_dp_synchronization:
32
- logger.info_once("Using CPU all reduce to syncronize DP padding between ranks.")
32
+ logger.info_once(
33
+ "Using CPU all reduce to synchronize DP padding between ranks."
34
+ )
35
device = "cpu"
36
group = get_dp_group().cpu_group
37
return device, group