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 14, 05:30 UTC
[Doc]: fix typos in Python comments (#24115)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Sep 3, 2025 at 06:14 UTC
02d411fdb232f31ce46c6f8076bfe9e5acf88fc9
11 files changed
+11
-11
.buildkite/nightly-benchmarks/scripts/compare-json-results.py
+1
-1
@@ -218,7 +218,7 @@ if __name__ == "__main__":
218
"--xaxis",
219
type=str,
220
default="# of max concurrency.",
221
- help="column name to use as X Axis in comparision graph",
221
+ help="column name to use as X Axis in comparison graph",
222
)
223
args = parser.parse_args()
224
benchmarks/benchmark_serving.py
+1
-1
@@ -1104,7 +1104,7 @@ def create_argument_parser():
1104
"--percentile-metrics",
1105
type=str,
1106
default="ttft,tpot,itl",
1107
- help="Comma-separated list of selected metrics to report percentils. "
1107
+ help="Comma-separated list of selected metrics to report percentiles. "
1108
"This argument specifies the metrics to report percentiles. "
1109
'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
1110
'Default value is "ttft,tpot,itl".',
benchmarks/benchmark_serving_structured_output.py
+1
-1
@@ -998,7 +998,7 @@ def create_argument_parser():
998
"--percentile-metrics",
999
type=str,
1000
default="ttft,tpot,itl",
1001
- help="Comma-separated list of selected metrics to report percentils. "
1001
+ help="Comma-separated list of selected metrics to report percentiles. "
1002
"This argument specifies the metrics to report percentiles. "
1003
'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
1004
'Default value is "ttft,tpot,itl".',
benchmarks/benchmark_throughput.py
+1
-1
@@ -719,7 +719,7 @@ def create_argument_parser():
719
"[length * (1 - range_ratio), length * (1 + range_ratio)].",
720
)
721
722
- # hf dtaset
722
+ # hf dataset
723
parser.add_argument(
724
"--hf-subset", type=str, default=None, help="Subset of the HF dataset."
725
)
tools/profiler/visualize_layerwise_profile.py
+1
-1
@@ -119,7 +119,7 @@ def attempt_to_make_names_unique(entries_and_traces):
119
if not all_the_same(trace_eles)), None)
120
121
if first_trace_difference is None:
122
- # can't create a unique name, leave them names as the
122
+ # can't create a unique name, leave the names as they
123
# are they will get aggregated by the pivot_table call
124
continue
125
vllm/compilation/collective_fusion.py
+1
-1
@@ -513,7 +513,7 @@ if flashinfer_comm is not None:
513
torch.ops._C.static_scaled_fp8_quant(
514
quant_out, norm_out, scale_factor)
515
if scale_factor is None or norm_out is not None:
516
- # we need to return allreduce outpput
516
+ # we need to return allreduce output
517
# in cases of non quant fused AR + RMS norm
518
# and fused AR + RMS norm + quant without fused add
519
allreduce_in.copy_(allreduce_out)
vllm/engine/multiprocessing/engine.py
+1
-1
@@ -49,7 +49,7 @@ class MQLLMEngine:
49
50
This class is used to wrap the
51
[`LLMEngine`][vllm.engine.llm_engine.LLMEngine] class to enable use
52
- in concurrnet manner. It runs a background loop and uses zeromq to
52
+ in concurrent manner. It runs a background loop and uses zeromq to
53
receive new requests and stream outputs incrementally via ipc.
54
55
The [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] generate or encode
vllm/model_executor/layers/quantization/utils/w8a8_utils.py
+1
-1
@@ -23,7 +23,7 @@ TORCH_DEVICE_IDENTITY = None
23
# The condition to determine if it is on a platform that supports
24
# torch._scaled_mm rowwise feature.
25
# The condition is determined once as the operations
26
-# are time consuming.
26
+# are time-consuming.
27
USE_ROWWISE_TORCH_SCALED_MM = (current_platform.is_rocm() and version.parse(
28
torch.__version__) >= version.parse("2.7")
29
and current_platform.has_device_capability(94))
vllm/model_executor/model_loader/default_loader.py
+1
-1
@@ -211,7 +211,7 @@ class DefaultModelLoader(BaseModelLoader):
211
212
if not USE_TPU_COMMONS:
213
# In PyTorch XLA, we should call `xm.mark_step`
214
- # requently so that not too many ops are accumulated
214
+ # frequently so that not too many ops are accumulated
215
# in the XLA program. import torch_xla.core.xla_model
216
# as xm
217
import torch_xla.core.xla_model as xm
vllm/v1/worker/xpu_worker.py
+1
-1
@@ -84,7 +84,7 @@ class XPUWorker(Worker):
84
"""Profiles the peak memory usage of the model to determine how many
85
KV blocks may be allocated without OOMs.
86
The engine will first conduct a profiling of the existing memory usage.
87
- Then, it calculate the maximum possible number of GPU and CPU blocks
87
+ Then, it calculates the maximum possible number of GPU and CPU blocks
88
that can be allocated with the remaining free memory.
89
.. tip::
90
You may limit the usage of GPU memory
vllm/worker/worker.py
+1
-1
@@ -234,7 +234,7 @@ class Worker(LocalOrDistributedWorkerBase):
234
KV blocks may be allocated without OOMs.
235
236
The engine will first conduct a profiling of the existing memory usage.
237
- Then, it calculate the maximum possible number of GPU and CPU blocks
237
+ Then, it calculates the maximum possible number of GPU and CPU blocks
238
that can be allocated with the remaining free memory.
239
240
Tip: