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 various files (#24821)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Sep 15, 2025 at 10:08 UTC
4979eb79da80a6dd7d4e52103053bf00b80c65cb
7 files changed
+7
-7
.buildkite/nightly-benchmarks/nightly-descriptions.md
+1
-1
@@ -8,7 +8,7 @@ This benchmark aims to:
8
9
Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html), scroll to the end.
10
11
-Latest reproduction guilde: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
11
+Latest reproduction guide: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
12
13
## Setup
14
vllm/model_executor/layers/quantization/moe_wna16.py
+1
-1
@@ -190,7 +190,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
190
group_size = self.quant_config.group_size
191
group_size_div_factor = 1
192
193
- # make intermediate_size and hidden_size diviable by group_size
193
+ # make intermediate_size and hidden_size divisible by group_size
194
# we reduce the group size to ensure that
195
# and we would repeat the loaded_weight later
196
while intermediate_size_per_partition % group_size or \
vllm/model_executor/layers/quantization/utils/marlin_utils_test.py
+1
-1
@@ -19,7 +19,7 @@ class MarlinWorkspace:
19
20
def __init__(self, out_features, min_thread_n, max_parallel):
21
assert (out_features % min_thread_n == 0), (
22
- "out_features = {} is undivisible by min_thread_n = {}".format(
22
+ "out_features = {} is indivisible by min_thread_n = {}".format(
23
out_features, min_thread_n))
24
25
max_workspace_size = ((out_features // min_thread_n) * max_parallel)
vllm/model_executor/layers/sampler.py
+1
-1
@@ -649,7 +649,7 @@ def _sample_with_torch(
649
else:
650
sampled_token_ids_tensor = None
651
652
- # Counterintiutively, having two loops here is actually faster.
652
+ # Counterintuitively, having two loops here is actually faster.
653
# The first loop can run without waiting on GPU<->CPU sync.
654
for sampling_type in SamplingType:
655
sample_indices = categorized_sample_indices[sampling_type]
vllm/model_executor/models/glm4_1v.py
+1
-1
@@ -1524,7 +1524,7 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
1524
return None
1525
1526
# The result multimodal_embeddings is tuple of tensors, with each
1527
- # tensor correspoending to a multimodal data item (image or video).
1527
+ # tensor corresponding to a multimodal data item (image or video).
1528
multimodal_embeddings: tuple[torch.Tensor, ...] = ()
1529
1530
# NOTE: It is important to iterate over the keys in this dictionary
vllm/model_executor/models/interns1.py
+1
-1
@@ -738,7 +738,7 @@ class InternS1ForConditionalGeneration(nn.Module, SupportsMultiModal,
738
return []
739
740
# The result multimodal_embeddings is tuple of tensors, with each
741
- # tensor correspoending to a multimodal data item (image or video).
741
+ # tensor corresponding to a multimodal data item (image or video).
742
multimodal_embeddings: tuple[torch.Tensor, ...] = ()
743
744
# NOTE: It is important to iterate over the keys in this dictionary
vllm/model_executor/models/ultravox.py
+1
-1
@@ -662,7 +662,7 @@ def pad_and_concat_to_dim3(
662
max_len = max(f.shape[-1] for f in features)
663
# Ensure all features have dim=3
664
features = [f.view(-1, *f.shape[-2:]) for f in features]
665
- # Pad and oncatenate:
665
+ # Pad and concatenate:
666
# [[B1, 80, M1], [B2, 80, M2]] -> [B1+B2, 80, max(M1, M2)]
667
features = [F.pad(f, (0, max_len - f.shape[-1])) for f in features]
668
return torch.cat(features)