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 (#24726)
Signed-off-by: Didier Durand <durand.didier@gmail.com>
Didier Durand committed
Sep 12, 2025 at 15:43 UTC
bcb06d7baf22a80383bcc25251a1dcebfe9a3919
11 files changed
+11
-11
benchmarks/kernels/benchmark_w8a8_block_fp8.py
+1
-1
@@ -56,7 +56,7 @@ def w8a8_block_matmul(
56
Bs: The per-block quantization scale for `B`.
57
block_size: The block size for per-block quantization.
58
It should be 2-dim, e.g., [128, 128].
59
- output_dytpe: The dtype of the returned tensor.
59
+ output_dtype: The dtype of the returned tensor.
60
61
Returns:
62
torch.Tensor: The result of matmul.
csrc/cpu/cpu_types_vxe.hpp
+1
-1
@@ -12,7 +12,7 @@ namespace vec_op {
12
#define vec_sub(a, b) ((a) - (b))
13
#define vec_mul(a, b) ((a) * (b))
14
#define vec_div(a, b) ((a) / (b))
15
-#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebaic
15
+#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic
16
#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left
17
18
// FIXME: FP16 is not fully supported in Torch-CPU
csrc/cpu/sgl-kernels/moe.cpp
+1
-1
@@ -215,7 +215,7 @@ int moe_align_block_size(
215
offsets[mb + 1] = sorted_id_size(sorted_ids + mb * BLOCK_M);
216
}
217
});
218
- // TODO: do we need to vecterize this ?
218
+ // TODO: do we need to vectorize this ?
219
for (int mb = 0; mb < num_token_blocks; ++mb) {
220
offsets[mb + 1] += offsets[mb];
221
}
docs/design/multiprocessing.md
+1
-1
@@ -8,7 +8,7 @@ page for information on known issues and how to solve them.
8
## Introduction
9
10
!!! important
11
- The source code references are to the state of the code at the time of writing in December, 2024.
11
+ The source code references are to the state of the code at the time of writing in December 2024.
12
13
The use of Python multiprocessing in vLLM is complicated by:
14
vllm/attention/backends/flash_attn.py
+1
-1
@@ -901,7 +901,7 @@ def _get_query_key_seq_metadata(
901
attn_metadata.encoder_seq_start_loc,
902
attn_metadata.max_encoder_seq_len)
903
elif attn_type == AttentionType.ENCODER:
904
- # For encoder attention both the query and the key are same i.e the
904
+ # For encoder attention both the query and the key are same i.e. the
905
# encoder sequence.
906
return (attn_metadata.encoder_seq_start_loc,
907
attn_metadata.max_encoder_seq_len,
vllm/benchmarks/datasets.py
+1
-1
@@ -551,7 +551,7 @@ class RandomDataset(BenchmarkDataset):
551
[6880, 6881] -> ['Ġcalls', 'here'] ->
552
[1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
553
To avoid uncontrolled change of the prompt length,
554
- the encoded sequence is truncated before being decode again.
554
+ the encoded sequence is truncated before being decoded again.
555
"""
556
# Build the inner sequence by sampling sequentially from the vocab
557
inner_seq = ((offset + index + np.arange(input_len))
vllm/entrypoints/openai/protocol.py
+1
-1
@@ -242,7 +242,7 @@ def get_logits_processors(processors: Optional[LogitsProcessors],
242
elif processors:
243
raise ValueError(
244
"The `logits_processors` argument is not supported by this "
245
- "server. See --logits-processor-pattern engine argugment "
245
+ "server. See --logits-processor-pattern engine argument "
246
"for more information.")
247
return None
248
vllm/model_executor/layers/mamba/mamba_mixer2.py
+1
-1
@@ -324,7 +324,7 @@ class MambaMixer2(MambaBase, CustomOp):
324
# - the weight already has a "weight_loader" attribute
325
# which set_weight_attrs will raise if we do not
326
# delete before trying to override it
327
- # - ditto for the otther two weights below
327
+ # - ditto for the other two weights below
328
delattr(self.conv1d.bias, "weight_loader")
329
set_weight_attrs(
330
self.conv1d.bias,
vllm/model_executor/models/minicpmv.py
+1
-1
@@ -1117,7 +1117,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
1117
1118
def _process_multimodal_inputs(self, modalities: dict):
1119
# The result multimodal_embeddings is tuple of tensors, with each
1120
- # tensor correspoending to a multimodal data item (image or video).
1120
+ # tensor corresponding to a multimodal data item (image or video).
1121
multimodal_embeddings: tuple[torch.Tensor, ...] = ()
1122
1123
# NOTE: It is important to iterate over the keys in this dictionary
vllm/v1/worker/gpu_model_runner.py
+1
-1
@@ -2659,7 +2659,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
2659
num_tokens += num_pad
2660
2661
# If cudagraph_mode.decode_mode() == FULL and
2662
- # cudagraph_mode.seperate_routine(). This means that we are using
2662
+ # cudagraph_mode.separate_routine(). This means that we are using
2663
# different graphs and/or modes for mixed prefill-decode batches vs.
2664
# uniform decode batches. A uniform decode batch means that all
2665
# requests have identical query length, except a potential virtual
vllm/v1/worker/tpu_input_batch.py
+1
-1
@@ -392,7 +392,7 @@ class InputBatch:
392
# NOTE: the following is unsafe
393
# self.token_ids_cpu[i1, ...], self.token_ids_cpu[i2, ...], =\
394
# self.token_ids_cpu[i2, ...], self.token_ids_cpu[i1, ...]
395
- # instead, we need to temporiarily copy the data for one of the indices
395
+ # instead, we need to temporarily copy the data for one of the indices
396
# TODO(lucas): optimize this by only copying valid indices
397
tmp = self.token_ids_cpu[i1, ...].copy()
398
self.token_ids_cpu[i1, ...] = self.token_ids_cpu[i2, ...]