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 11, 17:30 UTC
Fix typos in comments across multiple files (#30345)
Signed-off-by: Wilson Wu <iwilsonwu@gmail.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Wilson Wu committed
Dec 10, 2025 at 12:05 UTC
3bdd426636cec97d4cd5cff0e1a057b45429e07c
7 files changed
+7
-7
csrc/cpu/cpu_attn_impl.hpp
+1
-1
@@ -186,7 +186,7 @@ struct AttentionMetadata {
186
// - Intermediate outputs: q_tile_size * head_dim * output_buffer_elem_size + 2
187
// * q_tile_size * 4, partial output, max + sum (float)
188
// Reduction scratchpad contains:
189
-// - flags: bool array to indicate wether the split is finished
189
+// - flags: bool array to indicate whether the split is finished
190
// - outputs: split_num * q_tile_size * head_dim * output_buffer_elem_size
191
// - max, sum: 2 * split_num * q_tile_size * 4
192
class AttentionScratchPad {
csrc/quantization/machete/machete_mainloop.cuh
+1
-1
@@ -617,7 +617,7 @@ struct MacheteCollectiveMma {
617
618
// Same as upstream, should be kept the same when possible, not formatted for
619
// easier comparison
620
- // with `SwapAB ? N : M -> M` since we dont support SwapAB
620
+ // with `SwapAB ? N : M -> M` since we don't support SwapAB
621
// clang-format off
622
template<class ProblemShape>
623
static bool
docs/features/nixl_connector_usage.md
+1
-1
@@ -22,7 +22,7 @@ python tools/install_nixl_from_source_ubuntu.py
22
NixlConnector uses NIXL library for underlying communication, which supports multiple transport backends. UCX (Unified Communication X) is the primary default transport library used by NIXL. Configure transport environment variables:
23
24
```bash
25
-# Example UCX configuration, adjust according to your enviroment
25
+# Example UCX configuration, adjust according to your environment
26
export UCX_TLS=all # or specify specific transports like "rc,ud,sm,^cuda_ipc" ..etc
27
export UCX_NET_DEVICES=all # or specify network devices like "mlx5_0:1,mlx5_1:1"
28
```
vllm/model_executor/layers/fused_moe/layer.py
+1
-1
@@ -881,7 +881,7 @@ class FusedMoE(CustomOp):
881
# Record that the clone will be used by shared_experts_stream
882
# to avoid gc issue from deallocation of hidden_states_clone
883
# For more details: https://docs.pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html # noqa: E501
884
- # NOTE: We dont need shared_output.record_stream(current_stream())
884
+ # NOTE: We don't need shared_output.record_stream(current_stream())
885
# because we synch the streams before using shared_output.
886
hidden_states_clone.record_stream(self.shared_experts_stream)
887
vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_nvfp4.py
+1
-1
@@ -28,7 +28,7 @@ class CompressedTensorsW4A16Fp4(CompressedTensorsScheme):
28
29
@classmethod
30
def get_min_capability(cls) -> int:
31
- # dont restrict as emulations
31
+ # don't restrict as emulations
32
return 80
33
34
def create_weights(
vllm/v1/worker/gpu_model_runner.py
+1
-1
@@ -4871,7 +4871,7 @@ class GPUModelRunner(
4871
# we need to adjust the cudagraph sizes to be a multiple of the uniform
4872
# decode query length to avoid: https://github.com/vllm-project/vllm/issues/28207
4873
# temp-fix: https://github.com/vllm-project/vllm/issues/28207#issuecomment-3504004536
4874
- # Will be removed in the near future when we have seperate cudagraph capture
4874
+ # Will be removed in the near future when we have separate cudagraph capture
4875
# sizes for decode and mixed prefill-decode.
4876
if (
4877
cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
vllm/v1/worker/utils.py
+1
-1
@@ -135,7 +135,7 @@ class AttentionGroup:
135
kv_cache_spec: KVCacheSpec
136
kv_cache_group_id: int
137
# When ubatching is enabled we will have a metadata builder for each ubatch
138
- # so that if they use internal persistant buffers for cudagraphs, and they
138
+ # so that if they use internal persistent buffers for cudagraphs, and they
139
# won't have to worry about conflicting with the other ubatches.
140
metadata_builders: list[AttentionMetadataBuilder] = field(
141
default_factory=lambda: []