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 6, 23:00 UTC
[Doc] Fix duplicate words in comments (#36713)
Signed-off-by: Hongbin10 <jdmjdm1998@163.com>
Hongbin Guo committed
Mar 11, 2026 at 12:28 UTC
4bf533623b6957ed27ba3a026df86ee9e2b9685e
6 files changed
+6
-6
vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py
+1
-1
@@ -264,7 +264,7 @@ class DefaultMoERunner(MoERunner):
264
)
265
266
# Record that the shared_experts_input will be used in the
267
- # shared_experts_stream to to avoid gc issue from
267
+ # shared_experts_stream to avoid gc issue from
268
# deallocation. For more details:
269
# https://docs.pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html # noqa: E501
270
# NOTE: We don't need shared_output.record_stream(current_stream())
vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
+1
-1
@@ -50,7 +50,7 @@ def swap_w13_to_w31(x: torch.Tensor) -> torch.Tensor:
50
def rotate_weights_for_fi_trtllm_fp8_per_tensor_moe(
51
gemm1_weights: torch.Tensor, gemm2_weights: torch.Tensor, is_gated_activation: bool
52
):
53
- """Shuffle weights for for FI TRT-LLM Format"""
53
+ """Shuffle weights for FI TRT-LLM Format"""
54
from flashinfer import reorder_rows_for_gated_act_gemm, shuffle_matrix_a
55
56
epilogue_tile_m = 128
vllm/model_executor/models/transformers/pooling.py
+1
-1
@@ -57,7 +57,7 @@ class SequenceClassificationMixin(SupportsCrossEncoding, VllmModelForPooling):
57
pooler_config = vllm_config.model_config.pooler_config
58
assert pooler_config is not None
59
60
- # Certain information about the the model and classifier can only be
60
+ # Certain information about the model and classifier can only be
61
# inferred from the `ForSequenceClassification` class. Therefore, we
62
# instantiate it on the "meta" device to avoid allocating GPU memory.
63
with torch.device("meta"):
vllm/multimodal/video.py
+1
-1
@@ -952,7 +952,7 @@ class OpenCVDynamicOpenPanguVideoBackend(VideoLoader, OpenCVVideoBackendMixin):
952
frame_recovery=frame_recovery,
953
)
954
955
- # Use transformers transformers.video_utils.VideoMetadata format
955
+ # Use transformers.video_utils.VideoMetadata format
956
metadata = cls.create_hf_metadata(
957
source=source,
958
video_backend="opencv_dynamic",
vllm/tokenizers/mistral.py
+1
-1
@@ -44,7 +44,7 @@ def maybe_serialize_tool_calls(request: "MistralChatCompletionRequest"):
44
# SEE: https://github.com/vllm-project/vllm/pull/9951
45
# Credits go to: @gcalmettes
46
# NOTE: There is currently a bug in pydantic where attributes
47
- # declared as iterables are replaced in in the instances by
47
+ # declared as iterables are replaced in the instances by
48
# pydantic-core ValidatorIterator instance. In particular, this
49
# affects tool_calls defined in ChatCompletionAssistantMessageParam
50
# model:
vllm/v1/worker/gpu_worker.py
+1
-1
@@ -1055,6 +1055,6 @@ def init_worker_distributed_environment(
1055
parallel_config.decode_context_parallel_size,
1056
)
1057
1058
- # Init ec connector here before KV caches caches init
1058
+ # Init ec connector here before KV caches init
1059
# NOTE: We do not init KV caches for Encoder-only instance in EPD disagg mode
1060
ensure_ec_transfer_initialized(vllm_config)