[Doc] Fix duplicate words in comments (#36713)

Signed-off-by: Hongbin10 <jdmjdm1998@163.com>

vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py

@@ -264,7 +264,7 @@ class DefaultMoERunner(MoERunner):
             )
 
             # Record that the shared_experts_input will be used in the
-            # shared_experts_stream to to avoid gc issue from
+            # shared_experts_stream to avoid gc issue from
             # deallocation. For more details:
             # https://docs.pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html # noqa: E501
             # NOTE: We don't need shared_output.record_stream(current_stream())

vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

@@ -50,7 +50,7 @@ def swap_w13_to_w31(x: torch.Tensor) -> torch.Tensor:
 def rotate_weights_for_fi_trtllm_fp8_per_tensor_moe(
     gemm1_weights: torch.Tensor, gemm2_weights: torch.Tensor, is_gated_activation: bool
 ):
-    """Shuffle weights for for FI TRT-LLM Format"""
+    """Shuffle weights for FI TRT-LLM Format"""
     from flashinfer import reorder_rows_for_gated_act_gemm, shuffle_matrix_a
 
     epilogue_tile_m = 128

vllm/model_executor/models/transformers/pooling.py

@@ -57,7 +57,7 @@ class SequenceClassificationMixin(SupportsCrossEncoding, VllmModelForPooling):
         pooler_config = vllm_config.model_config.pooler_config
         assert pooler_config is not None
 
-        # Certain information about the the model and classifier can only be
+        # Certain information about the model and classifier can only be
         # inferred from the `ForSequenceClassification` class. Therefore, we
         # instantiate it on the "meta" device to avoid allocating GPU memory.
         with torch.device("meta"):

vllm/multimodal/video.py

@@ -952,7 +952,7 @@ class OpenCVDynamicOpenPanguVideoBackend(VideoLoader, OpenCVVideoBackendMixin):
             frame_recovery=frame_recovery,
         )
 
-        # Use transformers transformers.video_utils.VideoMetadata format
+        # Use transformers.video_utils.VideoMetadata format
         metadata = cls.create_hf_metadata(
             source=source,
             video_backend="opencv_dynamic",

vllm/tokenizers/mistral.py

@@ -44,7 +44,7 @@ def maybe_serialize_tool_calls(request: "MistralChatCompletionRequest"):
     # SEE: https://github.com/vllm-project/vllm/pull/9951
     # Credits go to: @gcalmettes
     # NOTE: There is currently a bug in pydantic where attributes
-    # declared as iterables are replaced in in the instances by
+    # declared as iterables are replaced in the instances by
     # pydantic-core ValidatorIterator instance. In particular, this
     # affects tool_calls defined in ChatCompletionAssistantMessageParam
     # model:

vllm/v1/worker/gpu_worker.py

@@ -1055,6 +1055,6 @@ def init_worker_distributed_environment(
         parallel_config.decode_context_parallel_size,
     )
 
-    # Init ec connector here before KV caches caches init
+    # Init ec connector here before KV caches init
     # NOTE: We do not init KV caches for Encoder-only instance in EPD disagg mode
     ensure_ec_transfer_initialized(vllm_config)