[Bugfix][CI] fix typos (#34934)

Signed-off-by: 1195343015 <1195343015@qq.com>
Signed-off-by: Jiayi Yan <66017932+1195343015@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

vllm/model_executor/models/qwen2_vl.py

@@ -916,7 +916,7 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
         self, max_pixels: int | None = None
     ) -> ImageSize:
         # NOTE: Simply processing a huge size with _get_vision_info might not give a
-        # size that maximizes the number of featrues, i.e., the number of (merged)
+        # size that maximizes the number of features, i.e., the number of (merged)
         # patches. This is because the number of patches limits the allowed aspect
         # ratios. For example, suppose the maximum number of patches is 1280. A square
         # image cannot be broken down into 1280 patches, so feeding a giant square image

vllm/model_executor/models/step3_vl.py

@@ -459,14 +459,14 @@ class Step3VLProcessor:
             image_inputs = {}
             text_inputs = self.tokenizer(text)
         else:
-            splitted_images_data = self._split_images(images)
+            split_images_data = self._split_images(images)
             pixel_values_lst = []
             patch_pixel_values_lst = []
             patch_newline_mask_lst = []
             image_repl_str_lst = []
             image_repl_ids_lst = []
             num_patches = []
-            for raw_img, img_patches, patch_newline_mask in splitted_images_data:
+            for raw_img, img_patches, patch_newline_mask in split_images_data:
                 pixel_values_lst.extend(self._convert_images_to_pixel_values([raw_img]))
 
                 if len(img_patches) > 0:

vllm/model_executor/models/step3p5.py

@@ -353,7 +353,7 @@ class FusedMoEBlock(nn.Module):
         if swiglu_limit not in (None, 0):
             swiglu_limit = float(swiglu_limit)
             assert swiglu_limit == 7.0, (
-                "Swiglu limit in fused moe block only suport 7.0 now."
+                "Swiglu limit in fused moe block only support 7.0 now."
             )
             activation = "swiglustep"
             logger.debug(

vllm/reasoning/ernie45_reasoning_parser.py

@@ -18,7 +18,7 @@ logger = init_logger(__name__)
 class Ernie45ReasoningParser(BaseThinkingReasoningParser):
     """
     Reasoning parser for Ernie45 thinking model.
-    The Ernie45 thinking model ouput format is
+    The Ernie45 thinking model output format is
         abc\n</think>\n\n<response>\ndef\n</response>\n
     or  abc\n</think>\ndef
     """
@@ -73,7 +73,7 @@ class Ernie45ReasoningParser(BaseThinkingReasoningParser):
         Extract reasoning content from a delta message.
         Handles streaming output where previous + delta = current.
         Uses token IDs for faster processing.
-        The Ernie45 thinking model ouput format is
+        The Ernie45 thinking model output format is
             abc\n</think>\n\n<response>\ndef\n</response>\n
         or  abc\n</think>\ndef
         - 'abc' goes to reasoning
@@ -148,7 +148,7 @@ class Ernie45ReasoningParser(BaseThinkingReasoningParser):
     ) -> tuple[str | None, str | None]:
         """
         Extract reasoning content from the model output.
-        The Ernie45 thinking model ouput format is
+        The Ernie45 thinking model output format is
             abc\n</think>\n\n\n<response>\ndef\n</response>\n
         or  abc\n</think>\ndef
         - 'abc' goes to reasoning

vllm/renderers/hf.py

@@ -564,7 +564,7 @@ def replace_vision_chunk_video_placeholder(
     mm_data: "MultiModalDataDict",
     video_placeholder: str | None,
 ) -> str | list[int]:
-    # get video placehoder, replace it with runtime video-chunk prompts
+    # get video placeholder, replace it with runtime video-chunk prompts
     if video_placeholder and isinstance(prompt_raw, str):
         video_prompts = build_video_prompts_from_mm_data(mm_data)
 

vllm/renderers/inputs/preprocess.py

 """
-Schemas and utilites for preprocessing inputs.
+Schemas and utilities for preprocessing inputs.
 """
 
 # SPDX-License-Identifier: Apache-2.0

vllm/renderers/inputs/tokenize.py

 """
-Schemas and utilites for tokenization inputs.
+Schemas and utilities for tokenization inputs.
 """
 
 # SPDX-License-Identifier: Apache-2.0

vllm/tokenizers/mistral.py

@@ -169,7 +169,7 @@ def _prepare_apply_chat_template_tools_and_messages(
                     tool.pop(tool_key)
                     logger.warning_once(
                         f"'{tool_key}' is not supported by mistral-common for tools. "
-                        "It has been poped from the tool definition."
+                        "It has been popped from the tool definition."
                     )
                 if tool["type"] == "function":
                     function_keys = list(tool["function"].keys())
@@ -178,7 +178,7 @@ def _prepare_apply_chat_template_tools_and_messages(
                             tool["function"].pop(function_key)
                             logger.warning_once(
                                 f"'{function_key}' is not supported by mistral-common "
-                                "for function tools. It has been poped from the "
+                                "for function tools. It has been popped from the "
                                 "function definition."
                             )
                 else:

vllm/transformers_utils/processors/ovis2_5.py

@@ -402,7 +402,7 @@ class Ovis2_5Processor(ProcessorMixin):
                 images = [images]
         elif video is not None:
             is_video = True
-            # type of vidoe in dummy_mm_data is np.ndarray
+            # type of video in dummy_mm_data is np.ndarray
             if isinstance(video, np.ndarray):
                 images = []
                 for i in range(video.shape[0]):

vllm/v1/attention/backends/cpu_attn.py

@@ -174,7 +174,7 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]
             query_start_loc = query_start_loc[: num_decodes + 1]
             block_table_tensor = block_table_tensor[:num_decodes]
 
-        sheduler_metadata = ops.cpu_attn_get_scheduler_metadata(
+        scheduler_metadata = ops.cpu_attn_get_scheduler_metadata(
             num_reqs=num_reqs,
             num_heads=self.num_heads,
             num_kv_heads=self.num_kv_heads,
@@ -197,7 +197,7 @@ class CPUAttentionMetadataBuilder(AttentionMetadataBuilder[CPUAttentionMetadata]
             seq_lens=seq_lens,
             block_table=block_table_tensor,
             slot_mapping=slot_mapping,
-            scheduler_metadata=sheduler_metadata,
+            scheduler_metadata=scheduler_metadata,
             causal=causal,
             use_sdpa_prefill=self.use_sdpa_prefill,
             num_decode_tokens=num_decode_tokens,

vllm/v1/attention/backends/mamba_attn.py

@@ -383,7 +383,7 @@ class BaseMambaAttentionMetadataBuilder(AttentionMetadataBuilder[M], abc.ABC):
 
             # Return a tensor of shape (#requests, #max blocks)
             state_indices_tensor = common_attn_metadata.block_table_tensor
-            # Additional cache-related varaiables:
+            # Additional cache-related variables:
             mamba_block_size = self.kv_cache_spec.block_size
             (
                 block_idx_last_computed_token,

vllm/v1/attention/backends/mla/flashmla_sparse.py

@@ -49,14 +49,14 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
-# For FP8 sparse attention we have two impelementations:
+# For FP8 sparse attention we have two implementations:
 # 1. Mixed batch mode: use the FP8 decode kernel for both prefill and decode this is
 #    done by treating all tokens as single batch.
 # 2. Separate prefill and decode mode: use the BF16 prefill kernel for prefill
 #    (upconverting the FP8 cache to BF16 then calling the prefill kernel) and using
 #    the FP8 decode kernel for decode.
 # Currently we use #1 when the number of heads per rank is low (i.e. TP) since the BF16
-# prefill kernel requires padding the numer of heads to 128 while the decode does not
+# prefill kernel requires padding the number of heads to 128 while the decode does not
 # so when the per ranke head count is below MIN_HEADS_FOR_BF16_PREFILL we use the mixed
 # batch mode (#2).
 MIN_HEADS_FOR_BF16_PREFILL = 32
@@ -126,7 +126,7 @@ class FlashMLASparseBackend(AttentionBackend):
         cache_dtype_str: str = "auto",
     ) -> tuple[int, ...]:
         if cache_dtype_str == "fp8_ds_mla":
-            # custom storage fromat is 656 bytes
+            # custom storage format is 656 bytes
             #  see FlashMLA readme.md for details
             return (num_blocks, block_size, 656)
         else:

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -370,7 +370,7 @@ class AiterFlashAttentionMetadata:
     slot_mapping: torch.Tensor
     block_table: torch.Tensor
 
-    # prefill and deocde split
+    # prefill and decode split
     num_decodes: int
     num_decode_tokens: int
     num_prefills: int
@@ -1099,7 +1099,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                 extend_tokens_slice = slice(
                     num_decode_tokens, num_decode_tokens + num_extend_tokens
                 )
-                extend_querys = query[extend_tokens_slice]
+                extend_queries = query[extend_tokens_slice]
                 extend_keys = key[extend_tokens_slice]
                 extend_values = value[extend_tokens_slice]
                 extend_outputs = output[extend_tokens_slice]
@@ -1110,7 +1110,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                     v_scale = attn_metadata.v_scale
                 self.extend_forward(
                     attn_metadata=attn_metadata,
-                    query=extend_querys,
+                    query=extend_queries,
                     key=extend_keys,
                     value=extend_values,
                     key_cache=key_cache,

vllm/v1/core/single_type_kv_cache_manager.py

@@ -863,7 +863,7 @@ class MambaManager(SingleTypeKVCacheManager):
         ):
             # Mamba can't rely on blocks generated by other requests in the current step
             # To put it in the next step, we return num_gpu_blocks + 1 so
-            # that kv_cache_manager will think there is no enough blocks to allocte now
+            # that kv_cache_manager will think there is no enough blocks to allocate now
             # and don't schedule it in the current step.
             return self.block_pool.num_gpu_blocks + 1
         if self.mamba_cache_mode != "align":

vllm/v1/engine/core.py

@@ -1724,11 +1724,11 @@ class DPEngineCoreProc(EngineCoreProc):
         """
         Send notifications to EngineCoreClient, which can then forward
         the notifications to other engine core processes. It is used for:
-        1) In scale up: new core engines to notify exisiting core engines
+        1) In scale up: new core engines to notify existing core engines
            that they are ready;
         2) In scale down: removing core engines to notify EngineCoreClient
            so EngineCoreClient can release their ray placement groups;
-        3) Both scale up/down: to notify EngineCoreClient that exisiting
+        3) Both scale up/down: to notify EngineCoreClient that existing
            core engines have already switched to the new parallel setup.
         """
         if vllm_config is None:

vllm/v1/engine/input_processor.py

@@ -194,7 +194,7 @@ class InputProcessor:
     @staticmethod
     def assign_request_id(request: EngineCoreRequest):
         """Replace the externally supplied request ID with an internal request ID
-        that adds 8 random characters in order to ensure uniquness.
+        that adds 8 random characters in order to ensure uniqueness.
         """
         if request.external_req_id is not None:
             raise ValueError(

vllm/v1/kv_offload/worker/cpu_gpu.py

@@ -197,7 +197,7 @@ class SingleDirectionOffloadingHandler(OffloadingHandler):
             transfer = self._transfers.popleft()
             transfer_time = (
                 transfer.start_event.elapsed_time(transfer.end_event) * 1e-3
-            )  # elapsed_time is in miliseconds
+            )  # elapsed_time is in milliseconds
             result = TransferResult(
                 job_id=transfer.job_id,
                 success=True,

vllm/v1/worker/gpu_model_runner.py

@@ -905,7 +905,7 @@ class GPUModelRunner(
         Args:
             scheduler_output: The scheduler output.
         """
-        # Attention free models have zero kv_cache_goups, however models
+        # Attention free models have zero kv_cache_groups, however models
         # like Mamba are also attention free but use the kv_cache for
         # keeping its internal state. This is why we check the number
         # of kv_cache groups instead of solely checking
@@ -1065,7 +1065,7 @@ class GPUModelRunner(
                 # of the request. for example:
                 # fist step: num_computed_tokens = 0, spec_tokens = [],
                 # prev_num_draft_len = 0.
-                # second step: num_computed_tokens = 100(prompt lenth),
+                # second step: num_computed_tokens = 100(prompt length),
                 # spec_tokens = [a,b], prev_num_draft_len = 0.
                 # third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
                 # prev_num_draft_len = 2.
@@ -1412,30 +1412,30 @@ class GPUModelRunner(
                 prev_draft_token_indices.extend(range(start, start + draft_len))
                 indices_match &= prev_index == flattened_index
                 max_flattened_index = max(max_flattened_index, flattened_index)
-        num_commmon_tokens = len(sample_flattened_indices)
+        num_common_tokens = len(sample_flattened_indices)
         total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
-        if num_commmon_tokens < total_without_spec:
+        if num_common_tokens < total_without_spec:
             # If not all requests are decodes from the last iteration,
             # We need to copy the input_ids_cpu to the GPU first.
             self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
             if self.enable_prompt_embeds:
                 self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
                 self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
-        if num_commmon_tokens == 0:
+        if num_common_tokens == 0:
             # No requests in common with the previous iteration
             # So input_ids.cpu will have all the input ids.
             return
-        if indices_match and max_flattened_index == (num_commmon_tokens - 1):
+        if indices_match and max_flattened_index == (num_common_tokens - 1):
             # Common-case optimization: the batch is unchanged
             # and no reordering happened.
             # The indices are both the same permutation of 0..N-1 so
             # we can copy directly using a single slice.
-            self.input_ids.gpu[:num_commmon_tokens].copy_(
-                self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
+            self.input_ids.gpu[:num_common_tokens].copy_(
+                self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
                 non_blocking=True,
             )
             if self.enable_prompt_embeds:
-                self.is_token_ids.gpu[:num_commmon_tokens] = True
+                self.is_token_ids.gpu[:num_common_tokens] = True
             return
         # Upload the index tensors asynchronously so the scatter can be non-blocking.
         sampled_tokens_index_tensor = torch.tensor(
@@ -4383,7 +4383,7 @@ class GPUModelRunner(
             self.model.compile(fullgraph=True, backend=backend)
             return
         # for other compilation modes, cudagraph behavior is controlled by
-        # CudagraphWraper and CudagraphDispatcher of vllm.
+        # CudagraphWrapper and CudagraphDispatcher of vllm.
 
         # wrap the model with full cudagraph wrapper if needed.
         cudagraph_mode = self.compilation_config.cudagraph_mode
@@ -4444,7 +4444,7 @@ class GPUModelRunner(
         :param weights_path: path to load weights from if weights_iterator is not
             provided. Use path of original model if neither is provided.
         :param is_checkpoint_format: set to False if weights have already been processed
-            into kernel format (repacking, renaming, ect.)
+            into kernel format (repacking, renaming, etc.)
         """
         # TODO(@kylesayrs): generalize to all runners and loaders
         # argument validation