fix some typos (#24071)

Signed-off-by: co63oc <co63oc@users.noreply.github.com>

benchmarks/benchmark_block_pool.py

@@ -57,7 +57,7 @@ def invoke_main() -> None:
         "--num-iteration",
         type=int,
         default=1000,
-        help="Number of iterations to run to stablize final data readings",
+        help="Number of iterations to run to stabilize final data readings",
     )
     parser.add_argument(
         "--allocate-blocks",

benchmarks/benchmark_ngram_proposer.py

@@ -77,7 +77,7 @@ def invoke_main() -> None:
         "--num-iteration",
         type=int,
         default=100,
-        help="Number of iterations to run to stablize final data readings",
+        help="Number of iterations to run to stabilize final data readings",
     )
     parser.add_argument(
         "--num-req", type=int, default=128, help="Number of requests in the batch"

csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu

@@ -181,7 +181,7 @@ struct W4A8GemmKernel {
     auto A_ptr = static_cast<MmaType const*>(A.const_data_ptr());
     auto B_ptr = static_cast<QuantType const*>(B.const_data_ptr());
     auto D_ptr = static_cast<ElementD*>(D.data_ptr());
-    // can we avoid harcode the 8 here
+    // can we avoid hardcode the 8 here
     auto S_ptr =
         static_cast<cutlass::Array<ElementScale, ScalePackSize> const*>(
             group_scales.const_data_ptr());

docs/configuration/optimization.md

@@ -210,7 +210,7 @@ vllm serve Qwen/Qwen2.5-VL-3B-Instruct --api-server-count 4 -dp 2
 
 !!! note
     API server scale-out disables [multi-modal IPC caching](#ipc-caching)
-    because it requires a one-to-one correspondance between API and engine core processes.
+    because it requires a one-to-one correspondence between API and engine core processes.
 
     This does not impact [multi-modal processor caching](#processor-caching).
 
@@ -227,7 +227,7 @@ to avoid repeatedly processing the same multi-modal inputs in `BaseMultiModalPro
 ### IPC Caching
 
 Multi-modal IPC caching is automatically enabled when
-there is a one-to-one correspondance between API (`P0`) and engine core (`P1`) processes,
+there is a one-to-one correspondence between API (`P0`) and engine core (`P1`) processes,
 to avoid repeatedly transferring the same multi-modal inputs between them.
 
 ### Configuration

docs/design/io_processor_plugins.md

@@ -2,7 +2,7 @@
 
 IO Processor plugins are a feature that allows pre and post processing of the model input and output for pooling models. The idea is that users are allowed to pass a custom input to vLLM that is converted into one or more model prompts and fed to the model `encode` method. One potential use-case of such plugins is that of using vLLM for generating multi-modal data. Say users feed an image to vLLM and get an image in output.
 
-When performing an inference with IO Processor plugins, the prompt type is defined by the plugin and the same is valid for the final request output. vLLM does not perform any validation of input/output data, and it is up to the plugin to ensure the correct data is being fed to the model and returned to the user. As of now these plugins support only pooling models and can be triggerd via the `encode` method in `LLM` and `AsyncLLM`, or in online serving mode via the `/pooling` endpoint.
+When performing an inference with IO Processor plugins, the prompt type is defined by the plugin and the same is valid for the final request output. vLLM does not perform any validation of input/output data, and it is up to the plugin to ensure the correct data is being fed to the model and returned to the user. As of now these plugins support only pooling models and can be triggered via the `encode` method in `LLM` and `AsyncLLM`, or in online serving mode via the `/pooling` endpoint.
 
 ## Writing an IO Processor Plugin
 

examples/offline_inference/prithvi_geospatial_mae_io_processor.py

@@ -12,7 +12,7 @@ from vllm.pooling_params import PoolingParams
 # multimodal data. In this specific case this example will take a geotiff
 # image as input, process it using the multimodal data processor, and
 # perform inference.
-# Reuirement - install plugin at:
+# Requirement - install plugin at:
 #   https://github.com/christian-pinto/prithvi_io_processor_plugin
 
 

examples/online_serving/prithvi_geospatial_mae.py

@@ -10,7 +10,7 @@ import requests
 # multimodal data. In this specific case this example will take a geotiff
 # image as input, process it using the multimodal data processor, and
 # perform inference.
-# Reuirements :
+# Requirements :
 # - install plugin at:
 #   https://github.com/christian-pinto/prithvi_io_processor_plugin
 # - start vllm in serving mode with the below args

tests/compile/piecewise/test_multiple_graphs.py

@@ -134,7 +134,7 @@ class SimpleModelWithTwoGraphs(ParentModel):
         # Test will fail without set_model_tag here with error:
         # "ValueError: too many values to unpack (expected 3)"
         # This is because CompiledAttention and CompiledAttentionTwo
-        # have different implmentations but the same torch.compile
+        # have different implementations but the same torch.compile
         # cache dir will be used as default prefix is 'model_tag'
         with set_model_tag("attn_one"):
             self.attn_one = CompiledAttention(

tests/kernels/moe/test_mxfp4_moe.py

@@ -224,7 +224,7 @@ def tg_mxfp4_moe(
     assert (w2_bias.dim() == 2 and w2_bias.shape[0] == num_experts
             and w2_bias.shape[1] == hidden_size)
 
-    # Swap w1 and w3 as the defenition of
+    # Swap w1 and w3 as the definition of
     # swiglu is different in the trtllm-gen
     w13_weight_scale_ = w13_weight_scale.clone()
     w13_weight_ = w13_weight.clone()

tests/models/multimodal/processing/test_mllama4.py

@@ -52,7 +52,7 @@ def test_profiling(model_id: str, max_model_len: int):
     chunks_per_image = prod(mm_data["patches_per_image"])
     total_num_patches = chunks_per_image * tokens_per_patch
     num_tiles = mm_data["aspect_ratios"][0][0] * mm_data["aspect_ratios"][0][
-        1]  # x-y seperator tokens
+        1]  # x-y separator tokens
     total_tokens = total_num_patches.item() + num_tiles.item(
     ) + 3  # image start, image, image end
 

tests/quantization/test_modelopt.py

@@ -27,7 +27,7 @@ def use_v0_only(monkeypatch):
                     reason="ModelOpt FP8 is not supported on this GPU type.")
 def test_modelopt_fp8_checkpoint_setup(vllm_runner):
     """Test ModelOpt FP8 checkpoint loading and structure validation."""
-    # TODO: provide a small publically available test checkpoint
+    # TODO: provide a small publicly available test checkpoint
     model_path = ("/home/scratch.omniml_data_1/zhiyu/ckpts/test_ckpts/"
                   "TinyLlama-1.1B-Chat-v1.0-fp8-0710")
 

tests/samplers/test_beam_search.py

@@ -82,7 +82,7 @@ def test_beam_search_with_concurrency_limit(
     beam_width: int,
 ) -> None:
     # example_prompts[1]&[3]&[7] fails due to unknown reason even without
-    # concurency limit. skip them for now.
+    # concurrency limit. skip them for now.
     example_prompts = (example_prompts[:8])
     concurrency_limit = 2
     assert len(example_prompts) > concurrency_limit

tests/v1/attention/test_chunked_local_attention.py

@@ -160,7 +160,7 @@ def test_local_attention_virtual_batches(test_data: LocalAttentionTestData):
         # Use torch.arange instead of torch.randint so we can assert on
         # block table tensor values. The block table will have shape
         # (num_batches, cdiv(max_seq_len, block_size)) and the values will be
-        # aranged from 0 to cdiv(max_seq_len, block_size)-1
+        # arranged from 0 to cdiv(max_seq_len, block_size)-1
         arange_block_indices=True,
     )
 

tests/v1/kv_connector/unit/test_shared_storage_connector.py

@@ -33,7 +33,7 @@ def _check_path_len(path):
 
 
 def _list_path(path):
-    """Return the list of foldername (hashes generatd) under the path"""
+    """Return the list of foldername (hashes generated) under the path"""
     return list(path.iterdir())
 
 
@@ -41,7 +41,7 @@ def run_test(tmp_path, processor, llm: LLM, question: str,
              image_urls: list[Image], expected_len: int, info: str):
     """
     One individual test to process the prompt and output base on 1 set of input
-    Then check if the length in the strorage path matches the expected length
+    Then check if the length in the storage path matches the expected length
     `info` introduces details or purpose of the individual test
     """
     print(f"***info: {info}***")
@@ -115,7 +115,7 @@ def test_shared_storage_connector_hashes(tmp_path):
     """
     Tests that SharedStorageConnector saves KV to the storage locations
     with proper hashes; that are unique for inputs with identical text but 
-    differnt images (same size), or same multiple images but different orders.
+    different images (same size), or same multiple images but different orders.
     """
     # Using tmp_path as the storage path to store KV
     print(f"KV storage path at: {str(tmp_path)}")
@@ -171,12 +171,12 @@ def test_shared_storage_connector_hashes(tmp_path):
                   img=[image_1],
                   expected_len=2,
                   info=("image_1 single input the 2nd time. "
-                        "It should not form aother new hash.")),
+                        "It should not form another new hash.")),
         InputCase(text=TEXT_PROMPTS[0],
                   img=[image_2],
                   expected_len=2,
                   info=("image_2 single input the 2nd time. "
-                        "It should not form aother new hash.")),
+                        "It should not form another new hash.")),
         InputCase(text=TEXT_PROMPTS[0],
                   img=[image_1, image_2],
                   expected_len=3,
@@ -189,12 +189,12 @@ def test_shared_storage_connector_hashes(tmp_path):
                   img=[image_1, image_2],
                   expected_len=4,
                   info=("[image_1, image_2] input the 2nd time. "
-                        "It should not form aother new hash.")),
+                        "It should not form another new hash.")),
         InputCase(text=TEXT_PROMPTS[0],
                   img=[image_2, image_1],
                   expected_len=4,
                   info=("[image_2, image_1] input the 2nd time. "
-                        "It should not form aother new hash.")),
+                        "It should not form another new hash.")),
         InputCase(text=TEXT_PROMPTS[0],
                   img=[],
                   expected_len=5,

tests/v1/logits_processors/test_custom_offline.py

@@ -81,7 +81,7 @@ def _run_test(kwargs: dict, logitproc_loaded: bool) -> None:
             target_token = params.extra_args[DUMMY_LOGITPROC_ARG]
             if not all(x == target_token for x in lp_toks):
                 raise AssertionError(
-                    f"Request {bdx} generated {lp_toks}, shoud all be "
+                    f"Request {bdx} generated {lp_toks}, should all be "
                     f"{target_token}")
         else:
             # This request does not exercise custom logitproc (or custom

vllm/benchmarks/serve.py

@@ -189,7 +189,7 @@ async def get_request(
         # NOTE: If we simply accumulate the random delta values
         # from the gamma distribution, their sum would have 1-2% gap
         # from target_total_delay_s. The purpose of the following logic is to
-        # close the gap for stablizing the throughput data
+        # close the gap for stabilizing the throughput data
         # from different random seeds.
         target_total_delay_s = total_requests / request_rate
         normalize_factor = target_total_delay_s / delay_ts[-1]

vllm/config/compilation.py

@@ -234,7 +234,7 @@ class CompilationConfig:
     - FULL_AND_PIECEWISE.
 
     PIECEWISE mode build piecewise cudagraph only, keeping the cudagraph
-    incompatiable ops (i.e. some attention ops) outside the cudagraph
+    incompatible ops (i.e. some attention ops) outside the cudagraph
     for general flexibility.
     This is the default mode.
 

vllm/config/parallel.py

@@ -87,7 +87,7 @@ class ParallelConfig:
     data_parallel_external_lb: bool = False
     """Whether to use "external" DP LB mode. Applies only to online serving
     and when data_parallel_size > 0. This is useful for a "one-pod-per-rank"
-    wide-EP setup in Kuberentes. Set implicitly when --data-parallel-rank
+    wide-EP setup in Kubernetes. Set implicitly when --data-parallel-rank
     is provided explicitly to vllm serve."""
     data_parallel_hybrid_lb: bool = False
     """Whether to use "hybrid" DP LB mode. Applies only to online serving

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -787,7 +787,7 @@ class NixlConnectorWorker:
         self.src_xfer_side_handle = self.nixl_wrapper.prep_xfer_dlist(
             "NIXL_INIT_AGENT", descs)
 
-        # TODO(mgoin): Hybrid memory allocator is currently diabled for
+        # TODO(mgoin): Hybrid memory allocator is currently disabled for
         # models with local attention (Llama 4). Can remove this once enabled.
         if self.vllm_config.model_config.hf_config.model_type == "llama4":
             from transformers import Llama4TextConfig

vllm/entrypoints/openai/serving_responses.py

@@ -717,7 +717,7 @@ class OpenAIServingResponses(OpenAIServing):
                             prev_msgs.append(msg)
             messages.extend(prev_msgs)
         # Append the new input.
-        # Reponses API supports simple text inputs without chat format.
+        # Responses API supports simple text inputs without chat format.
         if isinstance(request.input, str):
             messages.append(get_user_message(request.input))
         else:

vllm/model_executor/layers/activation.py

@@ -362,7 +362,7 @@ class ReLUSquaredActivation(CustomOp):
         return torch.square(F.relu(x))
 
     def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
-        #TODO : implement cuda kenrels
+        #TODO : implement cuda kernels
         return self.forward_native(x)
 
 

vllm/model_executor/layers/quantization/compressed_tensors/transform/module.py

@@ -83,7 +83,7 @@ class HadamardTransform(torch.nn.Module):
             # do not fold into weight in order to utilize FWHT
             self.scales[part_id] = 1 / math.sqrt(data.size(0))
 
-            # FUTURE: avoid runtime tranpose by processing weights
+            # FUTURE: avoid runtime transpose by processing weights
             # prior to apply
 
     def forward(self, value: Tensor, part_id: int = 0) -> Tensor:

vllm/model_executor/layers/quantization/mxfp4.py

@@ -310,7 +310,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
             w13_bias = layer.w13_bias.data.to(torch.float32)
             w2_bias = layer.w2_bias.data.to(torch.float32)
 
-            # Swap w1 and w3 as the defenition of
+            # Swap w1 and w3 as the definition of
             # swiglu is different in the trtllm-gen
             def swap_every_two_rows(x, axis=-1):
                 shape = x.shape

vllm/model_executor/models/gemma3n_mm.py

@@ -179,7 +179,7 @@ class Gemma3nMultiModalProcessor(BaseMultiModalProcessor[Gemma3nProcessingInfo]
     ) -> BatchFeature:
 
         # HF Transformers audio processor no longer accepts `audios` key.
-        # We pop `audios` and replace it with `audio` key to surpress
+        # We pop `audios` and replace it with `audio` key to suppress
         # the warning.
         if 'audios' in mm_data:
             mm_data['audio'] = mm_data.pop('audios')

vllm/model_executor/models/interns1.py

@@ -492,7 +492,7 @@ class InternS1ForConditionalGeneration(nn.Module, SupportsMultiModal,
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
-        # transformers InternVLProcessor uses <IMG_CONTEXT> as the seperator
+        # transformers InternVLProcessor uses <IMG_CONTEXT> as the separator
         # refer to https://github.com/huggingface/transformers/blob/f90de364c2484c7c325bbe05befdcf487bd75b63/src/transformers/models/internvl/processing_internvl.py#L116
         if modality.startswith("image"):
             return '<IMG_CONTEXT>'

vllm/third_party/pynvml.py

@@ -3533,7 +3533,7 @@ def nvmlDeviceGetMPSComputeRunningProcesses_v3(handle):
         return []
     elif (ret == NVML_ERROR_INSUFFICIENT_SIZE):
         # typical case
-        # oversize the array incase more processes are created
+        # oversize the array in case more processes are created
         c_count.value = c_count.value * 2 + 5
         proc_array = c_nvmlProcessInfo_v3_t * c_count.value
         c_procs = proc_array()

vllm/v1/attention/backends/flash_attn.py

@@ -167,7 +167,7 @@ class FlashAttentionMetadataBuilder(
     # work for mixed prefill-decode and uniform-decode. But for non-spec decodes
     # the graphs would not work for mixed prefill-decode; sorta the inverse
     # of UNIFORM_SINGLE_TOKEN_DECODE.
-    # Theres probably a better way to describe this using `AttentionCGSupport`
+    # There's probably a better way to describe this using `AttentionCGSupport`
     # but for now just set it to `UNIFORM_BATCH` to get use to drop down
     # to FULL_AND_PIECEWISE.
     # TODO(luka, lucas): audit FA2 as part of:

vllm/v1/attention/backends/flashinfer.py

@@ -291,7 +291,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 paged_kv_indices_buffer=paged_kv_indices,
                 paged_kv_last_page_len_buffer=paged_kv_last_page_len,
                 # Tensor cores are enabled by default because the perf would be
-                # atleast as good as cuda cores for all attention ops in latest
+                # at least as good as cuda cores for all attention ops in latest
                 # gpus.
                 use_tensor_cores=True,
             )

vllm/v1/core/kv_cache_utils.py

@@ -217,7 +217,7 @@ class FreeKVCacheBlockQueue:
         # Create a fake head and a tail block for the doubly linked list to
         # reduce branching in the code
         #
-        # The implementation garenteed that the fake head and tail
+        # The implementation guaranteed that the fake head and tail
         # are NEVER got popped, so we could safely assume each real blocks
         # in the queue has prev and next blocks.
         self.fake_free_list_head = KVCacheBlock(block_id=-1)

vllm/v1/worker/gpu_input_batch.py

@@ -584,7 +584,7 @@ class InputBatch:
 
             if self.is_pooling_model:
                 last_req_index -= 1
-                # Samping state not used by pooling models.
+                # Sampling state not used by pooling models.
                 continue
 
             # Autoregressive models require detailed tracking of condense

vllm/v1/worker/gpu_model_runner.py

@@ -2776,7 +2776,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             self.attn_groups.append(
                 create_attn_groups(attn_backends, kv_cache_spec))
 
-        # Calculate reorder batch threshold (if neeeded)
+        # Calculate reorder batch threshold (if needed)
         self.calculate_reorder_batch_threshold()
 
     def initialize_cudagraph_capture(self) -> None:

vllm/v1/worker/kv_connector_model_runner_mixin.py

@@ -82,7 +82,7 @@ class KVConnectorModelRunnerMixin:
             scheduler_output) if has_kv_transfer_group() else nullcontext()
 
     # This context manager must be used within an active forward context.
-    # It encapsulates the entire KV conector lifecycle within execute_model
+    # It encapsulates the entire KV connector lifecycle within execute_model
     @staticmethod
     @contextmanager
     def _get_kv_connector_output(