fix some typos (#24071)

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

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(