[CI] change spell checker from codespell to typos (#18711)

Signed-off-by: Andy Xie <andy.xning@gmail.com>

vllm/model_executor/layers/quantization/utils/int8_utils.py

@@ -219,7 +219,7 @@ def per_token_group_quant_int8(
     quantized tensor along with the scaling factor used for quantization.
 
     Args:
-        x: The input tenosr with ndim >= 2.
+        x: The input tensor with ndim >= 2.
         group_size: The group size used for quantization.
         eps: The minimum to avoid dividing zero.
         dtype: The dype of output tensor. Note that only `torch.int8`

vllm/model_executor/model_loader/bitsandbytes_loader.py

@@ -401,7 +401,7 @@ class BitsAndBytesModelLoader(BaseModelLoader):
                         self.target_modules.append(
                             name.replace(rep_name, sub_name))
                 # Add original module name even if the module has stacked map,
-                # in case model has a mixture of disk-merged and disk-splitted
+                # in case model has a mixture of disk-merged and disk-split
                 # weights with same last name.
                 self.target_modules.append(name)
 

vllm/model_executor/models/baichuan.py

@@ -131,7 +131,7 @@ class BaiChuanAttention(nn.Module):
         self.num_heads = (self.total_num_heads //
                           tensor_model_parallel_world_size)
         self.head_dim = hidden_size // self.total_num_heads
-        self.postion_embedding = position_embedding
+        self.position_embedding = position_embedding
         self.rope_theta = rope_theta
         self.max_position_embeddings = max_position_embeddings
 
@@ -151,7 +151,7 @@ class BaiChuanAttention(nn.Module):
             quant_config=quant_config,
         )
         # Create the alibi slopes and slice them.
-        if self.postion_embedding == "ALIBI":
+        if self.position_embedding == "ALIBI":
             tp_rank = get_tensor_model_parallel_rank()
             head_start = tp_rank * self.num_heads
             head_end = (tp_rank + 1) * self.num_heads
@@ -187,7 +187,7 @@ class BaiChuanAttention(nn.Module):
     ) -> torch.Tensor:
         qkv, _ = self.W_pack(hidden_states)
         q, k, v = qkv.chunk(chunks=3, dim=-1)
-        if self.postion_embedding != "ALIBI":
+        if self.position_embedding != "ALIBI":
             q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)
         output, _ = self.o_proj(attn_output)

vllm/model_executor/models/deepseek_vl2.py

@@ -344,7 +344,7 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
             self.image_newline = nn.Parameter(
                 torch.randn(self.projector_config.n_embed) * embed_std)
             # This is a typo in original implementation
-            self.view_seperator = nn.Parameter(
+            self.view_separator = nn.Parameter(
                 torch.randn(self.projector_config.n_embed) * embed_std)
         else:
             raise ValueError(
@@ -549,13 +549,13 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
             if self.global_view_pos == "head":
                 global_local_features = torch.cat([
                     global_features,
-                    self.view_seperator[None, :],
+                    self.view_separator[None, :],
                     local_features,
                 ])
             else:
                 global_local_features = torch.cat([
                     local_features,
-                    self.view_seperator[None, :],
+                    self.view_separator[None, :],
                     global_features,
                 ])
 

vllm/model_executor/models/eagle.py

@@ -197,7 +197,7 @@ class EAGLE(nn.Module):
         return logits
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
-        # This implementation is incompitable with https://huggingface.co/yuhuili/EAGLE-LLaMA3-Instruct-8B
+        # This implementation is incompatible with https://huggingface.co/yuhuili/EAGLE-LLaMA3-Instruct-8B
         # due to missing lm_head weights and its config being that of a
         # Llama model. Here's a compatible version with the same weights:
         # https://huggingface.co/abhigoyal/EAGLE-LLaMA3-Instruct-8B-vllm

vllm/model_executor/models/gemma3_mm.py

@@ -634,13 +634,13 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
         kwargs["has_images"] = True
         # NOTE(woosuk): Here, we distinguish the sequences by the position id 0.
         # This is a HACK. Fix this.
-        start_idices = (positions == 0).cpu().nonzero()
-        num_seqs = len(start_idices)
+        start_indices = (positions == 0).cpu().nonzero()
+        num_seqs = len(start_indices)
         seq_lens = []
         for i in range(num_seqs):
-            start_idx = start_idices[i].item()
+            start_idx = start_indices[i].item()
             if i < num_seqs - 1:
-                end_idx = start_idices[i + 1].item()
+                end_idx = start_indices[i + 1].item()
             else:
                 end_idx = len(input_ids)
             seq_lens.append(end_idx - start_idx)

vllm/model_executor/models/llama4.py

@@ -52,7 +52,7 @@ class Llama4MoE(nn.Module):
         renormalize: bool,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         router_scores, router_indices = fast_topk(gating_output, topk, dim=-1)
-        # psuedo-standard is that the router scores are floats
+        # pseudo-standard is that the router scores are floats
         router_scores = torch.sigmoid(router_scores.float())
         return (router_scores, router_indices.to(torch.int32))
 

vllm/model_executor/models/mixtral_quant.py

@@ -114,9 +114,9 @@ class MixtralMoE(nn.Module):
                 f"Tensor parallel size {self.tp_size} is greater than "
                 f"the number of experts {self.num_total_experts}.")
         # Split experts equally between ranks
-        self.expert_indicies = np.array_split(range(
-            self.num_total_experts), self.tp_size)[self.rank].tolist()
-        if not self.expert_indicies:
+        self.expert_indices = np.array_split(range(self.num_total_experts),
+                                             self.tp_size)[self.rank].tolist()
+        if not self.expert_indices:
             raise ValueError(
                 f"Rank {self.rank} has no experts assigned to it.")
 
@@ -125,7 +125,7 @@ class MixtralMoE(nn.Module):
                        config.hidden_size,
                        config.intermediate_size,
                        quant_config=quant_config)
-            if idx in self.expert_indicies else None
+            if idx in self.expert_indices else None
             for idx in range(self.num_total_experts)
         ])
         self.gate = ReplicatedLinear(config.hidden_size,
@@ -146,7 +146,7 @@ class MixtralMoE(nn.Module):
         routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
 
         final_hidden_states = None
-        for expert_idx in self.expert_indicies:
+        for expert_idx in self.expert_indices:
             expert_layer = self.experts[expert_idx]
             expert_mask = (selected_experts == expert_idx)
             expert_weights = (routing_weights * expert_mask).sum(dim=-1,

vllm/model_executor/models/ovis.py

@@ -283,7 +283,7 @@ class OvisProcessingInfo(BaseProcessingInfo):
     def get_image_size_with_most_features(self) -> ImageSize:
         height, width = self.get_hf_processor().get_image_size()
         hs = self.get_hf_config().visual_tokenizer_config.hidden_stride
-        # NOTE(Isotr0py): 9 is `max_partion` hardcoded in original code
+        # NOTE(Isotr0py): 9 is `max_partition` hardcoded in original code
         # https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/modeling_ovis.py#L96
         return ImageSize(width=width * hs * 9, height=height * hs * 9)
 

vllm/model_executor/models/phi3_small.py

@@ -145,7 +145,7 @@ class Phi3SmallSelfAttention(nn.Module):
         self.num_q_per_kv = self.num_heads // self.num_key_value_heads
         if self.tp_size > 1:
             assert self.num_key_value_heads % self.tp_size == 0
-        self.num_kv_heads_per_partion = max(
+        self.num_kv_heads_per_partition = max(
             1, self.num_key_value_heads // self.tp_size)
         self.num_heads_per_partition = self.num_heads // self.tp_size
 
@@ -212,7 +212,7 @@ class Phi3SmallSelfAttention(nn.Module):
             bs_params = {
                 'max_seqlen': self.max_position_embeddings,
                 'num_heads': self.num_heads_per_partition,
-                "num_kv_heads": self.num_kv_heads_per_partion,
+                "num_kv_heads": self.num_kv_heads_per_partition,
                 "block_size": self.sparse_block_size,
                 "local_blocks": self.local_blocks,
                 "vert_stride": self.vert_stride,
@@ -222,7 +222,7 @@ class Phi3SmallSelfAttention(nn.Module):
         self.attn = Attention(self.num_heads_per_partition,
                               self.head_dim,
                               self.scale,
-                              num_kv_heads=self.num_kv_heads_per_partion,
+                              num_kv_heads=self.num_kv_heads_per_partition,
                               cache_config=cache_config,
                               quant_config=quant_config,
                               blocksparse_params=bs_params,
@@ -243,8 +243,8 @@ class Phi3SmallSelfAttention(nn.Module):
         # NOTE: this is required by RotaryEmbed, which indeed does not have to
         # TODO: allow 3D QK for rotary forward
         q = q.reshape(-1, self.head_dim * self.num_heads_per_partition)
-        k = k.reshape(-1, self.head_dim * self.num_kv_heads_per_partion)
-        v = v.reshape(-1, self.head_dim * self.num_kv_heads_per_partion)
+        k = k.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)
+        v = v.reshape(-1, self.head_dim * self.num_kv_heads_per_partition)
 
         q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)

vllm/model_executor/models/phi4mm_audio.py

@@ -126,7 +126,7 @@ class ConformerEncoderLayer(nn.Module):
             (Multi-Head Attention),
             1 = typical Multi-Head Attention,
             1 < attn_group_sizes < attention_heads = Grouped-Query Attention
-            attn_group_sizes = attenion_heads = Multi-Query Attention
+            attn_group_sizes = attention_heads = Multi-Query Attention
     """
 
     def __init__(
@@ -318,7 +318,7 @@ class TransformerEncoderBase(abc.ABC, nn.Module):
             1 = typical Multi-Head Attention,
             1 < attention_group_size < attention_heads = Grouped-Query 
             Attention
-            attention_group_size = attenion_heads = Multi-Query Attention
+            attention_group_size = attention_heads = Multi-Query Attention
     """
 
     def __init__(
@@ -744,7 +744,7 @@ class ConformerEncoder(TransformerEncoderBase):
             1 = typical Multi-Head Attention,
             1 < attention_group_size < attention_heads = Grouped-Query
             Attention
-            attention_group_size = attenion_heads = Multi-Query Attention
+            attention_group_size = attention_heads = Multi-Query Attention
     """
 
     extra_multi_layer_output_idxs: list[int]

vllm/model_executor/models/phimoe.py

@@ -147,15 +147,15 @@ class mp(torch.autograd.Function):
 
         grad_at_output = grad_at_output * multiplier
 
-        grad_at_scores_expaned = masked_gates * grad_at_output.mul(-1)
-        grad_at_scores_expaned.scatter_add_(
+        grad_at_scores_expanded = masked_gates * grad_at_output.mul(-1)
+        grad_at_scores_expanded.scatter_add_(
             dim=-1,
             index=selected_experts,
             src=grad_at_output,
         )
 
         return (
-            grad_at_scores_expaned,
+            grad_at_scores_expanded,
             None,
             None,
             None,

vllm/multimodal/utils.py

@@ -324,7 +324,7 @@ def merge_and_sort_multimodal_metadata(
     Returns:
         list[str]: List of item modalities in order of their positions in the
         input sequence.
-        list[PlaceholderRange]: Sorted list of all PlaceholdeRanges from
+        list[PlaceholderRange]: Sorted list of all PlaceholderRanges from
         mm_positions.
         Optional[list[str]]: Sorted list of all hashes from mm_hashes if given,
         None otherwise.

vllm/transformers_utils/processors/ovis.py

@@ -68,7 +68,7 @@ class OvisProcessor(ProcessorMixin):
     """
 
     attributes = ["image_processor", "tokenizer"]
-    valid_kwargs = ["chat_template", "image_pad_token", "image_segement_len"]
+    valid_kwargs = ["chat_template", "image_pad_token", "image_segment_len"]
 
     image_processor_class = "AutoImageProcessor"
     tokenizer_class = "AutoTokenizer"

vllm/worker/hpu_model_runner.py

@@ -886,7 +886,7 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
             num_decode_tokens=0,
             slot_mapping=slot_mapping,
             multi_modal_placeholder_index_maps=
-            None,  # FIXME(kzawora): mutli-modality will not work here
+            None,  # FIXME(kzawora): multi-modality will not work here
             enable_kv_scales_calculation=False,
         )
         multi_modal_kwargs = MultiModalKwargs.batch(multi_modal_kwargs_list)

vllm/worker/multi_step_model_runner.py

@@ -277,7 +277,7 @@ class StatefulModelInput(BroadcastableModelInput):
         assert fmi.input_tokens.shape[0] >= self.num_seqs
         fmi_new_input_tokens: torch.Tensor = fmi.input_tokens[:self.num_seqs]
 
-        # Update frozen_model_input::input_positons.
+        # Update frozen_model_input::input_positions.
         assert fmi.input_positions is not None
         assert fmi.input_positions.shape[0] >= self.num_seqs
         fmi_new_input_positions: torch.Tensor = fmi.input_positions[:self.

vllm/worker/tpu_model_runner.py

@@ -798,9 +798,9 @@ class ModelWrapper(nn.Module):
         """
         batch_size, seq_len = token_ids.shape
         # Calculate the positions to sample from.
-        start_indicies = torch.arange(
+        start_indices = torch.arange(
             batch_size, dtype=torch.int32, device=input_lens.device) * seq_len
-        logits_indices = start_indicies + input_lens - 1
+        logits_indices = start_indices + input_lens - 1
         attn_metadata = get_forward_context().attn_metadata
 
         # FIXME(woosuk): This is a temporary hack to avoid using the existing
@@ -822,14 +822,14 @@ class ModelWrapper(nn.Module):
             num_kv_heads, num_blocks, block_size, _ = kv_caches[0][0].shape
             slot_mapping = attn_metadata.slot_mapping
             slot_mapping = slot_mapping.flatten()
-            head_indicies = torch.arange(0,
+            head_indices = torch.arange(0,
                                         num_kv_heads,
                                         device=slot_mapping.device,
                                         dtype=slot_mapping.dtype)
-            head_indicies *= block_size * num_blocks
+            head_indices *= block_size * num_blocks
             slot_mapping = slot_mapping.repeat_interleave(num_kv_heads).view(
                 -1, num_kv_heads)
-            slot_mapping = slot_mapping + head_indicies.view(1, -1)
+            slot_mapping = slot_mapping + head_indices.view(1, -1)
             slot_mapping = slot_mapping.flatten()
             attn_metadata.slot_mapping = slot_mapping