[Docs] Fix warnings in mkdocs build (continued) (#23743)

Signed-off-by: Zerohertz <ohg3417@gmail.com>
Signed-off-by: Hyogeun Oh (오효근) <ohg3417@gmail.com>

vllm/core/block/naive_block.py

@@ -207,7 +207,7 @@ class NaiveBlockAllocator(BlockAllocator):
 
         Args:
             absolute_id (int): The absolute block id for the block 
-            in whole allocator.
+                in whole allocator.
 
         Returns:
             int: The zero-offset block id on certain device.

vllm/core/block/prefix_caching_block.py

@@ -61,7 +61,7 @@ class PrefixCachingBlockAllocator(BlockAllocator):
     Args:
         num_blocks (int): The total number of blocks to manage.
         block_size (int): The size of each block in tokens.
-        block_ids(Optional[Iterable[int]], optional): An optional iterable of
+        block_ids (Optional[Iterable[int]], optional): An optional iterable of
             block IDs. If not provided, block IDs will be assigned sequentially
             from 0 to num_blocks - 1.
     """

vllm/core/scheduler.py

@@ -657,7 +657,7 @@ class Scheduler:
                 `budget.num_batched_tokens` has not enough capacity to schedule
                 all tokens.
             partial_prefill_metadata: information about the partial prefills
-            that are currently running
+                that are currently running
 
         Returns:
             SchedulerRunningOutputs.

vllm/v1/attention/backends/cpu_attn.py

@@ -491,7 +491,8 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
             query: shape = [num_tokens, num_heads * head_size]
             key: shape = [num_tokens, num_kv_heads * head_size]
             value: shape = [num_tokens, num_kv_heads * head_size]
-            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size * num_kv_heads * head_size]
                 NOTE: kv_cache will be an empty tensor with shape [0]
                 for profiling run.
             attn_metadata: Metadata for attention.

vllm/v1/attention/backends/flash_attn.py

@@ -438,7 +438,8 @@ class FlashAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/flashinfer.py

@@ -637,11 +637,9 @@ class FlashInferImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache: shape -
-            # NHD: [num_blocks, 2, block_size, num_kv_heads, head_size]
-            # HND: [num_blocks, 2,  num_kv_heads, block_size, head_size]
-
-
+            kv_cache: KV cache tensor with different possible shapes:
+                - NHD: [num_blocks, 2, block_size, num_kv_heads, head_size]
+                - HND: [num_blocks, 2, num_kv_heads, block_size, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/flex_attention.py

@@ -689,7 +689,8 @@ class FlexAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/pallas.py

@@ -235,7 +235,8 @@ class PallasAttentionBackendImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads * head_size]
             key: shape = [num_tokens, num_kv_heads * head_size]
             value: shape = [num_tokens, num_kv_heads * head_size]
-            kv_cache = [num_blocks, block_size, num_kv_heads * 2, head_size]
+            kv_cache: shape =
+                [num_blocks, block_size, num_kv_heads * 2, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]
@@ -329,7 +330,7 @@ def write_to_kv_cache(
     Args:
         key: shape = [num_tokens, num_kv_heads, head_size]
         value: shape = [num_tokens, num_kv_heads, head_size]
-        kv_cache = [num_blocks, block_size, num_kv_heads * 2, head_size]
+        kv_cache: shape = [num_blocks, block_size, num_kv_heads * 2, head_size]
         num_slices_per_kv_cache_update_block: int
     """
     _, page_size, num_combined_kv_heads, head_size = kv_cache.shape

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -429,7 +429,8 @@ class AiterFlashAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/tree_attn.py

@@ -362,7 +362,8 @@ class TreeAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/triton_attn.py

@@ -285,7 +285,8 @@ class TritonAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/attention/backends/xformers.py

@@ -330,7 +330,8 @@ class XFormersAttentionImpl(AttentionImpl):
             query: shape = [num_tokens, num_heads, head_size]
             key: shape = [num_tokens, num_kv_heads, head_size]
             value: shape = [num_tokens, num_kv_heads, head_size]
-            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+            kv_cache: shape =
+                [2, num_blocks, block_size, num_kv_heads, head_size]
             attn_metadata: Metadata for attention.
         Returns:
             shape = [num_tokens, num_heads * head_size]

vllm/v1/core/encoder_cache_manager.py

@@ -255,9 +255,9 @@ def compute_encoder_budget(
 
     Returns:
         - Compute budget for encoder execution, measured in number of tokens
-          from the input sequence.
+            from the input sequence.
         - Space budget for encoder cache size, measured in number of tokens
-          from the input sequence.
+            from the input sequence.
     """
     if mm_registry.supports_multimodal_inputs(model_config):
         max_tokens_by_modality = mm_registry \
@@ -303,9 +303,9 @@ def compute_mm_encoder_budget(
 
     Returns:
         - Compute budget for encoder execution, measured in number of tokens
-          from the input sequence.
+            from the input sequence.
         - Space budget for encoder cache size, measured in number of tokens
-          from the input sequence.
+            from the input sequence.
     """
 
     if not max_tokens_by_modality:

vllm/v1/core/kv_cache_coordinator.py

@@ -119,7 +119,8 @@ class KVCacheCoordinator(ABC):
 
         Args:
             request: The request.
-            num_tokens: The total number of tokens that need to be cached 
+            num_computed_tokens: The total number of tokens
+                that need to be cached
                 (including tokens that are already cached).
         """
         for manager in self.single_type_managers:

vllm/v1/core/kv_cache_manager.py

@@ -54,14 +54,15 @@ class KVCacheBlocks:
     def get_block_ids(
         self,
         allow_none: bool = False,
-    ):
+    ) -> Optional[tuple[list[int], ...]]:
         """
         Converts the KVCacheBlocks instance to block_ids.
 
         Returns:
-            tuple[list[int], ...]: A tuple of lists where
-            * the outer tuple corresponds to KV cache groups
-            * each inner list contains the block_ids of the blocks in that group
+            tuple[list[int], ...]: A tuple of lists where:
+                - the outer tuple corresponds to KV cache groups
+                - each inner list contains the block_ids of the blocks in that
+                  group
         """
         if allow_none and all(len(group) == 0 for group in self.blocks):
             return None

vllm/v1/executor/ray_distributed_executor.py

@@ -8,6 +8,7 @@ from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator
 from vllm.executor.ray_distributed_executor import (  # noqa
     RayDistributedExecutor as RayDistributedExecutorV0)
 from vllm.logger import init_logger
+from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.engine import ReconfigureDistributedRequest, ReconfigureRankType
 from vllm.v1.executor.abstract import Executor
 from vllm.v1.outputs import ModelRunnerOutput
@@ -64,7 +65,7 @@ class RayDistributedExecutor(RayDistributedExecutorV0, Executor):
 
     def execute_model(
         self,
-        scheduler_output,
+        scheduler_output: SchedulerOutput,
     ) -> Union[ModelRunnerOutput, Future[ModelRunnerOutput]]:
         """Execute the model on the Ray workers.
 

vllm/v1/metrics/prometheus.py

@@ -36,7 +36,7 @@ def setup_multiprocess_prometheus():
                        "and vLLM will properly handle cleanup.")
 
 
-def get_prometheus_registry():
+def get_prometheus_registry() -> CollectorRegistry:
     """Get the appropriate prometheus registry based on multiprocessing 
     configuration.
     

vllm/v1/sample/logits_processor/interface.py

@@ -91,7 +91,7 @@ class LogitsProcessor(ABC):
         to each forward pass.
 
         Args:
-            batch_update is non-None iff there have been
-            changes to the batch makeup.
+            batch_update: Non-None iff there have been changes
+                to the batch makeup.
         """
         raise NotImplementedError

vllm/v1/sample/rejection_sampler.py

@@ -68,7 +68,7 @@ class RejectionSampler(nn.Module):
                 different requests are flattened into a single tensor because
                 this is the shape of the output logits.
                 NOTE: `target_logits` can be updated in place to save memory.
-            bonus_token_ids_tensor (torch.Tensor):
+            bonus_token_ids (torch.Tensor):
                 A tensor containing bonus tokens. Shape is [batch_size, 1].
                 Bonus tokens are added to the end of the sequence if all
                 proposed tokens are accepted. We generate the bonus tokens

vllm/v1/sample/tpu/sampler.py

@@ -89,7 +89,7 @@ class Sampler(nn.Module):
         Gather logprobs for topk and sampled/prompt token.
 
         Args:
-          logits: (num tokens) x (vocab) tensor
+          logprobs: (num tokens) x (vocab) tensor
           num_logprobs: minimum number of logprobs to
                         retain per token
           token_ids: prompt tokens (if prompt logprobs)

vllm/v1/structured_output/backend_types.py

@@ -110,7 +110,7 @@ class StructuredOutputBackend(ABC):
 
         Args:
             request_type (StructuredOutputOptions): The type of structured
-              output request.
+                output request.
             grammar_spec (str): The grammar specification to compile.
 
         Returns:
@@ -124,7 +124,7 @@ class StructuredOutputBackend(ABC):
 
         Args:
             max_num_seqs (int): The maximum number of sequences for which
-              to allocate the bitmask.
+                to allocate the bitmask.
         """
 
     @abstractmethod

vllm/v1/worker/gpu_input_batch.py

@@ -525,9 +525,6 @@ class InputBatch:
         Any consecutive empty indices at the very end of the list are not
         filled.
 
-        Args:
-          empty_req_indices: empty indices which may be filled.
-
         Returns:
           swaps: list of (from,to) swap tuples for moved requests
           empty_req_indices: indices not filled by condensation

vllm/v1/worker/gpu_model_runner.py

@@ -2955,7 +2955,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         Args:
             kv_cache_config: The KV cache config
             kv_cache_raw_tensors: The KV cache buffer of each layer, with
-            correct size but uninitialized shape.
+                correct size but uninitialized shape.
         Returns:
             Dict[str, torch.Tensor]: A map between layer names to their
             corresponding memory buffer for KV cache.

vllm/v1/worker/tpu_model_runner.py

@@ -552,7 +552,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         return kv_cache_spec
 
     def _get_slot_mapping_metadata(self, num_reqs,
-                                   num_scheduled_tokens_per_req):
+                                   num_scheduled_tokens_per_req) -> np.ndarray:
         """
         Computes metadata for mapping slots to blocks in the key-value (KV)
         cache for a batch of requests.
@@ -565,15 +565,15 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
         Args:
             num_reqs (int): Number of requests in the current batch.
             num_scheduled_tokens_per_req (int or np.ndarray): Number of tokens
-            to be scheduled for each request.
+                to be scheduled for each request.
 
         Returns:
             np.ndarray: A 2D array of shape (total_block_len, 3), where each row
-            contains:
+                contains:
                 - kv_cache_start_index (int): The starting index in the KV cache
-                    for the corresponding slice.
+                  for the corresponding slice.
                 - new_kv_start_index (int): The starting index in the new KV
-                    cache for the corresponding slice.
+                  cache for the corresponding slice.
                 - slice_len (int): The length of the slice.
         """
         slices_start = self.input_batch.num_computed_tokens_cpu[:num_reqs]

vllm/v1/worker/utils.py

@@ -172,10 +172,10 @@ def scatter_mm_placeholders(
 
     Args:
         embeds: The multimodal embeddings.
-          Shape: `(num_embeds, embed_dim)`
+            Shape: `(num_embeds, embed_dim)`
         is_embed: A boolean mask indicating which positions in the placeholder
-          tokens need to be filled with multimodal embeddings.
-          Shape: `(num_placeholders, num_embeds)`
+            tokens need to be filled with multimodal embeddings.
+            Shape: `(num_placeholders, num_embeds)`
     """
     if is_embed is None:
         return embeds
@@ -278,7 +278,7 @@ def bind_kv_cache(
     Args:
         kv_caches: The allocated kv_caches with layer names as keys.
         forward_context: The global forward context containing all Attention
-        layers with layer names as keys.
+            layers with layer names as keys.
         runner_kv_caches: The kv_cache declared by ModelRunner.
     """
     # Bind kv_caches to ModelRunner

vllm/v1/worker/worker_base.py

@@ -37,7 +37,7 @@ class WorkerBase(WorkerBaseV0):
             rank: Global rank in distributed setup
             distributed_init_method: Distributed initialization method
             is_driver_worker: Whether this worker handles driver
-            responsibilities
+                responsibilities
         """
         # Configuration storage
         super().__init__(vllm_config=vllm_config)