[Doc] Convert Sphinx directives ( `{class}`, `{meth}`, `{attr}`, ...) to...

[Doc]  Convert Sphinx directives ( `{class}`, `{meth}`, `{attr}`, ...) to MkDocs format for better documentation linking (#18663)
Signed-off-by: Zerohertz <ohg3417@gmail.com>

vllm/compilation/compiler_interface.py

@@ -39,7 +39,8 @@ class CompilerInterface:
         Gather all the relevant information from the vLLM config,
         to compute a hash so that we can cache the compiled model.
 
-        See {meth}`VllmConfig.compute_hash` to check what information
+        See [`VllmConfig.compute_hash`][vllm.config.VllmConfig.compute_hash]
+        to check what information
         is already considered by default. This function should only
         consider the information that is specific to the compiler.
         """

vllm/config.py

@@ -2986,7 +2986,7 @@ class PoolerConfig:
     pooling_type: Optional[str] = None
     """
     The pooling method of the pooling model. This should be a key in
-    {class}`vllm.model_executor.layers.pooler.PoolingType`.
+    [`vllm.model_executor.layers.pooler.PoolingType`][].
     """
 
     normalize: Optional[bool] = None
@@ -3697,23 +3697,27 @@ class CompilationConfig:
     """Configuration for compilation. It has three parts:
 
     - Top-level Compilation control:
-        - {attr}`level`
-        - {attr}`debug_dump_path`
-        - {attr}`cache_dir`
-        - {attr}`backend`
-        - {attr}`custom_ops`
-        - {attr}`splitting_ops`
+        - [`level`][vllm.config.CompilationConfig.level]
+        - [`debug_dump_path`][vllm.config.CompilationConfig.debug_dump_path]
+        - [`cache_dir`][vllm.config.CompilationConfig.cache_dir]
+        - [`backend`][vllm.config.CompilationConfig.backend]
+        - [`custom_ops`][vllm.config.CompilationConfig.custom_ops]
+        - [`splitting_ops`][vllm.config.CompilationConfig.splitting_ops]
     - CudaGraph capture:
-        - {attr}`use_cudagraph`
-        - {attr}`cudagraph_capture_sizes`
-        - {attr}`cudagraph_num_of_warmups`
-        - {attr}`cudagraph_copy_inputs`
-        - {attr}`full_cuda_graph`
+        - [`use_cudagraph`][vllm.config.CompilationConfig.use_cudagraph]
+        - [`cudagraph_capture_sizes`]
+        [vllm.config.CompilationConfig.cudagraph_capture_sizes]
+        - [`cudagraph_num_of_warmups`]
+        [vllm.config.CompilationConfig.cudagraph_num_of_warmups]
+        - [`cudagraph_copy_inputs`]
+        [vllm.config.CompilationConfig.cudagraph_copy_inputs]
+        - [`full_cuda_graph`][vllm.config.CompilationConfig.full_cuda_graph]
     - Inductor compilation:
-        - {attr}`use_inductor`
-        - {attr}`compile_sizes`
-        - {attr}`inductor_compile_config`
-        - {attr}`inductor_passes`
+        - [`use_inductor`][vllm.config.CompilationConfig.use_inductor]
+        - [`compile_sizes`][vllm.config.CompilationConfig.compile_sizes]
+        - [`inductor_compile_config`]
+        [vllm.config.CompilationConfig.inductor_compile_config]
+        - [`inductor_passes`][vllm.config.CompilationConfig.inductor_passes]
         - custom inductor passes
 
     Why we have different sizes for cudagraph and inductor:

vllm/connections.py

@@ -167,4 +167,7 @@ class HTTPConnection:
 
 
 global_http_connection = HTTPConnection()
-"""The global {class}`HTTPConnection` instance used by vLLM."""
+"""
+The global [`HTTPConnection`][vllm.connections.HTTPConnection] instance used
+by vLLM.
+"""

vllm/engine/async_llm_engine.py

@@ -475,7 +475,8 @@ class _AsyncLLMEngine(LLMEngine):
             *,
             inputs: Optional[PromptType] = None,  # DEPRECATED
     ) -> None:
-        """Async version of {meth}`add_request`."""
+        """Async version of
+        [`add_request`][vllm.engine.llm_engine.LLMEngine.add_request]."""
         if inputs is not None:
             prompt = inputs
         assert prompt is not None and params is not None
@@ -582,20 +583,21 @@ async def build_guided_decoding_logits_processor_async(
 
 
 class AsyncLLMEngine(EngineClient):
-    """An asynchronous wrapper for {class}`LLMEngine`.
+    """An asynchronous wrapper for [`LLMEngine`][vllm.LLMEngine].
 
-    This class is used to wrap the {class}`LLMEngine` class to make it
-    asynchronous. It uses asyncio to create a background loop that keeps
-    processing incoming requests. The {class}`LLMEngine` is kicked by the
-    generate method when there are requests in the waiting queue. The generate
-    method yields the outputs from the {class}`LLMEngine` to the caller.
+    This class is used to wrap the [`LLMEngine`][vllm.LLMEngine] class to
+    make it asynchronous. It uses asyncio to create a background loop that keeps
+    processing incoming requests. The [`LLMEngine`][vllm.LLMEngine] is kicked
+    by the generate method when there are requests in the waiting queue. The
+    generate method yields the outputs from the [`LLMEngine`][vllm.LLMEngine]
+    to the caller.
 
     Args:
         log_requests: Whether to log the requests.
         start_engine_loop: If True, the background task to run the engine
             will be automatically started in the generate call.
-        *args: Arguments for {class}`LLMEngine`.
-        **kwargs: Arguments for {class}`LLMEngine`.
+        *args: Arguments for [`LLMEngine`][vllm.LLMEngine].
+        **kwargs: Arguments for [`LLMEngine`][vllm.LLMEngine].
     """
 
     _engine_class: Type[_AsyncLLMEngine] = _AsyncLLMEngine
@@ -985,8 +987,9 @@ class AsyncLLMEngine(EngineClient):
         from the LLMEngine to the caller.
 
         Args:
-            prompt: The prompt to the LLM. See {class}`~vllm.inputs.PromptType`
-                for more details about the format of each input.
+            prompt: The prompt to the LLM. See
+                [`PromptType`][vllm.inputs.PromptType] for more details about
+                the format of each input.
             sampling_params: The sampling parameters of the request.
             request_id: The unique id of the request.
             lora_request: LoRA request to use for generation, if any.
@@ -1003,7 +1006,7 @@ class AsyncLLMEngine(EngineClient):
         Details:
             - If the engine is not running, start the background loop,
               which iteratively invokes
-              {meth}`~vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step`
+              [`engine_step`][vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step]
               to process the waiting requests.
             - Add the request to the engine's `RequestTracker`.
               On the next background loop, this request will be sent to
@@ -1075,8 +1078,9 @@ class AsyncLLMEngine(EngineClient):
         from the LLMEngine to the caller.
 
         Args:
-            prompt: The prompt to the LLM. See {class}`~vllm.inputs.PromptType`
-                for more details about the format of each input.
+            prompt: The prompt to the LLM. See
+                [`PromptType`][vllm.inputs.PromptType] for more details about
+                the format of each input.
             pooling_params: The pooling parameters of the request.
             request_id: The unique id of the request.
             lora_request: LoRA request to use for generation, if any.
@@ -1091,7 +1095,7 @@ class AsyncLLMEngine(EngineClient):
         Details:
             - If the engine is not running, start the background loop,
                 which iteratively invokes
-            {meth}`~vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step`
+                [`vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step`][]
                 to process the waiting requests.
             - Add the request to the engine's `RequestTracker`.
                 On the next background loop, this request will be sent to

vllm/engine/llm_engine.py

@@ -130,11 +130,11 @@ class LLMEngine:
     iteration-level scheduling and efficient memory management to maximize the
     serving throughput.
 
-    The [LLM][vllm.LLM] class wraps this class for offline batched inference
-    and the [AsyncLLMEngine][] class wraps this class for online serving.
+    The [`LLM`][vllm.LLM] class wraps this class for offline batched inference
+    and the [`AsyncLLMEngine`][vllm.engine.async_llm_engine.AsyncLLMEngine]
+    class wraps this class for online serving.
 
-    The config arguments are derived from [EngineArgs][vllm.EngineArgs]. (See
-    [engine-args][])
+    The config arguments are derived from [`EngineArgs`][vllm.EngineArgs].
 
     Args:
         vllm_config: The configuration for initializing and running vLLM.

vllm/engine/multiprocessing/client.py

@@ -492,8 +492,9 @@ class MQLLMEngineClient(EngineClient):
         from the LLMEngine to the caller.
 
         Args:
-            prompt: The prompt to the LLM. See {class}`~vllm.inputs.PromptType`
-                for more details about the format of each input.
+            prompt: The prompt to the LLM. See
+                [`PromptType`][vllm.inputs.PromptType] for more details about
+                the format of each input.
             sampling_params: The sampling parameters of the request.
             request_id: The unique id of the request.
             lora_request: LoRA request to use for generation, if any.
@@ -561,8 +562,9 @@ class MQLLMEngineClient(EngineClient):
         from the LLMEngine to the caller.
 
         Args:
-            prompt: The prompt to the LLM. See {class}`~vllm.inputs.PromptType`
-                for more details about the format of each input.
+            prompt: The prompt to the LLM. See
+                [`PromptType`][vllm.inputs.PromptType] for more details about
+                the format of each input.
             pooling_params: The pooling parameters of the request.
             request_id: The unique id of the request.
             lora_request: LoRA request to use for generation, if any.

vllm/engine/multiprocessing/engine.py

@@ -42,19 +42,22 @@ HEALTHY_RESPONSE = (pickle.dumps(VLLM_RPC_SUCCESS_STR), )
 
 
 class MQLLMEngine:
-    """A multiprocessing wrapper for {class}`LLMEngine`.
+    """A multiprocessing wrapper for
+    [`LLMEngine`][vllm.engine.llm_engine.LLMEngine].
 
-    This class is used to wrap the {class}`LLMEngine` class to enable use
+    This class is used to wrap the
+    [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] class to enable use
     in concurrnet manner. It runs a background loop and uses zeromq to
     receive new requests and stream outputs incrementally via ipc.
 
-    The {class}`LLMEngine` generate or encode process is kicked off when a new
-    RPCProcessRequest is received by the input_socket.
+    The [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] generate or encode
+    process is kicked off when a new RPCProcessRequest is received by the
+    input_socket.
 
     The self.engine_loop checks the input_socket for new requests,
     adds them to the LLMEngine if there are any, calls the internal
-    {class}`LLMEngine.step()`, and sends the RequestOutputs back over
-    the output_socket.
+    [`LLMEngine.step()`][vllm.engine.llm_engine.LLMEngine.step], and sends
+    the RequestOutputs back over the output_socket.
 
     If use_async_sockets is set, the logic associated with reading new
     requests from the socket and sending data to the socket is passed
@@ -65,8 +68,8 @@ class MQLLMEngine:
         ipc_path: Base path for zeromq interprocess messaging
         use_async_sockets: Whether to make send/recv async with GPU
         log_requests: Whether to log the requests.
-        *args: Arguments for {class}`LLMEngine`.
-        **kwargs: Arguments for {class}`LLMEngine`.
+        *args: Arguments for [`LLMEngine`][vllm.engine.llm_engine.LLMEngine].
+        **kwargs: Arguments for [`LLMEngine`][vllm.engine.llm_engine.LLMEngine].
     """
 
     def __init__(self,

vllm/engine/output_processor/multi_step.py

@@ -56,8 +56,11 @@ class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
         scheduled computation.
 
         Args:
-          seq_group: the outputs are associated with this {class}`SequenceGroup`
-          outputs: the {class}`SequenceGroupOutput`s for all scheduler steps
+          seq_group: the outputs are associated with this
+              [`SequenceGroup`][vllm.sequence.SequenceGroup]
+          outputs: the
+              [`SequenceGroupOutput`][vllm.sequence.SequenceGroupOutput]s
+              for all scheduler steps
         """
         for output in outputs:
             # Concatenate single-step prompt logprob processing results.

vllm/engine/output_processor/single_step.py

@@ -19,17 +19,21 @@ logger = init_logger(__name__)
 def single_step_process_prompt_logprob(
         sg_output_proc: SequenceGroupOutputProcessor, seq_group: SequenceGroup,
         output: CompletionSequenceGroupOutput) -> None:
-    """Process prompt logprobs associated with the {class}`SequenceGroupOutput`
-    for a given step.
+    """Process prompt logprobs associated with the
+    [`SequenceGroupOutput`][vllm.sequence.SequenceGroupOutput] for a given step.
 
     Do nothing if the output has no prompt logprobs.
 
     Account for the fact that transformers do not compute first-token logprobs.
     
     Args:
-      sg_output_proc: {class}`SequenceGroupOutputProcessor` instance
-      seq_group: the output is associated with this {class}`SequenceGroup`
-      output: the {class}`SequenceGroupOutput` for a single scheduler step
+      sg_output_proc:
+          [`SequenceGroupOutputProcessor`][vllm.engine.output_processor.interfaces.SequenceGroupOutputProcessor]
+          instance
+      seq_group: the output is associated with this
+          [`SequenceGroup`][vllm.sequence.SequenceGroup]
+      output: the [`SequenceGroupOutput`][vllm.sequence.SequenceGroupOutput]
+          for a single scheduler step
     """
     prompt_logprobs = output.prompt_logprobs
 
@@ -103,8 +107,11 @@ class SingleStepOutputProcessor(SequenceGroupOutputProcessor):
         scheduled computation.
         
         Args:
-          seq_group: the output is associated with this {class}`SequenceGroup`
-          outputs: the {class}`SequenceGroupOutput` for a single scheduler step
+          seq_group: the output is associated with this
+              [`SequenceGroup`][vllm.sequence.SequenceGroup]
+          outputs: the
+              [`SequenceGroupOutput`][vllm.sequence.SequenceGroupOutput]
+              for a single scheduler step
         """
         assert len(outputs) == 1, "Single step should only have 1 output."
         output = outputs[0]

vllm/entrypoints/llm.py

@@ -129,8 +129,7 @@ class LLM:
         compilation_config: Either an integer or a dictionary. If it is an
             integer, it is used as the level of compilation optimization. If it
             is a dictionary, it can specify the full compilation configuration.
-        **kwargs: Arguments for [EngineArgs][vllm.EngineArgs]. (See
-            [engine-args][])
+        **kwargs: Arguments for [`EngineArgs`][vllm.EngineArgs].
 
     Note:
         This class is intended to be used for offline inference. For online
@@ -494,7 +493,7 @@ class LLM:
                 `self` argument, in addition to the arguments passed in `args`
                 and `kwargs`. The `self` argument will be the worker object.
             timeout: Maximum time in seconds to wait for execution. Raises a
-                {exc}`TimeoutError` on timeout. `None` means wait indefinitely.
+                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
             args: Positional arguments to pass to the worker method.
             kwargs: Keyword arguments to pass to the worker method.
 

vllm/entrypoints/openai/serving_engine.py

@@ -582,7 +582,8 @@ class OpenAIServing:
         add_special_tokens: bool = True,
     ) -> TextTokensPrompt:
         """
-        A simpler implementation of {meth}`_tokenize_prompt_input_or_inputs`
+        A simpler implementation of
+        [`_tokenize_prompt_input_or_inputs`][vllm.entrypoints.openai.serving_engine.OpenAIServing._tokenize_prompt_input_or_inputs]
         that assumes single input.
         """
         return next(
@@ -603,7 +604,8 @@ class OpenAIServing:
         add_special_tokens: bool = True,
     ) -> Iterator[TextTokensPrompt]:
         """
-        A simpler implementation of {meth}`_tokenize_prompt_input_or_inputs`
+        A simpler implementation of
+        [`_tokenize_prompt_input_or_inputs`][vllm.entrypoints.openai.serving_engine.OpenAIServing._tokenize_prompt_input_or_inputs]
         that assumes multiple inputs.
         """
         for text in prompt_inputs:

vllm/executor/executor_base.py

@@ -74,7 +74,7 @@ class ExecutorBase(ABC):
                 `self` argument, in addition to the arguments passed in `args`
                 and `kwargs`. The `self` argument will be the worker object.
             timeout: Maximum time in seconds to wait for execution. Raises a
-                {exc}`TimeoutError` on timeout. `None` means wait indefinitely.
+                [`TimeoutError`][] on timeout. `None` means wait indefinitely.
             args: Positional arguments to pass to the worker method.
             kwargs: Keyword arguments to pass to the worker method.
 

vllm/inputs/__init__.py

@@ -10,8 +10,9 @@ from .registry import (DummyData, InputContext, InputProcessingContext,
 
 INPUT_REGISTRY = InputRegistry()
 """
-The global {class}`~InputRegistry` which is used by {class}`~vllm.LLMEngine`
-to dispatch data processing according to the target model.
+The global [`InputRegistry`][vllm.inputs.registry.InputRegistry] which is used
+by [`LLMEngine`][vllm.LLMEngine] to dispatch data processing according to the
+target model.
 """
 
 __all__ = [

vllm/inputs/data.py

@@ -80,22 +80,24 @@ SingletonPrompt = Union[str, TextPrompt, TokensPrompt, EmbedsPrompt]
 """
 Set of possible schemas for a single prompt:
 
-- A text prompt ({class}`str` or {class}`TextPrompt`)
-- A tokenized prompt ({class}`TokensPrompt`)
-- An embeddings prompt ({class}`EmbedsPrompt`)
+- A text prompt ([`str`][] or [`TextPrompt`][vllm.inputs.data.TextPrompt])
+- A tokenized prompt ([`TokensPrompt`][vllm.inputs.data.TokensPrompt])
+- An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.data.EmbedsPrompt])
 
 Note that "singleton" is as opposed to a data structure
 which encapsulates multiple prompts, i.e. of the sort
 which may be utilized for encoder/decoder models when
 the user desires to express both the encoder & decoder
-prompts explicitly, i.e. {class}`ExplicitEncoderDecoderPrompt`
+prompts explicitly, i.e. 
+[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
 
-A prompt of type {class}`SingletonPrompt` may be employed
-as (1) input to a decoder-only model, (2) input to
+A prompt of type [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] may be 
+employed as (1) input to a decoder-only model, (2) input to
 the encoder of an encoder/decoder model, in the scenario
 where the decoder-prompt is not specified explicitly, or
 (3) as a member of a larger data structure encapsulating
-more than one prompt, i.e. {class}`ExplicitEncoderDecoderPrompt`
+more than one prompt, i.e. 
+[`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
 """
 
 
@@ -126,18 +128,20 @@ class ExplicitEncoderDecoderPrompt(TypedDict, Generic[_T1_co, _T2_co]):
     comprising an explicit encoder prompt and a decoder prompt.
 
     The encoder and decoder prompts, respectively, may be formatted
-    according to any of the {class}`SingletonPrompt` schemas,
+    according to any of the
+    [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] schemas,
     and are not required to have the same schema.
 
     Only the encoder prompt may have multi-modal data. mm_processor_kwargs
     should be at the top-level, and should not be set in the encoder/decoder
     prompts, since they are agnostic to the encoder/decoder.
 
-    Note that an {class}`ExplicitEncoderDecoderPrompt` may not
-    be used as an input to a decoder-only model,
+    Note that an
+    [`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
+    may not be used as an input to a decoder-only model,
     and that the `encoder_prompt` and `decoder_prompt`
     fields of this data structure themselves must be
-    {class}`SingletonPrompt` instances.
+    [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] instances.
     """
 
     encoder_prompt: _T1_co
@@ -152,11 +156,11 @@ PromptType = Union[SingletonPrompt, ExplicitEncoderDecoderPrompt]
 Set of possible schemas for an LLM input, including
 both decoder-only and encoder/decoder input types:
 
-- A text prompt ({class}`str` or {class}`TextPrompt`)
-- A tokenized prompt ({class}`TokensPrompt`)
-- An embeddings prompt ({class}`EmbedsPrompt`)
+- A text prompt ([`str`][] or [`TextPrompt`][vllm.inputs.data.TextPrompt])
+- A tokenized prompt ([`TokensPrompt`][vllm.inputs.data.TokensPrompt])
+- An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.data.EmbedsPrompt])
 - A single data structure containing both an encoder and a decoder prompt
-  ({class}`ExplicitEncoderDecoderPrompt`)
+  ([`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt])
 """
 
 
@@ -189,7 +193,8 @@ def token_inputs(
     prompt: Optional[str] = None,
     cache_salt: Optional[str] = None,
 ) -> TokenInputs:
-    """Construct {class}`TokenInputs` from optional values."""
+    """Construct [`TokenInputs`][vllm.inputs.data.TokenInputs] from optional
+    values."""
     inputs = TokenInputs(type="token", prompt_token_ids=prompt_token_ids)
 
     if prompt is not None:
@@ -221,7 +226,8 @@ def embeds_inputs(
     prompt_embeds: torch.Tensor,
     cache_salt: Optional[str] = None,
 ) -> EmbedsInputs:
-    """Construct :class:`EmbedsInputs` from optional values."""
+    """Construct [`EmbedsInputs`][vllm.inputs.data.EmbedsInputs] from optional
+    values."""
     inputs = EmbedsInputs(type="embeds", prompt_embeds=prompt_embeds)
 
     if cache_salt is not None:
@@ -232,7 +238,7 @@ def embeds_inputs(
 
 DecoderOnlyInputs = Union[TokenInputs, EmbedsInputs, "MultiModalInputs"]
 """
-The inputs in {class}`~vllm.LLMEngine` before they are
+The inputs in [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] before they are
 passed to the model executor.
 This specifies the data required for decoder-only models.
 """
@@ -240,11 +246,12 @@ This specifies the data required for decoder-only models.
 
 class EncoderDecoderInputs(TypedDict):
     """
-    The inputs in {class}`~vllm.LLMEngine` before they are
-    passed to the model executor.
+    The inputs in [`LLMEngine`][vllm.engine.llm_engine.LLMEngine] before they
+    are passed to the model executor.
 
     This specifies the required data for encoder-decoder models.
     """
+
     encoder: Union[TokenInputs, "MultiModalInputs"]
     """The inputs for the encoder portion."""
 
@@ -254,13 +261,13 @@ class EncoderDecoderInputs(TypedDict):
 
 SingletonInputs = Union[TokenInputs, EmbedsInputs, "MultiModalInputs"]
 """
-A processed {class}`SingletonPrompt` which can be passed to
-{class}`vllm.sequence.Sequence`.
+A processed [`SingletonPrompt`][vllm.inputs.data.SingletonPrompt] which can be 
+passed to [`vllm.sequence.Sequence`][].
 """
 
 ProcessorInputs = Union[DecoderOnlyInputs, EncoderDecoderInputs]
 """
-The inputs to {data}`vllm.inputs.InputProcessor`.
+The outputs from [`vllm.inputs.preprocess.InputPreprocessor`][].
 """
 
 _T1 = TypeVar("_T1", bound=SingletonPrompt, default=SingletonPrompt)
@@ -277,7 +284,8 @@ def build_explicit_enc_dec_prompt(
     return ExplicitEncoderDecoderPrompt(
         encoder_prompt=encoder_prompt,
         decoder_prompt=decoder_prompt,
-        mm_processor_kwargs=mm_processor_kwargs)
+        mm_processor_kwargs=mm_processor_kwargs,
+    )
 
 
 def zip_enc_dec_prompts(
@@ -288,7 +296,8 @@ def zip_enc_dec_prompts(
 ) -> list[ExplicitEncoderDecoderPrompt[_T1, _T2]]:
     """
     Zip encoder and decoder prompts together into a list of
-    {class}`ExplicitEncoderDecoderPrompt` instances.
+    [`ExplicitEncoderDecoderPrompt`][vllm.inputs.data.ExplicitEncoderDecoderPrompt]
+    instances.
 
     ``mm_processor_kwargs`` may also be provided; if a dict is passed, the same
     dictionary will be used for every encoder/decoder prompt. If an iterable is
@@ -299,9 +308,10 @@ def zip_enc_dec_prompts(
     if isinstance(mm_processor_kwargs, dict):
         return [
             build_explicit_enc_dec_prompt(
-                encoder_prompt, decoder_prompt,
-                cast(dict[str, Any], mm_processor_kwargs))
-            for (encoder_prompt,
+                encoder_prompt,
+                decoder_prompt,
+                cast(dict[str, Any], mm_processor_kwargs),
+            ) for (encoder_prompt,
                    decoder_prompt) in zip(enc_prompts, dec_prompts)
         ]
     return [

vllm/inputs/parse.py

@@ -23,13 +23,13 @@ class ParsedTokens(TypedDict):
 
 @overload
 def parse_and_batch_prompt(
-        prompt: Union[str, list[str]]) -> Sequence[ParsedText]:
+    prompt: Union[str, list[str]], ) -> Sequence[ParsedText]:
     ...
 
 
 @overload
 def parse_and_batch_prompt(
-        prompt: Union[list[int], list[list[int]]]) -> Sequence[ParsedTokens]:
+    prompt: Union[list[int], list[list[int]]], ) -> Sequence[ParsedTokens]:
     ...
 
 
@@ -86,7 +86,7 @@ class ParsedTokensPrompt(TypedDict):
 
 
 class ParsedEmbedsPrompt(TypedDict):
-    type: Literal['embeds']
+    type: Literal["embeds"]
     content: EmbedsPrompt
 
 
@@ -133,7 +133,7 @@ def parse_singleton_prompt(prompt: SingletonPrompt) -> ParsedSingletonPrompt:
 
 
 def is_explicit_encoder_decoder_prompt(
-        prompt: PromptType) -> TypeIs[ExplicitEncoderDecoderPrompt]:
+    prompt: PromptType, ) -> TypeIs[ExplicitEncoderDecoderPrompt]:
     return isinstance(prompt, dict) and "encoder_prompt" in prompt
 
 

vllm/inputs/preprocess.py

@@ -67,11 +67,11 @@ class InputPreprocessor:
         return self.tokenizer.get_lora_tokenizer(lora_request).eos_token_id
 
     def get_decoder_start_token_id(self) -> Optional[int]:
-        '''
+        """
         Obtain the decoder start token id employed by an encoder/decoder
         model. Returns None for non-encoder/decoder models or if the
         model config is unavailable.
-        '''
+        """
 
         if not self.model_config.is_encoder_decoder:
             logger.warning_once(
@@ -79,14 +79,14 @@ class InputPreprocessor:
                 "this is not an encoder/decoder model.")
             return None
 
-        if (self.model_config is None or self.model_config.hf_config is None):
+        if self.model_config is None or self.model_config.hf_config is None:
             logger.warning_once(
                 "Using None for decoder start token id because "
                 "model config is not available.")
             return None
 
         dec_start_token_id = getattr(self.model_config.hf_config,
-                                     'decoder_start_token_id', None)
+                                     "decoder_start_token_id", None)
         if dec_start_token_id is None:
             logger.warning_once(
                 "Falling back on <BOS> for decoder start token "
@@ -97,7 +97,7 @@ class InputPreprocessor:
         return dec_start_token_id
 
     def _get_default_enc_dec_decoder_prompt(self) -> list[int]:
-        '''
+        """
         Specifically for encoder/decoder models:
         generate a default decoder prompt for when
         the user specifies only the encoder prompt.
@@ -126,7 +126,7 @@ class InputPreprocessor:
         Returns:
 
         * prompt_token_ids
-        '''
+        """
 
         bos_token_id = self.get_bos_token_id()
         assert bos_token_id is not None
@@ -224,7 +224,10 @@ class InputPreprocessor:
         lora_request: Optional[LoRARequest],
         tokenization_kwargs: Optional[dict[str, Any]] = None,
     ) -> list[int]:
-        """Async version of {meth}`_tokenize_prompt`."""
+        """
+        Async version of
+        [`_tokenize_prompt`][vllm.inputs.preprocess.InputPreprocessor._tokenize_prompt].
+        """
         tokenizer = self.get_tokenizer_group()
         tokenization_kwargs = self._get_tokenization_kw(tokenization_kwargs)
 
@@ -287,7 +290,10 @@ class InputPreprocessor:
         lora_request: Optional[LoRARequest],
         return_mm_hashes: bool = False,
     ) -> MultiModalInputs:
-        """Async version of {meth}`_process_multimodal`."""
+        """
+        Async version of
+        [`_process_multimodal`][vllm.inputs.preprocess.InputPreprocessor._process_multimodal].
+        """
         tokenizer = await self._get_mm_tokenizer_async(lora_request)
 
         mm_processor = self.mm_registry.create_processor(self.model_config,
@@ -472,7 +478,7 @@ class InputPreprocessor:
 
         Returns:
 
-        * {class}`SingletonInputs` instance
+        * [`SingletonInputs`][vllm.inputs.data.SingletonInputs] instance
         """
         parsed = parse_singleton_prompt(prompt)
 
@@ -508,7 +514,10 @@ class InputPreprocessor:
         lora_request: Optional[LoRARequest] = None,
         return_mm_hashes: bool = False,
     ) -> SingletonInputs:
-        """Async version of {meth}`_prompt_to_llm_inputs`."""
+        """
+        Async version of
+        [`_prompt_to_llm_inputs`][vllm.inputs.preprocess.InputPreprocessor._prompt_to_llm_inputs].
+        """
         parsed = parse_singleton_prompt(prompt)
 
         if parsed["type"] == "embeds":
@@ -644,7 +653,9 @@ class InputPreprocessor:
     ) -> EncoderDecoderInputs:
         """
         For encoder/decoder models only:
-        Process an input prompt into an {class}`EncoderDecoderInputs` instance.
+        Process an input prompt into an
+        [`EncoderDecoderInputs`][vllm.inputs.data.EncoderDecoderInputs]
+        instance.
 
         There are two types of input prompts:
         singleton prompts which carry only the
@@ -670,7 +681,8 @@ class InputPreprocessor:
 
         Returns:
 
-        * {class}`EncoderDecoderInputs` instance
+        * [`EncoderDecoderInputs`][vllm.inputs.data.EncoderDecoderInputs]
+          instance
         """
         encoder_inputs: SingletonInputs
         decoder_inputs: Optional[SingletonInputs]
@@ -710,7 +722,10 @@ class InputPreprocessor:
         prompt: PromptType,
         tokenization_kwargs: Optional[dict[str, Any]] = None,
     ) -> EncoderDecoderInputs:
-        """Async version of {meth}`_process_encoder_decoder_prompt`."""
+        """
+        Async version of
+        [`_process_encoder_decoder_prompt`][vllm.inputs.preprocess.InputPreprocessor._process_encoder_decoder_prompt].
+        """
         encoder_inputs: SingletonInputs
         decoder_inputs: Optional[SingletonInputs]
 
@@ -778,7 +793,8 @@ class InputPreprocessor:
     ) -> DecoderOnlyInputs:
         """
         For decoder-only models:
-        Process an input prompt into an {class}`DecoderOnlyInputs` instance.
+        Process an input prompt into a
+        [`DecoderOnlyInputs`][vllm.inputs.data.DecoderOnlyInputs] instance.
 
         Arguments:
 
@@ -789,7 +805,7 @@ class InputPreprocessor:
 
         Returns:
 
-        * {class}`DecoderOnlyInputs` instance
+        * [`DecoderOnlyInputs`][vllm.inputs.data.DecoderOnlyInputs] instance
         """
 
         prompt_comps = self._prompt_to_llm_inputs(
@@ -812,7 +828,10 @@ class InputPreprocessor:
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         return_mm_hashes: bool = False,
     ) -> DecoderOnlyInputs:
-        """Async version of {meth}`_process_decoder_only_prompt`."""
+        """
+        Async version of
+        [`_process_decoder_only_prompt`][vllm.inputs.preprocess.InputPreprocessor._process_decoder_only_prompt].
+        """
         prompt_comps = await self._prompt_to_llm_inputs_async(
             prompt,
             tokenization_kwargs=tokenization_kwargs,
@@ -863,7 +882,10 @@ class InputPreprocessor:
         prompt_adapter_request: Optional[PromptAdapterRequest] = None,
         return_mm_hashes: bool = False,
     ) -> ProcessorInputs:
-        """Async version of {meth}`preprocess`."""
+        """
+        Async version of
+        [`preprocess`][vllm.inputs.preprocess.InputPreprocessor.preprocess].
+        """
         if self.model_config.is_encoder_decoder:
             assert not return_mm_hashes, (
                 "Multimodal hashes for encoder-decoder models should not be ",

vllm/inputs/registry.py

@@ -38,7 +38,7 @@ class InputContext:
     ) -> _C:
         """
         Get the HuggingFace configuration
-        ({class}`transformers.PretrainedConfig`) of the model,
+        (`transformers.PretrainedConfig`) of the model,
         additionally checking its type.
 
         Raises:
@@ -79,7 +79,7 @@ class InputContext:
     ) -> _P:
         """
         Get the HuggingFace processor
-        ({class}`transformers.ProcessorMixin`) of the model,
+        (`transformers.ProcessorMixin`) of the model,
         additionally checking its type.
 
         Raises:

vllm/logger.py

@@ -68,22 +68,22 @@ class _VllmLogger(Logger):
     """
     Note:
         This class is just to provide type information.
-        We actually patch the methods directly on the {class}`logging.Logger`
+        We actually patch the methods directly on the [`logging.Logger`][]
         instance to avoid conflicting with other libraries such as
         `intel_extension_for_pytorch.utils._logger`.
     """
 
     def info_once(self, msg: str, *args: Hashable) -> None:
         """
-        As {meth}`info`, but subsequent calls with the same message
-        are silently dropped.
+        As [`info`][logging.Logger.info], but subsequent calls with
+        the same message are silently dropped.
         """
         _print_info_once(self, msg, *args)
 
     def warning_once(self, msg: str, *args: Hashable) -> None:
         """
-        As {meth}`warning`, but subsequent calls with the same message
-        are silently dropped.
+        As [`warning`][logging.Logger.warning], but subsequent calls with
+        the same message are silently dropped.
         """
         _print_warning_once(self, msg, *args)
 

vllm/model_executor/layers/sampler.py

@@ -238,7 +238,9 @@ class Sampler(nn.Module):
             * Defer Pythonization of sampling result & logprobs
             tensor
             * Encapsulate arguments required for deferred Pythonization
-          in the {class}`SamplerOutput` structure
+            in the
+            [`SamplerOutput`][vllm.model_executor.layers.sampler.SamplerOutput]
+            structure
 
         Args:
             logits: (num_tokens, vocab_size).

vllm/model_executor/models/interfaces.py

@@ -226,9 +226,11 @@ class SupportsPP(Protocol):
         intermediate_tensors: Optional["IntermediateTensors"],
     ) -> Union[Tensor, "IntermediateTensors"]:
         """
-        Accept {class}`IntermediateTensors` when PP rank > 0.
+        Accept [`IntermediateTensors`][vllm.sequence.IntermediateTensors] when
+        PP rank > 0.
 
-        Return {class}`IntermediateTensors` only for the last PP rank.
+        Return [`IntermediateTensors`][vllm.sequence.IntermediateTensors] only
+        for the last PP rank.
         """
         ...