Data models

Cross-process boundaries demand explicit, versioned dataclasses. This page enumerates the most important ones with the file they live in.

Sampling and pooling

Class	File	Role
SamplingParams	vllm/sampling_params.py	Per-request decoding configuration
StructuredOutputsParams	vllm/sampling_params.py	Grammar / regex / json constraints
BeamSearchParams	vllm/sampling_params.py	Beam search options
RequestOutputKind	vllm/sampling_params.py	Streaming kind (cumulative/delta/final)
GuidedDecodingParams	(alias for StructuredOutputsParams)	Backwards-compatible name
PoolingParams	vllm/pooling_params.py	Pooling-task options

Inputs

Class	File	Role
PromptType	vllm/inputs/data.py	Union: TokensPrompt / TextPrompt / EmbedsPrompt / multimodal
TokensPrompt	vllm/inputs/data.py	Pre-tokenized prompt
TextPrompt	vllm/inputs/data.py	Plain text prompt
EmbedsPrompt	vllm/inputs/data.py	Direct embedding inputs
MultiModalDataDict	vllm/multimodal/inputs.py	Per-modality data (images, video, audio)
MultiModalKwargs	vllm/multimodal/inputs.py	Processed kwargs returned to layers

Outputs

Class	File	Role
RequestOutput	vllm/outputs.py	Generation result: outputs, logprobs, finished
CompletionOutput	vllm/outputs.py	One sampled completion's tokens / text
EmbeddingRequestOutput	vllm/outputs.py	Embedding pooling result
ClassificationRequestOutput	vllm/outputs.py	Classification pooling result
ScoringRequestOutput	vllm/outputs.py	Score / rerank result
PoolingRequestOutput	vllm/outputs.py	Generic pooling result
Logprob, PromptLogprobs	vllm/logprobs.py	Log-prob containers

V1 engine internals

Class	File	Role
EngineCoreRequest	vllm/v1/engine/__init__.py	Cross-process request payload
EngineCoreOutput	vllm/v1/engine/__init__.py	Per-step output to AsyncLLM
EngineCoreOutputs	vllm/v1/engine/__init__.py	Batched outputs
Request	vllm/v1/request.py	Engine-side request state
RequestStatus, FinishReason	vllm/v1/request.py	Lifecycle markers
KVCacheBlocks, KVCacheBlock	vllm/v1/core/kv_cache_utils.py	Cache addressing
BlockHash, BlockHashWithGroupId	vllm/v1/core/kv_cache_utils.py	Prefix cache keys
KVCacheConfig	vllm/v1/kv_cache_interface.py	Per-group KV cache layout
KVCacheSpec	vllm/v1/kv_cache_interface.py	Backend-declared shape
SchedulerOutput, NewRequestData	vllm/v1/core/sched/output.py	Step plan
ModelRunnerOutput	vllm/v1/outputs.py	Worker → scheduler results
SchedulerStats	vllm/v1/core/sched/output.py	Per-step counters
IterationStats, PrefixCacheStats	vllm/v1/metrics/stats.py	Engine-level stats
SpecDecodeMetadata	vllm/v1/spec_decode/metadata.py	Per-step spec-decode info
SpecDecodingStats	vllm/v1/spec_decode/metrics.py	Spec-decode metrics
SamplingMetadata	vllm/v1/sample/metadata.py	Sampler input batch

Worker-side payloads

Class	File	Role
ExecuteModelRequest	vllm/sequence.py (V0) / vllm/v1/...	Cross-rank request payload
IntermediateTensors	vllm/sequence.py	PP carry-over tensors
WorkerProc	vllm/v1/executor/multiproc_executor.py	Per-rank worker process
RPCBaseRequest	vllm/v1/engine/coordinator/zmq_*.py	ZMQ message base

LoRA

Class	File	Role
LoRARequest	vllm/lora/request.py	Per-request adapter selection
LoRAModel	vllm/lora/lora_model.py	Adapter tensors
PunicaWrapper	vllm/lora/punica_wrapper/__init__.py	Per-shape kernel dispatch

Adapters & multimodal

Class	File	Role
MultiModalRegistry	vllm/multimodal/registry.py	Per-modality plugins
BaseMultiModalProcessor	vllm/multimodal/processing.py	HF-processor adapter
MultiModalCache	vllm/multimodal/cache.py	Encoder cache

Connectors

Class	File	Role
KVConnectorBase_V1	vllm/distributed/kv_transfer/kv_connector/v1/base.py	KV transport interface
ECConnectorBase	vllm/distributed/ec_transfer/ec_connector/base.py	Encoder-cache transport
KVConnectorStats	vllm/distributed/kv_transfer/kv_connector/v1/metrics.py	Per-connector metrics
EventPublisher	vllm/distributed/kv_events.py	KV event broadcasts

For the request lifecycle that ties these classes together, see Engine core and Scheduler.