Configuration cheatsheet

This page is the index of VllmConfig and its sub-configs. For the architecture and how the configs flow through the engine, see Configuration.

VllmConfig (vllm/config/vllm.py)

The aggregate dataclass — every other config below is a field on it.

Field	Type
model_config	ModelConfig
cache_config	CacheConfig
parallel_config	ParallelConfig
scheduler_config	SchedulerConfig
device_config	DeviceConfig
lora_config	LoRAConfig
speculative_config	SpeculativeConfig
multimodal_config	MultiModalConfig
pooler_config	PoolerConfig
decoding_config	StructuredOutputsConfig
observability_config	ObservabilityConfig
kv_transfer_config	KVTransferConfig
kv_events_config	KVEventsConfig
compilation_config	CompilationConfig
quant_config	QuantizationConfig (resolved at runtime)
kernel_config	KernelConfig
load_config	LoadConfig
attention_config	AttentionConfig
mamba_config	MambaConfig
profiler_config	ProfilerConfig
online_quantization_config_args	OnlineQuantizationConfigArgs
reasoning_config	ReasoningConfig
speech_to_text_config	SpeechToTextConfig
offload_config	OffloadConfig
weight_transfer_config	WeightTransferConfig
ec_transfer_config	ECTransferConfig

set_current_vllm_config(config) / get_current_vllm_config() provide thread-local access for layers and kernels.

ModelConfig (vllm/config/model.py)

Field	Notes
model	HF id or local path
tokenizer	Override tokenizer path
tokenizer_mode	auto, slow, mistral, custom
runner	auto, generate, pool, draft
convert	Convert architecture on load
dtype	auto, float16, bfloat16, float32
kv_cache_dtype	auto, fp8_e5m2, fp8_e4m3, ...
quantization	Selects quant method (see Quantization page)
seed	RNG seed
max_model_len	Override context window
max_logprobs, disable_sliding_window, disable_cascade_attn
enable_prompt_embeds	Allow prompt-embedding inputs
served_model_name	Name reported to /v1/models
task	Pre-set runner task
revision, code_revision, rope_scaling, rope_theta
trust_remote_code	HF-style trust toggle
mm_processor_kwargs	Multi-modal processor kwargs

ParallelConfig (vllm/config/parallel.py)

Field	Notes
tensor_parallel_size	TP degree
pipeline_parallel_size	PP degree
data_parallel_size	Across-replica DP
data_parallel_rank, data_parallel_external_lb	DP coordination
enable_expert_parallel	EP for MoE
expert_parallel_size	EP degree
eplb_config	EPLBConfig (rebalancing)
distributed_executor_backend	mp, ray, uni, external_launcher
worker_extension_cls	Plugin class for worker extensions
enable_dbo	Dual-batch overlap
batch_invariant	Force deterministic execution
nccl_backend	nccl, gloo, ...

SchedulerConfig (vllm/config/scheduler.py)

Field	Notes
max_num_seqs	Max requests in a step
max_num_batched_tokens	Max tokens per step
max_model_len	Mirror of ModelConfig.max_model_len
enable_chunked_prefill	Chunk long prefills
long_prefill_token_threshold	When to chunk
policy	fcfs, priority
async_scheduling	Overlap schedule with execute
cuda_graph_sizes	List of batch sizes to capture
enable_swap	CPU swap (V0 holdover)
is_multimodal_model	Forces MM-aware scheduling

CacheConfig (vllm/config/cache.py)

Field	Notes
block_size	KV block size in tokens
gpu_memory_utilization	Fraction of GPU mem to use
swap_space	CPU swap GB (V0)
cache_dtype	auto, fp8, ...
num_gpu_blocks_override	Manual block count
enable_prefix_caching	Toggle prefix caching
prefix_cache_hash_algo	sha256, builtin, xxhash
cpu_offload_gb	KV offload to CPU
prefix_cache_salt	Default per-engine salt

CompilationConfig (vllm/config/compilation.py)

Field	Notes
level	torch.compile mode
cuda_graph_mode	NONE, FULL, PIECEWISE, FULL_AND_PIECEWISE
pass_config	Per-FX pass switches (fusion, custom ops)
inductor_compile_config	Forwarded to inductor
splitting_ops	Ops to split graphs around
compile_sizes, capture_sizes	Shapes to specialize

LoRAConfig (vllm/config/lora.py)

Field	Notes
max_loras	Max active adapters
max_lora_rank	Max rank
lora_extra_vocab_size	Extra vocab tokens
lora_dtype	dtype for LoRA tensors
enable_lora_bias	Train bias terms
default_mm_loras	Default multimodal LoRAs

SpeculativeConfig (vllm/config/speculative.py)

Selects the proposer (method), draft model, number of speculative tokens, plus method-specific options. See Speculative decoding.

KernelConfig (vllm/config/kernel.py)

Per-kernel feature flags. Use these to opt into experimental kernels (deepgemm, flashinfer cutlass MoE, AITER paths, ...).

ObservabilityConfig (vllm/config/observability.py)

Field	Notes
otlp_traces_endpoint	OTLP exporter URL
collect_detailed_traces	Per-request span attributes
dump_engine_exception_path	Where to dump on EngineCore exceptions
enable_kv_cache_metrics	Extra KV gauges

Where --help lives

vllm/utils/argparse_utils.py::FlexibleArgumentParser slices vllm serve --help per sub-config:

vllm serve --help=ModelConfig
vllm serve --help=SchedulerConfig
vllm serve --help=CompilationConfig
vllm serve --help=all

The argparse generator at vllm/engine/arg_utils.py emits a flag per dataclass field. Field docstrings (parsed by vllm/config/utils.py::get_attr_docs) become --help text.

Environment variables

vllm/envs.py documents every supported environment variable (over 200), including VLLM_USE_V1, VLLM_FLASH_ATTN_VERSION, VLLM_USE_TRITON_FLASH_ATTN, VLLM_USE_DEEP_GEMM, VLLM_LOGGING_LEVEL, VLLM_DO_NOT_TRACK, VLLM_TORCH_PROFILER_DIR, etc. Per-platform plugin packages may add more.