vLLM Engine Schema

Engine version: 0.19.1
Discovered at: 2026-07-03T16:25:37+02:00
Discovery method: dataclasses.fields(EngineArgs) + msgspec.json.schema(SamplingParams) + config/*.py declarative-constraint overlay
Schema version: 1.0.0

Summary: 185 engine parameters, 35 sampling parameters.

Discovery limitations

• sampling_params - constraints (e.g. temperature>=0, top_p in (0,1]) live in imperative _verify_args() and are not introspectable from field metadata
• engine_params - per-field descriptions unavailable (vLLM EngineArgs has only a class docstring)

Engine Parameters

Field	Type	Default	Description
model	str	Qwen/Qwen3-0.6B
enable_return_routed_experts	bool	false
model_weights	str	""
served_model_name	`str	list[str]	None`
tokenizer	`str	None`	-
hf_config_path	`str	None`	-
runner	str	auto
convert	str	auto
skip_tokenizer_init	bool	false
enable_prompt_embeds	bool	false
tokenizer_mode	`Literal['auto', 'hf', 'slow', 'mistral', 'deepseek_v32']	str`	auto
trust_remote_code	bool	false
allowed_local_media_path	str	""
allowed_media_domains	`list[str]	None`	-
download_dir	`str	None`	-
safetensors_load_strategy	`str	None`	-
load_format	`str	Any`	auto
config_format	str	auto
dtype	str	auto
kv_cache_dtype	str	auto
seed	int	0
max_model_len	int	-
cudagraph_capture_sizes	`list[int]	None`	-
max_cudagraph_capture_size	`int	None`	-
distributed_executor_backend	`str	Literal['ray', 'mp', 'uni', 'external_launcher']	type[Any]
pipeline_parallel_size	int	1
master_addr	str	127.0.0.1
master_port	int	29501
nnodes	int	1
node_rank	int	0
distributed_timeout_seconds	`int	None`	-
tensor_parallel_size	int	1
prefill_context_parallel_size	int	1
decode_context_parallel_size	int	1
dcp_comm_backend	str	ag_rs
dcp_kv_cache_interleave_size	int	1
cp_kv_cache_interleave_size	int	1
data_parallel_size	int	1
data_parallel_rank	`int	None`	-
data_parallel_start_rank	`int	None`	-
data_parallel_size_local	`int	None`	-
data_parallel_address	`str	None`	-
data_parallel_rpc_port	`int	None`	-
data_parallel_hybrid_lb	bool	false
data_parallel_external_lb	bool	false
data_parallel_backend	str	mp
enable_expert_parallel	bool	false
enable_ep_weight_filter	bool	false
moe_backend	str	auto
all2all_backend	str	allgather_reducescatter
enable_elastic_ep	bool	false
enable_dbo	bool	false
ubatch_size	int	0
dbo_decode_token_threshold	int	32
dbo_prefill_token_threshold	int	512
disable_nccl_for_dp_synchronization	`bool	None`	-
eplb_config	-	-
enable_eplb	bool	false
expert_placement_strategy	str	linear
_api_process_count	int	1
_api_process_rank	int	0
max_parallel_loading_workers	`int	None`	-
block_size	`int	None`	-
enable_prefix_caching	`bool	None`	-
prefix_caching_hash_algo	str	sha256
disable_sliding_window	bool	false
disable_cascade_attn	bool	true
offload_backend	str	auto
cpu_offload_gb	float	0
cpu_offload_params	set[str]	[]
offload_group_size	int	0
offload_num_in_group	int	1
offload_prefetch_step	int	1
offload_params	set[str]	[]
gpu_memory_utilization	float	0.9
kv_cache_memory_bytes	`int	None`	-
max_num_batched_tokens	`int	None`	-
max_num_partial_prefills	int	1
max_long_partial_prefills	int	1
long_prefill_token_threshold	int	0
max_num_seqs	`int	None`	-
max_logprobs	int	20
logprobs_mode	str	raw_logprobs
disable_log_stats	bool	false
aggregate_engine_logging	bool	false
revision	`str	None`	-
code_revision	`str	None`	-
hf_token	`bool	str	None`
hf_overrides	`dict[str, Any]	Callable[[typing.Any], Any]`	{}
tokenizer_revision	`str	None`	-
quantization	`Any	str	None`
allow_deprecated_quantization	bool	false
enforce_eager	bool	false
disable_custom_all_reduce	bool	false
language_model_only	bool	false
limit_mm_per_prompt	`dict[str, int	dict[str, int]]`	{}
enable_mm_embeds	bool	false
interleave_mm_strings	bool	false
media_io_kwargs	dict[str, dict[str, Any]]	{}
mm_processor_kwargs	`dict[str, Any]	None`	-
mm_processor_cache_gb	float	4
mm_processor_cache_type	str	lru
mm_shm_cache_max_object_size_mb	int	128
mm_encoder_only	bool	false
mm_encoder_tp_mode	str	weights
mm_encoder_attn_backend	`AttentionBackendEnum	str	None`
io_processor_plugin	`str	None`	-
renderer_num_workers	int	1
skip_mm_profiling	bool	false
video_pruning_rate	`float	None`	-
mm_tensor_ipc	str	direct_rpc
enable_lora	bool	false
max_loras	int	1
max_lora_rank	int	16
default_mm_loras	`dict[str, str]	None`	-
fully_sharded_loras	bool	false
max_cpu_loras	`int	None`	-
lora_dtype	`str	dtype	None`
lora_target_modules	`list[str]	None`	-
enable_tower_connector_lora	bool	false
specialize_active_lora	bool	false
ray_workers_use_nsight	bool	false
num_gpu_blocks_override	`int	None`	-
model_loader_extra_config	dict	{}
ignore_patterns	`str	list[str]`	["original/**/*"]
enable_chunked_prefill	`bool	None`	-
disable_chunked_mm_input	bool	false
scheduler_reserve_full_isl	bool	true
disable_hybrid_kv_cache_manager	`bool	None`	-
structured_outputs_config	-	-
reasoning_parser	str	""
reasoning_parser_plugin	`str	None`	-
speculative_config	-	-
show_hidden_metrics_for_version	`str	None`	-
otlp_traces_endpoint	`str	None`	-
collect_detailed_traces	`list[Literal['model', 'worker', 'all']]	None`	-
kv_cache_metrics	bool	false
kv_cache_metrics_sample	float	0.01
cudagraph_metrics	bool	false
enable_layerwise_nvtx_tracing	bool	false
enable_mfu_metrics	bool	false
enable_logging_iteration_details	bool	false
enable_mm_processor_stats	bool	false
scheduling_policy	str	fcfs
scheduler_cls	`str	type[object]	None`
pooler_config	-	-
compilation_config	-	-
attention_config	-	-
kernel_config	-	-
enable_flashinfer_autotune	bool	-
worker_cls	str	auto
worker_extension_cls	str	""
profiler_config	-	-
kv_transfer_config	-	-
kv_events_config	-	-
ec_transfer_config	-	-
reasoning_config	-	-
generation_config	str	auto
enable_sleep_mode	bool	false
override_generation_config	dict[str, Any]	{}
model_impl	str	auto
override_attention_dtype	`str	None`	-
attention_backend	`AttentionBackendEnum	None`	-
calculate_kv_scales	bool	false
kv_cache_dtype_skip_layers	list[str]	[]
mamba_cache_dtype	str	auto
mamba_ssm_cache_dtype	str	auto
mamba_block_size	`int	None`	-
mamba_cache_mode	str	none
additional_config	dict[str, Any]	{}
use_tqdm_on_load	bool	true
pt_load_map_location	`str	dict[str, str]`	cpu
logits_processors	`list[str	type[LogitsProcessor]]	None`
async_scheduling	`bool	None`	-
stream_interval	int	1
kv_sharing_fast_prefill	bool	false
optimization_level	OptimizationLevel	2
performance_mode	str	balanced
kv_offloading_size	`float	None`	-
kv_offloading_backend	str	native
tokens_only	bool	false
shutdown_timeout	int	0
weight_transfer_config	-	-
fail_on_environ_validation	bool	false
gdn_prefill_backend	str	-

Sampling Parameters

Field	Type	Default	Description
n	integer	1
presence_penalty	number	0.0
frequency_penalty	number	0.0
repetition_penalty	number	1.0
temperature	number	1.0
top_p	number	1.0
top_k	integer	0
min_p	number	0.0
seed	`int	None`	-
stop	`str	list[str]	None`
stop_token_ids	`list[int]	None`	-
ignore_eos	boolean	false
max_tokens	`int	None`	16
min_tokens	integer	0
logprobs	`int	None`	-
prompt_logprobs	`int	None`	-
flat_logprobs	boolean	false
detokenize	boolean	true
skip_special_tokens	boolean	true
spaces_between_special_tokens	boolean	true
include_stop_str_in_output	boolean	false
output_kind	RequestOutputKind	0
skip_clone	boolean	false
output_text_buffer_length	integer	0
_eos_token_id	`int	None`	-
_all_stop_token_ids	array	[]
structured_outputs	`StructuredOutputsParams	None`	-
logit_bias	`dict[int, float]	None`	-
allowed_token_ids	`list[int]	None`	-
extra_args	`dict[str, Any]	None`	-
bad_words	`list[str]	None`	-
_bad_words_token_ids	`list[list[int]]	None`	-
skip_reading_prefix_cache	`bool	None`	-
thinking_token_budget	`int	None`	-
repetition_detection	`RepetitionDetectionParams	None`	-