vLLM Parameter Curation

Engine version: 0.7.3
Discovered at: 2026-05-06 22:57:22+02:00
Discovery method: dataclasses.fields(EngineArgs) + msgspec.json.schema(SamplingParams)

Summary: 32/135 parameters curated (104 engine + 31 sampling discovered)

Delta vs previous: deferred until first probe-pass cycle.

Engine Parameters

Field	Type	Default	Curated?
additional_config	dict[str, Any] | None	null	-
allowed_local_media_path	str	``	-
block_size	int | None	null	yes
calculate_kv_scales	bool | None	null	-
code_revision	str | None	null	-
collect_detailed_traces	str | None	null	-
compilation_config	CompilationConfig | None	null	yes
config_format	ConfigFormat	auto	-
cpu_offload_gb	float	0	yes
device	str	auto	-
disable_async_output_proc	bool	False	-
disable_custom_all_reduce	bool	False	yes
disable_log_stats	bool	False	-
disable_logprobs_during_spec_decoding	bool | None	null	-
disable_mm_preprocessor_cache	bool	False	-
disable_sliding_window	bool	False	-
distributed_executor_backend	str | type[ExecutorBase] | None	null	yes
download_dir	str | None	null	-
dtype	str	auto	yes
enable_chunked_prefill	bool | None	null	yes
enable_lora	bool	False	-
enable_lora_bias	bool	False	-
enable_prefix_caching	bool | None	null	yes
enable_prompt_adapter	bool	False	-
enable_sleep_mode	bool	False	-
enforce_eager	bool | None	null	yes
fully_sharded_loras	bool	False	-
generation_config	str | None	null	-
gpu_memory_utilization	float	0.9	yes
guided_decoding_backend	str	xgrammar	-
hf_overrides	dict[str, Any] | Callable[[<class 'transformers.configuration_utils.PretrainedConfig'>], PretrainedConfig] | None	null	-
ignore_patterns	str | list[str] | None	null	-
kv_cache_dtype	str	auto	yes
kv_transfer_config	KVTransferConfig | None	null	-
limit_mm_per_prompt	Mapping[str, int] | None	null	-
load_format	str	auto	-
logits_processor_pattern	str | None	null	-
long_lora_scaling_factors	tuple[float] | None	null	-
long_prefill_token_threshold	int | None	0	-
lora_dtype	str | dtype | None	auto	-
lora_extra_vocab_size	int	256	-
max_cpu_loras	int | None	null	-
max_logprobs	int	20	-
max_long_partial_prefills	int | None	1	-
max_lora_rank	int	16	-
max_loras	int	1	-
max_model_len	int | None	null	yes
max_num_batched_tokens	int | None	null	yes
max_num_partial_prefills	int | None	1	-
max_num_seqs	int | None	null	yes
max_parallel_loading_workers	int | None	null	-
max_prompt_adapter_token	int	0	-
max_prompt_adapters	int	1	-
max_seq_len_to_capture	int	8192	yes
mm_processor_kwargs	dict[str, Any] | None	null	-
model	str	facebook/opt-125m	yes
model_impl	str	auto	-
model_loader_extra_config	dict | None	null	-
multi_step_stream_outputs	bool	True	-
ngram_prompt_lookup_max	int | None	null	-
ngram_prompt_lookup_min	int | None	null	-
num_gpu_blocks_override	int | None	null	-
num_lookahead_slots	int	0	-
num_scheduler_steps	int	1	yes
num_speculative_tokens	int | None	null	yes
otlp_traces_endpoint	str | None	null	-
override_generation_config	dict[str, Any] | None	null	-
override_neuron_config	dict[str, Any] | None	null	-
override_pooler_config	PoolerConfig | None	null	-
pipeline_parallel_size	int	1	yes
preemption_mode	str | None	null	-
qlora_adapter_name_or_path	str | None	null	-
quantization	str | None	null	yes
ray_workers_use_nsight	bool	False	-
revision	str | None	null	-
rope_scaling	dict[str, Any] | None	null	-
rope_theta	float | None	null	-
scheduler_cls	str | type[object]	vllm.core.scheduler.Scheduler	-
scheduler_delay_factor	float	0.0	-
scheduling_policy	Literal['fcfs', 'priority']	fcfs	-
seed	int	0	-
served_model_name	str | list[str] | None	null	-
skip_tokenizer_init	bool	False	-
spec_decoding_acceptance_method	str	rejection_sampler	-
speculative_disable_by_batch_size	int | None	null	-
speculative_disable_mqa_scorer	bool | None	False	-
speculative_draft_tensor_parallel_size	int | None	null	-
speculative_max_model_len	int | None	null	-
speculative_model	str | None	null	-
speculative_model_quantization	str | None	null	-
swap_space	float	4	yes
task	Literal['auto', 'generate', 'embedding', 'embed', 'classify', 'score', 'reward', 'transcription']	auto	-
tensor_parallel_size	int	1	yes
tokenizer	str | None	null	-
tokenizer_mode	str	auto	-
tokenizer_pool_extra_config	dict[str, Any] | None	null	-
tokenizer_pool_size	int	0	-
tokenizer_pool_type	str | type[ForwardRef('BaseTokenizerGroup')]	ray	-
tokenizer_revision	str | None	null	-
trust_remote_code	bool	False	-
typical_acceptance_sampler_posterior_alpha	float | None	null	-
typical_acceptance_sampler_posterior_threshold	float | None	null	-
use_v2_block_manager	bool	True	-
worker_cls	str	auto	-

Sampling Parameters

Field	Type	Default	Curated?
_all_stop_token_ids	array	[]	-
_real_n	unknown	null	-
allowed_token_ids	unknown	null	-
bad_words	unknown	null	-
best_of	unknown	null	-
detokenize	boolean	True	-
frequency_penalty	number	0.0	yes
guided_decoding	unknown	null	-
ignore_eos	boolean	False	yes
include_stop_str_in_output	boolean	False	-
logit_bias	unknown	null	-
logits_processors	unknown	null	-
logprobs	unknown	null	-
max_tokens	unknown	16	-
min_p	number	0.0	yes
min_tokens	integer	0	yes
n	integer	1	yes
output_kind	unknown	0	-
output_text_buffer_length	integer	0	-
presence_penalty	number	0.0	yes
prompt_logprobs	unknown	null	-
repetition_penalty	number	1.0	yes
seed	unknown	null	-
skip_special_tokens	boolean	True	-
spaces_between_special_tokens	boolean	True	-
stop	unknown	null	-
stop_token_ids	unknown	null	-
temperature	number	1.0	yes
top_k	integer	-1	yes
top_p	number	1.0	yes
truncate_prompt_tokens	unknown	null	-