vLLM Parameter Curation

Engine version: 0.19.1
Discovered at: 2026-06-22 16:20:19.046420+00:00
Discovery method: dataclasses.fields(EngineArgs) + msgspec.json.schema(SamplingParams) + config/*.py declarative-constraint overlay

Summary: 37/220 parameters curated (185 engine + 35 sampling discovered)

Engine Parameters

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

Sampling Parameters

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