TensorRT-LLM Parameter Curation

Engine version: 0.21.0
Discovered at: 2026-05-06 20:20:57+02:00
Discovery method: TrtLlmArgs.model_json_schema() + dataclasses.fields(SamplingParams)

Summary: 18/107 parameters curated (60 engine + 47 sampling discovered)

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

Engine Parameters

Field	Type	Default	Curated?
auto_parallel	boolean	False	-
auto_parallel_world_size	integer | None	null	-
backend	string | None	null	yes
batched_logits_processor	Optional[tensorrt_llm.sampling_params.BatchedLogitsProcessor]	null	-
batching_type	BatchingType | None	null	-
build_config	Optional[tensorrt_llm.builder.BuildConfig]	null	-
cache_transceiver_config	CacheTransceiverConfig | None	null	-
calib_config	CalibConfig | None	null	-
context_parallel_size	integer	1	-
cp_config	object | None	null	-
decoding_config	Optional[DecodingConfig]	null	-
dtype	string	auto	yes
embedding_parallel_mode	string	SHARDING_ALONG_VOCAB	-
enable_attention_dp	boolean	False	-
enable_build_cache	Union[tensorrt_llm.llmapi.build_cache.BuildCacheConfig, bool]	False	-
enable_chunked_prefill	boolean	False	-
enable_lora	boolean	False	-
enable_prompt_adapter	boolean	False	-
enable_tqdm	boolean	False	-
extended_runtime_perf_knob_config	ExtendedRuntimePerfKnobConfig | None	null	-
fast_build	boolean	False	yes
garbage_collection_gen0_threshold	integer	20000	-
gather_generation_logits	boolean	False	-
gpus_per_node	integer | None	null	-
guided_decoding_backend	string | None	null	-
iter_stats_max_iterations	integer | None	null	-
kv_cache_config	KvCacheConfig	null	-
load_format	Literal['auto', 'dummy']	auto	-
lora_config	LoraConfig | None	null	-
max_batch_size	integer | None	null	yes
max_beam_width	integer | None	null	-
max_cpu_loras	integer	4	-
max_input_len	integer | None	null	yes
max_lora_rank	integer | None	null	-
max_loras	integer	4	-
max_num_tokens	integer | None	null	yes
max_prompt_adapter_token	integer	0	-
max_seq_len	integer | None	null	yes
model	string	null	-
moe_cluster_parallel_size	integer | None	null	-
moe_expert_parallel_size	integer | None	null	-
moe_tensor_parallel_size	integer | None	null	-
normalize_log_probs	boolean	False	-
num_postprocess_workers	integer	0	-
peft_cache_config	PeftCacheConfig | None	null	-
pipeline_parallel_size	integer	1	yes
postprocess_tokenizer_dir	string | None	null	-
quant_config	QuantConfig | None	null	-
reasoning_parser	string | None	null	-
request_stats_max_iterations	integer | None	null	-
revision	string | None	null	-
scheduler_config	SchedulerConfig	null	-
skip_tokenizer_init	boolean	False	-
speculative_config	LookaheadDecodingConfig | MedusaDecodingConfig | EagleDecodingConfig | MTPDecodingConfig | NGramDecodingConfig | DraftTargetDecodingConfig | None	null	-
tensor_parallel_size	integer	1	yes
tokenizer	string | None	null	-
tokenizer_mode	Literal['auto', 'slow']	auto	-
tokenizer_revision	string | None	null	-
trust_remote_code	boolean	False	-
workspace	string | None	null	-

Sampling Parameters

Field	Type	Default	Curated?
add_special_tokens	bool	True	-
additional_model_outputs	list[AdditionalModelOutput] | None	null	-
apply_batched_logits_processor	bool	False	-
bad	str | list[str] | None	null	-
bad_token_ids	list[int] | None	null	-
beam_search_diversity_rate	float | None	null	-
beam_width_array	list[int] | None	null	-
best_of	int | None	null	-
detokenize	bool	True	-
early_stopping	int | None	null	-
embedding_bias	Tensor | None	null	-
end_id	int | None	null	-
exclude_input_from_output	bool	True	-
frequency_penalty	float | None	null	-
guided_decoding	GuidedDecodingParams | None	null	-
ignore_eos	bool	False	yes
include_stop_str_in_output	bool	False	-
length_penalty	float | None	null	-
logits_processor	LogitsProcessor | list[LogitsProcessor] | None	null	-
logprobs	int | None	null	-
lookahead_config	LookaheadDecodingConfig | None	null	-
max_tokens	int	32	yes
min_p	float | None	null	yes
min_tokens	int | None	null	yes
n	int	1	yes
no_repeat_ngram_size	int | None	null	-
pad_id	int | None	null	-
presence_penalty	float | None	null	-
prompt_logprobs	int | None	null	-
repetition_penalty	float | None	null	yes
return_context_logits	bool	False	-
return_encoder_output	bool	False	-
return_generation_logits	bool	False	-
return_perf_metrics	bool	False	-
seed	int | None	null	-
skip_special_tokens	bool	True	-
spaces_between_special_tokens	bool	True	-
stop	str | list[str] | None	null	-
stop_token_ids	list[int] | None	null	-
temperature	float | None	null	yes
top_k	int | None	null	yes
top_p	float | None	null	yes
top_p_decay	float | None	null	-
top_p_min	float | None	null	-
top_p_reset_ids	int | None	null	-
truncate_prompt_tokens	int | None	null	-
use_beam_search	bool	False	-