Models

class tensorrt_llm.models.BaichuanForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of BaichuanConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a BaichuanForCausalLM object from give parameters

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, device: str = 'cuda', calib_dataset: str = 'cnn_dailymail', calib_batches: int = 512, calib_batch_size: int = 1, calib_max_seq_length: int = 512, random_seed: int = 1234, tokenizer_max_seq_length: int = 2048, **kwargs)

class tensorrt_llm.models.BertForQuestionAnswering(*args, **kwargs)

Bases: BertBase

forward(input_ids=None, input_lengths=None, token_type_ids=None, position_ids=None, hidden_states=None, max_input_length=None)

class tensorrt_llm.models.BertForSequenceClassification(*args, **kwargs)

Bases: BertBase

forward(input_ids, input_lengths, token_type_ids=None, position_ids=None, hidden_states=None, max_input_length=None)

class tensorrt_llm.models.BertModel(*args, **kwargs)

Bases: BertBase

forward(input_ids=None, input_lengths=None, position_ids=None, token_type_ids=None, hidden_states=None, max_input_length=None)

class tensorrt_llm.models.BloomForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

class tensorrt_llm.models.BloomModel(config: PretrainedConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None, attention_params=None)

class tensorrt_llm.models.ChatGLMConfig(*, chatglm_version: str = 'chatglm3', add_bias_linear: bool = False, add_qkv_bias: bool = True, apply_query_key_layer_scaling: bool = False, apply_residual_connection_post_layernorm: bool = False, rmsnorm: bool = True, rotary_pct: float = 0.5, rotary_base: float = 10000.0, rotary_scaling: dict | None = None, **kwargs)

Bases: PretrainedConfig

classmethod from_hugging_face(hf_config_or_dir: str | transformers.PretrainedConfig, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

to_dict()

class tensorrt_llm.models.ChatGLMForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of ChatGLMConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a LLaMAForCausalLM object from give parameters

prepare_inputs(*args, **kwargs)

See PretrainedModel.prepare_inputs for the detailed parameter list.

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, device: str = 'cuda', calib_dataset: str = 'cnn_dailymail', calib_batches: int = 512, calib_batch_size: int = 1, calib_max_seq_length: int = 512, random_seed: int = 1234, tokenizer_max_seq_length: int = 2048, **kwargs)

class tensorrt_llm.models.ChatGLMModel(config: ChatGLMConfig)

Bases: Module

forward(input_ids: Tensor = None, position_ids: Tensor = None, use_cache: bool = False, attention_mask: Tensor = None, kv_cache_params: KeyValueCacheParams = None, attention_params: AttentionParams = None)

class tensorrt_llm.models.CogVLMConfig(*, mlp_bias: bool = False, attn_bias: bool = False, rotary_base: float = 10000.0, rotary_scaling: dict | None = None, **kwargs)

Bases: PretrainedConfig

to_dict()

class tensorrt_llm.models.CogVLMForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM, TopModelMixin

config_class

alias of CogVLMConfig

default_plugin_config(**kwargs)

Return the default plugin config for this model, when the plugin_config value is not given in to_trt() call. If users need to set different plugin configs, they can start from the return object and change it.

classmethod from_hugging_face(hf_model_dir, dtype='float16', mapping: Mapping | None = None, quant_mode: QuantMode | None = None, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

classmethod quantize(hf_model_dir, output_dir, quant_config: QuantConfig, *, dtype='float16', mapping: Mapping | None = None, calib_batches=512, calib_batch_size=1, random_seed=1234, tokenizer_max_seq_length=2048, **kwargs)

class tensorrt_llm.models.CohereForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of CohereConfig

classmethod from_hugging_face(hf_model_or_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a CohereForCausalLM object from give parameters

class tensorrt_llm.models.DbrxConfig(*, bias: bool = False, clip_qkv: float | None = None, rotary_base: float = 500000.0, rotary_scaling: dict | None = None, moe: MoeConfig | dict | None = None, **kwargs)

Bases: PretrainedConfig

to_dict()

class tensorrt_llm.models.DbrxForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of DbrxConfig

class tensorrt_llm.models.DecoderModel(*args, **kwargs)

Bases: PretrainedModel

check_config(config: PretrainedConfig)

forward(decoder_input_ids: Tensor, encoder_output: Tensor, position_ids=None, token_type_ids=None, use_cache=False, attention_mask_params=None, last_token_ids=None, kv_cache_params=None, attention_params=None, hidden_states=None, lora_params: LoraParams = None, cross_kv_cache_gen: Tensor | None = None, cross_kv_reuse: Tensor | None = None)

precompute_relative_attention_bias(build_config)

prepare_inputs(max_batch_size, max_beam_width, max_decoder_input_len, max_seq_len, max_encoder_input_len, gather_context_logits: bool = False, gather_generation_logits: bool = False, lora_target_modules: List[str] = None, use_cache=True, *args, **kwargs)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

class tensorrt_llm.models.DeepseekForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

classmethod from_hugging_face(hf_model, model_dir, dtype: str = 'auto', mapping: Mapping | None = None, override_fields={}, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

class tensorrt_llm.models.DeepseekV2ForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

classmethod from_hugging_face(hf_model, model_dir, dtype: str = 'auto', mapping: Mapping | None = None, override_fields={}, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

class tensorrt_llm.models.DiT(*args, **kwargs)

Bases: PretrainedModel

check_config(config: PretrainedConfig)

forward(latent, timestep, label)

Forward pass of DiT. latent: (N, C, H, W) timestep: (N,) label: (N,)

forward_with_cfg(x, t, y)

Forward pass with classifier-free guidance.

forward_without_cfg(x, t, y)

Forward pass without classifier-free guidance.

prepare_inputs(max_batch_size, **kwargs)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

unpatchify(x: Tensor)

class tensorrt_llm.models.EagleForCausalLM(*args, **kwargs)

Bases: LLaMAForCausalLM

config_class

alias of EagleConfig

forward(*args, **kwargs)

prepare_inputs(*args, **kwargs)

Inputs needed:

device_request_types: [bs] draft_tokens: [bs, max_draft_len] draft_lens: [bs] spec_decoding_generation_lengths: [bs] spec_decoding_position_offsets: [bs, max_gen_tokens] spec_decoding_packed_mask: [bs, max_draft_len, packed_length] ** eagle_temperature: [bs] rand_data_sample: [bs] rand_data_validation: [bs, max_draft_tokens]

** The mask is tricky since the boolean mask will need to be

packed in runtime. So, the last dim will be:

packed_length = ceil((max_draft_tokens+1)/32)

class tensorrt_llm.models.EncoderModel(*args, **kwargs)

Bases: PretrainedModel

check_config(config: PretrainedConfig)

forward(input_ids: Tensor, input_lengths=None, position_ids=None, token_type_ids=None, hidden_states=None, max_input_length=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None, attention_mask=None, lora_params: LoraParams = None)

precompute_relative_attention_bias(build_config)

prepare_inputs(max_batch_size, max_input_len, prompt_embedding_table_size: int = 0, lora_target_modules: List[str] = None, *args, **kwargs)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

use_prompt_tuning()

Enable p tuning when build the TRT engine, call this before to_trt

class tensorrt_llm.models.FalconConfig(*, bias: bool = False, parallel_attention: bool = False, num_ln_in_parallel_attn: int | None = None, new_decoder_architecture: bool = False, rotary_base: float = 10000.0, **kwargs)

Bases: PretrainedConfig

classmethod from_hugging_face(hf_config_or_dir: str | transformers.PretrainedConfig, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

to_dict()

class tensorrt_llm.models.FalconForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

check_config(config)

config_class

alias of FalconConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a FalconForCausalLM object from give parameters

class tensorrt_llm.models.FalconModel(config: FalconConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, hidden_states=None)

class tensorrt_llm.models.GPTConfig(*, gpt_variant: str = 'gpt2', bias: bool = True, q_scaling: float = 1.0, embedding_scale: float | None = None, apply_query_key_layer_scaling: bool = False, rotary_pct: float = 1.0, rotary_base: float = 10000.0, rotary_scaling: dict | None = None, inner_layernorm: bool = False, norm_before_bmm1: bool = False, moe: MoeConfig | dict | None = None, **kwargs)

Bases: PretrainedConfig

classmethod from_hugging_face(hf_config_or_dir: str | transformers.PretrainedConfig, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

classmethod from_nemo(nemo_ckpt_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

to_dict()

class tensorrt_llm.models.GPTForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of GPTConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a LLaMAForCausalLM object from give parameters

classmethod from_nemo(nemo_ckpt_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, device: str = 'cuda', calib_dataset: str = 'cnn_dailymail', calib_batches: int = 512, calib_batch_size: int = 1, calib_max_seq_length: int = 512, random_seed: int = 1234, tokenizer_max_seq_length: int = 2048, **kwargs)

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

class tensorrt_llm.models.GPTJConfig(*, rotary_dim: int = 64, **kwargs)

Bases: PretrainedConfig

This is the configuration class to store the configuration of GPTJ model.

classmethod from_hugging_face(hf_config_or_dir: str | transformers.PretrainedConfig, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

to_dict()

class tensorrt_llm.models.GPTJForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of GPTJConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config=None, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

class tensorrt_llm.models.GPTJModel(config: GPTJConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None)

class tensorrt_llm.models.GPTModel(config: GPTConfig)

Bases: Module

forward(input_ids, position_ids, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, hidden_states=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None, lora_params=None, spec_decoding_params=None)

class tensorrt_llm.models.GPTNeoXForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

class tensorrt_llm.models.GPTNeoXModel(config: PretrainedConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None)

class tensorrt_llm.models.GemmaConfig(*, architecture: str, rotary_base: float = 10000.0, rotary_scaling: dict | None = None, attn_bias: bool = False, mlp_bias: bool = False, share_embedding_table: Literal[True] = True, position_embedding_type: PositionEmbeddingType = PositionEmbeddingType.rope_gpt_neox, query_pre_attn_scalar: int | None = None, final_logit_softcapping: float | None = None, attn_logit_softcapping: float | None = None, mapping: Mapping | dict | None = None, **kwargs)

Bases: PretrainedConfig

GEMMA2_ADDED_FIELDS = {'attn_logit_softcapping', 'final_logit_softcapping', 'query_pre_attn_scalar'}

GEMMA_ADDED_FIELDS = {'attn_bias', 'inter_layernorms', 'mlp_bias', 'rotary_base', 'rotary_scaling'}

VERBATIM = {'attn_logit_softcapping', 'final_logit_softcapping', 'hidden_act', 'hidden_size', 'intermediate_size', 'max_position_embeddings', 'num_attention_heads', 'num_hidden_layers', 'query_pre_attn_scalar', 'use_parallel_embedding', 'vocab_size'}

classmethod from_hugging_face(hf_config_or_dir: HfConfigOrDir, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs) → GemmaConfig

gemma2_config()

property is_gemma_2: bool

to_dict()

Serialize the fields added in GemmaConfig

class tensorrt_llm.models.GemmaForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

NATIVE_QUANT_FLOW = {QuantAlgo.W4A16, QuantAlgo.W8A16, QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TENSOR_PLUGIN, QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN, QuantAlgo.W8A8_SQ_PER_TENSOR_PER_TOKEN_PLUGIN, QuantAlgo.W8A8_SQ_PER_TENSOR_PLUGIN}

classmethod assert_valid_quant_algo(quant_algo: QuantAlgo | None)

config_class

alias of GemmaConfig

classmethod from_hugging_face(hf_model_dir: HfConfigOrDir, dtype='float16', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, load_model_on_cpu: bool = True, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'float16', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, gemma_config_kwargs: Dict[str, Any] = None, **quantize_kwargs: Dict[str, Any])

class tensorrt_llm.models.LLaMAConfig(*, mlp_bias: bool = False, attn_bias: bool = False, rotary_base: float = 10000.0, rotary_scaling: dict | None = None, residual_mlp: bool = False, disable_weight_only_quant_plugin: bool = False, moe: MoeConfig | dict | None = None, remove_duplicated_kv_heads: bool = False, **kwargs)

Bases: PretrainedConfig

classmethod from_hugging_face(hf_config_or_dir: str | transformers.PretrainedConfig, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

classmethod from_meta_ckpt(meta_ckpt_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

to_dict()

class tensorrt_llm.models.LLaMAForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of LLaMAConfig

default_plugin_config(**kwargs)

Return the default plugin config for this model, when the plugin_config value is not given in to_trt() call. If users need to set different plugin configs, they can start from the return object and change it.

classmethod from_hugging_face(hf_model_or_dir: str | PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a LLaMAForCausalLM object from give parameters

classmethod from_meta_ckpt(meta_ckpt_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, device: str = 'cuda', calib_dataset: str = 'cnn_dailymail', calib_batches: int = 512, calib_batch_size: int = 1, calib_max_seq_length: int = 512, random_seed: int = 1234, tokenizer_max_seq_length: int = 2048, **kwargs)

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

class tensorrt_llm.models.LLaMAModel(config: LLaMAConfig)

Bases: Module

forward(input_ids, position_ids=None, use_cache=False, attention_mask=None, spec_decoding_params=None, kv_cache_params=None, attention_params=None, hidden_states=None, hidden_states_for_embed=None, prompt_embedding_table: Tensor | None = None, prompt_tasks: Tensor | None = None, prompt_vocab_size: Tensor | None = None, lora_params=None)

class tensorrt_llm.models.MLLaMAModel(*args, **kwargs)

Bases: PretrainedModel

config_class

alias of MLLaMAConfig

forward(decoder_input_ids: Tensor, encoder_output: Tensor, use_cache=False, attention_mask_params=None, last_token_ids=None, kv_cache_params=None, attention_params=None, hidden_states=None, lora_params: LoraParams = None, cross_kv_cache_gen: Tensor | None = None, cross_kv_reuse: Tensor | None = None, prompt_embedding_table: Tensor | None = None, prompt_tasks: Tensor | None = None, prompt_vocab_size: Tensor | None = None, skip_cross_attn_blocks: Tensor | None = None)

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create a MLLaMAModel object from give parameters

precompute_relative_attention_bias(build_config)

prepare_inputs(max_batch_size, max_beam_width, max_decoder_input_len, max_seq_len, max_encoder_input_len, gather_context_logits: bool = False, gather_generation_logits: bool = False, lora_target_modules: List[str] = None, prompt_embedding_table_size: int = 0, use_cache=True, *args, **kwargs)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

class tensorrt_llm.models.MPTForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

check_config(config)

class tensorrt_llm.models.MPTModel(config: PretrainedConfig)

Bases: Module

forward(input_ids, position_ids, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None)

class tensorrt_llm.models.MambaForCausalLM(*args, **kwargs)

Bases: PretrainedModel

config_class

alias of MambaConfig

forward(input_ids, conv_states, ssm_states, host_request_types, last_token_ids, last_token_ids_for_logits, host_context_lengths, slot_mapping: Tensor | None = None)

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

prepare_inputs(max_batch_size, max_input_len, max_seq_len, max_num_tokens, use_cache, max_beam_width: int = 1, opt_num_tokens: int = None, opt_batch_size: int = 0, prompt_embedding_table_size: int = 0, max_draft_len: int = 0, gather_context_logits: bool = False, gather_generation_logits: bool = False, lora_target_modules: List[str] = None, speculative_decoding_draft_tokens_external: bool = False)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

class tensorrt_llm.models.MedusaConfig(*, num_medusa_heads: int = 4, num_medusa_layers: int = 1, max_draft_len: int = 63, **kwargs)

Bases: object

to_dict()

class tensorrt_llm.models.MedusaForCausalLm(*args, **kwargs)

Bases: PretrainedModel

config_class

alias of MedusaConfig

class tensorrt_llm.models.OPTForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

check_config(config)

class tensorrt_llm.models.OPTModel(config: PretrainedConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None)

class tensorrt_llm.models.Phi3ForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

config_class

alias of Phi3Config

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

use_lora(lora_config: LoraConfig)

Load lora weights from the give config to the module :param lora_config: the lora config

class tensorrt_llm.models.Phi3Model(config: PretrainedConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None, lora_params=None)

class tensorrt_llm.models.PhiForCausalLM(*args, **kwargs)

Bases: DecoderModelForCausalLM

check_config(config)

config_class

alias of PhiConfig

classmethod from_hugging_face(hf_model_or_dir: str | transformers.PreTrainedModel, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, **kwargs)

Create LLM object and load weights from hugging face :param hf_model_dir: the hugging face model directory :param dtype: str, the default weights data type when loading from the hugging face model :param mapping: Mapping, specify the multi-gpu parallel strategy, when it’s None, single GPU is used

class tensorrt_llm.models.PhiModel(config: PretrainedConfig)

Bases: Module

forward(input_ids: Tensor, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, prompt_embedding_table=None, prompt_tasks=None, prompt_vocab_size=None)

class tensorrt_llm.models.PretrainedConfig(*, architecture: str, dtype: str, hidden_size: int, num_hidden_layers: int, num_attention_heads: int, vocab_size: int | None = None, hidden_act: str = 'gelu', logits_dtype: str = 'float32', norm_epsilon: float = 1e-05, position_embedding_type: PositionEmbeddingType | str = PositionEmbeddingType.learned_absolute, max_position_embeddings: int | None = None, rotary_embedding_dim: int | None = None, num_key_value_heads: int | None = None, intermediate_size: int | None = None, mapping: Mapping | dict | None = None, quantization: QuantConfig | dict | None = None, use_parallel_embedding: bool = False, embedding_sharding_dim: int = 0, share_embedding_table: bool = False, head_size: int | None = None, qk_layernorm: bool = False, runtime_defaults: RuntimeDefaultsIn = None, **kwargs)

Bases: object

static create_runtime_defaults(defaults: RuntimeDefaultsIn = None) → RuntimeDefaults | None

for_each_rank() → Generator[Self, None, None]

classmethod from_checkpoint(ckpt_dir: str)

classmethod from_dict(config: dict)

classmethod from_json_file(config_file: str)

get_config_group(group_cls: Type[CG]) → CG

get_quant_cfg(module_name: str)

has_config_group(group_cls: Type[CG]) → bool

property kv_dtype

property quant_algo

property quant_mode

set_if_not_exist(key, value)

set_rank(rank)

to_dict()

to_json_file(config_file: str)

to_layer_quant_config(config_file: str)

class tensorrt_llm.models.PretrainedModel(*args, **kwargs)

Bases: Module, GenerationMixin, TopModelMixin

check_config(config)

classmethod from_checkpoint(ckpt_dir: str, rank: int | None = None, config: PretrainedConfig | None = None, *, preprocess_weights_hook: Callable[[Dict[str, Tensor]], Dict[str, Tensor]] | None = None)

classmethod from_config(config: PretrainedConfig)

load(weights, from_pruned=False)

prepare_inputs(max_batch_size, max_input_len, max_seq_len, max_num_tokens, use_cache, max_beam_width: int = 1, opt_num_tokens: int = None, prompt_embedding_table_size: int = 0, position_encoding_2d: bool = False, max_draft_len: int = 0, speculative_decoding_draft_tokens_external: bool = False, spec_decoding_is_generation_length_variable: bool = False, gather_context_logits: bool = False, gather_generation_logits: bool = False, lora_target_modules: List[str] = None, opt_batch_size: int = 0, num_hidden_layers: int = None, mrope_rotary_sin_cos_size: int = None)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

classmethod quantize(hf_model_dir: str, output_dir: str, dtype: str = 'auto', mapping: Mapping | None = None, quant_config: QuantConfig | None = None, *, device: str = 'cuda', calib_dataset: str = 'cnn_dailymail', calib_batches: int = 512, calib_batch_size: int = 1, calib_max_seq_length: int = 512, random_seed: int = 1234, tokenizer_max_seq_length: int = 2048, **kwargs)

release()

save_checkpoint(output_dir, save_config=True)

class tensorrt_llm.models.ReDrafterForCausalLM(*args, **kwargs)

Bases: LLaMAForCausalLM

forward(*args, **kwargs)

0. run base model, get logits, hidden_states

prepare_inputs(*args, **kwargs)

Inputs needed:

Assuming, max_gen_tokens = 1 + nb*(bl - 1), counting true token device_request_types: [bs] draft_tokens: [bs, nb, bl] draft_indices: [bs, nb, bl] draft_probs: [bs, nb, bl-1, V] spec_decoding_generation_lengths: [bs] spec_decoding_position_offsets: [bs, max_gen_tokens] spec_decoding_packed_mask: [bs, max_gen_tokens, packed_length] ** redrafter_inverted_temperature: [bs] rand_data_sample: [bs] rand_data_validation: [bs, nb, bl-1]

** The mask is tricky since the boolean mask will need to be

packed in runtime. So, the last dim will be:

packed_length = ceil(max_gen_tokens/32)

class tensorrt_llm.models.RecurrentGemmaForCausalLM(*args, **kwargs)

Bases: PretrainedModel

forward(input_ids, position_ids=None, use_cache=False, attention_mask=None, kv_cache_params=None, attention_params=None, conv_states=None, rnn_states=None, host_request_types=None, last_token_ids=None, last_token_ids_for_logits=None, host_context_lengths=None, slot_mapping=None)

prepare_inputs(max_batch_size, max_input_len, max_seq_len, max_num_tokens, use_cache, max_beam_width: int = 1, opt_num_tokens: int = None, opt_batch_size: int = 0, prompt_embedding_table_size: int = 0, max_draft_len: int = 0, gather_context_logits: bool = False, gather_generation_logits: bool = False, lora_target_modules: List[str] = None, speculative_decoding_draft_tokens_external: bool = False)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()

prepare_recurrent_inputs(max_batch_size, num_profiles, mapping)

tensorrt_llm.models.RobertaForQuestionAnswering

alias of BertForQuestionAnswering

tensorrt_llm.models.RobertaForSequenceClassification

alias of BertForSequenceClassification

tensorrt_llm.models.RobertaModel

alias of BertModel

class tensorrt_llm.models.SpeculativeDecodingMode(value, names=<not given>, *values, module=None, qualname=None, type=None, start=1, boundary=None)

Bases: IntFlag

DRAFT_TOKENS_EXTERNAL = 2

EAGLE = 32

EXPLICIT_DRAFT_TOKENS = 16

LOOKAHEAD_DECODING = 8

MEDUSA = 4

NONE = 1

static from_arguments(args: Namespace)

class tensorrt_llm.models.WhisperEncoder(*args, **kwargs)

Bases: PretrainedModel

forward(input_features: Tensor, input_lengths=None, position_ids=None)

precompute_relative_attention_bias(build_config)

prepare_inputs(max_batch_size=16)

@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the ranges of the dimensions of when using TRT dynamic shapes.

@return: a list contains values which can be fed into the self.forward()