Mila
Deep Neural Network Library
Loading...
Searching...
No Matches
Mila::Dnn::Serialization::WeightsMetadata Struct Referenceexport

The model configuration a weights file carries. More...

Public Attributes

std::string activation
std::string architecture
bool attention_output_gate = false
std::string attention_type
uint32_t embedding_dim
uint32_t expert_hidden_dim = 0
float final_logit_softcapping
uint32_t full_attention_interval = 0
uint32_t global_head_dim
uint32_t global_rotary_dim
uint32_t head_dim
uint32_t hidden_dim
bool key_equals_value
uint32_t linear_conv_kernel_dim = 0
uint32_t linear_head_dim = 0
uint32_t linear_num_key_heads = 0
uint32_t linear_num_value_heads = 0
uint32_t max_seq_length
std::string model_name
float norm_epsilon
std::string norm_type
uint32_t num_experts = 0
uint32_t num_global_kv_heads
uint32_t num_heads
uint32_t num_kv_heads
uint32_t num_layers
float partial_rotary_factor = 0.0f
std::string positional_encoding
float rope_theta
float rope_theta_global
float rope_theta_local
uint32_t sliding_window_pattern
bool tie_word_embeddings = false
uint32_t top_k_experts = 0
bool use_bias
uint32_t vocab_size
uint32_t window

Detailed Description

The model configuration a weights file carries.


The documentation for this struct was generated from the following file: