switch-base-256 / config.json
ybelkada
add model weights
038d8e2
{
"add_router_probs": false,
"architectures": [
"SwitchTransformersForConditionalGeneration"
],
"d_ff": 3072,
"d_kv": 64,
"d_model": 768,
"decoder_sparse_step": 2,
"dense_act_fn": "relu",
"dropout_rate": 0.1,
"encoder_sparse_step": 2,
"eos_token_id": 1,
"expert_capacity": 64,
"feed_forward_proj": "relu",
"initializer_factor": 1.0,
"is_encoder_decoder": true,
"is_gated_act": false,
"layer_norm_epsilon": 1e-06,
"model_type": "switch_transformers",
"num_decoder_layers": 12,
"num_experts": 256,
"num_heads": 12,
"num_layers": 12,
"num_selected_experts": 1,
"num_sparse_decoder_layers": 6,
"num_sparse_encoder_layers": 6,
"pad_token_id": 0,
"relative_attention_max_distance": 128,
"relative_attention_num_buckets": 32,
"router_aux_loss_coef": 0.001,
"router_bias": false,
"router_dtype": "float32",
"router_ignore_padding_tokens": false,
"router_jitter_noise": 0.01,
"router_type": "tokens_masked",
"router_z_loss_coef": 0.001,
"torch_dtype": "float32",
"transformers_version": "4.25.0.dev0",
"use_cache": true,
"vocab_size": 32128
}