Translation
Transformers
PyTorch
nllb-moe
feature-extraction
nllb-moe-54b / config.json
ArthurZ's picture
ArthurZ HF staff
Update config.json
00d1629
raw
history blame
1.41 kB
{
"_name_or_path": "/home/arthur_huggingface_co/fairseq/weights/checkpoints/hf-converted-moe-54b",
"activation_dropout": 0.0,
"activation_function": "relu",
"architectures": [
"NllbMoeModel"
],
"attention_dropout": 0.1,
"batch_prioritized_routing": false,
"bos_token_id": 0,
"d_model": 2048,
"decoder_attention_heads": 16,
"decoder_ffn_dim": 8192,
"decoder_layerdrop": 0,
"decoder_layers": 24,
"decoder_sparse_step": 4,
"decoder_start_token_id": 2,
"dropout": 0.1,
"encoder_attention_heads": 16,
"encoder_ffn_dim": 8192,
"encoder_layerdrop": 0,
"encoder_layers": 24,
"encoder_sparse_step": 4,
"eos_token_id": 2,
"expert_capacity": 64,
"init_std": 0.02,
"is_encoder_decoder": true,
"max_length": 200,
"max_position_embeddings": 1024,
"model_type": "nllb-moe",
"moe_eval_capacity_token_fraction": 1.0,
"moe_token_dropout": 0.2,
"normalize_router_prob_before_dropping": false,
"num_experts": 128,
"num_hidden_layers": 24,
"pad_token_id": 1,
"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,
"scale_embedding": true,
"second_expert_policy": "all",
"torch_dtype": "float32",
"transformers_version": "4.27.0.dev0",
"use_cache": true,
"vocab_size": 256206
}