config.json

{
  "_name_or_path": "../fairseq-moe/fairseq-moe-15b",
  "activation_function": "gelu",
  "architectures": [
    "FSGPTMoEForCausalLM"
  ],
  "attention_dropout": 0.1,
  "attention_heads": 12,
  "batch_prioritized_routing": false,
  "bos_token_id": 50257,
  "classifier_dropout": 0.1,
  "d_model": 768,
  "embed_dropout": 0.0,
  "eos_token_id": 50259,
  "expert_capacity": null,
  "ffn_dim": 3072,
  "hidden_size": 768,
  "intermediate_size": 3072,
  "max_position_embeddings": 2048,
  "moe_eval_capacity_token_fraction": -1.0,
  "moe_token_dropout": 0.2,
  "no_scale_embedding": false,
  "normalize_router_prob_before_dropping": false,
  "num_experts": 512,
  "num_heads": 12,
  "num_layers": 12,
  "output_router_logits": true,
  "pad_token_id": 50258,
  "resid_dropout": 0.1,
  "router_bias": false,
  "router_dtype": "float32",
  "router_ignore_padding_tokens": true,
  "second_expert_policy": "all",
  "sparse_step": 2,
  "tokenizer_class": "GPT2Tokenizer",
  "torch_dtype": "bfloat16",
  "transformers_version": "4.33.2",
  "unk_token_id": 3,
  "use_cache": true,
  "vocab_size": 51200
}