DeepSeek-Flash-Mini / config.json
nowordsxiaomu's picture
Initial release: DeepSeek-Flash-Mini nano (15M MoE, MLA+MTP)
5e6d9f5 verified
Raw
History Blame Contribute Delete
690 Bytes
{
"vocab_size": 8192,
"dim": 256,
"n_layers": 6,
"max_seq_len": 512,
"norm_eps": 1e-05,
"dropout": 0.0,
"tie_embeddings": true,
"n_heads": 8,
"q_lora_rank": 0,
"kv_lora_rank": 64,
"qk_nope_head_dim": 32,
"qk_rope_head_dim": 16,
"v_head_dim": 32,
"attn_impl": "naive",
"rope_theta": 10000.0,
"rope_scaling": 1.0,
"n_dense_layers": 1,
"dense_inter_dim": 704,
"moe_inter_dim": 256,
"n_routed_experts": 8,
"n_shared_experts": 1,
"n_activated_experts": 2,
"n_expert_groups": 1,
"n_limited_groups": 1,
"score_func": "sigmoid",
"route_scale": 1.0,
"bias_update_speed": 0.001,
"aux_loss_alpha": 0.001,
"n_mtp": 1,
"mtp_loss_weight": 0.3
}