Text Generation
Transformers
Safetensors
English
Chinese
ernie5_5_moe
ERNIE5.5
MoE
fp8
ERNIE-5.5-Flash-Base / config.json
mrfakename's picture PJC0819's picture
Duplicate from baidu/ERNIE-5.5-Flash-Base
79864fb
Raw History Blame Contribute Delete
3.81 kB
{
"activation_situ_beta": 4.0,
"activation_situ_linear_beta": 25.0,
"add_full_attention_sink_bias": true,
"architectures": [
"Ernie5_5_MoeForCausalLM"
],
"attention_bias": false,
"bos_token_id": 1,
"compress_ratios": [
128,
128,
128,
128,
128,
128,
128,
-2,
128,
128,
128,
128,
128,
128,
-2,
128,
128,
128,
128,
128,
128,
-2,
128,
128,
128,
128,
128,
128,
-2,
128,
128,
128,
128,
128,
128,
-2,
128,
128,
128,
128,
128,
128,
-2,
-2
],
"compress_rope_theta": 160000.0,
"csa_dense_mode": true,
"enable_hyper_connections": true,
"eos_token_id": 2,
"experimental_attention_variant": "dsv4_hybrid",
"first_k_dense_replace": 1,
"gated_attn_act": "sigmoid",
"gated_attn_use_q_lora": true,
"gated_linear_unit": true,
"hc_eps": 1e-06,
"hc_mult": 4,
"hc_sinkhorn_iters": 20,
"hca_rope_type": "yarn",
"head_dim": 512,
"hidden_act": "situ",
"hidden_size": 4096,
"hybrid_mla_attention": "mqa_dsa",
"hybrid_mla_kv_lora_rank": 512,
"hybrid_mla_num_attention_heads": 64,
"hybrid_mla_num_key_value_heads": 64,
"hybrid_mla_q_lora_rank": 1024,
"hybrid_mla_qk_nope_head_dim": 192,
"hybrid_mla_qk_rope_head_dim": 64,
"hybrid_mla_v_head_dim": 256,
"index_head_dim": 128,
"index_n_heads": 64,
"index_topk": 2048,
"intermediate_size": 14336,
"kv_lora_rank": 512,
"latent_moe_use_norm": true,
"max_position_embeddings": 1048576,
"model_type": "ernie5_5_moe",
"moe_group_experts": false,
"moe_intermediate_size": 2048,
"moe_latent_size": 2048,
"moe_layer_freq": 1,
"moe_shared_expert_gate": false,
"moe_split_feature_routing": true,
"moe_token_dispatcher_type": "deepep",
"mqa_split_kv_b_proj": true,
"mtp_shared_last_layer": true,
"multimax": "lm_head",
"n_group": 1,
"n_routed_experts": 512,
"n_shared_experts": 1,
"norm_topk_prob": true,
"num_attention_heads": 64,
"num_experts_per_tok": 10,
"num_hidden_layers": 43,
"num_key_value_heads": 1,
"num_nextn_predict_layers": 1,
"o_groups": 8,
"o_lora_rank": 1024,
"pad_token_id": 0,
"q_lora_rank": 1024,
"qk_nope_head_dim": 448,
"qk_norm_type": "per_head",
"qk_pos_emb_head_dim": 64,
"qk_rope_head_dim": 64,
"rms_norm_eps": 1e-06,
"rope_interleave": true,
"rope_scaling": {
"beta_fast": 32,
"beta_slow": 1,
"factor": 16,
"original_max_position_embeddings": 65536,
"type": "yarn"
},
"rope_theta": 8000000,
"rope_type": "rope",
"rotary_base": 8000000,
"rotary_percent": 1.0,
"routed_scaling_factor": 3.0,
"routed_scaling_factor_learnable": true,
"scoring_func": "sigmoid",
"sliding_window": 128,
"tie_word_embeddings": false,
"topk_group": 1,
"topk_method": "noaux_tc",
"torch_dtype": "bfloat16",
"transformers_version": "4.46.1",
"use_attn_merger": true,
"use_bias": false,
"use_cache": true,
"use_dense_mtp": false,
"use_fast_hadamard": true,
"use_gated_attn": true,
"use_mla": false,
"use_qk_norm": true,
"use_vha_attention": true,
"v_head_dim": 512,
"vocab_size": 201216,
"quantization_config": {
"activation_scheme": "dynamic",
"fmt": "e4m3",
"quant_method": "fp8",
"scale_fmt": "ue8m0",
"weight_block_size": [
128,
128
]
}
}