{
"model_id": "DeepSeek-VL2",
"text_decoder": {
"model_id": "deepseek",
"num_layers": 30,
"pipeline_num_layers": [13, 17],
"hidden_size": 2560,
"num_attention_heads": 32,
"num_query_groups": 32,
"ffn_hidden_size": 12288,
"moe_intermediate_size": 1536,
"kv_channels": 128,
"hidden_dropout": 0.0,
"attention_dropout": 0.0,
"layernorm_epsilon": 1e-06,
"normalization": "RMSNorm",
"qk_layernorm": true,
"add_bias_linear": false,
"add_qkv_bias": false,
"bias_activation_fusion": false,
"gated_linear_unit": true,
"init_method_std": 0.01,
"attention_softmax_in_fp32": true,
"bias_dropout_fusion":false,
"apply_rope_fusion": true,
"max_position_embeddings": 4096,
"fp16": false,
"bf16": true,
"params_dtype": "bf16",
"fp16_lm_cross_entropy": false,
"rotary_percent": 1.0,
"position_embedding_type": "rope",
"shape_order": "SBH",
"parallel_output": true,
"initializer_factor": 0.1,
"persist_layer_norm": true,
"activation_func": "silu",
"vocab_size": 129280,
"rotary_base": 10000,
"first_k_dense_replace": 1,
"moe_layer_freq": 1,
"num_moe_experts": 72,
"n_shared_experts": 2,
"n_group": 1,
"topk_group": 1,
"norm_topk_prob": true,
"moe_router_topk": 6,
"moe_router_load_balancing_type": "noaux_tc",
"moe_router_score_function": "sigmoid",
"moe_router_enable_expert_bias": true,
"moe_router_pre_softmax": true,
"routed_scaling_factor": 2.0,
"router_gating_in_fp32": true,
"mla_mm_split": false,
"mla_fa_without_pad": false,
"padded_base_length": 128,
"mla_up_proj_tp_overlap": false,
"recompute_mla_up_proj": false,
"mla_zero_memory": true,
"schedules_method": "None",
"recompute_granularity": "full",
"recompute_method": "uniform",
"recompute_num_layers": 1
},
"image_encoder": {
"tile_tag": "2D",
"global_view_pos": "head",
"vision_encoder": {
"model_id": "SigLip",
"num_layers": 27,
"pipeline_num_layers": [27, 0],
"params_dtype": "bf16",
"model_name": "siglip_so400m_patch14_384",
"image_size": 384,
"select_layer": -1,
"freeze": false
},
"vision_projector": {
"model_id": "DeepSeekVL2MLP",
"projector_type": "downsample_mlp_gelu",
"input_dim": 1152,
"n_embed": 2560,
"depth": 2,
"mlp_ratio": 1,
"downsample_ratio": 2,
"freeze": false
}
},
"text_encoder": null,
"video_encoder": null
}