deepstack/llava_3b_siglip_so/config.json

{
  "_mlp_class": "LLaMAMLP",
  "_name_or_path": "/home/hk-project-p0022189/tum_yvc3016/yingqi/workspace/ckpts/mtgv/MobileLLaMA-2.7B-Chat",
  "_norm_class": "RMSNorm",
  "architectures": [
    "LlavaLlamaForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bias": false,
  "block_size": 2048,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "freeze_mm_mlp_adapter": false,
  "gelu_approximate": "none",
  "hidden_act": "silu",
  "hidden_size": 2560,
  "image_aspect_ratio": "pad",
  "initializer_range": 0.02,
  "intermediate_size": 6912,
  "lm_head_bias": false,
  "max_position_embeddings": 2048,
  "mm_hidden_size": 1152,
  "mm_patch_merge_type": "flat",
  "mm_projector_lr": null,
  "mm_projector_type": "mlp2x_gelu",
  "mm_use_im_patch_token": false,
  "mm_use_im_start_end": false,
  "mm_vision_select_feature": "patch",
  "mm_vision_select_layer": -2,
  "mm_vision_selection_config": {
    "fusion_location": "pre_proj",
    "layer_configs": [
      {
        "llm_layer": 0,
        "proj": [
          0
        ],
        "vit_layer": [
          -1
        ]
      }
    ],
    "llm_fusion_strategy": "direct_sum",
    "projectors_configs": [
      {
        "type": "mlp2x_gelu"
      }
    ],
    "vit_fusion_strategy": "channel_avg_pooling"
  },
  "mm_vision_tower": "/home/hk-project-p0022189/tum_yvc3016/yingqi/workspace/ckpts/google/siglip-so400m-patch14-384",
  "model_type": "llama",
  "n_embd": 2560,
  "n_head": 16,
  "n_layer": 32,
  "n_query_groups": 16,
  "name": "huggy_llama_3b",
  "norm_eps": 1e-06,
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 32,
  "org": "meta",
  "pad_token_id": 0,
  "padded_vocab_size": 32000,
  "padding_multiple": 64,
  "parallel_residual": false,
  "pretraining_tp": 1,
  "rms_norm_eps": 1e-06,
  "rope_base": 10000,
  "rope_condense_ratio": 1,
  "rope_scaling": null,
  "rope_theta": 10000.0,
  "rotary_percentage": 1.0,
  "shared_attention_norm": false,
  "tie_word_embeddings": false,
  "tokenizer_model_max_length": 2048,
  "tokenizer_padding_side": "right",
  "torch_dtype": "bfloat16",
  "transformers_version": "4.39.2",
  "tune_mm_mlp_adapter": false,
  "use_cache": true,
  "use_mm_proj": true,
  "vocab_size": 32000
}