Hugging Face's logo

euhidaman
/
MicroVLM-V

Image-to-Text
Transformers
PyTorch
vision-language
multimodal
episodic-memory
fiber-alignment
qwen2
deit
Model card Files
xet
Community
MicroVLM-V
File size: 3,165 Bytes 
8514404
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
 
{
  "qwen2_5_config": {
    "hidden_size": 896,
    "num_hidden_layers": 24,
    "num_attention_heads": 14,
    "num_key_value_heads": 2,
    "intermediate_size": 4864,
    "vocab_size": 151936,
    "max_position_embeddings": 32768,
    "rope_theta": 1000000.0,
    "rms_norm_eps": 1e-06,
    "tie_word_embeddings": true,
    "use_sliding_window": false,
    "sliding_window": null,
    "max_window_layers": 21,
    "head_dim": 64,
    "kv_head_dim": 128,
    "model_type": "qwen2"
  },
  "deit_tiny_config": {
    "image_size": 224,
    "patch_size": 16,
    "num_channels": 3,
    "hidden_size": 192,
    "num_hidden_layers": 12,
    "num_attention_heads": 3,
    "intermediate_size": 768,
    "hidden_dropout_prob": 0.0,
    "attention_probs_dropout_prob": 0.0,
    "layer_norm_eps": 1e-06,
    "num_patches": 196,
    "encoder_stride": 16
  },
  "quantization_config": {
    "memory_quantization": {
      "bits": 1.58,
      "method": "bitnet_158",
      "description": "1.58-bit quantization for episodic memory and final model"
    },
    "training_quantization": {
      "weight_bits": 4,
      "activation_bits": 4,
      "method": "symmetric",
      "description": "4-bit quantization for weights and activations during training"
    },
    "inference_quantization": {
      "model_bits": 1.58,
      "method": "bitnet_158",
      "description": "1.58-bit quantization for final deployed model"
    }
  },
  "model_dimensions": {
    "qwen_hidden_dim": 896,
    "deit_embed_dim": 192,
    "vision_hidden_size": 192,
    "language_hidden_size": 896,
    "num_patches": 196,
    "k_prefix": 25,
    "adapter_projection_dim": 896,
    "alignment_dim": 128,
    "memory_size": 64,
    "memory_dim": 896,
    "memory_target_layers": 6,
    "memory_num_heads": 4,
    "memory_num_layers": 6,
    "head_dim": 64,
    "scope_hidden_dim": 256,
    "itm_hidden_dim": 256,
    "fusion_layers": [
      9,
      11
    ],
    "num_fusion_heads": 2,
    "fusion_dim": 384,
    "w_logvar_setting": 1,
    "deterministic_memory": false
  },
  "itc_config": {
    "use_itc_queue": true,
    "itc_queue_size": 256,
    "itc_embed_dim": 128,
    "description": "FIBER-style ITC queue for better negative sampling (lightweight version)"
  },
  "estimated_sizes": {
    "qwen_original_mb": 1976,
    "deit_original_mb": 22.8,
    "qwen_4bit_mb": 247.0,
    "deit_4bit_mb": 2.85,
    "qwen_158bit_mb": 98.80000000000001,
    "adapter_mb": 0.777728,
    "memory_original_mb": 0.229376,
    "memory_158bit_mb": 0.011468800000000001,
    "wm_projection_mb": 11.010048,
    "scopenet_mb": 1.180672,
    "fusion_mb": 0.009216,
    "alignment_mb": 0.557056,
    "itm_head_mb": 1.11616,
    "overhead_total_mb": 14.880256,
    "total_fp32_mb": 2013.6802559999999,
    "total_4bit_qwen_mb": 273.28025599999995,
    "total_fully_quantized_mb": 116.53025600000001
  },
  "metadata": {
    "qwen_model": "Qwen/Qwen2.5-0.5B",
    "deit_model": "facebook/deit-tiny-patch16-224",
    "target_total_size_mb": 500,
    "description": "MicroVLM-V compact model configuration with 4-bit quantization for <1GB deployment",
    "quantization_default": "4-bit Qwen, FP16 DeiT, FP32 overhead"
  }
}