Hugging Face's logo

mcrovero
/
gemma-3-4b-it-unsloth-bnb-4bit-custom-merged

Image-Text-to-Text
Transformers
English
gemma3
text-generation-inference
unsloth
conversational
4-bit precision
bitsandbytes
Model card Files Community
mcrovero commited on
Commit
f9a2d67
·
verified ·
1 Parent(s): cee44a9

Update config.json

Browse files
Files changed (1) hide show
  1. config.json +136 -30
config.json CHANGED
@@ -1,37 +1,143 @@
1
  {
2
  "architectures": [
3
- "Gemma3ForCausalLM"
4
  ],
5
- "attention_bias": false,
6
- "attention_dropout": 0.0,
7
- "attn_logit_softcapping": null,
8
  "bos_token_id": 2,
9
- "cache_implementation": "hybrid",
10
- "eos_token_id": 1,
11
- "final_logit_softcapping": null,
12
- "head_dim": 256,
13
- "hidden_activation": "gelu_pytorch_tanh",
14
- "hidden_size": 2560,
15
  "initializer_range": 0.02,
16
- "intermediate_size": 10240,
17
- "max_position_embeddings": 131072,
18
- "model_type": "gemma3_text",
19
- "num_attention_heads": 8,
20
- "num_hidden_layers": 34,
21
- "num_key_value_heads": 4,
22
  "pad_token_id": 0,
23
- "query_pre_attn_scalar": 256,
24
- "rms_norm_eps": 1e-06,
25
- "rope_local_base_freq": 10000.0,
26
- "rope_scaling": {
27
- "factor": 8.0,
28
- "rope_type": "linear"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29
  },
30
- "rope_theta": 1000000.0,
31
- "sliding_window": 1024,
32
- "sliding_window_pattern": 6,
33
- "torch_dtype": "float16",
34
- "transformers_version": "4.50.0.dev0",
35
- "use_cache": true,
36
- "vocab_size": 262208
37
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  {
2
  "architectures": [
3
+ "Gemma3ForConditionalGeneration"
4
  ],
5
+ "boi_token_index": 255999,
 
 
6
  "bos_token_id": 2,
7
+ "eoi_token_index": 256000,
8
+ "eos_token_id": 106,
9
+ "image_token_index": 262144,
 
 
 
10
  "initializer_range": 0.02,
11
+ "mm_tokens_per_image": 256,
12
+ "model_type": "gemma3",
 
 
 
 
13
  "pad_token_id": 0,
14
+ "quantization_config": {
15
+ "_load_in_4bit": true,
16
+ "_load_in_8bit": false,
17
+ "bnb_4bit_compute_dtype": "bfloat16",
18
+ "bnb_4bit_quant_storage": "uint8",
19
+ "bnb_4bit_quant_type": "nf4",
20
+ "bnb_4bit_use_double_quant": true,
21
+ "llm_int8_enable_fp32_cpu_offload": false,
22
+ "llm_int8_has_fp16_weight": false,
23
+ "llm_int8_skip_modules": [
24
+ "lm_head",
25
+ "multi_modal_projector",
26
+ "merger",
27
+ "modality_projection",
28
+ "language_model.model.layers.5.mlp",
29
+ "language_model.model.layers.6.mlp",
30
+ "vision_tower.vision_model.encoder.layers.23.self_attn",
31
+ "language_model.model.layers.4.mlp",
32
+ "language_model.model.layers.1.mlp",
33
+ "vision_tower.vision_model.encoder.layers.24.self_attn",
34
+ "vision_tower.vision_model.encoder.layers.22.self_attn",
35
+ "vision_tower.vision_model.encoder.layers.21.self_attn",
36
+ "vision_tower.vision_model.encoder.layers.19.self_attn",
37
+ "vision_tower.vision_model.encoder.layers.25.self_attn",
38
+ "vision_tower.vision_model.encoder.layers.26.self_attn",
39
+ "language_model.model.layers.3.mlp",
40
+ "vision_tower.vision_model.encoder.layers.25.mlp",
41
+ "vision_tower.vision_model.encoder.layers.19.mlp",
42
+ "vision_tower.vision_model.encoder.layers.20.self_attn",
43
+ "vision_tower.vision_model.encoder.layers.17.mlp",
44
+ "vision_tower.vision_model.encoder.layers.22.mlp",
45
+ "vision_tower.vision_model.encoder.layers.18.self_attn",
46
+ "language_model.model.layers.0.mlp",
47
+ "vision_tower.vision_model.encoder.layers.20.mlp",
48
+ "vision_tower.vision_model.encoder.layers.21.mlp",
49
+ "language_model.model.layers.2.self_attn",
50
+ "vision_tower.vision_model.encoder.layers.15.mlp",
51
+ "vision_tower.vision_model.encoder.layers.17.self_attn",
52
+ "vision_tower.vision_model.encoder.layers.24.mlp",
53
+ "vision_tower.vision_model.encoder.layers.16.mlp",
54
+ "vision_tower.vision_model.encoder.layers.16.self_attn",
55
+ "vision_tower.vision_model.encoder.layers.18.mlp",
56
+ "vision_tower.vision_model.encoder.layers.13.mlp",
57
+ "vision_tower.vision_model.encoder.layers.23.mlp",
58
+ "vision_tower.vision_model.encoder.layers.10.mlp",
59
+ "vision_tower.vision_model.encoder.layers.14.mlp",
60
+ "vision_tower.vision_model.encoder.layers.15.self_attn",
61
+ "vision_tower.vision_model.encoder.layers.11.self_attn",
62
+ "vision_tower.vision_model.encoder.layers.14.self_attn",
63
+ "vision_tower.vision_model.encoder.layers.7.mlp",
64
+ "vision_tower.vision_model.encoder.layers.5.mlp",
65
+ "vision_tower.vision_model.encoder.layers.9.mlp",
66
+ "vision_tower.vision_model.encoder.layers.11.mlp",
67
+ "vision_tower.vision_model.encoder.layers.12.mlp",
68
+ "vision_tower.vision_model.encoder.layers.10.self_attn",
69
+ "vision_tower.vision_model.encoder.layers.13.self_attn",
70
+ "vision_tower.vision_model.encoder.layers.8.mlp",
71
+ "vision_tower.vision_model.encoder.layers.12.self_attn",
72
+ "vision_tower.vision_model.encoder.layers.8.self_attn",
73
+ "vision_tower.vision_model.encoder.layers.4.mlp",
74
+ "vision_tower.vision_model.encoder.layers.4.self_attn",
75
+ "vision_tower.vision_model.encoder.layers.9.self_attn",
76
+ "vision_tower.vision_model.encoder.layers.2.mlp",
77
+ "vision_tower.vision_model.encoder.layers.7.self_attn",
78
+ "vision_tower.vision_model.encoder.layers.3.mlp",
79
+ "vision_tower.vision_model.encoder.layers.6.mlp",
80
+ "vision_tower.vision_model.encoder.layers.1.mlp",
81
+ "vision_tower.vision_model.encoder.layers.6.self_attn",
82
+ "vision_tower.vision_model.encoder.layers.5.self_attn",
83
+ "vision_tower.vision_model.encoder.layers.1.self_attn",
84
+ "vision_tower.vision_model.encoder.layers.3.self_attn",
85
+ "vision_tower.vision_model.encoder.layers.0.mlp",
86
+ "vision_tower.vision_model.encoder.layers.0.self_attn",
87
+ "vision_tower.vision_model.encoder.layers.2.self_attn",
88
+ "vision_tower.vision_model.encoder.layers.26.mlp"
89
+ ],
90
+ "llm_int8_threshold": 6.0,
91
+ "load_in_4bit": true,
92
+ "load_in_8bit": false,
93
+ "quant_method": "bitsandbytes"
94
  },
95
+ "text_config": {
96
+ "attention_bias": false,
97
+ "attention_dropout": 0.0,
98
+ "attn_logit_softcapping": null,
99
+ "cache_implementation": "hybrid",
100
+ "final_logit_softcapping": null,
101
+ "head_dim": 256,
102
+ "hidden_activation": "gelu_pytorch_tanh",
103
+ "hidden_size": 2560,
104
+ "initializer_range": 0.02,
105
+ "intermediate_size": 10240,
106
+ "max_position_embeddings": 131072,
107
+ "model_type": "gemma3_text",
108
+ "num_attention_heads": 8,
109
+ "num_hidden_layers": 34,
110
+ "num_key_value_heads": 4,
111
+ "query_pre_attn_scalar": 256,
112
+ "rms_norm_eps": 1e-06,
113
+ "rope_local_base_freq": 10000.0,
114
+ "rope_scaling": {
115
+ "factor": 8.0,
116
+ "rope_type": "linear"
117
+ },
118
+ "rope_theta": 1000000.0,
119
+ "sliding_window": 1024,
120
+ "sliding_window_pattern": 6,
121
+ "torch_dtype": "bfloat16",
122
+ "use_cache": true,
123
+ "vocab_size": 262208
124
+ },
125
+ "torch_dtype": "bfloat16",
126
+ "transformers_version": "4.52.0.dev0",
127
+ "unsloth_fixed": true,
128
+ "vision_config": {
129
+ "attention_dropout": 0.0,
130
+ "hidden_act": "gelu_pytorch_tanh",
131
+ "hidden_size": 1152,
132
+ "image_size": 896,
133
+ "intermediate_size": 4304,
134
+ "layer_norm_eps": 1e-06,
135
+ "model_type": "siglip_vision_model",
136
+ "num_attention_heads": 16,
137
+ "num_channels": 3,
138
+ "num_hidden_layers": 27,
139
+ "patch_size": 14,
140
+ "torch_dtype": "bfloat16",
141
+ "vision_use_head": false
142
+ }
143
+ }