config
Browse files- config.json +169 -0
config.json
ADDED
@@ -0,0 +1,169 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": ["Gemma3nForConditionalGeneration"],
|
3 |
+
"audio_config": {
|
4 |
+
"conf_attention_chunk_size": 12,
|
5 |
+
"conf_attention_context_left": 13,
|
6 |
+
"conf_attention_context_right": 0,
|
7 |
+
"conf_attention_logit_cap": 50.0,
|
8 |
+
"conf_conv_kernel_size": 5,
|
9 |
+
"conf_num_attention_heads": 8,
|
10 |
+
"conf_num_hidden_layers": 12,
|
11 |
+
"conf_reduction_factor": 4,
|
12 |
+
"conf_residual_weight": 0.5,
|
13 |
+
"gradient_clipping": 10000000000.0,
|
14 |
+
"hidden_size": 1536,
|
15 |
+
"input_feat_size": 128,
|
16 |
+
"model_type": "gemma3n_audio",
|
17 |
+
"rms_norm_eps": 1e-6,
|
18 |
+
"sscp_conv_channel_size": [128, 32],
|
19 |
+
"sscp_conv_group_norm_eps": 0.001,
|
20 |
+
"sscp_conv_kernel_size": [
|
21 |
+
[3, 3],
|
22 |
+
[3, 3]
|
23 |
+
],
|
24 |
+
"sscp_conv_stride_size": [
|
25 |
+
[2, 2],
|
26 |
+
[2, 2]
|
27 |
+
],
|
28 |
+
"torch_dtype": "bfloat16",
|
29 |
+
"vocab_offset": 262272,
|
30 |
+
"vocab_size": 128
|
31 |
+
},
|
32 |
+
"audio_soft_tokens_per_image": 188,
|
33 |
+
"audio_token_id": 262273,
|
34 |
+
"boa_token_id": 256000,
|
35 |
+
"boi_token_id": 255999,
|
36 |
+
"bos_token_id": 2,
|
37 |
+
"eoa_token_id": 262272,
|
38 |
+
"eoi_token_id": 262144,
|
39 |
+
"eos_token_id": 106,
|
40 |
+
"image_token_id": 262145,
|
41 |
+
"initializer_range": 0.02,
|
42 |
+
"model_type": "gemma3n",
|
43 |
+
"hidden_size": 2048,
|
44 |
+
"num_attention_heads": 8,
|
45 |
+
"pad_token_id": 0,
|
46 |
+
"quantization_config": {
|
47 |
+
"_load_in_4bit": true,
|
48 |
+
"_load_in_8bit": false,
|
49 |
+
"bnb_4bit_compute_dtype": "bfloat16",
|
50 |
+
"bnb_4bit_quant_storage": "uint8",
|
51 |
+
"bnb_4bit_quant_type": "nf4",
|
52 |
+
"bnb_4bit_use_double_quant": true,
|
53 |
+
"llm_int8_enable_fp32_cpu_offload": false,
|
54 |
+
"llm_int8_has_fp16_weight": false,
|
55 |
+
"llm_int8_skip_modules": [
|
56 |
+
"embed_tokens",
|
57 |
+
"embedding",
|
58 |
+
"lm_head",
|
59 |
+
"multi_modal_projector",
|
60 |
+
"merger",
|
61 |
+
"modality_projection",
|
62 |
+
"correction_coefs",
|
63 |
+
"prediction_coefs",
|
64 |
+
"modality_router",
|
65 |
+
"embedding_projection",
|
66 |
+
"input_proj_linear",
|
67 |
+
"per_layer_projection",
|
68 |
+
"audio_tower",
|
69 |
+
"embed_vision",
|
70 |
+
"embed_audio",
|
71 |
+
"altup_projections",
|
72 |
+
"altup_unembed_projections"
|
73 |
+
],
|
74 |
+
"llm_int8_threshold": 6.0,
|
75 |
+
"load_in_4bit": true,
|
76 |
+
"load_in_8bit": false,
|
77 |
+
"quant_method": "bitsandbytes"
|
78 |
+
},
|
79 |
+
"text_config": {
|
80 |
+
"activation_sparsity_pattern": [
|
81 |
+
0.95, 0.95, 0.95, 0.95, 0.95, 0.95, 0.95, 0.95, 0.95, 0.95, 0.0, 0.0, 0.0,
|
82 |
+
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
|
83 |
+
0.0, 0.0
|
84 |
+
],
|
85 |
+
"altup_active_idx": 0,
|
86 |
+
"altup_coef_clip": 120.0,
|
87 |
+
"altup_correct_scale": true,
|
88 |
+
"altup_num_inputs": 4,
|
89 |
+
"attention_bias": false,
|
90 |
+
"attention_dropout": 0.0,
|
91 |
+
"final_logit_softcapping": 30.0,
|
92 |
+
"head_dim": 256,
|
93 |
+
"hidden_activation": "gelu_pytorch_tanh",
|
94 |
+
"hidden_size": 2048,
|
95 |
+
"hidden_size_per_layer_input": 256,
|
96 |
+
"initializer_range": 0.02,
|
97 |
+
"intermediate_size": [
|
98 |
+
8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192,
|
99 |
+
8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192,
|
100 |
+
8192, 8192, 8192, 8192, 8192, 8192
|
101 |
+
],
|
102 |
+
"laurel_rank": 64,
|
103 |
+
"layer_types": [
|
104 |
+
"sliding_attention",
|
105 |
+
"sliding_attention",
|
106 |
+
"sliding_attention",
|
107 |
+
"sliding_attention",
|
108 |
+
"full_attention",
|
109 |
+
"sliding_attention",
|
110 |
+
"sliding_attention",
|
111 |
+
"sliding_attention",
|
112 |
+
"sliding_attention",
|
113 |
+
"full_attention",
|
114 |
+
"sliding_attention",
|
115 |
+
"sliding_attention",
|
116 |
+
"sliding_attention",
|
117 |
+
"sliding_attention",
|
118 |
+
"full_attention",
|
119 |
+
"sliding_attention",
|
120 |
+
"sliding_attention",
|
121 |
+
"sliding_attention",
|
122 |
+
"sliding_attention",
|
123 |
+
"full_attention",
|
124 |
+
"sliding_attention",
|
125 |
+
"sliding_attention",
|
126 |
+
"sliding_attention",
|
127 |
+
"sliding_attention",
|
128 |
+
"full_attention",
|
129 |
+
"sliding_attention",
|
130 |
+
"sliding_attention",
|
131 |
+
"sliding_attention",
|
132 |
+
"sliding_attention",
|
133 |
+
"full_attention"
|
134 |
+
],
|
135 |
+
"max_position_embeddings": 32768,
|
136 |
+
"model_type": "gemma3n_text",
|
137 |
+
"num_attention_heads": 8,
|
138 |
+
"num_hidden_layers": 30,
|
139 |
+
"num_key_value_heads": 2,
|
140 |
+
"num_kv_shared_layers": 10,
|
141 |
+
"rms_norm_eps": 1e-6,
|
142 |
+
"rope_local_base_freq": 10000.0,
|
143 |
+
"rope_scaling": null,
|
144 |
+
"rope_theta": 1000000.0,
|
145 |
+
"sliding_window": 512,
|
146 |
+
"torch_dtype": "bfloat16",
|
147 |
+
"use_cache": true,
|
148 |
+
"vocab_size": 262400,
|
149 |
+
"vocab_size_per_layer_input": 262144
|
150 |
+
},
|
151 |
+
"torch_dtype": "bfloat16",
|
152 |
+
"transformers_version": "4.53.1",
|
153 |
+
"unsloth_fixed": true,
|
154 |
+
"vision_config": {
|
155 |
+
"architecture": "mobilenetv5_300m_enc",
|
156 |
+
"do_pooling": false,
|
157 |
+
"hidden_size": 2048,
|
158 |
+
"initializer_range": 0.02,
|
159 |
+
"label_names": ["LABEL_0", "LABEL_1"],
|
160 |
+
"model_args": null,
|
161 |
+
"model_type": "gemma3n_vision",
|
162 |
+
"num_classes": 2,
|
163 |
+
"rms_norm_eps": 1e-6,
|
164 |
+
"torch_dtype": "bfloat16",
|
165 |
+
"vocab_offset": 262144,
|
166 |
+
"vocab_size": 128
|
167 |
+
},
|
168 |
+
"vision_soft_tokens_per_image": 256
|
169 |
+
}
|