Hugging Face's logo

RedHatAI
/
gemma-3n-E4B-it-FP8-dynamic

Image-Text-to-Text
Transformers
Safetensors
gemma3n
vllm
vision
audio
fp8
conversational
compressed-tensors
Model card Files
xet
Community
Shubhra Pandit commited on
Commit
f0e9963
·
1 Parent(s): 0369f14

Upload model files

Browse files
Files changed (16) hide show
  1. .gitattributes +2 -0
  2. chat_template.jinja +49 -0
  3. config.json +652 -0
  4. generation_config.json +13 -0
  5. model-00001-of-00004.safetensors +3 -0
  6. model-00002-of-00004.safetensors +3 -0
  7. model-00003-of-00004.safetensors +3 -0
  8. model-00004-of-00004.safetensors +3 -0
  9. model.safetensors.index.json +0 -0
  10. preprocessor_config.json +51 -0
  11. processor_config.json +5 -0
  12. recipe.yaml +8 -0
  13. special_tokens_map.json +36 -0
  14. tokenizer.json +3 -0
  15. tokenizer.model +3 -0
  16. tokenizer_config.json +0 -0
.gitattributes CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
37
+
chat_template.jinja ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {{ bos_token }}
2
+ {%- if messages[0]['role'] == 'system' -%}
3
+ {%- if messages[0]['content'] is string -%}
4
+ {%- set first_user_prefix = messages[0]['content'] + '
5
+
6
+ ' -%}
7
+ {%- else -%}
8
+ {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
9
+
10
+ ' -%}
11
+ {%- endif -%}
12
+ {%- set loop_messages = messages[1:] -%}
13
+ {%- else -%}
14
+ {%- set first_user_prefix = "" -%}
15
+ {%- set loop_messages = messages -%}
16
+ {%- endif -%}
17
+ {%- for message in loop_messages -%}
18
+ {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
19
+ {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
20
+ {%- endif -%}
21
+ {%- if (message['role'] == 'assistant') -%}
22
+ {%- set role = "model" -%}
23
+ {%- else -%}
24
+ {%- set role = message['role'] -%}
25
+ {%- endif -%}
26
+ {{ '<start_of_turn>' + role + '
27
+ ' + (first_user_prefix if loop.first else "") }}
28
+ {%- if message['content'] is string -%}
29
+ {{ message['content'] | trim }}
30
+ {%- elif message['content'] is iterable -%}
31
+ {%- for item in message['content'] -%}
32
+ {%- if item['type'] == 'audio' -%}
33
+ {{ '<audio_soft_token>' }}
34
+ {%- elif item['type'] == 'image' -%}
35
+ {{ '<image_soft_token>' }}
36
+ {%- elif item['type'] == 'text' -%}
37
+ {{ item['text'] | trim }}
38
+ {%- endif -%}
39
+ {%- endfor -%}
40
+ {%- else -%}
41
+ {{ raise_exception("Invalid content type") }}
42
+ {%- endif -%}
43
+ {{ '<end_of_turn>
44
+ ' }}
45
+ {%- endfor -%}
46
+ {%- if add_generation_prompt -%}
47
+ {{'<start_of_turn>model
48
+ '}}
49
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,652 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma3nForConditionalGeneration"
4
+ ],
5
+ "audio_config": {
6
+ "conf_attention_chunk_size": 12,
7
+ "conf_attention_context_left": 13,
8
+ "conf_attention_context_right": 0,
9
+ "conf_attention_logit_cap": 50.0,
10
+ "conf_conv_kernel_size": 5,
11
+ "conf_num_attention_heads": 8,
12
+ "conf_num_hidden_layers": 12,
13
+ "conf_reduction_factor": 4,
14
+ "conf_residual_weight": 0.5,
15
+ "gradient_clipping": 10000000000.0,
16
+ "hidden_size": 1536,
17
+ "input_feat_size": 128,
18
+ "model_type": "gemma3n_audio",
19
+ "rms_norm_eps": 1e-06,
20
+ "sscp_conv_channel_size": [
21
+ 128,
22
+ 32
23
+ ],
24
+ "sscp_conv_group_norm_eps": 0.001,
25
+ "sscp_conv_kernel_size": [
26
+ [
27
+ 3,
28
+ 3
29
+ ],
30
+ [
31
+ 3,
32
+ 3
33
+ ]
34
+ ],
35
+ "sscp_conv_stride_size": [
36
+ [
37
+ 2,
38
+ 2
39
+ ],
40
+ [
41
+ 2,
42
+ 2
43
+ ]
44
+ ],
45
+ "torch_dtype": "bfloat16",
46
+ "vocab_offset": 262272,
47
+ "vocab_size": 128
48
+ },
49
+ "audio_soft_tokens_per_image": 188,
50
+ "audio_token_id": 262273,
51
+ "boa_token_id": 256000,
52
+ "boi_token_id": 255999,
53
+ "eoa_token_id": 262272,
54
+ "eoi_token_id": 262144,
55
+ "eos_token_id": [
56
+ 1,
57
+ 106
58
+ ],
59
+ "image_token_id": 262145,
60
+ "initializer_range": 0.02,
61
+ "model_type": "gemma3n",
62
+ "quantization_config": {
63
+ "config_groups": {
64
+ "group_0": {
65
+ "input_activations": {
66
+ "actorder": null,
67
+ "block_structure": null,
68
+ "dynamic": true,
69
+ "group_size": null,
70
+ "num_bits": 8,
71
+ "observer": null,
72
+ "observer_kwargs": {},
73
+ "strategy": "token",
74
+ "symmetric": true,
75
+ "type": "float"
76
+ },
77
+ "output_activations": null,
78
+ "targets": [
79
+ "Linear"
80
+ ],
81
+ "weights": {
82
+ "actorder": null,
83
+ "block_structure": null,
84
+ "dynamic": false,
85
+ "group_size": null,
86
+ "num_bits": 8,
87
+ "observer": "minmax",
88
+ "observer_kwargs": {},
89
+ "strategy": "channel",
90
+ "symmetric": true,
91
+ "type": "float"
92
+ }
93
+ }
94
+ },
95
+ "format": "float-quantized",
96
+ "global_compression_ratio": null,
97
+ "ignore": [
98
+ "model.language_model.layers.0.altup.correction_coefs",
99
+ "model.language_model.layers.0.altup.prediction_coefs",
100
+ "model.language_model.layers.0.altup.modality_router",
101
+ "model.language_model.layers.0.laurel.linear_left",
102
+ "model.language_model.layers.0.laurel.linear_right",
103
+ "model.language_model.layers.0.per_layer_input_gate",
104
+ "model.language_model.layers.0.per_layer_projection",
105
+ "model.language_model.layers.1.altup.correction_coefs",
106
+ "model.language_model.layers.1.altup.prediction_coefs",
107
+ "model.language_model.layers.1.altup.modality_router",
108
+ "model.language_model.layers.1.laurel.linear_left",
109
+ "model.language_model.layers.1.laurel.linear_right",
110
+ "model.language_model.layers.1.per_layer_input_gate",
111
+ "model.language_model.layers.1.per_layer_projection",
112
+ "model.language_model.layers.2.altup.correction_coefs",
113
+ "model.language_model.layers.2.altup.prediction_coefs",
114
+ "model.language_model.layers.2.altup.modality_router",
115
+ "model.language_model.layers.2.laurel.linear_left",
116
+ "model.language_model.layers.2.laurel.linear_right",
117
+ "model.language_model.layers.2.per_layer_input_gate",
118
+ "model.language_model.layers.2.per_layer_projection",
119
+ "model.language_model.layers.3.altup.correction_coefs",
120
+ "model.language_model.layers.3.altup.prediction_coefs",
121
+ "model.language_model.layers.3.altup.modality_router",
122
+ "model.language_model.layers.3.laurel.linear_left",
123
+ "model.language_model.layers.3.laurel.linear_right",
124
+ "model.language_model.layers.3.per_layer_input_gate",
125
+ "model.language_model.layers.3.per_layer_projection",
126
+ "model.language_model.layers.4.altup.correction_coefs",
127
+ "model.language_model.layers.4.altup.prediction_coefs",
128
+ "model.language_model.layers.4.altup.modality_router",
129
+ "model.language_model.layers.4.laurel.linear_left",
130
+ "model.language_model.layers.4.laurel.linear_right",
131
+ "model.language_model.layers.4.per_layer_input_gate",
132
+ "model.language_model.layers.4.per_layer_projection",
133
+ "model.language_model.layers.5.altup.correction_coefs",
134
+ "model.language_model.layers.5.altup.prediction_coefs",
135
+ "model.language_model.layers.5.altup.modality_router",
136
+ "model.language_model.layers.5.laurel.linear_left",
137
+ "model.language_model.layers.5.laurel.linear_right",
138
+ "model.language_model.layers.5.per_layer_input_gate",
139
+ "model.language_model.layers.5.per_layer_projection",
140
+ "model.language_model.layers.6.altup.correction_coefs",
141
+ "model.language_model.layers.6.altup.prediction_coefs",
142
+ "model.language_model.layers.6.altup.modality_router",
143
+ "model.language_model.layers.6.laurel.linear_left",
144
+ "model.language_model.layers.6.laurel.linear_right",
145
+ "model.language_model.layers.6.per_layer_input_gate",
146
+ "model.language_model.layers.6.per_layer_projection",
147
+ "model.language_model.layers.7.altup.correction_coefs",
148
+ "model.language_model.layers.7.altup.prediction_coefs",
149
+ "model.language_model.layers.7.altup.modality_router",
150
+ "model.language_model.layers.7.laurel.linear_left",
151
+ "model.language_model.layers.7.laurel.linear_right",
152
+ "model.language_model.layers.7.per_layer_input_gate",
153
+ "model.language_model.layers.7.per_layer_projection",
154
+ "model.language_model.layers.8.altup.correction_coefs",
155
+ "model.language_model.layers.8.altup.prediction_coefs",
156
+ "model.language_model.layers.8.altup.modality_router",
157
+ "model.language_model.layers.8.laurel.linear_left",
158
+ "model.language_model.layers.8.laurel.linear_right",
159
+ "model.language_model.layers.8.per_layer_input_gate",
160
+ "model.language_model.layers.8.per_layer_projection",
161
+ "model.language_model.layers.9.altup.correction_coefs",
162
+ "model.language_model.layers.9.altup.prediction_coefs",
163
+ "model.language_model.layers.9.altup.modality_router",
164
+ "model.language_model.layers.9.laurel.linear_left",
165
+ "model.language_model.layers.9.laurel.linear_right",
166
+ "model.language_model.layers.9.per_layer_input_gate",
167
+ "model.language_model.layers.9.per_layer_projection",
168
+ "model.language_model.layers.10.altup.correction_coefs",
169
+ "model.language_model.layers.10.altup.prediction_coefs",
170
+ "model.language_model.layers.10.altup.modality_router",
171
+ "model.language_model.layers.10.laurel.linear_left",
172
+ "model.language_model.layers.10.laurel.linear_right",
173
+ "model.language_model.layers.10.per_layer_input_gate",
174
+ "model.language_model.layers.10.per_layer_projection",
175
+ "model.language_model.layers.11.altup.correction_coefs",
176
+ "model.language_model.layers.11.altup.prediction_coefs",
177
+ "model.language_model.layers.11.altup.modality_router",
178
+ "model.language_model.layers.11.laurel.linear_left",
179
+ "model.language_model.layers.11.laurel.linear_right",
180
+ "model.language_model.layers.11.per_layer_input_gate",
181
+ "model.language_model.layers.11.per_layer_projection",
182
+ "model.language_model.layers.12.altup.correction_coefs",
183
+ "model.language_model.layers.12.altup.prediction_coefs",
184
+ "model.language_model.layers.12.altup.modality_router",
185
+ "model.language_model.layers.12.laurel.linear_left",
186
+ "model.language_model.layers.12.laurel.linear_right",
187
+ "model.language_model.layers.12.per_layer_input_gate",
188
+ "model.language_model.layers.12.per_layer_projection",
189
+ "model.language_model.layers.13.altup.correction_coefs",
190
+ "model.language_model.layers.13.altup.prediction_coefs",
191
+ "model.language_model.layers.13.altup.modality_router",
192
+ "model.language_model.layers.13.laurel.linear_left",
193
+ "model.language_model.layers.13.laurel.linear_right",
194
+ "model.language_model.layers.13.per_layer_input_gate",
195
+ "model.language_model.layers.13.per_layer_projection",
196
+ "model.language_model.layers.14.altup.correction_coefs",
197
+ "model.language_model.layers.14.altup.prediction_coefs",
198
+ "model.language_model.layers.14.altup.modality_router",
199
+ "model.language_model.layers.14.laurel.linear_left",
200
+ "model.language_model.layers.14.laurel.linear_right",
201
+ "model.language_model.layers.14.per_layer_input_gate",
202
+ "model.language_model.layers.14.per_layer_projection",
203
+ "model.language_model.layers.15.altup.correction_coefs",
204
+ "model.language_model.layers.15.altup.prediction_coefs",
205
+ "model.language_model.layers.15.altup.modality_router",
206
+ "model.language_model.layers.15.laurel.linear_left",
207
+ "model.language_model.layers.15.laurel.linear_right",
208
+ "model.language_model.layers.15.per_layer_input_gate",
209
+ "model.language_model.layers.15.per_layer_projection",
210
+ "model.language_model.layers.16.altup.correction_coefs",
211
+ "model.language_model.layers.16.altup.prediction_coefs",
212
+ "model.language_model.layers.16.altup.modality_router",
213
+ "model.language_model.layers.16.laurel.linear_left",
214
+ "model.language_model.layers.16.laurel.linear_right",
215
+ "model.language_model.layers.16.per_layer_input_gate",
216
+ "model.language_model.layers.16.per_layer_projection",
217
+ "model.language_model.layers.17.altup.correction_coefs",
218
+ "model.language_model.layers.17.altup.prediction_coefs",
219
+ "model.language_model.layers.17.altup.modality_router",
220
+ "model.language_model.layers.17.laurel.linear_left",
221
+ "model.language_model.layers.17.laurel.linear_right",
222
+ "model.language_model.layers.17.per_layer_input_gate",
223
+ "model.language_model.layers.17.per_layer_projection",
224
+ "model.language_model.layers.18.altup.correction_coefs",
225
+ "model.language_model.layers.18.altup.prediction_coefs",
226
+ "model.language_model.layers.18.altup.modality_router",
227
+ "model.language_model.layers.18.laurel.linear_left",
228
+ "model.language_model.layers.18.laurel.linear_right",
229
+ "model.language_model.layers.18.per_layer_input_gate",
230
+ "model.language_model.layers.18.per_layer_projection",
231
+ "model.language_model.layers.19.altup.correction_coefs",
232
+ "model.language_model.layers.19.altup.prediction_coefs",
233
+ "model.language_model.layers.19.altup.modality_router",
234
+ "model.language_model.layers.19.laurel.linear_left",
235
+ "model.language_model.layers.19.laurel.linear_right",
236
+ "model.language_model.layers.19.per_layer_input_gate",
237
+ "model.language_model.layers.19.per_layer_projection",
238
+ "model.language_model.layers.20.altup.correction_coefs",
239
+ "model.language_model.layers.20.altup.prediction_coefs",
240
+ "model.language_model.layers.20.altup.modality_router",
241
+ "model.language_model.layers.20.laurel.linear_left",
242
+ "model.language_model.layers.20.laurel.linear_right",
243
+ "model.language_model.layers.20.per_layer_input_gate",
244
+ "model.language_model.layers.20.per_layer_projection",
245
+ "model.language_model.layers.21.altup.correction_coefs",
246
+ "model.language_model.layers.21.altup.prediction_coefs",
247
+ "model.language_model.layers.21.altup.modality_router",
248
+ "model.language_model.layers.21.laurel.linear_left",
249
+ "model.language_model.layers.21.laurel.linear_right",
250
+ "model.language_model.layers.21.per_layer_input_gate",
251
+ "model.language_model.layers.21.per_layer_projection",
252
+ "model.language_model.layers.22.altup.correction_coefs",
253
+ "model.language_model.layers.22.altup.prediction_coefs",
254
+ "model.language_model.layers.22.altup.modality_router",
255
+ "model.language_model.layers.22.laurel.linear_left",
256
+ "model.language_model.layers.22.laurel.linear_right",
257
+ "model.language_model.layers.22.per_layer_input_gate",
258
+ "model.language_model.layers.22.per_layer_projection",
259
+ "model.language_model.layers.23.altup.correction_coefs",
260
+ "model.language_model.layers.23.altup.prediction_coefs",
261
+ "model.language_model.layers.23.altup.modality_router",
262
+ "model.language_model.layers.23.laurel.linear_left",
263
+ "model.language_model.layers.23.laurel.linear_right",
264
+ "model.language_model.layers.23.per_layer_input_gate",
265
+ "model.language_model.layers.23.per_layer_projection",
266
+ "model.language_model.layers.24.altup.correction_coefs",
267
+ "model.language_model.layers.24.altup.prediction_coefs",
268
+ "model.language_model.layers.24.altup.modality_router",
269
+ "model.language_model.layers.24.laurel.linear_left",
270
+ "model.language_model.layers.24.laurel.linear_right",
271
+ "model.language_model.layers.24.per_layer_input_gate",
272
+ "model.language_model.layers.24.per_layer_projection",
273
+ "model.language_model.layers.25.altup.correction_coefs",
274
+ "model.language_model.layers.25.altup.prediction_coefs",
275
+ "model.language_model.layers.25.altup.modality_router",
276
+ "model.language_model.layers.25.laurel.linear_left",
277
+ "model.language_model.layers.25.laurel.linear_right",
278
+ "model.language_model.layers.25.per_layer_input_gate",
279
+ "model.language_model.layers.25.per_layer_projection",
280
+ "model.language_model.layers.26.altup.correction_coefs",
281
+ "model.language_model.layers.26.altup.prediction_coefs",
282
+ "model.language_model.layers.26.altup.modality_router",
283
+ "model.language_model.layers.26.laurel.linear_left",
284
+ "model.language_model.layers.26.laurel.linear_right",
285
+ "model.language_model.layers.26.per_layer_input_gate",
286
+ "model.language_model.layers.26.per_layer_projection",
287
+ "model.language_model.layers.27.altup.correction_coefs",
288
+ "model.language_model.layers.27.altup.prediction_coefs",
289
+ "model.language_model.layers.27.altup.modality_router",
290
+ "model.language_model.layers.27.laurel.linear_left",
291
+ "model.language_model.layers.27.laurel.linear_right",
292
+ "model.language_model.layers.27.per_layer_input_gate",
293
+ "model.language_model.layers.27.per_layer_projection",
294
+ "model.language_model.layers.28.altup.correction_coefs",
295
+ "model.language_model.layers.28.altup.prediction_coefs",
296
+ "model.language_model.layers.28.altup.modality_router",
297
+ "model.language_model.layers.28.laurel.linear_left",
298
+ "model.language_model.layers.28.laurel.linear_right",
299
+ "model.language_model.layers.28.per_layer_input_gate",
300
+ "model.language_model.layers.28.per_layer_projection",
301
+ "model.language_model.layers.29.altup.correction_coefs",
302
+ "model.language_model.layers.29.altup.prediction_coefs",
303
+ "model.language_model.layers.29.altup.modality_router",
304
+ "model.language_model.layers.29.laurel.linear_left",
305
+ "model.language_model.layers.29.laurel.linear_right",
306
+ "model.language_model.layers.29.per_layer_input_gate",
307
+ "model.language_model.layers.29.per_layer_projection",
308
+ "model.language_model.layers.30.altup.correction_coefs",
309
+ "model.language_model.layers.30.altup.prediction_coefs",
310
+ "model.language_model.layers.30.altup.modality_router",
311
+ "model.language_model.layers.30.laurel.linear_left",
312
+ "model.language_model.layers.30.laurel.linear_right",
313
+ "model.language_model.layers.30.per_layer_input_gate",
314
+ "model.language_model.layers.30.per_layer_projection",
315
+ "model.language_model.layers.31.altup.correction_coefs",
316
+ "model.language_model.layers.31.altup.prediction_coefs",
317
+ "model.language_model.layers.31.altup.modality_router",
318
+ "model.language_model.layers.31.laurel.linear_left",
319
+ "model.language_model.layers.31.laurel.linear_right",
320
+ "model.language_model.layers.31.per_layer_input_gate",
321
+ "model.language_model.layers.31.per_layer_projection",
322
+ "model.language_model.layers.32.altup.correction_coefs",
323
+ "model.language_model.layers.32.altup.prediction_coefs",
324
+ "model.language_model.layers.32.altup.modality_router",
325
+ "model.language_model.layers.32.laurel.linear_left",
326
+ "model.language_model.layers.32.laurel.linear_right",
327
+ "model.language_model.layers.32.per_layer_input_gate",
328
+ "model.language_model.layers.32.per_layer_projection",
329
+ "model.language_model.layers.33.altup.correction_coefs",
330
+ "model.language_model.layers.33.altup.prediction_coefs",
331
+ "model.language_model.layers.33.altup.modality_router",
332
+ "model.language_model.layers.33.laurel.linear_left",
333
+ "model.language_model.layers.33.laurel.linear_right",
334
+ "model.language_model.layers.33.per_layer_input_gate",
335
+ "model.language_model.layers.33.per_layer_projection",
336
+ "model.language_model.layers.34.altup.correction_coefs",
337
+ "model.language_model.layers.34.altup.prediction_coefs",
338
+ "model.language_model.layers.34.altup.modality_router",
339
+ "model.language_model.layers.34.laurel.linear_left",
340
+ "model.language_model.layers.34.laurel.linear_right",
341
+ "model.language_model.layers.34.per_layer_input_gate",
342
+ "model.language_model.layers.34.per_layer_projection",
343
+ "model.language_model.per_layer_model_projection",
344
+ "model.language_model.altup_projections.0",
345
+ "model.language_model.altup_projections.1",
346
+ "model.language_model.altup_projections.2",
347
+ "model.language_model.altup_unembed_projections.0",
348
+ "model.language_model.altup_unembed_projections.1",
349
+ "model.language_model.altup_unembed_projections.2",
350
+ "model.audio_tower.subsample_conv_projection.input_proj_linear",
351
+ "model.audio_tower.conformer.0.ffw_layer_start.ffw_layer_1",
352
+ "model.audio_tower.conformer.0.ffw_layer_start.ffw_layer_2",
353
+ "model.audio_tower.conformer.0.attention.attn.relative_position_embedding.pos_proj",
354
+ "model.audio_tower.conformer.0.attention.attn.q_proj",
355
+ "model.audio_tower.conformer.0.attention.attn.k_proj",
356
+ "model.audio_tower.conformer.0.attention.attn.v_proj",
357
+ "model.audio_tower.conformer.0.attention.post",
358
+ "model.audio_tower.conformer.0.lconv1d.linear_start",
359
+ "model.audio_tower.conformer.0.lconv1d.linear_end",
360
+ "model.audio_tower.conformer.0.ffw_layer_end.ffw_layer_1",
361
+ "model.audio_tower.conformer.0.ffw_layer_end.ffw_layer_2",
362
+ "model.audio_tower.conformer.1.ffw_layer_start.ffw_layer_1",
363
+ "model.audio_tower.conformer.1.ffw_layer_start.ffw_layer_2",
364
+ "model.audio_tower.conformer.1.attention.attn.relative_position_embedding.pos_proj",
365
+ "model.audio_tower.conformer.1.attention.attn.q_proj",
366
+ "model.audio_tower.conformer.1.attention.attn.k_proj",
367
+ "model.audio_tower.conformer.1.attention.attn.v_proj",
368
+ "model.audio_tower.conformer.1.attention.post",
369
+ "model.audio_tower.conformer.1.lconv1d.linear_start",
370
+ "model.audio_tower.conformer.1.lconv1d.linear_end",
371
+ "model.audio_tower.conformer.1.ffw_layer_end.ffw_layer_1",
372
+ "model.audio_tower.conformer.1.ffw_layer_end.ffw_layer_2",
373
+ "model.audio_tower.conformer.2.ffw_layer_start.ffw_layer_1",
374
+ "model.audio_tower.conformer.2.ffw_layer_start.ffw_layer_2",
375
+ "model.audio_tower.conformer.2.attention.attn.relative_position_embedding.pos_proj",
376
+ "model.audio_tower.conformer.2.attention.attn.q_proj",
377
+ "model.audio_tower.conformer.2.attention.attn.k_proj",
378
+ "model.audio_tower.conformer.2.attention.attn.v_proj",
379
+ "model.audio_tower.conformer.2.attention.post",
380
+ "model.audio_tower.conformer.2.lconv1d.linear_start",
381
+ "model.audio_tower.conformer.2.lconv1d.linear_end",
382
+ "model.audio_tower.conformer.2.ffw_layer_end.ffw_layer_1",
383
+ "model.audio_tower.conformer.2.ffw_layer_end.ffw_layer_2",
384
+ "model.audio_tower.conformer.3.ffw_layer_start.ffw_layer_1",
385
+ "model.audio_tower.conformer.3.ffw_layer_start.ffw_layer_2",
386
+ "model.audio_tower.conformer.3.attention.attn.relative_position_embedding.pos_proj",
387
+ "model.audio_tower.conformer.3.attention.attn.q_proj",
388
+ "model.audio_tower.conformer.3.attention.attn.k_proj",
389
+ "model.audio_tower.conformer.3.attention.attn.v_proj",
390
+ "model.audio_tower.conformer.3.attention.post",
391
+ "model.audio_tower.conformer.3.lconv1d.linear_start",
392
+ "model.audio_tower.conformer.3.lconv1d.linear_end",
393
+ "model.audio_tower.conformer.3.ffw_layer_end.ffw_layer_1",
394
+ "model.audio_tower.conformer.3.ffw_layer_end.ffw_layer_2",
395
+ "model.audio_tower.conformer.4.ffw_layer_start.ffw_layer_1",
396
+ "model.audio_tower.conformer.4.ffw_layer_start.ffw_layer_2",
397
+ "model.audio_tower.conformer.4.attention.attn.relative_position_embedding.pos_proj",
398
+ "model.audio_tower.conformer.4.attention.attn.q_proj",
399
+ "model.audio_tower.conformer.4.attention.attn.k_proj",
400
+ "model.audio_tower.conformer.4.attention.attn.v_proj",
401
+ "model.audio_tower.conformer.4.attention.post",
402
+ "model.audio_tower.conformer.4.lconv1d.linear_start",
403
+ "model.audio_tower.conformer.4.lconv1d.linear_end",
404
+ "model.audio_tower.conformer.4.ffw_layer_end.ffw_layer_1",
405
+ "model.audio_tower.conformer.4.ffw_layer_end.ffw_layer_2",
406
+ "model.audio_tower.conformer.5.ffw_layer_start.ffw_layer_1",
407
+ "model.audio_tower.conformer.5.ffw_layer_start.ffw_layer_2",
408
+ "model.audio_tower.conformer.5.attention.attn.relative_position_embedding.pos_proj",
409
+ "model.audio_tower.conformer.5.attention.attn.q_proj",
410
+ "model.audio_tower.conformer.5.attention.attn.k_proj",
411
+ "model.audio_tower.conformer.5.attention.attn.v_proj",
412
+ "model.audio_tower.conformer.5.attention.post",
413
+ "model.audio_tower.conformer.5.lconv1d.linear_start",
414
+ "model.audio_tower.conformer.5.lconv1d.linear_end",
415
+ "model.audio_tower.conformer.5.ffw_layer_end.ffw_layer_1",
416
+ "model.audio_tower.conformer.5.ffw_layer_end.ffw_layer_2",
417
+ "model.audio_tower.conformer.6.ffw_layer_start.ffw_layer_1",
418
+ "model.audio_tower.conformer.6.ffw_layer_start.ffw_layer_2",
419
+ "model.audio_tower.conformer.6.attention.attn.relative_position_embedding.pos_proj",
420
+ "model.audio_tower.conformer.6.attention.attn.q_proj",
421
+ "model.audio_tower.conformer.6.attention.attn.k_proj",
422
+ "model.audio_tower.conformer.6.attention.attn.v_proj",
423
+ "model.audio_tower.conformer.6.attention.post",
424
+ "model.audio_tower.conformer.6.lconv1d.linear_start",
425
+ "model.audio_tower.conformer.6.lconv1d.linear_end",
426
+ "model.audio_tower.conformer.6.ffw_layer_end.ffw_layer_1",
427
+ "model.audio_tower.conformer.6.ffw_layer_end.ffw_layer_2",
428
+ "model.audio_tower.conformer.7.ffw_layer_start.ffw_layer_1",
429
+ "model.audio_tower.conformer.7.ffw_layer_start.ffw_layer_2",
430
+ "model.audio_tower.conformer.7.attention.attn.relative_position_embedding.pos_proj",
431
+ "model.audio_tower.conformer.7.attention.attn.q_proj",
432
+ "model.audio_tower.conformer.7.attention.attn.k_proj",
433
+ "model.audio_tower.conformer.7.attention.attn.v_proj",
434
+ "model.audio_tower.conformer.7.attention.post",
435
+ "model.audio_tower.conformer.7.lconv1d.linear_start",
436
+ "model.audio_tower.conformer.7.lconv1d.linear_end",
437
+ "model.audio_tower.conformer.7.ffw_layer_end.ffw_layer_1",
438
+ "model.audio_tower.conformer.7.ffw_layer_end.ffw_layer_2",
439
+ "model.audio_tower.conformer.8.ffw_layer_start.ffw_layer_1",
440
+ "model.audio_tower.conformer.8.ffw_layer_start.ffw_layer_2",
441
+ "model.audio_tower.conformer.8.attention.attn.relative_position_embedding.pos_proj",
442
+ "model.audio_tower.conformer.8.attention.attn.q_proj",
443
+ "model.audio_tower.conformer.8.attention.attn.k_proj",
444
+ "model.audio_tower.conformer.8.attention.attn.v_proj",
445
+ "model.audio_tower.conformer.8.attention.post",
446
+ "model.audio_tower.conformer.8.lconv1d.linear_start",
447
+ "model.audio_tower.conformer.8.lconv1d.linear_end",
448
+ "model.audio_tower.conformer.8.ffw_layer_end.ffw_layer_1",
449
+ "model.audio_tower.conformer.8.ffw_layer_end.ffw_layer_2",
450
+ "model.audio_tower.conformer.9.ffw_layer_start.ffw_layer_1",
451
+ "model.audio_tower.conformer.9.ffw_layer_start.ffw_layer_2",
452
+ "model.audio_tower.conformer.9.attention.attn.relative_position_embedding.pos_proj",
453
+ "model.audio_tower.conformer.9.attention.attn.q_proj",
454
+ "model.audio_tower.conformer.9.attention.attn.k_proj",
455
+ "model.audio_tower.conformer.9.attention.attn.v_proj",
456
+ "model.audio_tower.conformer.9.attention.post",
457
+ "model.audio_tower.conformer.9.lconv1d.linear_start",
458
+ "model.audio_tower.conformer.9.lconv1d.linear_end",
459
+ "model.audio_tower.conformer.9.ffw_layer_end.ffw_layer_1",
460
+ "model.audio_tower.conformer.9.ffw_layer_end.ffw_layer_2",
461
+ "model.audio_tower.conformer.10.ffw_layer_start.ffw_layer_1",
462
+ "model.audio_tower.conformer.10.ffw_layer_start.ffw_layer_2",
463
+ "model.audio_tower.conformer.10.attention.attn.relative_position_embedding.pos_proj",
464
+ "model.audio_tower.conformer.10.attention.attn.q_proj",
465
+ "model.audio_tower.conformer.10.attention.attn.k_proj",
466
+ "model.audio_tower.conformer.10.attention.attn.v_proj",
467
+ "model.audio_tower.conformer.10.attention.post",
468
+ "model.audio_tower.conformer.10.lconv1d.linear_start",
469
+ "model.audio_tower.conformer.10.lconv1d.linear_end",
470
+ "model.audio_tower.conformer.10.ffw_layer_end.ffw_layer_1",
471
+ "model.audio_tower.conformer.10.ffw_layer_end.ffw_layer_2",
472
+ "model.audio_tower.conformer.11.ffw_layer_start.ffw_layer_1",
473
+ "model.audio_tower.conformer.11.ffw_layer_start.ffw_layer_2",
474
+ "model.audio_tower.conformer.11.attention.attn.relative_position_embedding.pos_proj",
475
+ "model.audio_tower.conformer.11.attention.attn.q_proj",
476
+ "model.audio_tower.conformer.11.attention.attn.k_proj",
477
+ "model.audio_tower.conformer.11.attention.attn.v_proj",
478
+ "model.audio_tower.conformer.11.attention.post",
479
+ "model.audio_tower.conformer.11.lconv1d.linear_start",
480
+ "model.audio_tower.conformer.11.lconv1d.linear_end",
481
+ "model.audio_tower.conformer.11.ffw_layer_end.ffw_layer_1",
482
+ "model.audio_tower.conformer.11.ffw_layer_end.ffw_layer_2",
483
+ "model.embed_vision.embedding_projection",
484
+ "model.embed_audio.embedding_projection",
485
+ "lm_head"
486
+ ],
487
+ "kv_cache_scheme": null,
488
+ "quant_method": "compressed-tensors",
489
+ "quantization_status": "compressed"
490
+ },
491
+ "text_config": {
492
+ "activation_sparsity_pattern": [
493
+ 0.95,
494
+ 0.95,
495
+ 0.95,
496
+ 0.95,
497
+ 0.95,
498
+ 0.95,
499
+ 0.95,
500
+ 0.95,
501
+ 0.95,
502
+ 0.95,
503
+ 0.0,
504
+ 0.0,
505
+ 0.0,
506
+ 0.0,
507
+ 0.0,
508
+ 0.0,
509
+ 0.0,
510
+ 0.0,
511
+ 0.0,
512
+ 0.0,
513
+ 0.0,
514
+ 0.0,
515
+ 0.0,
516
+ 0.0,
517
+ 0.0,
518
+ 0.0,
519
+ 0.0,
520
+ 0.0,
521
+ 0.0,
522
+ 0.0,
523
+ 0.0,
524
+ 0.0,
525
+ 0.0,
526
+ 0.0,
527
+ 0.0
528
+ ],
529
+ "altup_active_idx": 0,
530
+ "altup_coef_clip": 120.0,
531
+ "altup_correct_scale": true,
532
+ "altup_num_inputs": 4,
533
+ "attention_bias": false,
534
+ "attention_dropout": 0.0,
535
+ "final_logit_softcapping": 30.0,
536
+ "head_dim": 256,
537
+ "hidden_activation": "gelu_pytorch_tanh",
538
+ "hidden_size": 2048,
539
+ "hidden_size_per_layer_input": 256,
540
+ "initializer_range": 0.02,
541
+ "intermediate_size": [
542
+ 16384,
543
+ 16384,
544
+ 16384,
545
+ 16384,
546
+ 16384,
547
+ 16384,
548
+ 16384,
549
+ 16384,
550
+ 16384,
551
+ 16384,
552
+ 16384,
553
+ 16384,
554
+ 16384,
555
+ 16384,
556
+ 16384,
557
+ 16384,
558
+ 16384,
559
+ 16384,
560
+ 16384,
561
+ 16384,
562
+ 16384,
563
+ 16384,
564
+ 16384,
565
+ 16384,
566
+ 16384,
567
+ 16384,
568
+ 16384,
569
+ 16384,
570
+ 16384,
571
+ 16384,
572
+ 16384,
573
+ 16384,
574
+ 16384,
575
+ 16384,
576
+ 16384
577
+ ],
578
+ "laurel_rank": 64,
579
+ "layer_types": [
580
+ "sliding_attention",
581
+ "sliding_attention",
582
+ "sliding_attention",
583
+ "sliding_attention",
584
+ "full_attention",
585
+ "sliding_attention",
586
+ "sliding_attention",
587
+ "sliding_attention",
588
+ "sliding_attention",
589
+ "full_attention",
590
+ "sliding_attention",
591
+ "sliding_attention",
592
+ "sliding_attention",
593
+ "sliding_attention",
594
+ "full_attention",
595
+ "sliding_attention",
596
+ "sliding_attention",
597
+ "sliding_attention",
598
+ "sliding_attention",
599
+ "full_attention",
600
+ "sliding_attention",
601
+ "sliding_attention",
602
+ "sliding_attention",
603
+ "sliding_attention",
604
+ "full_attention",
605
+ "sliding_attention",
606
+ "sliding_attention",
607
+ "sliding_attention",
608
+ "sliding_attention",
609
+ "full_attention",
610
+ "sliding_attention",
611
+ "sliding_attention",
612
+ "sliding_attention",
613
+ "sliding_attention",
614
+ "full_attention"
615
+ ],
616
+ "max_position_embeddings": 32768,
617
+ "model_type": "gemma3n_text",
618
+ "num_attention_heads": 8,
619
+ "num_hidden_layers": 35,
620
+ "num_key_value_heads": 2,
621
+ "num_kv_shared_layers": 15,
622
+ "rms_norm_eps": 1e-06,
623
+ "rope_local_base_freq": 10000.0,
624
+ "rope_scaling": null,
625
+ "rope_theta": 1000000.0,
626
+ "sliding_window": 512,
627
+ "torch_dtype": "bfloat16",
628
+ "use_cache": true,
629
+ "vocab_size": 262400,
630
+ "vocab_size_per_layer_input": 262144
631
+ },
632
+ "torch_dtype": "bfloat16",
633
+ "transformers_version": "4.53.2",
634
+ "vision_config": {
635
+ "architecture": "mobilenetv5_300m_enc",
636
+ "do_pooling": false,
637
+ "hidden_size": 2048,
638
+ "initializer_range": 0.02,
639
+ "label_names": [
640
+ "LABEL_0",
641
+ "LABEL_1"
642
+ ],
643
+ "model_args": null,
644
+ "model_type": "gemma3n_vision",
645
+ "num_classes": 2,
646
+ "rms_norm_eps": 1e-06,
647
+ "torch_dtype": "bfloat16",
648
+ "vocab_offset": 262144,
649
+ "vocab_size": 128
650
+ },
651
+ "vision_soft_tokens_per_image": 256
652
+ }
generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "cache_implementation": "hybrid",
4
+ "do_sample": true,
5
+ "eos_token_id": [
6
+ 1,
7
+ 106
8
+ ],
9
+ "pad_token_id": 0,
10
+ "top_k": 64,
11
+ "top_p": 0.95,
12
+ "transformers_version": "4.53.2"
13
+ }
model-00001-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3916e50e5293a8ad10c7c2de039534e1875de2a03c3b5df7ce0519fb2c649a4e
3
+ size 4976858608
model-00002-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:37dcdf65554a6121307ceb00634ee332c24b773028a6d4d5f9158e512c9f20f9
3
+ size 672934152
model-00003-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5597d11cd53e6830d026588b38c7491ebd104820fde3fd7d4a0b743591f352e
3
+ size 4995580968
model-00004-of-00004.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b215d940b1442e27d2a33070fd871c5e08973599517ae35226ae02e058afa581
3
+ size 1167419608
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
preprocessor_config.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": null,
3
+ "data_format": "channels_first",
4
+ "default_to_square": false,
5
+ "device": null,
6
+ "disable_grouping": null,
7
+ "dither": 0.0,
8
+ "do_center_crop": null,
9
+ "do_convert_rgb": null,
10
+ "do_normalize": false,
11
+ "do_rescale": true,
12
+ "do_resize": true,
13
+ "feature_size": 128,
14
+ "fft_length": 1024,
15
+ "fft_overdrive": true,
16
+ "frame_length": 512,
17
+ "hop_length": 160,
18
+ "image_mean": [
19
+ 0.5,
20
+ 0.5,
21
+ 0.5
22
+ ],
23
+ "image_processor_type": "SiglipImageProcessorFast",
24
+ "image_seq_length": 256,
25
+ "image_std": [
26
+ 0.5,
27
+ 0.5,
28
+ 0.5
29
+ ],
30
+ "input_data_format": null,
31
+ "input_scale_factor": 1.0,
32
+ "max_frequency": 7600.0,
33
+ "mel_floor": 1e-05,
34
+ "min_frequency": 125.0,
35
+ "padding_side": "right",
36
+ "padding_value": 0.0,
37
+ "per_bin_mean": null,
38
+ "per_bin_stddev": null,
39
+ "preemphasis": 0.97,
40
+ "preemphasis_htk_flavor": true,
41
+ "processor_class": "Gemma3nProcessor",
42
+ "resample": 2,
43
+ "rescale_factor": 0.00392156862745098,
44
+ "return_attention_mask": true,
45
+ "return_tensors": null,
46
+ "sampling_rate": 16000,
47
+ "size": {
48
+ "height": 768,
49
+ "width": 768
50
+ }
51
+ }
processor_config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "audio_seq_length": 188,
3
+ "image_seq_length": 256,
4
+ "processor_class": "Gemma3nProcessor"
5
+ }
recipe.yaml ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ default_stage:
2
+ default_modifiers:
3
+ QuantizationModifier:
4
+ targets: [Linear]
5
+ ignore: ['re:.*embed_audio.*', 're:.*embed_vision.*', 're:.*audio_tower.*', 're:.*vision_tower.*',
6
+ 're:.*altup.*', 're:.*lm_head.*', 're:.*laurel.*', 're:model\.language_model\.layers\.\d+\.per_layer_input_gate',
7
+ 're:model\.language_model\.layers\.\d+\.per_layer_projection', model.language_model.per_layer_model_projection]
8
+ scheme: FP8_DYNAMIC
special_tokens_map.json ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audio_token": "<audio_soft_token>",
3
+ "boa_token": "<start_of_audio>",
4
+ "boi_token": "<start_of_image>",
5
+ "bos_token": {
6
+ "content": "<bos>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false
11
+ },
12
+ "eoa_token": "<end_of_audio>",
13
+ "eoi_token": "<end_of_image>",
14
+ "eos_token": {
15
+ "content": "<eos>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false
20
+ },
21
+ "image_token": "<image_soft_token>",
22
+ "pad_token": {
23
+ "content": "<pad>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false
28
+ },
29
+ "unk_token": {
30
+ "content": "<unk>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false
35
+ }
36
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b6c35ee648c07754b44cd9e371c75d4caa05c4504910b7ad29b1847ee9d8ba5d
3
+ size 33442553
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ea5f0cc48abfbfc04d14562270a32e02149a3e7035f368cc5a462786f4a59961
3
+ size 4696020
tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff