Hugging Face's logo

RedHatAI
/
gemma-3n-E4B-it-quantized.w4a16

Image-Text-to-Text
Transformers
Safetensors
gemma3n
vllm
vision
audio
int4
conversational
compressed-tensors
Model card Files
xet
Community
1
Shubhra Pandit commited on
Commit
24d7bfb
·
1 Parent(s): a294598

Upload model files

Browse files
Files changed (16) hide show
  1. .gitattributes +2 -0
  2. chat_template.jinja +49 -0
  3. config.json +641 -0
  4. generation_config.json +13 -0
  5. lighteval_results/model_args.yaml +12 -0
  6. model-00001-of-00003.safetensors +3 -0
  7. model-00002-of-00003.safetensors +3 -0
  8. model-00003-of-00003.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 +12 -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 +1,641 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": null,
66
+ "output_activations": null,
67
+ "targets": [
68
+ "Linear"
69
+ ],
70
+ "weights": {
71
+ "actorder": null,
72
+ "block_structure": null,
73
+ "dynamic": false,
74
+ "group_size": 128,
75
+ "num_bits": 4,
76
+ "observer": "minmax",
77
+ "observer_kwargs": {},
78
+ "strategy": "group",
79
+ "symmetric": true,
80
+ "type": "int"
81
+ }
82
+ }
83
+ },
84
+ "format": "pack-quantized",
85
+ "global_compression_ratio": null,
86
+ "ignore": [
87
+ "model.language_model.layers.0.altup.correction_coefs",
88
+ "model.language_model.layers.0.altup.prediction_coefs",
89
+ "model.language_model.layers.0.altup.modality_router",
90
+ "model.language_model.layers.0.laurel.linear_left",
91
+ "model.language_model.layers.0.laurel.linear_right",
92
+ "model.language_model.layers.0.per_layer_input_gate",
93
+ "model.language_model.layers.0.per_layer_projection",
94
+ "model.language_model.layers.1.altup.correction_coefs",
95
+ "model.language_model.layers.1.altup.prediction_coefs",
96
+ "model.language_model.layers.1.altup.modality_router",
97
+ "model.language_model.layers.1.laurel.linear_left",
98
+ "model.language_model.layers.1.laurel.linear_right",
99
+ "model.language_model.layers.1.per_layer_input_gate",
100
+ "model.language_model.layers.1.per_layer_projection",
101
+ "model.language_model.layers.2.altup.correction_coefs",
102
+ "model.language_model.layers.2.altup.prediction_coefs",
103
+ "model.language_model.layers.2.altup.modality_router",
104
+ "model.language_model.layers.2.laurel.linear_left",
105
+ "model.language_model.layers.2.laurel.linear_right",
106
+ "model.language_model.layers.2.per_layer_input_gate",
107
+ "model.language_model.layers.2.per_layer_projection",
108
+ "model.language_model.layers.3.altup.correction_coefs",
109
+ "model.language_model.layers.3.altup.prediction_coefs",
110
+ "model.language_model.layers.3.altup.modality_router",
111
+ "model.language_model.layers.3.laurel.linear_left",
112
+ "model.language_model.layers.3.laurel.linear_right",
113
+ "model.language_model.layers.3.per_layer_input_gate",
114
+ "model.language_model.layers.3.per_layer_projection",
115
+ "model.language_model.layers.4.altup.correction_coefs",
116
+ "model.language_model.layers.4.altup.prediction_coefs",
117
+ "model.language_model.layers.4.altup.modality_router",
118
+ "model.language_model.layers.4.laurel.linear_left",
119
+ "model.language_model.layers.4.laurel.linear_right",
120
+ "model.language_model.layers.4.per_layer_input_gate",
121
+ "model.language_model.layers.4.per_layer_projection",
122
+ "model.language_model.layers.5.altup.correction_coefs",
123
+ "model.language_model.layers.5.altup.prediction_coefs",
124
+ "model.language_model.layers.5.altup.modality_router",
125
+ "model.language_model.layers.5.laurel.linear_left",
126
+ "model.language_model.layers.5.laurel.linear_right",
127
+ "model.language_model.layers.5.per_layer_input_gate",
128
+ "model.language_model.layers.5.per_layer_projection",
129
+ "model.language_model.layers.6.altup.correction_coefs",
130
+ "model.language_model.layers.6.altup.prediction_coefs",
131
+ "model.language_model.layers.6.altup.modality_router",
132
+ "model.language_model.layers.6.laurel.linear_left",
133
+ "model.language_model.layers.6.laurel.linear_right",
134
+ "model.language_model.layers.6.per_layer_input_gate",
135
+ "model.language_model.layers.6.per_layer_projection",
136
+ "model.language_model.layers.7.altup.correction_coefs",
137
+ "model.language_model.layers.7.altup.prediction_coefs",
138
+ "model.language_model.layers.7.altup.modality_router",
139
+ "model.language_model.layers.7.laurel.linear_left",
140
+ "model.language_model.layers.7.laurel.linear_right",
141
+ "model.language_model.layers.7.per_layer_input_gate",
142
+ "model.language_model.layers.7.per_layer_projection",
143
+ "model.language_model.layers.8.altup.correction_coefs",
144
+ "model.language_model.layers.8.altup.prediction_coefs",
145
+ "model.language_model.layers.8.altup.modality_router",
146
+ "model.language_model.layers.8.laurel.linear_left",
147
+ "model.language_model.layers.8.laurel.linear_right",
148
+ "model.language_model.layers.8.per_layer_input_gate",
149
+ "model.language_model.layers.8.per_layer_projection",
150
+ "model.language_model.layers.9.altup.correction_coefs",
151
+ "model.language_model.layers.9.altup.prediction_coefs",
152
+ "model.language_model.layers.9.altup.modality_router",
153
+ "model.language_model.layers.9.laurel.linear_left",
154
+ "model.language_model.layers.9.laurel.linear_right",
155
+ "model.language_model.layers.9.per_layer_input_gate",
156
+ "model.language_model.layers.9.per_layer_projection",
157
+ "model.language_model.layers.10.altup.correction_coefs",
158
+ "model.language_model.layers.10.altup.prediction_coefs",
159
+ "model.language_model.layers.10.altup.modality_router",
160
+ "model.language_model.layers.10.laurel.linear_left",
161
+ "model.language_model.layers.10.laurel.linear_right",
162
+ "model.language_model.layers.10.per_layer_input_gate",
163
+ "model.language_model.layers.10.per_layer_projection",
164
+ "model.language_model.layers.11.altup.correction_coefs",
165
+ "model.language_model.layers.11.altup.prediction_coefs",
166
+ "model.language_model.layers.11.altup.modality_router",
167
+ "model.language_model.layers.11.laurel.linear_left",
168
+ "model.language_model.layers.11.laurel.linear_right",
169
+ "model.language_model.layers.11.per_layer_input_gate",
170
+ "model.language_model.layers.11.per_layer_projection",
171
+ "model.language_model.layers.12.altup.correction_coefs",
172
+ "model.language_model.layers.12.altup.prediction_coefs",
173
+ "model.language_model.layers.12.altup.modality_router",
174
+ "model.language_model.layers.12.laurel.linear_left",
175
+ "model.language_model.layers.12.laurel.linear_right",
176
+ "model.language_model.layers.12.per_layer_input_gate",
177
+ "model.language_model.layers.12.per_layer_projection",
178
+ "model.language_model.layers.13.altup.correction_coefs",
179
+ "model.language_model.layers.13.altup.prediction_coefs",
180
+ "model.language_model.layers.13.altup.modality_router",
181
+ "model.language_model.layers.13.laurel.linear_left",
182
+ "model.language_model.layers.13.laurel.linear_right",
183
+ "model.language_model.layers.13.per_layer_input_gate",
184
+ "model.language_model.layers.13.per_layer_projection",
185
+ "model.language_model.layers.14.altup.correction_coefs",
186
+ "model.language_model.layers.14.altup.prediction_coefs",
187
+ "model.language_model.layers.14.altup.modality_router",
188
+ "model.language_model.layers.14.laurel.linear_left",
189
+ "model.language_model.layers.14.laurel.linear_right",
190
+ "model.language_model.layers.14.per_layer_input_gate",
191
+ "model.language_model.layers.14.per_layer_projection",
192
+ "model.language_model.layers.15.altup.correction_coefs",
193
+ "model.language_model.layers.15.altup.prediction_coefs",
194
+ "model.language_model.layers.15.altup.modality_router",
195
+ "model.language_model.layers.15.laurel.linear_left",
196
+ "model.language_model.layers.15.laurel.linear_right",
197
+ "model.language_model.layers.15.per_layer_input_gate",
198
+ "model.language_model.layers.15.per_layer_projection",
199
+ "model.language_model.layers.16.altup.correction_coefs",
200
+ "model.language_model.layers.16.altup.prediction_coefs",
201
+ "model.language_model.layers.16.altup.modality_router",
202
+ "model.language_model.layers.16.laurel.linear_left",
203
+ "model.language_model.layers.16.laurel.linear_right",
204
+ "model.language_model.layers.16.per_layer_input_gate",
205
+ "model.language_model.layers.16.per_layer_projection",
206
+ "model.language_model.layers.17.altup.correction_coefs",
207
+ "model.language_model.layers.17.altup.prediction_coefs",
208
+ "model.language_model.layers.17.altup.modality_router",
209
+ "model.language_model.layers.17.laurel.linear_left",
210
+ "model.language_model.layers.17.laurel.linear_right",
211
+ "model.language_model.layers.17.per_layer_input_gate",
212
+ "model.language_model.layers.17.per_layer_projection",
213
+ "model.language_model.layers.18.altup.correction_coefs",
214
+ "model.language_model.layers.18.altup.prediction_coefs",
215
+ "model.language_model.layers.18.altup.modality_router",
216
+ "model.language_model.layers.18.laurel.linear_left",
217
+ "model.language_model.layers.18.laurel.linear_right",
218
+ "model.language_model.layers.18.per_layer_input_gate",
219
+ "model.language_model.layers.18.per_layer_projection",
220
+ "model.language_model.layers.19.altup.correction_coefs",
221
+ "model.language_model.layers.19.altup.prediction_coefs",
222
+ "model.language_model.layers.19.altup.modality_router",
223
+ "model.language_model.layers.19.laurel.linear_left",
224
+ "model.language_model.layers.19.laurel.linear_right",
225
+ "model.language_model.layers.19.per_layer_input_gate",
226
+ "model.language_model.layers.19.per_layer_projection",
227
+ "model.language_model.layers.20.altup.correction_coefs",
228
+ "model.language_model.layers.20.altup.prediction_coefs",
229
+ "model.language_model.layers.20.altup.modality_router",
230
+ "model.language_model.layers.20.laurel.linear_left",
231
+ "model.language_model.layers.20.laurel.linear_right",
232
+ "model.language_model.layers.20.per_layer_input_gate",
233
+ "model.language_model.layers.20.per_layer_projection",
234
+ "model.language_model.layers.21.altup.correction_coefs",
235
+ "model.language_model.layers.21.altup.prediction_coefs",
236
+ "model.language_model.layers.21.altup.modality_router",
237
+ "model.language_model.layers.21.laurel.linear_left",
238
+ "model.language_model.layers.21.laurel.linear_right",
239
+ "model.language_model.layers.21.per_layer_input_gate",
240
+ "model.language_model.layers.21.per_layer_projection",
241
+ "model.language_model.layers.22.altup.correction_coefs",
242
+ "model.language_model.layers.22.altup.prediction_coefs",
243
+ "model.language_model.layers.22.altup.modality_router",
244
+ "model.language_model.layers.22.laurel.linear_left",
245
+ "model.language_model.layers.22.laurel.linear_right",
246
+ "model.language_model.layers.22.per_layer_input_gate",
247
+ "model.language_model.layers.22.per_layer_projection",
248
+ "model.language_model.layers.23.altup.correction_coefs",
249
+ "model.language_model.layers.23.altup.prediction_coefs",
250
+ "model.language_model.layers.23.altup.modality_router",
251
+ "model.language_model.layers.23.laurel.linear_left",
252
+ "model.language_model.layers.23.laurel.linear_right",
253
+ "model.language_model.layers.23.per_layer_input_gate",
254
+ "model.language_model.layers.23.per_layer_projection",
255
+ "model.language_model.layers.24.altup.correction_coefs",
256
+ "model.language_model.layers.24.altup.prediction_coefs",
257
+ "model.language_model.layers.24.altup.modality_router",
258
+ "model.language_model.layers.24.laurel.linear_left",
259
+ "model.language_model.layers.24.laurel.linear_right",
260
+ "model.language_model.layers.24.per_layer_input_gate",
261
+ "model.language_model.layers.24.per_layer_projection",
262
+ "model.language_model.layers.25.altup.correction_coefs",
263
+ "model.language_model.layers.25.altup.prediction_coefs",
264
+ "model.language_model.layers.25.altup.modality_router",
265
+ "model.language_model.layers.25.laurel.linear_left",
266
+ "model.language_model.layers.25.laurel.linear_right",
267
+ "model.language_model.layers.25.per_layer_input_gate",
268
+ "model.language_model.layers.25.per_layer_projection",
269
+ "model.language_model.layers.26.altup.correction_coefs",
270
+ "model.language_model.layers.26.altup.prediction_coefs",
271
+ "model.language_model.layers.26.altup.modality_router",
272
+ "model.language_model.layers.26.laurel.linear_left",
273
+ "model.language_model.layers.26.laurel.linear_right",
274
+ "model.language_model.layers.26.per_layer_input_gate",
275
+ "model.language_model.layers.26.per_layer_projection",
276
+ "model.language_model.layers.27.altup.correction_coefs",
277
+ "model.language_model.layers.27.altup.prediction_coefs",
278
+ "model.language_model.layers.27.altup.modality_router",
279
+ "model.language_model.layers.27.laurel.linear_left",
280
+ "model.language_model.layers.27.laurel.linear_right",
281
+ "model.language_model.layers.27.per_layer_input_gate",
282
+ "model.language_model.layers.27.per_layer_projection",
283
+ "model.language_model.layers.28.altup.correction_coefs",
284
+ "model.language_model.layers.28.altup.prediction_coefs",
285
+ "model.language_model.layers.28.altup.modality_router",
286
+ "model.language_model.layers.28.laurel.linear_left",
287
+ "model.language_model.layers.28.laurel.linear_right",
288
+ "model.language_model.layers.28.per_layer_input_gate",
289
+ "model.language_model.layers.28.per_layer_projection",
290
+ "model.language_model.layers.29.altup.correction_coefs",
291
+ "model.language_model.layers.29.altup.prediction_coefs",
292
+ "model.language_model.layers.29.altup.modality_router",
293
+ "model.language_model.layers.29.laurel.linear_left",
294
+ "model.language_model.layers.29.laurel.linear_right",
295
+ "model.language_model.layers.29.per_layer_input_gate",
296
+ "model.language_model.layers.29.per_layer_projection",
297
+ "model.language_model.layers.30.altup.correction_coefs",
298
+ "model.language_model.layers.30.altup.prediction_coefs",
299
+ "model.language_model.layers.30.altup.modality_router",
300
+ "model.language_model.layers.30.laurel.linear_left",
301
+ "model.language_model.layers.30.laurel.linear_right",
302
+ "model.language_model.layers.30.per_layer_input_gate",
303
+ "model.language_model.layers.30.per_layer_projection",
304
+ "model.language_model.layers.31.altup.correction_coefs",
305
+ "model.language_model.layers.31.altup.prediction_coefs",
306
+ "model.language_model.layers.31.altup.modality_router",
307
+ "model.language_model.layers.31.laurel.linear_left",
308
+ "model.language_model.layers.31.laurel.linear_right",
309
+ "model.language_model.layers.31.per_layer_input_gate",
310
+ "model.language_model.layers.31.per_layer_projection",
311
+ "model.language_model.layers.32.altup.correction_coefs",
312
+ "model.language_model.layers.32.altup.prediction_coefs",
313
+ "model.language_model.layers.32.altup.modality_router",
314
+ "model.language_model.layers.32.laurel.linear_left",
315
+ "model.language_model.layers.32.laurel.linear_right",
316
+ "model.language_model.layers.32.per_layer_input_gate",
317
+ "model.language_model.layers.32.per_layer_projection",
318
+ "model.language_model.layers.33.altup.correction_coefs",
319
+ "model.language_model.layers.33.altup.prediction_coefs",
320
+ "model.language_model.layers.33.altup.modality_router",
321
+ "model.language_model.layers.33.laurel.linear_left",
322
+ "model.language_model.layers.33.laurel.linear_right",
323
+ "model.language_model.layers.33.per_layer_input_gate",
324
+ "model.language_model.layers.33.per_layer_projection",
325
+ "model.language_model.layers.34.altup.correction_coefs",
326
+ "model.language_model.layers.34.altup.prediction_coefs",
327
+ "model.language_model.layers.34.altup.modality_router",
328
+ "model.language_model.layers.34.laurel.linear_left",
329
+ "model.language_model.layers.34.laurel.linear_right",
330
+ "model.language_model.layers.34.per_layer_input_gate",
331
+ "model.language_model.layers.34.per_layer_projection",
332
+ "model.language_model.per_layer_model_projection",
333
+ "model.language_model.altup_projections.0",
334
+ "model.language_model.altup_projections.1",
335
+ "model.language_model.altup_projections.2",
336
+ "model.language_model.altup_unembed_projections.0",
337
+ "model.language_model.altup_unembed_projections.1",
338
+ "model.language_model.altup_unembed_projections.2",
339
+ "model.audio_tower.subsample_conv_projection.input_proj_linear",
340
+ "model.audio_tower.conformer.0.ffw_layer_start.ffw_layer_1",
341
+ "model.audio_tower.conformer.0.ffw_layer_start.ffw_layer_2",
342
+ "model.audio_tower.conformer.0.attention.attn.relative_position_embedding.pos_proj",
343
+ "model.audio_tower.conformer.0.attention.attn.q_proj",
344
+ "model.audio_tower.conformer.0.attention.attn.k_proj",
345
+ "model.audio_tower.conformer.0.attention.attn.v_proj",
346
+ "model.audio_tower.conformer.0.attention.post",
347
+ "model.audio_tower.conformer.0.lconv1d.linear_start",
348
+ "model.audio_tower.conformer.0.lconv1d.linear_end",
349
+ "model.audio_tower.conformer.0.ffw_layer_end.ffw_layer_1",
350
+ "model.audio_tower.conformer.0.ffw_layer_end.ffw_layer_2",
351
+ "model.audio_tower.conformer.1.ffw_layer_start.ffw_layer_1",
352
+ "model.audio_tower.conformer.1.ffw_layer_start.ffw_layer_2",
353
+ "model.audio_tower.conformer.1.attention.attn.relative_position_embedding.pos_proj",
354
+ "model.audio_tower.conformer.1.attention.attn.q_proj",
355
+ "model.audio_tower.conformer.1.attention.attn.k_proj",
356
+ "model.audio_tower.conformer.1.attention.attn.v_proj",
357
+ "model.audio_tower.conformer.1.attention.post",
358
+ "model.audio_tower.conformer.1.lconv1d.linear_start",
359
+ "model.audio_tower.conformer.1.lconv1d.linear_end",
360
+ "model.audio_tower.conformer.1.ffw_layer_end.ffw_layer_1",
361
+ "model.audio_tower.conformer.1.ffw_layer_end.ffw_layer_2",
362
+ "model.audio_tower.conformer.2.ffw_layer_start.ffw_layer_1",
363
+ "model.audio_tower.conformer.2.ffw_layer_start.ffw_layer_2",
364
+ "model.audio_tower.conformer.2.attention.attn.relative_position_embedding.pos_proj",
365
+ "model.audio_tower.conformer.2.attention.attn.q_proj",
366
+ "model.audio_tower.conformer.2.attention.attn.k_proj",
367
+ "model.audio_tower.conformer.2.attention.attn.v_proj",
368
+ "model.audio_tower.conformer.2.attention.post",
369
+ "model.audio_tower.conformer.2.lconv1d.linear_start",
370
+ "model.audio_tower.conformer.2.lconv1d.linear_end",
371
+ "model.audio_tower.conformer.2.ffw_layer_end.ffw_layer_1",
372
+ "model.audio_tower.conformer.2.ffw_layer_end.ffw_layer_2",
373
+ "model.audio_tower.conformer.3.ffw_layer_start.ffw_layer_1",
374
+ "model.audio_tower.conformer.3.ffw_layer_start.ffw_layer_2",
375
+ "model.audio_tower.conformer.3.attention.attn.relative_position_embedding.pos_proj",
376
+ "model.audio_tower.conformer.3.attention.attn.q_proj",
377
+ "model.audio_tower.conformer.3.attention.attn.k_proj",
378
+ "model.audio_tower.conformer.3.attention.attn.v_proj",
379
+ "model.audio_tower.conformer.3.attention.post",
380
+ "model.audio_tower.conformer.3.lconv1d.linear_start",
381
+ "model.audio_tower.conformer.3.lconv1d.linear_end",
382
+ "model.audio_tower.conformer.3.ffw_layer_end.ffw_layer_1",
383
+ "model.audio_tower.conformer.3.ffw_layer_end.ffw_layer_2",
384
+ "model.audio_tower.conformer.4.ffw_layer_start.ffw_layer_1",
385
+ "model.audio_tower.conformer.4.ffw_layer_start.ffw_layer_2",
386
+ "model.audio_tower.conformer.4.attention.attn.relative_position_embedding.pos_proj",
387
+ "model.audio_tower.conformer.4.attention.attn.q_proj",
388
+ "model.audio_tower.conformer.4.attention.attn.k_proj",
389
+ "model.audio_tower.conformer.4.attention.attn.v_proj",
390
+ "model.audio_tower.conformer.4.attention.post",
391
+ "model.audio_tower.conformer.4.lconv1d.linear_start",
392
+ "model.audio_tower.conformer.4.lconv1d.linear_end",
393
+ "model.audio_tower.conformer.4.ffw_layer_end.ffw_layer_1",
394
+ "model.audio_tower.conformer.4.ffw_layer_end.ffw_layer_2",
395
+ "model.audio_tower.conformer.5.ffw_layer_start.ffw_layer_1",
396
+ "model.audio_tower.conformer.5.ffw_layer_start.ffw_layer_2",
397
+ "model.audio_tower.conformer.5.attention.attn.relative_position_embedding.pos_proj",
398
+ "model.audio_tower.conformer.5.attention.attn.q_proj",
399
+ "model.audio_tower.conformer.5.attention.attn.k_proj",
400
+ "model.audio_tower.conformer.5.attention.attn.v_proj",
401
+ "model.audio_tower.conformer.5.attention.post",
402
+ "model.audio_tower.conformer.5.lconv1d.linear_start",
403
+ "model.audio_tower.conformer.5.lconv1d.linear_end",
404
+ "model.audio_tower.conformer.5.ffw_layer_end.ffw_layer_1",
405
+ "model.audio_tower.conformer.5.ffw_layer_end.ffw_layer_2",
406
+ "model.audio_tower.conformer.6.ffw_layer_start.ffw_layer_1",
407
+ "model.audio_tower.conformer.6.ffw_layer_start.ffw_layer_2",
408
+ "model.audio_tower.conformer.6.attention.attn.relative_position_embedding.pos_proj",
409
+ "model.audio_tower.conformer.6.attention.attn.q_proj",
410
+ "model.audio_tower.conformer.6.attention.attn.k_proj",
411
+ "model.audio_tower.conformer.6.attention.attn.v_proj",
412
+ "model.audio_tower.conformer.6.attention.post",
413
+ "model.audio_tower.conformer.6.lconv1d.linear_start",
414
+ "model.audio_tower.conformer.6.lconv1d.linear_end",
415
+ "model.audio_tower.conformer.6.ffw_layer_end.ffw_layer_1",
416
+ "model.audio_tower.conformer.6.ffw_layer_end.ffw_layer_2",
417
+ "model.audio_tower.conformer.7.ffw_layer_start.ffw_layer_1",
418
+ "model.audio_tower.conformer.7.ffw_layer_start.ffw_layer_2",
419
+ "model.audio_tower.conformer.7.attention.attn.relative_position_embedding.pos_proj",
420
+ "model.audio_tower.conformer.7.attention.attn.q_proj",
421
+ "model.audio_tower.conformer.7.attention.attn.k_proj",
422
+ "model.audio_tower.conformer.7.attention.attn.v_proj",
423
+ "model.audio_tower.conformer.7.attention.post",
424
+ "model.audio_tower.conformer.7.lconv1d.linear_start",
425
+ "model.audio_tower.conformer.7.lconv1d.linear_end",
426
+ "model.audio_tower.conformer.7.ffw_layer_end.ffw_layer_1",
427
+ "model.audio_tower.conformer.7.ffw_layer_end.ffw_layer_2",
428
+ "model.audio_tower.conformer.8.ffw_layer_start.ffw_layer_1",
429
+ "model.audio_tower.conformer.8.ffw_layer_start.ffw_layer_2",
430
+ "model.audio_tower.conformer.8.attention.attn.relative_position_embedding.pos_proj",
431
+ "model.audio_tower.conformer.8.attention.attn.q_proj",
432
+ "model.audio_tower.conformer.8.attention.attn.k_proj",
433
+ "model.audio_tower.conformer.8.attention.attn.v_proj",
434
+ "model.audio_tower.conformer.8.attention.post",
435
+ "model.audio_tower.conformer.8.lconv1d.linear_start",
436
+ "model.audio_tower.conformer.8.lconv1d.linear_end",
437
+ "model.audio_tower.conformer.8.ffw_layer_end.ffw_layer_1",
438
+ "model.audio_tower.conformer.8.ffw_layer_end.ffw_layer_2",
439
+ "model.audio_tower.conformer.9.ffw_layer_start.ffw_layer_1",
440
+ "model.audio_tower.conformer.9.ffw_layer_start.ffw_layer_2",
441
+ "model.audio_tower.conformer.9.attention.attn.relative_position_embedding.pos_proj",
442
+ "model.audio_tower.conformer.9.attention.attn.q_proj",
443
+ "model.audio_tower.conformer.9.attention.attn.k_proj",
444
+ "model.audio_tower.conformer.9.attention.attn.v_proj",
445
+ "model.audio_tower.conformer.9.attention.post",
446
+ "model.audio_tower.conformer.9.lconv1d.linear_start",
447
+ "model.audio_tower.conformer.9.lconv1d.linear_end",
448
+ "model.audio_tower.conformer.9.ffw_layer_end.ffw_layer_1",
449
+ "model.audio_tower.conformer.9.ffw_layer_end.ffw_layer_2",
450
+ "model.audio_tower.conformer.10.ffw_layer_start.ffw_layer_1",
451
+ "model.audio_tower.conformer.10.ffw_layer_start.ffw_layer_2",
452
+ "model.audio_tower.conformer.10.attention.attn.relative_position_embedding.pos_proj",
453
+ "model.audio_tower.conformer.10.attention.attn.q_proj",
454
+ "model.audio_tower.conformer.10.attention.attn.k_proj",
455
+ "model.audio_tower.conformer.10.attention.attn.v_proj",
456
+ "model.audio_tower.conformer.10.attention.post",
457
+ "model.audio_tower.conformer.10.lconv1d.linear_start",
458
+ "model.audio_tower.conformer.10.lconv1d.linear_end",
459
+ "model.audio_tower.conformer.10.ffw_layer_end.ffw_layer_1",
460
+ "model.audio_tower.conformer.10.ffw_layer_end.ffw_layer_2",
461
+ "model.audio_tower.conformer.11.ffw_layer_start.ffw_layer_1",
462
+ "model.audio_tower.conformer.11.ffw_layer_start.ffw_layer_2",
463
+ "model.audio_tower.conformer.11.attention.attn.relative_position_embedding.pos_proj",
464
+ "model.audio_tower.conformer.11.attention.attn.q_proj",
465
+ "model.audio_tower.conformer.11.attention.attn.k_proj",
466
+ "model.audio_tower.conformer.11.attention.attn.v_proj",
467
+ "model.audio_tower.conformer.11.attention.post",
468
+ "model.audio_tower.conformer.11.lconv1d.linear_start",
469
+ "model.audio_tower.conformer.11.lconv1d.linear_end",
470
+ "model.audio_tower.conformer.11.ffw_layer_end.ffw_layer_1",
471
+ "model.audio_tower.conformer.11.ffw_layer_end.ffw_layer_2",
472
+ "model.embed_vision.embedding_projection",
473
+ "model.embed_audio.embedding_projection",
474
+ "lm_head"
475
+ ],
476
+ "kv_cache_scheme": null,
477
+ "quant_method": "compressed-tensors",
478
+ "quantization_status": "compressed"
479
+ },
480
+ "text_config": {
481
+ "activation_sparsity_pattern": [
482
+ 0.95,
483
+ 0.95,
484
+ 0.95,
485
+ 0.95,
486
+ 0.95,
487
+ 0.95,
488
+ 0.95,
489
+ 0.95,
490
+ 0.95,
491
+ 0.95,
492
+ 0.0,
493
+ 0.0,
494
+ 0.0,
495
+ 0.0,
496
+ 0.0,
497
+ 0.0,
498
+ 0.0,
499
+ 0.0,
500
+ 0.0,
501
+ 0.0,
502
+ 0.0,
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
+ ],
518
+ "altup_active_idx": 0,
519
+ "altup_coef_clip": 120.0,
520
+ "altup_correct_scale": true,
521
+ "altup_num_inputs": 4,
522
+ "attention_bias": false,
523
+ "attention_dropout": 0.0,
524
+ "final_logit_softcapping": 30.0,
525
+ "head_dim": 256,
526
+ "hidden_activation": "gelu_pytorch_tanh",
527
+ "hidden_size": 2048,
528
+ "hidden_size_per_layer_input": 256,
529
+ "initializer_range": 0.02,
530
+ "intermediate_size": [
531
+ 16384,
532
+ 16384,
533
+ 16384,
534
+ 16384,
535
+ 16384,
536
+ 16384,
537
+ 16384,
538
+ 16384,
539
+ 16384,
540
+ 16384,
541
+ 16384,
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
+ ],
567
+ "laurel_rank": 64,
568
+ "layer_types": [
569
+ "sliding_attention",
570
+ "sliding_attention",
571
+ "sliding_attention",
572
+ "sliding_attention",
573
+ "full_attention",
574
+ "sliding_attention",
575
+ "sliding_attention",
576
+ "sliding_attention",
577
+ "sliding_attention",
578
+ "full_attention",
579
+ "sliding_attention",
580
+ "sliding_attention",
581
+ "sliding_attention",
582
+ "sliding_attention",
583
+ "full_attention",
584
+ "sliding_attention",
585
+ "sliding_attention",
586
+ "sliding_attention",
587
+ "sliding_attention",
588
+ "full_attention",
589
+ "sliding_attention",
590
+ "sliding_attention",
591
+ "sliding_attention",
592
+ "sliding_attention",
593
+ "full_attention",
594
+ "sliding_attention",
595
+ "sliding_attention",
596
+ "sliding_attention",
597
+ "sliding_attention",
598
+ "full_attention",
599
+ "sliding_attention",
600
+ "sliding_attention",
601
+ "sliding_attention",
602
+ "sliding_attention",
603
+ "full_attention"
604
+ ],
605
+ "max_position_embeddings": 32768,
606
+ "model_type": "gemma3n_text",
607
+ "num_attention_heads": 8,
608
+ "num_hidden_layers": 35,
609
+ "num_key_value_heads": 2,
610
+ "num_kv_shared_layers": 15,
611
+ "rms_norm_eps": 1e-06,
612
+ "rope_local_base_freq": 10000.0,
613
+ "rope_scaling": null,
614
+ "rope_theta": 1000000.0,
615
+ "sliding_window": 512,
616
+ "torch_dtype": "bfloat16",
617
+ "use_cache": true,
618
+ "vocab_size": 262400,
619
+ "vocab_size_per_layer_input": 262144
620
+ },
621
+ "torch_dtype": "bfloat16",
622
+ "transformers_version": "4.53.2",
623
+ "vision_config": {
624
+ "architecture": "mobilenetv5_300m_enc",
625
+ "do_pooling": false,
626
+ "hidden_size": 2048,
627
+ "initializer_range": 0.02,
628
+ "label_names": [
629
+ "LABEL_0",
630
+ "LABEL_1"
631
+ ],
632
+ "model_args": null,
633
+ "model_type": "gemma3n_vision",
634
+ "num_classes": 2,
635
+ "rms_norm_eps": 1e-06,
636
+ "torch_dtype": "bfloat16",
637
+ "vocab_offset": 262144,
638
+ "vocab_size": 128
639
+ },
640
+ "vision_soft_tokens_per_image": 256
641
+ }
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
+ }
lighteval_results/model_args.yaml ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model_parameters:
2
+ model_name: "/network/shubhra/gemma_3n_quantized/gemma-3n-E4B-it-W4A16__ns512_df0.01_fix"
3
+ dtype: auto
4
+ gpu_memory_utilization: 0.9
5
+ tensor_parallel_size: 1
6
+ max_model_length: 32768
7
+ generation_parameters:
8
+ temperature: 0.6
9
+ top_k: 20
10
+ min_p: 0.0
11
+ top_p: 0.95
12
+ max_new_tokens: 32768
model-00001-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f6528ee02245d643390c30c4d2fb717cee1f6612053f4cb0403f01c4a9aad2c
3
+ size 3762706944
model-00002-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5597d11cd53e6830d026588b38c7491ebd104820fde3fd7d4a0b743591f352e
3
+ size 4995580968
model-00003-of-00003.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 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ default_stage:
2
+ default_modifiers:
3
+ GPTQModifier:
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: W4A16
9
+ sequential_update: true
10
+ block_size: 128
11
+ dampening_frac: 0.01
12
+ offload_hessians: false
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 sha256:57cce74cb255a2400a0aaf1be4bb84afcce027f7809d149a5cde1a7a93fa56cd
3
+ size 33442819
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