Hugging Face's logo

HIT-TMG
/
Lychee-FD

Safetensors
Chinese
English
speech
audio
full-duplex
realtime
speech-to-speech
vllm
Model card Files
xet
 Community
7

Upload folder using huggingface_hub

#1
by xyidealist - opened
base: refs/heads/main
←
from: refs/pr/1
Discussion Files changed
+365811
-0
Files changed (18) hide show
  1. README.md +181 -0
  2. lychee_full_duplex_v1.5/added_tokens.json +0 -0
  3. lychee_full_duplex_v1.5/chat_template.jinja +54 -0
  4. lychee_full_duplex_v1.5/config.json +375 -0
  5. lychee_full_duplex_v1.5/configuration_step_audio_2.py +128 -0
  6. lychee_full_duplex_v1.5/generation_config.json +4 -0
  7. lychee_full_duplex_v1.5/merges.txt +0 -0
  8. lychee_full_duplex_v1.5/model-00001-of-00006.safetensors +3 -0
  9. lychee_full_duplex_v1.5/model-00002-of-00006.safetensors +3 -0
  10. lychee_full_duplex_v1.5/model-00003-of-00006.safetensors +3 -0
  11. lychee_full_duplex_v1.5/model-00004-of-00006.safetensors +3 -0
  12. lychee_full_duplex_v1.5/model-00005-of-00006.safetensors +3 -0
  13. lychee_full_duplex_v1.5/model-00006-of-00006.safetensors +3 -0
  14. lychee_full_duplex_v1.5/model.safetensors.index.json +990 -0
  15. lychee_full_duplex_v1.5/modeling_step_audio_2.py +425 -0
  16. lychee_full_duplex_v1.5/special_tokens_map.json +60 -0
  17. lychee_full_duplex_v1.5/tokenizer_config.json +0 -0
  18. lychee_full_duplex_v1.5/vocab.json +0 -0
README.md CHANGED
@@ -1,3 +1,184 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ language:
4
+ - zh
5
+ - en
6
+ tags:
7
+ - speech
8
+ - audio
9
+ - full-duplex
10
+ - realtime
11
+ - speech-to-speech
12
+ - vllm
13
  ---
14
+
15
+ # Lychee-FD
16
+
17
+ Full-duplex realtime speech interaction model and Docker demo.
18
+
19
+ [![GitHub](https://img.shields.io/badge/GitHub-Lychee--FD-black?logo=github)](https://github.com/HITsz-TMG/Lychee-FD)
20
+ [![Paper](https://img.shields.io/badge/Paper-ACL%202026-red?logo=adobeacrobatreader)](https://aclanthology.org/2026.acl-long.419.pdf)
21
+ [![License](https://img.shields.io/badge/License-Apache--2.0-green)](https://github.com/HITsz-TMG/Lychee-FD/blob/main/LICENSE)
22
+
23
+ This repository hosts the Lychee-FD full-duplex model checkpoint. The Docker
24
+ runtime and demo code are maintained in the GitHub repository:
25
+
26
+ ```text
27
+ https://github.com/HITsz-TMG/Lychee-FD
28
+ ```
29
+
30
+ ## Required Weights
31
+
32
+ The Docker demo needs two weight directories:
33
+
34
+ | Component | Source | Expected local directory |
35
+ | --- | --- | --- |
36
+ | Lychee-FD full-duplex model | `HIT-TMG/Lychee-FD`, folder `lychee_full_duplex_v1.5/` | `lychee_full_duplex_v1.5/` |
37
+ | Token2Wav vocoder | `stepfun-ai/Step-Audio-2-mini`, folder `token2wav/` | `token2wav/` |
38
+
39
+ The Lychee-FD checkpoint and Token2Wav checkpoint are intentionally separated.
40
+ This model repository provides the Lychee-FD full-duplex checkpoint. Please
41
+ download the Token2Wav folder from Step-Audio-2-mini separately.
42
+
43
+ ## Download Weights
44
+
45
+ Create one local model root and put both directories under it:
46
+
47
+ ```text
48
+ /path/to/model-root/
49
+ lychee_full_duplex_v1.5/
50
+ token2wav/
51
+ ```
52
+
53
+ Download the Lychee-FD checkpoint:
54
+
55
+ ```bash
56
+ huggingface-cli download HIT-TMG/Lychee-FD \
57
+ --include "lychee_full_duplex_v1.5/*" \
58
+ --local-dir /path/to/model-root
59
+ ```
60
+
61
+ Download Token2Wav from Step-Audio-2-mini:
62
+
63
+ ```bash
64
+ huggingface-cli download stepfun-ai/Step-Audio-2-mini \
65
+ --include "token2wav/*" \
66
+ --local-dir /path/to/model-root
67
+ ```
68
+
69
+ After downloading, the directory should look like:
70
+
71
+ ```text
72
+ /path/to/model-root/
73
+ lychee_full_duplex_v1.5/
74
+ token2wav/
75
+ ```
76
+
77
+ ## Docker Quick Start
78
+
79
+ Clone the GitHub code repository:
80
+
81
+ ```bash
82
+ git clone https://github.com/HITsz-TMG/Lychee-FD.git
83
+ cd Lychee-FD
84
+ ```
85
+
86
+ Create a local Docker environment file:
87
+
88
+ ```bash
89
+ cp .env.docker.example .env
90
+ ```
91
+
92
+ Edit `.env` and replace the model paths:
93
+
94
+ ```dotenv
95
+ LYCHEE_FD_IMAGE=ghcr.io/idealistxy/lychee-fd:latest
96
+
97
+ # Host path on your machine.
98
+ HOST_MODEL_ROOT=/path/to/model-root
99
+
100
+ # Container paths. The host model root is mounted as /models.
101
+ CONTAINER_MODEL_ROOT=/models
102
+ ALLOWED_MODEL_ROOT=/models
103
+
104
+ # Main Lychee-FD checkpoint.
105
+ STEPAUDIO_MODEL_PATH=/models/lychee_full_duplex_v1.5
106
+
107
+ # Token2Wav checkpoint downloaded from stepfun-ai/Step-Audio-2-mini.
108
+ STEPAUDIO_T2W_MODEL_PATH=/models/token2wav
109
+ ```
110
+
111
+ Important path rule:
112
+
113
+ - `HOST_MODEL_ROOT` is the path on your host machine.
114
+ - `STEPAUDIO_MODEL_PATH` and `STEPAUDIO_T2W_MODEL_PATH` are paths inside the
115
+ container.
116
+ - By default, Docker mounts `HOST_MODEL_ROOT` to `/models`, so model paths inside
117
+ `.env` should usually start with `/models`.
118
+
119
+ Start the demo:
120
+
121
+ ```bash
122
+ docker compose pull
123
+ docker compose up
124
+ ```
125
+
126
+ Open the frontend:
127
+
128
+ ```text
129
+ http://127.0.0.1:8084
130
+ ```
131
+
132
+ For a remote server, replace `127.0.0.1` with the server IP.
133
+
134
+ ## GPU Settings
135
+
136
+ By default, Token2Wav and the main backend use separate GPUs:
137
+
138
+ ```dotenv
139
+ TOKEN2WAV_CUDA_VISIBLE_DEVICES=0
140
+ BACKEND_CUDA_VISIBLE_DEVICES=1
141
+ ```
142
+
143
+ For a single-GPU machine, set both to `0`:
144
+
145
+ ```dotenv
146
+ TOKEN2WAV_CUDA_VISIBLE_DEVICES=0
147
+ BACKEND_CUDA_VISIBLE_DEVICES=0
148
+ ```
149
+
150
+ If GPU memory is limited, reduce the vLLM maximum context length:
151
+
152
+ ```dotenv
153
+ STEPAUDIO_VLLM_MAX_MODEL_LEN=8192
154
+ ```
155
+
156
+ ## Model Presets
157
+
158
+ The frontend model list is loaded from `model_presets_dev.json` in the GitHub
159
+ repository. If you customize the model list, make sure the preset path uses the
160
+ container path:
161
+
162
+ ```json
163
+ {
164
+ "name": "lychee_full_duplex_v1.5",
165
+ "model_path": "/models/lychee_full_duplex_v1.5",
166
+ "backend_type": "vllm",
167
+ "mode": "stable"
168
+ }
169
+ ```
170
+
171
+ After editing presets:
172
+
173
+ ```bash
174
+ docker compose restart frontend
175
+ ```
176
+
177
+ ## Notes
178
+
179
+ - The Docker image contains the runtime environment and demo code, but it does
180
+ not include model weights.
181
+ - The Token2Wav checkpoint is provided by `stepfun-ai/Step-Audio-2-mini`; please
182
+ follow the license and usage terms of the upstream model repository.
183
+ - For source code, Docker Compose files, and detailed serving instructions, see
184
+ the GitHub repository: `https://github.com/HITsz-TMG/Lychee-FD`.
lychee_full_duplex_v1.5/added_tokens.json ADDED
The diff for this file is too large to render. See raw diff
 
lychee_full_duplex_v1.5/chat_template.jinja ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0]['role'] == 'system' %}
4
+ {{- messages[0]['content'] }}
5
+ {%- else %}
6
+ {{- 'You are a helpful assistant.' }}
7
+ {%- endif %}
8
+ {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
9
+ {%- for tool in tools %}
10
+ {{- "\n" }}
11
+ {{- tool | tojson }}
12
+ {%- endfor %}
13
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
14
+ {%- else %}
15
+ {%- if messages[0]['role'] == 'system' %}
16
+ {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
17
+ {%- else %}
18
+ {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
19
+ {%- endif %}
20
+ {%- endif %}
21
+ {%- for message in messages %}
22
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
23
+ {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
24
+ {%- elif message.role == "assistant" %}
25
+ {{- '<|im_start|>' + message.role }}
26
+ {%- if message.content %}
27
+ {{- '\n' + message.content }}
28
+ {%- endif %}
29
+ {%- for tool_call in message.tool_calls %}
30
+ {%- if tool_call.function is defined %}
31
+ {%- set tool_call = tool_call.function %}
32
+ {%- endif %}
33
+ {{- '\n<tool_call>\n{"name": "' }}
34
+ {{- tool_call.name }}
35
+ {{- '", "arguments": ' }}
36
+ {{- tool_call.arguments | tojson }}
37
+ {{- '}\n</tool_call>' }}
38
+ {%- endfor %}
39
+ {{- '<|im_end|>\n' }}
40
+ {%- elif message.role == "tool" %}
41
+ {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
42
+ {{- '<|im_start|>user' }}
43
+ {%- endif %}
44
+ {{- '\n<tool_response>\n' }}
45
+ {{- message.content }}
46
+ {{- '\n</tool_response>' }}
47
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
48
+ {{- '<|im_end|>\n' }}
49
+ {%- endif %}
50
+ {%- endif %}
51
+ {%- endfor %}
52
+ {%- if add_generation_prompt %}
53
+ {{- '<|im_start|>assistant\n' }}
54
+ {%- endif %}
lychee_full_duplex_v1.5/config.json ADDED
@@ -0,0 +1,375 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adding_text_hiddenstates": false,
3
+ "architectures": [
4
+ "StepAudio2FullDuplex"
5
+ ],
6
+ "audio_encoder_config": {
7
+ "adapter_stride": 2,
8
+ "kernel_size": 3,
9
+ "llm_dim": 3584,
10
+ "model_type": "step_audio_2_encoder",
11
+ "n_audio_ctx": 1500,
12
+ "n_audio_head": 20,
13
+ "n_audio_layer": 32,
14
+ "n_audio_state": 1280,
15
+ "n_codebook_size": 4096,
16
+ "n_mels": 128
17
+ },
18
+ "audio_pad_token_id": 158360,
19
+ "auto_map": {
20
+ "AutoConfig": "configuration_step_audio_2.StepAudio2Config",
21
+ "AutoModelForCausalLM": "modeling_step_audio_2.StepAudio2ForCausalLM"
22
+ },
23
+ "control_branch_layer": 8,
24
+ "control_layer_config": {
25
+ "_name_or_path": "",
26
+ "add_cross_attention": false,
27
+ "architectures": [
28
+ "Qwen2ForCausalLM"
29
+ ],
30
+ "attention_dropout": 0.0,
31
+ "bad_words_ids": null,
32
+ "begin_suppress_tokens": null,
33
+ "bos_token_id": null,
34
+ "chunk_size_feed_forward": 0,
35
+ "cross_attention_hidden_size": null,
36
+ "decoder_start_token_id": null,
37
+ "diversity_penalty": 0.0,
38
+ "do_sample": false,
39
+ "early_stopping": false,
40
+ "encoder_no_repeat_ngram_size": 0,
41
+ "eos_token_id": null,
42
+ "exponential_decay_length_penalty": null,
43
+ "finetuning_task": null,
44
+ "forced_bos_token_id": null,
45
+ "forced_eos_token_id": null,
46
+ "hidden_act": "silu",
47
+ "hidden_size": 3584,
48
+ "id2label": {
49
+ "0": "LABEL_0",
50
+ "1": "LABEL_1"
51
+ },
52
+ "initializer_range": 0.02,
53
+ "intermediate_size": 18944,
54
+ "is_decoder": false,
55
+ "is_encoder_decoder": false,
56
+ "label2id": {
57
+ "LABEL_0": 0,
58
+ "LABEL_1": 1
59
+ },
60
+ "layer_types": [
61
+ "full_attention",
62
+ "full_attention",
63
+ "full_attention",
64
+ "full_attention"
65
+ ],
66
+ "length_penalty": 1.0,
67
+ "max_length": 20,
68
+ "max_position_embeddings": 16384,
69
+ "max_window_layers": 28,
70
+ "min_length": 0,
71
+ "model_type": "qwen2",
72
+ "no_repeat_ngram_size": 0,
73
+ "num_attention_groups": 4,
74
+ "num_attention_heads": 28,
75
+ "num_beam_groups": 1,
76
+ "num_beams": 1,
77
+ "num_hidden_layers": 4,
78
+ "num_key_value_heads": 4,
79
+ "num_return_sequences": 1,
80
+ "output_attentions": false,
81
+ "output_hidden_states": false,
82
+ "output_scores": false,
83
+ "pad_token_id": null,
84
+ "prefix": null,
85
+ "problem_type": null,
86
+ "pruned_heads": {},
87
+ "remove_invalid_values": false,
88
+ "repetition_penalty": 1.0,
89
+ "return_dict": true,
90
+ "return_dict_in_generate": false,
91
+ "rms_norm_eps": 1e-06,
92
+ "rope_scaling": null,
93
+ "rope_theta": 1000000.0,
94
+ "sep_token_id": null,
95
+ "sliding_window": null,
96
+ "suppress_tokens": null,
97
+ "task_specific_params": null,
98
+ "temperature": 1.0,
99
+ "tf_legacy_loss": false,
100
+ "tie_encoder_decoder": false,
101
+ "tie_word_embeddings": false,
102
+ "tokenizer_class": null,
103
+ "top_k": 50,
104
+ "top_p": 1.0,
105
+ "torch_dtype": "bfloat16",
106
+ "torchscript": false,
107
+ "typical_p": 1.0,
108
+ "use_bfloat16": false,
109
+ "use_cache": true,
110
+ "use_sliding_window": false,
111
+ "vocab_size": 158363
112
+ },
113
+ "control_token_chunk_size": 10,
114
+ "control_token_focal_loss_beta": 2.0,
115
+ "control_token_ids_max": 158356,
116
+ "control_token_ids_min": 158352,
117
+ "control_token_use_focal_loss": false,
118
+ "detect_token_id": 158356,
119
+ "end_bc_token_id": 158353,
120
+ "keep_bc_token_id": 158362,
121
+ "keep_listening_token_id": 158354,
122
+ "keep_speaking_token_id": 158355,
123
+ "max_window_layers": null,
124
+ "merge_layer_config": {
125
+ "_name_or_path": "",
126
+ "add_cross_attention": false,
127
+ "architectures": [
128
+ "Qwen2ForCausalLM"
129
+ ],
130
+ "attention_dropout": 0.0,
131
+ "bad_words_ids": null,
132
+ "begin_suppress_tokens": null,
133
+ "bos_token_id": null,
134
+ "chunk_size_feed_forward": 0,
135
+ "cross_attention_hidden_size": null,
136
+ "decoder_start_token_id": null,
137
+ "diversity_penalty": 0.0,
138
+ "do_sample": false,
139
+ "early_stopping": false,
140
+ "encoder_no_repeat_ngram_size": 0,
141
+ "eos_token_id": null,
142
+ "exponential_decay_length_penalty": null,
143
+ "finetuning_task": null,
144
+ "forced_bos_token_id": null,
145
+ "forced_eos_token_id": null,
146
+ "hidden_act": "silu",
147
+ "hidden_size": 3584,
148
+ "id2label": {
149
+ "0": "LABEL_0",
150
+ "1": "LABEL_1"
151
+ },
152
+ "initializer_range": 0.02,
153
+ "intermediate_size": 18944,
154
+ "is_decoder": false,
155
+ "is_encoder_decoder": false,
156
+ "label2id": {
157
+ "LABEL_0": 0,
158
+ "LABEL_1": 1
159
+ },
160
+ "layer_types": [
161
+ "full_attention",
162
+ "full_attention",
163
+ "full_attention",
164
+ "full_attention"
165
+ ],
166
+ "length_penalty": 1.0,
167
+ "max_length": 20,
168
+ "max_position_embeddings": 16384,
169
+ "max_window_layers": 28,
170
+ "min_length": 0,
171
+ "model_type": "qwen2",
172
+ "no_repeat_ngram_size": 0,
173
+ "num_attention_groups": 4,
174
+ "num_attention_heads": 28,
175
+ "num_beam_groups": 1,
176
+ "num_beams": 1,
177
+ "num_hidden_layers": 4,
178
+ "num_key_value_heads": 4,
179
+ "num_return_sequences": 1,
180
+ "output_attentions": false,
181
+ "output_hidden_states": false,
182
+ "output_scores": false,
183
+ "pad_token_id": null,
184
+ "prefix": null,
185
+ "problem_type": null,
186
+ "pruned_heads": {},
187
+ "remove_invalid_values": false,
188
+ "repetition_penalty": 1.0,
189
+ "return_dict": true,
190
+ "return_dict_in_generate": false,
191
+ "rms_norm_eps": 1e-06,
192
+ "rope_scaling": null,
193
+ "rope_theta": 1000000.0,
194
+ "sep_token_id": null,
195
+ "sliding_window": null,
196
+ "suppress_tokens": null,
197
+ "task_specific_params": null,
198
+ "temperature": 1.0,
199
+ "tf_legacy_loss": false,
200
+ "tie_encoder_decoder": false,
201
+ "tie_word_embeddings": false,
202
+ "tokenizer_class": null,
203
+ "top_k": 50,
204
+ "top_p": 1.0,
205
+ "torch_dtype": "bfloat16",
206
+ "torchscript": false,
207
+ "typical_p": 1.0,
208
+ "use_bfloat16": false,
209
+ "use_cache": true,
210
+ "use_sliding_window": false,
211
+ "vocab_size": 158363
212
+ },
213
+ "model_type": "step_audio_2_full_duplex",
214
+ "no_stoken_label": false,
215
+ "no_text_label": false,
216
+ "sleep_token_id": 158357,
217
+ "sliding_window": 2048,
218
+ "start_bc_token_id": 158362,
219
+ "start_listening_token_id": 158353,
220
+ "start_speaking_token_id": 158352,
221
+ "stoken_delay_num": 10,
222
+ "stoken_delay_token_id": 158361,
223
+ "stoken_layer_config": {
224
+ "_name_or_path": "",
225
+ "add_cross_attention": false,
226
+ "architectures": [
227
+ "Qwen2ForCausalLM"
228
+ ],
229
+ "attention_dropout": 0.0,
230
+ "bad_words_ids": null,
231
+ "begin_suppress_tokens": null,
232
+ "bos_token_id": null,
233
+ "chunk_size_feed_forward": 0,
234
+ "cross_attention_hidden_size": null,
235
+ "decoder_start_token_id": null,
236
+ "diversity_penalty": 0.0,
237
+ "do_sample": false,
238
+ "early_stopping": false,
239
+ "encoder_no_repeat_ngram_size": 0,
240
+ "eos_token_id": null,
241
+ "exponential_decay_length_penalty": null,
242
+ "finetuning_task": null,
243
+ "forced_bos_token_id": null,
244
+ "forced_eos_token_id": null,
245
+ "hidden_act": "silu",
246
+ "hidden_size": 3584,
247
+ "id2label": {
248
+ "0": "LABEL_0",
249
+ "1": "LABEL_1"
250
+ },
251
+ "initializer_range": 0.02,
252
+ "intermediate_size": 18944,
253
+ "is_decoder": false,
254
+ "is_encoder_decoder": false,
255
+ "label2id": {
256
+ "LABEL_0": 0,
257
+ "LABEL_1": 1
258
+ },
259
+ "layer_types": [
260
+ "full_attention",
261
+ "full_attention",
262
+ "full_attention",
263
+ "full_attention"
264
+ ],
265
+ "length_penalty": 1.0,
266
+ "max_length": 20,
267
+ "max_position_embeddings": 16384,
268
+ "max_window_layers": 28,
269
+ "min_length": 0,
270
+ "model_type": "qwen2",
271
+ "no_repeat_ngram_size": 0,
272
+ "num_attention_groups": 4,
273
+ "num_attention_heads": 28,
274
+ "num_beam_groups": 1,
275
+ "num_beams": 1,
276
+ "num_hidden_layers": 4,
277
+ "num_key_value_heads": 4,
278
+ "num_return_sequences": 1,
279
+ "output_attentions": false,
280
+ "output_hidden_states": false,
281
+ "output_scores": false,
282
+ "pad_token_id": null,
283
+ "prefix": null,
284
+ "problem_type": null,
285
+ "pruned_heads": {},
286
+ "remove_invalid_values": false,
287
+ "repetition_penalty": 1.0,
288
+ "return_dict": true,
289
+ "return_dict_in_generate": false,
290
+ "rms_norm_eps": 1e-06,
291
+ "rope_scaling": null,
292
+ "rope_theta": 1000000.0,
293
+ "sep_token_id": null,
294
+ "sliding_window": null,
295
+ "suppress_tokens": null,
296
+ "task_specific_params": null,
297
+ "temperature": 1.0,
298
+ "tf_legacy_loss": false,
299
+ "tie_encoder_decoder": false,
300
+ "tie_word_embeddings": false,
301
+ "tokenizer_class": null,
302
+ "top_k": 50,
303
+ "top_p": 1.0,
304
+ "torch_dtype": "bfloat16",
305
+ "torchscript": false,
306
+ "typical_p": 1.0,
307
+ "use_bfloat16": false,
308
+ "use_cache": true,
309
+ "use_sliding_window": false,
310
+ "vocab_size": 158363
311
+ },
312
+ "stoken_pad_token_id": 158359,
313
+ "stoken_token_ids_max": 158352,
314
+ "stoken_token_ids_min": 151694,
315
+ "text_config": {
316
+ "architectures": [
317
+ "Qwen2ForCausalLM"
318
+ ],
319
+ "attention_dropout": 0.0,
320
+ "hidden_act": "silu",
321
+ "hidden_size": 3584,
322
+ "initializer_range": 0.02,
323
+ "intermediate_size": 18944,
324
+ "layer_types": [
325
+ "full_attention",
326
+ "full_attention",
327
+ "full_attention",
328
+ "full_attention",
329
+ "full_attention",
330
+ "full_attention",
331
+ "full_attention",
332
+ "full_attention",
333
+ "full_attention",
334
+ "full_attention",
335
+ "full_attention",
336
+ "full_attention",
337
+ "full_attention",
338
+ "full_attention",
339
+ "full_attention",
340
+ "full_attention",
341
+ "full_attention",
342
+ "full_attention",
343
+ "full_attention",
344
+ "full_attention",
345
+ "full_attention",
346
+ "full_attention",
347
+ "full_attention",
348
+ "full_attention",
349
+ "full_attention",
350
+ "full_attention",
351
+ "full_attention",
352
+ "full_attention"
353
+ ],
354
+ "max_position_embeddings": 16384,
355
+ "max_window_layers": 28,
356
+ "model_type": "qwen2",
357
+ "num_attention_groups": 4,
358
+ "num_attention_heads": 28,
359
+ "num_hidden_layers": 28,
360
+ "num_key_value_heads": 4,
361
+ "rms_norm_eps": 1e-06,
362
+ "rope_scaling": null,
363
+ "rope_theta": 1000000.0,
364
+ "sliding_window": null,
365
+ "torch_dtype": "bfloat16",
366
+ "use_cache": true,
367
+ "use_sliding_window": false,
368
+ "vocab_size": 158363
369
+ },
370
+ "text_pad_token_id": 158358,
371
+ "text_sleep_token_id": null,
372
+ "torch_dtype": "bfloat16",
373
+ "transformers_version": "4.53.1",
374
+ "use_sliding_window": false
375
+ }
lychee_full_duplex_v1.5/configuration_step_audio_2.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Union
2
+
3
+ from transformers import Qwen2Config
4
+ from transformers.configuration_utils import PretrainedConfig
5
+
6
+
7
+ class StepAudio2EncoderConfig(PretrainedConfig):
8
+ model_type = "step_audio_2_encoder"
9
+
10
+ def __init__(
11
+ self,
12
+ n_mels=128,
13
+ n_audio_ctx=1500,
14
+ n_audio_state=512,
15
+ n_audio_head=8,
16
+ n_audio_layer=6,
17
+ llm_dim=4096,
18
+ kernel_size=3,
19
+ adapter_stride=2,
20
+ **kwargs,
21
+ ):
22
+ self.n_mels = n_mels
23
+ self.n_audio_ctx = n_audio_ctx
24
+ self.n_audio_state = n_audio_state
25
+ self.n_audio_head = n_audio_head
26
+ self.n_audio_layer = n_audio_layer
27
+ self.llm_dim = llm_dim
28
+ self.kernel_size = kernel_size
29
+ self.adapter_stride = adapter_stride
30
+ super().__init__(**kwargs)
31
+
32
+ class StepAudio2TextConfig(PretrainedConfig):
33
+ model_type = "step_audio_2_text"
34
+
35
+ def __init__(
36
+ self,
37
+ vocab_size=64012,
38
+ hidden_size=4096,
39
+ intermediate_size=11008,
40
+ num_hidden_layers=48,
41
+ num_attention_heads=32,
42
+ num_attention_groups=4,
43
+ num_key_value_heads=4,
44
+ hidden_act="silu",
45
+ max_position_embeddings=8192,
46
+ initializer_range=0.02,
47
+ rms_norm_eps=1e-6,
48
+ rope_theta=1000000.0,
49
+ rope_scaling=None,
50
+ eos_token_id=None,
51
+ **kwargs
52
+ ):
53
+
54
+ if eos_token_id is not None:
55
+ if isinstance(eos_token_id, list):
56
+ eos_token_id = list(set([151643, 151645, 151665] + eos_token_id))
57
+ else:
58
+ eos_token_id = [151643, 151645, 151665, eos_token_id]
59
+ else:
60
+ eos_token_id = [151643, 151645, 151665]
61
+
62
+ super().__init__(
63
+ eos_token_id=eos_token_id,
64
+ **kwargs)
65
+
66
+ self.vocab_size = vocab_size
67
+ self.hidden_size = hidden_size
68
+ self.intermediate_size = intermediate_size
69
+ self.num_hidden_layers = num_hidden_layers
70
+ self.num_attention_heads = num_attention_heads
71
+ self.num_attention_groups = num_attention_groups
72
+ self.num_key_value_heads = num_key_value_heads
73
+ assert self.num_attention_groups == self.num_key_value_heads, "num_attention_groups must be equal to num_key_value_heads"
74
+ self.hidden_act = hidden_act
75
+ self.max_position_embeddings = max_position_embeddings
76
+ self.initializer_range = initializer_range
77
+ self.rms_norm_eps = rms_norm_eps
78
+ self.rope_theta = rope_theta
79
+ self.rope_scaling = rope_scaling
80
+
81
+ self.text_config = Qwen2Config(
82
+ vocab_size=vocab_size,
83
+ hidden_size=hidden_size,
84
+ intermediate_size=intermediate_size,
85
+ num_hidden_layers=num_hidden_layers,
86
+ num_attention_heads=num_attention_heads,
87
+ num_key_value_heads=num_key_value_heads,
88
+ hidden_act=hidden_act,
89
+ max_position_embeddings=max_position_embeddings,
90
+ initializer_range=initializer_range,
91
+ rms_norm_eps=rms_norm_eps,
92
+ rope_theta=rope_theta,
93
+ rope_scaling=rope_scaling,
94
+ architectures=["Qwen2ForCausalLM"],
95
+ torch_dtype=getattr(self, "torch_dtype", "bfloat16"),
96
+ )
97
+
98
+ class StepAudio2Config(PretrainedConfig):
99
+ model_type = "step_audio_2"
100
+ architectures = ["StepAudio2ForCausalLM"]
101
+
102
+ def __init__(
103
+ self,
104
+ audio_encoder_config :Optional[Union[dict, StepAudio2EncoderConfig]] = None,
105
+ text_config: Optional[Union[dict, StepAudio2TextConfig]] = None,
106
+ use_sliding_window: bool = False,
107
+ sliding_window: Optional[int] = 2048,
108
+ max_window_layers: Optional[int] = None,
109
+ **kwargs
110
+ ):
111
+ kwargs.setdefault("use_sliding_window", use_sliding_window)
112
+ kwargs.setdefault("sliding_window", sliding_window)
113
+ if max_window_layers is None:
114
+ max_window_layers = kwargs.get("num_hidden_layers", None)
115
+ kwargs.setdefault("max_window_layers", max_window_layers)
116
+ super().__init__(**kwargs)
117
+
118
+ if text_config is None:
119
+ text_config = StepAudio2TextConfig().text_config
120
+ elif isinstance(text_config, dict):
121
+ text_config = StepAudio2TextConfig(**text_config).text_config
122
+
123
+ self.text_config = text_config
124
+
125
+ if audio_encoder_config is None:
126
+ self.audio_encoder_config = StepAudio2EncoderConfig()
127
+ elif isinstance(audio_encoder_config, dict):
128
+ self.audio_encoder_config = StepAudio2EncoderConfig(**audio_encoder_config)
lychee_full_duplex_v1.5/generation_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "transformers_version": "4.53.1"
4
+ }
lychee_full_duplex_v1.5/merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
lychee_full_duplex_v1.5/model-00001-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1a7783ef988d3f617c2f38d4f397c0090f06dc064f17ae33c99110d480bc1597
3
+ size 4922812008
lychee_full_duplex_v1.5/model-00002-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a25a737023da1ac3f1e9e893873c8ad86e550be8dcaef6f21db157748ea82a9
3
+ size 4932751008
lychee_full_duplex_v1.5/model-00003-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2f03a2753a20ba27f29f310857ea8ec5c0ee3242298877cb7388b543a8b017ac
3
+ size 4330865200
lychee_full_duplex_v1.5/model-00004-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:378dcebcc0c5b9296e89dd15eb45947cb8e5379e87587df37aa380c2b303543b
3
+ size 4931196440
lychee_full_duplex_v1.5/model-00005-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c53dce8bcda72e94b2c997b7185965f3b2a424aa8f064a85e279507cafc13821
3
+ size 4997303504
lychee_full_duplex_v1.5/model-00006-of-00006.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9946c7c66ecee67050ba064a9619ea3e7e889bb256a8d20d41ad26bdde2d048
3
+ size 1509268128
lychee_full_duplex_v1.5/model.safetensors.index.json ADDED
@@ -0,0 +1,990 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_parameters": 12812043520,
4
+ "total_size": 25624087040
5
+ },
6
+ "weight_map": {
7
+ "adapter.conv.bias": "model-00006-of-00006.safetensors",
8
+ "adapter.conv.weight": "model-00006-of-00006.safetensors",
9
+ "adapter.linear1.bias": "model-00006-of-00006.safetensors",
10
+ "adapter.linear1.weight": "model-00006-of-00006.safetensors",
11
+ "adapter.linear2.bias": "model-00006-of-00006.safetensors",
12
+ "adapter.linear2.weight": "model-00006-of-00006.safetensors",
13
+ "control_model.embed_tokens.weight": "model-00004-of-00006.safetensors",
14
+ "control_model.layers.0.input_layernorm.weight": "model-00004-of-00006.safetensors",
15
+ "control_model.layers.0.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
16
+ "control_model.layers.0.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
17
+ "control_model.layers.0.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
18
+ "control_model.layers.0.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
19
+ "control_model.layers.0.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
20
+ "control_model.layers.0.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
21
+ "control_model.layers.0.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
22
+ "control_model.layers.0.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
23
+ "control_model.layers.0.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
24
+ "control_model.layers.0.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
25
+ "control_model.layers.0.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
26
+ "control_model.layers.1.input_layernorm.weight": "model-00005-of-00006.safetensors",
27
+ "control_model.layers.1.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
28
+ "control_model.layers.1.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
29
+ "control_model.layers.1.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
30
+ "control_model.layers.1.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
31
+ "control_model.layers.1.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
32
+ "control_model.layers.1.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
33
+ "control_model.layers.1.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
34
+ "control_model.layers.1.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
35
+ "control_model.layers.1.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
36
+ "control_model.layers.1.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
37
+ "control_model.layers.1.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
38
+ "control_model.layers.2.input_layernorm.weight": "model-00005-of-00006.safetensors",
39
+ "control_model.layers.2.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
40
+ "control_model.layers.2.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
41
+ "control_model.layers.2.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
42
+ "control_model.layers.2.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
43
+ "control_model.layers.2.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
44
+ "control_model.layers.2.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
45
+ "control_model.layers.2.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
46
+ "control_model.layers.2.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
47
+ "control_model.layers.2.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
48
+ "control_model.layers.2.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
49
+ "control_model.layers.2.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
50
+ "control_model.layers.3.input_layernorm.weight": "model-00005-of-00006.safetensors",
51
+ "control_model.layers.3.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
52
+ "control_model.layers.3.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
53
+ "control_model.layers.3.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
54
+ "control_model.layers.3.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
55
+ "control_model.layers.3.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
56
+ "control_model.layers.3.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
57
+ "control_model.layers.3.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
58
+ "control_model.layers.3.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
59
+ "control_model.layers.3.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
60
+ "control_model.layers.3.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
61
+ "control_model.layers.3.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
62
+ "control_model.norm.weight": "model-00005-of-00006.safetensors",
63
+ "encoder.after_norm.bias": "model-00006-of-00006.safetensors",
64
+ "encoder.after_norm.weight": "model-00006-of-00006.safetensors",
65
+ "encoder.blocks.0.attn.key.weight": "model-00005-of-00006.safetensors",
66
+ "encoder.blocks.0.attn.out.bias": "model-00005-of-00006.safetensors",
67
+ "encoder.blocks.0.attn.out.weight": "model-00005-of-00006.safetensors",
68
+ "encoder.blocks.0.attn.query.bias": "model-00005-of-00006.safetensors",
69
+ "encoder.blocks.0.attn.query.weight": "model-00005-of-00006.safetensors",
70
+ "encoder.blocks.0.attn.value.bias": "model-00005-of-00006.safetensors",
71
+ "encoder.blocks.0.attn.value.weight": "model-00005-of-00006.safetensors",
72
+ "encoder.blocks.0.attn_ln.bias": "model-00005-of-00006.safetensors",
73
+ "encoder.blocks.0.attn_ln.weight": "model-00005-of-00006.safetensors",
74
+ "encoder.blocks.0.mlp.0.bias": "model-00005-of-00006.safetensors",
75
+ "encoder.blocks.0.mlp.0.weight": "model-00005-of-00006.safetensors",
76
+ "encoder.blocks.0.mlp.2.bias": "model-00005-of-00006.safetensors",
77
+ "encoder.blocks.0.mlp.2.weight": "model-00005-of-00006.safetensors",
78
+ "encoder.blocks.0.mlp_ln.bias": "model-00005-of-00006.safetensors",
79
+ "encoder.blocks.0.mlp_ln.weight": "model-00005-of-00006.safetensors",
80
+ "encoder.blocks.1.attn.key.weight": "model-00005-of-00006.safetensors",
81
+ "encoder.blocks.1.attn.out.bias": "model-00005-of-00006.safetensors",
82
+ "encoder.blocks.1.attn.out.weight": "model-00005-of-00006.safetensors",
83
+ "encoder.blocks.1.attn.query.bias": "model-00005-of-00006.safetensors",
84
+ "encoder.blocks.1.attn.query.weight": "model-00005-of-00006.safetensors",
85
+ "encoder.blocks.1.attn.value.bias": "model-00005-of-00006.safetensors",
86
+ "encoder.blocks.1.attn.value.weight": "model-00005-of-00006.safetensors",
87
+ "encoder.blocks.1.attn_ln.bias": "model-00005-of-00006.safetensors",
88
+ "encoder.blocks.1.attn_ln.weight": "model-00005-of-00006.safetensors",
89
+ "encoder.blocks.1.mlp.0.bias": "model-00005-of-00006.safetensors",
90
+ "encoder.blocks.1.mlp.0.weight": "model-00005-of-00006.safetensors",
91
+ "encoder.blocks.1.mlp.2.bias": "model-00005-of-00006.safetensors",
92
+ "encoder.blocks.1.mlp.2.weight": "model-00005-of-00006.safetensors",
93
+ "encoder.blocks.1.mlp_ln.bias": "model-00005-of-00006.safetensors",
94
+ "encoder.blocks.1.mlp_ln.weight": "model-00005-of-00006.safetensors",
95
+ "encoder.blocks.10.attn.key.weight": "model-00005-of-00006.safetensors",
96
+ "encoder.blocks.10.attn.out.bias": "model-00005-of-00006.safetensors",
97
+ "encoder.blocks.10.attn.out.weight": "model-00005-of-00006.safetensors",
98
+ "encoder.blocks.10.attn.query.bias": "model-00005-of-00006.safetensors",
99
+ "encoder.blocks.10.attn.query.weight": "model-00005-of-00006.safetensors",
100
+ "encoder.blocks.10.attn.value.bias": "model-00005-of-00006.safetensors",
101
+ "encoder.blocks.10.attn.value.weight": "model-00005-of-00006.safetensors",
102
+ "encoder.blocks.10.attn_ln.bias": "model-00005-of-00006.safetensors",
103
+ "encoder.blocks.10.attn_ln.weight": "model-00005-of-00006.safetensors",
104
+ "encoder.blocks.10.mlp.0.bias": "model-00005-of-00006.safetensors",
105
+ "encoder.blocks.10.mlp.0.weight": "model-00005-of-00006.safetensors",
106
+ "encoder.blocks.10.mlp.2.bias": "model-00005-of-00006.safetensors",
107
+ "encoder.blocks.10.mlp.2.weight": "model-00005-of-00006.safetensors",
108
+ "encoder.blocks.10.mlp_ln.bias": "model-00005-of-00006.safetensors",
109
+ "encoder.blocks.10.mlp_ln.weight": "model-00005-of-00006.safetensors",
110
+ "encoder.blocks.11.attn.key.weight": "model-00005-of-00006.safetensors",
111
+ "encoder.blocks.11.attn.out.bias": "model-00005-of-00006.safetensors",
112
+ "encoder.blocks.11.attn.out.weight": "model-00005-of-00006.safetensors",
113
+ "encoder.blocks.11.attn.query.bias": "model-00005-of-00006.safetensors",
114
+ "encoder.blocks.11.attn.query.weight": "model-00005-of-00006.safetensors",
115
+ "encoder.blocks.11.attn.value.bias": "model-00005-of-00006.safetensors",
116
+ "encoder.blocks.11.attn.value.weight": "model-00005-of-00006.safetensors",
117
+ "encoder.blocks.11.attn_ln.bias": "model-00005-of-00006.safetensors",
118
+ "encoder.blocks.11.attn_ln.weight": "model-00005-of-00006.safetensors",
119
+ "encoder.blocks.11.mlp.0.bias": "model-00005-of-00006.safetensors",
120
+ "encoder.blocks.11.mlp.0.weight": "model-00005-of-00006.safetensors",
121
+ "encoder.blocks.11.mlp.2.bias": "model-00005-of-00006.safetensors",
122
+ "encoder.blocks.11.mlp.2.weight": "model-00005-of-00006.safetensors",
123
+ "encoder.blocks.11.mlp_ln.bias": "model-00005-of-00006.safetensors",
124
+ "encoder.blocks.11.mlp_ln.weight": "model-00005-of-00006.safetensors",
125
+ "encoder.blocks.12.attn.key.weight": "model-00005-of-00006.safetensors",
126
+ "encoder.blocks.12.attn.out.bias": "model-00005-of-00006.safetensors",
127
+ "encoder.blocks.12.attn.out.weight": "model-00005-of-00006.safetensors",
128
+ "encoder.blocks.12.attn.query.bias": "model-00005-of-00006.safetensors",
129
+ "encoder.blocks.12.attn.query.weight": "model-00005-of-00006.safetensors",
130
+ "encoder.blocks.12.attn.value.bias": "model-00005-of-00006.safetensors",
131
+ "encoder.blocks.12.attn.value.weight": "model-00005-of-00006.safetensors",
132
+ "encoder.blocks.12.attn_ln.bias": "model-00005-of-00006.safetensors",
133
+ "encoder.blocks.12.attn_ln.weight": "model-00005-of-00006.safetensors",
134
+ "encoder.blocks.12.mlp.0.bias": "model-00005-of-00006.safetensors",
135
+ "encoder.blocks.12.mlp.0.weight": "model-00005-of-00006.safetensors",
136
+ "encoder.blocks.12.mlp.2.bias": "model-00005-of-00006.safetensors",
137
+ "encoder.blocks.12.mlp.2.weight": "model-00005-of-00006.safetensors",
138
+ "encoder.blocks.12.mlp_ln.bias": "model-00005-of-00006.safetensors",
139
+ "encoder.blocks.12.mlp_ln.weight": "model-00005-of-00006.safetensors",
140
+ "encoder.blocks.13.attn.key.weight": "model-00005-of-00006.safetensors",
141
+ "encoder.blocks.13.attn.out.bias": "model-00005-of-00006.safetensors",
142
+ "encoder.blocks.13.attn.out.weight": "model-00005-of-00006.safetensors",
143
+ "encoder.blocks.13.attn.query.bias": "model-00005-of-00006.safetensors",
144
+ "encoder.blocks.13.attn.query.weight": "model-00005-of-00006.safetensors",
145
+ "encoder.blocks.13.attn.value.bias": "model-00005-of-00006.safetensors",
146
+ "encoder.blocks.13.attn.value.weight": "model-00005-of-00006.safetensors",
147
+ "encoder.blocks.13.attn_ln.bias": "model-00005-of-00006.safetensors",
148
+ "encoder.blocks.13.attn_ln.weight": "model-00005-of-00006.safetensors",
149
+ "encoder.blocks.13.mlp.0.bias": "model-00005-of-00006.safetensors",
150
+ "encoder.blocks.13.mlp.0.weight": "model-00005-of-00006.safetensors",
151
+ "encoder.blocks.13.mlp.2.bias": "model-00005-of-00006.safetensors",
152
+ "encoder.blocks.13.mlp.2.weight": "model-00005-of-00006.safetensors",
153
+ "encoder.blocks.13.mlp_ln.bias": "model-00005-of-00006.safetensors",
154
+ "encoder.blocks.13.mlp_ln.weight": "model-00005-of-00006.safetensors",
155
+ "encoder.blocks.14.attn.key.weight": "model-00005-of-00006.safetensors",
156
+ "encoder.blocks.14.attn.out.bias": "model-00005-of-00006.safetensors",
157
+ "encoder.blocks.14.attn.out.weight": "model-00005-of-00006.safetensors",
158
+ "encoder.blocks.14.attn.query.bias": "model-00005-of-00006.safetensors",
159
+ "encoder.blocks.14.attn.query.weight": "model-00005-of-00006.safetensors",
160
+ "encoder.blocks.14.attn.value.bias": "model-00005-of-00006.safetensors",
161
+ "encoder.blocks.14.attn.value.weight": "model-00005-of-00006.safetensors",
162
+ "encoder.blocks.14.attn_ln.bias": "model-00005-of-00006.safetensors",
163
+ "encoder.blocks.14.attn_ln.weight": "model-00005-of-00006.safetensors",
164
+ "encoder.blocks.14.mlp.0.bias": "model-00005-of-00006.safetensors",
165
+ "encoder.blocks.14.mlp.0.weight": "model-00005-of-00006.safetensors",
166
+ "encoder.blocks.14.mlp.2.bias": "model-00005-of-00006.safetensors",
167
+ "encoder.blocks.14.mlp.2.weight": "model-00005-of-00006.safetensors",
168
+ "encoder.blocks.14.mlp_ln.bias": "model-00005-of-00006.safetensors",
169
+ "encoder.blocks.14.mlp_ln.weight": "model-00005-of-00006.safetensors",
170
+ "encoder.blocks.15.attn.key.weight": "model-00005-of-00006.safetensors",
171
+ "encoder.blocks.15.attn.out.bias": "model-00005-of-00006.safetensors",
172
+ "encoder.blocks.15.attn.out.weight": "model-00005-of-00006.safetensors",
173
+ "encoder.blocks.15.attn.query.bias": "model-00005-of-00006.safetensors",
174
+ "encoder.blocks.15.attn.query.weight": "model-00005-of-00006.safetensors",
175
+ "encoder.blocks.15.attn.value.bias": "model-00005-of-00006.safetensors",
176
+ "encoder.blocks.15.attn.value.weight": "model-00005-of-00006.safetensors",
177
+ "encoder.blocks.15.attn_ln.bias": "model-00005-of-00006.safetensors",
178
+ "encoder.blocks.15.attn_ln.weight": "model-00005-of-00006.safetensors",
179
+ "encoder.blocks.15.mlp.0.bias": "model-00005-of-00006.safetensors",
180
+ "encoder.blocks.15.mlp.0.weight": "model-00005-of-00006.safetensors",
181
+ "encoder.blocks.15.mlp.2.bias": "model-00005-of-00006.safetensors",
182
+ "encoder.blocks.15.mlp.2.weight": "model-00005-of-00006.safetensors",
183
+ "encoder.blocks.15.mlp_ln.bias": "model-00005-of-00006.safetensors",
184
+ "encoder.blocks.15.mlp_ln.weight": "model-00005-of-00006.safetensors",
185
+ "encoder.blocks.16.attn.key.weight": "model-00005-of-00006.safetensors",
186
+ "encoder.blocks.16.attn.out.bias": "model-00005-of-00006.safetensors",
187
+ "encoder.blocks.16.attn.out.weight": "model-00005-of-00006.safetensors",
188
+ "encoder.blocks.16.attn.query.bias": "model-00005-of-00006.safetensors",
189
+ "encoder.blocks.16.attn.query.weight": "model-00005-of-00006.safetensors",
190
+ "encoder.blocks.16.attn.value.bias": "model-00005-of-00006.safetensors",
191
+ "encoder.blocks.16.attn.value.weight": "model-00005-of-00006.safetensors",
192
+ "encoder.blocks.16.attn_ln.bias": "model-00005-of-00006.safetensors",
193
+ "encoder.blocks.16.attn_ln.weight": "model-00005-of-00006.safetensors",
194
+ "encoder.blocks.16.mlp.0.bias": "model-00005-of-00006.safetensors",
195
+ "encoder.blocks.16.mlp.0.weight": "model-00005-of-00006.safetensors",
196
+ "encoder.blocks.16.mlp.2.bias": "model-00005-of-00006.safetensors",
197
+ "encoder.blocks.16.mlp.2.weight": "model-00005-of-00006.safetensors",
198
+ "encoder.blocks.16.mlp_ln.bias": "model-00005-of-00006.safetensors",
199
+ "encoder.blocks.16.mlp_ln.weight": "model-00005-of-00006.safetensors",
200
+ "encoder.blocks.17.attn.key.weight": "model-00005-of-00006.safetensors",
201
+ "encoder.blocks.17.attn.out.bias": "model-00005-of-00006.safetensors",
202
+ "encoder.blocks.17.attn.out.weight": "model-00005-of-00006.safetensors",
203
+ "encoder.blocks.17.attn.query.bias": "model-00005-of-00006.safetensors",
204
+ "encoder.blocks.17.attn.query.weight": "model-00005-of-00006.safetensors",
205
+ "encoder.blocks.17.attn.value.bias": "model-00005-of-00006.safetensors",
206
+ "encoder.blocks.17.attn.value.weight": "model-00005-of-00006.safetensors",
207
+ "encoder.blocks.17.attn_ln.bias": "model-00005-of-00006.safetensors",
208
+ "encoder.blocks.17.attn_ln.weight": "model-00005-of-00006.safetensors",
209
+ "encoder.blocks.17.mlp.0.bias": "model-00005-of-00006.safetensors",
210
+ "encoder.blocks.17.mlp.0.weight": "model-00005-of-00006.safetensors",
211
+ "encoder.blocks.17.mlp.2.bias": "model-00005-of-00006.safetensors",
212
+ "encoder.blocks.17.mlp.2.weight": "model-00005-of-00006.safetensors",
213
+ "encoder.blocks.17.mlp_ln.bias": "model-00005-of-00006.safetensors",
214
+ "encoder.blocks.17.mlp_ln.weight": "model-00005-of-00006.safetensors",
215
+ "encoder.blocks.18.attn.key.weight": "model-00005-of-00006.safetensors",
216
+ "encoder.blocks.18.attn.out.bias": "model-00005-of-00006.safetensors",
217
+ "encoder.blocks.18.attn.out.weight": "model-00005-of-00006.safetensors",
218
+ "encoder.blocks.18.attn.query.bias": "model-00005-of-00006.safetensors",
219
+ "encoder.blocks.18.attn.query.weight": "model-00005-of-00006.safetensors",
220
+ "encoder.blocks.18.attn.value.bias": "model-00005-of-00006.safetensors",
221
+ "encoder.blocks.18.attn.value.weight": "model-00005-of-00006.safetensors",
222
+ "encoder.blocks.18.attn_ln.bias": "model-00005-of-00006.safetensors",
223
+ "encoder.blocks.18.attn_ln.weight": "model-00005-of-00006.safetensors",
224
+ "encoder.blocks.18.mlp.0.bias": "model-00005-of-00006.safetensors",
225
+ "encoder.blocks.18.mlp.0.weight": "model-00005-of-00006.safetensors",
226
+ "encoder.blocks.18.mlp.2.bias": "model-00005-of-00006.safetensors",
227
+ "encoder.blocks.18.mlp.2.weight": "model-00005-of-00006.safetensors",
228
+ "encoder.blocks.18.mlp_ln.bias": "model-00005-of-00006.safetensors",
229
+ "encoder.blocks.18.mlp_ln.weight": "model-00005-of-00006.safetensors",
230
+ "encoder.blocks.19.attn.key.weight": "model-00005-of-00006.safetensors",
231
+ "encoder.blocks.19.attn.out.bias": "model-00005-of-00006.safetensors",
232
+ "encoder.blocks.19.attn.out.weight": "model-00005-of-00006.safetensors",
233
+ "encoder.blocks.19.attn.query.bias": "model-00005-of-00006.safetensors",
234
+ "encoder.blocks.19.attn.query.weight": "model-00005-of-00006.safetensors",
235
+ "encoder.blocks.19.attn.value.bias": "model-00005-of-00006.safetensors",
236
+ "encoder.blocks.19.attn.value.weight": "model-00005-of-00006.safetensors",
237
+ "encoder.blocks.19.attn_ln.bias": "model-00005-of-00006.safetensors",
238
+ "encoder.blocks.19.attn_ln.weight": "model-00005-of-00006.safetensors",
239
+ "encoder.blocks.19.mlp.0.bias": "model-00005-of-00006.safetensors",
240
+ "encoder.blocks.19.mlp.0.weight": "model-00005-of-00006.safetensors",
241
+ "encoder.blocks.19.mlp.2.bias": "model-00005-of-00006.safetensors",
242
+ "encoder.blocks.19.mlp.2.weight": "model-00005-of-00006.safetensors",
243
+ "encoder.blocks.19.mlp_ln.bias": "model-00005-of-00006.safetensors",
244
+ "encoder.blocks.19.mlp_ln.weight": "model-00005-of-00006.safetensors",
245
+ "encoder.blocks.2.attn.key.weight": "model-00005-of-00006.safetensors",
246
+ "encoder.blocks.2.attn.out.bias": "model-00005-of-00006.safetensors",
247
+ "encoder.blocks.2.attn.out.weight": "model-00005-of-00006.safetensors",
248
+ "encoder.blocks.2.attn.query.bias": "model-00005-of-00006.safetensors",
249
+ "encoder.blocks.2.attn.query.weight": "model-00005-of-00006.safetensors",
250
+ "encoder.blocks.2.attn.value.bias": "model-00005-of-00006.safetensors",
251
+ "encoder.blocks.2.attn.value.weight": "model-00005-of-00006.safetensors",
252
+ "encoder.blocks.2.attn_ln.bias": "model-00005-of-00006.safetensors",
253
+ "encoder.blocks.2.attn_ln.weight": "model-00005-of-00006.safetensors",
254
+ "encoder.blocks.2.mlp.0.bias": "model-00005-of-00006.safetensors",
255
+ "encoder.blocks.2.mlp.0.weight": "model-00005-of-00006.safetensors",
256
+ "encoder.blocks.2.mlp.2.bias": "model-00005-of-00006.safetensors",
257
+ "encoder.blocks.2.mlp.2.weight": "model-00005-of-00006.safetensors",
258
+ "encoder.blocks.2.mlp_ln.bias": "model-00005-of-00006.safetensors",
259
+ "encoder.blocks.2.mlp_ln.weight": "model-00005-of-00006.safetensors",
260
+ "encoder.blocks.20.attn.key.weight": "model-00005-of-00006.safetensors",
261
+ "encoder.blocks.20.attn.out.bias": "model-00005-of-00006.safetensors",
262
+ "encoder.blocks.20.attn.out.weight": "model-00005-of-00006.safetensors",
263
+ "encoder.blocks.20.attn.query.bias": "model-00005-of-00006.safetensors",
264
+ "encoder.blocks.20.attn.query.weight": "model-00005-of-00006.safetensors",
265
+ "encoder.blocks.20.attn.value.bias": "model-00005-of-00006.safetensors",
266
+ "encoder.blocks.20.attn.value.weight": "model-00005-of-00006.safetensors",
267
+ "encoder.blocks.20.attn_ln.bias": "model-00005-of-00006.safetensors",
268
+ "encoder.blocks.20.attn_ln.weight": "model-00005-of-00006.safetensors",
269
+ "encoder.blocks.20.mlp.0.bias": "model-00005-of-00006.safetensors",
270
+ "encoder.blocks.20.mlp.0.weight": "model-00005-of-00006.safetensors",
271
+ "encoder.blocks.20.mlp.2.bias": "model-00005-of-00006.safetensors",
272
+ "encoder.blocks.20.mlp.2.weight": "model-00005-of-00006.safetensors",
273
+ "encoder.blocks.20.mlp_ln.bias": "model-00005-of-00006.safetensors",
274
+ "encoder.blocks.20.mlp_ln.weight": "model-00005-of-00006.safetensors",
275
+ "encoder.blocks.21.attn.key.weight": "model-00005-of-00006.safetensors",
276
+ "encoder.blocks.21.attn.out.bias": "model-00005-of-00006.safetensors",
277
+ "encoder.blocks.21.attn.out.weight": "model-00005-of-00006.safetensors",
278
+ "encoder.blocks.21.attn.query.bias": "model-00005-of-00006.safetensors",
279
+ "encoder.blocks.21.attn.query.weight": "model-00005-of-00006.safetensors",
280
+ "encoder.blocks.21.attn.value.bias": "model-00005-of-00006.safetensors",
281
+ "encoder.blocks.21.attn.value.weight": "model-00005-of-00006.safetensors",
282
+ "encoder.blocks.21.attn_ln.bias": "model-00005-of-00006.safetensors",
283
+ "encoder.blocks.21.attn_ln.weight": "model-00005-of-00006.safetensors",
284
+ "encoder.blocks.21.mlp.0.bias": "model-00005-of-00006.safetensors",
285
+ "encoder.blocks.21.mlp.0.weight": "model-00005-of-00006.safetensors",
286
+ "encoder.blocks.21.mlp.2.bias": "model-00005-of-00006.safetensors",
287
+ "encoder.blocks.21.mlp.2.weight": "model-00005-of-00006.safetensors",
288
+ "encoder.blocks.21.mlp_ln.bias": "model-00005-of-00006.safetensors",
289
+ "encoder.blocks.21.mlp_ln.weight": "model-00005-of-00006.safetensors",
290
+ "encoder.blocks.22.attn.key.weight": "model-00005-of-00006.safetensors",
291
+ "encoder.blocks.22.attn.out.bias": "model-00005-of-00006.safetensors",
292
+ "encoder.blocks.22.attn.out.weight": "model-00005-of-00006.safetensors",
293
+ "encoder.blocks.22.attn.query.bias": "model-00005-of-00006.safetensors",
294
+ "encoder.blocks.22.attn.query.weight": "model-00005-of-00006.safetensors",
295
+ "encoder.blocks.22.attn.value.bias": "model-00005-of-00006.safetensors",
296
+ "encoder.blocks.22.attn.value.weight": "model-00005-of-00006.safetensors",
297
+ "encoder.blocks.22.attn_ln.bias": "model-00005-of-00006.safetensors",
298
+ "encoder.blocks.22.attn_ln.weight": "model-00005-of-00006.safetensors",
299
+ "encoder.blocks.22.mlp.0.bias": "model-00005-of-00006.safetensors",
300
+ "encoder.blocks.22.mlp.0.weight": "model-00005-of-00006.safetensors",
301
+ "encoder.blocks.22.mlp.2.bias": "model-00005-of-00006.safetensors",
302
+ "encoder.blocks.22.mlp.2.weight": "model-00005-of-00006.safetensors",
303
+ "encoder.blocks.22.mlp_ln.bias": "model-00005-of-00006.safetensors",
304
+ "encoder.blocks.22.mlp_ln.weight": "model-00005-of-00006.safetensors",
305
+ "encoder.blocks.23.attn.key.weight": "model-00005-of-00006.safetensors",
306
+ "encoder.blocks.23.attn.out.bias": "model-00006-of-00006.safetensors",
307
+ "encoder.blocks.23.attn.out.weight": "model-00006-of-00006.safetensors",
308
+ "encoder.blocks.23.attn.query.bias": "model-00005-of-00006.safetensors",
309
+ "encoder.blocks.23.attn.query.weight": "model-00005-of-00006.safetensors",
310
+ "encoder.blocks.23.attn.value.bias": "model-00005-of-00006.safetensors",
311
+ "encoder.blocks.23.attn.value.weight": "model-00005-of-00006.safetensors",
312
+ "encoder.blocks.23.attn_ln.bias": "model-00006-of-00006.safetensors",
313
+ "encoder.blocks.23.attn_ln.weight": "model-00006-of-00006.safetensors",
314
+ "encoder.blocks.23.mlp.0.bias": "model-00006-of-00006.safetensors",
315
+ "encoder.blocks.23.mlp.0.weight": "model-00006-of-00006.safetensors",
316
+ "encoder.blocks.23.mlp.2.bias": "model-00006-of-00006.safetensors",
317
+ "encoder.blocks.23.mlp.2.weight": "model-00006-of-00006.safetensors",
318
+ "encoder.blocks.23.mlp_ln.bias": "model-00006-of-00006.safetensors",
319
+ "encoder.blocks.23.mlp_ln.weight": "model-00006-of-00006.safetensors",
320
+ "encoder.blocks.24.attn.key.weight": "model-00006-of-00006.safetensors",
321
+ "encoder.blocks.24.attn.out.bias": "model-00006-of-00006.safetensors",
322
+ "encoder.blocks.24.attn.out.weight": "model-00006-of-00006.safetensors",
323
+ "encoder.blocks.24.attn.query.bias": "model-00006-of-00006.safetensors",
324
+ "encoder.blocks.24.attn.query.weight": "model-00006-of-00006.safetensors",
325
+ "encoder.blocks.24.attn.value.bias": "model-00006-of-00006.safetensors",
326
+ "encoder.blocks.24.attn.value.weight": "model-00006-of-00006.safetensors",
327
+ "encoder.blocks.24.attn_ln.bias": "model-00006-of-00006.safetensors",
328
+ "encoder.blocks.24.attn_ln.weight": "model-00006-of-00006.safetensors",
329
+ "encoder.blocks.24.mlp.0.bias": "model-00006-of-00006.safetensors",
330
+ "encoder.blocks.24.mlp.0.weight": "model-00006-of-00006.safetensors",
331
+ "encoder.blocks.24.mlp.2.bias": "model-00006-of-00006.safetensors",
332
+ "encoder.blocks.24.mlp.2.weight": "model-00006-of-00006.safetensors",
333
+ "encoder.blocks.24.mlp_ln.bias": "model-00006-of-00006.safetensors",
334
+ "encoder.blocks.24.mlp_ln.weight": "model-00006-of-00006.safetensors",
335
+ "encoder.blocks.25.attn.key.weight": "model-00006-of-00006.safetensors",
336
+ "encoder.blocks.25.attn.out.bias": "model-00006-of-00006.safetensors",
337
+ "encoder.blocks.25.attn.out.weight": "model-00006-of-00006.safetensors",
338
+ "encoder.blocks.25.attn.query.bias": "model-00006-of-00006.safetensors",
339
+ "encoder.blocks.25.attn.query.weight": "model-00006-of-00006.safetensors",
340
+ "encoder.blocks.25.attn.value.bias": "model-00006-of-00006.safetensors",
341
+ "encoder.blocks.25.attn.value.weight": "model-00006-of-00006.safetensors",
342
+ "encoder.blocks.25.attn_ln.bias": "model-00006-of-00006.safetensors",
343
+ "encoder.blocks.25.attn_ln.weight": "model-00006-of-00006.safetensors",
344
+ "encoder.blocks.25.mlp.0.bias": "model-00006-of-00006.safetensors",
345
+ "encoder.blocks.25.mlp.0.weight": "model-00006-of-00006.safetensors",
346
+ "encoder.blocks.25.mlp.2.bias": "model-00006-of-00006.safetensors",
347
+ "encoder.blocks.25.mlp.2.weight": "model-00006-of-00006.safetensors",
348
+ "encoder.blocks.25.mlp_ln.bias": "model-00006-of-00006.safetensors",
349
+ "encoder.blocks.25.mlp_ln.weight": "model-00006-of-00006.safetensors",
350
+ "encoder.blocks.26.attn.key.weight": "model-00006-of-00006.safetensors",
351
+ "encoder.blocks.26.attn.out.bias": "model-00006-of-00006.safetensors",
352
+ "encoder.blocks.26.attn.out.weight": "model-00006-of-00006.safetensors",
353
+ "encoder.blocks.26.attn.query.bias": "model-00006-of-00006.safetensors",
354
+ "encoder.blocks.26.attn.query.weight": "model-00006-of-00006.safetensors",
355
+ "encoder.blocks.26.attn.value.bias": "model-00006-of-00006.safetensors",
356
+ "encoder.blocks.26.attn.value.weight": "model-00006-of-00006.safetensors",
357
+ "encoder.blocks.26.attn_ln.bias": "model-00006-of-00006.safetensors",
358
+ "encoder.blocks.26.attn_ln.weight": "model-00006-of-00006.safetensors",
359
+ "encoder.blocks.26.mlp.0.bias": "model-00006-of-00006.safetensors",
360
+ "encoder.blocks.26.mlp.0.weight": "model-00006-of-00006.safetensors",
361
+ "encoder.blocks.26.mlp.2.bias": "model-00006-of-00006.safetensors",
362
+ "encoder.blocks.26.mlp.2.weight": "model-00006-of-00006.safetensors",
363
+ "encoder.blocks.26.mlp_ln.bias": "model-00006-of-00006.safetensors",
364
+ "encoder.blocks.26.mlp_ln.weight": "model-00006-of-00006.safetensors",
365
+ "encoder.blocks.27.attn.key.weight": "model-00006-of-00006.safetensors",
366
+ "encoder.blocks.27.attn.out.bias": "model-00006-of-00006.safetensors",
367
+ "encoder.blocks.27.attn.out.weight": "model-00006-of-00006.safetensors",
368
+ "encoder.blocks.27.attn.query.bias": "model-00006-of-00006.safetensors",
369
+ "encoder.blocks.27.attn.query.weight": "model-00006-of-00006.safetensors",
370
+ "encoder.blocks.27.attn.value.bias": "model-00006-of-00006.safetensors",
371
+ "encoder.blocks.27.attn.value.weight": "model-00006-of-00006.safetensors",
372
+ "encoder.blocks.27.attn_ln.bias": "model-00006-of-00006.safetensors",
373
+ "encoder.blocks.27.attn_ln.weight": "model-00006-of-00006.safetensors",
374
+ "encoder.blocks.27.mlp.0.bias": "model-00006-of-00006.safetensors",
375
+ "encoder.blocks.27.mlp.0.weight": "model-00006-of-00006.safetensors",
376
+ "encoder.blocks.27.mlp.2.bias": "model-00006-of-00006.safetensors",
377
+ "encoder.blocks.27.mlp.2.weight": "model-00006-of-00006.safetensors",
378
+ "encoder.blocks.27.mlp_ln.bias": "model-00006-of-00006.safetensors",
379
+ "encoder.blocks.27.mlp_ln.weight": "model-00006-of-00006.safetensors",
380
+ "encoder.blocks.28.attn.key.weight": "model-00006-of-00006.safetensors",
381
+ "encoder.blocks.28.attn.out.bias": "model-00006-of-00006.safetensors",
382
+ "encoder.blocks.28.attn.out.weight": "model-00006-of-00006.safetensors",
383
+ "encoder.blocks.28.attn.query.bias": "model-00006-of-00006.safetensors",
384
+ "encoder.blocks.28.attn.query.weight": "model-00006-of-00006.safetensors",
385
+ "encoder.blocks.28.attn.value.bias": "model-00006-of-00006.safetensors",
386
+ "encoder.blocks.28.attn.value.weight": "model-00006-of-00006.safetensors",
387
+ "encoder.blocks.28.attn_ln.bias": "model-00006-of-00006.safetensors",
388
+ "encoder.blocks.28.attn_ln.weight": "model-00006-of-00006.safetensors",
389
+ "encoder.blocks.28.mlp.0.bias": "model-00006-of-00006.safetensors",
390
+ "encoder.blocks.28.mlp.0.weight": "model-00006-of-00006.safetensors",
391
+ "encoder.blocks.28.mlp.2.bias": "model-00006-of-00006.safetensors",
392
+ "encoder.blocks.28.mlp.2.weight": "model-00006-of-00006.safetensors",
393
+ "encoder.blocks.28.mlp_ln.bias": "model-00006-of-00006.safetensors",
394
+ "encoder.blocks.28.mlp_ln.weight": "model-00006-of-00006.safetensors",
395
+ "encoder.blocks.29.attn.key.weight": "model-00006-of-00006.safetensors",
396
+ "encoder.blocks.29.attn.out.bias": "model-00006-of-00006.safetensors",
397
+ "encoder.blocks.29.attn.out.weight": "model-00006-of-00006.safetensors",
398
+ "encoder.blocks.29.attn.query.bias": "model-00006-of-00006.safetensors",
399
+ "encoder.blocks.29.attn.query.weight": "model-00006-of-00006.safetensors",
400
+ "encoder.blocks.29.attn.value.bias": "model-00006-of-00006.safetensors",
401
+ "encoder.blocks.29.attn.value.weight": "model-00006-of-00006.safetensors",
402
+ "encoder.blocks.29.attn_ln.bias": "model-00006-of-00006.safetensors",
403
+ "encoder.blocks.29.attn_ln.weight": "model-00006-of-00006.safetensors",
404
+ "encoder.blocks.29.mlp.0.bias": "model-00006-of-00006.safetensors",
405
+ "encoder.blocks.29.mlp.0.weight": "model-00006-of-00006.safetensors",
406
+ "encoder.blocks.29.mlp.2.bias": "model-00006-of-00006.safetensors",
407
+ "encoder.blocks.29.mlp.2.weight": "model-00006-of-00006.safetensors",
408
+ "encoder.blocks.29.mlp_ln.bias": "model-00006-of-00006.safetensors",
409
+ "encoder.blocks.29.mlp_ln.weight": "model-00006-of-00006.safetensors",
410
+ "encoder.blocks.3.attn.key.weight": "model-00005-of-00006.safetensors",
411
+ "encoder.blocks.3.attn.out.bias": "model-00005-of-00006.safetensors",
412
+ "encoder.blocks.3.attn.out.weight": "model-00005-of-00006.safetensors",
413
+ "encoder.blocks.3.attn.query.bias": "model-00005-of-00006.safetensors",
414
+ "encoder.blocks.3.attn.query.weight": "model-00005-of-00006.safetensors",
415
+ "encoder.blocks.3.attn.value.bias": "model-00005-of-00006.safetensors",
416
+ "encoder.blocks.3.attn.value.weight": "model-00005-of-00006.safetensors",
417
+ "encoder.blocks.3.attn_ln.bias": "model-00005-of-00006.safetensors",
418
+ "encoder.blocks.3.attn_ln.weight": "model-00005-of-00006.safetensors",
419
+ "encoder.blocks.3.mlp.0.bias": "model-00005-of-00006.safetensors",
420
+ "encoder.blocks.3.mlp.0.weight": "model-00005-of-00006.safetensors",
421
+ "encoder.blocks.3.mlp.2.bias": "model-00005-of-00006.safetensors",
422
+ "encoder.blocks.3.mlp.2.weight": "model-00005-of-00006.safetensors",
423
+ "encoder.blocks.3.mlp_ln.bias": "model-00005-of-00006.safetensors",
424
+ "encoder.blocks.3.mlp_ln.weight": "model-00005-of-00006.safetensors",
425
+ "encoder.blocks.30.attn.key.weight": "model-00006-of-00006.safetensors",
426
+ "encoder.blocks.30.attn.out.bias": "model-00006-of-00006.safetensors",
427
+ "encoder.blocks.30.attn.out.weight": "model-00006-of-00006.safetensors",
428
+ "encoder.blocks.30.attn.query.bias": "model-00006-of-00006.safetensors",
429
+ "encoder.blocks.30.attn.query.weight": "model-00006-of-00006.safetensors",
430
+ "encoder.blocks.30.attn.value.bias": "model-00006-of-00006.safetensors",
431
+ "encoder.blocks.30.attn.value.weight": "model-00006-of-00006.safetensors",
432
+ "encoder.blocks.30.attn_ln.bias": "model-00006-of-00006.safetensors",
433
+ "encoder.blocks.30.attn_ln.weight": "model-00006-of-00006.safetensors",
434
+ "encoder.blocks.30.mlp.0.bias": "model-00006-of-00006.safetensors",
435
+ "encoder.blocks.30.mlp.0.weight": "model-00006-of-00006.safetensors",
436
+ "encoder.blocks.30.mlp.2.bias": "model-00006-of-00006.safetensors",
437
+ "encoder.blocks.30.mlp.2.weight": "model-00006-of-00006.safetensors",
438
+ "encoder.blocks.30.mlp_ln.bias": "model-00006-of-00006.safetensors",
439
+ "encoder.blocks.30.mlp_ln.weight": "model-00006-of-00006.safetensors",
440
+ "encoder.blocks.31.attn.key.weight": "model-00006-of-00006.safetensors",
441
+ "encoder.blocks.31.attn.out.bias": "model-00006-of-00006.safetensors",
442
+ "encoder.blocks.31.attn.out.weight": "model-00006-of-00006.safetensors",
443
+ "encoder.blocks.31.attn.query.bias": "model-00006-of-00006.safetensors",
444
+ "encoder.blocks.31.attn.query.weight": "model-00006-of-00006.safetensors",
445
+ "encoder.blocks.31.attn.value.bias": "model-00006-of-00006.safetensors",
446
+ "encoder.blocks.31.attn.value.weight": "model-00006-of-00006.safetensors",
447
+ "encoder.blocks.31.attn_ln.bias": "model-00006-of-00006.safetensors",
448
+ "encoder.blocks.31.attn_ln.weight": "model-00006-of-00006.safetensors",
449
+ "encoder.blocks.31.mlp.0.bias": "model-00006-of-00006.safetensors",
450
+ "encoder.blocks.31.mlp.0.weight": "model-00006-of-00006.safetensors",
451
+ "encoder.blocks.31.mlp.2.bias": "model-00006-of-00006.safetensors",
452
+ "encoder.blocks.31.mlp.2.weight": "model-00006-of-00006.safetensors",
453
+ "encoder.blocks.31.mlp_ln.bias": "model-00006-of-00006.safetensors",
454
+ "encoder.blocks.31.mlp_ln.weight": "model-00006-of-00006.safetensors",
455
+ "encoder.blocks.4.attn.key.weight": "model-00005-of-00006.safetensors",
456
+ "encoder.blocks.4.attn.out.bias": "model-00005-of-00006.safetensors",
457
+ "encoder.blocks.4.attn.out.weight": "model-00005-of-00006.safetensors",
458
+ "encoder.blocks.4.attn.query.bias": "model-00005-of-00006.safetensors",
459
+ "encoder.blocks.4.attn.query.weight": "model-00005-of-00006.safetensors",
460
+ "encoder.blocks.4.attn.value.bias": "model-00005-of-00006.safetensors",
461
+ "encoder.blocks.4.attn.value.weight": "model-00005-of-00006.safetensors",
462
+ "encoder.blocks.4.attn_ln.bias": "model-00005-of-00006.safetensors",
463
+ "encoder.blocks.4.attn_ln.weight": "model-00005-of-00006.safetensors",
464
+ "encoder.blocks.4.mlp.0.bias": "model-00005-of-00006.safetensors",
465
+ "encoder.blocks.4.mlp.0.weight": "model-00005-of-00006.safetensors",
466
+ "encoder.blocks.4.mlp.2.bias": "model-00005-of-00006.safetensors",
467
+ "encoder.blocks.4.mlp.2.weight": "model-00005-of-00006.safetensors",
468
+ "encoder.blocks.4.mlp_ln.bias": "model-00005-of-00006.safetensors",
469
+ "encoder.blocks.4.mlp_ln.weight": "model-00005-of-00006.safetensors",
470
+ "encoder.blocks.5.attn.key.weight": "model-00005-of-00006.safetensors",
471
+ "encoder.blocks.5.attn.out.bias": "model-00005-of-00006.safetensors",
472
+ "encoder.blocks.5.attn.out.weight": "model-00005-of-00006.safetensors",
473
+ "encoder.blocks.5.attn.query.bias": "model-00005-of-00006.safetensors",
474
+ "encoder.blocks.5.attn.query.weight": "model-00005-of-00006.safetensors",
475
+ "encoder.blocks.5.attn.value.bias": "model-00005-of-00006.safetensors",
476
+ "encoder.blocks.5.attn.value.weight": "model-00005-of-00006.safetensors",
477
+ "encoder.blocks.5.attn_ln.bias": "model-00005-of-00006.safetensors",
478
+ "encoder.blocks.5.attn_ln.weight": "model-00005-of-00006.safetensors",
479
+ "encoder.blocks.5.mlp.0.bias": "model-00005-of-00006.safetensors",
480
+ "encoder.blocks.5.mlp.0.weight": "model-00005-of-00006.safetensors",
481
+ "encoder.blocks.5.mlp.2.bias": "model-00005-of-00006.safetensors",
482
+ "encoder.blocks.5.mlp.2.weight": "model-00005-of-00006.safetensors",
483
+ "encoder.blocks.5.mlp_ln.bias": "model-00005-of-00006.safetensors",
484
+ "encoder.blocks.5.mlp_ln.weight": "model-00005-of-00006.safetensors",
485
+ "encoder.blocks.6.attn.key.weight": "model-00005-of-00006.safetensors",
486
+ "encoder.blocks.6.attn.out.bias": "model-00005-of-00006.safetensors",
487
+ "encoder.blocks.6.attn.out.weight": "model-00005-of-00006.safetensors",
488
+ "encoder.blocks.6.attn.query.bias": "model-00005-of-00006.safetensors",
489
+ "encoder.blocks.6.attn.query.weight": "model-00005-of-00006.safetensors",
490
+ "encoder.blocks.6.attn.value.bias": "model-00005-of-00006.safetensors",
491
+ "encoder.blocks.6.attn.value.weight": "model-00005-of-00006.safetensors",
492
+ "encoder.blocks.6.attn_ln.bias": "model-00005-of-00006.safetensors",
493
+ "encoder.blocks.6.attn_ln.weight": "model-00005-of-00006.safetensors",
494
+ "encoder.blocks.6.mlp.0.bias": "model-00005-of-00006.safetensors",
495
+ "encoder.blocks.6.mlp.0.weight": "model-00005-of-00006.safetensors",
496
+ "encoder.blocks.6.mlp.2.bias": "model-00005-of-00006.safetensors",
497
+ "encoder.blocks.6.mlp.2.weight": "model-00005-of-00006.safetensors",
498
+ "encoder.blocks.6.mlp_ln.bias": "model-00005-of-00006.safetensors",
499
+ "encoder.blocks.6.mlp_ln.weight": "model-00005-of-00006.safetensors",
500
+ "encoder.blocks.7.attn.key.weight": "model-00005-of-00006.safetensors",
501
+ "encoder.blocks.7.attn.out.bias": "model-00005-of-00006.safetensors",
502
+ "encoder.blocks.7.attn.out.weight": "model-00005-of-00006.safetensors",
503
+ "encoder.blocks.7.attn.query.bias": "model-00005-of-00006.safetensors",
504
+ "encoder.blocks.7.attn.query.weight": "model-00005-of-00006.safetensors",
505
+ "encoder.blocks.7.attn.value.bias": "model-00005-of-00006.safetensors",
506
+ "encoder.blocks.7.attn.value.weight": "model-00005-of-00006.safetensors",
507
+ "encoder.blocks.7.attn_ln.bias": "model-00005-of-00006.safetensors",
508
+ "encoder.blocks.7.attn_ln.weight": "model-00005-of-00006.safetensors",
509
+ "encoder.blocks.7.mlp.0.bias": "model-00005-of-00006.safetensors",
510
+ "encoder.blocks.7.mlp.0.weight": "model-00005-of-00006.safetensors",
511
+ "encoder.blocks.7.mlp.2.bias": "model-00005-of-00006.safetensors",
512
+ "encoder.blocks.7.mlp.2.weight": "model-00005-of-00006.safetensors",
513
+ "encoder.blocks.7.mlp_ln.bias": "model-00005-of-00006.safetensors",
514
+ "encoder.blocks.7.mlp_ln.weight": "model-00005-of-00006.safetensors",
515
+ "encoder.blocks.8.attn.key.weight": "model-00005-of-00006.safetensors",
516
+ "encoder.blocks.8.attn.out.bias": "model-00005-of-00006.safetensors",
517
+ "encoder.blocks.8.attn.out.weight": "model-00005-of-00006.safetensors",
518
+ "encoder.blocks.8.attn.query.bias": "model-00005-of-00006.safetensors",
519
+ "encoder.blocks.8.attn.query.weight": "model-00005-of-00006.safetensors",
520
+ "encoder.blocks.8.attn.value.bias": "model-00005-of-00006.safetensors",
521
+ "encoder.blocks.8.attn.value.weight": "model-00005-of-00006.safetensors",
522
+ "encoder.blocks.8.attn_ln.bias": "model-00005-of-00006.safetensors",
523
+ "encoder.blocks.8.attn_ln.weight": "model-00005-of-00006.safetensors",
524
+ "encoder.blocks.8.mlp.0.bias": "model-00005-of-00006.safetensors",
525
+ "encoder.blocks.8.mlp.0.weight": "model-00005-of-00006.safetensors",
526
+ "encoder.blocks.8.mlp.2.bias": "model-00005-of-00006.safetensors",
527
+ "encoder.blocks.8.mlp.2.weight": "model-00005-of-00006.safetensors",
528
+ "encoder.blocks.8.mlp_ln.bias": "model-00005-of-00006.safetensors",
529
+ "encoder.blocks.8.mlp_ln.weight": "model-00005-of-00006.safetensors",
530
+ "encoder.blocks.9.attn.key.weight": "model-00005-of-00006.safetensors",
531
+ "encoder.blocks.9.attn.out.bias": "model-00005-of-00006.safetensors",
532
+ "encoder.blocks.9.attn.out.weight": "model-00005-of-00006.safetensors",
533
+ "encoder.blocks.9.attn.query.bias": "model-00005-of-00006.safetensors",
534
+ "encoder.blocks.9.attn.query.weight": "model-00005-of-00006.safetensors",
535
+ "encoder.blocks.9.attn.value.bias": "model-00005-of-00006.safetensors",
536
+ "encoder.blocks.9.attn.value.weight": "model-00005-of-00006.safetensors",
537
+ "encoder.blocks.9.attn_ln.bias": "model-00005-of-00006.safetensors",
538
+ "encoder.blocks.9.attn_ln.weight": "model-00005-of-00006.safetensors",
539
+ "encoder.blocks.9.mlp.0.bias": "model-00005-of-00006.safetensors",
540
+ "encoder.blocks.9.mlp.0.weight": "model-00005-of-00006.safetensors",
541
+ "encoder.blocks.9.mlp.2.bias": "model-00005-of-00006.safetensors",
542
+ "encoder.blocks.9.mlp.2.weight": "model-00005-of-00006.safetensors",
543
+ "encoder.blocks.9.mlp_ln.bias": "model-00005-of-00006.safetensors",
544
+ "encoder.blocks.9.mlp_ln.weight": "model-00005-of-00006.safetensors",
545
+ "encoder.conv1.bias": "model-00005-of-00006.safetensors",
546
+ "encoder.conv1.weight": "model-00005-of-00006.safetensors",
547
+ "encoder.conv2.bias": "model-00005-of-00006.safetensors",
548
+ "encoder.conv2.weight": "model-00005-of-00006.safetensors",
549
+ "encoder.positional_embedding.weight": "model-00005-of-00006.safetensors",
550
+ "lm_head.weight": "model-00006-of-00006.safetensors",
551
+ "merge_model.embed_tokens.weight": "model-00005-of-00006.safetensors",
552
+ "merge_model.layers.0.input_layernorm.weight": "model-00005-of-00006.safetensors",
553
+ "merge_model.layers.0.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
554
+ "merge_model.layers.0.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
555
+ "merge_model.layers.0.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
556
+ "merge_model.layers.0.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
557
+ "merge_model.layers.0.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
558
+ "merge_model.layers.0.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
559
+ "merge_model.layers.0.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
560
+ "merge_model.layers.0.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
561
+ "merge_model.layers.0.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
562
+ "merge_model.layers.0.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
563
+ "merge_model.layers.0.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
564
+ "merge_model.layers.1.input_layernorm.weight": "model-00005-of-00006.safetensors",
565
+ "merge_model.layers.1.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
566
+ "merge_model.layers.1.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
567
+ "merge_model.layers.1.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
568
+ "merge_model.layers.1.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
569
+ "merge_model.layers.1.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
570
+ "merge_model.layers.1.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
571
+ "merge_model.layers.1.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
572
+ "merge_model.layers.1.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
573
+ "merge_model.layers.1.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
574
+ "merge_model.layers.1.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
575
+ "merge_model.layers.1.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
576
+ "merge_model.layers.2.input_layernorm.weight": "model-00005-of-00006.safetensors",
577
+ "merge_model.layers.2.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
578
+ "merge_model.layers.2.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
579
+ "merge_model.layers.2.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
580
+ "merge_model.layers.2.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
581
+ "merge_model.layers.2.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
582
+ "merge_model.layers.2.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
583
+ "merge_model.layers.2.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
584
+ "merge_model.layers.2.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
585
+ "merge_model.layers.2.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
586
+ "merge_model.layers.2.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
587
+ "merge_model.layers.2.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
588
+ "merge_model.layers.3.input_layernorm.weight": "model-00005-of-00006.safetensors",
589
+ "merge_model.layers.3.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
590
+ "merge_model.layers.3.mlp.gate_proj.weight": "model-00005-of-00006.safetensors",
591
+ "merge_model.layers.3.mlp.up_proj.weight": "model-00005-of-00006.safetensors",
592
+ "merge_model.layers.3.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
593
+ "merge_model.layers.3.self_attn.k_proj.bias": "model-00005-of-00006.safetensors",
594
+ "merge_model.layers.3.self_attn.k_proj.weight": "model-00005-of-00006.safetensors",
595
+ "merge_model.layers.3.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
596
+ "merge_model.layers.3.self_attn.q_proj.bias": "model-00005-of-00006.safetensors",
597
+ "merge_model.layers.3.self_attn.q_proj.weight": "model-00005-of-00006.safetensors",
598
+ "merge_model.layers.3.self_attn.v_proj.bias": "model-00005-of-00006.safetensors",
599
+ "merge_model.layers.3.self_attn.v_proj.weight": "model-00005-of-00006.safetensors",
600
+ "merge_model.norm.weight": "model-00005-of-00006.safetensors",
601
+ "model.embed_tokens.weight": "model-00001-of-00006.safetensors",
602
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00006.safetensors",
603
+ "model.layers.0.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
604
+ "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
605
+ "model.layers.0.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
606
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
607
+ "model.layers.0.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
608
+ "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
609
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
610
+ "model.layers.0.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
611
+ "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
612
+ "model.layers.0.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
613
+ "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
614
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00006.safetensors",
615
+ "model.layers.1.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
616
+ "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
617
+ "model.layers.1.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
618
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
619
+ "model.layers.1.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
620
+ "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
621
+ "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
622
+ "model.layers.1.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
623
+ "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
624
+ "model.layers.1.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
625
+ "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
626
+ "model.layers.10.input_layernorm.weight": "model-00002-of-00006.safetensors",
627
+ "model.layers.10.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
628
+ "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
629
+ "model.layers.10.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
630
+ "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
631
+ "model.layers.10.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
632
+ "model.layers.10.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
633
+ "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
634
+ "model.layers.10.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
635
+ "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
636
+ "model.layers.10.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
637
+ "model.layers.10.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
638
+ "model.layers.11.input_layernorm.weight": "model-00002-of-00006.safetensors",
639
+ "model.layers.11.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
640
+ "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
641
+ "model.layers.11.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
642
+ "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
643
+ "model.layers.11.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
644
+ "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
645
+ "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
646
+ "model.layers.11.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
647
+ "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
648
+ "model.layers.11.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
649
+ "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
650
+ "model.layers.12.input_layernorm.weight": "model-00002-of-00006.safetensors",
651
+ "model.layers.12.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
652
+ "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
653
+ "model.layers.12.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
654
+ "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
655
+ "model.layers.12.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
656
+ "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
657
+ "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
658
+ "model.layers.12.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
659
+ "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
660
+ "model.layers.12.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
661
+ "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
662
+ "model.layers.13.input_layernorm.weight": "model-00002-of-00006.safetensors",
663
+ "model.layers.13.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
664
+ "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
665
+ "model.layers.13.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
666
+ "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
667
+ "model.layers.13.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
668
+ "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
669
+ "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
670
+ "model.layers.13.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
671
+ "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
672
+ "model.layers.13.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
673
+ "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
674
+ "model.layers.14.input_layernorm.weight": "model-00002-of-00006.safetensors",
675
+ "model.layers.14.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
676
+ "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
677
+ "model.layers.14.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
678
+ "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
679
+ "model.layers.14.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
680
+ "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
681
+ "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
682
+ "model.layers.14.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
683
+ "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
684
+ "model.layers.14.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
685
+ "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
686
+ "model.layers.15.input_layernorm.weight": "model-00002-of-00006.safetensors",
687
+ "model.layers.15.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
688
+ "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
689
+ "model.layers.15.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
690
+ "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
691
+ "model.layers.15.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
692
+ "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
693
+ "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
694
+ "model.layers.15.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
695
+ "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
696
+ "model.layers.15.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
697
+ "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
698
+ "model.layers.16.input_layernorm.weight": "model-00002-of-00006.safetensors",
699
+ "model.layers.16.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
700
+ "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
701
+ "model.layers.16.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
702
+ "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
703
+ "model.layers.16.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
704
+ "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
705
+ "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
706
+ "model.layers.16.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
707
+ "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
708
+ "model.layers.16.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
709
+ "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
710
+ "model.layers.17.input_layernorm.weight": "model-00002-of-00006.safetensors",
711
+ "model.layers.17.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
712
+ "model.layers.17.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
713
+ "model.layers.17.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
714
+ "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
715
+ "model.layers.17.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
716
+ "model.layers.17.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
717
+ "model.layers.17.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
718
+ "model.layers.17.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
719
+ "model.layers.17.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
720
+ "model.layers.17.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
721
+ "model.layers.17.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
722
+ "model.layers.18.input_layernorm.weight": "model-00003-of-00006.safetensors",
723
+ "model.layers.18.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
724
+ "model.layers.18.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
725
+ "model.layers.18.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
726
+ "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
727
+ "model.layers.18.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
728
+ "model.layers.18.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
729
+ "model.layers.18.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
730
+ "model.layers.18.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
731
+ "model.layers.18.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
732
+ "model.layers.18.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
733
+ "model.layers.18.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
734
+ "model.layers.19.input_layernorm.weight": "model-00003-of-00006.safetensors",
735
+ "model.layers.19.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
736
+ "model.layers.19.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
737
+ "model.layers.19.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
738
+ "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
739
+ "model.layers.19.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
740
+ "model.layers.19.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
741
+ "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
742
+ "model.layers.19.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
743
+ "model.layers.19.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
744
+ "model.layers.19.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
745
+ "model.layers.19.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
746
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00006.safetensors",
747
+ "model.layers.2.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
748
+ "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
749
+ "model.layers.2.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
750
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
751
+ "model.layers.2.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
752
+ "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
753
+ "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
754
+ "model.layers.2.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
755
+ "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
756
+ "model.layers.2.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
757
+ "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
758
+ "model.layers.20.input_layernorm.weight": "model-00003-of-00006.safetensors",
759
+ "model.layers.20.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
760
+ "model.layers.20.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
761
+ "model.layers.20.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
762
+ "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
763
+ "model.layers.20.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
764
+ "model.layers.20.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
765
+ "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
766
+ "model.layers.20.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
767
+ "model.layers.20.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
768
+ "model.layers.20.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
769
+ "model.layers.20.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
770
+ "model.layers.21.input_layernorm.weight": "model-00003-of-00006.safetensors",
771
+ "model.layers.21.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
772
+ "model.layers.21.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
773
+ "model.layers.21.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
774
+ "model.layers.21.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
775
+ "model.layers.21.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
776
+ "model.layers.21.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
777
+ "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
778
+ "model.layers.21.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
779
+ "model.layers.21.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
780
+ "model.layers.21.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
781
+ "model.layers.21.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
782
+ "model.layers.22.input_layernorm.weight": "model-00003-of-00006.safetensors",
783
+ "model.layers.22.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
784
+ "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
785
+ "model.layers.22.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
786
+ "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
787
+ "model.layers.22.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
788
+ "model.layers.22.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
789
+ "model.layers.22.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
790
+ "model.layers.22.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
791
+ "model.layers.22.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
792
+ "model.layers.22.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
793
+ "model.layers.22.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
794
+ "model.layers.23.input_layernorm.weight": "model-00003-of-00006.safetensors",
795
+ "model.layers.23.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
796
+ "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
797
+ "model.layers.23.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
798
+ "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
799
+ "model.layers.23.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
800
+ "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
801
+ "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
802
+ "model.layers.23.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
803
+ "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
804
+ "model.layers.23.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
805
+ "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
806
+ "model.layers.24.input_layernorm.weight": "model-00003-of-00006.safetensors",
807
+ "model.layers.24.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
808
+ "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
809
+ "model.layers.24.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
810
+ "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
811
+ "model.layers.24.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
812
+ "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
813
+ "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
814
+ "model.layers.24.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
815
+ "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
816
+ "model.layers.24.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
817
+ "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
818
+ "model.layers.25.input_layernorm.weight": "model-00003-of-00006.safetensors",
819
+ "model.layers.25.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
820
+ "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
821
+ "model.layers.25.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
822
+ "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
823
+ "model.layers.25.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
824
+ "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
825
+ "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
826
+ "model.layers.25.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
827
+ "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
828
+ "model.layers.25.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
829
+ "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
830
+ "model.layers.26.input_layernorm.weight": "model-00003-of-00006.safetensors",
831
+ "model.layers.26.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
832
+ "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
833
+ "model.layers.26.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
834
+ "model.layers.26.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
835
+ "model.layers.26.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
836
+ "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
837
+ "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
838
+ "model.layers.26.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
839
+ "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
840
+ "model.layers.26.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
841
+ "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
842
+ "model.layers.27.input_layernorm.weight": "model-00003-of-00006.safetensors",
843
+ "model.layers.27.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
844
+ "model.layers.27.mlp.gate_proj.weight": "model-00003-of-00006.safetensors",
845
+ "model.layers.27.mlp.up_proj.weight": "model-00003-of-00006.safetensors",
846
+ "model.layers.27.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
847
+ "model.layers.27.self_attn.k_proj.bias": "model-00003-of-00006.safetensors",
848
+ "model.layers.27.self_attn.k_proj.weight": "model-00003-of-00006.safetensors",
849
+ "model.layers.27.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
850
+ "model.layers.27.self_attn.q_proj.bias": "model-00003-of-00006.safetensors",
851
+ "model.layers.27.self_attn.q_proj.weight": "model-00003-of-00006.safetensors",
852
+ "model.layers.27.self_attn.v_proj.bias": "model-00003-of-00006.safetensors",
853
+ "model.layers.27.self_attn.v_proj.weight": "model-00003-of-00006.safetensors",
854
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00006.safetensors",
855
+ "model.layers.3.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
856
+ "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
857
+ "model.layers.3.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
858
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
859
+ "model.layers.3.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
860
+ "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
861
+ "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
862
+ "model.layers.3.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
863
+ "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
864
+ "model.layers.3.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
865
+ "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
866
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00006.safetensors",
867
+ "model.layers.4.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
868
+ "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
869
+ "model.layers.4.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
870
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
871
+ "model.layers.4.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
872
+ "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
873
+ "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
874
+ "model.layers.4.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
875
+ "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
876
+ "model.layers.4.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
877
+ "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
878
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00006.safetensors",
879
+ "model.layers.5.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
880
+ "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
881
+ "model.layers.5.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
882
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
883
+ "model.layers.5.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
884
+ "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
885
+ "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
886
+ "model.layers.5.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
887
+ "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
888
+ "model.layers.5.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
889
+ "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
890
+ "model.layers.6.input_layernorm.weight": "model-00001-of-00006.safetensors",
891
+ "model.layers.6.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
892
+ "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
893
+ "model.layers.6.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
894
+ "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
895
+ "model.layers.6.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
896
+ "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
897
+ "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
898
+ "model.layers.6.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
899
+ "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
900
+ "model.layers.6.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
901
+ "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
902
+ "model.layers.7.input_layernorm.weight": "model-00001-of-00006.safetensors",
903
+ "model.layers.7.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
904
+ "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00006.safetensors",
905
+ "model.layers.7.mlp.up_proj.weight": "model-00001-of-00006.safetensors",
906
+ "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
907
+ "model.layers.7.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
908
+ "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
909
+ "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
910
+ "model.layers.7.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
911
+ "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
912
+ "model.layers.7.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
913
+ "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
914
+ "model.layers.8.input_layernorm.weight": "model-00002-of-00006.safetensors",
915
+ "model.layers.8.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
916
+ "model.layers.8.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
917
+ "model.layers.8.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
918
+ "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
919
+ "model.layers.8.self_attn.k_proj.bias": "model-00001-of-00006.safetensors",
920
+ "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00006.safetensors",
921
+ "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
922
+ "model.layers.8.self_attn.q_proj.bias": "model-00001-of-00006.safetensors",
923
+ "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00006.safetensors",
924
+ "model.layers.8.self_attn.v_proj.bias": "model-00001-of-00006.safetensors",
925
+ "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00006.safetensors",
926
+ "model.layers.9.input_layernorm.weight": "model-00002-of-00006.safetensors",
927
+ "model.layers.9.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
928
+ "model.layers.9.mlp.gate_proj.weight": "model-00002-of-00006.safetensors",
929
+ "model.layers.9.mlp.up_proj.weight": "model-00002-of-00006.safetensors",
930
+ "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
931
+ "model.layers.9.self_attn.k_proj.bias": "model-00002-of-00006.safetensors",
932
+ "model.layers.9.self_attn.k_proj.weight": "model-00002-of-00006.safetensors",
933
+ "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
934
+ "model.layers.9.self_attn.q_proj.bias": "model-00002-of-00006.safetensors",
935
+ "model.layers.9.self_attn.q_proj.weight": "model-00002-of-00006.safetensors",
936
+ "model.layers.9.self_attn.v_proj.bias": "model-00002-of-00006.safetensors",
937
+ "model.layers.9.self_attn.v_proj.weight": "model-00002-of-00006.safetensors",
938
+ "model.norm.weight": "model-00003-of-00006.safetensors",
939
+ "stoken_model.embed_tokens.weight": "model-00004-of-00006.safetensors",
940
+ "stoken_model.layers.0.input_layernorm.weight": "model-00004-of-00006.safetensors",
941
+ "stoken_model.layers.0.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
942
+ "stoken_model.layers.0.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
943
+ "stoken_model.layers.0.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
944
+ "stoken_model.layers.0.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
945
+ "stoken_model.layers.0.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
946
+ "stoken_model.layers.0.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
947
+ "stoken_model.layers.0.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
948
+ "stoken_model.layers.0.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
949
+ "stoken_model.layers.0.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
950
+ "stoken_model.layers.0.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
951
+ "stoken_model.layers.0.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
952
+ "stoken_model.layers.1.input_layernorm.weight": "model-00004-of-00006.safetensors",
953
+ "stoken_model.layers.1.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
954
+ "stoken_model.layers.1.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
955
+ "stoken_model.layers.1.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
956
+ "stoken_model.layers.1.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
957
+ "stoken_model.layers.1.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
958
+ "stoken_model.layers.1.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
959
+ "stoken_model.layers.1.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
960
+ "stoken_model.layers.1.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
961
+ "stoken_model.layers.1.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
962
+ "stoken_model.layers.1.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
963
+ "stoken_model.layers.1.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
964
+ "stoken_model.layers.2.input_layernorm.weight": "model-00004-of-00006.safetensors",
965
+ "stoken_model.layers.2.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
966
+ "stoken_model.layers.2.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
967
+ "stoken_model.layers.2.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
968
+ "stoken_model.layers.2.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
969
+ "stoken_model.layers.2.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
970
+ "stoken_model.layers.2.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
971
+ "stoken_model.layers.2.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
972
+ "stoken_model.layers.2.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
973
+ "stoken_model.layers.2.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
974
+ "stoken_model.layers.2.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
975
+ "stoken_model.layers.2.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
976
+ "stoken_model.layers.3.input_layernorm.weight": "model-00004-of-00006.safetensors",
977
+ "stoken_model.layers.3.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
978
+ "stoken_model.layers.3.mlp.gate_proj.weight": "model-00004-of-00006.safetensors",
979
+ "stoken_model.layers.3.mlp.up_proj.weight": "model-00004-of-00006.safetensors",
980
+ "stoken_model.layers.3.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
981
+ "stoken_model.layers.3.self_attn.k_proj.bias": "model-00004-of-00006.safetensors",
982
+ "stoken_model.layers.3.self_attn.k_proj.weight": "model-00004-of-00006.safetensors",
983
+ "stoken_model.layers.3.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
984
+ "stoken_model.layers.3.self_attn.q_proj.bias": "model-00004-of-00006.safetensors",
985
+ "stoken_model.layers.3.self_attn.q_proj.weight": "model-00004-of-00006.safetensors",
986
+ "stoken_model.layers.3.self_attn.v_proj.bias": "model-00004-of-00006.safetensors",
987
+ "stoken_model.layers.3.self_attn.v_proj.weight": "model-00004-of-00006.safetensors",
988
+ "stoken_model.norm.weight": "model-00004-of-00006.safetensors"
989
+ }
990
+ }
lychee_full_duplex_v1.5/modeling_step_audio_2.py ADDED
@@ -0,0 +1,425 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Iterable, Optional, Tuple
2
+
3
+ import librosa
4
+ import torch
5
+ import torch.nn.functional as F
6
+ import torchaudio
7
+ from torch import Tensor, nn
8
+ from transformers import PreTrainedModel, Qwen2Model
9
+ from transformers.generation.utils import GenerationMixin
10
+ from transformers.modeling_outputs import CausalLMOutputWithPast
11
+
12
+ from .configuration_step_audio_2 import StepAudio2Config
13
+
14
+
15
+ def _mel_filters(n_mels: int) -> torch.Tensor:
16
+ """Load the mel filterbank matrix for projecting STFT into a Mel spectrogram."""
17
+ assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}"
18
+ if n_mels == 128:
19
+ return torch.from_numpy(librosa.filters.mel(sr=16000, n_fft=400, n_mels=128))
20
+ else:
21
+ return torch.from_numpy(librosa.filters.mel(sr=16000, n_fft=400, n_mels=80))
22
+
23
+
24
+ def load_audio(file_path, target_rate=16000, max_length=None):
25
+ """
26
+ Open an audio file and read as mono waveform, resampling as necessary
27
+ If max_length is provided, truncate the audio to that length
28
+ """
29
+ waveform, sample_rate = torchaudio.load(file_path)
30
+ if sample_rate != target_rate:
31
+ waveform = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_rate)(waveform)
32
+ audio = waveform[0] # get the first channel
33
+
34
+ # Truncate audio if it exceeds max_length
35
+ if max_length is not None and audio.shape[0] > max_length:
36
+ audio = audio[:max_length]
37
+
38
+ return audio
39
+
40
+ def log_mel_spectrogram(audio, n_mels=128, padding=479, device=None):
41
+ """
42
+ Compute the log-Mel spectrogram with specific padding for StepAudio
43
+ """
44
+ if not torch.is_tensor(audio):
45
+ if isinstance(audio, str):
46
+ audio = load_audio(audio)
47
+ audio = torch.from_numpy(audio)
48
+ if device is not None:
49
+ audio = audio.to(device)
50
+ if padding > 0:
51
+ audio = F.pad(audio, (0, padding))
52
+ window = torch.hann_window(400).to(audio.device)
53
+ stft = torch.stft(audio, 400, 160, window=window, return_complex=True)
54
+ magnitudes = stft[..., :-1].abs() ** 2
55
+ filters = _mel_filters(n_mels)
56
+ mel_spec = filters @ magnitudes
57
+
58
+ log_spec = torch.clamp(mel_spec, min=1e-10).log10()
59
+ log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
60
+ log_spec = (log_spec + 4.0) / 4.0
61
+ return log_spec
62
+
63
+ def compute_token_num(max_feature_len):
64
+ # First, audio goes through encoder:
65
+ # 1. conv1: kernel=3, stride=1, padding=1 -> size unchanged
66
+ # 2. conv2: kernel=3, stride=2, padding=1 -> size/2
67
+ # 3. avg_pooler: kernel=2, stride=2 -> size/2
68
+ max_feature_len = max_feature_len - 2 # remove padding
69
+ encoder_output_dim = (max_feature_len + 1) // 2 // 2 # after conv2 and avg_pooler
70
+
71
+ # Then through adaptor (parameters from config file):
72
+ padding = 1
73
+ kernel_size = 3 # from config: audio_encoder_config.kernel_size
74
+ stride = 2 # from config: audio_encoder_config.adapter_stride
75
+ adapter_output_dim = (encoder_output_dim + 2 * padding - kernel_size) // stride + 1
76
+ return adapter_output_dim
77
+
78
+ def make_non_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor:
79
+ """Make mask tensor containing indices of non-padded part.
80
+
81
+ The sequences in a batch may have different lengths. To enable
82
+ batch computing, padding is need to make all sequence in same
83
+ size. To avoid the padding part pass value to context dependent
84
+ block such as attention or convolution , this padding part is
85
+ masked.
86
+
87
+ 1 for non-padded part and 0 for padded part.
88
+
89
+ Parameters
90
+ ----------
91
+ lengths (torch.Tensor): Batch of lengths (B,).
92
+
93
+ Returns:
94
+ -------
95
+ torch.Tensor: Mask tensor containing indices of padded part (B, max_T).
96
+
97
+ Examples:
98
+ >>> import torch
99
+ >>> import s3tokenizer
100
+ >>> lengths = torch.tensor([5, 3, 2])
101
+ >>> masks = s3tokenizer.make_non_pad_mask(lengths)
102
+ masks = [[1, 1, 1, 1, 1],
103
+ [1, 1, 1, 0, 0],
104
+ [1, 1, 0, 0, 0]]
105
+ """
106
+ batch_size = lengths.size(0)
107
+ max_len = max_len if max_len > 0 else lengths.max().item()
108
+ seq_range = torch.arange(0,
109
+ max_len,
110
+ dtype=torch.int64,
111
+ device=lengths.device)
112
+ seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len)
113
+ seq_length_expand = lengths.unsqueeze(-1)
114
+ mask = seq_range_expand >= seq_length_expand
115
+ return ~mask
116
+
117
+ def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
118
+ """Convert bool-tensor to float-tensor for flash attention.
119
+
120
+ Parameters
121
+ ----------
122
+ lengths (torch.Tensor): Batch of lengths (B, ?).
123
+
124
+ Returns:
125
+ -------
126
+ torch.Tensor: Mask tensor containing indices of padded part (B, ?).
127
+
128
+ Examples:
129
+ >>> import torch
130
+ >>> import s3tokenizer
131
+ >>> lengths = torch.tensor([5, 3, 2])
132
+ >>> masks = s3tokenizer.make_non_pad_mask(lengths)
133
+ masks = [[1, 1, 1, 1, 1],
134
+ [1, 1, 1, 0, 0],
135
+ [1, 1, 0, 0, 0]]
136
+ >>> new_masks = s3tokenizer.mask_to_bias(masks, torch.float32)
137
+ new_masks = [[-0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00],
138
+ [-0.0000e+00, -0.0000e+00, -0.0000e+00, -1.0000e+10, -1.0000e+10],
139
+ [-0.0000e+00, -0.0000e+00, -1.0000e+10, -1.0000e+10, -1.0000e+10]]
140
+ """
141
+ assert mask.dtype == torch.bool
142
+ assert dtype in [torch.float32, torch.bfloat16, torch.float16]
143
+ mask = mask.to(dtype)
144
+ # attention mask bias
145
+ # NOTE(Mddct): torch.finfo jit issues
146
+ # chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min
147
+ mask = (1.0 - mask) * -1.0e+10
148
+ return mask
149
+
150
+ class LayerNorm(nn.LayerNorm):
151
+ def forward(self, input: Tensor) -> Tensor:
152
+ return super().forward(input).type(input.dtype)
153
+
154
+ class Linear(nn.Linear):
155
+ def forward(self, input: Tensor) -> Tensor:
156
+ return F.linear(
157
+ input,
158
+ self.weight.to(input.dtype),
159
+ None if self.bias is None else self.bias.to(input.dtype),
160
+ )
161
+
162
+ class Conv1d(nn.Conv1d):
163
+ def _conv_forward(
164
+ self, input: Tensor, weight: Tensor, bias: Optional[Tensor]
165
+ ) -> Tensor:
166
+ return super()._conv_forward(
167
+ input, weight.to(input.dtype), None if bias is None else bias.to(input.dtype)
168
+ )
169
+
170
+ class MultiHeadAttention(nn.Module):
171
+ def __init__(self, n_state: int, n_head: int):
172
+ super().__init__()
173
+ self.n_head = n_head
174
+ self.query = Linear(n_state, n_state)
175
+ self.key = Linear(n_state, n_state, bias=False)
176
+ self.value = Linear(n_state, n_state)
177
+ self.out = Linear(n_state, n_state)
178
+
179
+ def forward(
180
+ self,
181
+ x: Tensor,
182
+ mask: Optional[Tensor] = None,
183
+ ):
184
+ q = self.query(x)
185
+ k = self.key(x)
186
+ v = self.value(x)
187
+
188
+ wv, qk = self.qkv_attention(q, k, v, mask)
189
+ return self.out(wv), qk
190
+
191
+ def qkv_attention(
192
+ self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None
193
+ ):
194
+ _, T, D = q.shape
195
+ scale = (D // self.n_head) ** -0.25
196
+ q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale
197
+ k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 3, 1) * scale
198
+ v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
199
+
200
+ qk = q @ k # (B, n_head, T, T)
201
+ if mask is not None:
202
+ qk = qk + mask
203
+ qk = qk.float()
204
+
205
+ w = F.softmax(qk, dim=-1).to(q.dtype)
206
+ return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()
207
+
208
+ class ResidualAttentionBlock(nn.Module):
209
+ def __init__(self, n_state: int, n_head: int):
210
+ super().__init__()
211
+
212
+ self.attn = MultiHeadAttention(n_state, n_head)
213
+ self.attn_ln = LayerNorm(n_state)
214
+
215
+ n_mlp = n_state * 4
216
+ self.mlp = nn.Sequential(
217
+ Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state)
218
+ )
219
+ self.mlp_ln = LayerNorm(n_state)
220
+
221
+ def forward(
222
+ self,
223
+ x: Tensor,
224
+ mask: Optional[Tensor] = None,
225
+ ):
226
+ x = x + self.attn(self.attn_ln(x.contiguous()), mask=mask)[0]
227
+ x = x + self.mlp(self.mlp_ln(x.contiguous()))
228
+ return x
229
+
230
+ class AudioEncoder(nn.Module):
231
+ def __init__(
232
+ self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int
233
+ ):
234
+ super().__init__()
235
+ self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
236
+ self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
237
+ self.positional_embedding = nn.Embedding(n_ctx, n_state)
238
+ self.positional_embedding.requires_grad_(False)
239
+ self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
240
+ [ResidualAttentionBlock(n_state, n_head) for _ in range(n_layer)]
241
+ )
242
+ self.avg_pooler = nn.AvgPool1d(2, stride=2)
243
+ self.after_norm = LayerNorm(n_state)
244
+ self.gradient_checkpointing = False
245
+
246
+ def forward(self, x: Tensor, x_len: Tensor) -> Tuple[Tensor, Tensor]:
247
+ T = x.size(-1)
248
+ x = F.gelu(self.conv1(x))
249
+ x = F.gelu(self.conv2(x))
250
+ x = x.permute(0, 2, 1) # (B, T // 2, n_state)
251
+ mask = make_non_pad_mask(x_len, T).unsqueeze(1) # (B, 1, T)
252
+ mask = mask_to_bias(mask[:, :, (T + 1) % 2::2], x.dtype) # (B, 1, T // 2)
253
+ x = (x + self.positional_embedding.weight[:x.shape[1], :]).to(x.dtype)
254
+ for block in self.blocks:
255
+ if self.gradient_checkpointing and self.training:
256
+ x = torch.utils.checkpoint.checkpoint(block, x, mask.unsqueeze(1))
257
+ else:
258
+ x = block(x, mask.unsqueeze(1))
259
+ x = x.permute(0, 2, 1)
260
+ x = self.avg_pooler(x)
261
+ x = x.permute(0, 2, 1)
262
+ x_len = (x_len + 1) // 2 // 2
263
+ x = self.after_norm(x.contiguous())
264
+ return x, x_len
265
+
266
+ class Adaptor(nn.Module):
267
+ def __init__(
268
+ self,
269
+ n_state: int = 1280,
270
+ n_hidden: int = 3072,
271
+ kernel_size: int = 7,
272
+ stride: int = 4
273
+ ):
274
+ super().__init__()
275
+ self.stride = stride
276
+ if self.stride != -1:
277
+ # print("self.stride: {}".format(self.stride))
278
+ self.conv = Conv1d(n_state, n_state, kernel_size, stride, padding=1)
279
+ self.linear1 = nn.Linear(n_state, 2048)
280
+ self.relu = nn.ReLU()
281
+ self.linear2 = nn.Linear(2048, n_hidden)
282
+ self.gradient_checkpointing = False
283
+
284
+ def forward(self, x: Tensor) -> Tuple[Tensor]:
285
+ T = x.size(-1)
286
+ if self.stride != -1:
287
+ if self.gradient_checkpointing and self.training:
288
+ x = torch.utils.checkpoint.checkpoint(self.conv, x.permute(0, 2, 1))
289
+ x = x.permute(0, 2, 1)
290
+ else:
291
+ x = x.permute(0, 2, 1)
292
+ x = F.gelu(self.conv(x))
293
+ x = x.permute(0, 2, 1)
294
+ if self.gradient_checkpointing and self.training:
295
+ x = torch.utils.checkpoint.checkpoint(self.linear1, x)
296
+ x = torch.utils.checkpoint.checkpoint(self.relu, x)
297
+ x = torch.utils.checkpoint.checkpoint(self.linear2, x)
298
+ else:
299
+ x = self.linear1(x)
300
+ x = self.relu(x)
301
+ x = self.linear2(x)
302
+ return x
303
+
304
+ class StepAudio2ForCausalLM(PreTrainedModel, GenerationMixin):
305
+ config_class = StepAudio2Config
306
+ main_input_name = "input_ids"
307
+ # Important: Add this attribute to make HF recognize it as a model with generation capability
308
+ # _keys_to_ignore_on_load_missing = ["lm_head.weight"]
309
+ supports_gradient_checkpointing = True # 新增，声明支持gradient checkpointing
310
+
311
+ def __init__(self, config: StepAudio2Config):
312
+ super().__init__(config)
313
+ if isinstance(config.torch_dtype, str):
314
+ dtype = getattr(torch, config.torch_dtype)
315
+ else:
316
+ dtype = config.torch_dtype
317
+ self.model = Qwen2Model(config.text_config)
318
+ self.bf16 = dtype==torch.bfloat16
319
+ self.encoder = AudioEncoder(
320
+ config.audio_encoder_config.n_mels, config.audio_encoder_config.n_audio_ctx, config.audio_encoder_config.n_audio_state,
321
+ config.audio_encoder_config.n_audio_head, config.audio_encoder_config.n_audio_layer
322
+ )
323
+ self.adapter = Adaptor(
324
+ config.audio_encoder_config.n_audio_state, config.audio_encoder_config.llm_dim,
325
+ config.audio_encoder_config.kernel_size, config.audio_encoder_config.adapter_stride
326
+ )
327
+ if self.bf16:
328
+ self.encoder = self.encoder.bfloat16()
329
+ self.adapter = self.adapter.bfloat16()
330
+ self.lm_head = torch.nn.Linear(
331
+ config.text_config.hidden_size,
332
+ config.text_config.vocab_size,
333
+ bias=False,
334
+ dtype=dtype
335
+ )
336
+ self.post_init()
337
+
338
+ def forward(
339
+ self,
340
+ input_ids=None,
341
+ wavs=None,
342
+ wav_lens=None,
343
+ attention_mask=None,
344
+ **kwargs
345
+ ):
346
+ hidden_states = self.model.embed_tokens(input_ids)
347
+ if wavs is not None:
348
+ if self.bf16:
349
+ wavs = wavs.bfloat16()
350
+ out, feat_lens = self.encoder(wavs, wav_lens)
351
+ out = self.adapter(out)
352
+ feat_lens = (feat_lens - 1) // 2 + 1
353
+ insert_location = torch.nonzero(input_ids == 151688)
354
+ insert_location[:,1] += 1
355
+ for idx in range(len(insert_location)):
356
+ i,s = insert_location[idx]
357
+ hidden_states[i][s : s+feat_lens[idx]] = out[idx][:feat_lens[idx]]
358
+
359
+ x = self.model(inputs_embeds=hidden_states, attention_mask=attention_mask)[0]
360
+ logits = self.lm_head(x)
361
+ return CausalLMOutputWithPast(
362
+ logits=logits,
363
+ past_key_values=None,
364
+ hidden_states=None,
365
+ attentions=None
366
+ )
367
+
368
+ def get_input_embeddings(self):
369
+ """Return the model's input embeddings - required for GenerationMixin"""
370
+ return self.model.embed_tokens
371
+
372
+ def get_output_embeddings(self):
373
+ """Return the model's output embeddings (LM head) - required for GenerationMixin"""
374
+ return self.lm_head
375
+
376
+ def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **kwargs):
377
+ """Prepare inputs for generation - required for GenerationMixin"""
378
+ # Keep the wavs and wav_lens from the initial call
379
+ wavs = kwargs.get("wavs", None)
380
+ wav_lens = kwargs.get("wav_lens", None)
381
+
382
+ # For generation steps after the first, we don't need to process audio again
383
+ # because the audio tokens have already been replaced in the input sequence
384
+ if "past_key_values" in kwargs and kwargs["past_key_values"] is not None:
385
+ # We're in a generation step, no need to process audio again
386
+ return {
387
+ "input_ids": input_ids,
388
+ "attention_mask": attention_mask,
389
+ "past_key_values": kwargs.get("past_key_values")
390
+ }
391
+
392
+ # First generation step, include audio processing
393
+ return {
394
+ "input_ids": input_ids,
395
+ "attention_mask": attention_mask,
396
+ "wavs": wavs,
397
+ "wav_lens": wav_lens
398
+ }
399
+
400
+ def _reorder_cache(self, past_key_values, beam_idx):
401
+ """Reorder the cache for beam search - required for GenerationMixin if using beam search"""
402
+ # If you're not using past_key_values or beam search, this can be a simple pass-through
403
+ # Otherwise implement according to your model's cache structure
404
+ return past_key_values
405
+
406
+ def _set_gradient_checkpointing(self, module, value=False):
407
+ # For Qwen2Model
408
+ if hasattr(self.model, 'gradient_checkpointing'):
409
+ self.model.gradient_checkpointing = value
410
+
411
+ # Add the missing _gradient_checkpointing_func method to Qwen2Model
412
+ # This is what Qwen2Model tries to use when gradient_checkpointing=True
413
+ if value and not hasattr(self.model, '_gradient_checkpointing_func'):
414
+ def _gradient_checkpointing_func(module_to_run, *args, **kwargs):
415
+ # This function wraps torch.utils.checkpoint.checkpoint
416
+ # and is used by Qwen2Model to perform checkpointing
417
+ return torch.utils.checkpoint.checkpoint(module_to_run, *args, **kwargs)
418
+
419
+ self.model._gradient_checkpointing_func = _gradient_checkpointing_func
420
+
421
+ # For custom encoder and adapter
422
+ if hasattr(self.encoder, 'gradient_checkpointing'):
423
+ self.encoder.gradient_checkpointing = value
424
+ if hasattr(self.adapter, 'gradient_checkpointing'):
425
+ self.adapter.gradient_checkpointing = value
lychee_full_duplex_v1.5/special_tokens_map.json ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|EOT|>",
4
+ "<|BOT|>",
5
+ "<|CALL_START|>",
6
+ "<|CALL_END|>",
7
+ "<|THINK_START|>",
8
+ "<|THINK_END|>",
9
+ "<|IMG_START|>",
10
+ "<|IMG_END|>",
11
+ "<|META_START|>",
12
+ "<|META_END|>",
13
+ "<im_patch>",
14
+ "<im_start>",
15
+ "<im_end>",
16
+ "<dream>",
17
+ "<dream_start>",
18
+ "<dream_end>",
19
+ "<|MASK_1e69f|>",
20
+ "<|UNMASK_1e69f|>",
21
+ "<video_start>",
22
+ "<video_end>",
23
+ "<patch_start>",
24
+ "<patch_end>",
25
+ "<patch_newline>",
26
+ "<audio_start>",
27
+ "<audio_end>",
28
+ "<audio_patch>",
29
+ "<audio_patch_pad>",
30
+ "<|SC|>",
31
+ "<tts_start>",
32
+ "<tts_end>",
33
+ "<tts_pad>",
34
+ "<|S-S|>",
35
+ "<|S-L|>",
36
+ "<|K-L|>",
37
+ "<|K-S|>",
38
+ "<|Detect|>",
39
+ "<|Sleep|>",
40
+ "<|TextPad|>",
41
+ "<|StokenPad|>",
42
+ "<|AudioPad|>",
43
+ "<|StokenDelay|>",
44
+ "<|BackChannel|>"
45
+ ],
46
+ "eos_token": {
47
+ "content": "<|EOT|>",
48
+ "lstrip": false,
49
+ "normalized": false,
50
+ "rstrip": false,
51
+ "single_word": false
52
+ },
53
+ "pad_token": {
54
+ "content": "<|endoftext|>",
55
+ "lstrip": false,
56
+ "normalized": false,
57
+ "rstrip": false,
58
+ "single_word": false
59
+ }
60
+ }
lychee_full_duplex_v1.5/tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff
 
lychee_full_duplex_v1.5/vocab.json ADDED
The diff for this file is too large to render. See raw diff