Hugging Face's logo

GreenBitAI
/
Phi-3-mini-128k-instruct-layer-mix-bpw-3.0-mlx

MLX
Safetensors
phi3
custom_code
Model card Files Community
yanghaojin commited on
Commit
8b201dd
•
1 Parent(s): 481be8d

Upload folder using huggingface_hub

Browse files
Files changed (10) hide show
  1. README.md +21 -0
  2. added_tokens.json +40 -0
  3. config.json +193 -0
  4. model.safetensors +3 -0
  5. model.safetensors.index.json +714 -0
  6. quant_strategy.json +1668 -0
  7. special_tokens_map.json +33 -0
  8. tokenizer.json +0 -0
  9. tokenizer.model +3 -0
  10. tokenizer_config.json +351 -0
README.md ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - mlx
5
+ ---
6
+
7
+ # GreenBitAI/Phi-3-mini-128k-instruct-layer-mix-bpw-3.0-mlx
8
+ This quantized low-bit model was converted to MLX format from [`GreenBitAI/Phi-3-mini-128k-instruct-layer-mix-bpw-3.0`]().
9
+ Refer to the [original model card](https://huggingface.co/GreenBitAI/Phi-3-mini-128k-instruct-layer-mix-bpw-3.0) for more details on the model.
10
+ ## Use with mlx
11
+
12
+ ```bash
13
+ pip install gbx-lm
14
+ ```
15
+
16
+ ```python
17
+ from gbx_lm import load, generate
18
+
19
+ model, tokenizer = load("GreenBitAI/Phi-3-mini-128k-instruct-layer-mix-bpw-3.0-mlx")
20
+ response = generate(model, tokenizer, prompt="hello", verbose=True)
21
+ ```
added_tokens.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "<|/code|>": 32014,
3
+ "<|/data|>": 32033,
4
+ "<|/inst|>": 32037,
5
+ "<|/query|>": 32031,
6
+ "<|/sys|>": 32035,
7
+ "<|assistant_mask|>": 32017,
8
+ "<|assistant|>": 32001,
9
+ "<|calc|>": 32012,
10
+ "<|code|>": 32013,
11
+ "<|continue|>": 32009,
12
+ "<|data|>": 32032,
13
+ "<|diff_marker|>": 32025,
14
+ "<|disc_sep|>": 32029,
15
+ "<|disc_start|>": 32028,
16
+ "<|disc_thread|><|query|>": 32030,
17
+ "<|endoftext|>": 32000,
18
+ "<|end|>": 32007,
19
+ "<|fim_middle|>": 32021,
20
+ "<|fim_prefix|>": 32020,
21
+ "<|fim_suffix|>": 32022,
22
+ "<|function_call|>": 32005,
23
+ "<|function_list|>": 32011,
24
+ "<|function_output|>": 32003,
25
+ "<|ghissue|>": 32026,
26
+ "<|ghreview|>": 32027,
27
+ "<|inst|>": 32036,
28
+ "<|ipynb_marker|>": 32024,
29
+ "<|message|>": 32019,
30
+ "<|meta_start|>": 32023,
31
+ "<|raw|>": 32008,
32
+ "<|resource|>": 32016,
33
+ "<|start|>": 32018,
34
+ "<|step|>": 32002,
35
+ "<|summary|>": 32015,
36
+ "<|system|>": 32006,
37
+ "<|sys|>": 32034,
38
+ "<|tag|>": 32004,
39
+ "<|user|>": 32010
40
+ }
config.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "Phi3ForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_phi3.Phi3Config",
9
+ "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM"
10
+ },
11
+ "bad_words_ids": null,
12
+ "begin_suppress_tokens": null,
13
+ "bos_token_id": 1,
14
+ "chunk_size_feed_forward": 0,
15
+ "cross_attention_hidden_size": null,
16
+ "decoder_start_token_id": null,
17
+ "diversity_penalty": 0.0,
18
+ "do_sample": false,
19
+ "early_stopping": false,
20
+ "embd_pdrop": 0.0,
21
+ "encoder_no_repeat_ngram_size": 0,
22
+ "eos_token_id": 32000,
23
+ "exponential_decay_length_penalty": null,
24
+ "finetuning_task": null,
25
+ "forced_bos_token_id": null,
26
+ "forced_eos_token_id": null,
27
+ "hidden_act": "silu",
28
+ "hidden_size": 3072,
29
+ "id2label": {
30
+ "0": "LABEL_0",
31
+ "1": "LABEL_1"
32
+ },
33
+ "initializer_range": 0.02,
34
+ "intermediate_size": 8192,
35
+ "is_decoder": false,
36
+ "is_encoder_decoder": false,
37
+ "label2id": {
38
+ "LABEL_0": 0,
39
+ "LABEL_1": 1
40
+ },
41
+ "length_penalty": 1.0,
42
+ "max_length": 20,
43
+ "max_position_embeddings": 131072,
44
+ "min_length": 0,
45
+ "model_type": "phi3",
46
+ "no_repeat_ngram_size": 0,
47
+ "num_attention_heads": 32,
48
+ "num_beam_groups": 1,
49
+ "num_beams": 1,
50
+ "num_hidden_layers": 32,
51
+ "num_key_value_heads": 32,
52
+ "num_return_sequences": 1,
53
+ "original_max_position_embeddings": 4096,
54
+ "output_attentions": false,
55
+ "output_hidden_states": false,
56
+ "output_scores": false,
57
+ "pad_token_id": 32000,
58
+ "prefix": null,
59
+ "problem_type": null,
60
+ "pruned_heads": {},
61
+ "quantization": {
62
+ "group_size": null,
63
+ "bits": null
64
+ },
65
+ "remove_invalid_values": false,
66
+ "repetition_penalty": 1.0,
67
+ "resid_pdrop": 0.0,
68
+ "return_dict": true,
69
+ "return_dict_in_generate": false,
70
+ "rms_norm_eps": 1e-05,
71
+ "rope_scaling": {
72
+ "long_factor": [
73
+ 1.0299999713897705,
74
+ 1.0499999523162842,
75
+ 1.0499999523162842,
76
+ 1.0799999237060547,
77
+ 1.2299998998641968,
78
+ 1.2299998998641968,
79
+ 1.2999999523162842,
80
+ 1.4499999284744263,
81
+ 1.5999999046325684,
82
+ 1.6499998569488525,
83
+ 1.8999998569488525,
84
+ 2.859999895095825,
85
+ 3.68999981880188,
86
+ 5.419999599456787,
87
+ 5.489999771118164,
88
+ 5.489999771118164,
89
+ 9.09000015258789,
90
+ 11.579999923706055,
91
+ 15.65999984741211,
92
+ 15.769999504089355,
93
+ 15.789999961853027,
94
+ 18.360000610351562,
95
+ 21.989999771118164,
96
+ 23.079999923706055,
97
+ 30.009998321533203,
98
+ 32.35000228881836,
99
+ 32.590003967285156,
100
+ 35.56000518798828,
101
+ 39.95000457763672,
102
+ 53.840003967285156,
103
+ 56.20000457763672,
104
+ 57.95000457763672,
105
+ 59.29000473022461,
106
+ 59.77000427246094,
107
+ 59.920005798339844,
108
+ 61.190006256103516,
109
+ 61.96000671386719,
110
+ 62.50000762939453,
111
+ 63.3700065612793,
112
+ 63.48000717163086,
113
+ 63.48000717163086,
114
+ 63.66000747680664,
115
+ 63.850006103515625,
116
+ 64.08000946044922,
117
+ 64.760009765625,
118
+ 64.80001068115234,
119
+ 64.81001281738281,
120
+ 64.81001281738281
121
+ ],
122
+ "short_factor": [
123
+ 1.05,
124
+ 1.05,
125
+ 1.05,
126
+ 1.1,
127
+ 1.1,
128
+ 1.1500000000000001,
129
+ 1.2000000000000002,
130
+ 1.2500000000000002,
131
+ 1.3000000000000003,
132
+ 1.3500000000000003,
133
+ 1.5000000000000004,
134
+ 2.000000000000001,
135
+ 2.000000000000001,
136
+ 2.000000000000001,
137
+ 2.000000000000001,
138
+ 2.000000000000001,
139
+ 2.000000000000001,
140
+ 2.000000000000001,
141
+ 2.000000000000001,
142
+ 2.000000000000001,
143
+ 2.000000000000001,
144
+ 2.000000000000001,
145
+ 2.000000000000001,
146
+ 2.000000000000001,
147
+ 2.000000000000001,
148
+ 2.000000000000001,
149
+ 2.000000000000001,
150
+ 2.000000000000001,
151
+ 2.000000000000001,
152
+ 2.000000000000001,
153
+ 2.000000000000001,
154
+ 2.000000000000001,
155
+ 2.0500000000000007,
156
+ 2.0500000000000007,
157
+ 2.0500000000000007,
158
+ 2.1000000000000005,
159
+ 2.1000000000000005,
160
+ 2.1000000000000005,
161
+ 2.1500000000000004,
162
+ 2.1500000000000004,
163
+ 2.3499999999999996,
164
+ 2.549999999999999,
165
+ 2.5999999999999988,
166
+ 2.5999999999999988,
167
+ 2.7499999999999982,
168
+ 2.849999999999998,
169
+ 2.849999999999998,
170
+ 2.9499999999999975
171
+ ],
172
+ "type": "su"
173
+ },
174
+ "rope_theta": 10000.0,
175
+ "sep_token_id": null,
176
+ "sliding_window": 262144,
177
+ "suppress_tokens": null,
178
+ "task_specific_params": null,
179
+ "temperature": 1.0,
180
+ "tf_legacy_loss": false,
181
+ "tie_encoder_decoder": false,
182
+ "tie_word_embeddings": false,
183
+ "tokenizer_class": null,
184
+ "top_k": 50,
185
+ "top_p": 1.0,
186
+ "torch_dtype": "float16",
187
+ "torchscript": false,
188
+ "transformers_version": "4.39.0.dev0",
189
+ "typical_p": 1.0,
190
+ "use_bfloat16": false,
191
+ "use_cache": true,
192
+ "vocab_size": 32064
193
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b3b6ecec5990d2d75be5eb52938f9fed282d1e0a40e7a5ea3c6a5867e5860342
3
+ size 1903515059
model.safetensors.index.json ADDED
@@ -0,0 +1,714 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 1903433728
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model.safetensors",
7
+ "model.embed_tokens.weight": "model.safetensors",
8
+ "model.layers.0.input_layernorm.weight": "model.safetensors",
9
+ "model.layers.0.mlp.down_proj.channel_scale": "model.safetensors",
10
+ "model.layers.0.mlp.down_proj.q_perm": "model.safetensors",
11
+ "model.layers.0.mlp.down_proj.qweight": "model.safetensors",
12
+ "model.layers.0.mlp.down_proj.scales": "model.safetensors",
13
+ "model.layers.0.mlp.down_proj.zeros": "model.safetensors",
14
+ "model.layers.0.mlp.gate_up_proj.channel_scale": "model.safetensors",
15
+ "model.layers.0.mlp.gate_up_proj.q_perm": "model.safetensors",
16
+ "model.layers.0.mlp.gate_up_proj.qweight": "model.safetensors",
17
+ "model.layers.0.mlp.gate_up_proj.scales": "model.safetensors",
18
+ "model.layers.0.mlp.gate_up_proj.zeros": "model.safetensors",
19
+ "model.layers.0.post_attention_layernorm.weight": "model.safetensors",
20
+ "model.layers.0.self_attn.o_proj.channel_scale": "model.safetensors",
21
+ "model.layers.0.self_attn.o_proj.q_perm": "model.safetensors",
22
+ "model.layers.0.self_attn.o_proj.qweight": "model.safetensors",
23
+ "model.layers.0.self_attn.o_proj.scales": "model.safetensors",
24
+ "model.layers.0.self_attn.o_proj.zeros": "model.safetensors",
25
+ "model.layers.0.self_attn.qkv_proj.channel_scale": "model.safetensors",
26
+ "model.layers.0.self_attn.qkv_proj.q_perm": "model.safetensors",
27
+ "model.layers.0.self_attn.qkv_proj.qweight": "model.safetensors",
28
+ "model.layers.0.self_attn.qkv_proj.scales": "model.safetensors",
29
+ "model.layers.0.self_attn.qkv_proj.zeros": "model.safetensors",
30
+ "model.layers.1.input_layernorm.weight": "model.safetensors",
31
+ "model.layers.1.mlp.down_proj.channel_scale": "model.safetensors",
32
+ "model.layers.1.mlp.down_proj.q_perm": "model.safetensors",
33
+ "model.layers.1.mlp.down_proj.qweight": "model.safetensors",
34
+ "model.layers.1.mlp.down_proj.scales": "model.safetensors",
35
+ "model.layers.1.mlp.down_proj.zeros": "model.safetensors",
36
+ "model.layers.1.mlp.gate_up_proj.channel_scale": "model.safetensors",
37
+ "model.layers.1.mlp.gate_up_proj.q_perm": "model.safetensors",
38
+ "model.layers.1.mlp.gate_up_proj.qweight": "model.safetensors",
39
+ "model.layers.1.mlp.gate_up_proj.scales": "model.safetensors",
40
+ "model.layers.1.mlp.gate_up_proj.zeros": "model.safetensors",
41
+ "model.layers.1.post_attention_layernorm.weight": "model.safetensors",
42
+ "model.layers.1.self_attn.o_proj.channel_scale": "model.safetensors",
43
+ "model.layers.1.self_attn.o_proj.q_perm": "model.safetensors",
44
+ "model.layers.1.self_attn.o_proj.qweight": "model.safetensors",
45
+ "model.layers.1.self_attn.o_proj.scales": "model.safetensors",
46
+ "model.layers.1.self_attn.o_proj.zeros": "model.safetensors",
47
+ "model.layers.1.self_attn.qkv_proj.channel_scale": "model.safetensors",
48
+ "model.layers.1.self_attn.qkv_proj.q_perm": "model.safetensors",
49
+ "model.layers.1.self_attn.qkv_proj.qweight": "model.safetensors",
50
+ "model.layers.1.self_attn.qkv_proj.scales": "model.safetensors",
51
+ "model.layers.1.self_attn.qkv_proj.zeros": "model.safetensors",
52
+ "model.layers.10.input_layernorm.weight": "model.safetensors",
53
+ "model.layers.10.mlp.down_proj.channel_scale": "model.safetensors",
54
+ "model.layers.10.mlp.down_proj.q_perm": "model.safetensors",
55
+ "model.layers.10.mlp.down_proj.qweight": "model.safetensors",
56
+ "model.layers.10.mlp.down_proj.scales": "model.safetensors",
57
+ "model.layers.10.mlp.down_proj.zeros": "model.safetensors",
58
+ "model.layers.10.mlp.gate_up_proj.channel_scale": "model.safetensors",
59
+ "model.layers.10.mlp.gate_up_proj.q_perm": "model.safetensors",
60
+ "model.layers.10.mlp.gate_up_proj.qweight": "model.safetensors",
61
+ "model.layers.10.mlp.gate_up_proj.scales": "model.safetensors",
62
+ "model.layers.10.mlp.gate_up_proj.zeros": "model.safetensors",
63
+ "model.layers.10.post_attention_layernorm.weight": "model.safetensors",
64
+ "model.layers.10.self_attn.o_proj.channel_scale": "model.safetensors",
65
+ "model.layers.10.self_attn.o_proj.q_perm": "model.safetensors",
66
+ "model.layers.10.self_attn.o_proj.qweight": "model.safetensors",
67
+ "model.layers.10.self_attn.o_proj.scales": "model.safetensors",
68
+ "model.layers.10.self_attn.o_proj.zeros": "model.safetensors",
69
+ "model.layers.10.self_attn.qkv_proj.channel_scale": "model.safetensors",
70
+ "model.layers.10.self_attn.qkv_proj.q_perm": "model.safetensors",
71
+ "model.layers.10.self_attn.qkv_proj.qweight": "model.safetensors",
72
+ "model.layers.10.self_attn.qkv_proj.scales": "model.safetensors",
73
+ "model.layers.10.self_attn.qkv_proj.zeros": "model.safetensors",
74
+ "model.layers.11.input_layernorm.weight": "model.safetensors",
75
+ "model.layers.11.mlp.down_proj.channel_scale": "model.safetensors",
76
+ "model.layers.11.mlp.down_proj.q_perm": "model.safetensors",
77
+ "model.layers.11.mlp.down_proj.qweight": "model.safetensors",
78
+ "model.layers.11.mlp.down_proj.scales": "model.safetensors",
79
+ "model.layers.11.mlp.down_proj.zeros": "model.safetensors",
80
+ "model.layers.11.mlp.gate_up_proj.channel_scale": "model.safetensors",
81
+ "model.layers.11.mlp.gate_up_proj.q_perm": "model.safetensors",
82
+ "model.layers.11.mlp.gate_up_proj.qweight": "model.safetensors",
83
+ "model.layers.11.mlp.gate_up_proj.scales": "model.safetensors",
84
+ "model.layers.11.mlp.gate_up_proj.zeros": "model.safetensors",
85
+ "model.layers.11.post_attention_layernorm.weight": "model.safetensors",
86
+ "model.layers.11.self_attn.o_proj.channel_scale": "model.safetensors",
87
+ "model.layers.11.self_attn.o_proj.q_perm": "model.safetensors",
88
+ "model.layers.11.self_attn.o_proj.qweight": "model.safetensors",
89
+ "model.layers.11.self_attn.o_proj.scales": "model.safetensors",
90
+ "model.layers.11.self_attn.o_proj.zeros": "model.safetensors",
91
+ "model.layers.11.self_attn.qkv_proj.channel_scale": "model.safetensors",
92
+ "model.layers.11.self_attn.qkv_proj.q_perm": "model.safetensors",
93
+ "model.layers.11.self_attn.qkv_proj.qweight": "model.safetensors",
94
+ "model.layers.11.self_attn.qkv_proj.scales": "model.safetensors",
95
+ "model.layers.11.self_attn.qkv_proj.zeros": "model.safetensors",
96
+ "model.layers.12.input_layernorm.weight": "model.safetensors",
97
+ "model.layers.12.mlp.down_proj.channel_scale": "model.safetensors",
98
+ "model.layers.12.mlp.down_proj.q_perm": "model.safetensors",
99
+ "model.layers.12.mlp.down_proj.qweight": "model.safetensors",
100
+ "model.layers.12.mlp.down_proj.scales": "model.safetensors",
101
+ "model.layers.12.mlp.down_proj.zeros": "model.safetensors",
102
+ "model.layers.12.mlp.gate_up_proj.channel_scale": "model.safetensors",
103
+ "model.layers.12.mlp.gate_up_proj.q_perm": "model.safetensors",
104
+ "model.layers.12.mlp.gate_up_proj.qweight": "model.safetensors",
105
+ "model.layers.12.mlp.gate_up_proj.scales": "model.safetensors",
106
+ "model.layers.12.mlp.gate_up_proj.zeros": "model.safetensors",
107
+ "model.layers.12.post_attention_layernorm.weight": "model.safetensors",
108
+ "model.layers.12.self_attn.o_proj.channel_scale": "model.safetensors",
109
+ "model.layers.12.self_attn.o_proj.q_perm": "model.safetensors",
110
+ "model.layers.12.self_attn.o_proj.qweight": "model.safetensors",
111
+ "model.layers.12.self_attn.o_proj.scales": "model.safetensors",
112
+ "model.layers.12.self_attn.o_proj.zeros": "model.safetensors",
113
+ "model.layers.12.self_attn.qkv_proj.channel_scale": "model.safetensors",
114
+ "model.layers.12.self_attn.qkv_proj.q_perm": "model.safetensors",
115
+ "model.layers.12.self_attn.qkv_proj.qweight": "model.safetensors",
116
+ "model.layers.12.self_attn.qkv_proj.scales": "model.safetensors",
117
+ "model.layers.12.self_attn.qkv_proj.zeros": "model.safetensors",
118
+ "model.layers.13.input_layernorm.weight": "model.safetensors",
119
+ "model.layers.13.mlp.down_proj.channel_scale": "model.safetensors",
120
+ "model.layers.13.mlp.down_proj.q_perm": "model.safetensors",
121
+ "model.layers.13.mlp.down_proj.qweight": "model.safetensors",
122
+ "model.layers.13.mlp.down_proj.scales": "model.safetensors",
123
+ "model.layers.13.mlp.down_proj.zeros": "model.safetensors",
124
+ "model.layers.13.mlp.gate_up_proj.channel_scale": "model.safetensors",
125
+ "model.layers.13.mlp.gate_up_proj.q_perm": "model.safetensors",
126
+ "model.layers.13.mlp.gate_up_proj.qweight": "model.safetensors",
127
+ "model.layers.13.mlp.gate_up_proj.scales": "model.safetensors",
128
+ "model.layers.13.mlp.gate_up_proj.zeros": "model.safetensors",
129
+ "model.layers.13.post_attention_layernorm.weight": "model.safetensors",
130
+ "model.layers.13.self_attn.o_proj.channel_scale": "model.safetensors",
131
+ "model.layers.13.self_attn.o_proj.q_perm": "model.safetensors",
132
+ "model.layers.13.self_attn.o_proj.qweight": "model.safetensors",
133
+ "model.layers.13.self_attn.o_proj.scales": "model.safetensors",
134
+ "model.layers.13.self_attn.o_proj.zeros": "model.safetensors",
135
+ "model.layers.13.self_attn.qkv_proj.channel_scale": "model.safetensors",
136
+ "model.layers.13.self_attn.qkv_proj.q_perm": "model.safetensors",
137
+ "model.layers.13.self_attn.qkv_proj.qweight": "model.safetensors",
138
+ "model.layers.13.self_attn.qkv_proj.scales": "model.safetensors",
139
+ "model.layers.13.self_attn.qkv_proj.zeros": "model.safetensors",
140
+ "model.layers.14.input_layernorm.weight": "model.safetensors",
141
+ "model.layers.14.mlp.down_proj.channel_scale": "model.safetensors",
142
+ "model.layers.14.mlp.down_proj.q_perm": "model.safetensors",
143
+ "model.layers.14.mlp.down_proj.qweight": "model.safetensors",
144
+ "model.layers.14.mlp.down_proj.scales": "model.safetensors",
145
+ "model.layers.14.mlp.down_proj.zeros": "model.safetensors",
146
+ "model.layers.14.mlp.gate_up_proj.channel_scale": "model.safetensors",
147
+ "model.layers.14.mlp.gate_up_proj.q_perm": "model.safetensors",
148
+ "model.layers.14.mlp.gate_up_proj.qweight": "model.safetensors",
149
+ "model.layers.14.mlp.gate_up_proj.scales": "model.safetensors",
150
+ "model.layers.14.mlp.gate_up_proj.zeros": "model.safetensors",
151
+ "model.layers.14.post_attention_layernorm.weight": "model.safetensors",
152
+ "model.layers.14.self_attn.o_proj.channel_scale": "model.safetensors",
153
+ "model.layers.14.self_attn.o_proj.q_perm": "model.safetensors",
154
+ "model.layers.14.self_attn.o_proj.qweight": "model.safetensors",
155
+ "model.layers.14.self_attn.o_proj.scales": "model.safetensors",
156
+ "model.layers.14.self_attn.o_proj.zeros": "model.safetensors",
157
+ "model.layers.14.self_attn.qkv_proj.channel_scale": "model.safetensors",
158
+ "model.layers.14.self_attn.qkv_proj.q_perm": "model.safetensors",
159
+ "model.layers.14.self_attn.qkv_proj.qweight": "model.safetensors",
160
+ "model.layers.14.self_attn.qkv_proj.scales": "model.safetensors",
161
+ "model.layers.14.self_attn.qkv_proj.zeros": "model.safetensors",
162
+ "model.layers.15.input_layernorm.weight": "model.safetensors",
163
+ "model.layers.15.mlp.down_proj.channel_scale": "model.safetensors",
164
+ "model.layers.15.mlp.down_proj.q_perm": "model.safetensors",
165
+ "model.layers.15.mlp.down_proj.qweight": "model.safetensors",
166
+ "model.layers.15.mlp.down_proj.scales": "model.safetensors",
167
+ "model.layers.15.mlp.down_proj.zeros": "model.safetensors",
168
+ "model.layers.15.mlp.gate_up_proj.channel_scale": "model.safetensors",
169
+ "model.layers.15.mlp.gate_up_proj.q_perm": "model.safetensors",
170
+ "model.layers.15.mlp.gate_up_proj.qweight": "model.safetensors",
171
+ "model.layers.15.mlp.gate_up_proj.scales": "model.safetensors",
172
+ "model.layers.15.mlp.gate_up_proj.zeros": "model.safetensors",
173
+ "model.layers.15.post_attention_layernorm.weight": "model.safetensors",
174
+ "model.layers.15.self_attn.o_proj.channel_scale": "model.safetensors",
175
+ "model.layers.15.self_attn.o_proj.q_perm": "model.safetensors",
176
+ "model.layers.15.self_attn.o_proj.qweight": "model.safetensors",
177
+ "model.layers.15.self_attn.o_proj.scales": "model.safetensors",
178
+ "model.layers.15.self_attn.o_proj.zeros": "model.safetensors",
179
+ "model.layers.15.self_attn.qkv_proj.channel_scale": "model.safetensors",
180
+ "model.layers.15.self_attn.qkv_proj.q_perm": "model.safetensors",
181
+ "model.layers.15.self_attn.qkv_proj.qweight": "model.safetensors",
182
+ "model.layers.15.self_attn.qkv_proj.scales": "model.safetensors",
183
+ "model.layers.15.self_attn.qkv_proj.zeros": "model.safetensors",
184
+ "model.layers.16.input_layernorm.weight": "model.safetensors",
185
+ "model.layers.16.mlp.down_proj.channel_scale": "model.safetensors",
186
+ "model.layers.16.mlp.down_proj.q_perm": "model.safetensors",
187
+ "model.layers.16.mlp.down_proj.qweight": "model.safetensors",
188
+ "model.layers.16.mlp.down_proj.scales": "model.safetensors",
189
+ "model.layers.16.mlp.down_proj.zeros": "model.safetensors",
190
+ "model.layers.16.mlp.gate_up_proj.channel_scale": "model.safetensors",
191
+ "model.layers.16.mlp.gate_up_proj.q_perm": "model.safetensors",
192
+ "model.layers.16.mlp.gate_up_proj.qweight": "model.safetensors",
193
+ "model.layers.16.mlp.gate_up_proj.scales": "model.safetensors",
194
+ "model.layers.16.mlp.gate_up_proj.zeros": "model.safetensors",
195
+ "model.layers.16.post_attention_layernorm.weight": "model.safetensors",
196
+ "model.layers.16.self_attn.o_proj.channel_scale": "model.safetensors",
197
+ "model.layers.16.self_attn.o_proj.q_perm": "model.safetensors",
198
+ "model.layers.16.self_attn.o_proj.qweight": "model.safetensors",
199
+ "model.layers.16.self_attn.o_proj.scales": "model.safetensors",
200
+ "model.layers.16.self_attn.o_proj.zeros": "model.safetensors",
201
+ "model.layers.16.self_attn.qkv_proj.channel_scale": "model.safetensors",
202
+ "model.layers.16.self_attn.qkv_proj.q_perm": "model.safetensors",
203
+ "model.layers.16.self_attn.qkv_proj.qweight": "model.safetensors",
204
+ "model.layers.16.self_attn.qkv_proj.scales": "model.safetensors",
205
+ "model.layers.16.self_attn.qkv_proj.zeros": "model.safetensors",
206
+ "model.layers.17.input_layernorm.weight": "model.safetensors",
207
+ "model.layers.17.mlp.down_proj.channel_scale": "model.safetensors",
208
+ "model.layers.17.mlp.down_proj.q_perm": "model.safetensors",
209
+ "model.layers.17.mlp.down_proj.qweight": "model.safetensors",
210
+ "model.layers.17.mlp.down_proj.scales": "model.safetensors",
211
+ "model.layers.17.mlp.down_proj.zeros": "model.safetensors",
212
+ "model.layers.17.mlp.gate_up_proj.channel_scale": "model.safetensors",
213
+ "model.layers.17.mlp.gate_up_proj.q_perm": "model.safetensors",
214
+ "model.layers.17.mlp.gate_up_proj.qweight": "model.safetensors",
215
+ "model.layers.17.mlp.gate_up_proj.scales": "model.safetensors",
216
+ "model.layers.17.mlp.gate_up_proj.zeros": "model.safetensors",
217
+ "model.layers.17.post_attention_layernorm.weight": "model.safetensors",
218
+ "model.layers.17.self_attn.o_proj.channel_scale": "model.safetensors",
219
+ "model.layers.17.self_attn.o_proj.q_perm": "model.safetensors",
220
+ "model.layers.17.self_attn.o_proj.qweight": "model.safetensors",
221
+ "model.layers.17.self_attn.o_proj.scales": "model.safetensors",
222
+ "model.layers.17.self_attn.o_proj.zeros": "model.safetensors",
223
+ "model.layers.17.self_attn.qkv_proj.channel_scale": "model.safetensors",
224
+ "model.layers.17.self_attn.qkv_proj.q_perm": "model.safetensors",
225
+ "model.layers.17.self_attn.qkv_proj.qweight": "model.safetensors",
226
+ "model.layers.17.self_attn.qkv_proj.scales": "model.safetensors",
227
+ "model.layers.17.self_attn.qkv_proj.zeros": "model.safetensors",
228
+ "model.layers.18.input_layernorm.weight": "model.safetensors",
229
+ "model.layers.18.mlp.down_proj.channel_scale": "model.safetensors",
230
+ "model.layers.18.mlp.down_proj.q_perm": "model.safetensors",
231
+ "model.layers.18.mlp.down_proj.qweight": "model.safetensors",
232
+ "model.layers.18.mlp.down_proj.scales": "model.safetensors",
233
+ "model.layers.18.mlp.down_proj.zeros": "model.safetensors",
234
+ "model.layers.18.mlp.gate_up_proj.channel_scale": "model.safetensors",
235
+ "model.layers.18.mlp.gate_up_proj.q_perm": "model.safetensors",
236
+ "model.layers.18.mlp.gate_up_proj.qweight": "model.safetensors",
237
+ "model.layers.18.mlp.gate_up_proj.scales": "model.safetensors",
238
+ "model.layers.18.mlp.gate_up_proj.zeros": "model.safetensors",
239
+ "model.layers.18.post_attention_layernorm.weight": "model.safetensors",
240
+ "model.layers.18.self_attn.o_proj.channel_scale": "model.safetensors",
241
+ "model.layers.18.self_attn.o_proj.q_perm": "model.safetensors",
242
+ "model.layers.18.self_attn.o_proj.qweight": "model.safetensors",
243
+ "model.layers.18.self_attn.o_proj.scales": "model.safetensors",
244
+ "model.layers.18.self_attn.o_proj.zeros": "model.safetensors",
245
+ "model.layers.18.self_attn.qkv_proj.channel_scale": "model.safetensors",
246
+ "model.layers.18.self_attn.qkv_proj.q_perm": "model.safetensors",
247
+ "model.layers.18.self_attn.qkv_proj.qweight": "model.safetensors",
248
+ "model.layers.18.self_attn.qkv_proj.scales": "model.safetensors",
249
+ "model.layers.18.self_attn.qkv_proj.zeros": "model.safetensors",
250
+ "model.layers.19.input_layernorm.weight": "model.safetensors",
251
+ "model.layers.19.mlp.down_proj.channel_scale": "model.safetensors",
252
+ "model.layers.19.mlp.down_proj.q_perm": "model.safetensors",
253
+ "model.layers.19.mlp.down_proj.qweight": "model.safetensors",
254
+ "model.layers.19.mlp.down_proj.scales": "model.safetensors",
255
+ "model.layers.19.mlp.down_proj.zeros": "model.safetensors",
256
+ "model.layers.19.mlp.gate_up_proj.channel_scale": "model.safetensors",
257
+ "model.layers.19.mlp.gate_up_proj.q_perm": "model.safetensors",
258
+ "model.layers.19.mlp.gate_up_proj.qweight": "model.safetensors",
259
+ "model.layers.19.mlp.gate_up_proj.scales": "model.safetensors",
260
+ "model.layers.19.mlp.gate_up_proj.zeros": "model.safetensors",
261
+ "model.layers.19.post_attention_layernorm.weight": "model.safetensors",
262
+ "model.layers.19.self_attn.o_proj.channel_scale": "model.safetensors",
263
+ "model.layers.19.self_attn.o_proj.q_perm": "model.safetensors",
264
+ "model.layers.19.self_attn.o_proj.qweight": "model.safetensors",
265
+ "model.layers.19.self_attn.o_proj.scales": "model.safetensors",
266
+ "model.layers.19.self_attn.o_proj.zeros": "model.safetensors",
267
+ "model.layers.19.self_attn.qkv_proj.channel_scale": "model.safetensors",
268
+ "model.layers.19.self_attn.qkv_proj.q_perm": "model.safetensors",
269
+ "model.layers.19.self_attn.qkv_proj.qweight": "model.safetensors",
270
+ "model.layers.19.self_attn.qkv_proj.scales": "model.safetensors",
271
+ "model.layers.19.self_attn.qkv_proj.zeros": "model.safetensors",
272
+ "model.layers.2.input_layernorm.weight": "model.safetensors",
273
+ "model.layers.2.mlp.down_proj.channel_scale": "model.safetensors",
274
+ "model.layers.2.mlp.down_proj.q_perm": "model.safetensors",
275
+ "model.layers.2.mlp.down_proj.qweight": "model.safetensors",
276
+ "model.layers.2.mlp.down_proj.scales": "model.safetensors",
277
+ "model.layers.2.mlp.down_proj.zeros": "model.safetensors",
278
+ "model.layers.2.mlp.gate_up_proj.channel_scale": "model.safetensors",
279
+ "model.layers.2.mlp.gate_up_proj.q_perm": "model.safetensors",
280
+ "model.layers.2.mlp.gate_up_proj.qweight": "model.safetensors",
281
+ "model.layers.2.mlp.gate_up_proj.scales": "model.safetensors",
282
+ "model.layers.2.mlp.gate_up_proj.zeros": "model.safetensors",
283
+ "model.layers.2.post_attention_layernorm.weight": "model.safetensors",
284
+ "model.layers.2.self_attn.o_proj.channel_scale": "model.safetensors",
285
+ "model.layers.2.self_attn.o_proj.q_perm": "model.safetensors",
286
+ "model.layers.2.self_attn.o_proj.qweight": "model.safetensors",
287
+ "model.layers.2.self_attn.o_proj.scales": "model.safetensors",
288
+ "model.layers.2.self_attn.o_proj.zeros": "model.safetensors",
289
+ "model.layers.2.self_attn.qkv_proj.channel_scale": "model.safetensors",
290
+ "model.layers.2.self_attn.qkv_proj.q_perm": "model.safetensors",
291
+ "model.layers.2.self_attn.qkv_proj.qweight": "model.safetensors",
292
+ "model.layers.2.self_attn.qkv_proj.scales": "model.safetensors",
293
+ "model.layers.2.self_attn.qkv_proj.zeros": "model.safetensors",
294
+ "model.layers.20.input_layernorm.weight": "model.safetensors",
295
+ "model.layers.20.mlp.down_proj.channel_scale": "model.safetensors",
296
+ "model.layers.20.mlp.down_proj.q_perm": "model.safetensors",
297
+ "model.layers.20.mlp.down_proj.qweight": "model.safetensors",
298
+ "model.layers.20.mlp.down_proj.scales": "model.safetensors",
299
+ "model.layers.20.mlp.down_proj.zeros": "model.safetensors",
300
+ "model.layers.20.mlp.gate_up_proj.channel_scale": "model.safetensors",
301
+ "model.layers.20.mlp.gate_up_proj.q_perm": "model.safetensors",
302
+ "model.layers.20.mlp.gate_up_proj.qweight": "model.safetensors",
303
+ "model.layers.20.mlp.gate_up_proj.scales": "model.safetensors",
304
+ "model.layers.20.mlp.gate_up_proj.zeros": "model.safetensors",
305
+ "model.layers.20.post_attention_layernorm.weight": "model.safetensors",
306
+ "model.layers.20.self_attn.o_proj.channel_scale": "model.safetensors",
307
+ "model.layers.20.self_attn.o_proj.q_perm": "model.safetensors",
308
+ "model.layers.20.self_attn.o_proj.qweight": "model.safetensors",
309
+ "model.layers.20.self_attn.o_proj.scales": "model.safetensors",
310
+ "model.layers.20.self_attn.o_proj.zeros": "model.safetensors",
311
+ "model.layers.20.self_attn.qkv_proj.channel_scale": "model.safetensors",
312
+ "model.layers.20.self_attn.qkv_proj.q_perm": "model.safetensors",
313
+ "model.layers.20.self_attn.qkv_proj.qweight": "model.safetensors",
314
+ "model.layers.20.self_attn.qkv_proj.scales": "model.safetensors",
315
+ "model.layers.20.self_attn.qkv_proj.zeros": "model.safetensors",
316
+ "model.layers.21.input_layernorm.weight": "model.safetensors",
317
+ "model.layers.21.mlp.down_proj.channel_scale": "model.safetensors",
318
+ "model.layers.21.mlp.down_proj.q_perm": "model.safetensors",
319
+ "model.layers.21.mlp.down_proj.qweight": "model.safetensors",
320
+ "model.layers.21.mlp.down_proj.scales": "model.safetensors",
321
+ "model.layers.21.mlp.down_proj.zeros": "model.safetensors",
322
+ "model.layers.21.mlp.gate_up_proj.channel_scale": "model.safetensors",
323
+ "model.layers.21.mlp.gate_up_proj.q_perm": "model.safetensors",
324
+ "model.layers.21.mlp.gate_up_proj.qweight": "model.safetensors",
325
+ "model.layers.21.mlp.gate_up_proj.scales": "model.safetensors",
326
+ "model.layers.21.mlp.gate_up_proj.zeros": "model.safetensors",
327
+ "model.layers.21.post_attention_layernorm.weight": "model.safetensors",
328
+ "model.layers.21.self_attn.o_proj.channel_scale": "model.safetensors",
329
+ "model.layers.21.self_attn.o_proj.q_perm": "model.safetensors",
330
+ "model.layers.21.self_attn.o_proj.qweight": "model.safetensors",
331
+ "model.layers.21.self_attn.o_proj.scales": "model.safetensors",
332
+ "model.layers.21.self_attn.o_proj.zeros": "model.safetensors",
333
+ "model.layers.21.self_attn.qkv_proj.channel_scale": "model.safetensors",
334
+ "model.layers.21.self_attn.qkv_proj.q_perm": "model.safetensors",
335
+ "model.layers.21.self_attn.qkv_proj.qweight": "model.safetensors",
336
+ "model.layers.21.self_attn.qkv_proj.scales": "model.safetensors",
337
+ "model.layers.21.self_attn.qkv_proj.zeros": "model.safetensors",
338
+ "model.layers.22.input_layernorm.weight": "model.safetensors",
339
+ "model.layers.22.mlp.down_proj.channel_scale": "model.safetensors",
340
+ "model.layers.22.mlp.down_proj.q_perm": "model.safetensors",
341
+ "model.layers.22.mlp.down_proj.qweight": "model.safetensors",
342
+ "model.layers.22.mlp.down_proj.scales": "model.safetensors",
343
+ "model.layers.22.mlp.down_proj.zeros": "model.safetensors",
344
+ "model.layers.22.mlp.gate_up_proj.channel_scale": "model.safetensors",
345
+ "model.layers.22.mlp.gate_up_proj.q_perm": "model.safetensors",
346
+ "model.layers.22.mlp.gate_up_proj.qweight": "model.safetensors",
347
+ "model.layers.22.mlp.gate_up_proj.scales": "model.safetensors",
348
+ "model.layers.22.mlp.gate_up_proj.zeros": "model.safetensors",
349
+ "model.layers.22.post_attention_layernorm.weight": "model.safetensors",
350
+ "model.layers.22.self_attn.o_proj.channel_scale": "model.safetensors",
351
+ "model.layers.22.self_attn.o_proj.q_perm": "model.safetensors",
352
+ "model.layers.22.self_attn.o_proj.qweight": "model.safetensors",
353
+ "model.layers.22.self_attn.o_proj.scales": "model.safetensors",
354
+ "model.layers.22.self_attn.o_proj.zeros": "model.safetensors",
355
+ "model.layers.22.self_attn.qkv_proj.channel_scale": "model.safetensors",
356
+ "model.layers.22.self_attn.qkv_proj.q_perm": "model.safetensors",
357
+ "model.layers.22.self_attn.qkv_proj.qweight": "model.safetensors",
358
+ "model.layers.22.self_attn.qkv_proj.scales": "model.safetensors",
359
+ "model.layers.22.self_attn.qkv_proj.zeros": "model.safetensors",
360
+ "model.layers.23.input_layernorm.weight": "model.safetensors",
361
+ "model.layers.23.mlp.down_proj.channel_scale": "model.safetensors",
362
+ "model.layers.23.mlp.down_proj.q_perm": "model.safetensors",
363
+ "model.layers.23.mlp.down_proj.qweight": "model.safetensors",
364
+ "model.layers.23.mlp.down_proj.scales": "model.safetensors",
365
+ "model.layers.23.mlp.down_proj.zeros": "model.safetensors",
366
+ "model.layers.23.mlp.gate_up_proj.channel_scale": "model.safetensors",
367
+ "model.layers.23.mlp.gate_up_proj.q_perm": "model.safetensors",
368
+ "model.layers.23.mlp.gate_up_proj.qweight": "model.safetensors",
369
+ "model.layers.23.mlp.gate_up_proj.scales": "model.safetensors",
370
+ "model.layers.23.mlp.gate_up_proj.zeros": "model.safetensors",
371
+ "model.layers.23.post_attention_layernorm.weight": "model.safetensors",
372
+ "model.layers.23.self_attn.o_proj.channel_scale": "model.safetensors",
373
+ "model.layers.23.self_attn.o_proj.q_perm": "model.safetensors",
374
+ "model.layers.23.self_attn.o_proj.qweight": "model.safetensors",
375
+ "model.layers.23.self_attn.o_proj.scales": "model.safetensors",
376
+ "model.layers.23.self_attn.o_proj.zeros": "model.safetensors",
377
+ "model.layers.23.self_attn.qkv_proj.channel_scale": "model.safetensors",
378
+ "model.layers.23.self_attn.qkv_proj.q_perm": "model.safetensors",
379
+ "model.layers.23.self_attn.qkv_proj.qweight": "model.safetensors",
380
+ "model.layers.23.self_attn.qkv_proj.scales": "model.safetensors",
381
+ "model.layers.23.self_attn.qkv_proj.zeros": "model.safetensors",
382
+ "model.layers.24.input_layernorm.weight": "model.safetensors",
383
+ "model.layers.24.mlp.down_proj.channel_scale": "model.safetensors",
384
+ "model.layers.24.mlp.down_proj.q_perm": "model.safetensors",
385
+ "model.layers.24.mlp.down_proj.qweight": "model.safetensors",
386
+ "model.layers.24.mlp.down_proj.scales": "model.safetensors",
387
+ "model.layers.24.mlp.down_proj.zeros": "model.safetensors",
388
+ "model.layers.24.mlp.gate_up_proj.channel_scale": "model.safetensors",
389
+ "model.layers.24.mlp.gate_up_proj.q_perm": "model.safetensors",
390
+ "model.layers.24.mlp.gate_up_proj.qweight": "model.safetensors",
391
+ "model.layers.24.mlp.gate_up_proj.scales": "model.safetensors",
392
+ "model.layers.24.mlp.gate_up_proj.zeros": "model.safetensors",
393
+ "model.layers.24.post_attention_layernorm.weight": "model.safetensors",
394
+ "model.layers.24.self_attn.o_proj.channel_scale": "model.safetensors",
395
+ "model.layers.24.self_attn.o_proj.q_perm": "model.safetensors",
396
+ "model.layers.24.self_attn.o_proj.qweight": "model.safetensors",
397
+ "model.layers.24.self_attn.o_proj.scales": "model.safetensors",
398
+ "model.layers.24.self_attn.o_proj.zeros": "model.safetensors",
399
+ "model.layers.24.self_attn.qkv_proj.channel_scale": "model.safetensors",
400
+ "model.layers.24.self_attn.qkv_proj.q_perm": "model.safetensors",
401
+ "model.layers.24.self_attn.qkv_proj.qweight": "model.safetensors",
402
+ "model.layers.24.self_attn.qkv_proj.scales": "model.safetensors",
403
+ "model.layers.24.self_attn.qkv_proj.zeros": "model.safetensors",
404
+ "model.layers.25.input_layernorm.weight": "model.safetensors",
405
+ "model.layers.25.mlp.down_proj.channel_scale": "model.safetensors",
406
+ "model.layers.25.mlp.down_proj.q_perm": "model.safetensors",
407
+ "model.layers.25.mlp.down_proj.qweight": "model.safetensors",
408
+ "model.layers.25.mlp.down_proj.scales": "model.safetensors",
409
+ "model.layers.25.mlp.down_proj.zeros": "model.safetensors",
410
+ "model.layers.25.mlp.gate_up_proj.channel_scale": "model.safetensors",
411
+ "model.layers.25.mlp.gate_up_proj.q_perm": "model.safetensors",
412
+ "model.layers.25.mlp.gate_up_proj.qweight": "model.safetensors",
413
+ "model.layers.25.mlp.gate_up_proj.scales": "model.safetensors",
414
+ "model.layers.25.mlp.gate_up_proj.zeros": "model.safetensors",
415
+ "model.layers.25.post_attention_layernorm.weight": "model.safetensors",
416
+ "model.layers.25.self_attn.o_proj.channel_scale": "model.safetensors",
417
+ "model.layers.25.self_attn.o_proj.q_perm": "model.safetensors",
418
+ "model.layers.25.self_attn.o_proj.qweight": "model.safetensors",
419
+ "model.layers.25.self_attn.o_proj.scales": "model.safetensors",
420
+ "model.layers.25.self_attn.o_proj.zeros": "model.safetensors",
421
+ "model.layers.25.self_attn.qkv_proj.channel_scale": "model.safetensors",
422
+ "model.layers.25.self_attn.qkv_proj.q_perm": "model.safetensors",
423
+ "model.layers.25.self_attn.qkv_proj.qweight": "model.safetensors",
424
+ "model.layers.25.self_attn.qkv_proj.scales": "model.safetensors",
425
+ "model.layers.25.self_attn.qkv_proj.zeros": "model.safetensors",
426
+ "model.layers.26.input_layernorm.weight": "model.safetensors",
427
+ "model.layers.26.mlp.down_proj.channel_scale": "model.safetensors",
428
+ "model.layers.26.mlp.down_proj.q_perm": "model.safetensors",
429
+ "model.layers.26.mlp.down_proj.qweight": "model.safetensors",
430
+ "model.layers.26.mlp.down_proj.scales": "model.safetensors",
431
+ "model.layers.26.mlp.down_proj.zeros": "model.safetensors",
432
+ "model.layers.26.mlp.gate_up_proj.channel_scale": "model.safetensors",
433
+ "model.layers.26.mlp.gate_up_proj.q_perm": "model.safetensors",
434
+ "model.layers.26.mlp.gate_up_proj.qweight": "model.safetensors",
435
+ "model.layers.26.mlp.gate_up_proj.scales": "model.safetensors",
436
+ "model.layers.26.mlp.gate_up_proj.zeros": "model.safetensors",
437
+ "model.layers.26.post_attention_layernorm.weight": "model.safetensors",
438
+ "model.layers.26.self_attn.o_proj.channel_scale": "model.safetensors",
439
+ "model.layers.26.self_attn.o_proj.q_perm": "model.safetensors",
440
+ "model.layers.26.self_attn.o_proj.qweight": "model.safetensors",
441
+ "model.layers.26.self_attn.o_proj.scales": "model.safetensors",
442
+ "model.layers.26.self_attn.o_proj.zeros": "model.safetensors",
443
+ "model.layers.26.self_attn.qkv_proj.channel_scale": "model.safetensors",
444
+ "model.layers.26.self_attn.qkv_proj.q_perm": "model.safetensors",
445
+ "model.layers.26.self_attn.qkv_proj.qweight": "model.safetensors",
446
+ "model.layers.26.self_attn.qkv_proj.scales": "model.safetensors",
447
+ "model.layers.26.self_attn.qkv_proj.zeros": "model.safetensors",
448
+ "model.layers.27.input_layernorm.weight": "model.safetensors",
449
+ "model.layers.27.mlp.down_proj.channel_scale": "model.safetensors",
450
+ "model.layers.27.mlp.down_proj.q_perm": "model.safetensors",
451
+ "model.layers.27.mlp.down_proj.qweight": "model.safetensors",
452
+ "model.layers.27.mlp.down_proj.scales": "model.safetensors",
453
+ "model.layers.27.mlp.down_proj.zeros": "model.safetensors",
454
+ "model.layers.27.mlp.gate_up_proj.channel_scale": "model.safetensors",
455
+ "model.layers.27.mlp.gate_up_proj.q_perm": "model.safetensors",
456
+ "model.layers.27.mlp.gate_up_proj.qweight": "model.safetensors",
457
+ "model.layers.27.mlp.gate_up_proj.scales": "model.safetensors",
458
+ "model.layers.27.mlp.gate_up_proj.zeros": "model.safetensors",
459
+ "model.layers.27.post_attention_layernorm.weight": "model.safetensors",
460
+ "model.layers.27.self_attn.o_proj.channel_scale": "model.safetensors",
461
+ "model.layers.27.self_attn.o_proj.q_perm": "model.safetensors",
462
+ "model.layers.27.self_attn.o_proj.qweight": "model.safetensors",
463
+ "model.layers.27.self_attn.o_proj.scales": "model.safetensors",
464
+ "model.layers.27.self_attn.o_proj.zeros": "model.safetensors",
465
+ "model.layers.27.self_attn.qkv_proj.channel_scale": "model.safetensors",
466
+ "model.layers.27.self_attn.qkv_proj.q_perm": "model.safetensors",
467
+ "model.layers.27.self_attn.qkv_proj.qweight": "model.safetensors",
468
+ "model.layers.27.self_attn.qkv_proj.scales": "model.safetensors",
469
+ "model.layers.27.self_attn.qkv_proj.zeros": "model.safetensors",
470
+ "model.layers.28.input_layernorm.weight": "model.safetensors",
471
+ "model.layers.28.mlp.down_proj.channel_scale": "model.safetensors",
472
+ "model.layers.28.mlp.down_proj.q_perm": "model.safetensors",
473
+ "model.layers.28.mlp.down_proj.qweight": "model.safetensors",
474
+ "model.layers.28.mlp.down_proj.scales": "model.safetensors",
475
+ "model.layers.28.mlp.down_proj.zeros": "model.safetensors",
476
+ "model.layers.28.mlp.gate_up_proj.channel_scale": "model.safetensors",
477
+ "model.layers.28.mlp.gate_up_proj.q_perm": "model.safetensors",
478
+ "model.layers.28.mlp.gate_up_proj.qweight": "model.safetensors",
479
+ "model.layers.28.mlp.gate_up_proj.scales": "model.safetensors",
480
+ "model.layers.28.mlp.gate_up_proj.zeros": "model.safetensors",
481
+ "model.layers.28.post_attention_layernorm.weight": "model.safetensors",
482
+ "model.layers.28.self_attn.o_proj.channel_scale": "model.safetensors",
483
+ "model.layers.28.self_attn.o_proj.q_perm": "model.safetensors",
484
+ "model.layers.28.self_attn.o_proj.qweight": "model.safetensors",
485
+ "model.layers.28.self_attn.o_proj.scales": "model.safetensors",
486
+ "model.layers.28.self_attn.o_proj.zeros": "model.safetensors",
487
+ "model.layers.28.self_attn.qkv_proj.channel_scale": "model.safetensors",
488
+ "model.layers.28.self_attn.qkv_proj.q_perm": "model.safetensors",
489
+ "model.layers.28.self_attn.qkv_proj.qweight": "model.safetensors",
490
+ "model.layers.28.self_attn.qkv_proj.scales": "model.safetensors",
491
+ "model.layers.28.self_attn.qkv_proj.zeros": "model.safetensors",
492
+ "model.layers.29.input_layernorm.weight": "model.safetensors",
493
+ "model.layers.29.mlp.down_proj.channel_scale": "model.safetensors",
494
+ "model.layers.29.mlp.down_proj.q_perm": "model.safetensors",
495
+ "model.layers.29.mlp.down_proj.qweight": "model.safetensors",
496
+ "model.layers.29.mlp.down_proj.scales": "model.safetensors",
497
+ "model.layers.29.mlp.down_proj.zeros": "model.safetensors",
498
+ "model.layers.29.mlp.gate_up_proj.channel_scale": "model.safetensors",
499
+ "model.layers.29.mlp.gate_up_proj.q_perm": "model.safetensors",
500
+ "model.layers.29.mlp.gate_up_proj.qweight": "model.safetensors",
501
+ "model.layers.29.mlp.gate_up_proj.scales": "model.safetensors",
502
+ "model.layers.29.mlp.gate_up_proj.zeros": "model.safetensors",
503
+ "model.layers.29.post_attention_layernorm.weight": "model.safetensors",
504
+ "model.layers.29.self_attn.o_proj.channel_scale": "model.safetensors",
505
+ "model.layers.29.self_attn.o_proj.q_perm": "model.safetensors",
506
+ "model.layers.29.self_attn.o_proj.qweight": "model.safetensors",
507
+ "model.layers.29.self_attn.o_proj.scales": "model.safetensors",
508
+ "model.layers.29.self_attn.o_proj.zeros": "model.safetensors",
509
+ "model.layers.29.self_attn.qkv_proj.channel_scale": "model.safetensors",
510
+ "model.layers.29.self_attn.qkv_proj.q_perm": "model.safetensors",
511
+ "model.layers.29.self_attn.qkv_proj.qweight": "model.safetensors",
512
+ "model.layers.29.self_attn.qkv_proj.scales": "model.safetensors",
513
+ "model.layers.29.self_attn.qkv_proj.zeros": "model.safetensors",
514
+ "model.layers.3.input_layernorm.weight": "model.safetensors",
515
+ "model.layers.3.mlp.down_proj.channel_scale": "model.safetensors",
516
+ "model.layers.3.mlp.down_proj.q_perm": "model.safetensors",
517
+ "model.layers.3.mlp.down_proj.qweight": "model.safetensors",
518
+ "model.layers.3.mlp.down_proj.scales": "model.safetensors",
519
+ "model.layers.3.mlp.down_proj.zeros": "model.safetensors",
520
+ "model.layers.3.mlp.gate_up_proj.channel_scale": "model.safetensors",
521
+ "model.layers.3.mlp.gate_up_proj.q_perm": "model.safetensors",
522
+ "model.layers.3.mlp.gate_up_proj.qweight": "model.safetensors",
523
+ "model.layers.3.mlp.gate_up_proj.scales": "model.safetensors",
524
+ "model.layers.3.mlp.gate_up_proj.zeros": "model.safetensors",
525
+ "model.layers.3.post_attention_layernorm.weight": "model.safetensors",
526
+ "model.layers.3.self_attn.o_proj.channel_scale": "model.safetensors",
527
+ "model.layers.3.self_attn.o_proj.q_perm": "model.safetensors",
528
+ "model.layers.3.self_attn.o_proj.qweight": "model.safetensors",
529
+ "model.layers.3.self_attn.o_proj.scales": "model.safetensors",
530
+ "model.layers.3.self_attn.o_proj.zeros": "model.safetensors",
531
+ "model.layers.3.self_attn.qkv_proj.channel_scale": "model.safetensors",
532
+ "model.layers.3.self_attn.qkv_proj.q_perm": "model.safetensors",
533
+ "model.layers.3.self_attn.qkv_proj.qweight": "model.safetensors",
534
+ "model.layers.3.self_attn.qkv_proj.scales": "model.safetensors",
535
+ "model.layers.3.self_attn.qkv_proj.zeros": "model.safetensors",
536
+ "model.layers.30.input_layernorm.weight": "model.safetensors",
537
+ "model.layers.30.mlp.down_proj.channel_scale": "model.safetensors",
538
+ "model.layers.30.mlp.down_proj.q_perm": "model.safetensors",
539
+ "model.layers.30.mlp.down_proj.qweight": "model.safetensors",
540
+ "model.layers.30.mlp.down_proj.scales": "model.safetensors",
541
+ "model.layers.30.mlp.down_proj.zeros": "model.safetensors",
542
+ "model.layers.30.mlp.gate_up_proj.channel_scale": "model.safetensors",
543
+ "model.layers.30.mlp.gate_up_proj.q_perm": "model.safetensors",
544
+ "model.layers.30.mlp.gate_up_proj.qweight": "model.safetensors",
545
+ "model.layers.30.mlp.gate_up_proj.scales": "model.safetensors",
546
+ "model.layers.30.mlp.gate_up_proj.zeros": "model.safetensors",
547
+ "model.layers.30.post_attention_layernorm.weight": "model.safetensors",
548
+ "model.layers.30.self_attn.o_proj.channel_scale": "model.safetensors",
549
+ "model.layers.30.self_attn.o_proj.q_perm": "model.safetensors",
550
+ "model.layers.30.self_attn.o_proj.qweight": "model.safetensors",
551
+ "model.layers.30.self_attn.o_proj.scales": "model.safetensors",
552
+ "model.layers.30.self_attn.o_proj.zeros": "model.safetensors",
553
+ "model.layers.30.self_attn.qkv_proj.channel_scale": "model.safetensors",
554
+ "model.layers.30.self_attn.qkv_proj.q_perm": "model.safetensors",
555
+ "model.layers.30.self_attn.qkv_proj.qweight": "model.safetensors",
556
+ "model.layers.30.self_attn.qkv_proj.scales": "model.safetensors",
557
+ "model.layers.30.self_attn.qkv_proj.zeros": "model.safetensors",
558
+ "model.layers.31.input_layernorm.weight": "model.safetensors",
559
+ "model.layers.31.mlp.down_proj.channel_scale": "model.safetensors",
560
+ "model.layers.31.mlp.down_proj.q_perm": "model.safetensors",
561
+ "model.layers.31.mlp.down_proj.qweight": "model.safetensors",
562
+ "model.layers.31.mlp.down_proj.scales": "model.safetensors",
563
+ "model.layers.31.mlp.down_proj.zeros": "model.safetensors",
564
+ "model.layers.31.mlp.gate_up_proj.channel_scale": "model.safetensors",
565
+ "model.layers.31.mlp.gate_up_proj.q_perm": "model.safetensors",
566
+ "model.layers.31.mlp.gate_up_proj.qweight": "model.safetensors",
567
+ "model.layers.31.mlp.gate_up_proj.scales": "model.safetensors",
568
+ "model.layers.31.mlp.gate_up_proj.zeros": "model.safetensors",
569
+ "model.layers.31.post_attention_layernorm.weight": "model.safetensors",
570
+ "model.layers.31.self_attn.o_proj.channel_scale": "model.safetensors",
571
+ "model.layers.31.self_attn.o_proj.q_perm": "model.safetensors",
572
+ "model.layers.31.self_attn.o_proj.qweight": "model.safetensors",
573
+ "model.layers.31.self_attn.o_proj.scales": "model.safetensors",
574
+ "model.layers.31.self_attn.o_proj.zeros": "model.safetensors",
575
+ "model.layers.31.self_attn.qkv_proj.channel_scale": "model.safetensors",
576
+ "model.layers.31.self_attn.qkv_proj.q_perm": "model.safetensors",
577
+ "model.layers.31.self_attn.qkv_proj.qweight": "model.safetensors",
578
+ "model.layers.31.self_attn.qkv_proj.scales": "model.safetensors",
579
+ "model.layers.31.self_attn.qkv_proj.zeros": "model.safetensors",
580
+ "model.layers.4.input_layernorm.weight": "model.safetensors",
581
+ "model.layers.4.mlp.down_proj.channel_scale": "model.safetensors",
582
+ "model.layers.4.mlp.down_proj.q_perm": "model.safetensors",
583
+ "model.layers.4.mlp.down_proj.qweight": "model.safetensors",
584
+ "model.layers.4.mlp.down_proj.scales": "model.safetensors",
585
+ "model.layers.4.mlp.down_proj.zeros": "model.safetensors",
586
+ "model.layers.4.mlp.gate_up_proj.channel_scale": "model.safetensors",
587
+ "model.layers.4.mlp.gate_up_proj.q_perm": "model.safetensors",
588
+ "model.layers.4.mlp.gate_up_proj.qweight": "model.safetensors",
589
+ "model.layers.4.mlp.gate_up_proj.scales": "model.safetensors",
590
+ "model.layers.4.mlp.gate_up_proj.zeros": "model.safetensors",
591
+ "model.layers.4.post_attention_layernorm.weight": "model.safetensors",
592
+ "model.layers.4.self_attn.o_proj.channel_scale": "model.safetensors",
593
+ "model.layers.4.self_attn.o_proj.q_perm": "model.safetensors",
594
+ "model.layers.4.self_attn.o_proj.qweight": "model.safetensors",
595
+ "model.layers.4.self_attn.o_proj.scales": "model.safetensors",
596
+ "model.layers.4.self_attn.o_proj.zeros": "model.safetensors",
597
+ "model.layers.4.self_attn.qkv_proj.channel_scale": "model.safetensors",
598
+ "model.layers.4.self_attn.qkv_proj.q_perm": "model.safetensors",
599
+ "model.layers.4.self_attn.qkv_proj.qweight": "model.safetensors",
600
+ "model.layers.4.self_attn.qkv_proj.scales": "model.safetensors",
601
+ "model.layers.4.self_attn.qkv_proj.zeros": "model.safetensors",
602
+ "model.layers.5.input_layernorm.weight": "model.safetensors",
603
+ "model.layers.5.mlp.down_proj.channel_scale": "model.safetensors",
604
+ "model.layers.5.mlp.down_proj.q_perm": "model.safetensors",
605
+ "model.layers.5.mlp.down_proj.qweight": "model.safetensors",
606
+ "model.layers.5.mlp.down_proj.scales": "model.safetensors",
607
+ "model.layers.5.mlp.down_proj.zeros": "model.safetensors",
608
+ "model.layers.5.mlp.gate_up_proj.channel_scale": "model.safetensors",
609
+ "model.layers.5.mlp.gate_up_proj.q_perm": "model.safetensors",
610
+ "model.layers.5.mlp.gate_up_proj.qweight": "model.safetensors",
611
+ "model.layers.5.mlp.gate_up_proj.scales": "model.safetensors",
612
+ "model.layers.5.mlp.gate_up_proj.zeros": "model.safetensors",
613
+ "model.layers.5.post_attention_layernorm.weight": "model.safetensors",
614
+ "model.layers.5.self_attn.o_proj.channel_scale": "model.safetensors",
615
+ "model.layers.5.self_attn.o_proj.q_perm": "model.safetensors",
616
+ "model.layers.5.self_attn.o_proj.qweight": "model.safetensors",
617
+ "model.layers.5.self_attn.o_proj.scales": "model.safetensors",
618
+ "model.layers.5.self_attn.o_proj.zeros": "model.safetensors",
619
+ "model.layers.5.self_attn.qkv_proj.channel_scale": "model.safetensors",
620
+ "model.layers.5.self_attn.qkv_proj.q_perm": "model.safetensors",
621
+ "model.layers.5.self_attn.qkv_proj.qweight": "model.safetensors",
622
+ "model.layers.5.self_attn.qkv_proj.scales": "model.safetensors",
623
+ "model.layers.5.self_attn.qkv_proj.zeros": "model.safetensors",
624
+ "model.layers.6.input_layernorm.weight": "model.safetensors",
625
+ "model.layers.6.mlp.down_proj.channel_scale": "model.safetensors",
626
+ "model.layers.6.mlp.down_proj.q_perm": "model.safetensors",
627
+ "model.layers.6.mlp.down_proj.qweight": "model.safetensors",
628
+ "model.layers.6.mlp.down_proj.scales": "model.safetensors",
629
+ "model.layers.6.mlp.down_proj.zeros": "model.safetensors",
630
+ "model.layers.6.mlp.gate_up_proj.channel_scale": "model.safetensors",
631
+ "model.layers.6.mlp.gate_up_proj.q_perm": "model.safetensors",
632
+ "model.layers.6.mlp.gate_up_proj.qweight": "model.safetensors",
633
+ "model.layers.6.mlp.gate_up_proj.scales": "model.safetensors",
634
+ "model.layers.6.mlp.gate_up_proj.zeros": "model.safetensors",
635
+ "model.layers.6.post_attention_layernorm.weight": "model.safetensors",
636
+ "model.layers.6.self_attn.o_proj.channel_scale": "model.safetensors",
637
+ "model.layers.6.self_attn.o_proj.q_perm": "model.safetensors",
638
+ "model.layers.6.self_attn.o_proj.qweight": "model.safetensors",
639
+ "model.layers.6.self_attn.o_proj.scales": "model.safetensors",
640
+ "model.layers.6.self_attn.o_proj.zeros": "model.safetensors",
641
+ "model.layers.6.self_attn.qkv_proj.channel_scale": "model.safetensors",
642
+ "model.layers.6.self_attn.qkv_proj.q_perm": "model.safetensors",
643
+ "model.layers.6.self_attn.qkv_proj.qweight": "model.safetensors",
644
+ "model.layers.6.self_attn.qkv_proj.scales": "model.safetensors",
645
+ "model.layers.6.self_attn.qkv_proj.zeros": "model.safetensors",
646
+ "model.layers.7.input_layernorm.weight": "model.safetensors",
647
+ "model.layers.7.mlp.down_proj.channel_scale": "model.safetensors",
648
+ "model.layers.7.mlp.down_proj.q_perm": "model.safetensors",
649
+ "model.layers.7.mlp.down_proj.qweight": "model.safetensors",
650
+ "model.layers.7.mlp.down_proj.scales": "model.safetensors",
651
+ "model.layers.7.mlp.down_proj.zeros": "model.safetensors",
652
+ "model.layers.7.mlp.gate_up_proj.channel_scale": "model.safetensors",
653
+ "model.layers.7.mlp.gate_up_proj.q_perm": "model.safetensors",
654
+ "model.layers.7.mlp.gate_up_proj.qweight": "model.safetensors",
655
+ "model.layers.7.mlp.gate_up_proj.scales": "model.safetensors",
656
+ "model.layers.7.mlp.gate_up_proj.zeros": "model.safetensors",
657
+ "model.layers.7.post_attention_layernorm.weight": "model.safetensors",
658
+ "model.layers.7.self_attn.o_proj.channel_scale": "model.safetensors",
659
+ "model.layers.7.self_attn.o_proj.q_perm": "model.safetensors",
660
+ "model.layers.7.self_attn.o_proj.qweight": "model.safetensors",
661
+ "model.layers.7.self_attn.o_proj.scales": "model.safetensors",
662
+ "model.layers.7.self_attn.o_proj.zeros": "model.safetensors",
663
+ "model.layers.7.self_attn.qkv_proj.channel_scale": "model.safetensors",
664
+ "model.layers.7.self_attn.qkv_proj.q_perm": "model.safetensors",
665
+ "model.layers.7.self_attn.qkv_proj.qweight": "model.safetensors",
666
+ "model.layers.7.self_attn.qkv_proj.scales": "model.safetensors",
667
+ "model.layers.7.self_attn.qkv_proj.zeros": "model.safetensors",
668
+ "model.layers.8.input_layernorm.weight": "model.safetensors",
669
+ "model.layers.8.mlp.down_proj.channel_scale": "model.safetensors",
670
+ "model.layers.8.mlp.down_proj.q_perm": "model.safetensors",
671
+ "model.layers.8.mlp.down_proj.qweight": "model.safetensors",
672
+ "model.layers.8.mlp.down_proj.scales": "model.safetensors",
673
+ "model.layers.8.mlp.down_proj.zeros": "model.safetensors",
674
+ "model.layers.8.mlp.gate_up_proj.channel_scale": "model.safetensors",
675
+ "model.layers.8.mlp.gate_up_proj.q_perm": "model.safetensors",
676
+ "model.layers.8.mlp.gate_up_proj.qweight": "model.safetensors",
677
+ "model.layers.8.mlp.gate_up_proj.scales": "model.safetensors",
678
+ "model.layers.8.mlp.gate_up_proj.zeros": "model.safetensors",
679
+ "model.layers.8.post_attention_layernorm.weight": "model.safetensors",
680
+ "model.layers.8.self_attn.o_proj.channel_scale": "model.safetensors",
681
+ "model.layers.8.self_attn.o_proj.q_perm": "model.safetensors",
682
+ "model.layers.8.self_attn.o_proj.qweight": "model.safetensors",
683
+ "model.layers.8.self_attn.o_proj.scales": "model.safetensors",
684
+ "model.layers.8.self_attn.o_proj.zeros": "model.safetensors",
685
+ "model.layers.8.self_attn.qkv_proj.channel_scale": "model.safetensors",
686
+ "model.layers.8.self_attn.qkv_proj.q_perm": "model.safetensors",
687
+ "model.layers.8.self_attn.qkv_proj.qweight": "model.safetensors",
688
+ "model.layers.8.self_attn.qkv_proj.scales": "model.safetensors",
689
+ "model.layers.8.self_attn.qkv_proj.zeros": "model.safetensors",
690
+ "model.layers.9.input_layernorm.weight": "model.safetensors",
691
+ "model.layers.9.mlp.down_proj.channel_scale": "model.safetensors",
692
+ "model.layers.9.mlp.down_proj.q_perm": "model.safetensors",
693
+ "model.layers.9.mlp.down_proj.qweight": "model.safetensors",
694
+ "model.layers.9.mlp.down_proj.scales": "model.safetensors",
695
+ "model.layers.9.mlp.down_proj.zeros": "model.safetensors",
696
+ "model.layers.9.mlp.gate_up_proj.channel_scale": "model.safetensors",
697
+ "model.layers.9.mlp.gate_up_proj.q_perm": "model.safetensors",
698
+ "model.layers.9.mlp.gate_up_proj.qweight": "model.safetensors",
699
+ "model.layers.9.mlp.gate_up_proj.scales": "model.safetensors",
700
+ "model.layers.9.mlp.gate_up_proj.zeros": "model.safetensors",
701
+ "model.layers.9.post_attention_layernorm.weight": "model.safetensors",
702
+ "model.layers.9.self_attn.o_proj.channel_scale": "model.safetensors",
703
+ "model.layers.9.self_attn.o_proj.q_perm": "model.safetensors",
704
+ "model.layers.9.self_attn.o_proj.qweight": "model.safetensors",
705
+ "model.layers.9.self_attn.o_proj.scales": "model.safetensors",
706
+ "model.layers.9.self_attn.o_proj.zeros": "model.safetensors",
707
+ "model.layers.9.self_attn.qkv_proj.channel_scale": "model.safetensors",
708
+ "model.layers.9.self_attn.qkv_proj.q_perm": "model.safetensors",
709
+ "model.layers.9.self_attn.qkv_proj.qweight": "model.safetensors",
710
+ "model.layers.9.self_attn.qkv_proj.scales": "model.safetensors",
711
+ "model.layers.9.self_attn.qkv_proj.zeros": "model.safetensors",
712
+ "model.norm.weight": "model.safetensors"
713
+ }
714
+ }
quant_strategy.json ADDED
@@ -0,0 +1,1668 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "measurement": {
3
+ "model.layers.0": {
4
+ "accuracy": 0.6637592315673828,
5
+ "total_bits": 357441024,
6
+ "o_proj": {
7
+ "group_size": {
8
+ "4": 128
9
+ },
10
+ "bits": [
11
+ 4
12
+ ],
13
+ "bits_prop": [
14
+ 1
15
+ ],
16
+ "scale_bits": 4
17
+ },
18
+ "down_proj": {
19
+ "group_size": {
20
+ "4": 128
21
+ },
22
+ "bits": [
23
+ 4
24
+ ],
25
+ "bits_prop": [
26
+ 1
27
+ ],
28
+ "scale_bits": 4
29
+ },
30
+ "qkv_proj": {
31
+ "group_size": {
32
+ "4": 128
33
+ },
34
+ "bits": [
35
+ 4
36
+ ],
37
+ "bits_prop": [
38
+ 1
39
+ ],
40
+ "scale_bits": 4
41
+ },
42
+ "gate_up_proj": {
43
+ "group_size": {
44
+ "2": 64
45
+ },
46
+ "bits": [
47
+ 2
48
+ ],
49
+ "bits_prop": [
50
+ 1
51
+ ],
52
+ "scale_bits": 4
53
+ }
54
+ },
55
+ "model.layers.1": {
56
+ "accuracy": 0.7465658187866211,
57
+ "total_bits": 357441024,
58
+ "o_proj": {
59
+ "group_size": {
60
+ "4": 128
61
+ },
62
+ "bits": [
63
+ 4
64
+ ],
65
+ "bits_prop": [
66
+ 1
67
+ ],
68
+ "scale_bits": 4
69
+ },
70
+ "down_proj": {
71
+ "group_size": {
72
+ "4": 128
73
+ },
74
+ "bits": [
75
+ 4
76
+ ],
77
+ "bits_prop": [
78
+ 1
79
+ ],
80
+ "scale_bits": 4
81
+ },
82
+ "qkv_proj": {
83
+ "group_size": {
84
+ "4": 128
85
+ },
86
+ "bits": [
87
+ 4
88
+ ],
89
+ "bits_prop": [
90
+ 1
91
+ ],
92
+ "scale_bits": 4
93
+ },
94
+ "gate_up_proj": {
95
+ "group_size": {
96
+ "2": 64
97
+ },
98
+ "bits": [
99
+ 2
100
+ ],
101
+ "bits_prop": [
102
+ 1
103
+ ],
104
+ "scale_bits": 4
105
+ }
106
+ },
107
+ "model.layers.2": {
108
+ "accuracy": 0.898890495300293,
109
+ "total_bits": 357441024,
110
+ "o_proj": {
111
+ "group_size": {
112
+ "4": 128
113
+ },
114
+ "bits": [
115
+ 4
116
+ ],
117
+ "bits_prop": [
118
+ 1
119
+ ],
120
+ "scale_bits": 4
121
+ },
122
+ "down_proj": {
123
+ "group_size": {
124
+ "4": 128
125
+ },
126
+ "bits": [
127
+ 4
128
+ ],
129
+ "bits_prop": [
130
+ 1
131
+ ],
132
+ "scale_bits": 4
133
+ },
134
+ "qkv_proj": {
135
+ "group_size": {
136
+ "4": 128
137
+ },
138
+ "bits": [
139
+ 4
140
+ ],
141
+ "bits_prop": [
142
+ 1
143
+ ],
144
+ "scale_bits": 4
145
+ },
146
+ "gate_up_proj": {
147
+ "group_size": {
148
+ "2": 64
149
+ },
150
+ "bits": [
151
+ 2
152
+ ],
153
+ "bits_prop": [
154
+ 1
155
+ ],
156
+ "scale_bits": 4
157
+ }
158
+ },
159
+ "model.layers.3": {
160
+ "accuracy": 0.8928320407867432,
161
+ "total_bits": 289320576,
162
+ "o_proj": {
163
+ "group_size": {
164
+ "2": 64
165
+ },
166
+ "bits": [
167
+ 2
168
+ ],
169
+ "bits_prop": [
170
+ 1
171
+ ],
172
+ "scale_bits": 4
173
+ },
174
+ "down_proj": {
175
+ "group_size": {
176
+ "2": 64
177
+ },
178
+ "bits": [
179
+ 2
180
+ ],
181
+ "bits_prop": [
182
+ 1
183
+ ],
184
+ "scale_bits": 4
185
+ },
186
+ "qkv_proj": {
187
+ "group_size": {
188
+ "4": 128
189
+ },
190
+ "bits": [
191
+ 4
192
+ ],
193
+ "bits_prop": [
194
+ 1
195
+ ],
196
+ "scale_bits": 4
197
+ },
198
+ "gate_up_proj": {
199
+ "group_size": {
200
+ "2": 64
201
+ },
202
+ "bits": [
203
+ 2
204
+ ],
205
+ "bits_prop": [
206
+ 1
207
+ ],
208
+ "scale_bits": 4
209
+ }
210
+ },
211
+ "model.layers.4": {
212
+ "accuracy": 0.8153390884399414,
213
+ "total_bits": 289320576,
214
+ "o_proj": {
215
+ "group_size": {
216
+ "2": 64
217
+ },
218
+ "bits": [
219
+ 2
220
+ ],
221
+ "bits_prop": [
222
+ 1
223
+ ],
224
+ "scale_bits": 4
225
+ },
226
+ "down_proj": {
227
+ "group_size": {
228
+ "2": 64
229
+ },
230
+ "bits": [
231
+ 2
232
+ ],
233
+ "bits_prop": [
234
+ 1
235
+ ],
236
+ "scale_bits": 4
237
+ },
238
+ "qkv_proj": {
239
+ "group_size": {
240
+ "4": 128
241
+ },
242
+ "bits": [
243
+ 4
244
+ ],
245
+ "bits_prop": [
246
+ 1
247
+ ],
248
+ "scale_bits": 4
249
+ },
250
+ "gate_up_proj": {
251
+ "group_size": {
252
+ "2": 64
253
+ },
254
+ "bits": [
255
+ 2
256
+ ],
257
+ "bits_prop": [
258
+ 1
259
+ ],
260
+ "scale_bits": 4
261
+ }
262
+ },
263
+ "model.layers.5": {
264
+ "accuracy": 0.9311076402664185,
265
+ "total_bits": 289320576,
266
+ "o_proj": {
267
+ "group_size": {
268
+ "2": 64
269
+ },
270
+ "bits": [
271
+ 2
272
+ ],
273
+ "bits_prop": [
274
+ 1
275
+ ],
276
+ "scale_bits": 4
277
+ },
278
+ "down_proj": {
279
+ "group_size": {
280
+ "2": 64
281
+ },
282
+ "bits": [
283
+ 2
284
+ ],
285
+ "bits_prop": [
286
+ 1
287
+ ],
288
+ "scale_bits": 4
289
+ },
290
+ "qkv_proj": {
291
+ "group_size": {
292
+ "4": 128
293
+ },
294
+ "bits": [
295
+ 4
296
+ ],
297
+ "bits_prop": [
298
+ 1
299
+ ],
300
+ "scale_bits": 4
301
+ },
302
+ "gate_up_proj": {
303
+ "group_size": {
304
+ "2": 64
305
+ },
306
+ "bits": [
307
+ 2
308
+ ],
309
+ "bits_prop": [
310
+ 1
311
+ ],
312
+ "scale_bits": 4
313
+ }
314
+ },
315
+ "model.layers.6": {
316
+ "accuracy": 0.9273414611816406,
317
+ "total_bits": 289320576,
318
+ "o_proj": {
319
+ "group_size": {
320
+ "2": 64
321
+ },
322
+ "bits": [
323
+ 2
324
+ ],
325
+ "bits_prop": [
326
+ 1
327
+ ],
328
+ "scale_bits": 4
329
+ },
330
+ "down_proj": {
331
+ "group_size": {
332
+ "2": 64
333
+ },
334
+ "bits": [
335
+ 2
336
+ ],
337
+ "bits_prop": [
338
+ 1
339
+ ],
340
+ "scale_bits": 4
341
+ },
342
+ "qkv_proj": {
343
+ "group_size": {
344
+ "4": 128
345
+ },
346
+ "bits": [
347
+ 4
348
+ ],
349
+ "bits_prop": [
350
+ 1
351
+ ],
352
+ "scale_bits": 4
353
+ },
354
+ "gate_up_proj": {
355
+ "group_size": {
356
+ "2": 64
357
+ },
358
+ "bits": [
359
+ 2
360
+ ],
361
+ "bits_prop": [
362
+ 1
363
+ ],
364
+ "scale_bits": 4
365
+ }
366
+ },
367
+ "model.layers.7": {
368
+ "accuracy": 0.8751392364501953,
369
+ "total_bits": 289320576,
370
+ "o_proj": {
371
+ "group_size": {
372
+ "2": 64
373
+ },
374
+ "bits": [
375
+ 2
376
+ ],
377
+ "bits_prop": [
378
+ 1
379
+ ],
380
+ "scale_bits": 4
381
+ },
382
+ "down_proj": {
383
+ "group_size": {
384
+ "2": 64
385
+ },
386
+ "bits": [
387
+ 2
388
+ ],
389
+ "bits_prop": [
390
+ 1
391
+ ],
392
+ "scale_bits": 4
393
+ },
394
+ "qkv_proj": {
395
+ "group_size": {
396
+ "4": 128
397
+ },
398
+ "bits": [
399
+ 4
400
+ ],
401
+ "bits_prop": [
402
+ 1
403
+ ],
404
+ "scale_bits": 4
405
+ },
406
+ "gate_up_proj": {
407
+ "group_size": {
408
+ "2": 64
409
+ },
410
+ "bits": [
411
+ 2
412
+ ],
413
+ "bits_prop": [
414
+ 1
415
+ ],
416
+ "scale_bits": 4
417
+ }
418
+ },
419
+ "model.layers.8": {
420
+ "accuracy": 0.872112512588501,
421
+ "total_bits": 283125504,
422
+ "o_proj": {
423
+ "group_size": {
424
+ "2": 64
425
+ },
426
+ "bits": [
427
+ 2
428
+ ],
429
+ "bits_prop": [
430
+ 1
431
+ ],
432
+ "scale_bits": 4
433
+ },
434
+ "down_proj": {
435
+ "group_size": {
436
+ "4": 128
437
+ },
438
+ "bits": [
439
+ 4
440
+ ],
441
+ "bits_prop": [
442
+ 1
443
+ ],
444
+ "scale_bits": 4
445
+ },
446
+ "qkv_proj": {
447
+ "group_size": {
448
+ "2": 64
449
+ },
450
+ "bits": [
451
+ 2
452
+ ],
453
+ "bits_prop": [
454
+ 1
455
+ ],
456
+ "scale_bits": 4
457
+ },
458
+ "gate_up_proj": {
459
+ "group_size": {
460
+ "2": 64
461
+ },
462
+ "bits": [
463
+ 2
464
+ ],
465
+ "bits_prop": [
466
+ 1
467
+ ],
468
+ "scale_bits": 4
469
+ }
470
+ },
471
+ "model.layers.9": {
472
+ "accuracy": 0.8808994293212891,
473
+ "total_bits": 233583360,
474
+ "o_proj": {
475
+ "group_size": {
476
+ "2": 64
477
+ },
478
+ "bits": [
479
+ 2
480
+ ],
481
+ "bits_prop": [
482
+ 1
483
+ ],
484
+ "scale_bits": 4
485
+ },
486
+ "down_proj": {
487
+ "group_size": {
488
+ "2": 64
489
+ },
490
+ "bits": [
491
+ 2
492
+ ],
493
+ "bits_prop": [
494
+ 1
495
+ ],
496
+ "scale_bits": 4
497
+ },
498
+ "qkv_proj": {
499
+ "group_size": {
500
+ "2": 64
501
+ },
502
+ "bits": [
503
+ 2
504
+ ],
505
+ "bits_prop": [
506
+ 1
507
+ ],
508
+ "scale_bits": 4
509
+ },
510
+ "gate_up_proj": {
511
+ "group_size": {
512
+ "2": 64
513
+ },
514
+ "bits": [
515
+ 2
516
+ ],
517
+ "bits_prop": [
518
+ 1
519
+ ],
520
+ "scale_bits": 4
521
+ }
522
+ },
523
+ "model.layers.10": {
524
+ "accuracy": 0.8999626636505127,
525
+ "total_bits": 283125504,
526
+ "o_proj": {
527
+ "group_size": {
528
+ "2": 64
529
+ },
530
+ "bits": [
531
+ 2
532
+ ],
533
+ "bits_prop": [
534
+ 1
535
+ ],
536
+ "scale_bits": 4
537
+ },
538
+ "down_proj": {
539
+ "group_size": {
540
+ "4": 128
541
+ },
542
+ "bits": [
543
+ 4
544
+ ],
545
+ "bits_prop": [
546
+ 1
547
+ ],
548
+ "scale_bits": 4
549
+ },
550
+ "qkv_proj": {
551
+ "group_size": {
552
+ "2": 64
553
+ },
554
+ "bits": [
555
+ 2
556
+ ],
557
+ "bits_prop": [
558
+ 1
559
+ ],
560
+ "scale_bits": 4
561
+ },
562
+ "gate_up_proj": {
563
+ "group_size": {
564
+ "2": 64
565
+ },
566
+ "bits": [
567
+ 2
568
+ ],
569
+ "bits_prop": [
570
+ 1
571
+ ],
572
+ "scale_bits": 4
573
+ }
574
+ },
575
+ "model.layers.11": {
576
+ "accuracy": 0.8984930515289307,
577
+ "total_bits": 283125504,
578
+ "o_proj": {
579
+ "group_size": {
580
+ "2": 64
581
+ },
582
+ "bits": [
583
+ 2
584
+ ],
585
+ "bits_prop": [
586
+ 1
587
+ ],
588
+ "scale_bits": 4
589
+ },
590
+ "down_proj": {
591
+ "group_size": {
592
+ "4": 128
593
+ },
594
+ "bits": [
595
+ 4
596
+ ],
597
+ "bits_prop": [
598
+ 1
599
+ ],
600
+ "scale_bits": 4
601
+ },
602
+ "qkv_proj": {
603
+ "group_size": {
604
+ "2": 64
605
+ },
606
+ "bits": [
607
+ 2
608
+ ],
609
+ "bits_prop": [
610
+ 1
611
+ ],
612
+ "scale_bits": 4
613
+ },
614
+ "gate_up_proj": {
615
+ "group_size": {
616
+ "2": 64
617
+ },
618
+ "bits": [
619
+ 2
620
+ ],
621
+ "bits_prop": [
622
+ 1
623
+ ],
624
+ "scale_bits": 4
625
+ }
626
+ },
627
+ "model.layers.12": {
628
+ "accuracy": 0.8965449333190918,
629
+ "total_bits": 252161664,
630
+ "o_proj": {
631
+ "group_size": {
632
+ "4": 128
633
+ },
634
+ "bits": [
635
+ 4
636
+ ],
637
+ "bits_prop": [
638
+ 1
639
+ ],
640
+ "scale_bits": 4
641
+ },
642
+ "down_proj": {
643
+ "group_size": {
644
+ "2": 64
645
+ },
646
+ "bits": [
647
+ 2
648
+ ],
649
+ "bits_prop": [
650
+ 1
651
+ ],
652
+ "scale_bits": 4
653
+ },
654
+ "qkv_proj": {
655
+ "group_size": {
656
+ "2": 64
657
+ },
658
+ "bits": [
659
+ 2
660
+ ],
661
+ "bits_prop": [
662
+ 1
663
+ ],
664
+ "scale_bits": 4
665
+ },
666
+ "gate_up_proj": {
667
+ "group_size": {
668
+ "2": 64
669
+ },
670
+ "bits": [
671
+ 2
672
+ ],
673
+ "bits_prop": [
674
+ 1
675
+ ],
676
+ "scale_bits": 4
677
+ }
678
+ },
679
+ "model.layers.13": {
680
+ "accuracy": 0.924396276473999,
681
+ "total_bits": 301703808,
682
+ "o_proj": {
683
+ "group_size": {
684
+ "4": 128
685
+ },
686
+ "bits": [
687
+ 4
688
+ ],
689
+ "bits_prop": [
690
+ 1
691
+ ],
692
+ "scale_bits": 4
693
+ },
694
+ "down_proj": {
695
+ "group_size": {
696
+ "4": 128
697
+ },
698
+ "bits": [
699
+ 4
700
+ ],
701
+ "bits_prop": [
702
+ 1
703
+ ],
704
+ "scale_bits": 4
705
+ },
706
+ "qkv_proj": {
707
+ "group_size": {
708
+ "2": 64
709
+ },
710
+ "bits": [
711
+ 2
712
+ ],
713
+ "bits_prop": [
714
+ 1
715
+ ],
716
+ "scale_bits": 4
717
+ },
718
+ "gate_up_proj": {
719
+ "group_size": {
720
+ "2": 64
721
+ },
722
+ "bits": [
723
+ 2
724
+ ],
725
+ "bits_prop": [
726
+ 1
727
+ ],
728
+ "scale_bits": 4
729
+ }
730
+ },
731
+ "model.layers.14": {
732
+ "accuracy": 0.9207086563110352,
733
+ "total_bits": 233583360,
734
+ "o_proj": {
735
+ "group_size": {
736
+ "2": 64
737
+ },
738
+ "bits": [
739
+ 2
740
+ ],
741
+ "bits_prop": [
742
+ 1
743
+ ],
744
+ "scale_bits": 4
745
+ },
746
+ "down_proj": {
747
+ "group_size": {
748
+ "2": 64
749
+ },
750
+ "bits": [
751
+ 2
752
+ ],
753
+ "bits_prop": [
754
+ 1
755
+ ],
756
+ "scale_bits": 4
757
+ },
758
+ "qkv_proj": {
759
+ "group_size": {
760
+ "2": 64
761
+ },
762
+ "bits": [
763
+ 2
764
+ ],
765
+ "bits_prop": [
766
+ 1
767
+ ],
768
+ "scale_bits": 4
769
+ },
770
+ "gate_up_proj": {
771
+ "group_size": {
772
+ "2": 64
773
+ },
774
+ "bits": [
775
+ 2
776
+ ],
777
+ "bits_prop": [
778
+ 1
779
+ ],
780
+ "scale_bits": 4
781
+ }
782
+ },
783
+ "model.layers.15": {
784
+ "accuracy": 0.9219722747802734,
785
+ "total_bits": 233583360,
786
+ "o_proj": {
787
+ "group_size": {
788
+ "2": 64
789
+ },
790
+ "bits": [
791
+ 2
792
+ ],
793
+ "bits_prop": [
794
+ 1
795
+ ],
796
+ "scale_bits": 4
797
+ },
798
+ "down_proj": {
799
+ "group_size": {
800
+ "2": 64
801
+ },
802
+ "bits": [
803
+ 2
804
+ ],
805
+ "bits_prop": [
806
+ 1
807
+ ],
808
+ "scale_bits": 4
809
+ },
810
+ "qkv_proj": {
811
+ "group_size": {
812
+ "2": 64
813
+ },
814
+ "bits": [
815
+ 2
816
+ ],
817
+ "bits_prop": [
818
+ 1
819
+ ],
820
+ "scale_bits": 4
821
+ },
822
+ "gate_up_proj": {
823
+ "group_size": {
824
+ "2": 64
825
+ },
826
+ "bits": [
827
+ 2
828
+ ],
829
+ "bits_prop": [
830
+ 1
831
+ ],
832
+ "scale_bits": 4
833
+ }
834
+ },
835
+ "model.layers.16": {
836
+ "accuracy": 0.9315023422241211,
837
+ "total_bits": 252161664,
838
+ "o_proj": {
839
+ "group_size": {
840
+ "4": 128
841
+ },
842
+ "bits": [
843
+ 4
844
+ ],
845
+ "bits_prop": [
846
+ 1
847
+ ],
848
+ "scale_bits": 4
849
+ },
850
+ "down_proj": {
851
+ "group_size": {
852
+ "2": 64
853
+ },
854
+ "bits": [
855
+ 2
856
+ ],
857
+ "bits_prop": [
858
+ 1
859
+ ],
860
+ "scale_bits": 4
861
+ },
862
+ "qkv_proj": {
863
+ "group_size": {
864
+ "2": 64
865
+ },
866
+ "bits": [
867
+ 2
868
+ ],
869
+ "bits_prop": [
870
+ 1
871
+ ],
872
+ "scale_bits": 4
873
+ },
874
+ "gate_up_proj": {
875
+ "group_size": {
876
+ "2": 64
877
+ },
878
+ "bits": [
879
+ 2
880
+ ],
881
+ "bits_prop": [
882
+ 1
883
+ ],
884
+ "scale_bits": 4
885
+ }
886
+ },
887
+ "model.layers.17": {
888
+ "accuracy": 0.9288318157196045,
889
+ "total_bits": 283125504,
890
+ "o_proj": {
891
+ "group_size": {
892
+ "2": 64
893
+ },
894
+ "bits": [
895
+ 2
896
+ ],
897
+ "bits_prop": [
898
+ 1
899
+ ],
900
+ "scale_bits": 4
901
+ },
902
+ "down_proj": {
903
+ "group_size": {
904
+ "4": 128
905
+ },
906
+ "bits": [
907
+ 4
908
+ ],
909
+ "bits_prop": [
910
+ 1
911
+ ],
912
+ "scale_bits": 4
913
+ },
914
+ "qkv_proj": {
915
+ "group_size": {
916
+ "2": 64
917
+ },
918
+ "bits": [
919
+ 2
920
+ ],
921
+ "bits_prop": [
922
+ 1
923
+ ],
924
+ "scale_bits": 4
925
+ },
926
+ "gate_up_proj": {
927
+ "group_size": {
928
+ "2": 64
929
+ },
930
+ "bits": [
931
+ 2
932
+ ],
933
+ "bits_prop": [
934
+ 1
935
+ ],
936
+ "scale_bits": 4
937
+ }
938
+ },
939
+ "model.layers.18": {
940
+ "accuracy": 0.9211232662200928,
941
+ "total_bits": 283125504,
942
+ "o_proj": {
943
+ "group_size": {
944
+ "2": 64
945
+ },
946
+ "bits": [
947
+ 2
948
+ ],
949
+ "bits_prop": [
950
+ 1
951
+ ],
952
+ "scale_bits": 4
953
+ },
954
+ "down_proj": {
955
+ "group_size": {
956
+ "4": 128
957
+ },
958
+ "bits": [
959
+ 4
960
+ ],
961
+ "bits_prop": [
962
+ 1
963
+ ],
964
+ "scale_bits": 4
965
+ },
966
+ "qkv_proj": {
967
+ "group_size": {
968
+ "2": 64
969
+ },
970
+ "bits": [
971
+ 2
972
+ ],
973
+ "bits_prop": [
974
+ 1
975
+ ],
976
+ "scale_bits": 4
977
+ },
978
+ "gate_up_proj": {
979
+ "group_size": {
980
+ "2": 64
981
+ },
982
+ "bits": [
983
+ 2
984
+ ],
985
+ "bits_prop": [
986
+ 1
987
+ ],
988
+ "scale_bits": 4
989
+ }
990
+ },
991
+ "model.layers.19": {
992
+ "accuracy": 0.9357051849365234,
993
+ "total_bits": 289320576,
994
+ "o_proj": {
995
+ "group_size": {
996
+ "2": 64
997
+ },
998
+ "bits": [
999
+ 2
1000
+ ],
1001
+ "bits_prop": [
1002
+ 1
1003
+ ],
1004
+ "scale_bits": 4
1005
+ },
1006
+ "down_proj": {
1007
+ "group_size": {
1008
+ "2": 64
1009
+ },
1010
+ "bits": [
1011
+ 2
1012
+ ],
1013
+ "bits_prop": [
1014
+ 1
1015
+ ],
1016
+ "scale_bits": 4
1017
+ },
1018
+ "qkv_proj": {
1019
+ "group_size": {
1020
+ "4": 128
1021
+ },
1022
+ "bits": [
1023
+ 4
1024
+ ],
1025
+ "bits_prop": [
1026
+ 1
1027
+ ],
1028
+ "scale_bits": 4
1029
+ },
1030
+ "gate_up_proj": {
1031
+ "group_size": {
1032
+ "2": 64
1033
+ },
1034
+ "bits": [
1035
+ 2
1036
+ ],
1037
+ "bits_prop": [
1038
+ 1
1039
+ ],
1040
+ "scale_bits": 4
1041
+ }
1042
+ },
1043
+ "model.layers.20": {
1044
+ "accuracy": 0.9271621704101562,
1045
+ "total_bits": 289320576,
1046
+ "o_proj": {
1047
+ "group_size": {
1048
+ "2": 64
1049
+ },
1050
+ "bits": [
1051
+ 2
1052
+ ],
1053
+ "bits_prop": [
1054
+ 1
1055
+ ],
1056
+ "scale_bits": 4
1057
+ },
1058
+ "down_proj": {
1059
+ "group_size": {
1060
+ "2": 64
1061
+ },
1062
+ "bits": [
1063
+ 2
1064
+ ],
1065
+ "bits_prop": [
1066
+ 1
1067
+ ],
1068
+ "scale_bits": 4
1069
+ },
1070
+ "qkv_proj": {
1071
+ "group_size": {
1072
+ "4": 128
1073
+ },
1074
+ "bits": [
1075
+ 4
1076
+ ],
1077
+ "bits_prop": [
1078
+ 1
1079
+ ],
1080
+ "scale_bits": 4
1081
+ },
1082
+ "gate_up_proj": {
1083
+ "group_size": {
1084
+ "2": 64
1085
+ },
1086
+ "bits": [
1087
+ 2
1088
+ ],
1089
+ "bits_prop": [
1090
+ 1
1091
+ ],
1092
+ "scale_bits": 4
1093
+ }
1094
+ },
1095
+ "model.layers.21": {
1096
+ "accuracy": 0.9496371746063232,
1097
+ "total_bits": 357441024,
1098
+ "o_proj": {
1099
+ "group_size": {
1100
+ "4": 128
1101
+ },
1102
+ "bits": [
1103
+ 4
1104
+ ],
1105
+ "bits_prop": [
1106
+ 1
1107
+ ],
1108
+ "scale_bits": 4
1109
+ },
1110
+ "down_proj": {
1111
+ "group_size": {
1112
+ "4": 128
1113
+ },
1114
+ "bits": [
1115
+ 4
1116
+ ],
1117
+ "bits_prop": [
1118
+ 1
1119
+ ],
1120
+ "scale_bits": 4
1121
+ },
1122
+ "qkv_proj": {
1123
+ "group_size": {
1124
+ "4": 128
1125
+ },
1126
+ "bits": [
1127
+ 4
1128
+ ],
1129
+ "bits_prop": [
1130
+ 1
1131
+ ],
1132
+ "scale_bits": 4
1133
+ },
1134
+ "gate_up_proj": {
1135
+ "group_size": {
1136
+ "2": 64
1137
+ },
1138
+ "bits": [
1139
+ 2
1140
+ ],
1141
+ "bits_prop": [
1142
+ 1
1143
+ ],
1144
+ "scale_bits": 4
1145
+ }
1146
+ },
1147
+ "model.layers.22": {
1148
+ "accuracy": 0.9459362030029297,
1149
+ "total_bits": 357441024,
1150
+ "o_proj": {
1151
+ "group_size": {
1152
+ "4": 128
1153
+ },
1154
+ "bits": [
1155
+ 4
1156
+ ],
1157
+ "bits_prop": [
1158
+ 1
1159
+ ],
1160
+ "scale_bits": 4
1161
+ },
1162
+ "down_proj": {
1163
+ "group_size": {
1164
+ "4": 128
1165
+ },
1166
+ "bits": [
1167
+ 4
1168
+ ],
1169
+ "bits_prop": [
1170
+ 1
1171
+ ],
1172
+ "scale_bits": 4
1173
+ },
1174
+ "qkv_proj": {
1175
+ "group_size": {
1176
+ "4": 128
1177
+ },
1178
+ "bits": [
1179
+ 4
1180
+ ],
1181
+ "bits_prop": [
1182
+ 1
1183
+ ],
1184
+ "scale_bits": 4
1185
+ },
1186
+ "gate_up_proj": {
1187
+ "group_size": {
1188
+ "2": 64
1189
+ },
1190
+ "bits": [
1191
+ 2
1192
+ ],
1193
+ "bits_prop": [
1194
+ 1
1195
+ ],
1196
+ "scale_bits": 4
1197
+ }
1198
+ },
1199
+ "model.layers.23": {
1200
+ "accuracy": 0.9806646108627319,
1201
+ "total_bits": 456530304,
1202
+ "o_proj": {
1203
+ "group_size": {
1204
+ "4": 128
1205
+ },
1206
+ "bits": [
1207
+ 4
1208
+ ],
1209
+ "bits_prop": [
1210
+ 1
1211
+ ],
1212
+ "scale_bits": 4
1213
+ },
1214
+ "down_proj": {
1215
+ "group_size": {
1216
+ "4": 128
1217
+ },
1218
+ "bits": [
1219
+ 4
1220
+ ],
1221
+ "bits_prop": [
1222
+ 1
1223
+ ],
1224
+ "scale_bits": 4
1225
+ },
1226
+ "qkv_proj": {
1227
+ "group_size": {
1228
+ "4": 128
1229
+ },
1230
+ "bits": [
1231
+ 4
1232
+ ],
1233
+ "bits_prop": [
1234
+ 1
1235
+ ],
1236
+ "scale_bits": 4
1237
+ },
1238
+ "gate_up_proj": {
1239
+ "group_size": {
1240
+ "4": 128
1241
+ },
1242
+ "bits": [
1243
+ 4
1244
+ ],
1245
+ "bits_prop": [
1246
+ 1
1247
+ ],
1248
+ "scale_bits": 4
1249
+ }
1250
+ },
1251
+ "model.layers.24": {
1252
+ "accuracy": 0.9800381660461426,
1253
+ "total_bits": 456530304,
1254
+ "o_proj": {
1255
+ "group_size": {
1256
+ "4": 128
1257
+ },
1258
+ "bits": [
1259
+ 4
1260
+ ],
1261
+ "bits_prop": [
1262
+ 1
1263
+ ],
1264
+ "scale_bits": 4
1265
+ },
1266
+ "down_proj": {
1267
+ "group_size": {
1268
+ "4": 128
1269
+ },
1270
+ "bits": [
1271
+ 4
1272
+ ],
1273
+ "bits_prop": [
1274
+ 1
1275
+ ],
1276
+ "scale_bits": 4
1277
+ },
1278
+ "qkv_proj": {
1279
+ "group_size": {
1280
+ "4": 128
1281
+ },
1282
+ "bits": [
1283
+ 4
1284
+ ],
1285
+ "bits_prop": [
1286
+ 1
1287
+ ],
1288
+ "scale_bits": 4
1289
+ },
1290
+ "gate_up_proj": {
1291
+ "group_size": {
1292
+ "4": 128
1293
+ },
1294
+ "bits": [
1295
+ 4
1296
+ ],
1297
+ "bits_prop": [
1298
+ 1
1299
+ ],
1300
+ "scale_bits": 4
1301
+ }
1302
+ },
1303
+ "model.layers.25": {
1304
+ "accuracy": 0.980277419090271,
1305
+ "total_bits": 456530304,
1306
+ "o_proj": {
1307
+ "group_size": {
1308
+ "4": 128
1309
+ },
1310
+ "bits": [
1311
+ 4
1312
+ ],
1313
+ "bits_prop": [
1314
+ 1
1315
+ ],
1316
+ "scale_bits": 4
1317
+ },
1318
+ "down_proj": {
1319
+ "group_size": {
1320
+ "4": 128
1321
+ },
1322
+ "bits": [
1323
+ 4
1324
+ ],
1325
+ "bits_prop": [
1326
+ 1
1327
+ ],
1328
+ "scale_bits": 4
1329
+ },
1330
+ "qkv_proj": {
1331
+ "group_size": {
1332
+ "4": 128
1333
+ },
1334
+ "bits": [
1335
+ 4
1336
+ ],
1337
+ "bits_prop": [
1338
+ 1
1339
+ ],
1340
+ "scale_bits": 4
1341
+ },
1342
+ "gate_up_proj": {
1343
+ "group_size": {
1344
+ "4": 128
1345
+ },
1346
+ "bits": [
1347
+ 4
1348
+ ],
1349
+ "bits_prop": [
1350
+ 1
1351
+ ],
1352
+ "scale_bits": 4
1353
+ }
1354
+ },
1355
+ "model.layers.26": {
1356
+ "accuracy": 0.9804829359054565,
1357
+ "total_bits": 456530304,
1358
+ "o_proj": {
1359
+ "group_size": {
1360
+ "4": 128
1361
+ },
1362
+ "bits": [
1363
+ 4
1364
+ ],
1365
+ "bits_prop": [
1366
+ 1
1367
+ ],
1368
+ "scale_bits": 4
1369
+ },
1370
+ "down_proj": {
1371
+ "group_size": {
1372
+ "4": 128
1373
+ },
1374
+ "bits": [
1375
+ 4
1376
+ ],
1377
+ "bits_prop": [
1378
+ 1
1379
+ ],
1380
+ "scale_bits": 4
1381
+ },
1382
+ "qkv_proj": {
1383
+ "group_size": {
1384
+ "4": 128
1385
+ },
1386
+ "bits": [
1387
+ 4
1388
+ ],
1389
+ "bits_prop": [
1390
+ 1
1391
+ ],
1392
+ "scale_bits": 4
1393
+ },
1394
+ "gate_up_proj": {
1395
+ "group_size": {
1396
+ "4": 128
1397
+ },
1398
+ "bits": [
1399
+ 4
1400
+ ],
1401
+ "bits_prop": [
1402
+ 1
1403
+ ],
1404
+ "scale_bits": 4
1405
+ }
1406
+ },
1407
+ "model.layers.27": {
1408
+ "accuracy": 0.9804953336715698,
1409
+ "total_bits": 456530304,
1410
+ "o_proj": {
1411
+ "group_size": {
1412
+ "4": 128
1413
+ },
1414
+ "bits": [
1415
+ 4
1416
+ ],
1417
+ "bits_prop": [
1418
+ 1
1419
+ ],
1420
+ "scale_bits": 4
1421
+ },
1422
+ "down_proj": {
1423
+ "group_size": {
1424
+ "4": 128
1425
+ },
1426
+ "bits": [
1427
+ 4
1428
+ ],
1429
+ "bits_prop": [
1430
+ 1
1431
+ ],
1432
+ "scale_bits": 4
1433
+ },
1434
+ "qkv_proj": {
1435
+ "group_size": {
1436
+ "4": 128
1437
+ },
1438
+ "bits": [
1439
+ 4
1440
+ ],
1441
+ "bits_prop": [
1442
+ 1
1443
+ ],
1444
+ "scale_bits": 4
1445
+ },
1446
+ "gate_up_proj": {
1447
+ "group_size": {
1448
+ "4": 128
1449
+ },
1450
+ "bits": [
1451
+ 4
1452
+ ],
1453
+ "bits_prop": [
1454
+ 1
1455
+ ],
1456
+ "scale_bits": 4
1457
+ }
1458
+ },
1459
+ "model.layers.28": {
1460
+ "accuracy": 0.9789243936538696,
1461
+ "total_bits": 456530304,
1462
+ "o_proj": {
1463
+ "group_size": {
1464
+ "4": 128
1465
+ },
1466
+ "bits": [
1467
+ 4
1468
+ ],
1469
+ "bits_prop": [
1470
+ 1
1471
+ ],
1472
+ "scale_bits": 4
1473
+ },
1474
+ "down_proj": {
1475
+ "group_size": {
1476
+ "4": 128
1477
+ },
1478
+ "bits": [
1479
+ 4
1480
+ ],
1481
+ "bits_prop": [
1482
+ 1
1483
+ ],
1484
+ "scale_bits": 4
1485
+ },
1486
+ "qkv_proj": {
1487
+ "group_size": {
1488
+ "4": 128
1489
+ },
1490
+ "bits": [
1491
+ 4
1492
+ ],
1493
+ "bits_prop": [
1494
+ 1
1495
+ ],
1496
+ "scale_bits": 4
1497
+ },
1498
+ "gate_up_proj": {
1499
+ "group_size": {
1500
+ "4": 128
1501
+ },
1502
+ "bits": [
1503
+ 4
1504
+ ],
1505
+ "bits_prop": [
1506
+ 1
1507
+ ],
1508
+ "scale_bits": 4
1509
+ }
1510
+ },
1511
+ "model.layers.29": {
1512
+ "accuracy": 0.9709323644638062,
1513
+ "total_bits": 456530304,
1514
+ "o_proj": {
1515
+ "group_size": {
1516
+ "4": 128
1517
+ },
1518
+ "bits": [
1519
+ 4
1520
+ ],
1521
+ "bits_prop": [
1522
+ 1
1523
+ ],
1524
+ "scale_bits": 4
1525
+ },
1526
+ "down_proj": {
1527
+ "group_size": {
1528
+ "4": 128
1529
+ },
1530
+ "bits": [
1531
+ 4
1532
+ ],
1533
+ "bits_prop": [
1534
+ 1
1535
+ ],
1536
+ "scale_bits": 4
1537
+ },
1538
+ "qkv_proj": {
1539
+ "group_size": {
1540
+ "4": 128
1541
+ },
1542
+ "bits": [
1543
+ 4
1544
+ ],
1545
+ "bits_prop": [
1546
+ 1
1547
+ ],
1548
+ "scale_bits": 4
1549
+ },
1550
+ "gate_up_proj": {
1551
+ "group_size": {
1552
+ "4": 128
1553
+ },
1554
+ "bits": [
1555
+ 4
1556
+ ],
1557
+ "bits_prop": [
1558
+ 1
1559
+ ],
1560
+ "scale_bits": 4
1561
+ }
1562
+ },
1563
+ "model.layers.30": {
1564
+ "accuracy": 0.9765185117721558,
1565
+ "total_bits": 456530304,
1566
+ "o_proj": {
1567
+ "group_size": {
1568
+ "4": 128
1569
+ },
1570
+ "bits": [
1571
+ 4
1572
+ ],
1573
+ "bits_prop": [
1574
+ 1
1575
+ ],
1576
+ "scale_bits": 4
1577
+ },
1578
+ "down_proj": {
1579
+ "group_size": {
1580
+ "4": 128
1581
+ },
1582
+ "bits": [
1583
+ 4
1584
+ ],
1585
+ "bits_prop": [
1586
+ 1
1587
+ ],
1588
+ "scale_bits": 4
1589
+ },
1590
+ "qkv_proj": {
1591
+ "group_size": {
1592
+ "4": 128
1593
+ },
1594
+ "bits": [
1595
+ 4
1596
+ ],
1597
+ "bits_prop": [
1598
+ 1
1599
+ ],
1600
+ "scale_bits": 4
1601
+ },
1602
+ "gate_up_proj": {
1603
+ "group_size": {
1604
+ "4": 128
1605
+ },
1606
+ "bits": [
1607
+ 4
1608
+ ],
1609
+ "bits_prop": [
1610
+ 1
1611
+ ],
1612
+ "scale_bits": 4
1613
+ }
1614
+ },
1615
+ "model.layers.31": {
1616
+ "accuracy": 0.9663764238357544,
1617
+ "total_bits": 456530304,
1618
+ "o_proj": {
1619
+ "group_size": {
1620
+ "4": 128
1621
+ },
1622
+ "bits": [
1623
+ 4
1624
+ ],
1625
+ "bits_prop": [
1626
+ 1
1627
+ ],
1628
+ "scale_bits": 4
1629
+ },
1630
+ "down_proj": {
1631
+ "group_size": {
1632
+ "4": 128
1633
+ },
1634
+ "bits": [
1635
+ 4
1636
+ ],
1637
+ "bits_prop": [
1638
+ 1
1639
+ ],
1640
+ "scale_bits": 4
1641
+ },
1642
+ "qkv_proj": {
1643
+ "group_size": {
1644
+ "4": 128
1645
+ },
1646
+ "bits": [
1647
+ 4
1648
+ ],
1649
+ "bits_prop": [
1650
+ 1
1651
+ ],
1652
+ "scale_bits": 4
1653
+ },
1654
+ "gate_up_proj": {
1655
+ "group_size": {
1656
+ "4": 128
1657
+ },
1658
+ "bits": [
1659
+ 4
1660
+ ],
1661
+ "bits_prop": [
1662
+ 1
1663
+ ],
1664
+ "scale_bits": 4
1665
+ }
1666
+ }
1667
+ }
1668
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|/inst|>"
4
+ ],
5
+ "bos_token": {
6
+ "content": "<s>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false
11
+ },
12
+ "eos_token": {
13
+ "content": "<|endoftext|>",
14
+ "lstrip": false,
15
+ "normalized": false,
16
+ "rstrip": false,
17
+ "single_word": false
18
+ },
19
+ "pad_token": {
20
+ "content": "<|endoftext|>",
21
+ "lstrip": false,
22
+ "normalized": false,
23
+ "rstrip": false,
24
+ "single_word": false
25
+ },
26
+ "unk_token": {
27
+ "content": "<unk>",
28
+ "lstrip": false,
29
+ "normalized": false,
30
+ "rstrip": false,
31
+ "single_word": false
32
+ }
33
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
tokenizer_config.json ADDED
@@ -0,0 +1,351 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": true,
5
+ "added_tokens_decoder": {
6
+ "0": {
7
+ "content": "<unk>",
8
+ "lstrip": false,
9
+ "normalized": false,
10
+ "rstrip": false,
11
+ "single_word": false,
12
+ "special": true
13
+ },
14
+ "1": {
15
+ "content": "<s>",
16
+ "lstrip": false,
17
+ "normalized": false,
18
+ "rstrip": false,
19
+ "single_word": false,
20
+ "special": true
21
+ },
22
+ "2": {
23
+ "content": "</s>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": true,
27
+ "single_word": false,
28
+ "special": false
29
+ },
30
+ "32000": {
31
+ "content": "<|endoftext|>",
32
+ "lstrip": false,
33
+ "normalized": false,
34
+ "rstrip": false,
35
+ "single_word": false,
36
+ "special": true
37
+ },
38
+ "32001": {
39
+ "content": "<|assistant|>",
40
+ "lstrip": false,
41
+ "normalized": false,
42
+ "rstrip": true,
43
+ "single_word": false,
44
+ "special": true
45
+ },
46
+ "32002": {
47
+ "content": "<|step|>",
48
+ "lstrip": false,
49
+ "normalized": false,
50
+ "rstrip": true,
51
+ "single_word": false,
52
+ "special": true
53
+ },
54
+ "32003": {
55
+ "content": "<|function_output|>",
56
+ "lstrip": false,
57
+ "normalized": false,
58
+ "rstrip": true,
59
+ "single_word": false,
60
+ "special": true
61
+ },
62
+ "32004": {
63
+ "content": "<|tag|>",
64
+ "lstrip": false,
65
+ "normalized": false,
66
+ "rstrip": true,
67
+ "single_word": false,
68
+ "special": true
69
+ },
70
+ "32005": {
71
+ "content": "<|function_call|>",
72
+ "lstrip": false,
73
+ "normalized": false,
74
+ "rstrip": true,
75
+ "single_word": false,
76
+ "special": true
77
+ },
78
+ "32006": {
79
+ "content": "<|system|>",
80
+ "lstrip": false,
81
+ "normalized": false,
82
+ "rstrip": true,
83
+ "single_word": false,
84
+ "special": true
85
+ },
86
+ "32007": {
87
+ "content": "<|end|>",
88
+ "lstrip": false,
89
+ "normalized": false,
90
+ "rstrip": true,
91
+ "single_word": false,
92
+ "special": true
93
+ },
94
+ "32008": {
95
+ "content": "<|raw|>",
96
+ "lstrip": false,
97
+ "normalized": false,
98
+ "rstrip": true,
99
+ "single_word": false,
100
+ "special": true
101
+ },
102
+ "32009": {
103
+ "content": "<|continue|>",
104
+ "lstrip": false,
105
+ "normalized": false,
106
+ "rstrip": true,
107
+ "single_word": false,
108
+ "special": true
109
+ },
110
+ "32010": {
111
+ "content": "<|user|>",
112
+ "lstrip": false,
113
+ "normalized": false,
114
+ "rstrip": true,
115
+ "single_word": false,
116
+ "special": true
117
+ },
118
+ "32011": {
119
+ "content": "<|function_list|>",
120
+ "lstrip": false,
121
+ "normalized": false,
122
+ "rstrip": true,
123
+ "single_word": false,
124
+ "special": true
125
+ },
126
+ "32012": {
127
+ "content": "<|calc|>",
128
+ "lstrip": false,
129
+ "normalized": false,
130
+ "rstrip": true,
131
+ "single_word": false,
132
+ "special": true
133
+ },
134
+ "32013": {
135
+ "content": "<|code|>",
136
+ "lstrip": false,
137
+ "normalized": false,
138
+ "rstrip": true,
139
+ "single_word": false,
140
+ "special": true
141
+ },
142
+ "32014": {
143
+ "content": "<|/code|>",
144
+ "lstrip": false,
145
+ "normalized": false,
146
+ "rstrip": true,
147
+ "single_word": false,
148
+ "special": true
149
+ },
150
+ "32015": {
151
+ "content": "<|summary|>",
152
+ "lstrip": false,
153
+ "normalized": false,
154
+ "rstrip": true,
155
+ "single_word": false,
156
+ "special": true
157
+ },
158
+ "32016": {
159
+ "content": "<|resource|>",
160
+ "lstrip": false,
161
+ "normalized": false,
162
+ "rstrip": true,
163
+ "single_word": false,
164
+ "special": true
165
+ },
166
+ "32017": {
167
+ "content": "<|assistant_mask|>",
168
+ "lstrip": false,
169
+ "normalized": false,
170
+ "rstrip": true,
171
+ "single_word": false,
172
+ "special": true
173
+ },
174
+ "32018": {
175
+ "content": "<|start|>",
176
+ "lstrip": false,
177
+ "normalized": false,
178
+ "rstrip": true,
179
+ "single_word": false,
180
+ "special": true
181
+ },
182
+ "32019": {
183
+ "content": "<|message|>",
184
+ "lstrip": false,
185
+ "normalized": false,
186
+ "rstrip": true,
187
+ "single_word": false,
188
+ "special": true
189
+ },
190
+ "32020": {
191
+ "content": "<|fim_prefix|>",
192
+ "lstrip": false,
193
+ "normalized": false,
194
+ "rstrip": true,
195
+ "single_word": false,
196
+ "special": true
197
+ },
198
+ "32021": {
199
+ "content": "<|fim_middle|>",
200
+ "lstrip": false,
201
+ "normalized": false,
202
+ "rstrip": true,
203
+ "single_word": false,
204
+ "special": true
205
+ },
206
+ "32022": {
207
+ "content": "<|fim_suffix|>",
208
+ "lstrip": false,
209
+ "normalized": false,
210
+ "rstrip": true,
211
+ "single_word": false,
212
+ "special": true
213
+ },
214
+ "32023": {
215
+ "content": "<|meta_start|>",
216
+ "lstrip": false,
217
+ "normalized": false,
218
+ "rstrip": true,
219
+ "single_word": false,
220
+ "special": true
221
+ },
222
+ "32024": {
223
+ "content": "<|ipynb_marker|>",
224
+ "lstrip": false,
225
+ "normalized": false,
226
+ "rstrip": true,
227
+ "single_word": false,
228
+ "special": true
229
+ },
230
+ "32025": {
231
+ "content": "<|diff_marker|>",
232
+ "lstrip": false,
233
+ "normalized": false,
234
+ "rstrip": true,
235
+ "single_word": false,
236
+ "special": true
237
+ },
238
+ "32026": {
239
+ "content": "<|ghissue|>",
240
+ "lstrip": false,
241
+ "normalized": false,
242
+ "rstrip": true,
243
+ "single_word": false,
244
+ "special": true
245
+ },
246
+ "32027": {
247
+ "content": "<|ghreview|>",
248
+ "lstrip": false,
249
+ "normalized": false,
250
+ "rstrip": true,
251
+ "single_word": false,
252
+ "special": true
253
+ },
254
+ "32028": {
255
+ "content": "<|disc_start|>",
256
+ "lstrip": false,
257
+ "normalized": false,
258
+ "rstrip": true,
259
+ "single_word": false,
260
+ "special": true
261
+ },
262
+ "32029": {
263
+ "content": "<|disc_sep|>",
264
+ "lstrip": false,
265
+ "normalized": false,
266
+ "rstrip": true,
267
+ "single_word": false,
268
+ "special": true
269
+ },
270
+ "32030": {
271
+ "content": "<|disc_thread|><|query|>",
272
+ "lstrip": false,
273
+ "normalized": false,
274
+ "rstrip": true,
275
+ "single_word": false,
276
+ "special": true
277
+ },
278
+ "32031": {
279
+ "content": "<|/query|>",
280
+ "lstrip": false,
281
+ "normalized": false,
282
+ "rstrip": true,
283
+ "single_word": false,
284
+ "special": true
285
+ },
286
+ "32032": {
287
+ "content": "<|data|>",
288
+ "lstrip": false,
289
+ "normalized": false,
290
+ "rstrip": true,
291
+ "single_word": false,
292
+ "special": true
293
+ },
294
+ "32033": {
295
+ "content": "<|/data|>",
296
+ "lstrip": false,
297
+ "normalized": false,
298
+ "rstrip": true,
299
+ "single_word": false,
300
+ "special": true
301
+ },
302
+ "32034": {
303
+ "content": "<|sys|>",
304
+ "lstrip": false,
305
+ "normalized": false,
306
+ "rstrip": true,
307
+ "single_word": false,
308
+ "special": true
309
+ },
310
+ "32035": {
311
+ "content": "<|/sys|>",
312
+ "lstrip": false,
313
+ "normalized": false,
314
+ "rstrip": true,
315
+ "single_word": false,
316
+ "special": true
317
+ },
318
+ "32036": {
319
+ "content": "<|inst|>",
320
+ "lstrip": false,
321
+ "normalized": false,
322
+ "rstrip": true,
323
+ "single_word": false,
324
+ "special": true
325
+ },
326
+ "32037": {
327
+ "content": "<|/inst|>",
328
+ "lstrip": false,
329
+ "normalized": false,
330
+ "rstrip": true,
331
+ "single_word": false,
332
+ "special": true
333
+ }
334
+ },
335
+ "additional_special_tokens": [
336
+ "<|/inst|>"
337
+ ],
338
+ "bos_token": "<s>",
339
+ "chat_template": "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
340
+ "clean_up_tokenization_spaces": false,
341
+ "eos_token": "<|endoftext|>",
342
+ "legacy": false,
343
+ "model_max_length": 131072,
344
+ "pad_token": "<|endoftext|>",
345
+ "padding_side": "left",
346
+ "sp_model_kwargs": {},
347
+ "spaces_between_special_tokens": false,
348
+ "tokenizer_class": "LlamaTokenizer",
349
+ "unk_token": "<unk>",
350
+ "use_default_system_prompt": false
351
+ }