NicoNico commited on
Commit
e7c7b0d
1 Parent(s): 6642376

Upload folder using huggingface_hub

Browse files
README.md ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - mlx
5
+ ---
6
+
7
+ # GreenBitAI/Mistral-7B-v0.1-layer-mix-bpw-2.5-mlx
8
+ This quantized low-bit model was converted to MLX format from [`GreenBitAI/Mistral-7B-v0.1-layer-mix-bpw-2.5`]().
9
+ Refer to the [original model card](https://huggingface.co/GreenBitAI/Mistral-7B-v0.1-layer-mix-bpw-2.5) 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/Mistral-7B-v0.1-layer-mix-bpw-2.5-mlx")
20
+ response = generate(model, tokenizer, prompt="hello", verbose=True)
21
+ ```
config.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "MistralForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "bad_words_ids": null,
8
+ "begin_suppress_tokens": null,
9
+ "bos_token_id": 1,
10
+ "chunk_size_feed_forward": 0,
11
+ "cross_attention_hidden_size": null,
12
+ "decoder_start_token_id": null,
13
+ "diversity_penalty": 0.0,
14
+ "do_sample": false,
15
+ "early_stopping": false,
16
+ "encoder_no_repeat_ngram_size": 0,
17
+ "eos_token_id": 2,
18
+ "exponential_decay_length_penalty": null,
19
+ "finetuning_task": null,
20
+ "forced_bos_token_id": null,
21
+ "forced_eos_token_id": null,
22
+ "hidden_act": "silu",
23
+ "hidden_size": 4096,
24
+ "id2label": {
25
+ "0": "LABEL_0",
26
+ "1": "LABEL_1"
27
+ },
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 14336,
30
+ "is_decoder": false,
31
+ "is_encoder_decoder": false,
32
+ "label2id": {
33
+ "LABEL_0": 0,
34
+ "LABEL_1": 1
35
+ },
36
+ "length_penalty": 1.0,
37
+ "max_length": 20,
38
+ "max_position_embeddings": 32768,
39
+ "min_length": 0,
40
+ "model_type": "mistral",
41
+ "no_repeat_ngram_size": 0,
42
+ "num_attention_heads": 32,
43
+ "num_beam_groups": 1,
44
+ "num_beams": 1,
45
+ "num_hidden_layers": 32,
46
+ "num_key_value_heads": 8,
47
+ "num_return_sequences": 1,
48
+ "output_attentions": false,
49
+ "output_hidden_states": false,
50
+ "output_scores": false,
51
+ "pad_token_id": null,
52
+ "prefix": null,
53
+ "problem_type": null,
54
+ "pruned_heads": {},
55
+ "quantization": {
56
+ "group_size": null,
57
+ "bits": null
58
+ },
59
+ "remove_invalid_values": false,
60
+ "repetition_penalty": 1.0,
61
+ "return_dict": true,
62
+ "return_dict_in_generate": false,
63
+ "rms_norm_eps": 1e-05,
64
+ "rope_theta": 10000.0,
65
+ "sep_token_id": null,
66
+ "sliding_window": 4096,
67
+ "suppress_tokens": null,
68
+ "task_specific_params": null,
69
+ "temperature": 1.0,
70
+ "tf_legacy_loss": false,
71
+ "tie_encoder_decoder": false,
72
+ "tie_word_embeddings": false,
73
+ "tokenizer_class": null,
74
+ "top_k": 50,
75
+ "top_p": 1.0,
76
+ "torch_dtype": "float16",
77
+ "torchscript": false,
78
+ "transformers_version": "4.39.3",
79
+ "typical_p": 1.0,
80
+ "use_bfloat16": false,
81
+ "use_cache": true,
82
+ "vocab_size": 32000
83
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22f8fd9b38e3a4a4bac0810d599055b4ed9aec38ffadaa0d525951a74359e152
3
+ size 3048616749
model.safetensors.index.json ADDED
@@ -0,0 +1,1194 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 3048480768
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_proj.channel_scale": "model.safetensors",
15
+ "model.layers.0.mlp.gate_proj.q_perm": "model.safetensors",
16
+ "model.layers.0.mlp.gate_proj.qweight": "model.safetensors",
17
+ "model.layers.0.mlp.gate_proj.scales": "model.safetensors",
18
+ "model.layers.0.mlp.gate_proj.zeros": "model.safetensors",
19
+ "model.layers.0.mlp.up_proj.channel_scale": "model.safetensors",
20
+ "model.layers.0.mlp.up_proj.q_perm": "model.safetensors",
21
+ "model.layers.0.mlp.up_proj.qweight": "model.safetensors",
22
+ "model.layers.0.mlp.up_proj.scales": "model.safetensors",
23
+ "model.layers.0.mlp.up_proj.zeros": "model.safetensors",
24
+ "model.layers.0.post_attention_layernorm.weight": "model.safetensors",
25
+ "model.layers.0.self_attn.k_proj.channel_scale": "model.safetensors",
26
+ "model.layers.0.self_attn.k_proj.q_perm": "model.safetensors",
27
+ "model.layers.0.self_attn.k_proj.qweight": "model.safetensors",
28
+ "model.layers.0.self_attn.k_proj.scales": "model.safetensors",
29
+ "model.layers.0.self_attn.k_proj.zeros": "model.safetensors",
30
+ "model.layers.0.self_attn.o_proj.channel_scale": "model.safetensors",
31
+ "model.layers.0.self_attn.o_proj.q_perm": "model.safetensors",
32
+ "model.layers.0.self_attn.o_proj.qweight": "model.safetensors",
33
+ "model.layers.0.self_attn.o_proj.scales": "model.safetensors",
34
+ "model.layers.0.self_attn.o_proj.zeros": "model.safetensors",
35
+ "model.layers.0.self_attn.q_proj.channel_scale": "model.safetensors",
36
+ "model.layers.0.self_attn.q_proj.q_perm": "model.safetensors",
37
+ "model.layers.0.self_attn.q_proj.qweight": "model.safetensors",
38
+ "model.layers.0.self_attn.q_proj.scales": "model.safetensors",
39
+ "model.layers.0.self_attn.q_proj.zeros": "model.safetensors",
40
+ "model.layers.0.self_attn.v_proj.channel_scale": "model.safetensors",
41
+ "model.layers.0.self_attn.v_proj.q_perm": "model.safetensors",
42
+ "model.layers.0.self_attn.v_proj.qweight": "model.safetensors",
43
+ "model.layers.0.self_attn.v_proj.scales": "model.safetensors",
44
+ "model.layers.0.self_attn.v_proj.zeros": "model.safetensors",
45
+ "model.layers.1.input_layernorm.weight": "model.safetensors",
46
+ "model.layers.1.mlp.down_proj.channel_scale": "model.safetensors",
47
+ "model.layers.1.mlp.down_proj.q_perm": "model.safetensors",
48
+ "model.layers.1.mlp.down_proj.qweight": "model.safetensors",
49
+ "model.layers.1.mlp.down_proj.scales": "model.safetensors",
50
+ "model.layers.1.mlp.down_proj.zeros": "model.safetensors",
51
+ "model.layers.1.mlp.gate_proj.channel_scale": "model.safetensors",
52
+ "model.layers.1.mlp.gate_proj.q_perm": "model.safetensors",
53
+ "model.layers.1.mlp.gate_proj.qweight": "model.safetensors",
54
+ "model.layers.1.mlp.gate_proj.scales": "model.safetensors",
55
+ "model.layers.1.mlp.gate_proj.zeros": "model.safetensors",
56
+ "model.layers.1.mlp.up_proj.channel_scale": "model.safetensors",
57
+ "model.layers.1.mlp.up_proj.q_perm": "model.safetensors",
58
+ "model.layers.1.mlp.up_proj.qweight": "model.safetensors",
59
+ "model.layers.1.mlp.up_proj.scales": "model.safetensors",
60
+ "model.layers.1.mlp.up_proj.zeros": "model.safetensors",
61
+ "model.layers.1.post_attention_layernorm.weight": "model.safetensors",
62
+ "model.layers.1.self_attn.k_proj.channel_scale": "model.safetensors",
63
+ "model.layers.1.self_attn.k_proj.q_perm": "model.safetensors",
64
+ "model.layers.1.self_attn.k_proj.qweight": "model.safetensors",
65
+ "model.layers.1.self_attn.k_proj.scales": "model.safetensors",
66
+ "model.layers.1.self_attn.k_proj.zeros": "model.safetensors",
67
+ "model.layers.1.self_attn.o_proj.channel_scale": "model.safetensors",
68
+ "model.layers.1.self_attn.o_proj.q_perm": "model.safetensors",
69
+ "model.layers.1.self_attn.o_proj.qweight": "model.safetensors",
70
+ "model.layers.1.self_attn.o_proj.scales": "model.safetensors",
71
+ "model.layers.1.self_attn.o_proj.zeros": "model.safetensors",
72
+ "model.layers.1.self_attn.q_proj.channel_scale": "model.safetensors",
73
+ "model.layers.1.self_attn.q_proj.q_perm": "model.safetensors",
74
+ "model.layers.1.self_attn.q_proj.qweight": "model.safetensors",
75
+ "model.layers.1.self_attn.q_proj.scales": "model.safetensors",
76
+ "model.layers.1.self_attn.q_proj.zeros": "model.safetensors",
77
+ "model.layers.1.self_attn.v_proj.channel_scale": "model.safetensors",
78
+ "model.layers.1.self_attn.v_proj.q_perm": "model.safetensors",
79
+ "model.layers.1.self_attn.v_proj.qweight": "model.safetensors",
80
+ "model.layers.1.self_attn.v_proj.scales": "model.safetensors",
81
+ "model.layers.1.self_attn.v_proj.zeros": "model.safetensors",
82
+ "model.layers.10.input_layernorm.weight": "model.safetensors",
83
+ "model.layers.10.mlp.down_proj.channel_scale": "model.safetensors",
84
+ "model.layers.10.mlp.down_proj.q_perm": "model.safetensors",
85
+ "model.layers.10.mlp.down_proj.qweight": "model.safetensors",
86
+ "model.layers.10.mlp.down_proj.scales": "model.safetensors",
87
+ "model.layers.10.mlp.down_proj.zeros": "model.safetensors",
88
+ "model.layers.10.mlp.gate_proj.channel_scale": "model.safetensors",
89
+ "model.layers.10.mlp.gate_proj.q_perm": "model.safetensors",
90
+ "model.layers.10.mlp.gate_proj.qweight": "model.safetensors",
91
+ "model.layers.10.mlp.gate_proj.scales": "model.safetensors",
92
+ "model.layers.10.mlp.gate_proj.zeros": "model.safetensors",
93
+ "model.layers.10.mlp.up_proj.channel_scale": "model.safetensors",
94
+ "model.layers.10.mlp.up_proj.q_perm": "model.safetensors",
95
+ "model.layers.10.mlp.up_proj.qweight": "model.safetensors",
96
+ "model.layers.10.mlp.up_proj.scales": "model.safetensors",
97
+ "model.layers.10.mlp.up_proj.zeros": "model.safetensors",
98
+ "model.layers.10.post_attention_layernorm.weight": "model.safetensors",
99
+ "model.layers.10.self_attn.k_proj.channel_scale": "model.safetensors",
100
+ "model.layers.10.self_attn.k_proj.q_perm": "model.safetensors",
101
+ "model.layers.10.self_attn.k_proj.qweight": "model.safetensors",
102
+ "model.layers.10.self_attn.k_proj.scales": "model.safetensors",
103
+ "model.layers.10.self_attn.k_proj.zeros": "model.safetensors",
104
+ "model.layers.10.self_attn.o_proj.channel_scale": "model.safetensors",
105
+ "model.layers.10.self_attn.o_proj.q_perm": "model.safetensors",
106
+ "model.layers.10.self_attn.o_proj.qweight": "model.safetensors",
107
+ "model.layers.10.self_attn.o_proj.scales": "model.safetensors",
108
+ "model.layers.10.self_attn.o_proj.zeros": "model.safetensors",
109
+ "model.layers.10.self_attn.q_proj.channel_scale": "model.safetensors",
110
+ "model.layers.10.self_attn.q_proj.q_perm": "model.safetensors",
111
+ "model.layers.10.self_attn.q_proj.qweight": "model.safetensors",
112
+ "model.layers.10.self_attn.q_proj.scales": "model.safetensors",
113
+ "model.layers.10.self_attn.q_proj.zeros": "model.safetensors",
114
+ "model.layers.10.self_attn.v_proj.channel_scale": "model.safetensors",
115
+ "model.layers.10.self_attn.v_proj.q_perm": "model.safetensors",
116
+ "model.layers.10.self_attn.v_proj.qweight": "model.safetensors",
117
+ "model.layers.10.self_attn.v_proj.scales": "model.safetensors",
118
+ "model.layers.10.self_attn.v_proj.zeros": "model.safetensors",
119
+ "model.layers.11.input_layernorm.weight": "model.safetensors",
120
+ "model.layers.11.mlp.down_proj.channel_scale": "model.safetensors",
121
+ "model.layers.11.mlp.down_proj.q_perm": "model.safetensors",
122
+ "model.layers.11.mlp.down_proj.qweight": "model.safetensors",
123
+ "model.layers.11.mlp.down_proj.scales": "model.safetensors",
124
+ "model.layers.11.mlp.down_proj.zeros": "model.safetensors",
125
+ "model.layers.11.mlp.gate_proj.channel_scale": "model.safetensors",
126
+ "model.layers.11.mlp.gate_proj.q_perm": "model.safetensors",
127
+ "model.layers.11.mlp.gate_proj.qweight": "model.safetensors",
128
+ "model.layers.11.mlp.gate_proj.scales": "model.safetensors",
129
+ "model.layers.11.mlp.gate_proj.zeros": "model.safetensors",
130
+ "model.layers.11.mlp.up_proj.channel_scale": "model.safetensors",
131
+ "model.layers.11.mlp.up_proj.q_perm": "model.safetensors",
132
+ "model.layers.11.mlp.up_proj.qweight": "model.safetensors",
133
+ "model.layers.11.mlp.up_proj.scales": "model.safetensors",
134
+ "model.layers.11.mlp.up_proj.zeros": "model.safetensors",
135
+ "model.layers.11.post_attention_layernorm.weight": "model.safetensors",
136
+ "model.layers.11.self_attn.k_proj.channel_scale": "model.safetensors",
137
+ "model.layers.11.self_attn.k_proj.q_perm": "model.safetensors",
138
+ "model.layers.11.self_attn.k_proj.qweight": "model.safetensors",
139
+ "model.layers.11.self_attn.k_proj.scales": "model.safetensors",
140
+ "model.layers.11.self_attn.k_proj.zeros": "model.safetensors",
141
+ "model.layers.11.self_attn.o_proj.channel_scale": "model.safetensors",
142
+ "model.layers.11.self_attn.o_proj.q_perm": "model.safetensors",
143
+ "model.layers.11.self_attn.o_proj.qweight": "model.safetensors",
144
+ "model.layers.11.self_attn.o_proj.scales": "model.safetensors",
145
+ "model.layers.11.self_attn.o_proj.zeros": "model.safetensors",
146
+ "model.layers.11.self_attn.q_proj.channel_scale": "model.safetensors",
147
+ "model.layers.11.self_attn.q_proj.q_perm": "model.safetensors",
148
+ "model.layers.11.self_attn.q_proj.qweight": "model.safetensors",
149
+ "model.layers.11.self_attn.q_proj.scales": "model.safetensors",
150
+ "model.layers.11.self_attn.q_proj.zeros": "model.safetensors",
151
+ "model.layers.11.self_attn.v_proj.channel_scale": "model.safetensors",
152
+ "model.layers.11.self_attn.v_proj.q_perm": "model.safetensors",
153
+ "model.layers.11.self_attn.v_proj.qweight": "model.safetensors",
154
+ "model.layers.11.self_attn.v_proj.scales": "model.safetensors",
155
+ "model.layers.11.self_attn.v_proj.zeros": "model.safetensors",
156
+ "model.layers.12.input_layernorm.weight": "model.safetensors",
157
+ "model.layers.12.mlp.down_proj.channel_scale": "model.safetensors",
158
+ "model.layers.12.mlp.down_proj.q_perm": "model.safetensors",
159
+ "model.layers.12.mlp.down_proj.qweight": "model.safetensors",
160
+ "model.layers.12.mlp.down_proj.scales": "model.safetensors",
161
+ "model.layers.12.mlp.down_proj.zeros": "model.safetensors",
162
+ "model.layers.12.mlp.gate_proj.channel_scale": "model.safetensors",
163
+ "model.layers.12.mlp.gate_proj.q_perm": "model.safetensors",
164
+ "model.layers.12.mlp.gate_proj.qweight": "model.safetensors",
165
+ "model.layers.12.mlp.gate_proj.scales": "model.safetensors",
166
+ "model.layers.12.mlp.gate_proj.zeros": "model.safetensors",
167
+ "model.layers.12.mlp.up_proj.channel_scale": "model.safetensors",
168
+ "model.layers.12.mlp.up_proj.q_perm": "model.safetensors",
169
+ "model.layers.12.mlp.up_proj.qweight": "model.safetensors",
170
+ "model.layers.12.mlp.up_proj.scales": "model.safetensors",
171
+ "model.layers.12.mlp.up_proj.zeros": "model.safetensors",
172
+ "model.layers.12.post_attention_layernorm.weight": "model.safetensors",
173
+ "model.layers.12.self_attn.k_proj.channel_scale": "model.safetensors",
174
+ "model.layers.12.self_attn.k_proj.q_perm": "model.safetensors",
175
+ "model.layers.12.self_attn.k_proj.qweight": "model.safetensors",
176
+ "model.layers.12.self_attn.k_proj.scales": "model.safetensors",
177
+ "model.layers.12.self_attn.k_proj.zeros": "model.safetensors",
178
+ "model.layers.12.self_attn.o_proj.channel_scale": "model.safetensors",
179
+ "model.layers.12.self_attn.o_proj.q_perm": "model.safetensors",
180
+ "model.layers.12.self_attn.o_proj.qweight": "model.safetensors",
181
+ "model.layers.12.self_attn.o_proj.scales": "model.safetensors",
182
+ "model.layers.12.self_attn.o_proj.zeros": "model.safetensors",
183
+ "model.layers.12.self_attn.q_proj.channel_scale": "model.safetensors",
184
+ "model.layers.12.self_attn.q_proj.q_perm": "model.safetensors",
185
+ "model.layers.12.self_attn.q_proj.qweight": "model.safetensors",
186
+ "model.layers.12.self_attn.q_proj.scales": "model.safetensors",
187
+ "model.layers.12.self_attn.q_proj.zeros": "model.safetensors",
188
+ "model.layers.12.self_attn.v_proj.channel_scale": "model.safetensors",
189
+ "model.layers.12.self_attn.v_proj.q_perm": "model.safetensors",
190
+ "model.layers.12.self_attn.v_proj.qweight": "model.safetensors",
191
+ "model.layers.12.self_attn.v_proj.scales": "model.safetensors",
192
+ "model.layers.12.self_attn.v_proj.zeros": "model.safetensors",
193
+ "model.layers.13.input_layernorm.weight": "model.safetensors",
194
+ "model.layers.13.mlp.down_proj.channel_scale": "model.safetensors",
195
+ "model.layers.13.mlp.down_proj.q_perm": "model.safetensors",
196
+ "model.layers.13.mlp.down_proj.qweight": "model.safetensors",
197
+ "model.layers.13.mlp.down_proj.scales": "model.safetensors",
198
+ "model.layers.13.mlp.down_proj.zeros": "model.safetensors",
199
+ "model.layers.13.mlp.gate_proj.channel_scale": "model.safetensors",
200
+ "model.layers.13.mlp.gate_proj.q_perm": "model.safetensors",
201
+ "model.layers.13.mlp.gate_proj.qweight": "model.safetensors",
202
+ "model.layers.13.mlp.gate_proj.scales": "model.safetensors",
203
+ "model.layers.13.mlp.gate_proj.zeros": "model.safetensors",
204
+ "model.layers.13.mlp.up_proj.channel_scale": "model.safetensors",
205
+ "model.layers.13.mlp.up_proj.q_perm": "model.safetensors",
206
+ "model.layers.13.mlp.up_proj.qweight": "model.safetensors",
207
+ "model.layers.13.mlp.up_proj.scales": "model.safetensors",
208
+ "model.layers.13.mlp.up_proj.zeros": "model.safetensors",
209
+ "model.layers.13.post_attention_layernorm.weight": "model.safetensors",
210
+ "model.layers.13.self_attn.k_proj.channel_scale": "model.safetensors",
211
+ "model.layers.13.self_attn.k_proj.q_perm": "model.safetensors",
212
+ "model.layers.13.self_attn.k_proj.qweight": "model.safetensors",
213
+ "model.layers.13.self_attn.k_proj.scales": "model.safetensors",
214
+ "model.layers.13.self_attn.k_proj.zeros": "model.safetensors",
215
+ "model.layers.13.self_attn.o_proj.channel_scale": "model.safetensors",
216
+ "model.layers.13.self_attn.o_proj.q_perm": "model.safetensors",
217
+ "model.layers.13.self_attn.o_proj.qweight": "model.safetensors",
218
+ "model.layers.13.self_attn.o_proj.scales": "model.safetensors",
219
+ "model.layers.13.self_attn.o_proj.zeros": "model.safetensors",
220
+ "model.layers.13.self_attn.q_proj.channel_scale": "model.safetensors",
221
+ "model.layers.13.self_attn.q_proj.q_perm": "model.safetensors",
222
+ "model.layers.13.self_attn.q_proj.qweight": "model.safetensors",
223
+ "model.layers.13.self_attn.q_proj.scales": "model.safetensors",
224
+ "model.layers.13.self_attn.q_proj.zeros": "model.safetensors",
225
+ "model.layers.13.self_attn.v_proj.channel_scale": "model.safetensors",
226
+ "model.layers.13.self_attn.v_proj.q_perm": "model.safetensors",
227
+ "model.layers.13.self_attn.v_proj.qweight": "model.safetensors",
228
+ "model.layers.13.self_attn.v_proj.scales": "model.safetensors",
229
+ "model.layers.13.self_attn.v_proj.zeros": "model.safetensors",
230
+ "model.layers.14.input_layernorm.weight": "model.safetensors",
231
+ "model.layers.14.mlp.down_proj.channel_scale": "model.safetensors",
232
+ "model.layers.14.mlp.down_proj.q_perm": "model.safetensors",
233
+ "model.layers.14.mlp.down_proj.qweight": "model.safetensors",
234
+ "model.layers.14.mlp.down_proj.scales": "model.safetensors",
235
+ "model.layers.14.mlp.down_proj.zeros": "model.safetensors",
236
+ "model.layers.14.mlp.gate_proj.channel_scale": "model.safetensors",
237
+ "model.layers.14.mlp.gate_proj.q_perm": "model.safetensors",
238
+ "model.layers.14.mlp.gate_proj.qweight": "model.safetensors",
239
+ "model.layers.14.mlp.gate_proj.scales": "model.safetensors",
240
+ "model.layers.14.mlp.gate_proj.zeros": "model.safetensors",
241
+ "model.layers.14.mlp.up_proj.channel_scale": "model.safetensors",
242
+ "model.layers.14.mlp.up_proj.q_perm": "model.safetensors",
243
+ "model.layers.14.mlp.up_proj.qweight": "model.safetensors",
244
+ "model.layers.14.mlp.up_proj.scales": "model.safetensors",
245
+ "model.layers.14.mlp.up_proj.zeros": "model.safetensors",
246
+ "model.layers.14.post_attention_layernorm.weight": "model.safetensors",
247
+ "model.layers.14.self_attn.k_proj.channel_scale": "model.safetensors",
248
+ "model.layers.14.self_attn.k_proj.q_perm": "model.safetensors",
249
+ "model.layers.14.self_attn.k_proj.qweight": "model.safetensors",
250
+ "model.layers.14.self_attn.k_proj.scales": "model.safetensors",
251
+ "model.layers.14.self_attn.k_proj.zeros": "model.safetensors",
252
+ "model.layers.14.self_attn.o_proj.channel_scale": "model.safetensors",
253
+ "model.layers.14.self_attn.o_proj.q_perm": "model.safetensors",
254
+ "model.layers.14.self_attn.o_proj.qweight": "model.safetensors",
255
+ "model.layers.14.self_attn.o_proj.scales": "model.safetensors",
256
+ "model.layers.14.self_attn.o_proj.zeros": "model.safetensors",
257
+ "model.layers.14.self_attn.q_proj.channel_scale": "model.safetensors",
258
+ "model.layers.14.self_attn.q_proj.q_perm": "model.safetensors",
259
+ "model.layers.14.self_attn.q_proj.qweight": "model.safetensors",
260
+ "model.layers.14.self_attn.q_proj.scales": "model.safetensors",
261
+ "model.layers.14.self_attn.q_proj.zeros": "model.safetensors",
262
+ "model.layers.14.self_attn.v_proj.channel_scale": "model.safetensors",
263
+ "model.layers.14.self_attn.v_proj.q_perm": "model.safetensors",
264
+ "model.layers.14.self_attn.v_proj.qweight": "model.safetensors",
265
+ "model.layers.14.self_attn.v_proj.scales": "model.safetensors",
266
+ "model.layers.14.self_attn.v_proj.zeros": "model.safetensors",
267
+ "model.layers.15.input_layernorm.weight": "model.safetensors",
268
+ "model.layers.15.mlp.down_proj.channel_scale": "model.safetensors",
269
+ "model.layers.15.mlp.down_proj.q_perm": "model.safetensors",
270
+ "model.layers.15.mlp.down_proj.qweight": "model.safetensors",
271
+ "model.layers.15.mlp.down_proj.scales": "model.safetensors",
272
+ "model.layers.15.mlp.down_proj.zeros": "model.safetensors",
273
+ "model.layers.15.mlp.gate_proj.channel_scale": "model.safetensors",
274
+ "model.layers.15.mlp.gate_proj.q_perm": "model.safetensors",
275
+ "model.layers.15.mlp.gate_proj.qweight": "model.safetensors",
276
+ "model.layers.15.mlp.gate_proj.scales": "model.safetensors",
277
+ "model.layers.15.mlp.gate_proj.zeros": "model.safetensors",
278
+ "model.layers.15.mlp.up_proj.channel_scale": "model.safetensors",
279
+ "model.layers.15.mlp.up_proj.q_perm": "model.safetensors",
280
+ "model.layers.15.mlp.up_proj.qweight": "model.safetensors",
281
+ "model.layers.15.mlp.up_proj.scales": "model.safetensors",
282
+ "model.layers.15.mlp.up_proj.zeros": "model.safetensors",
283
+ "model.layers.15.post_attention_layernorm.weight": "model.safetensors",
284
+ "model.layers.15.self_attn.k_proj.channel_scale": "model.safetensors",
285
+ "model.layers.15.self_attn.k_proj.q_perm": "model.safetensors",
286
+ "model.layers.15.self_attn.k_proj.qweight": "model.safetensors",
287
+ "model.layers.15.self_attn.k_proj.scales": "model.safetensors",
288
+ "model.layers.15.self_attn.k_proj.zeros": "model.safetensors",
289
+ "model.layers.15.self_attn.o_proj.channel_scale": "model.safetensors",
290
+ "model.layers.15.self_attn.o_proj.q_perm": "model.safetensors",
291
+ "model.layers.15.self_attn.o_proj.qweight": "model.safetensors",
292
+ "model.layers.15.self_attn.o_proj.scales": "model.safetensors",
293
+ "model.layers.15.self_attn.o_proj.zeros": "model.safetensors",
294
+ "model.layers.15.self_attn.q_proj.channel_scale": "model.safetensors",
295
+ "model.layers.15.self_attn.q_proj.q_perm": "model.safetensors",
296
+ "model.layers.15.self_attn.q_proj.qweight": "model.safetensors",
297
+ "model.layers.15.self_attn.q_proj.scales": "model.safetensors",
298
+ "model.layers.15.self_attn.q_proj.zeros": "model.safetensors",
299
+ "model.layers.15.self_attn.v_proj.channel_scale": "model.safetensors",
300
+ "model.layers.15.self_attn.v_proj.q_perm": "model.safetensors",
301
+ "model.layers.15.self_attn.v_proj.qweight": "model.safetensors",
302
+ "model.layers.15.self_attn.v_proj.scales": "model.safetensors",
303
+ "model.layers.15.self_attn.v_proj.zeros": "model.safetensors",
304
+ "model.layers.16.input_layernorm.weight": "model.safetensors",
305
+ "model.layers.16.mlp.down_proj.channel_scale": "model.safetensors",
306
+ "model.layers.16.mlp.down_proj.q_perm": "model.safetensors",
307
+ "model.layers.16.mlp.down_proj.qweight": "model.safetensors",
308
+ "model.layers.16.mlp.down_proj.scales": "model.safetensors",
309
+ "model.layers.16.mlp.down_proj.zeros": "model.safetensors",
310
+ "model.layers.16.mlp.gate_proj.channel_scale": "model.safetensors",
311
+ "model.layers.16.mlp.gate_proj.q_perm": "model.safetensors",
312
+ "model.layers.16.mlp.gate_proj.qweight": "model.safetensors",
313
+ "model.layers.16.mlp.gate_proj.scales": "model.safetensors",
314
+ "model.layers.16.mlp.gate_proj.zeros": "model.safetensors",
315
+ "model.layers.16.mlp.up_proj.channel_scale": "model.safetensors",
316
+ "model.layers.16.mlp.up_proj.q_perm": "model.safetensors",
317
+ "model.layers.16.mlp.up_proj.qweight": "model.safetensors",
318
+ "model.layers.16.mlp.up_proj.scales": "model.safetensors",
319
+ "model.layers.16.mlp.up_proj.zeros": "model.safetensors",
320
+ "model.layers.16.post_attention_layernorm.weight": "model.safetensors",
321
+ "model.layers.16.self_attn.k_proj.channel_scale": "model.safetensors",
322
+ "model.layers.16.self_attn.k_proj.q_perm": "model.safetensors",
323
+ "model.layers.16.self_attn.k_proj.qweight": "model.safetensors",
324
+ "model.layers.16.self_attn.k_proj.scales": "model.safetensors",
325
+ "model.layers.16.self_attn.k_proj.zeros": "model.safetensors",
326
+ "model.layers.16.self_attn.o_proj.channel_scale": "model.safetensors",
327
+ "model.layers.16.self_attn.o_proj.q_perm": "model.safetensors",
328
+ "model.layers.16.self_attn.o_proj.qweight": "model.safetensors",
329
+ "model.layers.16.self_attn.o_proj.scales": "model.safetensors",
330
+ "model.layers.16.self_attn.o_proj.zeros": "model.safetensors",
331
+ "model.layers.16.self_attn.q_proj.channel_scale": "model.safetensors",
332
+ "model.layers.16.self_attn.q_proj.q_perm": "model.safetensors",
333
+ "model.layers.16.self_attn.q_proj.qweight": "model.safetensors",
334
+ "model.layers.16.self_attn.q_proj.scales": "model.safetensors",
335
+ "model.layers.16.self_attn.q_proj.zeros": "model.safetensors",
336
+ "model.layers.16.self_attn.v_proj.channel_scale": "model.safetensors",
337
+ "model.layers.16.self_attn.v_proj.q_perm": "model.safetensors",
338
+ "model.layers.16.self_attn.v_proj.qweight": "model.safetensors",
339
+ "model.layers.16.self_attn.v_proj.scales": "model.safetensors",
340
+ "model.layers.16.self_attn.v_proj.zeros": "model.safetensors",
341
+ "model.layers.17.input_layernorm.weight": "model.safetensors",
342
+ "model.layers.17.mlp.down_proj.channel_scale": "model.safetensors",
343
+ "model.layers.17.mlp.down_proj.q_perm": "model.safetensors",
344
+ "model.layers.17.mlp.down_proj.qweight": "model.safetensors",
345
+ "model.layers.17.mlp.down_proj.scales": "model.safetensors",
346
+ "model.layers.17.mlp.down_proj.zeros": "model.safetensors",
347
+ "model.layers.17.mlp.gate_proj.channel_scale": "model.safetensors",
348
+ "model.layers.17.mlp.gate_proj.q_perm": "model.safetensors",
349
+ "model.layers.17.mlp.gate_proj.qweight": "model.safetensors",
350
+ "model.layers.17.mlp.gate_proj.scales": "model.safetensors",
351
+ "model.layers.17.mlp.gate_proj.zeros": "model.safetensors",
352
+ "model.layers.17.mlp.up_proj.channel_scale": "model.safetensors",
353
+ "model.layers.17.mlp.up_proj.q_perm": "model.safetensors",
354
+ "model.layers.17.mlp.up_proj.qweight": "model.safetensors",
355
+ "model.layers.17.mlp.up_proj.scales": "model.safetensors",
356
+ "model.layers.17.mlp.up_proj.zeros": "model.safetensors",
357
+ "model.layers.17.post_attention_layernorm.weight": "model.safetensors",
358
+ "model.layers.17.self_attn.k_proj.channel_scale": "model.safetensors",
359
+ "model.layers.17.self_attn.k_proj.q_perm": "model.safetensors",
360
+ "model.layers.17.self_attn.k_proj.qweight": "model.safetensors",
361
+ "model.layers.17.self_attn.k_proj.scales": "model.safetensors",
362
+ "model.layers.17.self_attn.k_proj.zeros": "model.safetensors",
363
+ "model.layers.17.self_attn.o_proj.channel_scale": "model.safetensors",
364
+ "model.layers.17.self_attn.o_proj.q_perm": "model.safetensors",
365
+ "model.layers.17.self_attn.o_proj.qweight": "model.safetensors",
366
+ "model.layers.17.self_attn.o_proj.scales": "model.safetensors",
367
+ "model.layers.17.self_attn.o_proj.zeros": "model.safetensors",
368
+ "model.layers.17.self_attn.q_proj.channel_scale": "model.safetensors",
369
+ "model.layers.17.self_attn.q_proj.q_perm": "model.safetensors",
370
+ "model.layers.17.self_attn.q_proj.qweight": "model.safetensors",
371
+ "model.layers.17.self_attn.q_proj.scales": "model.safetensors",
372
+ "model.layers.17.self_attn.q_proj.zeros": "model.safetensors",
373
+ "model.layers.17.self_attn.v_proj.channel_scale": "model.safetensors",
374
+ "model.layers.17.self_attn.v_proj.q_perm": "model.safetensors",
375
+ "model.layers.17.self_attn.v_proj.qweight": "model.safetensors",
376
+ "model.layers.17.self_attn.v_proj.scales": "model.safetensors",
377
+ "model.layers.17.self_attn.v_proj.zeros": "model.safetensors",
378
+ "model.layers.18.input_layernorm.weight": "model.safetensors",
379
+ "model.layers.18.mlp.down_proj.channel_scale": "model.safetensors",
380
+ "model.layers.18.mlp.down_proj.q_perm": "model.safetensors",
381
+ "model.layers.18.mlp.down_proj.qweight": "model.safetensors",
382
+ "model.layers.18.mlp.down_proj.scales": "model.safetensors",
383
+ "model.layers.18.mlp.down_proj.zeros": "model.safetensors",
384
+ "model.layers.18.mlp.gate_proj.channel_scale": "model.safetensors",
385
+ "model.layers.18.mlp.gate_proj.q_perm": "model.safetensors",
386
+ "model.layers.18.mlp.gate_proj.qweight": "model.safetensors",
387
+ "model.layers.18.mlp.gate_proj.scales": "model.safetensors",
388
+ "model.layers.18.mlp.gate_proj.zeros": "model.safetensors",
389
+ "model.layers.18.mlp.up_proj.channel_scale": "model.safetensors",
390
+ "model.layers.18.mlp.up_proj.q_perm": "model.safetensors",
391
+ "model.layers.18.mlp.up_proj.qweight": "model.safetensors",
392
+ "model.layers.18.mlp.up_proj.scales": "model.safetensors",
393
+ "model.layers.18.mlp.up_proj.zeros": "model.safetensors",
394
+ "model.layers.18.post_attention_layernorm.weight": "model.safetensors",
395
+ "model.layers.18.self_attn.k_proj.channel_scale": "model.safetensors",
396
+ "model.layers.18.self_attn.k_proj.q_perm": "model.safetensors",
397
+ "model.layers.18.self_attn.k_proj.qweight": "model.safetensors",
398
+ "model.layers.18.self_attn.k_proj.scales": "model.safetensors",
399
+ "model.layers.18.self_attn.k_proj.zeros": "model.safetensors",
400
+ "model.layers.18.self_attn.o_proj.channel_scale": "model.safetensors",
401
+ "model.layers.18.self_attn.o_proj.q_perm": "model.safetensors",
402
+ "model.layers.18.self_attn.o_proj.qweight": "model.safetensors",
403
+ "model.layers.18.self_attn.o_proj.scales": "model.safetensors",
404
+ "model.layers.18.self_attn.o_proj.zeros": "model.safetensors",
405
+ "model.layers.18.self_attn.q_proj.channel_scale": "model.safetensors",
406
+ "model.layers.18.self_attn.q_proj.q_perm": "model.safetensors",
407
+ "model.layers.18.self_attn.q_proj.qweight": "model.safetensors",
408
+ "model.layers.18.self_attn.q_proj.scales": "model.safetensors",
409
+ "model.layers.18.self_attn.q_proj.zeros": "model.safetensors",
410
+ "model.layers.18.self_attn.v_proj.channel_scale": "model.safetensors",
411
+ "model.layers.18.self_attn.v_proj.q_perm": "model.safetensors",
412
+ "model.layers.18.self_attn.v_proj.qweight": "model.safetensors",
413
+ "model.layers.18.self_attn.v_proj.scales": "model.safetensors",
414
+ "model.layers.18.self_attn.v_proj.zeros": "model.safetensors",
415
+ "model.layers.19.input_layernorm.weight": "model.safetensors",
416
+ "model.layers.19.mlp.down_proj.channel_scale": "model.safetensors",
417
+ "model.layers.19.mlp.down_proj.q_perm": "model.safetensors",
418
+ "model.layers.19.mlp.down_proj.qweight": "model.safetensors",
419
+ "model.layers.19.mlp.down_proj.scales": "model.safetensors",
420
+ "model.layers.19.mlp.down_proj.zeros": "model.safetensors",
421
+ "model.layers.19.mlp.gate_proj.channel_scale": "model.safetensors",
422
+ "model.layers.19.mlp.gate_proj.q_perm": "model.safetensors",
423
+ "model.layers.19.mlp.gate_proj.qweight": "model.safetensors",
424
+ "model.layers.19.mlp.gate_proj.scales": "model.safetensors",
425
+ "model.layers.19.mlp.gate_proj.zeros": "model.safetensors",
426
+ "model.layers.19.mlp.up_proj.channel_scale": "model.safetensors",
427
+ "model.layers.19.mlp.up_proj.q_perm": "model.safetensors",
428
+ "model.layers.19.mlp.up_proj.qweight": "model.safetensors",
429
+ "model.layers.19.mlp.up_proj.scales": "model.safetensors",
430
+ "model.layers.19.mlp.up_proj.zeros": "model.safetensors",
431
+ "model.layers.19.post_attention_layernorm.weight": "model.safetensors",
432
+ "model.layers.19.self_attn.k_proj.channel_scale": "model.safetensors",
433
+ "model.layers.19.self_attn.k_proj.q_perm": "model.safetensors",
434
+ "model.layers.19.self_attn.k_proj.qweight": "model.safetensors",
435
+ "model.layers.19.self_attn.k_proj.scales": "model.safetensors",
436
+ "model.layers.19.self_attn.k_proj.zeros": "model.safetensors",
437
+ "model.layers.19.self_attn.o_proj.channel_scale": "model.safetensors",
438
+ "model.layers.19.self_attn.o_proj.q_perm": "model.safetensors",
439
+ "model.layers.19.self_attn.o_proj.qweight": "model.safetensors",
440
+ "model.layers.19.self_attn.o_proj.scales": "model.safetensors",
441
+ "model.layers.19.self_attn.o_proj.zeros": "model.safetensors",
442
+ "model.layers.19.self_attn.q_proj.channel_scale": "model.safetensors",
443
+ "model.layers.19.self_attn.q_proj.q_perm": "model.safetensors",
444
+ "model.layers.19.self_attn.q_proj.qweight": "model.safetensors",
445
+ "model.layers.19.self_attn.q_proj.scales": "model.safetensors",
446
+ "model.layers.19.self_attn.q_proj.zeros": "model.safetensors",
447
+ "model.layers.19.self_attn.v_proj.channel_scale": "model.safetensors",
448
+ "model.layers.19.self_attn.v_proj.q_perm": "model.safetensors",
449
+ "model.layers.19.self_attn.v_proj.qweight": "model.safetensors",
450
+ "model.layers.19.self_attn.v_proj.scales": "model.safetensors",
451
+ "model.layers.19.self_attn.v_proj.zeros": "model.safetensors",
452
+ "model.layers.2.input_layernorm.weight": "model.safetensors",
453
+ "model.layers.2.mlp.down_proj.channel_scale": "model.safetensors",
454
+ "model.layers.2.mlp.down_proj.q_perm": "model.safetensors",
455
+ "model.layers.2.mlp.down_proj.qweight": "model.safetensors",
456
+ "model.layers.2.mlp.down_proj.scales": "model.safetensors",
457
+ "model.layers.2.mlp.down_proj.zeros": "model.safetensors",
458
+ "model.layers.2.mlp.gate_proj.channel_scale": "model.safetensors",
459
+ "model.layers.2.mlp.gate_proj.q_perm": "model.safetensors",
460
+ "model.layers.2.mlp.gate_proj.qweight": "model.safetensors",
461
+ "model.layers.2.mlp.gate_proj.scales": "model.safetensors",
462
+ "model.layers.2.mlp.gate_proj.zeros": "model.safetensors",
463
+ "model.layers.2.mlp.up_proj.channel_scale": "model.safetensors",
464
+ "model.layers.2.mlp.up_proj.q_perm": "model.safetensors",
465
+ "model.layers.2.mlp.up_proj.qweight": "model.safetensors",
466
+ "model.layers.2.mlp.up_proj.scales": "model.safetensors",
467
+ "model.layers.2.mlp.up_proj.zeros": "model.safetensors",
468
+ "model.layers.2.post_attention_layernorm.weight": "model.safetensors",
469
+ "model.layers.2.self_attn.k_proj.channel_scale": "model.safetensors",
470
+ "model.layers.2.self_attn.k_proj.q_perm": "model.safetensors",
471
+ "model.layers.2.self_attn.k_proj.qweight": "model.safetensors",
472
+ "model.layers.2.self_attn.k_proj.scales": "model.safetensors",
473
+ "model.layers.2.self_attn.k_proj.zeros": "model.safetensors",
474
+ "model.layers.2.self_attn.o_proj.channel_scale": "model.safetensors",
475
+ "model.layers.2.self_attn.o_proj.q_perm": "model.safetensors",
476
+ "model.layers.2.self_attn.o_proj.qweight": "model.safetensors",
477
+ "model.layers.2.self_attn.o_proj.scales": "model.safetensors",
478
+ "model.layers.2.self_attn.o_proj.zeros": "model.safetensors",
479
+ "model.layers.2.self_attn.q_proj.channel_scale": "model.safetensors",
480
+ "model.layers.2.self_attn.q_proj.q_perm": "model.safetensors",
481
+ "model.layers.2.self_attn.q_proj.qweight": "model.safetensors",
482
+ "model.layers.2.self_attn.q_proj.scales": "model.safetensors",
483
+ "model.layers.2.self_attn.q_proj.zeros": "model.safetensors",
484
+ "model.layers.2.self_attn.v_proj.channel_scale": "model.safetensors",
485
+ "model.layers.2.self_attn.v_proj.q_perm": "model.safetensors",
486
+ "model.layers.2.self_attn.v_proj.qweight": "model.safetensors",
487
+ "model.layers.2.self_attn.v_proj.scales": "model.safetensors",
488
+ "model.layers.2.self_attn.v_proj.zeros": "model.safetensors",
489
+ "model.layers.20.input_layernorm.weight": "model.safetensors",
490
+ "model.layers.20.mlp.down_proj.channel_scale": "model.safetensors",
491
+ "model.layers.20.mlp.down_proj.q_perm": "model.safetensors",
492
+ "model.layers.20.mlp.down_proj.qweight": "model.safetensors",
493
+ "model.layers.20.mlp.down_proj.scales": "model.safetensors",
494
+ "model.layers.20.mlp.down_proj.zeros": "model.safetensors",
495
+ "model.layers.20.mlp.gate_proj.channel_scale": "model.safetensors",
496
+ "model.layers.20.mlp.gate_proj.q_perm": "model.safetensors",
497
+ "model.layers.20.mlp.gate_proj.qweight": "model.safetensors",
498
+ "model.layers.20.mlp.gate_proj.scales": "model.safetensors",
499
+ "model.layers.20.mlp.gate_proj.zeros": "model.safetensors",
500
+ "model.layers.20.mlp.up_proj.channel_scale": "model.safetensors",
501
+ "model.layers.20.mlp.up_proj.q_perm": "model.safetensors",
502
+ "model.layers.20.mlp.up_proj.qweight": "model.safetensors",
503
+ "model.layers.20.mlp.up_proj.scales": "model.safetensors",
504
+ "model.layers.20.mlp.up_proj.zeros": "model.safetensors",
505
+ "model.layers.20.post_attention_layernorm.weight": "model.safetensors",
506
+ "model.layers.20.self_attn.k_proj.channel_scale": "model.safetensors",
507
+ "model.layers.20.self_attn.k_proj.q_perm": "model.safetensors",
508
+ "model.layers.20.self_attn.k_proj.qweight": "model.safetensors",
509
+ "model.layers.20.self_attn.k_proj.scales": "model.safetensors",
510
+ "model.layers.20.self_attn.k_proj.zeros": "model.safetensors",
511
+ "model.layers.20.self_attn.o_proj.channel_scale": "model.safetensors",
512
+ "model.layers.20.self_attn.o_proj.q_perm": "model.safetensors",
513
+ "model.layers.20.self_attn.o_proj.qweight": "model.safetensors",
514
+ "model.layers.20.self_attn.o_proj.scales": "model.safetensors",
515
+ "model.layers.20.self_attn.o_proj.zeros": "model.safetensors",
516
+ "model.layers.20.self_attn.q_proj.channel_scale": "model.safetensors",
517
+ "model.layers.20.self_attn.q_proj.q_perm": "model.safetensors",
518
+ "model.layers.20.self_attn.q_proj.qweight": "model.safetensors",
519
+ "model.layers.20.self_attn.q_proj.scales": "model.safetensors",
520
+ "model.layers.20.self_attn.q_proj.zeros": "model.safetensors",
521
+ "model.layers.20.self_attn.v_proj.channel_scale": "model.safetensors",
522
+ "model.layers.20.self_attn.v_proj.q_perm": "model.safetensors",
523
+ "model.layers.20.self_attn.v_proj.qweight": "model.safetensors",
524
+ "model.layers.20.self_attn.v_proj.scales": "model.safetensors",
525
+ "model.layers.20.self_attn.v_proj.zeros": "model.safetensors",
526
+ "model.layers.21.input_layernorm.weight": "model.safetensors",
527
+ "model.layers.21.mlp.down_proj.channel_scale": "model.safetensors",
528
+ "model.layers.21.mlp.down_proj.q_perm": "model.safetensors",
529
+ "model.layers.21.mlp.down_proj.qweight": "model.safetensors",
530
+ "model.layers.21.mlp.down_proj.scales": "model.safetensors",
531
+ "model.layers.21.mlp.down_proj.zeros": "model.safetensors",
532
+ "model.layers.21.mlp.gate_proj.channel_scale": "model.safetensors",
533
+ "model.layers.21.mlp.gate_proj.q_perm": "model.safetensors",
534
+ "model.layers.21.mlp.gate_proj.qweight": "model.safetensors",
535
+ "model.layers.21.mlp.gate_proj.scales": "model.safetensors",
536
+ "model.layers.21.mlp.gate_proj.zeros": "model.safetensors",
537
+ "model.layers.21.mlp.up_proj.channel_scale": "model.safetensors",
538
+ "model.layers.21.mlp.up_proj.q_perm": "model.safetensors",
539
+ "model.layers.21.mlp.up_proj.qweight": "model.safetensors",
540
+ "model.layers.21.mlp.up_proj.scales": "model.safetensors",
541
+ "model.layers.21.mlp.up_proj.zeros": "model.safetensors",
542
+ "model.layers.21.post_attention_layernorm.weight": "model.safetensors",
543
+ "model.layers.21.self_attn.k_proj.channel_scale": "model.safetensors",
544
+ "model.layers.21.self_attn.k_proj.q_perm": "model.safetensors",
545
+ "model.layers.21.self_attn.k_proj.qweight": "model.safetensors",
546
+ "model.layers.21.self_attn.k_proj.scales": "model.safetensors",
547
+ "model.layers.21.self_attn.k_proj.zeros": "model.safetensors",
548
+ "model.layers.21.self_attn.o_proj.channel_scale": "model.safetensors",
549
+ "model.layers.21.self_attn.o_proj.q_perm": "model.safetensors",
550
+ "model.layers.21.self_attn.o_proj.qweight": "model.safetensors",
551
+ "model.layers.21.self_attn.o_proj.scales": "model.safetensors",
552
+ "model.layers.21.self_attn.o_proj.zeros": "model.safetensors",
553
+ "model.layers.21.self_attn.q_proj.channel_scale": "model.safetensors",
554
+ "model.layers.21.self_attn.q_proj.q_perm": "model.safetensors",
555
+ "model.layers.21.self_attn.q_proj.qweight": "model.safetensors",
556
+ "model.layers.21.self_attn.q_proj.scales": "model.safetensors",
557
+ "model.layers.21.self_attn.q_proj.zeros": "model.safetensors",
558
+ "model.layers.21.self_attn.v_proj.channel_scale": "model.safetensors",
559
+ "model.layers.21.self_attn.v_proj.q_perm": "model.safetensors",
560
+ "model.layers.21.self_attn.v_proj.qweight": "model.safetensors",
561
+ "model.layers.21.self_attn.v_proj.scales": "model.safetensors",
562
+ "model.layers.21.self_attn.v_proj.zeros": "model.safetensors",
563
+ "model.layers.22.input_layernorm.weight": "model.safetensors",
564
+ "model.layers.22.mlp.down_proj.channel_scale": "model.safetensors",
565
+ "model.layers.22.mlp.down_proj.q_perm": "model.safetensors",
566
+ "model.layers.22.mlp.down_proj.qweight": "model.safetensors",
567
+ "model.layers.22.mlp.down_proj.scales": "model.safetensors",
568
+ "model.layers.22.mlp.down_proj.zeros": "model.safetensors",
569
+ "model.layers.22.mlp.gate_proj.channel_scale": "model.safetensors",
570
+ "model.layers.22.mlp.gate_proj.q_perm": "model.safetensors",
571
+ "model.layers.22.mlp.gate_proj.qweight": "model.safetensors",
572
+ "model.layers.22.mlp.gate_proj.scales": "model.safetensors",
573
+ "model.layers.22.mlp.gate_proj.zeros": "model.safetensors",
574
+ "model.layers.22.mlp.up_proj.channel_scale": "model.safetensors",
575
+ "model.layers.22.mlp.up_proj.q_perm": "model.safetensors",
576
+ "model.layers.22.mlp.up_proj.qweight": "model.safetensors",
577
+ "model.layers.22.mlp.up_proj.scales": "model.safetensors",
578
+ "model.layers.22.mlp.up_proj.zeros": "model.safetensors",
579
+ "model.layers.22.post_attention_layernorm.weight": "model.safetensors",
580
+ "model.layers.22.self_attn.k_proj.channel_scale": "model.safetensors",
581
+ "model.layers.22.self_attn.k_proj.q_perm": "model.safetensors",
582
+ "model.layers.22.self_attn.k_proj.qweight": "model.safetensors",
583
+ "model.layers.22.self_attn.k_proj.scales": "model.safetensors",
584
+ "model.layers.22.self_attn.k_proj.zeros": "model.safetensors",
585
+ "model.layers.22.self_attn.o_proj.channel_scale": "model.safetensors",
586
+ "model.layers.22.self_attn.o_proj.q_perm": "model.safetensors",
587
+ "model.layers.22.self_attn.o_proj.qweight": "model.safetensors",
588
+ "model.layers.22.self_attn.o_proj.scales": "model.safetensors",
589
+ "model.layers.22.self_attn.o_proj.zeros": "model.safetensors",
590
+ "model.layers.22.self_attn.q_proj.channel_scale": "model.safetensors",
591
+ "model.layers.22.self_attn.q_proj.q_perm": "model.safetensors",
592
+ "model.layers.22.self_attn.q_proj.qweight": "model.safetensors",
593
+ "model.layers.22.self_attn.q_proj.scales": "model.safetensors",
594
+ "model.layers.22.self_attn.q_proj.zeros": "model.safetensors",
595
+ "model.layers.22.self_attn.v_proj.channel_scale": "model.safetensors",
596
+ "model.layers.22.self_attn.v_proj.q_perm": "model.safetensors",
597
+ "model.layers.22.self_attn.v_proj.qweight": "model.safetensors",
598
+ "model.layers.22.self_attn.v_proj.scales": "model.safetensors",
599
+ "model.layers.22.self_attn.v_proj.zeros": "model.safetensors",
600
+ "model.layers.23.input_layernorm.weight": "model.safetensors",
601
+ "model.layers.23.mlp.down_proj.channel_scale": "model.safetensors",
602
+ "model.layers.23.mlp.down_proj.q_perm": "model.safetensors",
603
+ "model.layers.23.mlp.down_proj.qweight": "model.safetensors",
604
+ "model.layers.23.mlp.down_proj.scales": "model.safetensors",
605
+ "model.layers.23.mlp.down_proj.zeros": "model.safetensors",
606
+ "model.layers.23.mlp.gate_proj.channel_scale": "model.safetensors",
607
+ "model.layers.23.mlp.gate_proj.q_perm": "model.safetensors",
608
+ "model.layers.23.mlp.gate_proj.qweight": "model.safetensors",
609
+ "model.layers.23.mlp.gate_proj.scales": "model.safetensors",
610
+ "model.layers.23.mlp.gate_proj.zeros": "model.safetensors",
611
+ "model.layers.23.mlp.up_proj.channel_scale": "model.safetensors",
612
+ "model.layers.23.mlp.up_proj.q_perm": "model.safetensors",
613
+ "model.layers.23.mlp.up_proj.qweight": "model.safetensors",
614
+ "model.layers.23.mlp.up_proj.scales": "model.safetensors",
615
+ "model.layers.23.mlp.up_proj.zeros": "model.safetensors",
616
+ "model.layers.23.post_attention_layernorm.weight": "model.safetensors",
617
+ "model.layers.23.self_attn.k_proj.channel_scale": "model.safetensors",
618
+ "model.layers.23.self_attn.k_proj.q_perm": "model.safetensors",
619
+ "model.layers.23.self_attn.k_proj.qweight": "model.safetensors",
620
+ "model.layers.23.self_attn.k_proj.scales": "model.safetensors",
621
+ "model.layers.23.self_attn.k_proj.zeros": "model.safetensors",
622
+ "model.layers.23.self_attn.o_proj.channel_scale": "model.safetensors",
623
+ "model.layers.23.self_attn.o_proj.q_perm": "model.safetensors",
624
+ "model.layers.23.self_attn.o_proj.qweight": "model.safetensors",
625
+ "model.layers.23.self_attn.o_proj.scales": "model.safetensors",
626
+ "model.layers.23.self_attn.o_proj.zeros": "model.safetensors",
627
+ "model.layers.23.self_attn.q_proj.channel_scale": "model.safetensors",
628
+ "model.layers.23.self_attn.q_proj.q_perm": "model.safetensors",
629
+ "model.layers.23.self_attn.q_proj.qweight": "model.safetensors",
630
+ "model.layers.23.self_attn.q_proj.scales": "model.safetensors",
631
+ "model.layers.23.self_attn.q_proj.zeros": "model.safetensors",
632
+ "model.layers.23.self_attn.v_proj.channel_scale": "model.safetensors",
633
+ "model.layers.23.self_attn.v_proj.q_perm": "model.safetensors",
634
+ "model.layers.23.self_attn.v_proj.qweight": "model.safetensors",
635
+ "model.layers.23.self_attn.v_proj.scales": "model.safetensors",
636
+ "model.layers.23.self_attn.v_proj.zeros": "model.safetensors",
637
+ "model.layers.24.input_layernorm.weight": "model.safetensors",
638
+ "model.layers.24.mlp.down_proj.channel_scale": "model.safetensors",
639
+ "model.layers.24.mlp.down_proj.q_perm": "model.safetensors",
640
+ "model.layers.24.mlp.down_proj.qweight": "model.safetensors",
641
+ "model.layers.24.mlp.down_proj.scales": "model.safetensors",
642
+ "model.layers.24.mlp.down_proj.zeros": "model.safetensors",
643
+ "model.layers.24.mlp.gate_proj.channel_scale": "model.safetensors",
644
+ "model.layers.24.mlp.gate_proj.q_perm": "model.safetensors",
645
+ "model.layers.24.mlp.gate_proj.qweight": "model.safetensors",
646
+ "model.layers.24.mlp.gate_proj.scales": "model.safetensors",
647
+ "model.layers.24.mlp.gate_proj.zeros": "model.safetensors",
648
+ "model.layers.24.mlp.up_proj.channel_scale": "model.safetensors",
649
+ "model.layers.24.mlp.up_proj.q_perm": "model.safetensors",
650
+ "model.layers.24.mlp.up_proj.qweight": "model.safetensors",
651
+ "model.layers.24.mlp.up_proj.scales": "model.safetensors",
652
+ "model.layers.24.mlp.up_proj.zeros": "model.safetensors",
653
+ "model.layers.24.post_attention_layernorm.weight": "model.safetensors",
654
+ "model.layers.24.self_attn.k_proj.channel_scale": "model.safetensors",
655
+ "model.layers.24.self_attn.k_proj.q_perm": "model.safetensors",
656
+ "model.layers.24.self_attn.k_proj.qweight": "model.safetensors",
657
+ "model.layers.24.self_attn.k_proj.scales": "model.safetensors",
658
+ "model.layers.24.self_attn.k_proj.zeros": "model.safetensors",
659
+ "model.layers.24.self_attn.o_proj.channel_scale": "model.safetensors",
660
+ "model.layers.24.self_attn.o_proj.q_perm": "model.safetensors",
661
+ "model.layers.24.self_attn.o_proj.qweight": "model.safetensors",
662
+ "model.layers.24.self_attn.o_proj.scales": "model.safetensors",
663
+ "model.layers.24.self_attn.o_proj.zeros": "model.safetensors",
664
+ "model.layers.24.self_attn.q_proj.channel_scale": "model.safetensors",
665
+ "model.layers.24.self_attn.q_proj.q_perm": "model.safetensors",
666
+ "model.layers.24.self_attn.q_proj.qweight": "model.safetensors",
667
+ "model.layers.24.self_attn.q_proj.scales": "model.safetensors",
668
+ "model.layers.24.self_attn.q_proj.zeros": "model.safetensors",
669
+ "model.layers.24.self_attn.v_proj.channel_scale": "model.safetensors",
670
+ "model.layers.24.self_attn.v_proj.q_perm": "model.safetensors",
671
+ "model.layers.24.self_attn.v_proj.qweight": "model.safetensors",
672
+ "model.layers.24.self_attn.v_proj.scales": "model.safetensors",
673
+ "model.layers.24.self_attn.v_proj.zeros": "model.safetensors",
674
+ "model.layers.25.input_layernorm.weight": "model.safetensors",
675
+ "model.layers.25.mlp.down_proj.channel_scale": "model.safetensors",
676
+ "model.layers.25.mlp.down_proj.q_perm": "model.safetensors",
677
+ "model.layers.25.mlp.down_proj.qweight": "model.safetensors",
678
+ "model.layers.25.mlp.down_proj.scales": "model.safetensors",
679
+ "model.layers.25.mlp.down_proj.zeros": "model.safetensors",
680
+ "model.layers.25.mlp.gate_proj.channel_scale": "model.safetensors",
681
+ "model.layers.25.mlp.gate_proj.q_perm": "model.safetensors",
682
+ "model.layers.25.mlp.gate_proj.qweight": "model.safetensors",
683
+ "model.layers.25.mlp.gate_proj.scales": "model.safetensors",
684
+ "model.layers.25.mlp.gate_proj.zeros": "model.safetensors",
685
+ "model.layers.25.mlp.up_proj.channel_scale": "model.safetensors",
686
+ "model.layers.25.mlp.up_proj.q_perm": "model.safetensors",
687
+ "model.layers.25.mlp.up_proj.qweight": "model.safetensors",
688
+ "model.layers.25.mlp.up_proj.scales": "model.safetensors",
689
+ "model.layers.25.mlp.up_proj.zeros": "model.safetensors",
690
+ "model.layers.25.post_attention_layernorm.weight": "model.safetensors",
691
+ "model.layers.25.self_attn.k_proj.channel_scale": "model.safetensors",
692
+ "model.layers.25.self_attn.k_proj.q_perm": "model.safetensors",
693
+ "model.layers.25.self_attn.k_proj.qweight": "model.safetensors",
694
+ "model.layers.25.self_attn.k_proj.scales": "model.safetensors",
695
+ "model.layers.25.self_attn.k_proj.zeros": "model.safetensors",
696
+ "model.layers.25.self_attn.o_proj.channel_scale": "model.safetensors",
697
+ "model.layers.25.self_attn.o_proj.q_perm": "model.safetensors",
698
+ "model.layers.25.self_attn.o_proj.qweight": "model.safetensors",
699
+ "model.layers.25.self_attn.o_proj.scales": "model.safetensors",
700
+ "model.layers.25.self_attn.o_proj.zeros": "model.safetensors",
701
+ "model.layers.25.self_attn.q_proj.channel_scale": "model.safetensors",
702
+ "model.layers.25.self_attn.q_proj.q_perm": "model.safetensors",
703
+ "model.layers.25.self_attn.q_proj.qweight": "model.safetensors",
704
+ "model.layers.25.self_attn.q_proj.scales": "model.safetensors",
705
+ "model.layers.25.self_attn.q_proj.zeros": "model.safetensors",
706
+ "model.layers.25.self_attn.v_proj.channel_scale": "model.safetensors",
707
+ "model.layers.25.self_attn.v_proj.q_perm": "model.safetensors",
708
+ "model.layers.25.self_attn.v_proj.qweight": "model.safetensors",
709
+ "model.layers.25.self_attn.v_proj.scales": "model.safetensors",
710
+ "model.layers.25.self_attn.v_proj.zeros": "model.safetensors",
711
+ "model.layers.26.input_layernorm.weight": "model.safetensors",
712
+ "model.layers.26.mlp.down_proj.channel_scale": "model.safetensors",
713
+ "model.layers.26.mlp.down_proj.q_perm": "model.safetensors",
714
+ "model.layers.26.mlp.down_proj.qweight": "model.safetensors",
715
+ "model.layers.26.mlp.down_proj.scales": "model.safetensors",
716
+ "model.layers.26.mlp.down_proj.zeros": "model.safetensors",
717
+ "model.layers.26.mlp.gate_proj.channel_scale": "model.safetensors",
718
+ "model.layers.26.mlp.gate_proj.q_perm": "model.safetensors",
719
+ "model.layers.26.mlp.gate_proj.qweight": "model.safetensors",
720
+ "model.layers.26.mlp.gate_proj.scales": "model.safetensors",
721
+ "model.layers.26.mlp.gate_proj.zeros": "model.safetensors",
722
+ "model.layers.26.mlp.up_proj.channel_scale": "model.safetensors",
723
+ "model.layers.26.mlp.up_proj.q_perm": "model.safetensors",
724
+ "model.layers.26.mlp.up_proj.qweight": "model.safetensors",
725
+ "model.layers.26.mlp.up_proj.scales": "model.safetensors",
726
+ "model.layers.26.mlp.up_proj.zeros": "model.safetensors",
727
+ "model.layers.26.post_attention_layernorm.weight": "model.safetensors",
728
+ "model.layers.26.self_attn.k_proj.channel_scale": "model.safetensors",
729
+ "model.layers.26.self_attn.k_proj.q_perm": "model.safetensors",
730
+ "model.layers.26.self_attn.k_proj.qweight": "model.safetensors",
731
+ "model.layers.26.self_attn.k_proj.scales": "model.safetensors",
732
+ "model.layers.26.self_attn.k_proj.zeros": "model.safetensors",
733
+ "model.layers.26.self_attn.o_proj.channel_scale": "model.safetensors",
734
+ "model.layers.26.self_attn.o_proj.q_perm": "model.safetensors",
735
+ "model.layers.26.self_attn.o_proj.qweight": "model.safetensors",
736
+ "model.layers.26.self_attn.o_proj.scales": "model.safetensors",
737
+ "model.layers.26.self_attn.o_proj.zeros": "model.safetensors",
738
+ "model.layers.26.self_attn.q_proj.channel_scale": "model.safetensors",
739
+ "model.layers.26.self_attn.q_proj.q_perm": "model.safetensors",
740
+ "model.layers.26.self_attn.q_proj.qweight": "model.safetensors",
741
+ "model.layers.26.self_attn.q_proj.scales": "model.safetensors",
742
+ "model.layers.26.self_attn.q_proj.zeros": "model.safetensors",
743
+ "model.layers.26.self_attn.v_proj.channel_scale": "model.safetensors",
744
+ "model.layers.26.self_attn.v_proj.q_perm": "model.safetensors",
745
+ "model.layers.26.self_attn.v_proj.qweight": "model.safetensors",
746
+ "model.layers.26.self_attn.v_proj.scales": "model.safetensors",
747
+ "model.layers.26.self_attn.v_proj.zeros": "model.safetensors",
748
+ "model.layers.27.input_layernorm.weight": "model.safetensors",
749
+ "model.layers.27.mlp.down_proj.channel_scale": "model.safetensors",
750
+ "model.layers.27.mlp.down_proj.q_perm": "model.safetensors",
751
+ "model.layers.27.mlp.down_proj.qweight": "model.safetensors",
752
+ "model.layers.27.mlp.down_proj.scales": "model.safetensors",
753
+ "model.layers.27.mlp.down_proj.zeros": "model.safetensors",
754
+ "model.layers.27.mlp.gate_proj.channel_scale": "model.safetensors",
755
+ "model.layers.27.mlp.gate_proj.q_perm": "model.safetensors",
756
+ "model.layers.27.mlp.gate_proj.qweight": "model.safetensors",
757
+ "model.layers.27.mlp.gate_proj.scales": "model.safetensors",
758
+ "model.layers.27.mlp.gate_proj.zeros": "model.safetensors",
759
+ "model.layers.27.mlp.up_proj.channel_scale": "model.safetensors",
760
+ "model.layers.27.mlp.up_proj.q_perm": "model.safetensors",
761
+ "model.layers.27.mlp.up_proj.qweight": "model.safetensors",
762
+ "model.layers.27.mlp.up_proj.scales": "model.safetensors",
763
+ "model.layers.27.mlp.up_proj.zeros": "model.safetensors",
764
+ "model.layers.27.post_attention_layernorm.weight": "model.safetensors",
765
+ "model.layers.27.self_attn.k_proj.channel_scale": "model.safetensors",
766
+ "model.layers.27.self_attn.k_proj.q_perm": "model.safetensors",
767
+ "model.layers.27.self_attn.k_proj.qweight": "model.safetensors",
768
+ "model.layers.27.self_attn.k_proj.scales": "model.safetensors",
769
+ "model.layers.27.self_attn.k_proj.zeros": "model.safetensors",
770
+ "model.layers.27.self_attn.o_proj.channel_scale": "model.safetensors",
771
+ "model.layers.27.self_attn.o_proj.q_perm": "model.safetensors",
772
+ "model.layers.27.self_attn.o_proj.qweight": "model.safetensors",
773
+ "model.layers.27.self_attn.o_proj.scales": "model.safetensors",
774
+ "model.layers.27.self_attn.o_proj.zeros": "model.safetensors",
775
+ "model.layers.27.self_attn.q_proj.channel_scale": "model.safetensors",
776
+ "model.layers.27.self_attn.q_proj.q_perm": "model.safetensors",
777
+ "model.layers.27.self_attn.q_proj.qweight": "model.safetensors",
778
+ "model.layers.27.self_attn.q_proj.scales": "model.safetensors",
779
+ "model.layers.27.self_attn.q_proj.zeros": "model.safetensors",
780
+ "model.layers.27.self_attn.v_proj.channel_scale": "model.safetensors",
781
+ "model.layers.27.self_attn.v_proj.q_perm": "model.safetensors",
782
+ "model.layers.27.self_attn.v_proj.qweight": "model.safetensors",
783
+ "model.layers.27.self_attn.v_proj.scales": "model.safetensors",
784
+ "model.layers.27.self_attn.v_proj.zeros": "model.safetensors",
785
+ "model.layers.28.input_layernorm.weight": "model.safetensors",
786
+ "model.layers.28.mlp.down_proj.channel_scale": "model.safetensors",
787
+ "model.layers.28.mlp.down_proj.q_perm": "model.safetensors",
788
+ "model.layers.28.mlp.down_proj.qweight": "model.safetensors",
789
+ "model.layers.28.mlp.down_proj.scales": "model.safetensors",
790
+ "model.layers.28.mlp.down_proj.zeros": "model.safetensors",
791
+ "model.layers.28.mlp.gate_proj.channel_scale": "model.safetensors",
792
+ "model.layers.28.mlp.gate_proj.q_perm": "model.safetensors",
793
+ "model.layers.28.mlp.gate_proj.qweight": "model.safetensors",
794
+ "model.layers.28.mlp.gate_proj.scales": "model.safetensors",
795
+ "model.layers.28.mlp.gate_proj.zeros": "model.safetensors",
796
+ "model.layers.28.mlp.up_proj.channel_scale": "model.safetensors",
797
+ "model.layers.28.mlp.up_proj.q_perm": "model.safetensors",
798
+ "model.layers.28.mlp.up_proj.qweight": "model.safetensors",
799
+ "model.layers.28.mlp.up_proj.scales": "model.safetensors",
800
+ "model.layers.28.mlp.up_proj.zeros": "model.safetensors",
801
+ "model.layers.28.post_attention_layernorm.weight": "model.safetensors",
802
+ "model.layers.28.self_attn.k_proj.channel_scale": "model.safetensors",
803
+ "model.layers.28.self_attn.k_proj.q_perm": "model.safetensors",
804
+ "model.layers.28.self_attn.k_proj.qweight": "model.safetensors",
805
+ "model.layers.28.self_attn.k_proj.scales": "model.safetensors",
806
+ "model.layers.28.self_attn.k_proj.zeros": "model.safetensors",
807
+ "model.layers.28.self_attn.o_proj.channel_scale": "model.safetensors",
808
+ "model.layers.28.self_attn.o_proj.q_perm": "model.safetensors",
809
+ "model.layers.28.self_attn.o_proj.qweight": "model.safetensors",
810
+ "model.layers.28.self_attn.o_proj.scales": "model.safetensors",
811
+ "model.layers.28.self_attn.o_proj.zeros": "model.safetensors",
812
+ "model.layers.28.self_attn.q_proj.channel_scale": "model.safetensors",
813
+ "model.layers.28.self_attn.q_proj.q_perm": "model.safetensors",
814
+ "model.layers.28.self_attn.q_proj.qweight": "model.safetensors",
815
+ "model.layers.28.self_attn.q_proj.scales": "model.safetensors",
816
+ "model.layers.28.self_attn.q_proj.zeros": "model.safetensors",
817
+ "model.layers.28.self_attn.v_proj.channel_scale": "model.safetensors",
818
+ "model.layers.28.self_attn.v_proj.q_perm": "model.safetensors",
819
+ "model.layers.28.self_attn.v_proj.qweight": "model.safetensors",
820
+ "model.layers.28.self_attn.v_proj.scales": "model.safetensors",
821
+ "model.layers.28.self_attn.v_proj.zeros": "model.safetensors",
822
+ "model.layers.29.input_layernorm.weight": "model.safetensors",
823
+ "model.layers.29.mlp.down_proj.channel_scale": "model.safetensors",
824
+ "model.layers.29.mlp.down_proj.q_perm": "model.safetensors",
825
+ "model.layers.29.mlp.down_proj.qweight": "model.safetensors",
826
+ "model.layers.29.mlp.down_proj.scales": "model.safetensors",
827
+ "model.layers.29.mlp.down_proj.zeros": "model.safetensors",
828
+ "model.layers.29.mlp.gate_proj.channel_scale": "model.safetensors",
829
+ "model.layers.29.mlp.gate_proj.q_perm": "model.safetensors",
830
+ "model.layers.29.mlp.gate_proj.qweight": "model.safetensors",
831
+ "model.layers.29.mlp.gate_proj.scales": "model.safetensors",
832
+ "model.layers.29.mlp.gate_proj.zeros": "model.safetensors",
833
+ "model.layers.29.mlp.up_proj.channel_scale": "model.safetensors",
834
+ "model.layers.29.mlp.up_proj.q_perm": "model.safetensors",
835
+ "model.layers.29.mlp.up_proj.qweight": "model.safetensors",
836
+ "model.layers.29.mlp.up_proj.scales": "model.safetensors",
837
+ "model.layers.29.mlp.up_proj.zeros": "model.safetensors",
838
+ "model.layers.29.post_attention_layernorm.weight": "model.safetensors",
839
+ "model.layers.29.self_attn.k_proj.channel_scale": "model.safetensors",
840
+ "model.layers.29.self_attn.k_proj.q_perm": "model.safetensors",
841
+ "model.layers.29.self_attn.k_proj.qweight": "model.safetensors",
842
+ "model.layers.29.self_attn.k_proj.scales": "model.safetensors",
843
+ "model.layers.29.self_attn.k_proj.zeros": "model.safetensors",
844
+ "model.layers.29.self_attn.o_proj.channel_scale": "model.safetensors",
845
+ "model.layers.29.self_attn.o_proj.q_perm": "model.safetensors",
846
+ "model.layers.29.self_attn.o_proj.qweight": "model.safetensors",
847
+ "model.layers.29.self_attn.o_proj.scales": "model.safetensors",
848
+ "model.layers.29.self_attn.o_proj.zeros": "model.safetensors",
849
+ "model.layers.29.self_attn.q_proj.channel_scale": "model.safetensors",
850
+ "model.layers.29.self_attn.q_proj.q_perm": "model.safetensors",
851
+ "model.layers.29.self_attn.q_proj.qweight": "model.safetensors",
852
+ "model.layers.29.self_attn.q_proj.scales": "model.safetensors",
853
+ "model.layers.29.self_attn.q_proj.zeros": "model.safetensors",
854
+ "model.layers.29.self_attn.v_proj.channel_scale": "model.safetensors",
855
+ "model.layers.29.self_attn.v_proj.q_perm": "model.safetensors",
856
+ "model.layers.29.self_attn.v_proj.qweight": "model.safetensors",
857
+ "model.layers.29.self_attn.v_proj.scales": "model.safetensors",
858
+ "model.layers.29.self_attn.v_proj.zeros": "model.safetensors",
859
+ "model.layers.3.input_layernorm.weight": "model.safetensors",
860
+ "model.layers.3.mlp.down_proj.channel_scale": "model.safetensors",
861
+ "model.layers.3.mlp.down_proj.q_perm": "model.safetensors",
862
+ "model.layers.3.mlp.down_proj.qweight": "model.safetensors",
863
+ "model.layers.3.mlp.down_proj.scales": "model.safetensors",
864
+ "model.layers.3.mlp.down_proj.zeros": "model.safetensors",
865
+ "model.layers.3.mlp.gate_proj.channel_scale": "model.safetensors",
866
+ "model.layers.3.mlp.gate_proj.q_perm": "model.safetensors",
867
+ "model.layers.3.mlp.gate_proj.qweight": "model.safetensors",
868
+ "model.layers.3.mlp.gate_proj.scales": "model.safetensors",
869
+ "model.layers.3.mlp.gate_proj.zeros": "model.safetensors",
870
+ "model.layers.3.mlp.up_proj.channel_scale": "model.safetensors",
871
+ "model.layers.3.mlp.up_proj.q_perm": "model.safetensors",
872
+ "model.layers.3.mlp.up_proj.qweight": "model.safetensors",
873
+ "model.layers.3.mlp.up_proj.scales": "model.safetensors",
874
+ "model.layers.3.mlp.up_proj.zeros": "model.safetensors",
875
+ "model.layers.3.post_attention_layernorm.weight": "model.safetensors",
876
+ "model.layers.3.self_attn.k_proj.channel_scale": "model.safetensors",
877
+ "model.layers.3.self_attn.k_proj.q_perm": "model.safetensors",
878
+ "model.layers.3.self_attn.k_proj.qweight": "model.safetensors",
879
+ "model.layers.3.self_attn.k_proj.scales": "model.safetensors",
880
+ "model.layers.3.self_attn.k_proj.zeros": "model.safetensors",
881
+ "model.layers.3.self_attn.o_proj.channel_scale": "model.safetensors",
882
+ "model.layers.3.self_attn.o_proj.q_perm": "model.safetensors",
883
+ "model.layers.3.self_attn.o_proj.qweight": "model.safetensors",
884
+ "model.layers.3.self_attn.o_proj.scales": "model.safetensors",
885
+ "model.layers.3.self_attn.o_proj.zeros": "model.safetensors",
886
+ "model.layers.3.self_attn.q_proj.channel_scale": "model.safetensors",
887
+ "model.layers.3.self_attn.q_proj.q_perm": "model.safetensors",
888
+ "model.layers.3.self_attn.q_proj.qweight": "model.safetensors",
889
+ "model.layers.3.self_attn.q_proj.scales": "model.safetensors",
890
+ "model.layers.3.self_attn.q_proj.zeros": "model.safetensors",
891
+ "model.layers.3.self_attn.v_proj.channel_scale": "model.safetensors",
892
+ "model.layers.3.self_attn.v_proj.q_perm": "model.safetensors",
893
+ "model.layers.3.self_attn.v_proj.qweight": "model.safetensors",
894
+ "model.layers.3.self_attn.v_proj.scales": "model.safetensors",
895
+ "model.layers.3.self_attn.v_proj.zeros": "model.safetensors",
896
+ "model.layers.30.input_layernorm.weight": "model.safetensors",
897
+ "model.layers.30.mlp.down_proj.channel_scale": "model.safetensors",
898
+ "model.layers.30.mlp.down_proj.q_perm": "model.safetensors",
899
+ "model.layers.30.mlp.down_proj.qweight": "model.safetensors",
900
+ "model.layers.30.mlp.down_proj.scales": "model.safetensors",
901
+ "model.layers.30.mlp.down_proj.zeros": "model.safetensors",
902
+ "model.layers.30.mlp.gate_proj.channel_scale": "model.safetensors",
903
+ "model.layers.30.mlp.gate_proj.q_perm": "model.safetensors",
904
+ "model.layers.30.mlp.gate_proj.qweight": "model.safetensors",
905
+ "model.layers.30.mlp.gate_proj.scales": "model.safetensors",
906
+ "model.layers.30.mlp.gate_proj.zeros": "model.safetensors",
907
+ "model.layers.30.mlp.up_proj.channel_scale": "model.safetensors",
908
+ "model.layers.30.mlp.up_proj.q_perm": "model.safetensors",
909
+ "model.layers.30.mlp.up_proj.qweight": "model.safetensors",
910
+ "model.layers.30.mlp.up_proj.scales": "model.safetensors",
911
+ "model.layers.30.mlp.up_proj.zeros": "model.safetensors",
912
+ "model.layers.30.post_attention_layernorm.weight": "model.safetensors",
913
+ "model.layers.30.self_attn.k_proj.channel_scale": "model.safetensors",
914
+ "model.layers.30.self_attn.k_proj.q_perm": "model.safetensors",
915
+ "model.layers.30.self_attn.k_proj.qweight": "model.safetensors",
916
+ "model.layers.30.self_attn.k_proj.scales": "model.safetensors",
917
+ "model.layers.30.self_attn.k_proj.zeros": "model.safetensors",
918
+ "model.layers.30.self_attn.o_proj.channel_scale": "model.safetensors",
919
+ "model.layers.30.self_attn.o_proj.q_perm": "model.safetensors",
920
+ "model.layers.30.self_attn.o_proj.qweight": "model.safetensors",
921
+ "model.layers.30.self_attn.o_proj.scales": "model.safetensors",
922
+ "model.layers.30.self_attn.o_proj.zeros": "model.safetensors",
923
+ "model.layers.30.self_attn.q_proj.channel_scale": "model.safetensors",
924
+ "model.layers.30.self_attn.q_proj.q_perm": "model.safetensors",
925
+ "model.layers.30.self_attn.q_proj.qweight": "model.safetensors",
926
+ "model.layers.30.self_attn.q_proj.scales": "model.safetensors",
927
+ "model.layers.30.self_attn.q_proj.zeros": "model.safetensors",
928
+ "model.layers.30.self_attn.v_proj.channel_scale": "model.safetensors",
929
+ "model.layers.30.self_attn.v_proj.q_perm": "model.safetensors",
930
+ "model.layers.30.self_attn.v_proj.qweight": "model.safetensors",
931
+ "model.layers.30.self_attn.v_proj.scales": "model.safetensors",
932
+ "model.layers.30.self_attn.v_proj.zeros": "model.safetensors",
933
+ "model.layers.31.input_layernorm.weight": "model.safetensors",
934
+ "model.layers.31.mlp.down_proj.channel_scale": "model.safetensors",
935
+ "model.layers.31.mlp.down_proj.q_perm": "model.safetensors",
936
+ "model.layers.31.mlp.down_proj.qweight": "model.safetensors",
937
+ "model.layers.31.mlp.down_proj.scales": "model.safetensors",
938
+ "model.layers.31.mlp.down_proj.zeros": "model.safetensors",
939
+ "model.layers.31.mlp.gate_proj.channel_scale": "model.safetensors",
940
+ "model.layers.31.mlp.gate_proj.q_perm": "model.safetensors",
941
+ "model.layers.31.mlp.gate_proj.qweight": "model.safetensors",
942
+ "model.layers.31.mlp.gate_proj.scales": "model.safetensors",
943
+ "model.layers.31.mlp.gate_proj.zeros": "model.safetensors",
944
+ "model.layers.31.mlp.up_proj.channel_scale": "model.safetensors",
945
+ "model.layers.31.mlp.up_proj.q_perm": "model.safetensors",
946
+ "model.layers.31.mlp.up_proj.qweight": "model.safetensors",
947
+ "model.layers.31.mlp.up_proj.scales": "model.safetensors",
948
+ "model.layers.31.mlp.up_proj.zeros": "model.safetensors",
949
+ "model.layers.31.post_attention_layernorm.weight": "model.safetensors",
950
+ "model.layers.31.self_attn.k_proj.channel_scale": "model.safetensors",
951
+ "model.layers.31.self_attn.k_proj.q_perm": "model.safetensors",
952
+ "model.layers.31.self_attn.k_proj.qweight": "model.safetensors",
953
+ "model.layers.31.self_attn.k_proj.scales": "model.safetensors",
954
+ "model.layers.31.self_attn.k_proj.zeros": "model.safetensors",
955
+ "model.layers.31.self_attn.o_proj.channel_scale": "model.safetensors",
956
+ "model.layers.31.self_attn.o_proj.q_perm": "model.safetensors",
957
+ "model.layers.31.self_attn.o_proj.qweight": "model.safetensors",
958
+ "model.layers.31.self_attn.o_proj.scales": "model.safetensors",
959
+ "model.layers.31.self_attn.o_proj.zeros": "model.safetensors",
960
+ "model.layers.31.self_attn.q_proj.channel_scale": "model.safetensors",
961
+ "model.layers.31.self_attn.q_proj.q_perm": "model.safetensors",
962
+ "model.layers.31.self_attn.q_proj.qweight": "model.safetensors",
963
+ "model.layers.31.self_attn.q_proj.scales": "model.safetensors",
964
+ "model.layers.31.self_attn.q_proj.zeros": "model.safetensors",
965
+ "model.layers.31.self_attn.v_proj.channel_scale": "model.safetensors",
966
+ "model.layers.31.self_attn.v_proj.q_perm": "model.safetensors",
967
+ "model.layers.31.self_attn.v_proj.qweight": "model.safetensors",
968
+ "model.layers.31.self_attn.v_proj.scales": "model.safetensors",
969
+ "model.layers.31.self_attn.v_proj.zeros": "model.safetensors",
970
+ "model.layers.4.input_layernorm.weight": "model.safetensors",
971
+ "model.layers.4.mlp.down_proj.channel_scale": "model.safetensors",
972
+ "model.layers.4.mlp.down_proj.q_perm": "model.safetensors",
973
+ "model.layers.4.mlp.down_proj.qweight": "model.safetensors",
974
+ "model.layers.4.mlp.down_proj.scales": "model.safetensors",
975
+ "model.layers.4.mlp.down_proj.zeros": "model.safetensors",
976
+ "model.layers.4.mlp.gate_proj.channel_scale": "model.safetensors",
977
+ "model.layers.4.mlp.gate_proj.q_perm": "model.safetensors",
978
+ "model.layers.4.mlp.gate_proj.qweight": "model.safetensors",
979
+ "model.layers.4.mlp.gate_proj.scales": "model.safetensors",
980
+ "model.layers.4.mlp.gate_proj.zeros": "model.safetensors",
981
+ "model.layers.4.mlp.up_proj.channel_scale": "model.safetensors",
982
+ "model.layers.4.mlp.up_proj.q_perm": "model.safetensors",
983
+ "model.layers.4.mlp.up_proj.qweight": "model.safetensors",
984
+ "model.layers.4.mlp.up_proj.scales": "model.safetensors",
985
+ "model.layers.4.mlp.up_proj.zeros": "model.safetensors",
986
+ "model.layers.4.post_attention_layernorm.weight": "model.safetensors",
987
+ "model.layers.4.self_attn.k_proj.channel_scale": "model.safetensors",
988
+ "model.layers.4.self_attn.k_proj.q_perm": "model.safetensors",
989
+ "model.layers.4.self_attn.k_proj.qweight": "model.safetensors",
990
+ "model.layers.4.self_attn.k_proj.scales": "model.safetensors",
991
+ "model.layers.4.self_attn.k_proj.zeros": "model.safetensors",
992
+ "model.layers.4.self_attn.o_proj.channel_scale": "model.safetensors",
993
+ "model.layers.4.self_attn.o_proj.q_perm": "model.safetensors",
994
+ "model.layers.4.self_attn.o_proj.qweight": "model.safetensors",
995
+ "model.layers.4.self_attn.o_proj.scales": "model.safetensors",
996
+ "model.layers.4.self_attn.o_proj.zeros": "model.safetensors",
997
+ "model.layers.4.self_attn.q_proj.channel_scale": "model.safetensors",
998
+ "model.layers.4.self_attn.q_proj.q_perm": "model.safetensors",
999
+ "model.layers.4.self_attn.q_proj.qweight": "model.safetensors",
1000
+ "model.layers.4.self_attn.q_proj.scales": "model.safetensors",
1001
+ "model.layers.4.self_attn.q_proj.zeros": "model.safetensors",
1002
+ "model.layers.4.self_attn.v_proj.channel_scale": "model.safetensors",
1003
+ "model.layers.4.self_attn.v_proj.q_perm": "model.safetensors",
1004
+ "model.layers.4.self_attn.v_proj.qweight": "model.safetensors",
1005
+ "model.layers.4.self_attn.v_proj.scales": "model.safetensors",
1006
+ "model.layers.4.self_attn.v_proj.zeros": "model.safetensors",
1007
+ "model.layers.5.input_layernorm.weight": "model.safetensors",
1008
+ "model.layers.5.mlp.down_proj.channel_scale": "model.safetensors",
1009
+ "model.layers.5.mlp.down_proj.q_perm": "model.safetensors",
1010
+ "model.layers.5.mlp.down_proj.qweight": "model.safetensors",
1011
+ "model.layers.5.mlp.down_proj.scales": "model.safetensors",
1012
+ "model.layers.5.mlp.down_proj.zeros": "model.safetensors",
1013
+ "model.layers.5.mlp.gate_proj.channel_scale": "model.safetensors",
1014
+ "model.layers.5.mlp.gate_proj.q_perm": "model.safetensors",
1015
+ "model.layers.5.mlp.gate_proj.qweight": "model.safetensors",
1016
+ "model.layers.5.mlp.gate_proj.scales": "model.safetensors",
1017
+ "model.layers.5.mlp.gate_proj.zeros": "model.safetensors",
1018
+ "model.layers.5.mlp.up_proj.channel_scale": "model.safetensors",
1019
+ "model.layers.5.mlp.up_proj.q_perm": "model.safetensors",
1020
+ "model.layers.5.mlp.up_proj.qweight": "model.safetensors",
1021
+ "model.layers.5.mlp.up_proj.scales": "model.safetensors",
1022
+ "model.layers.5.mlp.up_proj.zeros": "model.safetensors",
1023
+ "model.layers.5.post_attention_layernorm.weight": "model.safetensors",
1024
+ "model.layers.5.self_attn.k_proj.channel_scale": "model.safetensors",
1025
+ "model.layers.5.self_attn.k_proj.q_perm": "model.safetensors",
1026
+ "model.layers.5.self_attn.k_proj.qweight": "model.safetensors",
1027
+ "model.layers.5.self_attn.k_proj.scales": "model.safetensors",
1028
+ "model.layers.5.self_attn.k_proj.zeros": "model.safetensors",
1029
+ "model.layers.5.self_attn.o_proj.channel_scale": "model.safetensors",
1030
+ "model.layers.5.self_attn.o_proj.q_perm": "model.safetensors",
1031
+ "model.layers.5.self_attn.o_proj.qweight": "model.safetensors",
1032
+ "model.layers.5.self_attn.o_proj.scales": "model.safetensors",
1033
+ "model.layers.5.self_attn.o_proj.zeros": "model.safetensors",
1034
+ "model.layers.5.self_attn.q_proj.channel_scale": "model.safetensors",
1035
+ "model.layers.5.self_attn.q_proj.q_perm": "model.safetensors",
1036
+ "model.layers.5.self_attn.q_proj.qweight": "model.safetensors",
1037
+ "model.layers.5.self_attn.q_proj.scales": "model.safetensors",
1038
+ "model.layers.5.self_attn.q_proj.zeros": "model.safetensors",
1039
+ "model.layers.5.self_attn.v_proj.channel_scale": "model.safetensors",
1040
+ "model.layers.5.self_attn.v_proj.q_perm": "model.safetensors",
1041
+ "model.layers.5.self_attn.v_proj.qweight": "model.safetensors",
1042
+ "model.layers.5.self_attn.v_proj.scales": "model.safetensors",
1043
+ "model.layers.5.self_attn.v_proj.zeros": "model.safetensors",
1044
+ "model.layers.6.input_layernorm.weight": "model.safetensors",
1045
+ "model.layers.6.mlp.down_proj.channel_scale": "model.safetensors",
1046
+ "model.layers.6.mlp.down_proj.q_perm": "model.safetensors",
1047
+ "model.layers.6.mlp.down_proj.qweight": "model.safetensors",
1048
+ "model.layers.6.mlp.down_proj.scales": "model.safetensors",
1049
+ "model.layers.6.mlp.down_proj.zeros": "model.safetensors",
1050
+ "model.layers.6.mlp.gate_proj.channel_scale": "model.safetensors",
1051
+ "model.layers.6.mlp.gate_proj.q_perm": "model.safetensors",
1052
+ "model.layers.6.mlp.gate_proj.qweight": "model.safetensors",
1053
+ "model.layers.6.mlp.gate_proj.scales": "model.safetensors",
1054
+ "model.layers.6.mlp.gate_proj.zeros": "model.safetensors",
1055
+ "model.layers.6.mlp.up_proj.channel_scale": "model.safetensors",
1056
+ "model.layers.6.mlp.up_proj.q_perm": "model.safetensors",
1057
+ "model.layers.6.mlp.up_proj.qweight": "model.safetensors",
1058
+ "model.layers.6.mlp.up_proj.scales": "model.safetensors",
1059
+ "model.layers.6.mlp.up_proj.zeros": "model.safetensors",
1060
+ "model.layers.6.post_attention_layernorm.weight": "model.safetensors",
1061
+ "model.layers.6.self_attn.k_proj.channel_scale": "model.safetensors",
1062
+ "model.layers.6.self_attn.k_proj.q_perm": "model.safetensors",
1063
+ "model.layers.6.self_attn.k_proj.qweight": "model.safetensors",
1064
+ "model.layers.6.self_attn.k_proj.scales": "model.safetensors",
1065
+ "model.layers.6.self_attn.k_proj.zeros": "model.safetensors",
1066
+ "model.layers.6.self_attn.o_proj.channel_scale": "model.safetensors",
1067
+ "model.layers.6.self_attn.o_proj.q_perm": "model.safetensors",
1068
+ "model.layers.6.self_attn.o_proj.qweight": "model.safetensors",
1069
+ "model.layers.6.self_attn.o_proj.scales": "model.safetensors",
1070
+ "model.layers.6.self_attn.o_proj.zeros": "model.safetensors",
1071
+ "model.layers.6.self_attn.q_proj.channel_scale": "model.safetensors",
1072
+ "model.layers.6.self_attn.q_proj.q_perm": "model.safetensors",
1073
+ "model.layers.6.self_attn.q_proj.qweight": "model.safetensors",
1074
+ "model.layers.6.self_attn.q_proj.scales": "model.safetensors",
1075
+ "model.layers.6.self_attn.q_proj.zeros": "model.safetensors",
1076
+ "model.layers.6.self_attn.v_proj.channel_scale": "model.safetensors",
1077
+ "model.layers.6.self_attn.v_proj.q_perm": "model.safetensors",
1078
+ "model.layers.6.self_attn.v_proj.qweight": "model.safetensors",
1079
+ "model.layers.6.self_attn.v_proj.scales": "model.safetensors",
1080
+ "model.layers.6.self_attn.v_proj.zeros": "model.safetensors",
1081
+ "model.layers.7.input_layernorm.weight": "model.safetensors",
1082
+ "model.layers.7.mlp.down_proj.channel_scale": "model.safetensors",
1083
+ "model.layers.7.mlp.down_proj.q_perm": "model.safetensors",
1084
+ "model.layers.7.mlp.down_proj.qweight": "model.safetensors",
1085
+ "model.layers.7.mlp.down_proj.scales": "model.safetensors",
1086
+ "model.layers.7.mlp.down_proj.zeros": "model.safetensors",
1087
+ "model.layers.7.mlp.gate_proj.channel_scale": "model.safetensors",
1088
+ "model.layers.7.mlp.gate_proj.q_perm": "model.safetensors",
1089
+ "model.layers.7.mlp.gate_proj.qweight": "model.safetensors",
1090
+ "model.layers.7.mlp.gate_proj.scales": "model.safetensors",
1091
+ "model.layers.7.mlp.gate_proj.zeros": "model.safetensors",
1092
+ "model.layers.7.mlp.up_proj.channel_scale": "model.safetensors",
1093
+ "model.layers.7.mlp.up_proj.q_perm": "model.safetensors",
1094
+ "model.layers.7.mlp.up_proj.qweight": "model.safetensors",
1095
+ "model.layers.7.mlp.up_proj.scales": "model.safetensors",
1096
+ "model.layers.7.mlp.up_proj.zeros": "model.safetensors",
1097
+ "model.layers.7.post_attention_layernorm.weight": "model.safetensors",
1098
+ "model.layers.7.self_attn.k_proj.channel_scale": "model.safetensors",
1099
+ "model.layers.7.self_attn.k_proj.q_perm": "model.safetensors",
1100
+ "model.layers.7.self_attn.k_proj.qweight": "model.safetensors",
1101
+ "model.layers.7.self_attn.k_proj.scales": "model.safetensors",
1102
+ "model.layers.7.self_attn.k_proj.zeros": "model.safetensors",
1103
+ "model.layers.7.self_attn.o_proj.channel_scale": "model.safetensors",
1104
+ "model.layers.7.self_attn.o_proj.q_perm": "model.safetensors",
1105
+ "model.layers.7.self_attn.o_proj.qweight": "model.safetensors",
1106
+ "model.layers.7.self_attn.o_proj.scales": "model.safetensors",
1107
+ "model.layers.7.self_attn.o_proj.zeros": "model.safetensors",
1108
+ "model.layers.7.self_attn.q_proj.channel_scale": "model.safetensors",
1109
+ "model.layers.7.self_attn.q_proj.q_perm": "model.safetensors",
1110
+ "model.layers.7.self_attn.q_proj.qweight": "model.safetensors",
1111
+ "model.layers.7.self_attn.q_proj.scales": "model.safetensors",
1112
+ "model.layers.7.self_attn.q_proj.zeros": "model.safetensors",
1113
+ "model.layers.7.self_attn.v_proj.channel_scale": "model.safetensors",
1114
+ "model.layers.7.self_attn.v_proj.q_perm": "model.safetensors",
1115
+ "model.layers.7.self_attn.v_proj.qweight": "model.safetensors",
1116
+ "model.layers.7.self_attn.v_proj.scales": "model.safetensors",
1117
+ "model.layers.7.self_attn.v_proj.zeros": "model.safetensors",
1118
+ "model.layers.8.input_layernorm.weight": "model.safetensors",
1119
+ "model.layers.8.mlp.down_proj.channel_scale": "model.safetensors",
1120
+ "model.layers.8.mlp.down_proj.q_perm": "model.safetensors",
1121
+ "model.layers.8.mlp.down_proj.qweight": "model.safetensors",
1122
+ "model.layers.8.mlp.down_proj.scales": "model.safetensors",
1123
+ "model.layers.8.mlp.down_proj.zeros": "model.safetensors",
1124
+ "model.layers.8.mlp.gate_proj.channel_scale": "model.safetensors",
1125
+ "model.layers.8.mlp.gate_proj.q_perm": "model.safetensors",
1126
+ "model.layers.8.mlp.gate_proj.qweight": "model.safetensors",
1127
+ "model.layers.8.mlp.gate_proj.scales": "model.safetensors",
1128
+ "model.layers.8.mlp.gate_proj.zeros": "model.safetensors",
1129
+ "model.layers.8.mlp.up_proj.channel_scale": "model.safetensors",
1130
+ "model.layers.8.mlp.up_proj.q_perm": "model.safetensors",
1131
+ "model.layers.8.mlp.up_proj.qweight": "model.safetensors",
1132
+ "model.layers.8.mlp.up_proj.scales": "model.safetensors",
1133
+ "model.layers.8.mlp.up_proj.zeros": "model.safetensors",
1134
+ "model.layers.8.post_attention_layernorm.weight": "model.safetensors",
1135
+ "model.layers.8.self_attn.k_proj.channel_scale": "model.safetensors",
1136
+ "model.layers.8.self_attn.k_proj.q_perm": "model.safetensors",
1137
+ "model.layers.8.self_attn.k_proj.qweight": "model.safetensors",
1138
+ "model.layers.8.self_attn.k_proj.scales": "model.safetensors",
1139
+ "model.layers.8.self_attn.k_proj.zeros": "model.safetensors",
1140
+ "model.layers.8.self_attn.o_proj.channel_scale": "model.safetensors",
1141
+ "model.layers.8.self_attn.o_proj.q_perm": "model.safetensors",
1142
+ "model.layers.8.self_attn.o_proj.qweight": "model.safetensors",
1143
+ "model.layers.8.self_attn.o_proj.scales": "model.safetensors",
1144
+ "model.layers.8.self_attn.o_proj.zeros": "model.safetensors",
1145
+ "model.layers.8.self_attn.q_proj.channel_scale": "model.safetensors",
1146
+ "model.layers.8.self_attn.q_proj.q_perm": "model.safetensors",
1147
+ "model.layers.8.self_attn.q_proj.qweight": "model.safetensors",
1148
+ "model.layers.8.self_attn.q_proj.scales": "model.safetensors",
1149
+ "model.layers.8.self_attn.q_proj.zeros": "model.safetensors",
1150
+ "model.layers.8.self_attn.v_proj.channel_scale": "model.safetensors",
1151
+ "model.layers.8.self_attn.v_proj.q_perm": "model.safetensors",
1152
+ "model.layers.8.self_attn.v_proj.qweight": "model.safetensors",
1153
+ "model.layers.8.self_attn.v_proj.scales": "model.safetensors",
1154
+ "model.layers.8.self_attn.v_proj.zeros": "model.safetensors",
1155
+ "model.layers.9.input_layernorm.weight": "model.safetensors",
1156
+ "model.layers.9.mlp.down_proj.channel_scale": "model.safetensors",
1157
+ "model.layers.9.mlp.down_proj.q_perm": "model.safetensors",
1158
+ "model.layers.9.mlp.down_proj.qweight": "model.safetensors",
1159
+ "model.layers.9.mlp.down_proj.scales": "model.safetensors",
1160
+ "model.layers.9.mlp.down_proj.zeros": "model.safetensors",
1161
+ "model.layers.9.mlp.gate_proj.channel_scale": "model.safetensors",
1162
+ "model.layers.9.mlp.gate_proj.q_perm": "model.safetensors",
1163
+ "model.layers.9.mlp.gate_proj.qweight": "model.safetensors",
1164
+ "model.layers.9.mlp.gate_proj.scales": "model.safetensors",
1165
+ "model.layers.9.mlp.gate_proj.zeros": "model.safetensors",
1166
+ "model.layers.9.mlp.up_proj.channel_scale": "model.safetensors",
1167
+ "model.layers.9.mlp.up_proj.q_perm": "model.safetensors",
1168
+ "model.layers.9.mlp.up_proj.qweight": "model.safetensors",
1169
+ "model.layers.9.mlp.up_proj.scales": "model.safetensors",
1170
+ "model.layers.9.mlp.up_proj.zeros": "model.safetensors",
1171
+ "model.layers.9.post_attention_layernorm.weight": "model.safetensors",
1172
+ "model.layers.9.self_attn.k_proj.channel_scale": "model.safetensors",
1173
+ "model.layers.9.self_attn.k_proj.q_perm": "model.safetensors",
1174
+ "model.layers.9.self_attn.k_proj.qweight": "model.safetensors",
1175
+ "model.layers.9.self_attn.k_proj.scales": "model.safetensors",
1176
+ "model.layers.9.self_attn.k_proj.zeros": "model.safetensors",
1177
+ "model.layers.9.self_attn.o_proj.channel_scale": "model.safetensors",
1178
+ "model.layers.9.self_attn.o_proj.q_perm": "model.safetensors",
1179
+ "model.layers.9.self_attn.o_proj.qweight": "model.safetensors",
1180
+ "model.layers.9.self_attn.o_proj.scales": "model.safetensors",
1181
+ "model.layers.9.self_attn.o_proj.zeros": "model.safetensors",
1182
+ "model.layers.9.self_attn.q_proj.channel_scale": "model.safetensors",
1183
+ "model.layers.9.self_attn.q_proj.q_perm": "model.safetensors",
1184
+ "model.layers.9.self_attn.q_proj.qweight": "model.safetensors",
1185
+ "model.layers.9.self_attn.q_proj.scales": "model.safetensors",
1186
+ "model.layers.9.self_attn.q_proj.zeros": "model.safetensors",
1187
+ "model.layers.9.self_attn.v_proj.channel_scale": "model.safetensors",
1188
+ "model.layers.9.self_attn.v_proj.q_perm": "model.safetensors",
1189
+ "model.layers.9.self_attn.v_proj.qweight": "model.safetensors",
1190
+ "model.layers.9.self_attn.v_proj.scales": "model.safetensors",
1191
+ "model.layers.9.self_attn.v_proj.zeros": "model.safetensors",
1192
+ "model.norm.weight": "model.safetensors"
1193
+ }
1194
+ }
quant_strategy.json ADDED
@@ -0,0 +1,2820 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "measurement": {
3
+ "model.layers.0": {
4
+ "accuracy": 0.8398913542429606,
5
+ "total_bits": 491152896,
6
+ "q_proj": {
7
+ "group_size": {
8
+ "2": 64
9
+ },
10
+ "bits": [
11
+ 2
12
+ ],
13
+ "bits_prop": [
14
+ 1
15
+ ],
16
+ "scale_bits": 4
17
+ },
18
+ "k_proj": {
19
+ "group_size": {
20
+ "2": 64
21
+ },
22
+ "bits": [
23
+ 2
24
+ ],
25
+ "bits_prop": [
26
+ 1
27
+ ],
28
+ "scale_bits": 4
29
+ },
30
+ "v_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
+ "o_proj": {
43
+ "group_size": {
44
+ "4": 128
45
+ },
46
+ "bits": [
47
+ 4
48
+ ],
49
+ "bits_prop": [
50
+ 1
51
+ ],
52
+ "scale_bits": 4
53
+ },
54
+ "up_proj": {
55
+ "group_size": {
56
+ "2": 64
57
+ },
58
+ "bits": [
59
+ 2
60
+ ],
61
+ "bits_prop": [
62
+ 1
63
+ ],
64
+ "scale_bits": 4
65
+ },
66
+ "gate_proj": {
67
+ "group_size": {
68
+ "2": 64
69
+ },
70
+ "bits": [
71
+ 2
72
+ ],
73
+ "bits_prop": [
74
+ 1
75
+ ],
76
+ "scale_bits": 4
77
+ },
78
+ "down_proj": {
79
+ "group_size": {
80
+ "2": 64
81
+ },
82
+ "bits": [
83
+ 2
84
+ ],
85
+ "bits_prop": [
86
+ 1
87
+ ],
88
+ "scale_bits": 4
89
+ }
90
+ },
91
+ "model.layers.1": {
92
+ "accuracy": 0.905671755472819,
93
+ "total_bits": 573724416,
94
+ "q_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
+ "k_proj": {
107
+ "group_size": {
108
+ "2": 64
109
+ },
110
+ "bits": [
111
+ 2
112
+ ],
113
+ "bits_prop": [
114
+ 1
115
+ ],
116
+ "scale_bits": 4
117
+ },
118
+ "v_proj": {
119
+ "group_size": {
120
+ "4": 128
121
+ },
122
+ "bits": [
123
+ 4
124
+ ],
125
+ "bits_prop": [
126
+ 1
127
+ ],
128
+ "scale_bits": 4
129
+ },
130
+ "o_proj": {
131
+ "group_size": {
132
+ "2": 64
133
+ },
134
+ "bits": [
135
+ 2
136
+ ],
137
+ "bits_prop": [
138
+ 1
139
+ ],
140
+ "scale_bits": 4
141
+ },
142
+ "up_proj": {
143
+ "group_size": {
144
+ "2": 64
145
+ },
146
+ "bits": [
147
+ 2
148
+ ],
149
+ "bits_prop": [
150
+ 1
151
+ ],
152
+ "scale_bits": 4
153
+ },
154
+ "gate_proj": {
155
+ "group_size": {
156
+ "2": 64
157
+ },
158
+ "bits": [
159
+ 2
160
+ ],
161
+ "bits_prop": [
162
+ 1
163
+ ],
164
+ "scale_bits": 4
165
+ },
166
+ "down_proj": {
167
+ "group_size": {
168
+ "4": 128
169
+ },
170
+ "bits": [
171
+ 4
172
+ ],
173
+ "bits_prop": [
174
+ 1
175
+ ],
176
+ "scale_bits": 4
177
+ }
178
+ },
179
+ "model.layers.2": {
180
+ "accuracy": 0.9026820063591003,
181
+ "total_bits": 458124288,
182
+ "q_proj": {
183
+ "group_size": {
184
+ "2": 64
185
+ },
186
+ "bits": [
187
+ 2
188
+ ],
189
+ "bits_prop": [
190
+ 1
191
+ ],
192
+ "scale_bits": 4
193
+ },
194
+ "k_proj": {
195
+ "group_size": {
196
+ "2": 64
197
+ },
198
+ "bits": [
199
+ 2
200
+ ],
201
+ "bits_prop": [
202
+ 1
203
+ ],
204
+ "scale_bits": 4
205
+ },
206
+ "v_proj": {
207
+ "group_size": {
208
+ "4": 128
209
+ },
210
+ "bits": [
211
+ 4
212
+ ],
213
+ "bits_prop": [
214
+ 1
215
+ ],
216
+ "scale_bits": 4
217
+ },
218
+ "o_proj": {
219
+ "group_size": {
220
+ "2": 64
221
+ },
222
+ "bits": [
223
+ 2
224
+ ],
225
+ "bits_prop": [
226
+ 1
227
+ ],
228
+ "scale_bits": 4
229
+ },
230
+ "up_proj": {
231
+ "group_size": {
232
+ "2": 64
233
+ },
234
+ "bits": [
235
+ 2
236
+ ],
237
+ "bits_prop": [
238
+ 1
239
+ ],
240
+ "scale_bits": 4
241
+ },
242
+ "gate_proj": {
243
+ "group_size": {
244
+ "2": 64
245
+ },
246
+ "bits": [
247
+ 2
248
+ ],
249
+ "bits_prop": [
250
+ 1
251
+ ],
252
+ "scale_bits": 4
253
+ },
254
+ "down_proj": {
255
+ "group_size": {
256
+ "2": 64
257
+ },
258
+ "bits": [
259
+ 2
260
+ ],
261
+ "bits_prop": [
262
+ 1
263
+ ],
264
+ "scale_bits": 4
265
+ }
266
+ },
267
+ "model.layers.3": {
268
+ "accuracy": 0.8770348230997721,
269
+ "total_bits": 458124288,
270
+ "q_proj": {
271
+ "group_size": {
272
+ "2": 64
273
+ },
274
+ "bits": [
275
+ 2
276
+ ],
277
+ "bits_prop": [
278
+ 1
279
+ ],
280
+ "scale_bits": 4
281
+ },
282
+ "k_proj": {
283
+ "group_size": {
284
+ "2": 64
285
+ },
286
+ "bits": [
287
+ 2
288
+ ],
289
+ "bits_prop": [
290
+ 1
291
+ ],
292
+ "scale_bits": 4
293
+ },
294
+ "v_proj": {
295
+ "group_size": {
296
+ "4": 128
297
+ },
298
+ "bits": [
299
+ 4
300
+ ],
301
+ "bits_prop": [
302
+ 1
303
+ ],
304
+ "scale_bits": 4
305
+ },
306
+ "o_proj": {
307
+ "group_size": {
308
+ "2": 64
309
+ },
310
+ "bits": [
311
+ 2
312
+ ],
313
+ "bits_prop": [
314
+ 1
315
+ ],
316
+ "scale_bits": 4
317
+ },
318
+ "up_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
+ "gate_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
+ "down_proj": {
343
+ "group_size": {
344
+ "2": 64
345
+ },
346
+ "bits": [
347
+ 2
348
+ ],
349
+ "bits_prop": [
350
+ 1
351
+ ],
352
+ "scale_bits": 4
353
+ }
354
+ },
355
+ "model.layers.4": {
356
+ "accuracy": 0.8886063098907471,
357
+ "total_bits": 458124288,
358
+ "q_proj": {
359
+ "group_size": {
360
+ "2": 64
361
+ },
362
+ "bits": [
363
+ 2
364
+ ],
365
+ "bits_prop": [
366
+ 1
367
+ ],
368
+ "scale_bits": 4
369
+ },
370
+ "k_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
+ "v_proj": {
383
+ "group_size": {
384
+ "4": 128
385
+ },
386
+ "bits": [
387
+ 4
388
+ ],
389
+ "bits_prop": [
390
+ 1
391
+ ],
392
+ "scale_bits": 4
393
+ },
394
+ "o_proj": {
395
+ "group_size": {
396
+ "2": 64
397
+ },
398
+ "bits": [
399
+ 2
400
+ ],
401
+ "bits_prop": [
402
+ 1
403
+ ],
404
+ "scale_bits": 4
405
+ },
406
+ "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
+ "gate_proj": {
419
+ "group_size": {
420
+ "2": 64
421
+ },
422
+ "bits": [
423
+ 2
424
+ ],
425
+ "bits_prop": [
426
+ 1
427
+ ],
428
+ "scale_bits": 4
429
+ },
430
+ "down_proj": {
431
+ "group_size": {
432
+ "2": 64
433
+ },
434
+ "bits": [
435
+ 2
436
+ ],
437
+ "bits_prop": [
438
+ 1
439
+ ],
440
+ "scale_bits": 4
441
+ }
442
+ },
443
+ "model.layers.5": {
444
+ "accuracy": 0.8565507729848225,
445
+ "total_bits": 491152896,
446
+ "q_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
+ "k_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
+ "v_proj": {
471
+ "group_size": {
472
+ "4": 128
473
+ },
474
+ "bits": [
475
+ 4
476
+ ],
477
+ "bits_prop": [
478
+ 1
479
+ ],
480
+ "scale_bits": 4
481
+ },
482
+ "o_proj": {
483
+ "group_size": {
484
+ "4": 128
485
+ },
486
+ "bits": [
487
+ 4
488
+ ],
489
+ "bits_prop": [
490
+ 1
491
+ ],
492
+ "scale_bits": 4
493
+ },
494
+ "up_proj": {
495
+ "group_size": {
496
+ "2": 64
497
+ },
498
+ "bits": [
499
+ 2
500
+ ],
501
+ "bits_prop": [
502
+ 1
503
+ ],
504
+ "scale_bits": 4
505
+ },
506
+ "gate_proj": {
507
+ "group_size": {
508
+ "2": 64
509
+ },
510
+ "bits": [
511
+ 2
512
+ ],
513
+ "bits_prop": [
514
+ 1
515
+ ],
516
+ "scale_bits": 4
517
+ },
518
+ "down_proj": {
519
+ "group_size": {
520
+ "2": 64
521
+ },
522
+ "bits": [
523
+ 2
524
+ ],
525
+ "bits_prop": [
526
+ 1
527
+ ],
528
+ "scale_bits": 4
529
+ }
530
+ },
531
+ "model.layers.6": {
532
+ "accuracy": 0.8628210226694744,
533
+ "total_bits": 466380288,
534
+ "q_proj": {
535
+ "group_size": {
536
+ "2": 64
537
+ },
538
+ "bits": [
539
+ 2
540
+ ],
541
+ "bits_prop": [
542
+ 1
543
+ ],
544
+ "scale_bits": 4
545
+ },
546
+ "k_proj": {
547
+ "group_size": {
548
+ "4": 128
549
+ },
550
+ "bits": [
551
+ 4
552
+ ],
553
+ "bits_prop": [
554
+ 1
555
+ ],
556
+ "scale_bits": 4
557
+ },
558
+ "v_proj": {
559
+ "group_size": {
560
+ "4": 128
561
+ },
562
+ "bits": [
563
+ 4
564
+ ],
565
+ "bits_prop": [
566
+ 1
567
+ ],
568
+ "scale_bits": 4
569
+ },
570
+ "o_proj": {
571
+ "group_size": {
572
+ "2": 64
573
+ },
574
+ "bits": [
575
+ 2
576
+ ],
577
+ "bits_prop": [
578
+ 1
579
+ ],
580
+ "scale_bits": 4
581
+ },
582
+ "up_proj": {
583
+ "group_size": {
584
+ "2": 64
585
+ },
586
+ "bits": [
587
+ 2
588
+ ],
589
+ "bits_prop": [
590
+ 1
591
+ ],
592
+ "scale_bits": 4
593
+ },
594
+ "gate_proj": {
595
+ "group_size": {
596
+ "2": 64
597
+ },
598
+ "bits": [
599
+ 2
600
+ ],
601
+ "bits_prop": [
602
+ 1
603
+ ],
604
+ "scale_bits": 4
605
+ },
606
+ "down_proj": {
607
+ "group_size": {
608
+ "2": 64
609
+ },
610
+ "bits": [
611
+ 2
612
+ ],
613
+ "bits_prop": [
614
+ 1
615
+ ],
616
+ "scale_bits": 4
617
+ }
618
+ },
619
+ "model.layers.7": {
620
+ "accuracy": 0.844833215077718,
621
+ "total_bits": 466380288,
622
+ "q_proj": {
623
+ "group_size": {
624
+ "2": 64
625
+ },
626
+ "bits": [
627
+ 2
628
+ ],
629
+ "bits_prop": [
630
+ 1
631
+ ],
632
+ "scale_bits": 4
633
+ },
634
+ "k_proj": {
635
+ "group_size": {
636
+ "4": 128
637
+ },
638
+ "bits": [
639
+ 4
640
+ ],
641
+ "bits_prop": [
642
+ 1
643
+ ],
644
+ "scale_bits": 4
645
+ },
646
+ "v_proj": {
647
+ "group_size": {
648
+ "4": 128
649
+ },
650
+ "bits": [
651
+ 4
652
+ ],
653
+ "bits_prop": [
654
+ 1
655
+ ],
656
+ "scale_bits": 4
657
+ },
658
+ "o_proj": {
659
+ "group_size": {
660
+ "2": 64
661
+ },
662
+ "bits": [
663
+ 2
664
+ ],
665
+ "bits_prop": [
666
+ 1
667
+ ],
668
+ "scale_bits": 4
669
+ },
670
+ "up_proj": {
671
+ "group_size": {
672
+ "2": 64
673
+ },
674
+ "bits": [
675
+ 2
676
+ ],
677
+ "bits_prop": [
678
+ 1
679
+ ],
680
+ "scale_bits": 4
681
+ },
682
+ "gate_proj": {
683
+ "group_size": {
684
+ "2": 64
685
+ },
686
+ "bits": [
687
+ 2
688
+ ],
689
+ "bits_prop": [
690
+ 1
691
+ ],
692
+ "scale_bits": 4
693
+ },
694
+ "down_proj": {
695
+ "group_size": {
696
+ "2": 64
697
+ },
698
+ "bits": [
699
+ 2
700
+ ],
701
+ "bits_prop": [
702
+ 1
703
+ ],
704
+ "scale_bits": 4
705
+ }
706
+ },
707
+ "model.layers.8": {
708
+ "accuracy": 0.8500169912974039,
709
+ "total_bits": 466380288,
710
+ "q_proj": {
711
+ "group_size": {
712
+ "2": 64
713
+ },
714
+ "bits": [
715
+ 2
716
+ ],
717
+ "bits_prop": [
718
+ 1
719
+ ],
720
+ "scale_bits": 4
721
+ },
722
+ "k_proj": {
723
+ "group_size": {
724
+ "4": 128
725
+ },
726
+ "bits": [
727
+ 4
728
+ ],
729
+ "bits_prop": [
730
+ 1
731
+ ],
732
+ "scale_bits": 4
733
+ },
734
+ "v_proj": {
735
+ "group_size": {
736
+ "4": 128
737
+ },
738
+ "bits": [
739
+ 4
740
+ ],
741
+ "bits_prop": [
742
+ 1
743
+ ],
744
+ "scale_bits": 4
745
+ },
746
+ "o_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
+ "up_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_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
+ "down_proj": {
783
+ "group_size": {
784
+ "2": 64
785
+ },
786
+ "bits": [
787
+ 2
788
+ ],
789
+ "bits_prop": [
790
+ 1
791
+ ],
792
+ "scale_bits": 4
793
+ }
794
+ },
795
+ "model.layers.9": {
796
+ "accuracy": 0.8500173886617024,
797
+ "total_bits": 466380288,
798
+ "q_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
+ "k_proj": {
811
+ "group_size": {
812
+ "4": 128
813
+ },
814
+ "bits": [
815
+ 4
816
+ ],
817
+ "bits_prop": [
818
+ 1
819
+ ],
820
+ "scale_bits": 4
821
+ },
822
+ "v_proj": {
823
+ "group_size": {
824
+ "4": 128
825
+ },
826
+ "bits": [
827
+ 4
828
+ ],
829
+ "bits_prop": [
830
+ 1
831
+ ],
832
+ "scale_bits": 4
833
+ },
834
+ "o_proj": {
835
+ "group_size": {
836
+ "2": 64
837
+ },
838
+ "bits": [
839
+ 2
840
+ ],
841
+ "bits_prop": [
842
+ 1
843
+ ],
844
+ "scale_bits": 4
845
+ },
846
+ "up_proj": {
847
+ "group_size": {
848
+ "2": 64
849
+ },
850
+ "bits": [
851
+ 2
852
+ ],
853
+ "bits_prop": [
854
+ 1
855
+ ],
856
+ "scale_bits": 4
857
+ },
858
+ "gate_proj": {
859
+ "group_size": {
860
+ "2": 64
861
+ },
862
+ "bits": [
863
+ 2
864
+ ],
865
+ "bits_prop": [
866
+ 1
867
+ ],
868
+ "scale_bits": 4
869
+ },
870
+ "down_proj": {
871
+ "group_size": {
872
+ "2": 64
873
+ },
874
+ "bits": [
875
+ 2
876
+ ],
877
+ "bits_prop": [
878
+ 1
879
+ ],
880
+ "scale_bits": 4
881
+ }
882
+ },
883
+ "model.layers.10": {
884
+ "accuracy": 0.8562391599019368,
885
+ "total_bits": 491152896,
886
+ "q_proj": {
887
+ "group_size": {
888
+ "2": 64
889
+ },
890
+ "bits": [
891
+ 2
892
+ ],
893
+ "bits_prop": [
894
+ 1
895
+ ],
896
+ "scale_bits": 4
897
+ },
898
+ "k_proj": {
899
+ "group_size": {
900
+ "2": 64
901
+ },
902
+ "bits": [
903
+ 2
904
+ ],
905
+ "bits_prop": [
906
+ 1
907
+ ],
908
+ "scale_bits": 4
909
+ },
910
+ "v_proj": {
911
+ "group_size": {
912
+ "4": 128
913
+ },
914
+ "bits": [
915
+ 4
916
+ ],
917
+ "bits_prop": [
918
+ 1
919
+ ],
920
+ "scale_bits": 4
921
+ },
922
+ "o_proj": {
923
+ "group_size": {
924
+ "4": 128
925
+ },
926
+ "bits": [
927
+ 4
928
+ ],
929
+ "bits_prop": [
930
+ 1
931
+ ],
932
+ "scale_bits": 4
933
+ },
934
+ "up_proj": {
935
+ "group_size": {
936
+ "2": 64
937
+ },
938
+ "bits": [
939
+ 2
940
+ ],
941
+ "bits_prop": [
942
+ 1
943
+ ],
944
+ "scale_bits": 4
945
+ },
946
+ "gate_proj": {
947
+ "group_size": {
948
+ "2": 64
949
+ },
950
+ "bits": [
951
+ 2
952
+ ],
953
+ "bits_prop": [
954
+ 1
955
+ ],
956
+ "scale_bits": 4
957
+ },
958
+ "down_proj": {
959
+ "group_size": {
960
+ "2": 64
961
+ },
962
+ "bits": [
963
+ 2
964
+ ],
965
+ "bits_prop": [
966
+ 1
967
+ ],
968
+ "scale_bits": 4
969
+ }
970
+ },
971
+ "model.layers.11": {
972
+ "accuracy": 0.847865104675293,
973
+ "total_bits": 491152896,
974
+ "q_proj": {
975
+ "group_size": {
976
+ "2": 64
977
+ },
978
+ "bits": [
979
+ 2
980
+ ],
981
+ "bits_prop": [
982
+ 1
983
+ ],
984
+ "scale_bits": 4
985
+ },
986
+ "k_proj": {
987
+ "group_size": {
988
+ "2": 64
989
+ },
990
+ "bits": [
991
+ 2
992
+ ],
993
+ "bits_prop": [
994
+ 1
995
+ ],
996
+ "scale_bits": 4
997
+ },
998
+ "v_proj": {
999
+ "group_size": {
1000
+ "4": 128
1001
+ },
1002
+ "bits": [
1003
+ 4
1004
+ ],
1005
+ "bits_prop": [
1006
+ 1
1007
+ ],
1008
+ "scale_bits": 4
1009
+ },
1010
+ "o_proj": {
1011
+ "group_size": {
1012
+ "4": 128
1013
+ },
1014
+ "bits": [
1015
+ 4
1016
+ ],
1017
+ "bits_prop": [
1018
+ 1
1019
+ ],
1020
+ "scale_bits": 4
1021
+ },
1022
+ "up_proj": {
1023
+ "group_size": {
1024
+ "2": 64
1025
+ },
1026
+ "bits": [
1027
+ 2
1028
+ ],
1029
+ "bits_prop": [
1030
+ 1
1031
+ ],
1032
+ "scale_bits": 4
1033
+ },
1034
+ "gate_proj": {
1035
+ "group_size": {
1036
+ "2": 64
1037
+ },
1038
+ "bits": [
1039
+ 2
1040
+ ],
1041
+ "bits_prop": [
1042
+ 1
1043
+ ],
1044
+ "scale_bits": 4
1045
+ },
1046
+ "down_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
+ },
1059
+ "model.layers.12": {
1060
+ "accuracy": 0.907697836558024,
1061
+ "total_bits": 615009024,
1062
+ "q_proj": {
1063
+ "group_size": {
1064
+ "2": 64
1065
+ },
1066
+ "bits": [
1067
+ 2
1068
+ ],
1069
+ "bits_prop": [
1070
+ 1
1071
+ ],
1072
+ "scale_bits": 4
1073
+ },
1074
+ "k_proj": {
1075
+ "group_size": {
1076
+ "4": 128
1077
+ },
1078
+ "bits": [
1079
+ 4
1080
+ ],
1081
+ "bits_prop": [
1082
+ 1
1083
+ ],
1084
+ "scale_bits": 4
1085
+ },
1086
+ "v_proj": {
1087
+ "group_size": {
1088
+ "4": 128
1089
+ },
1090
+ "bits": [
1091
+ 4
1092
+ ],
1093
+ "bits_prop": [
1094
+ 1
1095
+ ],
1096
+ "scale_bits": 4
1097
+ },
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
+ "up_proj": {
1111
+ "group_size": {
1112
+ "2": 64
1113
+ },
1114
+ "bits": [
1115
+ 2
1116
+ ],
1117
+ "bits_prop": [
1118
+ 1
1119
+ ],
1120
+ "scale_bits": 4
1121
+ },
1122
+ "gate_proj": {
1123
+ "group_size": {
1124
+ "2": 64
1125
+ },
1126
+ "bits": [
1127
+ 2
1128
+ ],
1129
+ "bits_prop": [
1130
+ 1
1131
+ ],
1132
+ "scale_bits": 4
1133
+ },
1134
+ "down_proj": {
1135
+ "group_size": {
1136
+ "4": 128
1137
+ },
1138
+ "bits": [
1139
+ 4
1140
+ ],
1141
+ "bits_prop": [
1142
+ 1
1143
+ ],
1144
+ "scale_bits": 4
1145
+ }
1146
+ },
1147
+ "model.layers.13": {
1148
+ "accuracy": 0.9013168017069499,
1149
+ "total_bits": 606753024,
1150
+ "q_proj": {
1151
+ "group_size": {
1152
+ "2": 64
1153
+ },
1154
+ "bits": [
1155
+ 2
1156
+ ],
1157
+ "bits_prop": [
1158
+ 1
1159
+ ],
1160
+ "scale_bits": 4
1161
+ },
1162
+ "k_proj": {
1163
+ "group_size": {
1164
+ "2": 64
1165
+ },
1166
+ "bits": [
1167
+ 2
1168
+ ],
1169
+ "bits_prop": [
1170
+ 1
1171
+ ],
1172
+ "scale_bits": 4
1173
+ },
1174
+ "v_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
+ "o_proj": {
1187
+ "group_size": {
1188
+ "4": 128
1189
+ },
1190
+ "bits": [
1191
+ 4
1192
+ ],
1193
+ "bits_prop": [
1194
+ 1
1195
+ ],
1196
+ "scale_bits": 4
1197
+ },
1198
+ "up_proj": {
1199
+ "group_size": {
1200
+ "2": 64
1201
+ },
1202
+ "bits": [
1203
+ 2
1204
+ ],
1205
+ "bits_prop": [
1206
+ 1
1207
+ ],
1208
+ "scale_bits": 4
1209
+ },
1210
+ "gate_proj": {
1211
+ "group_size": {
1212
+ "2": 64
1213
+ },
1214
+ "bits": [
1215
+ 2
1216
+ ],
1217
+ "bits_prop": [
1218
+ 1
1219
+ ],
1220
+ "scale_bits": 4
1221
+ },
1222
+ "down_proj": {
1223
+ "group_size": {
1224
+ "4": 128
1225
+ },
1226
+ "bits": [
1227
+ 4
1228
+ ],
1229
+ "bits_prop": [
1230
+ 1
1231
+ ],
1232
+ "scale_bits": 4
1233
+ }
1234
+ },
1235
+ "model.layers.14": {
1236
+ "accuracy": 0.9045030275980632,
1237
+ "total_bits": 606753024,
1238
+ "q_proj": {
1239
+ "group_size": {
1240
+ "2": 64
1241
+ },
1242
+ "bits": [
1243
+ 2
1244
+ ],
1245
+ "bits_prop": [
1246
+ 1
1247
+ ],
1248
+ "scale_bits": 4
1249
+ },
1250
+ "k_proj": {
1251
+ "group_size": {
1252
+ "2": 64
1253
+ },
1254
+ "bits": [
1255
+ 2
1256
+ ],
1257
+ "bits_prop": [
1258
+ 1
1259
+ ],
1260
+ "scale_bits": 4
1261
+ },
1262
+ "v_proj": {
1263
+ "group_size": {
1264
+ "4": 128
1265
+ },
1266
+ "bits": [
1267
+ 4
1268
+ ],
1269
+ "bits_prop": [
1270
+ 1
1271
+ ],
1272
+ "scale_bits": 4
1273
+ },
1274
+ "o_proj": {
1275
+ "group_size": {
1276
+ "4": 128
1277
+ },
1278
+ "bits": [
1279
+ 4
1280
+ ],
1281
+ "bits_prop": [
1282
+ 1
1283
+ ],
1284
+ "scale_bits": 4
1285
+ },
1286
+ "up_proj": {
1287
+ "group_size": {
1288
+ "2": 64
1289
+ },
1290
+ "bits": [
1291
+ 2
1292
+ ],
1293
+ "bits_prop": [
1294
+ 1
1295
+ ],
1296
+ "scale_bits": 4
1297
+ },
1298
+ "gate_proj": {
1299
+ "group_size": {
1300
+ "2": 64
1301
+ },
1302
+ "bits": [
1303
+ 2
1304
+ ],
1305
+ "bits_prop": [
1306
+ 1
1307
+ ],
1308
+ "scale_bits": 4
1309
+ },
1310
+ "down_proj": {
1311
+ "group_size": {
1312
+ "4": 128
1313
+ },
1314
+ "bits": [
1315
+ 4
1316
+ ],
1317
+ "bits_prop": [
1318
+ 1
1319
+ ],
1320
+ "scale_bits": 4
1321
+ }
1322
+ },
1323
+ "model.layers.15": {
1324
+ "accuracy": 0.9138115247090658,
1325
+ "total_bits": 648037632,
1326
+ "q_proj": {
1327
+ "group_size": {
1328
+ "4": 128
1329
+ },
1330
+ "bits": [
1331
+ 4
1332
+ ],
1333
+ "bits_prop": [
1334
+ 1
1335
+ ],
1336
+ "scale_bits": 4
1337
+ },
1338
+ "k_proj": {
1339
+ "group_size": {
1340
+ "4": 128
1341
+ },
1342
+ "bits": [
1343
+ 4
1344
+ ],
1345
+ "bits_prop": [
1346
+ 1
1347
+ ],
1348
+ "scale_bits": 4
1349
+ },
1350
+ "v_proj": {
1351
+ "group_size": {
1352
+ "4": 128
1353
+ },
1354
+ "bits": [
1355
+ 4
1356
+ ],
1357
+ "bits_prop": [
1358
+ 1
1359
+ ],
1360
+ "scale_bits": 4
1361
+ },
1362
+ "o_proj": {
1363
+ "group_size": {
1364
+ "4": 128
1365
+ },
1366
+ "bits": [
1367
+ 4
1368
+ ],
1369
+ "bits_prop": [
1370
+ 1
1371
+ ],
1372
+ "scale_bits": 4
1373
+ },
1374
+ "up_proj": {
1375
+ "group_size": {
1376
+ "2": 64
1377
+ },
1378
+ "bits": [
1379
+ 2
1380
+ ],
1381
+ "bits_prop": [
1382
+ 1
1383
+ ],
1384
+ "scale_bits": 4
1385
+ },
1386
+ "gate_proj": {
1387
+ "group_size": {
1388
+ "2": 64
1389
+ },
1390
+ "bits": [
1391
+ 2
1392
+ ],
1393
+ "bits_prop": [
1394
+ 1
1395
+ ],
1396
+ "scale_bits": 4
1397
+ },
1398
+ "down_proj": {
1399
+ "group_size": {
1400
+ "4": 128
1401
+ },
1402
+ "bits": [
1403
+ 4
1404
+ ],
1405
+ "bits_prop": [
1406
+ 1
1407
+ ],
1408
+ "scale_bits": 4
1409
+ }
1410
+ },
1411
+ "model.layers.16": {
1412
+ "accuracy": 0.8774760564168294,
1413
+ "total_bits": 573724416,
1414
+ "q_proj": {
1415
+ "group_size": {
1416
+ "2": 64
1417
+ },
1418
+ "bits": [
1419
+ 2
1420
+ ],
1421
+ "bits_prop": [
1422
+ 1
1423
+ ],
1424
+ "scale_bits": 4
1425
+ },
1426
+ "k_proj": {
1427
+ "group_size": {
1428
+ "2": 64
1429
+ },
1430
+ "bits": [
1431
+ 2
1432
+ ],
1433
+ "bits_prop": [
1434
+ 1
1435
+ ],
1436
+ "scale_bits": 4
1437
+ },
1438
+ "v_proj": {
1439
+ "group_size": {
1440
+ "4": 128
1441
+ },
1442
+ "bits": [
1443
+ 4
1444
+ ],
1445
+ "bits_prop": [
1446
+ 1
1447
+ ],
1448
+ "scale_bits": 4
1449
+ },
1450
+ "o_proj": {
1451
+ "group_size": {
1452
+ "2": 64
1453
+ },
1454
+ "bits": [
1455
+ 2
1456
+ ],
1457
+ "bits_prop": [
1458
+ 1
1459
+ ],
1460
+ "scale_bits": 4
1461
+ },
1462
+ "up_proj": {
1463
+ "group_size": {
1464
+ "2": 64
1465
+ },
1466
+ "bits": [
1467
+ 2
1468
+ ],
1469
+ "bits_prop": [
1470
+ 1
1471
+ ],
1472
+ "scale_bits": 4
1473
+ },
1474
+ "gate_proj": {
1475
+ "group_size": {
1476
+ "2": 64
1477
+ },
1478
+ "bits": [
1479
+ 2
1480
+ ],
1481
+ "bits_prop": [
1482
+ 1
1483
+ ],
1484
+ "scale_bits": 4
1485
+ },
1486
+ "down_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
+ },
1499
+ "model.layers.17": {
1500
+ "accuracy": 0.8854626019795735,
1501
+ "total_bits": 573724416,
1502
+ "q_proj": {
1503
+ "group_size": {
1504
+ "2": 64
1505
+ },
1506
+ "bits": [
1507
+ 2
1508
+ ],
1509
+ "bits_prop": [
1510
+ 1
1511
+ ],
1512
+ "scale_bits": 4
1513
+ },
1514
+ "k_proj": {
1515
+ "group_size": {
1516
+ "2": 64
1517
+ },
1518
+ "bits": [
1519
+ 2
1520
+ ],
1521
+ "bits_prop": [
1522
+ 1
1523
+ ],
1524
+ "scale_bits": 4
1525
+ },
1526
+ "v_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
+ "o_proj": {
1539
+ "group_size": {
1540
+ "2": 64
1541
+ },
1542
+ "bits": [
1543
+ 2
1544
+ ],
1545
+ "bits_prop": [
1546
+ 1
1547
+ ],
1548
+ "scale_bits": 4
1549
+ },
1550
+ "up_proj": {
1551
+ "group_size": {
1552
+ "2": 64
1553
+ },
1554
+ "bits": [
1555
+ 2
1556
+ ],
1557
+ "bits_prop": [
1558
+ 1
1559
+ ],
1560
+ "scale_bits": 4
1561
+ },
1562
+ "gate_proj": {
1563
+ "group_size": {
1564
+ "2": 64
1565
+ },
1566
+ "bits": [
1567
+ 2
1568
+ ],
1569
+ "bits_prop": [
1570
+ 1
1571
+ ],
1572
+ "scale_bits": 4
1573
+ },
1574
+ "down_proj": {
1575
+ "group_size": {
1576
+ "4": 128
1577
+ },
1578
+ "bits": [
1579
+ 4
1580
+ ],
1581
+ "bits_prop": [
1582
+ 1
1583
+ ],
1584
+ "scale_bits": 4
1585
+ }
1586
+ },
1587
+ "model.layers.18": {
1588
+ "accuracy": 0.882357120513916,
1589
+ "total_bits": 573724416,
1590
+ "q_proj": {
1591
+ "group_size": {
1592
+ "2": 64
1593
+ },
1594
+ "bits": [
1595
+ 2
1596
+ ],
1597
+ "bits_prop": [
1598
+ 1
1599
+ ],
1600
+ "scale_bits": 4
1601
+ },
1602
+ "k_proj": {
1603
+ "group_size": {
1604
+ "2": 64
1605
+ },
1606
+ "bits": [
1607
+ 2
1608
+ ],
1609
+ "bits_prop": [
1610
+ 1
1611
+ ],
1612
+ "scale_bits": 4
1613
+ },
1614
+ "v_proj": {
1615
+ "group_size": {
1616
+ "4": 128
1617
+ },
1618
+ "bits": [
1619
+ 4
1620
+ ],
1621
+ "bits_prop": [
1622
+ 1
1623
+ ],
1624
+ "scale_bits": 4
1625
+ },
1626
+ "o_proj": {
1627
+ "group_size": {
1628
+ "2": 64
1629
+ },
1630
+ "bits": [
1631
+ 2
1632
+ ],
1633
+ "bits_prop": [
1634
+ 1
1635
+ ],
1636
+ "scale_bits": 4
1637
+ },
1638
+ "up_proj": {
1639
+ "group_size": {
1640
+ "2": 64
1641
+ },
1642
+ "bits": [
1643
+ 2
1644
+ ],
1645
+ "bits_prop": [
1646
+ 1
1647
+ ],
1648
+ "scale_bits": 4
1649
+ },
1650
+ "gate_proj": {
1651
+ "group_size": {
1652
+ "2": 64
1653
+ },
1654
+ "bits": [
1655
+ 2
1656
+ ],
1657
+ "bits_prop": [
1658
+ 1
1659
+ ],
1660
+ "scale_bits": 4
1661
+ },
1662
+ "down_proj": {
1663
+ "group_size": {
1664
+ "4": 128
1665
+ },
1666
+ "bits": [
1667
+ 4
1668
+ ],
1669
+ "bits_prop": [
1670
+ 1
1671
+ ],
1672
+ "scale_bits": 4
1673
+ }
1674
+ },
1675
+ "model.layers.19": {
1676
+ "accuracy": 0.870351791381836,
1677
+ "total_bits": 573724416,
1678
+ "q_proj": {
1679
+ "group_size": {
1680
+ "2": 64
1681
+ },
1682
+ "bits": [
1683
+ 2
1684
+ ],
1685
+ "bits_prop": [
1686
+ 1
1687
+ ],
1688
+ "scale_bits": 4
1689
+ },
1690
+ "k_proj": {
1691
+ "group_size": {
1692
+ "2": 64
1693
+ },
1694
+ "bits": [
1695
+ 2
1696
+ ],
1697
+ "bits_prop": [
1698
+ 1
1699
+ ],
1700
+ "scale_bits": 4
1701
+ },
1702
+ "v_proj": {
1703
+ "group_size": {
1704
+ "4": 128
1705
+ },
1706
+ "bits": [
1707
+ 4
1708
+ ],
1709
+ "bits_prop": [
1710
+ 1
1711
+ ],
1712
+ "scale_bits": 4
1713
+ },
1714
+ "o_proj": {
1715
+ "group_size": {
1716
+ "2": 64
1717
+ },
1718
+ "bits": [
1719
+ 2
1720
+ ],
1721
+ "bits_prop": [
1722
+ 1
1723
+ ],
1724
+ "scale_bits": 4
1725
+ },
1726
+ "up_proj": {
1727
+ "group_size": {
1728
+ "2": 64
1729
+ },
1730
+ "bits": [
1731
+ 2
1732
+ ],
1733
+ "bits_prop": [
1734
+ 1
1735
+ ],
1736
+ "scale_bits": 4
1737
+ },
1738
+ "gate_proj": {
1739
+ "group_size": {
1740
+ "2": 64
1741
+ },
1742
+ "bits": [
1743
+ 2
1744
+ ],
1745
+ "bits_prop": [
1746
+ 1
1747
+ ],
1748
+ "scale_bits": 4
1749
+ },
1750
+ "down_proj": {
1751
+ "group_size": {
1752
+ "4": 128
1753
+ },
1754
+ "bits": [
1755
+ 4
1756
+ ],
1757
+ "bits_prop": [
1758
+ 1
1759
+ ],
1760
+ "scale_bits": 4
1761
+ }
1762
+ },
1763
+ "model.layers.20": {
1764
+ "accuracy": 0.8932943344116211,
1765
+ "total_bits": 573724416,
1766
+ "q_proj": {
1767
+ "group_size": {
1768
+ "2": 64
1769
+ },
1770
+ "bits": [
1771
+ 2
1772
+ ],
1773
+ "bits_prop": [
1774
+ 1
1775
+ ],
1776
+ "scale_bits": 4
1777
+ },
1778
+ "k_proj": {
1779
+ "group_size": {
1780
+ "2": 64
1781
+ },
1782
+ "bits": [
1783
+ 2
1784
+ ],
1785
+ "bits_prop": [
1786
+ 1
1787
+ ],
1788
+ "scale_bits": 4
1789
+ },
1790
+ "v_proj": {
1791
+ "group_size": {
1792
+ "4": 128
1793
+ },
1794
+ "bits": [
1795
+ 4
1796
+ ],
1797
+ "bits_prop": [
1798
+ 1
1799
+ ],
1800
+ "scale_bits": 4
1801
+ },
1802
+ "o_proj": {
1803
+ "group_size": {
1804
+ "2": 64
1805
+ },
1806
+ "bits": [
1807
+ 2
1808
+ ],
1809
+ "bits_prop": [
1810
+ 1
1811
+ ],
1812
+ "scale_bits": 4
1813
+ },
1814
+ "up_proj": {
1815
+ "group_size": {
1816
+ "2": 64
1817
+ },
1818
+ "bits": [
1819
+ 2
1820
+ ],
1821
+ "bits_prop": [
1822
+ 1
1823
+ ],
1824
+ "scale_bits": 4
1825
+ },
1826
+ "gate_proj": {
1827
+ "group_size": {
1828
+ "2": 64
1829
+ },
1830
+ "bits": [
1831
+ 2
1832
+ ],
1833
+ "bits_prop": [
1834
+ 1
1835
+ ],
1836
+ "scale_bits": 4
1837
+ },
1838
+ "down_proj": {
1839
+ "group_size": {
1840
+ "4": 128
1841
+ },
1842
+ "bits": [
1843
+ 4
1844
+ ],
1845
+ "bits_prop": [
1846
+ 1
1847
+ ],
1848
+ "scale_bits": 4
1849
+ }
1850
+ },
1851
+ "model.layers.21": {
1852
+ "accuracy": 0.8841193517049153,
1853
+ "total_bits": 573724416,
1854
+ "q_proj": {
1855
+ "group_size": {
1856
+ "2": 64
1857
+ },
1858
+ "bits": [
1859
+ 2
1860
+ ],
1861
+ "bits_prop": [
1862
+ 1
1863
+ ],
1864
+ "scale_bits": 4
1865
+ },
1866
+ "k_proj": {
1867
+ "group_size": {
1868
+ "2": 64
1869
+ },
1870
+ "bits": [
1871
+ 2
1872
+ ],
1873
+ "bits_prop": [
1874
+ 1
1875
+ ],
1876
+ "scale_bits": 4
1877
+ },
1878
+ "v_proj": {
1879
+ "group_size": {
1880
+ "4": 128
1881
+ },
1882
+ "bits": [
1883
+ 4
1884
+ ],
1885
+ "bits_prop": [
1886
+ 1
1887
+ ],
1888
+ "scale_bits": 4
1889
+ },
1890
+ "o_proj": {
1891
+ "group_size": {
1892
+ "2": 64
1893
+ },
1894
+ "bits": [
1895
+ 2
1896
+ ],
1897
+ "bits_prop": [
1898
+ 1
1899
+ ],
1900
+ "scale_bits": 4
1901
+ },
1902
+ "up_proj": {
1903
+ "group_size": {
1904
+ "2": 64
1905
+ },
1906
+ "bits": [
1907
+ 2
1908
+ ],
1909
+ "bits_prop": [
1910
+ 1
1911
+ ],
1912
+ "scale_bits": 4
1913
+ },
1914
+ "gate_proj": {
1915
+ "group_size": {
1916
+ "2": 64
1917
+ },
1918
+ "bits": [
1919
+ 2
1920
+ ],
1921
+ "bits_prop": [
1922
+ 1
1923
+ ],
1924
+ "scale_bits": 4
1925
+ },
1926
+ "down_proj": {
1927
+ "group_size": {
1928
+ "4": 128
1929
+ },
1930
+ "bits": [
1931
+ 4
1932
+ ],
1933
+ "bits_prop": [
1934
+ 1
1935
+ ],
1936
+ "scale_bits": 4
1937
+ }
1938
+ },
1939
+ "model.layers.22": {
1940
+ "accuracy": 0.8901621500651041,
1941
+ "total_bits": 573724416,
1942
+ "q_proj": {
1943
+ "group_size": {
1944
+ "2": 64
1945
+ },
1946
+ "bits": [
1947
+ 2
1948
+ ],
1949
+ "bits_prop": [
1950
+ 1
1951
+ ],
1952
+ "scale_bits": 4
1953
+ },
1954
+ "k_proj": {
1955
+ "group_size": {
1956
+ "2": 64
1957
+ },
1958
+ "bits": [
1959
+ 2
1960
+ ],
1961
+ "bits_prop": [
1962
+ 1
1963
+ ],
1964
+ "scale_bits": 4
1965
+ },
1966
+ "v_proj": {
1967
+ "group_size": {
1968
+ "4": 128
1969
+ },
1970
+ "bits": [
1971
+ 4
1972
+ ],
1973
+ "bits_prop": [
1974
+ 1
1975
+ ],
1976
+ "scale_bits": 4
1977
+ },
1978
+ "o_proj": {
1979
+ "group_size": {
1980
+ "2": 64
1981
+ },
1982
+ "bits": [
1983
+ 2
1984
+ ],
1985
+ "bits_prop": [
1986
+ 1
1987
+ ],
1988
+ "scale_bits": 4
1989
+ },
1990
+ "up_proj": {
1991
+ "group_size": {
1992
+ "2": 64
1993
+ },
1994
+ "bits": [
1995
+ 2
1996
+ ],
1997
+ "bits_prop": [
1998
+ 1
1999
+ ],
2000
+ "scale_bits": 4
2001
+ },
2002
+ "gate_proj": {
2003
+ "group_size": {
2004
+ "2": 64
2005
+ },
2006
+ "bits": [
2007
+ 2
2008
+ ],
2009
+ "bits_prop": [
2010
+ 1
2011
+ ],
2012
+ "scale_bits": 4
2013
+ },
2014
+ "down_proj": {
2015
+ "group_size": {
2016
+ "4": 128
2017
+ },
2018
+ "bits": [
2019
+ 4
2020
+ ],
2021
+ "bits_prop": [
2022
+ 1
2023
+ ],
2024
+ "scale_bits": 4
2025
+ }
2026
+ },
2027
+ "model.layers.23": {
2028
+ "accuracy": 0.8752395311991374,
2029
+ "total_bits": 573724416,
2030
+ "q_proj": {
2031
+ "group_size": {
2032
+ "2": 64
2033
+ },
2034
+ "bits": [
2035
+ 2
2036
+ ],
2037
+ "bits_prop": [
2038
+ 1
2039
+ ],
2040
+ "scale_bits": 4
2041
+ },
2042
+ "k_proj": {
2043
+ "group_size": {
2044
+ "2": 64
2045
+ },
2046
+ "bits": [
2047
+ 2
2048
+ ],
2049
+ "bits_prop": [
2050
+ 1
2051
+ ],
2052
+ "scale_bits": 4
2053
+ },
2054
+ "v_proj": {
2055
+ "group_size": {
2056
+ "4": 128
2057
+ },
2058
+ "bits": [
2059
+ 4
2060
+ ],
2061
+ "bits_prop": [
2062
+ 1
2063
+ ],
2064
+ "scale_bits": 4
2065
+ },
2066
+ "o_proj": {
2067
+ "group_size": {
2068
+ "2": 64
2069
+ },
2070
+ "bits": [
2071
+ 2
2072
+ ],
2073
+ "bits_prop": [
2074
+ 1
2075
+ ],
2076
+ "scale_bits": 4
2077
+ },
2078
+ "up_proj": {
2079
+ "group_size": {
2080
+ "2": 64
2081
+ },
2082
+ "bits": [
2083
+ 2
2084
+ ],
2085
+ "bits_prop": [
2086
+ 1
2087
+ ],
2088
+ "scale_bits": 4
2089
+ },
2090
+ "gate_proj": {
2091
+ "group_size": {
2092
+ "2": 64
2093
+ },
2094
+ "bits": [
2095
+ 2
2096
+ ],
2097
+ "bits_prop": [
2098
+ 1
2099
+ ],
2100
+ "scale_bits": 4
2101
+ },
2102
+ "down_proj": {
2103
+ "group_size": {
2104
+ "4": 128
2105
+ },
2106
+ "bits": [
2107
+ 4
2108
+ ],
2109
+ "bits_prop": [
2110
+ 1
2111
+ ],
2112
+ "scale_bits": 4
2113
+ }
2114
+ },
2115
+ "model.layers.24": {
2116
+ "accuracy": 0.8780385653177897,
2117
+ "total_bits": 573724416,
2118
+ "q_proj": {
2119
+ "group_size": {
2120
+ "2": 64
2121
+ },
2122
+ "bits": [
2123
+ 2
2124
+ ],
2125
+ "bits_prop": [
2126
+ 1
2127
+ ],
2128
+ "scale_bits": 4
2129
+ },
2130
+ "k_proj": {
2131
+ "group_size": {
2132
+ "2": 64
2133
+ },
2134
+ "bits": [
2135
+ 2
2136
+ ],
2137
+ "bits_prop": [
2138
+ 1
2139
+ ],
2140
+ "scale_bits": 4
2141
+ },
2142
+ "v_proj": {
2143
+ "group_size": {
2144
+ "4": 128
2145
+ },
2146
+ "bits": [
2147
+ 4
2148
+ ],
2149
+ "bits_prop": [
2150
+ 1
2151
+ ],
2152
+ "scale_bits": 4
2153
+ },
2154
+ "o_proj": {
2155
+ "group_size": {
2156
+ "2": 64
2157
+ },
2158
+ "bits": [
2159
+ 2
2160
+ ],
2161
+ "bits_prop": [
2162
+ 1
2163
+ ],
2164
+ "scale_bits": 4
2165
+ },
2166
+ "up_proj": {
2167
+ "group_size": {
2168
+ "2": 64
2169
+ },
2170
+ "bits": [
2171
+ 2
2172
+ ],
2173
+ "bits_prop": [
2174
+ 1
2175
+ ],
2176
+ "scale_bits": 4
2177
+ },
2178
+ "gate_proj": {
2179
+ "group_size": {
2180
+ "2": 64
2181
+ },
2182
+ "bits": [
2183
+ 2
2184
+ ],
2185
+ "bits_prop": [
2186
+ 1
2187
+ ],
2188
+ "scale_bits": 4
2189
+ },
2190
+ "down_proj": {
2191
+ "group_size": {
2192
+ "4": 128
2193
+ },
2194
+ "bits": [
2195
+ 4
2196
+ ],
2197
+ "bits_prop": [
2198
+ 1
2199
+ ],
2200
+ "scale_bits": 4
2201
+ }
2202
+ },
2203
+ "model.layers.25": {
2204
+ "accuracy": 0.88067626953125,
2205
+ "total_bits": 573724416,
2206
+ "q_proj": {
2207
+ "group_size": {
2208
+ "2": 64
2209
+ },
2210
+ "bits": [
2211
+ 2
2212
+ ],
2213
+ "bits_prop": [
2214
+ 1
2215
+ ],
2216
+ "scale_bits": 4
2217
+ },
2218
+ "k_proj": {
2219
+ "group_size": {
2220
+ "2": 64
2221
+ },
2222
+ "bits": [
2223
+ 2
2224
+ ],
2225
+ "bits_prop": [
2226
+ 1
2227
+ ],
2228
+ "scale_bits": 4
2229
+ },
2230
+ "v_proj": {
2231
+ "group_size": {
2232
+ "4": 128
2233
+ },
2234
+ "bits": [
2235
+ 4
2236
+ ],
2237
+ "bits_prop": [
2238
+ 1
2239
+ ],
2240
+ "scale_bits": 4
2241
+ },
2242
+ "o_proj": {
2243
+ "group_size": {
2244
+ "2": 64
2245
+ },
2246
+ "bits": [
2247
+ 2
2248
+ ],
2249
+ "bits_prop": [
2250
+ 1
2251
+ ],
2252
+ "scale_bits": 4
2253
+ },
2254
+ "up_proj": {
2255
+ "group_size": {
2256
+ "2": 64
2257
+ },
2258
+ "bits": [
2259
+ 2
2260
+ ],
2261
+ "bits_prop": [
2262
+ 1
2263
+ ],
2264
+ "scale_bits": 4
2265
+ },
2266
+ "gate_proj": {
2267
+ "group_size": {
2268
+ "2": 64
2269
+ },
2270
+ "bits": [
2271
+ 2
2272
+ ],
2273
+ "bits_prop": [
2274
+ 1
2275
+ ],
2276
+ "scale_bits": 4
2277
+ },
2278
+ "down_proj": {
2279
+ "group_size": {
2280
+ "4": 128
2281
+ },
2282
+ "bits": [
2283
+ 4
2284
+ ],
2285
+ "bits_prop": [
2286
+ 1
2287
+ ],
2288
+ "scale_bits": 4
2289
+ }
2290
+ },
2291
+ "model.layers.26": {
2292
+ "accuracy": 0.8708769480387369,
2293
+ "total_bits": 573724416,
2294
+ "q_proj": {
2295
+ "group_size": {
2296
+ "2": 64
2297
+ },
2298
+ "bits": [
2299
+ 2
2300
+ ],
2301
+ "bits_prop": [
2302
+ 1
2303
+ ],
2304
+ "scale_bits": 4
2305
+ },
2306
+ "k_proj": {
2307
+ "group_size": {
2308
+ "2": 64
2309
+ },
2310
+ "bits": [
2311
+ 2
2312
+ ],
2313
+ "bits_prop": [
2314
+ 1
2315
+ ],
2316
+ "scale_bits": 4
2317
+ },
2318
+ "v_proj": {
2319
+ "group_size": {
2320
+ "4": 128
2321
+ },
2322
+ "bits": [
2323
+ 4
2324
+ ],
2325
+ "bits_prop": [
2326
+ 1
2327
+ ],
2328
+ "scale_bits": 4
2329
+ },
2330
+ "o_proj": {
2331
+ "group_size": {
2332
+ "2": 64
2333
+ },
2334
+ "bits": [
2335
+ 2
2336
+ ],
2337
+ "bits_prop": [
2338
+ 1
2339
+ ],
2340
+ "scale_bits": 4
2341
+ },
2342
+ "up_proj": {
2343
+ "group_size": {
2344
+ "2": 64
2345
+ },
2346
+ "bits": [
2347
+ 2
2348
+ ],
2349
+ "bits_prop": [
2350
+ 1
2351
+ ],
2352
+ "scale_bits": 4
2353
+ },
2354
+ "gate_proj": {
2355
+ "group_size": {
2356
+ "2": 64
2357
+ },
2358
+ "bits": [
2359
+ 2
2360
+ ],
2361
+ "bits_prop": [
2362
+ 1
2363
+ ],
2364
+ "scale_bits": 4
2365
+ },
2366
+ "down_proj": {
2367
+ "group_size": {
2368
+ "4": 128
2369
+ },
2370
+ "bits": [
2371
+ 4
2372
+ ],
2373
+ "bits_prop": [
2374
+ 1
2375
+ ],
2376
+ "scale_bits": 4
2377
+ }
2378
+ },
2379
+ "model.layers.27": {
2380
+ "accuracy": 0.8813017209370931,
2381
+ "total_bits": 573724416,
2382
+ "q_proj": {
2383
+ "group_size": {
2384
+ "2": 64
2385
+ },
2386
+ "bits": [
2387
+ 2
2388
+ ],
2389
+ "bits_prop": [
2390
+ 1
2391
+ ],
2392
+ "scale_bits": 4
2393
+ },
2394
+ "k_proj": {
2395
+ "group_size": {
2396
+ "2": 64
2397
+ },
2398
+ "bits": [
2399
+ 2
2400
+ ],
2401
+ "bits_prop": [
2402
+ 1
2403
+ ],
2404
+ "scale_bits": 4
2405
+ },
2406
+ "v_proj": {
2407
+ "group_size": {
2408
+ "4": 128
2409
+ },
2410
+ "bits": [
2411
+ 4
2412
+ ],
2413
+ "bits_prop": [
2414
+ 1
2415
+ ],
2416
+ "scale_bits": 4
2417
+ },
2418
+ "o_proj": {
2419
+ "group_size": {
2420
+ "2": 64
2421
+ },
2422
+ "bits": [
2423
+ 2
2424
+ ],
2425
+ "bits_prop": [
2426
+ 1
2427
+ ],
2428
+ "scale_bits": 4
2429
+ },
2430
+ "up_proj": {
2431
+ "group_size": {
2432
+ "2": 64
2433
+ },
2434
+ "bits": [
2435
+ 2
2436
+ ],
2437
+ "bits_prop": [
2438
+ 1
2439
+ ],
2440
+ "scale_bits": 4
2441
+ },
2442
+ "gate_proj": {
2443
+ "group_size": {
2444
+ "2": 64
2445
+ },
2446
+ "bits": [
2447
+ 2
2448
+ ],
2449
+ "bits_prop": [
2450
+ 1
2451
+ ],
2452
+ "scale_bits": 4
2453
+ },
2454
+ "down_proj": {
2455
+ "group_size": {
2456
+ "4": 128
2457
+ },
2458
+ "bits": [
2459
+ 4
2460
+ ],
2461
+ "bits_prop": [
2462
+ 1
2463
+ ],
2464
+ "scale_bits": 4
2465
+ }
2466
+ },
2467
+ "model.layers.28": {
2468
+ "accuracy": 0.8748725255330403,
2469
+ "total_bits": 573724416,
2470
+ "q_proj": {
2471
+ "group_size": {
2472
+ "2": 64
2473
+ },
2474
+ "bits": [
2475
+ 2
2476
+ ],
2477
+ "bits_prop": [
2478
+ 1
2479
+ ],
2480
+ "scale_bits": 4
2481
+ },
2482
+ "k_proj": {
2483
+ "group_size": {
2484
+ "2": 64
2485
+ },
2486
+ "bits": [
2487
+ 2
2488
+ ],
2489
+ "bits_prop": [
2490
+ 1
2491
+ ],
2492
+ "scale_bits": 4
2493
+ },
2494
+ "v_proj": {
2495
+ "group_size": {
2496
+ "4": 128
2497
+ },
2498
+ "bits": [
2499
+ 4
2500
+ ],
2501
+ "bits_prop": [
2502
+ 1
2503
+ ],
2504
+ "scale_bits": 4
2505
+ },
2506
+ "o_proj": {
2507
+ "group_size": {
2508
+ "2": 64
2509
+ },
2510
+ "bits": [
2511
+ 2
2512
+ ],
2513
+ "bits_prop": [
2514
+ 1
2515
+ ],
2516
+ "scale_bits": 4
2517
+ },
2518
+ "up_proj": {
2519
+ "group_size": {
2520
+ "2": 64
2521
+ },
2522
+ "bits": [
2523
+ 2
2524
+ ],
2525
+ "bits_prop": [
2526
+ 1
2527
+ ],
2528
+ "scale_bits": 4
2529
+ },
2530
+ "gate_proj": {
2531
+ "group_size": {
2532
+ "2": 64
2533
+ },
2534
+ "bits": [
2535
+ 2
2536
+ ],
2537
+ "bits_prop": [
2538
+ 1
2539
+ ],
2540
+ "scale_bits": 4
2541
+ },
2542
+ "down_proj": {
2543
+ "group_size": {
2544
+ "4": 128
2545
+ },
2546
+ "bits": [
2547
+ 4
2548
+ ],
2549
+ "bits_prop": [
2550
+ 1
2551
+ ],
2552
+ "scale_bits": 4
2553
+ }
2554
+ },
2555
+ "model.layers.29": {
2556
+ "accuracy": 0.8845431009928385,
2557
+ "total_bits": 573724416,
2558
+ "q_proj": {
2559
+ "group_size": {
2560
+ "2": 64
2561
+ },
2562
+ "bits": [
2563
+ 2
2564
+ ],
2565
+ "bits_prop": [
2566
+ 1
2567
+ ],
2568
+ "scale_bits": 4
2569
+ },
2570
+ "k_proj": {
2571
+ "group_size": {
2572
+ "2": 64
2573
+ },
2574
+ "bits": [
2575
+ 2
2576
+ ],
2577
+ "bits_prop": [
2578
+ 1
2579
+ ],
2580
+ "scale_bits": 4
2581
+ },
2582
+ "v_proj": {
2583
+ "group_size": {
2584
+ "4": 128
2585
+ },
2586
+ "bits": [
2587
+ 4
2588
+ ],
2589
+ "bits_prop": [
2590
+ 1
2591
+ ],
2592
+ "scale_bits": 4
2593
+ },
2594
+ "o_proj": {
2595
+ "group_size": {
2596
+ "2": 64
2597
+ },
2598
+ "bits": [
2599
+ 2
2600
+ ],
2601
+ "bits_prop": [
2602
+ 1
2603
+ ],
2604
+ "scale_bits": 4
2605
+ },
2606
+ "up_proj": {
2607
+ "group_size": {
2608
+ "2": 64
2609
+ },
2610
+ "bits": [
2611
+ 2
2612
+ ],
2613
+ "bits_prop": [
2614
+ 1
2615
+ ],
2616
+ "scale_bits": 4
2617
+ },
2618
+ "gate_proj": {
2619
+ "group_size": {
2620
+ "2": 64
2621
+ },
2622
+ "bits": [
2623
+ 2
2624
+ ],
2625
+ "bits_prop": [
2626
+ 1
2627
+ ],
2628
+ "scale_bits": 4
2629
+ },
2630
+ "down_proj": {
2631
+ "group_size": {
2632
+ "4": 128
2633
+ },
2634
+ "bits": [
2635
+ 4
2636
+ ],
2637
+ "bits_prop": [
2638
+ 1
2639
+ ],
2640
+ "scale_bits": 4
2641
+ }
2642
+ },
2643
+ "model.layers.30": {
2644
+ "accuracy": 0.8821986516316731,
2645
+ "total_bits": 573724416,
2646
+ "q_proj": {
2647
+ "group_size": {
2648
+ "2": 64
2649
+ },
2650
+ "bits": [
2651
+ 2
2652
+ ],
2653
+ "bits_prop": [
2654
+ 1
2655
+ ],
2656
+ "scale_bits": 4
2657
+ },
2658
+ "k_proj": {
2659
+ "group_size": {
2660
+ "2": 64
2661
+ },
2662
+ "bits": [
2663
+ 2
2664
+ ],
2665
+ "bits_prop": [
2666
+ 1
2667
+ ],
2668
+ "scale_bits": 4
2669
+ },
2670
+ "v_proj": {
2671
+ "group_size": {
2672
+ "4": 128
2673
+ },
2674
+ "bits": [
2675
+ 4
2676
+ ],
2677
+ "bits_prop": [
2678
+ 1
2679
+ ],
2680
+ "scale_bits": 4
2681
+ },
2682
+ "o_proj": {
2683
+ "group_size": {
2684
+ "2": 64
2685
+ },
2686
+ "bits": [
2687
+ 2
2688
+ ],
2689
+ "bits_prop": [
2690
+ 1
2691
+ ],
2692
+ "scale_bits": 4
2693
+ },
2694
+ "up_proj": {
2695
+ "group_size": {
2696
+ "2": 64
2697
+ },
2698
+ "bits": [
2699
+ 2
2700
+ ],
2701
+ "bits_prop": [
2702
+ 1
2703
+ ],
2704
+ "scale_bits": 4
2705
+ },
2706
+ "gate_proj": {
2707
+ "group_size": {
2708
+ "2": 64
2709
+ },
2710
+ "bits": [
2711
+ 2
2712
+ ],
2713
+ "bits_prop": [
2714
+ 1
2715
+ ],
2716
+ "scale_bits": 4
2717
+ },
2718
+ "down_proj": {
2719
+ "group_size": {
2720
+ "4": 128
2721
+ },
2722
+ "bits": [
2723
+ 4
2724
+ ],
2725
+ "bits_prop": [
2726
+ 1
2727
+ ],
2728
+ "scale_bits": 4
2729
+ }
2730
+ },
2731
+ "model.layers.31": {
2732
+ "accuracy": 0.8758713404337564,
2733
+ "total_bits": 573728256,
2734
+ "q_proj": {
2735
+ "group_size": {
2736
+ "2": 64
2737
+ },
2738
+ "bits": [
2739
+ 2
2740
+ ],
2741
+ "bits_prop": [
2742
+ 1
2743
+ ],
2744
+ "scale_bits": 4
2745
+ },
2746
+ "k_proj": {
2747
+ "group_size": {
2748
+ "2": 64
2749
+ },
2750
+ "bits": [
2751
+ 2
2752
+ ],
2753
+ "bits_prop": [
2754
+ 1
2755
+ ],
2756
+ "scale_bits": 4
2757
+ },
2758
+ "v_proj": {
2759
+ "group_size": {
2760
+ "4": 128
2761
+ },
2762
+ "bits": [
2763
+ 4
2764
+ ],
2765
+ "bits_prop": [
2766
+ 1
2767
+ ],
2768
+ "scale_bits": 4
2769
+ },
2770
+ "o_proj": {
2771
+ "group_size": {
2772
+ "2": 64
2773
+ },
2774
+ "bits": [
2775
+ 2
2776
+ ],
2777
+ "bits_prop": [
2778
+ 1
2779
+ ],
2780
+ "scale_bits": 4
2781
+ },
2782
+ "up_proj": {
2783
+ "group_size": {
2784
+ "4": 128
2785
+ },
2786
+ "bits": [
2787
+ 4
2788
+ ],
2789
+ "bits_prop": [
2790
+ 1
2791
+ ],
2792
+ "scale_bits": 4
2793
+ },
2794
+ "gate_proj": {
2795
+ "group_size": {
2796
+ "2": 64
2797
+ },
2798
+ "bits": [
2799
+ 2
2800
+ ],
2801
+ "bits_prop": [
2802
+ 1
2803
+ ],
2804
+ "scale_bits": 4
2805
+ },
2806
+ "down_proj": {
2807
+ "group_size": {
2808
+ "2": 64
2809
+ },
2810
+ "bits": [
2811
+ 2
2812
+ ],
2813
+ "bits_prop": [
2814
+ 1
2815
+ ],
2816
+ "scale_bits": 4
2817
+ }
2818
+ }
2819
+ }
2820
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "unk_token": {
17
+ "content": "<unk>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ }
23
+ }
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:dadfd56d766715c61d2ef780a525ab43b8e6da4de6865bda3d95fdef5e134055
3
+ size 493443
tokenizer_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": false,
27
+ "single_word": false,
28
+ "special": true
29
+ }
30
+ },
31
+ "additional_special_tokens": [],
32
+ "bos_token": "<s>",
33
+ "clean_up_tokenization_spaces": false,
34
+ "eos_token": "</s>",
35
+ "legacy": true,
36
+ "model_max_length": 1000000000000000019884624838656,
37
+ "pad_token": null,
38
+ "sp_model_kwargs": {},
39
+ "spaces_between_special_tokens": false,
40
+ "tokenizer_class": "LlamaTokenizer",
41
+ "unk_token": "<unk>",
42
+ "use_default_system_prompt": false
43
+ }