Hugging Face's logo

RedHatAI
/
Mixtral-8x7B-v0.1-quantized.w4a16

Safetensors
English
mixtral
Mixture of Experts
w4a16
int4
vllm
compressed-tensors
Model card Files
xet
Community
Eldar Kurtic commited on
Commit
f28e05f
·
1 Parent(s): a6b032f

add model

Browse files
Files changed (12) hide show
  1. README.md +208 -0
  2. config.json +1033 -0
  3. generation_config.json +6 -0
  4. model-00001-of-00005.safetensors +3 -0
  5. model-00002-of-00005.safetensors +3 -0
  6. model-00003-of-00005.safetensors +3 -0
  7. model-00004-of-00005.safetensors +3 -0
  8. model-00005-of-00005.safetensors +3 -0
  9. model.safetensors.index.json +0 -0
  10. special_tokens_map.json +23 -0
  11. tokenizer.json +0 -0
  12. tokenizer_config.json +43 -0
README.md ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ tags:
6
+ - moe
7
+ - w4a16
8
+ - int4
9
+ - vllm
10
+ ---
11
+
12
+ # Mixtral-8x7B-v0.1-quantized.w4a16
13
+
14
+ ## Model Overview
15
+ - **Model Architecture:** Mixtral-8x7B-v0.1
16
+ - **Input:** Text
17
+ - **Output:** Text
18
+ - **Model Optimizations:**
19
+ - **Weight quantization:** INT4
20
+ - **Activation quantization:** None
21
+ - **Release Date:** 3/1/2025
22
+ - **Version:** 1.0
23
+ - **Model Developers:** Neural Magic
24
+
25
+ Quantized version of [Mixtral-8x7B-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1).
26
+ It achieves an average score of 67.74 on the [OpenLLM](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) benchmark (version 1), whereas the unquantized model achieves 68.59.
27
+
28
+ ### Model Optimizations
29
+
30
+ This model was obtained by only quantizing the weights to INT4 data type, ready for inference with vLLM >= 0.5.2.
31
+ This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%. Only the weights of the linear operators within transformers blocks are quantized, except the MLP routers.
32
+
33
+ ## Deployment
34
+
35
+ ### Use with vLLM
36
+
37
+ This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below.
38
+
39
+ ```python
40
+ from transformers import AutoTokenizer
41
+ from vllm import LLM, SamplingParams
42
+
43
+ max_model_len, tp_size = 4096, 2
44
+ model_name = "neuralmagic-ent/Mixtral-8x7B-v0.1-quantized.w4a16"
45
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
46
+ llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True)
47
+ sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])
48
+
49
+ messages_list = [
50
+ [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}],
51
+ ]
52
+
53
+ prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]
54
+
55
+ outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)
56
+
57
+ generated_text = [output.outputs[0].text for output in outputs]
58
+ print(generated_text)
59
+ ```
60
+
61
+ vLLM also supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details.
62
+
63
+ ## Creation
64
+
65
+ This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below with the following command:
66
+
67
+ ```bash
68
+ python quantize.py --model_path mistralai/Mixtral-8x7B-v0.1 --quant_path "output_dir" --calib_size 1024 --dampening_frac 0.1 --observer mse --actorder False
69
+ ```
70
+
71
+
72
+ ```python
73
+ from datasets import load_dataset
74
+ from transformers import AutoTokenizer
75
+ from llmcompressor.modifiers.quantization import GPTQModifier
76
+ from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot, apply
77
+ import argparse
78
+ from compressed_tensors.quantization import QuantizationScheme, QuantizationArgs, QuantizationType, QuantizationStrategy
79
+
80
+ def parse_actorder(value):
81
+ # Interpret the input value for --actorder
82
+ if value.lower() == "false":
83
+ return False
84
+ elif value.lower() == "group":
85
+ return "group"
86
+ else:
87
+ raise argparse.ArgumentTypeError("Invalid value for --actorder. Use 'group' or 'False'.")
88
+
89
+
90
+ parser = argparse.ArgumentParser()
91
+ parser.add_argument('--model_path', type=str)
92
+ parser.add_argument('--quant_path', type=str)
93
+ parser.add_argument('--num_bits', type=int, default=4)
94
+ parser.add_argument('--sequential_update', type=bool, default=True)
95
+ parser.add_argument('--calib_size', type=int, default=256)
96
+ parser.add_argument('--dampening_frac', type=float, default=0.05)
97
+ parser.add_argument('--observer', type=str, default="minmax")
98
+ parser.add_argument(
99
+ '--actorder',
100
+ type=parse_actorder,
101
+ default=False, # Default value is False
102
+ help="Specify actorder as 'group' (string) or False (boolean)."
103
+ )
104
+
105
+ args = parser.parse_args()
106
+
107
+
108
+ model = SparseAutoModelForCausalLM.from_pretrained(
109
+ args.model_path,
110
+ device_map="auto",
111
+ torch_dtype="auto",
112
+ use_cache=False,
113
+ )
114
+ tokenizer = AutoTokenizer.from_pretrained(args.model_path)
115
+
116
+
117
+ NUM_CALIBRATION_SAMPLES = args.calib_size
118
+ DATASET_ID = "garage-bAInd/Open-Platypus"
119
+ DATASET_SPLIT = "train"
120
+ ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
121
+ ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
122
+
123
+ def preprocess(example):
124
+ concat_txt = example["instruction"] + "\n" + example["output"]
125
+ return {"text": concat_txt}
126
+
127
+ ds = ds.map(preprocess)
128
+
129
+ def tokenize(sample):
130
+ return tokenizer(
131
+ sample["text"],
132
+ padding=False,
133
+ truncation=False,
134
+ add_special_tokens=True,
135
+ )
136
+
137
+
138
+ ds = ds.map(tokenize, remove_columns=ds.column_names)
139
+
140
+ quant_scheme = QuantizationScheme(
141
+ targets=["Linear"],
142
+ weights=QuantizationArgs(
143
+ num_bits=args.num_bits,
144
+ type=QuantizationType.INT,
145
+ symmetric=True,
146
+ group_size=128,
147
+ strategy=QuantizationStrategy.GROUP,
148
+ observer=args.observer,
149
+ actorder=args.actorder
150
+ ),
151
+ input_activations=None,
152
+ output_activations=None,
153
+ )
154
+
155
+ recipe = [
156
+ GPTQModifier(
157
+ targets=["Linear"],
158
+ ignore=["lm_head", "re:.*block_sparse_moe.gate"],
159
+ sequential_update=args.sequential_update,
160
+ dampening_frac=args.dampening_frac,
161
+ config_groups={"group_0": quant_scheme},
162
+ )
163
+ ]
164
+ oneshot(
165
+ model=model,
166
+ dataset=ds,
167
+ recipe=recipe,
168
+ num_calibration_samples=args.calib_size,
169
+ )
170
+
171
+ # Save to disk compressed.
172
+ SAVE_DIR = args.quant_path
173
+ model.save_pretrained(SAVE_DIR, save_compressed=True)
174
+ tokenizer.save_pretrained(SAVE_DIR)
175
+ ```
176
+
177
+ ## Evaluation
178
+
179
+ The model was evaluated on OpenLLM Leaderboard [V1](https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard) using the following command:
180
+
181
+ OpenLLM Leaderboard V1:
182
+ ```
183
+ lm_eval \
184
+ --model vllm \
185
+ --model_args pretrained="neuralmagic-ent/Mixtral-8x7B-v0.1-quantized.w4a16",dtype=auto,add_bos_token=True,max_model_len=4096,tensor_parallel_size=1,gpu_memory_utilization=0.8,enable_chunked_prefill=True,trust_remote_code=True \
186
+ --tasks openllm \
187
+ --write_out \
188
+ --batch_size auto \
189
+ --output_path output_dir \
190
+ --show_config
191
+ ```
192
+
193
+
194
+ ### Accuracy
195
+
196
+ #### OpenLLM Leaderboard V1 evaluation scores
197
+
198
+ | Metric | mistralai/Mixtral-8x7B-v0.1 | neuralmagic-ent/Mixtral-8x7B-v0.1-quantized.w4a16 |
199
+ |-----------------------------------------|:---------------------------------:|:-------------------------------------------:|
200
+ | ARC-Challenge (Acc-Norm, 25-shot) | 66.55 | 65.61 |
201
+ | GSM8K (Strict-Match, 5-shot) | 59.89 | 58.07 |
202
+ | HellaSwag (Acc-Norm, 10-shot) | 86.65 | 85.21 |
203
+ | MMLU (Acc, 5-shot) | 70.33 | 69.23 |
204
+ | TruthfulQA (MC2, 0-shot) | 46.65 | 47.18 |
205
+ | Winogrande (Acc, 5-shot) | 81.45 | 81.14 |
206
+ | **Average Score** | **68.59** | **67.74** |
207
+ | **Recovery** | **100.00** | **98.76** |
208
+
config.json ADDED
@@ -0,0 +1,1033 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/home/eldar/.cache/huggingface/hub/models--mistralai--Mixtral-8x7B-v0.1/snapshots/ffe1a706bacbd5abddc5ff99432ee38f7e0662fb",
3
+ "architectures": [
4
+ "MixtralForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 1,
8
+ "eos_token_id": 2,
9
+ "hidden_act": "silu",
10
+ "hidden_size": 4096,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 14336,
13
+ "max_position_embeddings": 32768,
14
+ "model_type": "mixtral",
15
+ "num_attention_heads": 32,
16
+ "num_experts_per_tok": 2,
17
+ "num_hidden_layers": 32,
18
+ "num_key_value_heads": 8,
19
+ "num_local_experts": 8,
20
+ "output_router_logits": false,
21
+ "quantization_config": {
22
+ "config_groups": {
23
+ "group_0": {
24
+ "input_activations": null,
25
+ "output_activations": null,
26
+ "targets": [
27
+ "Linear"
28
+ ],
29
+ "weights": {
30
+ "actorder": null,
31
+ "block_structure": null,
32
+ "dynamic": false,
33
+ "group_size": 128,
34
+ "num_bits": 4,
35
+ "observer": "mse",
36
+ "observer_kwargs": {},
37
+ "strategy": "group",
38
+ "symmetric": true,
39
+ "type": "int"
40
+ }
41
+ }
42
+ },
43
+ "format": "pack-quantized",
44
+ "global_compression_ratio": 2.0968847636762966,
45
+ "ignore": [
46
+ "model.layers.0.block_sparse_moe.gate",
47
+ "model.layers.1.block_sparse_moe.gate",
48
+ "model.layers.2.block_sparse_moe.gate",
49
+ "model.layers.3.block_sparse_moe.gate",
50
+ "model.layers.4.block_sparse_moe.gate",
51
+ "model.layers.5.block_sparse_moe.gate",
52
+ "model.layers.6.block_sparse_moe.gate",
53
+ "model.layers.7.block_sparse_moe.gate",
54
+ "model.layers.8.block_sparse_moe.gate",
55
+ "model.layers.9.block_sparse_moe.gate",
56
+ "model.layers.10.block_sparse_moe.gate",
57
+ "model.layers.11.block_sparse_moe.gate",
58
+ "model.layers.12.block_sparse_moe.gate",
59
+ "model.layers.13.block_sparse_moe.gate",
60
+ "model.layers.14.block_sparse_moe.gate",
61
+ "model.layers.15.block_sparse_moe.gate",
62
+ "model.layers.16.block_sparse_moe.gate",
63
+ "model.layers.17.block_sparse_moe.gate",
64
+ "model.layers.18.block_sparse_moe.gate",
65
+ "model.layers.19.block_sparse_moe.gate",
66
+ "model.layers.20.block_sparse_moe.gate",
67
+ "model.layers.21.block_sparse_moe.gate",
68
+ "model.layers.22.block_sparse_moe.gate",
69
+ "model.layers.23.block_sparse_moe.gate",
70
+ "model.layers.24.block_sparse_moe.gate",
71
+ "model.layers.25.block_sparse_moe.gate",
72
+ "model.layers.26.block_sparse_moe.gate",
73
+ "model.layers.27.block_sparse_moe.gate",
74
+ "model.layers.28.block_sparse_moe.gate",
75
+ "model.layers.29.block_sparse_moe.gate",
76
+ "model.layers.30.block_sparse_moe.gate",
77
+ "model.layers.31.block_sparse_moe.gate",
78
+ "lm_head"
79
+ ],
80
+ "kv_cache_scheme": null,
81
+ "quant_method": "compressed-tensors",
82
+ "quantization_status": "compressed",
83
+ "sparsity_config": {
84
+ "format": "dense",
85
+ "global_sparsity": 0.14580331375755926,
86
+ "ignore": [
87
+ "model.layers.0.self_attn.v_proj",
88
+ "model.layers.0.self_attn.o_proj",
89
+ "model.layers.0.block_sparse_moe.gate",
90
+ "model.layers.0.block_sparse_moe.experts.0.w1",
91
+ "model.layers.0.block_sparse_moe.experts.0.w2",
92
+ "model.layers.0.block_sparse_moe.experts.0.w3",
93
+ "model.layers.0.block_sparse_moe.experts.1.w1",
94
+ "model.layers.0.block_sparse_moe.experts.1.w2",
95
+ "model.layers.0.block_sparse_moe.experts.1.w3",
96
+ "model.layers.0.block_sparse_moe.experts.2.w1",
97
+ "model.layers.0.block_sparse_moe.experts.2.w2",
98
+ "model.layers.0.block_sparse_moe.experts.2.w3",
99
+ "model.layers.0.block_sparse_moe.experts.3.w1",
100
+ "model.layers.0.block_sparse_moe.experts.3.w2",
101
+ "model.layers.0.block_sparse_moe.experts.3.w3",
102
+ "model.layers.0.block_sparse_moe.experts.4.w1",
103
+ "model.layers.0.block_sparse_moe.experts.4.w2",
104
+ "model.layers.0.block_sparse_moe.experts.4.w3",
105
+ "model.layers.0.block_sparse_moe.experts.5.w1",
106
+ "model.layers.0.block_sparse_moe.experts.5.w2",
107
+ "model.layers.0.block_sparse_moe.experts.5.w3",
108
+ "model.layers.0.block_sparse_moe.experts.6.w1",
109
+ "model.layers.0.block_sparse_moe.experts.6.w2",
110
+ "model.layers.0.block_sparse_moe.experts.6.w3",
111
+ "model.layers.0.block_sparse_moe.experts.7.w1",
112
+ "model.layers.0.block_sparse_moe.experts.7.w2",
113
+ "model.layers.0.block_sparse_moe.experts.7.w3",
114
+ "model.layers.1.self_attn.q_proj",
115
+ "model.layers.1.self_attn.k_proj",
116
+ "model.layers.1.self_attn.v_proj",
117
+ "model.layers.1.self_attn.o_proj",
118
+ "model.layers.1.block_sparse_moe.gate",
119
+ "model.layers.1.block_sparse_moe.experts.0.w1",
120
+ "model.layers.1.block_sparse_moe.experts.0.w2",
121
+ "model.layers.1.block_sparse_moe.experts.0.w3",
122
+ "model.layers.1.block_sparse_moe.experts.1.w1",
123
+ "model.layers.1.block_sparse_moe.experts.1.w2",
124
+ "model.layers.1.block_sparse_moe.experts.1.w3",
125
+ "model.layers.1.block_sparse_moe.experts.2.w1",
126
+ "model.layers.1.block_sparse_moe.experts.2.w2",
127
+ "model.layers.1.block_sparse_moe.experts.2.w3",
128
+ "model.layers.1.block_sparse_moe.experts.3.w1",
129
+ "model.layers.1.block_sparse_moe.experts.3.w2",
130
+ "model.layers.1.block_sparse_moe.experts.3.w3",
131
+ "model.layers.1.block_sparse_moe.experts.4.w1",
132
+ "model.layers.1.block_sparse_moe.experts.4.w2",
133
+ "model.layers.1.block_sparse_moe.experts.4.w3",
134
+ "model.layers.1.block_sparse_moe.experts.5.w1",
135
+ "model.layers.1.block_sparse_moe.experts.5.w2",
136
+ "model.layers.1.block_sparse_moe.experts.5.w3",
137
+ "model.layers.1.block_sparse_moe.experts.6.w1",
138
+ "model.layers.1.block_sparse_moe.experts.6.w2",
139
+ "model.layers.1.block_sparse_moe.experts.6.w3",
140
+ "model.layers.1.block_sparse_moe.experts.7.w1",
141
+ "model.layers.1.block_sparse_moe.experts.7.w2",
142
+ "model.layers.1.block_sparse_moe.experts.7.w3",
143
+ "model.layers.2.self_attn.q_proj",
144
+ "model.layers.2.self_attn.k_proj",
145
+ "model.layers.2.self_attn.v_proj",
146
+ "model.layers.2.self_attn.o_proj",
147
+ "model.layers.2.block_sparse_moe.gate",
148
+ "model.layers.2.block_sparse_moe.experts.0.w1",
149
+ "model.layers.2.block_sparse_moe.experts.0.w2",
150
+ "model.layers.2.block_sparse_moe.experts.0.w3",
151
+ "model.layers.2.block_sparse_moe.experts.1.w1",
152
+ "model.layers.2.block_sparse_moe.experts.1.w2",
153
+ "model.layers.2.block_sparse_moe.experts.1.w3",
154
+ "model.layers.2.block_sparse_moe.experts.2.w1",
155
+ "model.layers.2.block_sparse_moe.experts.2.w2",
156
+ "model.layers.2.block_sparse_moe.experts.2.w3",
157
+ "model.layers.2.block_sparse_moe.experts.3.w1",
158
+ "model.layers.2.block_sparse_moe.experts.3.w2",
159
+ "model.layers.2.block_sparse_moe.experts.3.w3",
160
+ "model.layers.2.block_sparse_moe.experts.4.w1",
161
+ "model.layers.2.block_sparse_moe.experts.4.w2",
162
+ "model.layers.2.block_sparse_moe.experts.4.w3",
163
+ "model.layers.2.block_sparse_moe.experts.5.w1",
164
+ "model.layers.2.block_sparse_moe.experts.5.w2",
165
+ "model.layers.2.block_sparse_moe.experts.5.w3",
166
+ "model.layers.2.block_sparse_moe.experts.6.w1",
167
+ "model.layers.2.block_sparse_moe.experts.6.w2",
168
+ "model.layers.2.block_sparse_moe.experts.6.w3",
169
+ "model.layers.2.block_sparse_moe.experts.7.w1",
170
+ "model.layers.2.block_sparse_moe.experts.7.w2",
171
+ "model.layers.2.block_sparse_moe.experts.7.w3",
172
+ "model.layers.3.self_attn.q_proj",
173
+ "model.layers.3.self_attn.k_proj",
174
+ "model.layers.3.self_attn.v_proj",
175
+ "model.layers.3.self_attn.o_proj",
176
+ "model.layers.3.block_sparse_moe.gate",
177
+ "model.layers.3.block_sparse_moe.experts.0.w1",
178
+ "model.layers.3.block_sparse_moe.experts.0.w2",
179
+ "model.layers.3.block_sparse_moe.experts.0.w3",
180
+ "model.layers.3.block_sparse_moe.experts.1.w1",
181
+ "model.layers.3.block_sparse_moe.experts.1.w2",
182
+ "model.layers.3.block_sparse_moe.experts.1.w3",
183
+ "model.layers.3.block_sparse_moe.experts.2.w1",
184
+ "model.layers.3.block_sparse_moe.experts.2.w2",
185
+ "model.layers.3.block_sparse_moe.experts.2.w3",
186
+ "model.layers.3.block_sparse_moe.experts.3.w1",
187
+ "model.layers.3.block_sparse_moe.experts.3.w2",
188
+ "model.layers.3.block_sparse_moe.experts.3.w3",
189
+ "model.layers.3.block_sparse_moe.experts.4.w1",
190
+ "model.layers.3.block_sparse_moe.experts.4.w2",
191
+ "model.layers.3.block_sparse_moe.experts.4.w3",
192
+ "model.layers.3.block_sparse_moe.experts.5.w1",
193
+ "model.layers.3.block_sparse_moe.experts.5.w2",
194
+ "model.layers.3.block_sparse_moe.experts.5.w3",
195
+ "model.layers.3.block_sparse_moe.experts.6.w1",
196
+ "model.layers.3.block_sparse_moe.experts.6.w2",
197
+ "model.layers.3.block_sparse_moe.experts.6.w3",
198
+ "model.layers.3.block_sparse_moe.experts.7.w1",
199
+ "model.layers.3.block_sparse_moe.experts.7.w2",
200
+ "model.layers.3.block_sparse_moe.experts.7.w3",
201
+ "model.layers.4.self_attn.q_proj",
202
+ "model.layers.4.self_attn.k_proj",
203
+ "model.layers.4.self_attn.v_proj",
204
+ "model.layers.4.self_attn.o_proj",
205
+ "model.layers.4.block_sparse_moe.gate",
206
+ "model.layers.4.block_sparse_moe.experts.0.w1",
207
+ "model.layers.4.block_sparse_moe.experts.0.w2",
208
+ "model.layers.4.block_sparse_moe.experts.0.w3",
209
+ "model.layers.4.block_sparse_moe.experts.1.w1",
210
+ "model.layers.4.block_sparse_moe.experts.1.w2",
211
+ "model.layers.4.block_sparse_moe.experts.1.w3",
212
+ "model.layers.4.block_sparse_moe.experts.2.w1",
213
+ "model.layers.4.block_sparse_moe.experts.2.w2",
214
+ "model.layers.4.block_sparse_moe.experts.2.w3",
215
+ "model.layers.4.block_sparse_moe.experts.3.w1",
216
+ "model.layers.4.block_sparse_moe.experts.3.w2",
217
+ "model.layers.4.block_sparse_moe.experts.3.w3",
218
+ "model.layers.4.block_sparse_moe.experts.4.w1",
219
+ "model.layers.4.block_sparse_moe.experts.4.w2",
220
+ "model.layers.4.block_sparse_moe.experts.4.w3",
221
+ "model.layers.4.block_sparse_moe.experts.5.w1",
222
+ "model.layers.4.block_sparse_moe.experts.5.w2",
223
+ "model.layers.4.block_sparse_moe.experts.5.w3",
224
+ "model.layers.4.block_sparse_moe.experts.6.w1",
225
+ "model.layers.4.block_sparse_moe.experts.6.w2",
226
+ "model.layers.4.block_sparse_moe.experts.6.w3",
227
+ "model.layers.4.block_sparse_moe.experts.7.w1",
228
+ "model.layers.4.block_sparse_moe.experts.7.w2",
229
+ "model.layers.4.block_sparse_moe.experts.7.w3",
230
+ "model.layers.5.self_attn.q_proj",
231
+ "model.layers.5.self_attn.k_proj",
232
+ "model.layers.5.self_attn.v_proj",
233
+ "model.layers.5.self_attn.o_proj",
234
+ "model.layers.5.block_sparse_moe.gate",
235
+ "model.layers.5.block_sparse_moe.experts.0.w1",
236
+ "model.layers.5.block_sparse_moe.experts.0.w2",
237
+ "model.layers.5.block_sparse_moe.experts.0.w3",
238
+ "model.layers.5.block_sparse_moe.experts.1.w1",
239
+ "model.layers.5.block_sparse_moe.experts.1.w2",
240
+ "model.layers.5.block_sparse_moe.experts.1.w3",
241
+ "model.layers.5.block_sparse_moe.experts.2.w1",
242
+ "model.layers.5.block_sparse_moe.experts.2.w2",
243
+ "model.layers.5.block_sparse_moe.experts.2.w3",
244
+ "model.layers.5.block_sparse_moe.experts.3.w1",
245
+ "model.layers.5.block_sparse_moe.experts.3.w2",
246
+ "model.layers.5.block_sparse_moe.experts.3.w3",
247
+ "model.layers.5.block_sparse_moe.experts.4.w1",
248
+ "model.layers.5.block_sparse_moe.experts.4.w2",
249
+ "model.layers.5.block_sparse_moe.experts.4.w3",
250
+ "model.layers.5.block_sparse_moe.experts.5.w1",
251
+ "model.layers.5.block_sparse_moe.experts.5.w2",
252
+ "model.layers.5.block_sparse_moe.experts.5.w3",
253
+ "model.layers.5.block_sparse_moe.experts.6.w1",
254
+ "model.layers.5.block_sparse_moe.experts.6.w2",
255
+ "model.layers.5.block_sparse_moe.experts.6.w3",
256
+ "model.layers.5.block_sparse_moe.experts.7.w1",
257
+ "model.layers.5.block_sparse_moe.experts.7.w2",
258
+ "model.layers.5.block_sparse_moe.experts.7.w3",
259
+ "model.layers.6.self_attn.q_proj",
260
+ "model.layers.6.self_attn.k_proj",
261
+ "model.layers.6.self_attn.v_proj",
262
+ "model.layers.6.self_attn.o_proj",
263
+ "model.layers.6.block_sparse_moe.gate",
264
+ "model.layers.6.block_sparse_moe.experts.0.w1",
265
+ "model.layers.6.block_sparse_moe.experts.0.w2",
266
+ "model.layers.6.block_sparse_moe.experts.0.w3",
267
+ "model.layers.6.block_sparse_moe.experts.1.w1",
268
+ "model.layers.6.block_sparse_moe.experts.1.w2",
269
+ "model.layers.6.block_sparse_moe.experts.1.w3",
270
+ "model.layers.6.block_sparse_moe.experts.2.w1",
271
+ "model.layers.6.block_sparse_moe.experts.2.w2",
272
+ "model.layers.6.block_sparse_moe.experts.2.w3",
273
+ "model.layers.6.block_sparse_moe.experts.3.w1",
274
+ "model.layers.6.block_sparse_moe.experts.3.w2",
275
+ "model.layers.6.block_sparse_moe.experts.3.w3",
276
+ "model.layers.6.block_sparse_moe.experts.4.w1",
277
+ "model.layers.6.block_sparse_moe.experts.4.w2",
278
+ "model.layers.6.block_sparse_moe.experts.4.w3",
279
+ "model.layers.6.block_sparse_moe.experts.5.w1",
280
+ "model.layers.6.block_sparse_moe.experts.5.w2",
281
+ "model.layers.6.block_sparse_moe.experts.5.w3",
282
+ "model.layers.6.block_sparse_moe.experts.6.w1",
283
+ "model.layers.6.block_sparse_moe.experts.6.w2",
284
+ "model.layers.6.block_sparse_moe.experts.6.w3",
285
+ "model.layers.6.block_sparse_moe.experts.7.w1",
286
+ "model.layers.6.block_sparse_moe.experts.7.w2",
287
+ "model.layers.6.block_sparse_moe.experts.7.w3",
288
+ "model.layers.7.self_attn.q_proj",
289
+ "model.layers.7.self_attn.k_proj",
290
+ "model.layers.7.self_attn.v_proj",
291
+ "model.layers.7.self_attn.o_proj",
292
+ "model.layers.7.block_sparse_moe.gate",
293
+ "model.layers.7.block_sparse_moe.experts.0.w1",
294
+ "model.layers.7.block_sparse_moe.experts.0.w2",
295
+ "model.layers.7.block_sparse_moe.experts.0.w3",
296
+ "model.layers.7.block_sparse_moe.experts.1.w1",
297
+ "model.layers.7.block_sparse_moe.experts.1.w2",
298
+ "model.layers.7.block_sparse_moe.experts.1.w3",
299
+ "model.layers.7.block_sparse_moe.experts.2.w1",
300
+ "model.layers.7.block_sparse_moe.experts.2.w2",
301
+ "model.layers.7.block_sparse_moe.experts.2.w3",
302
+ "model.layers.7.block_sparse_moe.experts.3.w1",
303
+ "model.layers.7.block_sparse_moe.experts.3.w2",
304
+ "model.layers.7.block_sparse_moe.experts.3.w3",
305
+ "model.layers.7.block_sparse_moe.experts.4.w1",
306
+ "model.layers.7.block_sparse_moe.experts.4.w2",
307
+ "model.layers.7.block_sparse_moe.experts.4.w3",
308
+ "model.layers.7.block_sparse_moe.experts.5.w1",
309
+ "model.layers.7.block_sparse_moe.experts.5.w2",
310
+ "model.layers.7.block_sparse_moe.experts.5.w3",
311
+ "model.layers.7.block_sparse_moe.experts.6.w1",
312
+ "model.layers.7.block_sparse_moe.experts.6.w2",
313
+ "model.layers.7.block_sparse_moe.experts.6.w3",
314
+ "model.layers.7.block_sparse_moe.experts.7.w1",
315
+ "model.layers.7.block_sparse_moe.experts.7.w2",
316
+ "model.layers.7.block_sparse_moe.experts.7.w3",
317
+ "model.layers.8.self_attn.q_proj",
318
+ "model.layers.8.self_attn.k_proj",
319
+ "model.layers.8.self_attn.v_proj",
320
+ "model.layers.8.self_attn.o_proj",
321
+ "model.layers.8.block_sparse_moe.gate",
322
+ "model.layers.8.block_sparse_moe.experts.0.w1",
323
+ "model.layers.8.block_sparse_moe.experts.0.w2",
324
+ "model.layers.8.block_sparse_moe.experts.0.w3",
325
+ "model.layers.8.block_sparse_moe.experts.1.w1",
326
+ "model.layers.8.block_sparse_moe.experts.1.w2",
327
+ "model.layers.8.block_sparse_moe.experts.1.w3",
328
+ "model.layers.8.block_sparse_moe.experts.2.w1",
329
+ "model.layers.8.block_sparse_moe.experts.2.w2",
330
+ "model.layers.8.block_sparse_moe.experts.2.w3",
331
+ "model.layers.8.block_sparse_moe.experts.3.w1",
332
+ "model.layers.8.block_sparse_moe.experts.3.w2",
333
+ "model.layers.8.block_sparse_moe.experts.3.w3",
334
+ "model.layers.8.block_sparse_moe.experts.4.w1",
335
+ "model.layers.8.block_sparse_moe.experts.4.w2",
336
+ "model.layers.8.block_sparse_moe.experts.4.w3",
337
+ "model.layers.8.block_sparse_moe.experts.5.w1",
338
+ "model.layers.8.block_sparse_moe.experts.5.w2",
339
+ "model.layers.8.block_sparse_moe.experts.5.w3",
340
+ "model.layers.8.block_sparse_moe.experts.6.w1",
341
+ "model.layers.8.block_sparse_moe.experts.6.w2",
342
+ "model.layers.8.block_sparse_moe.experts.6.w3",
343
+ "model.layers.8.block_sparse_moe.experts.7.w1",
344
+ "model.layers.8.block_sparse_moe.experts.7.w2",
345
+ "model.layers.8.block_sparse_moe.experts.7.w3",
346
+ "model.layers.9.self_attn.q_proj",
347
+ "model.layers.9.self_attn.k_proj",
348
+ "model.layers.9.self_attn.v_proj",
349
+ "model.layers.9.self_attn.o_proj",
350
+ "model.layers.9.block_sparse_moe.gate",
351
+ "model.layers.9.block_sparse_moe.experts.0.w1",
352
+ "model.layers.9.block_sparse_moe.experts.0.w2",
353
+ "model.layers.9.block_sparse_moe.experts.0.w3",
354
+ "model.layers.9.block_sparse_moe.experts.1.w1",
355
+ "model.layers.9.block_sparse_moe.experts.1.w2",
356
+ "model.layers.9.block_sparse_moe.experts.1.w3",
357
+ "model.layers.9.block_sparse_moe.experts.2.w1",
358
+ "model.layers.9.block_sparse_moe.experts.2.w2",
359
+ "model.layers.9.block_sparse_moe.experts.2.w3",
360
+ "model.layers.9.block_sparse_moe.experts.3.w1",
361
+ "model.layers.9.block_sparse_moe.experts.3.w2",
362
+ "model.layers.9.block_sparse_moe.experts.3.w3",
363
+ "model.layers.9.block_sparse_moe.experts.4.w1",
364
+ "model.layers.9.block_sparse_moe.experts.4.w2",
365
+ "model.layers.9.block_sparse_moe.experts.4.w3",
366
+ "model.layers.9.block_sparse_moe.experts.5.w1",
367
+ "model.layers.9.block_sparse_moe.experts.5.w2",
368
+ "model.layers.9.block_sparse_moe.experts.5.w3",
369
+ "model.layers.9.block_sparse_moe.experts.6.w1",
370
+ "model.layers.9.block_sparse_moe.experts.6.w2",
371
+ "model.layers.9.block_sparse_moe.experts.6.w3",
372
+ "model.layers.9.block_sparse_moe.experts.7.w1",
373
+ "model.layers.9.block_sparse_moe.experts.7.w2",
374
+ "model.layers.9.block_sparse_moe.experts.7.w3",
375
+ "model.layers.10.self_attn.q_proj",
376
+ "model.layers.10.self_attn.k_proj",
377
+ "model.layers.10.self_attn.v_proj",
378
+ "model.layers.10.self_attn.o_proj",
379
+ "model.layers.10.block_sparse_moe.gate",
380
+ "model.layers.10.block_sparse_moe.experts.0.w1",
381
+ "model.layers.10.block_sparse_moe.experts.0.w2",
382
+ "model.layers.10.block_sparse_moe.experts.0.w3",
383
+ "model.layers.10.block_sparse_moe.experts.1.w1",
384
+ "model.layers.10.block_sparse_moe.experts.1.w2",
385
+ "model.layers.10.block_sparse_moe.experts.1.w3",
386
+ "model.layers.10.block_sparse_moe.experts.2.w1",
387
+ "model.layers.10.block_sparse_moe.experts.2.w2",
388
+ "model.layers.10.block_sparse_moe.experts.2.w3",
389
+ "model.layers.10.block_sparse_moe.experts.3.w1",
390
+ "model.layers.10.block_sparse_moe.experts.3.w2",
391
+ "model.layers.10.block_sparse_moe.experts.3.w3",
392
+ "model.layers.10.block_sparse_moe.experts.4.w1",
393
+ "model.layers.10.block_sparse_moe.experts.4.w2",
394
+ "model.layers.10.block_sparse_moe.experts.4.w3",
395
+ "model.layers.10.block_sparse_moe.experts.5.w1",
396
+ "model.layers.10.block_sparse_moe.experts.5.w2",
397
+ "model.layers.10.block_sparse_moe.experts.5.w3",
398
+ "model.layers.10.block_sparse_moe.experts.6.w1",
399
+ "model.layers.10.block_sparse_moe.experts.6.w2",
400
+ "model.layers.10.block_sparse_moe.experts.6.w3",
401
+ "model.layers.10.block_sparse_moe.experts.7.w1",
402
+ "model.layers.10.block_sparse_moe.experts.7.w2",
403
+ "model.layers.10.block_sparse_moe.experts.7.w3",
404
+ "model.layers.11.self_attn.q_proj",
405
+ "model.layers.11.self_attn.k_proj",
406
+ "model.layers.11.self_attn.v_proj",
407
+ "model.layers.11.self_attn.o_proj",
408
+ "model.layers.11.block_sparse_moe.gate",
409
+ "model.layers.11.block_sparse_moe.experts.0.w1",
410
+ "model.layers.11.block_sparse_moe.experts.0.w2",
411
+ "model.layers.11.block_sparse_moe.experts.0.w3",
412
+ "model.layers.11.block_sparse_moe.experts.1.w1",
413
+ "model.layers.11.block_sparse_moe.experts.1.w2",
414
+ "model.layers.11.block_sparse_moe.experts.1.w3",
415
+ "model.layers.11.block_sparse_moe.experts.2.w1",
416
+ "model.layers.11.block_sparse_moe.experts.2.w2",
417
+ "model.layers.11.block_sparse_moe.experts.2.w3",
418
+ "model.layers.11.block_sparse_moe.experts.3.w1",
419
+ "model.layers.11.block_sparse_moe.experts.3.w2",
420
+ "model.layers.11.block_sparse_moe.experts.3.w3",
421
+ "model.layers.11.block_sparse_moe.experts.4.w1",
422
+ "model.layers.11.block_sparse_moe.experts.4.w2",
423
+ "model.layers.11.block_sparse_moe.experts.4.w3",
424
+ "model.layers.11.block_sparse_moe.experts.5.w1",
425
+ "model.layers.11.block_sparse_moe.experts.5.w2",
426
+ "model.layers.11.block_sparse_moe.experts.5.w3",
427
+ "model.layers.11.block_sparse_moe.experts.6.w1",
428
+ "model.layers.11.block_sparse_moe.experts.6.w2",
429
+ "model.layers.11.block_sparse_moe.experts.6.w3",
430
+ "model.layers.11.block_sparse_moe.experts.7.w1",
431
+ "model.layers.11.block_sparse_moe.experts.7.w2",
432
+ "model.layers.11.block_sparse_moe.experts.7.w3",
433
+ "model.layers.12.self_attn.q_proj",
434
+ "model.layers.12.self_attn.k_proj",
435
+ "model.layers.12.self_attn.v_proj",
436
+ "model.layers.12.self_attn.o_proj",
437
+ "model.layers.12.block_sparse_moe.gate",
438
+ "model.layers.12.block_sparse_moe.experts.0.w1",
439
+ "model.layers.12.block_sparse_moe.experts.0.w2",
440
+ "model.layers.12.block_sparse_moe.experts.0.w3",
441
+ "model.layers.12.block_sparse_moe.experts.1.w1",
442
+ "model.layers.12.block_sparse_moe.experts.1.w2",
443
+ "model.layers.12.block_sparse_moe.experts.1.w3",
444
+ "model.layers.12.block_sparse_moe.experts.2.w1",
445
+ "model.layers.12.block_sparse_moe.experts.2.w2",
446
+ "model.layers.12.block_sparse_moe.experts.2.w3",
447
+ "model.layers.12.block_sparse_moe.experts.3.w1",
448
+ "model.layers.12.block_sparse_moe.experts.3.w2",
449
+ "model.layers.12.block_sparse_moe.experts.3.w3",
450
+ "model.layers.12.block_sparse_moe.experts.4.w1",
451
+ "model.layers.12.block_sparse_moe.experts.4.w2",
452
+ "model.layers.12.block_sparse_moe.experts.4.w3",
453
+ "model.layers.12.block_sparse_moe.experts.5.w1",
454
+ "model.layers.12.block_sparse_moe.experts.5.w2",
455
+ "model.layers.12.block_sparse_moe.experts.5.w3",
456
+ "model.layers.12.block_sparse_moe.experts.6.w1",
457
+ "model.layers.12.block_sparse_moe.experts.6.w2",
458
+ "model.layers.12.block_sparse_moe.experts.6.w3",
459
+ "model.layers.12.block_sparse_moe.experts.7.w1",
460
+ "model.layers.12.block_sparse_moe.experts.7.w2",
461
+ "model.layers.12.block_sparse_moe.experts.7.w3",
462
+ "model.layers.13.self_attn.q_proj",
463
+ "model.layers.13.self_attn.k_proj",
464
+ "model.layers.13.self_attn.v_proj",
465
+ "model.layers.13.self_attn.o_proj",
466
+ "model.layers.13.block_sparse_moe.gate",
467
+ "model.layers.13.block_sparse_moe.experts.0.w1",
468
+ "model.layers.13.block_sparse_moe.experts.0.w2",
469
+ "model.layers.13.block_sparse_moe.experts.0.w3",
470
+ "model.layers.13.block_sparse_moe.experts.1.w1",
471
+ "model.layers.13.block_sparse_moe.experts.1.w2",
472
+ "model.layers.13.block_sparse_moe.experts.1.w3",
473
+ "model.layers.13.block_sparse_moe.experts.2.w1",
474
+ "model.layers.13.block_sparse_moe.experts.2.w2",
475
+ "model.layers.13.block_sparse_moe.experts.2.w3",
476
+ "model.layers.13.block_sparse_moe.experts.3.w1",
477
+ "model.layers.13.block_sparse_moe.experts.3.w2",
478
+ "model.layers.13.block_sparse_moe.experts.3.w3",
479
+ "model.layers.13.block_sparse_moe.experts.4.w1",
480
+ "model.layers.13.block_sparse_moe.experts.4.w2",
481
+ "model.layers.13.block_sparse_moe.experts.4.w3",
482
+ "model.layers.13.block_sparse_moe.experts.5.w1",
483
+ "model.layers.13.block_sparse_moe.experts.5.w2",
484
+ "model.layers.13.block_sparse_moe.experts.5.w3",
485
+ "model.layers.13.block_sparse_moe.experts.6.w1",
486
+ "model.layers.13.block_sparse_moe.experts.6.w2",
487
+ "model.layers.13.block_sparse_moe.experts.6.w3",
488
+ "model.layers.13.block_sparse_moe.experts.7.w1",
489
+ "model.layers.13.block_sparse_moe.experts.7.w2",
490
+ "model.layers.13.block_sparse_moe.experts.7.w3",
491
+ "model.layers.14.self_attn.q_proj",
492
+ "model.layers.14.self_attn.k_proj",
493
+ "model.layers.14.self_attn.v_proj",
494
+ "model.layers.14.self_attn.o_proj",
495
+ "model.layers.14.block_sparse_moe.gate",
496
+ "model.layers.14.block_sparse_moe.experts.0.w1",
497
+ "model.layers.14.block_sparse_moe.experts.0.w2",
498
+ "model.layers.14.block_sparse_moe.experts.0.w3",
499
+ "model.layers.14.block_sparse_moe.experts.1.w1",
500
+ "model.layers.14.block_sparse_moe.experts.1.w2",
501
+ "model.layers.14.block_sparse_moe.experts.1.w3",
502
+ "model.layers.14.block_sparse_moe.experts.2.w1",
503
+ "model.layers.14.block_sparse_moe.experts.2.w2",
504
+ "model.layers.14.block_sparse_moe.experts.2.w3",
505
+ "model.layers.14.block_sparse_moe.experts.3.w1",
506
+ "model.layers.14.block_sparse_moe.experts.3.w2",
507
+ "model.layers.14.block_sparse_moe.experts.3.w3",
508
+ "model.layers.14.block_sparse_moe.experts.4.w1",
509
+ "model.layers.14.block_sparse_moe.experts.4.w2",
510
+ "model.layers.14.block_sparse_moe.experts.4.w3",
511
+ "model.layers.14.block_sparse_moe.experts.5.w1",
512
+ "model.layers.14.block_sparse_moe.experts.5.w2",
513
+ "model.layers.14.block_sparse_moe.experts.5.w3",
514
+ "model.layers.14.block_sparse_moe.experts.6.w1",
515
+ "model.layers.14.block_sparse_moe.experts.6.w2",
516
+ "model.layers.14.block_sparse_moe.experts.6.w3",
517
+ "model.layers.14.block_sparse_moe.experts.7.w1",
518
+ "model.layers.14.block_sparse_moe.experts.7.w2",
519
+ "model.layers.14.block_sparse_moe.experts.7.w3",
520
+ "model.layers.15.self_attn.q_proj",
521
+ "model.layers.15.self_attn.k_proj",
522
+ "model.layers.15.self_attn.v_proj",
523
+ "model.layers.15.self_attn.o_proj",
524
+ "model.layers.15.block_sparse_moe.gate",
525
+ "model.layers.15.block_sparse_moe.experts.0.w1",
526
+ "model.layers.15.block_sparse_moe.experts.0.w2",
527
+ "model.layers.15.block_sparse_moe.experts.0.w3",
528
+ "model.layers.15.block_sparse_moe.experts.1.w1",
529
+ "model.layers.15.block_sparse_moe.experts.1.w2",
530
+ "model.layers.15.block_sparse_moe.experts.1.w3",
531
+ "model.layers.15.block_sparse_moe.experts.2.w1",
532
+ "model.layers.15.block_sparse_moe.experts.2.w2",
533
+ "model.layers.15.block_sparse_moe.experts.2.w3",
534
+ "model.layers.15.block_sparse_moe.experts.3.w1",
535
+ "model.layers.15.block_sparse_moe.experts.3.w2",
536
+ "model.layers.15.block_sparse_moe.experts.3.w3",
537
+ "model.layers.15.block_sparse_moe.experts.4.w1",
538
+ "model.layers.15.block_sparse_moe.experts.4.w2",
539
+ "model.layers.15.block_sparse_moe.experts.4.w3",
540
+ "model.layers.15.block_sparse_moe.experts.5.w1",
541
+ "model.layers.15.block_sparse_moe.experts.5.w2",
542
+ "model.layers.15.block_sparse_moe.experts.5.w3",
543
+ "model.layers.15.block_sparse_moe.experts.6.w1",
544
+ "model.layers.15.block_sparse_moe.experts.6.w2",
545
+ "model.layers.15.block_sparse_moe.experts.6.w3",
546
+ "model.layers.15.block_sparse_moe.experts.7.w1",
547
+ "model.layers.15.block_sparse_moe.experts.7.w2",
548
+ "model.layers.15.block_sparse_moe.experts.7.w3",
549
+ "model.layers.16.self_attn.q_proj",
550
+ "model.layers.16.self_attn.k_proj",
551
+ "model.layers.16.self_attn.v_proj",
552
+ "model.layers.16.self_attn.o_proj",
553
+ "model.layers.16.block_sparse_moe.gate",
554
+ "model.layers.16.block_sparse_moe.experts.0.w1",
555
+ "model.layers.16.block_sparse_moe.experts.0.w2",
556
+ "model.layers.16.block_sparse_moe.experts.0.w3",
557
+ "model.layers.16.block_sparse_moe.experts.1.w1",
558
+ "model.layers.16.block_sparse_moe.experts.1.w2",
559
+ "model.layers.16.block_sparse_moe.experts.1.w3",
560
+ "model.layers.16.block_sparse_moe.experts.2.w1",
561
+ "model.layers.16.block_sparse_moe.experts.2.w2",
562
+ "model.layers.16.block_sparse_moe.experts.2.w3",
563
+ "model.layers.16.block_sparse_moe.experts.3.w1",
564
+ "model.layers.16.block_sparse_moe.experts.3.w2",
565
+ "model.layers.16.block_sparse_moe.experts.3.w3",
566
+ "model.layers.16.block_sparse_moe.experts.4.w1",
567
+ "model.layers.16.block_sparse_moe.experts.4.w2",
568
+ "model.layers.16.block_sparse_moe.experts.4.w3",
569
+ "model.layers.16.block_sparse_moe.experts.5.w1",
570
+ "model.layers.16.block_sparse_moe.experts.5.w2",
571
+ "model.layers.16.block_sparse_moe.experts.5.w3",
572
+ "model.layers.16.block_sparse_moe.experts.6.w1",
573
+ "model.layers.16.block_sparse_moe.experts.6.w2",
574
+ "model.layers.16.block_sparse_moe.experts.6.w3",
575
+ "model.layers.16.block_sparse_moe.experts.7.w1",
576
+ "model.layers.16.block_sparse_moe.experts.7.w2",
577
+ "model.layers.16.block_sparse_moe.experts.7.w3",
578
+ "model.layers.17.self_attn.q_proj",
579
+ "model.layers.17.self_attn.k_proj",
580
+ "model.layers.17.self_attn.v_proj",
581
+ "model.layers.17.self_attn.o_proj",
582
+ "model.layers.17.block_sparse_moe.gate",
583
+ "model.layers.17.block_sparse_moe.experts.0.w1",
584
+ "model.layers.17.block_sparse_moe.experts.0.w2",
585
+ "model.layers.17.block_sparse_moe.experts.0.w3",
586
+ "model.layers.17.block_sparse_moe.experts.1.w1",
587
+ "model.layers.17.block_sparse_moe.experts.1.w2",
588
+ "model.layers.17.block_sparse_moe.experts.1.w3",
589
+ "model.layers.17.block_sparse_moe.experts.2.w1",
590
+ "model.layers.17.block_sparse_moe.experts.2.w2",
591
+ "model.layers.17.block_sparse_moe.experts.2.w3",
592
+ "model.layers.17.block_sparse_moe.experts.3.w1",
593
+ "model.layers.17.block_sparse_moe.experts.3.w2",
594
+ "model.layers.17.block_sparse_moe.experts.3.w3",
595
+ "model.layers.17.block_sparse_moe.experts.4.w1",
596
+ "model.layers.17.block_sparse_moe.experts.4.w2",
597
+ "model.layers.17.block_sparse_moe.experts.4.w3",
598
+ "model.layers.17.block_sparse_moe.experts.5.w1",
599
+ "model.layers.17.block_sparse_moe.experts.5.w2",
600
+ "model.layers.17.block_sparse_moe.experts.5.w3",
601
+ "model.layers.17.block_sparse_moe.experts.6.w1",
602
+ "model.layers.17.block_sparse_moe.experts.6.w2",
603
+ "model.layers.17.block_sparse_moe.experts.6.w3",
604
+ "model.layers.17.block_sparse_moe.experts.7.w1",
605
+ "model.layers.17.block_sparse_moe.experts.7.w2",
606
+ "model.layers.17.block_sparse_moe.experts.7.w3",
607
+ "model.layers.18.self_attn.q_proj",
608
+ "model.layers.18.self_attn.k_proj",
609
+ "model.layers.18.self_attn.v_proj",
610
+ "model.layers.18.self_attn.o_proj",
611
+ "model.layers.18.block_sparse_moe.gate",
612
+ "model.layers.18.block_sparse_moe.experts.0.w1",
613
+ "model.layers.18.block_sparse_moe.experts.0.w2",
614
+ "model.layers.18.block_sparse_moe.experts.0.w3",
615
+ "model.layers.18.block_sparse_moe.experts.1.w1",
616
+ "model.layers.18.block_sparse_moe.experts.1.w2",
617
+ "model.layers.18.block_sparse_moe.experts.1.w3",
618
+ "model.layers.18.block_sparse_moe.experts.2.w1",
619
+ "model.layers.18.block_sparse_moe.experts.2.w2",
620
+ "model.layers.18.block_sparse_moe.experts.2.w3",
621
+ "model.layers.18.block_sparse_moe.experts.3.w1",
622
+ "model.layers.18.block_sparse_moe.experts.3.w2",
623
+ "model.layers.18.block_sparse_moe.experts.3.w3",
624
+ "model.layers.18.block_sparse_moe.experts.4.w1",
625
+ "model.layers.18.block_sparse_moe.experts.4.w2",
626
+ "model.layers.18.block_sparse_moe.experts.4.w3",
627
+ "model.layers.18.block_sparse_moe.experts.5.w1",
628
+ "model.layers.18.block_sparse_moe.experts.5.w2",
629
+ "model.layers.18.block_sparse_moe.experts.5.w3",
630
+ "model.layers.18.block_sparse_moe.experts.6.w1",
631
+ "model.layers.18.block_sparse_moe.experts.6.w2",
632
+ "model.layers.18.block_sparse_moe.experts.6.w3",
633
+ "model.layers.18.block_sparse_moe.experts.7.w1",
634
+ "model.layers.18.block_sparse_moe.experts.7.w2",
635
+ "model.layers.18.block_sparse_moe.experts.7.w3",
636
+ "model.layers.19.self_attn.q_proj",
637
+ "model.layers.19.self_attn.k_proj",
638
+ "model.layers.19.self_attn.v_proj",
639
+ "model.layers.19.self_attn.o_proj",
640
+ "model.layers.19.block_sparse_moe.gate",
641
+ "model.layers.19.block_sparse_moe.experts.0.w1",
642
+ "model.layers.19.block_sparse_moe.experts.0.w2",
643
+ "model.layers.19.block_sparse_moe.experts.0.w3",
644
+ "model.layers.19.block_sparse_moe.experts.1.w1",
645
+ "model.layers.19.block_sparse_moe.experts.1.w2",
646
+ "model.layers.19.block_sparse_moe.experts.1.w3",
647
+ "model.layers.19.block_sparse_moe.experts.2.w1",
648
+ "model.layers.19.block_sparse_moe.experts.2.w2",
649
+ "model.layers.19.block_sparse_moe.experts.2.w3",
650
+ "model.layers.19.block_sparse_moe.experts.3.w1",
651
+ "model.layers.19.block_sparse_moe.experts.3.w2",
652
+ "model.layers.19.block_sparse_moe.experts.3.w3",
653
+ "model.layers.19.block_sparse_moe.experts.4.w1",
654
+ "model.layers.19.block_sparse_moe.experts.4.w2",
655
+ "model.layers.19.block_sparse_moe.experts.4.w3",
656
+ "model.layers.19.block_sparse_moe.experts.5.w1",
657
+ "model.layers.19.block_sparse_moe.experts.5.w2",
658
+ "model.layers.19.block_sparse_moe.experts.5.w3",
659
+ "model.layers.19.block_sparse_moe.experts.6.w1",
660
+ "model.layers.19.block_sparse_moe.experts.6.w2",
661
+ "model.layers.19.block_sparse_moe.experts.6.w3",
662
+ "model.layers.19.block_sparse_moe.experts.7.w1",
663
+ "model.layers.19.block_sparse_moe.experts.7.w2",
664
+ "model.layers.19.block_sparse_moe.experts.7.w3",
665
+ "model.layers.20.self_attn.q_proj",
666
+ "model.layers.20.self_attn.k_proj",
667
+ "model.layers.20.self_attn.v_proj",
668
+ "model.layers.20.self_attn.o_proj",
669
+ "model.layers.20.block_sparse_moe.gate",
670
+ "model.layers.20.block_sparse_moe.experts.0.w1",
671
+ "model.layers.20.block_sparse_moe.experts.0.w2",
672
+ "model.layers.20.block_sparse_moe.experts.0.w3",
673
+ "model.layers.20.block_sparse_moe.experts.1.w1",
674
+ "model.layers.20.block_sparse_moe.experts.1.w2",
675
+ "model.layers.20.block_sparse_moe.experts.1.w3",
676
+ "model.layers.20.block_sparse_moe.experts.2.w1",
677
+ "model.layers.20.block_sparse_moe.experts.2.w2",
678
+ "model.layers.20.block_sparse_moe.experts.2.w3",
679
+ "model.layers.20.block_sparse_moe.experts.3.w1",
680
+ "model.layers.20.block_sparse_moe.experts.3.w2",
681
+ "model.layers.20.block_sparse_moe.experts.3.w3",
682
+ "model.layers.20.block_sparse_moe.experts.4.w1",
683
+ "model.layers.20.block_sparse_moe.experts.4.w2",
684
+ "model.layers.20.block_sparse_moe.experts.4.w3",
685
+ "model.layers.20.block_sparse_moe.experts.5.w1",
686
+ "model.layers.20.block_sparse_moe.experts.5.w2",
687
+ "model.layers.20.block_sparse_moe.experts.5.w3",
688
+ "model.layers.20.block_sparse_moe.experts.6.w1",
689
+ "model.layers.20.block_sparse_moe.experts.6.w2",
690
+ "model.layers.20.block_sparse_moe.experts.6.w3",
691
+ "model.layers.20.block_sparse_moe.experts.7.w1",
692
+ "model.layers.20.block_sparse_moe.experts.7.w2",
693
+ "model.layers.20.block_sparse_moe.experts.7.w3",
694
+ "model.layers.21.self_attn.q_proj",
695
+ "model.layers.21.self_attn.k_proj",
696
+ "model.layers.21.self_attn.v_proj",
697
+ "model.layers.21.self_attn.o_proj",
698
+ "model.layers.21.block_sparse_moe.gate",
699
+ "model.layers.21.block_sparse_moe.experts.0.w1",
700
+ "model.layers.21.block_sparse_moe.experts.0.w2",
701
+ "model.layers.21.block_sparse_moe.experts.0.w3",
702
+ "model.layers.21.block_sparse_moe.experts.1.w1",
703
+ "model.layers.21.block_sparse_moe.experts.1.w2",
704
+ "model.layers.21.block_sparse_moe.experts.1.w3",
705
+ "model.layers.21.block_sparse_moe.experts.2.w1",
706
+ "model.layers.21.block_sparse_moe.experts.2.w2",
707
+ "model.layers.21.block_sparse_moe.experts.2.w3",
708
+ "model.layers.21.block_sparse_moe.experts.3.w1",
709
+ "model.layers.21.block_sparse_moe.experts.3.w2",
710
+ "model.layers.21.block_sparse_moe.experts.3.w3",
711
+ "model.layers.21.block_sparse_moe.experts.4.w1",
712
+ "model.layers.21.block_sparse_moe.experts.4.w2",
713
+ "model.layers.21.block_sparse_moe.experts.4.w3",
714
+ "model.layers.21.block_sparse_moe.experts.5.w1",
715
+ "model.layers.21.block_sparse_moe.experts.5.w2",
716
+ "model.layers.21.block_sparse_moe.experts.5.w3",
717
+ "model.layers.21.block_sparse_moe.experts.6.w1",
718
+ "model.layers.21.block_sparse_moe.experts.6.w2",
719
+ "model.layers.21.block_sparse_moe.experts.6.w3",
720
+ "model.layers.21.block_sparse_moe.experts.7.w1",
721
+ "model.layers.21.block_sparse_moe.experts.7.w2",
722
+ "model.layers.21.block_sparse_moe.experts.7.w3",
723
+ "model.layers.22.self_attn.q_proj",
724
+ "model.layers.22.self_attn.k_proj",
725
+ "model.layers.22.self_attn.v_proj",
726
+ "model.layers.22.self_attn.o_proj",
727
+ "model.layers.22.block_sparse_moe.gate",
728
+ "model.layers.22.block_sparse_moe.experts.0.w1",
729
+ "model.layers.22.block_sparse_moe.experts.0.w2",
730
+ "model.layers.22.block_sparse_moe.experts.0.w3",
731
+ "model.layers.22.block_sparse_moe.experts.1.w1",
732
+ "model.layers.22.block_sparse_moe.experts.1.w2",
733
+ "model.layers.22.block_sparse_moe.experts.1.w3",
734
+ "model.layers.22.block_sparse_moe.experts.2.w1",
735
+ "model.layers.22.block_sparse_moe.experts.2.w2",
736
+ "model.layers.22.block_sparse_moe.experts.2.w3",
737
+ "model.layers.22.block_sparse_moe.experts.3.w1",
738
+ "model.layers.22.block_sparse_moe.experts.3.w2",
739
+ "model.layers.22.block_sparse_moe.experts.3.w3",
740
+ "model.layers.22.block_sparse_moe.experts.4.w1",
741
+ "model.layers.22.block_sparse_moe.experts.4.w2",
742
+ "model.layers.22.block_sparse_moe.experts.4.w3",
743
+ "model.layers.22.block_sparse_moe.experts.5.w1",
744
+ "model.layers.22.block_sparse_moe.experts.5.w2",
745
+ "model.layers.22.block_sparse_moe.experts.5.w3",
746
+ "model.layers.22.block_sparse_moe.experts.6.w1",
747
+ "model.layers.22.block_sparse_moe.experts.6.w2",
748
+ "model.layers.22.block_sparse_moe.experts.6.w3",
749
+ "model.layers.22.block_sparse_moe.experts.7.w1",
750
+ "model.layers.22.block_sparse_moe.experts.7.w2",
751
+ "model.layers.22.block_sparse_moe.experts.7.w3",
752
+ "model.layers.23.self_attn.q_proj",
753
+ "model.layers.23.self_attn.k_proj",
754
+ "model.layers.23.self_attn.v_proj",
755
+ "model.layers.23.self_attn.o_proj",
756
+ "model.layers.23.block_sparse_moe.gate",
757
+ "model.layers.23.block_sparse_moe.experts.0.w1",
758
+ "model.layers.23.block_sparse_moe.experts.0.w2",
759
+ "model.layers.23.block_sparse_moe.experts.0.w3",
760
+ "model.layers.23.block_sparse_moe.experts.1.w1",
761
+ "model.layers.23.block_sparse_moe.experts.1.w2",
762
+ "model.layers.23.block_sparse_moe.experts.1.w3",
763
+ "model.layers.23.block_sparse_moe.experts.2.w1",
764
+ "model.layers.23.block_sparse_moe.experts.2.w2",
765
+ "model.layers.23.block_sparse_moe.experts.2.w3",
766
+ "model.layers.23.block_sparse_moe.experts.3.w1",
767
+ "model.layers.23.block_sparse_moe.experts.3.w2",
768
+ "model.layers.23.block_sparse_moe.experts.3.w3",
769
+ "model.layers.23.block_sparse_moe.experts.4.w1",
770
+ "model.layers.23.block_sparse_moe.experts.4.w2",
771
+ "model.layers.23.block_sparse_moe.experts.4.w3",
772
+ "model.layers.23.block_sparse_moe.experts.5.w1",
773
+ "model.layers.23.block_sparse_moe.experts.5.w2",
774
+ "model.layers.23.block_sparse_moe.experts.5.w3",
775
+ "model.layers.23.block_sparse_moe.experts.6.w1",
776
+ "model.layers.23.block_sparse_moe.experts.6.w2",
777
+ "model.layers.23.block_sparse_moe.experts.6.w3",
778
+ "model.layers.23.block_sparse_moe.experts.7.w1",
779
+ "model.layers.23.block_sparse_moe.experts.7.w2",
780
+ "model.layers.23.block_sparse_moe.experts.7.w3",
781
+ "model.layers.24.self_attn.q_proj",
782
+ "model.layers.24.self_attn.k_proj",
783
+ "model.layers.24.self_attn.v_proj",
784
+ "model.layers.24.self_attn.o_proj",
785
+ "model.layers.24.block_sparse_moe.gate",
786
+ "model.layers.24.block_sparse_moe.experts.0.w1",
787
+ "model.layers.24.block_sparse_moe.experts.0.w2",
788
+ "model.layers.24.block_sparse_moe.experts.0.w3",
789
+ "model.layers.24.block_sparse_moe.experts.1.w1",
790
+ "model.layers.24.block_sparse_moe.experts.1.w2",
791
+ "model.layers.24.block_sparse_moe.experts.1.w3",
792
+ "model.layers.24.block_sparse_moe.experts.2.w1",
793
+ "model.layers.24.block_sparse_moe.experts.2.w2",
794
+ "model.layers.24.block_sparse_moe.experts.2.w3",
795
+ "model.layers.24.block_sparse_moe.experts.3.w1",
796
+ "model.layers.24.block_sparse_moe.experts.3.w2",
797
+ "model.layers.24.block_sparse_moe.experts.3.w3",
798
+ "model.layers.24.block_sparse_moe.experts.4.w1",
799
+ "model.layers.24.block_sparse_moe.experts.4.w2",
800
+ "model.layers.24.block_sparse_moe.experts.4.w3",
801
+ "model.layers.24.block_sparse_moe.experts.5.w1",
802
+ "model.layers.24.block_sparse_moe.experts.5.w2",
803
+ "model.layers.24.block_sparse_moe.experts.5.w3",
804
+ "model.layers.24.block_sparse_moe.experts.6.w1",
805
+ "model.layers.24.block_sparse_moe.experts.6.w2",
806
+ "model.layers.24.block_sparse_moe.experts.6.w3",
807
+ "model.layers.24.block_sparse_moe.experts.7.w1",
808
+ "model.layers.24.block_sparse_moe.experts.7.w2",
809
+ "model.layers.24.block_sparse_moe.experts.7.w3",
810
+ "model.layers.25.self_attn.q_proj",
811
+ "model.layers.25.self_attn.k_proj",
812
+ "model.layers.25.self_attn.v_proj",
813
+ "model.layers.25.self_attn.o_proj",
814
+ "model.layers.25.block_sparse_moe.gate",
815
+ "model.layers.25.block_sparse_moe.experts.0.w1",
816
+ "model.layers.25.block_sparse_moe.experts.0.w2",
817
+ "model.layers.25.block_sparse_moe.experts.0.w3",
818
+ "model.layers.25.block_sparse_moe.experts.1.w1",
819
+ "model.layers.25.block_sparse_moe.experts.1.w2",
820
+ "model.layers.25.block_sparse_moe.experts.1.w3",
821
+ "model.layers.25.block_sparse_moe.experts.2.w1",
822
+ "model.layers.25.block_sparse_moe.experts.2.w2",
823
+ "model.layers.25.block_sparse_moe.experts.2.w3",
824
+ "model.layers.25.block_sparse_moe.experts.3.w1",
825
+ "model.layers.25.block_sparse_moe.experts.3.w2",
826
+ "model.layers.25.block_sparse_moe.experts.3.w3",
827
+ "model.layers.25.block_sparse_moe.experts.4.w1",
828
+ "model.layers.25.block_sparse_moe.experts.4.w2",
829
+ "model.layers.25.block_sparse_moe.experts.4.w3",
830
+ "model.layers.25.block_sparse_moe.experts.5.w1",
831
+ "model.layers.25.block_sparse_moe.experts.5.w2",
832
+ "model.layers.25.block_sparse_moe.experts.5.w3",
833
+ "model.layers.25.block_sparse_moe.experts.6.w1",
834
+ "model.layers.25.block_sparse_moe.experts.6.w2",
835
+ "model.layers.25.block_sparse_moe.experts.6.w3",
836
+ "model.layers.25.block_sparse_moe.experts.7.w1",
837
+ "model.layers.25.block_sparse_moe.experts.7.w2",
838
+ "model.layers.25.block_sparse_moe.experts.7.w3",
839
+ "model.layers.26.self_attn.q_proj",
840
+ "model.layers.26.self_attn.k_proj",
841
+ "model.layers.26.self_attn.v_proj",
842
+ "model.layers.26.self_attn.o_proj",
843
+ "model.layers.26.block_sparse_moe.gate",
844
+ "model.layers.26.block_sparse_moe.experts.0.w1",
845
+ "model.layers.26.block_sparse_moe.experts.0.w2",
846
+ "model.layers.26.block_sparse_moe.experts.0.w3",
847
+ "model.layers.26.block_sparse_moe.experts.1.w1",
848
+ "model.layers.26.block_sparse_moe.experts.1.w2",
849
+ "model.layers.26.block_sparse_moe.experts.1.w3",
850
+ "model.layers.26.block_sparse_moe.experts.2.w1",
851
+ "model.layers.26.block_sparse_moe.experts.2.w2",
852
+ "model.layers.26.block_sparse_moe.experts.2.w3",
853
+ "model.layers.26.block_sparse_moe.experts.3.w1",
854
+ "model.layers.26.block_sparse_moe.experts.3.w2",
855
+ "model.layers.26.block_sparse_moe.experts.3.w3",
856
+ "model.layers.26.block_sparse_moe.experts.4.w1",
857
+ "model.layers.26.block_sparse_moe.experts.4.w2",
858
+ "model.layers.26.block_sparse_moe.experts.4.w3",
859
+ "model.layers.26.block_sparse_moe.experts.5.w1",
860
+ "model.layers.26.block_sparse_moe.experts.5.w2",
861
+ "model.layers.26.block_sparse_moe.experts.5.w3",
862
+ "model.layers.26.block_sparse_moe.experts.6.w1",
863
+ "model.layers.26.block_sparse_moe.experts.6.w2",
864
+ "model.layers.26.block_sparse_moe.experts.6.w3",
865
+ "model.layers.26.block_sparse_moe.experts.7.w1",
866
+ "model.layers.26.block_sparse_moe.experts.7.w2",
867
+ "model.layers.26.block_sparse_moe.experts.7.w3",
868
+ "model.layers.27.self_attn.q_proj",
869
+ "model.layers.27.self_attn.k_proj",
870
+ "model.layers.27.self_attn.v_proj",
871
+ "model.layers.27.self_attn.o_proj",
872
+ "model.layers.27.block_sparse_moe.gate",
873
+ "model.layers.27.block_sparse_moe.experts.0.w1",
874
+ "model.layers.27.block_sparse_moe.experts.0.w2",
875
+ "model.layers.27.block_sparse_moe.experts.0.w3",
876
+ "model.layers.27.block_sparse_moe.experts.1.w1",
877
+ "model.layers.27.block_sparse_moe.experts.1.w2",
878
+ "model.layers.27.block_sparse_moe.experts.1.w3",
879
+ "model.layers.27.block_sparse_moe.experts.2.w1",
880
+ "model.layers.27.block_sparse_moe.experts.2.w2",
881
+ "model.layers.27.block_sparse_moe.experts.2.w3",
882
+ "model.layers.27.block_sparse_moe.experts.3.w1",
883
+ "model.layers.27.block_sparse_moe.experts.3.w2",
884
+ "model.layers.27.block_sparse_moe.experts.3.w3",
885
+ "model.layers.27.block_sparse_moe.experts.4.w1",
886
+ "model.layers.27.block_sparse_moe.experts.4.w2",
887
+ "model.layers.27.block_sparse_moe.experts.4.w3",
888
+ "model.layers.27.block_sparse_moe.experts.5.w1",
889
+ "model.layers.27.block_sparse_moe.experts.5.w2",
890
+ "model.layers.27.block_sparse_moe.experts.5.w3",
891
+ "model.layers.27.block_sparse_moe.experts.6.w1",
892
+ "model.layers.27.block_sparse_moe.experts.6.w2",
893
+ "model.layers.27.block_sparse_moe.experts.6.w3",
894
+ "model.layers.27.block_sparse_moe.experts.7.w1",
895
+ "model.layers.27.block_sparse_moe.experts.7.w2",
896
+ "model.layers.27.block_sparse_moe.experts.7.w3",
897
+ "model.layers.28.self_attn.q_proj",
898
+ "model.layers.28.self_attn.k_proj",
899
+ "model.layers.28.self_attn.v_proj",
900
+ "model.layers.28.self_attn.o_proj",
901
+ "model.layers.28.block_sparse_moe.gate",
902
+ "model.layers.28.block_sparse_moe.experts.0.w1",
903
+ "model.layers.28.block_sparse_moe.experts.0.w2",
904
+ "model.layers.28.block_sparse_moe.experts.0.w3",
905
+ "model.layers.28.block_sparse_moe.experts.1.w1",
906
+ "model.layers.28.block_sparse_moe.experts.1.w2",
907
+ "model.layers.28.block_sparse_moe.experts.1.w3",
908
+ "model.layers.28.block_sparse_moe.experts.2.w1",
909
+ "model.layers.28.block_sparse_moe.experts.2.w2",
910
+ "model.layers.28.block_sparse_moe.experts.2.w3",
911
+ "model.layers.28.block_sparse_moe.experts.3.w1",
912
+ "model.layers.28.block_sparse_moe.experts.3.w2",
913
+ "model.layers.28.block_sparse_moe.experts.3.w3",
914
+ "model.layers.28.block_sparse_moe.experts.4.w1",
915
+ "model.layers.28.block_sparse_moe.experts.4.w2",
916
+ "model.layers.28.block_sparse_moe.experts.4.w3",
917
+ "model.layers.28.block_sparse_moe.experts.5.w1",
918
+ "model.layers.28.block_sparse_moe.experts.5.w2",
919
+ "model.layers.28.block_sparse_moe.experts.5.w3",
920
+ "model.layers.28.block_sparse_moe.experts.6.w1",
921
+ "model.layers.28.block_sparse_moe.experts.6.w2",
922
+ "model.layers.28.block_sparse_moe.experts.6.w3",
923
+ "model.layers.28.block_sparse_moe.experts.7.w1",
924
+ "model.layers.28.block_sparse_moe.experts.7.w2",
925
+ "model.layers.28.block_sparse_moe.experts.7.w3",
926
+ "model.layers.29.self_attn.q_proj",
927
+ "model.layers.29.self_attn.k_proj",
928
+ "model.layers.29.self_attn.v_proj",
929
+ "model.layers.29.self_attn.o_proj",
930
+ "model.layers.29.block_sparse_moe.gate",
931
+ "model.layers.29.block_sparse_moe.experts.0.w1",
932
+ "model.layers.29.block_sparse_moe.experts.0.w2",
933
+ "model.layers.29.block_sparse_moe.experts.0.w3",
934
+ "model.layers.29.block_sparse_moe.experts.1.w1",
935
+ "model.layers.29.block_sparse_moe.experts.1.w2",
936
+ "model.layers.29.block_sparse_moe.experts.1.w3",
937
+ "model.layers.29.block_sparse_moe.experts.2.w1",
938
+ "model.layers.29.block_sparse_moe.experts.2.w2",
939
+ "model.layers.29.block_sparse_moe.experts.2.w3",
940
+ "model.layers.29.block_sparse_moe.experts.3.w1",
941
+ "model.layers.29.block_sparse_moe.experts.3.w2",
942
+ "model.layers.29.block_sparse_moe.experts.3.w3",
943
+ "model.layers.29.block_sparse_moe.experts.4.w1",
944
+ "model.layers.29.block_sparse_moe.experts.4.w2",
945
+ "model.layers.29.block_sparse_moe.experts.4.w3",
946
+ "model.layers.29.block_sparse_moe.experts.5.w1",
947
+ "model.layers.29.block_sparse_moe.experts.5.w2",
948
+ "model.layers.29.block_sparse_moe.experts.5.w3",
949
+ "model.layers.29.block_sparse_moe.experts.6.w1",
950
+ "model.layers.29.block_sparse_moe.experts.6.w2",
951
+ "model.layers.29.block_sparse_moe.experts.6.w3",
952
+ "model.layers.29.block_sparse_moe.experts.7.w1",
953
+ "model.layers.29.block_sparse_moe.experts.7.w2",
954
+ "model.layers.29.block_sparse_moe.experts.7.w3",
955
+ "model.layers.30.self_attn.q_proj",
956
+ "model.layers.30.self_attn.k_proj",
957
+ "model.layers.30.self_attn.v_proj",
958
+ "model.layers.30.self_attn.o_proj",
959
+ "model.layers.30.block_sparse_moe.gate",
960
+ "model.layers.30.block_sparse_moe.experts.0.w1",
961
+ "model.layers.30.block_sparse_moe.experts.0.w2",
962
+ "model.layers.30.block_sparse_moe.experts.0.w3",
963
+ "model.layers.30.block_sparse_moe.experts.1.w1",
964
+ "model.layers.30.block_sparse_moe.experts.1.w2",
965
+ "model.layers.30.block_sparse_moe.experts.1.w3",
966
+ "model.layers.30.block_sparse_moe.experts.2.w1",
967
+ "model.layers.30.block_sparse_moe.experts.2.w2",
968
+ "model.layers.30.block_sparse_moe.experts.2.w3",
969
+ "model.layers.30.block_sparse_moe.experts.3.w1",
970
+ "model.layers.30.block_sparse_moe.experts.3.w2",
971
+ "model.layers.30.block_sparse_moe.experts.3.w3",
972
+ "model.layers.30.block_sparse_moe.experts.4.w1",
973
+ "model.layers.30.block_sparse_moe.experts.4.w2",
974
+ "model.layers.30.block_sparse_moe.experts.4.w3",
975
+ "model.layers.30.block_sparse_moe.experts.5.w1",
976
+ "model.layers.30.block_sparse_moe.experts.5.w2",
977
+ "model.layers.30.block_sparse_moe.experts.5.w3",
978
+ "model.layers.30.block_sparse_moe.experts.6.w1",
979
+ "model.layers.30.block_sparse_moe.experts.6.w2",
980
+ "model.layers.30.block_sparse_moe.experts.6.w3",
981
+ "model.layers.30.block_sparse_moe.experts.7.w1",
982
+ "model.layers.30.block_sparse_moe.experts.7.w2",
983
+ "model.layers.30.block_sparse_moe.experts.7.w3",
984
+ "model.layers.31.self_attn.q_proj",
985
+ "model.layers.31.self_attn.k_proj",
986
+ "model.layers.31.self_attn.v_proj",
987
+ "model.layers.31.self_attn.o_proj",
988
+ "model.layers.31.block_sparse_moe.gate",
989
+ "model.layers.31.block_sparse_moe.experts.0.w1",
990
+ "model.layers.31.block_sparse_moe.experts.0.w2",
991
+ "model.layers.31.block_sparse_moe.experts.0.w3",
992
+ "model.layers.31.block_sparse_moe.experts.1.w1",
993
+ "model.layers.31.block_sparse_moe.experts.1.w2",
994
+ "model.layers.31.block_sparse_moe.experts.1.w3",
995
+ "model.layers.31.block_sparse_moe.experts.2.w1",
996
+ "model.layers.31.block_sparse_moe.experts.2.w2",
997
+ "model.layers.31.block_sparse_moe.experts.2.w3",
998
+ "model.layers.31.block_sparse_moe.experts.3.w1",
999
+ "model.layers.31.block_sparse_moe.experts.3.w2",
1000
+ "model.layers.31.block_sparse_moe.experts.3.w3",
1001
+ "model.layers.31.block_sparse_moe.experts.4.w1",
1002
+ "model.layers.31.block_sparse_moe.experts.4.w2",
1003
+ "model.layers.31.block_sparse_moe.experts.4.w3",
1004
+ "model.layers.31.block_sparse_moe.experts.5.w1",
1005
+ "model.layers.31.block_sparse_moe.experts.5.w2",
1006
+ "model.layers.31.block_sparse_moe.experts.5.w3",
1007
+ "model.layers.31.block_sparse_moe.experts.6.w1",
1008
+ "model.layers.31.block_sparse_moe.experts.6.w2",
1009
+ "model.layers.31.block_sparse_moe.experts.6.w3",
1010
+ "model.layers.31.block_sparse_moe.experts.7.w1",
1011
+ "model.layers.31.block_sparse_moe.experts.7.w2",
1012
+ "model.layers.31.block_sparse_moe.experts.7.w3",
1013
+ "lm_head"
1014
+ ],
1015
+ "registry_requires_subclass": false,
1016
+ "sparsity_structure": "unstructured",
1017
+ "targets": [
1018
+ "model.layers.0.self_attn.q_proj",
1019
+ "model.layers.0.self_attn.k_proj"
1020
+ ]
1021
+ }
1022
+ },
1023
+ "rms_norm_eps": 1e-05,
1024
+ "rope_theta": 1000000.0,
1025
+ "router_aux_loss_coef": 0.02,
1026
+ "router_jitter_noise": 0.0,
1027
+ "sliding_window": null,
1028
+ "tie_word_embeddings": false,
1029
+ "torch_dtype": "bfloat16",
1030
+ "transformers_version": "4.46.2",
1031
+ "use_cache": false,
1032
+ "vocab_size": 32000
1033
+ }
generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "transformers_version": "4.46.2"
6
+ }
model-00001-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bde549ae91e28de9910d65eccbdeef0deeac9673c84a546f1449a916e64c4817
3
+ size 4986084728
model-00002-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7fa87573827abeadfabcd3a3c1b1714018de0ddfec73e9bddfda4271781518f4
3
+ size 4974749080
model-00003-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:264fcbf59ebb0d313ed93ba35ed45616db6278d6c9a484a0443007ecb1f453a3
3
+ size 4996460000
model-00004-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd4e4abd73414b3ad7181e4117a5aa3698912f1abd5c621b8bdf0dfd3fadc6ae
3
+ size 4996460000
model-00005-of-00005.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f65d8ac389ce922c63b429f4f77800e27c82fcefe642e0a0386255d7b98af977
3
+ size 4518812800
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
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_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "add_prefix_space": null,
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
+ }