xianchaowu
commited on
Commit
·
47787a3
1
Parent(s):
f05e11e
upload lazy lora for llama1-33b
Browse files- README.md +157 -0
- adapter_config.json +455 -0
- adapter_model.bin +3 -0
- usage.py +51 -0
README.md
CHANGED
|
@@ -1,3 +1,160 @@
|
|
| 1 |
---
|
| 2 |
license: llama2
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: llama2
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
## Lazy LoRA
|
| 6 |
+
|
| 7 |
+
### Benefits
|
| 8 |
+
|
| 9 |
+
0. using the (former, since 33b model is not included in llama-2 for the public)[Meta's LLaMA-1 models](https://huggingface.co/huggyllama/llama-30b).
|
| 10 |
+
1. support [4-bit qlora](https://arxiv.org/abs/2305.14314), extreme GPU memory and inference time saving;
|
| 11 |
+
2. comparable (slightly worse, mainly due to 4-bit) MMLU evaluation dataset results, llama1-33b's 57.8% to our 56.97% (-0.83%).
|
| 12 |
+
3. This lazy-lora adapter is based on [Meta's LLaMA-1](https://huggingface.co/huggyllama/llama-30b), and using the [oasst1 dataset](https://huggingface.co/datasets/OpenAssistant/oasst1), following [Guanaco](https://huggingface.co/timdettmers/guanaco-65b).
|
| 13 |
+
|
| 14 |
+
### Introduction
|
| 15 |
+
Determine the rank of LoRA layers by the singular values of pretrained weight matrices.
|
| 16 |
+
Also, combines:
|
| 17 |
+
1. LoRA: [LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS](https://arxiv.org/abs/2106.09685)
|
| 18 |
+
2. Prefix Tuning: [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https://aclanthology.org/2021.acl-long.3
|
| 19 |
+
53/), [P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks](https://arxiv.or
|
| 20 |
+
g/pdf/2110.07602.pdf)
|
| 21 |
+
3. Prompt Tuning: [The Power of Scale for Parameter-Efficient Prompt Tuning](https://arxiv.org/abs/2104.08691)
|
| 22 |
+
4. LLaMA adapter: [LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention] (https://arxiv.org/abs/2303.16199)
|
| 23 |
+
in one model.
|
| 24 |
+
|
| 25 |
+
This allows you to perform LoRA (additional low rank adapters inserted to each linear layer), and prompt learning (additional virtual tokens attached to the input and to the attention layers acting as `past_key_values`)
|
| 26 |
+
|
| 27 |
+
## Usage:
|
| 28 |
+
```python
|
| 29 |
+
import sys
|
| 30 |
+
sys.path.insert(1, '/workspace/asr/peft/src')
|
| 31 |
+
# TODO set this path to the lazy-lora source code path,
|
| 32 |
+
# or you can install it from source code:
|
| 33 |
+
# TODO, please install lazylora for usage:
|
| 34 |
+
# git clone [email protected]:Xianchao-Wu/peft.git
|
| 35 |
+
# cd peft
|
| 36 |
+
# python setup.py install
|
| 37 |
+
|
| 38 |
+
from transformers import (AutoTokenizer,
|
| 39 |
+
AutoModelForCausalLM, BitsAndBytesConfig)
|
| 40 |
+
from peft import PeftModel, PeftConfig
|
| 41 |
+
import os
|
| 42 |
+
import torch
|
| 43 |
+
|
| 44 |
+
#import ipdb; ipdb.set_trace()
|
| 45 |
+
cache_dir="/workspace/asr/peft/qlora"
|
| 46 |
+
# TODO set this cache_dir to the path where you
|
| 47 |
+
# stored (or, want to store) llama1-33b model
|
| 48 |
+
|
| 49 |
+
lazylora_dir=os.getcwd()
|
| 50 |
+
# the path that contains 'adapter_config.json'
|
| 51 |
+
# and 'adapter_model.bin'
|
| 52 |
+
|
| 53 |
+
config = PeftConfig.from_pretrained(lazylora_dir)
|
| 54 |
+
|
| 55 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 56 |
+
config.base_model_name_or_path,
|
| 57 |
+
cache_dir=cache_dir,
|
| 58 |
+
use_auth_token=True
|
| 59 |
+
)
|
| 60 |
+
|
| 61 |
+
bnb_config = BitsAndBytesConfig(
|
| 62 |
+
load_in_4bit=True,
|
| 63 |
+
bnb_4bit_use_double_quant=True,
|
| 64 |
+
bnb_4bit_quant_type='nf4',
|
| 65 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
| 66 |
+
)
|
| 67 |
+
|
| 68 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 69 |
+
config.base_model_name_or_path,
|
| 70 |
+
quantization_config=bnb_config,
|
| 71 |
+
device_map="auto",
|
| 72 |
+
cache_dir=cache_dir,
|
| 73 |
+
use_auth_token=True
|
| 74 |
+
)
|
| 75 |
+
#model.print_trainable_parameters()
|
| 76 |
+
print(sum(p.numel() for p in model.parameters()))
|
| 77 |
+
# 16,477,866,496 -> half-size of 33B due to 4-bit loading
|
| 78 |
+
|
| 79 |
+
model = PeftModel.from_pretrained(model, lazylora_dir)
|
| 80 |
+
print('after adding lazy lora parameters:')
|
| 81 |
+
model.print_trainable_parameters()
|
| 82 |
+
# trainable params: 0 || all params: 16,965,645,824 || trainable%: 0.0
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
## MMLU result:
|
| 86 |
+
|
| 87 |
+
```json
|
| 88 |
+
{"mmlu_loss": 2.6712945443520275,
|
| 89 |
+
"mmlu_eval_accuracy_college_chemistry": 0.125,
|
| 90 |
+
"mmlu_eval_accuracy_philosophy": 0.7647058823529411,
|
| 91 |
+
"mmlu_eval_accuracy_virology": 0.3888888888888889,
|
| 92 |
+
"mmlu_eval_accuracy_high_school_european_history": 0.8333333333333334,
|
| 93 |
+
"mmlu_eval_accuracy_astronomy": 0.6875,
|
| 94 |
+
"mmlu_eval_accuracy_elementary_mathematics": 0.34146341463414637,
|
| 95 |
+
"mmlu_eval_accuracy_business_ethics": 0.5454545454545454,
|
| 96 |
+
"mmlu_eval_accuracy_computer_security": 0.8181818181818182,
|
| 97 |
+
"mmlu_eval_accuracy_anatomy": 0.5,
|
| 98 |
+
"mmlu_eval_accuracy_high_school_physics": 0.23529411764705882,
|
| 99 |
+
"mmlu_eval_accuracy_high_school_government_and_politics": 0.7619047619047619,
|
| 100 |
+
"mmlu_eval_accuracy_global_facts": 0.4,
|
| 101 |
+
"mmlu_eval_accuracy_logical_fallacies": 0.6666666666666666,
|
| 102 |
+
"mmlu_eval_accuracy_security_studies": 0.7037037037037037,
|
| 103 |
+
"mmlu_eval_accuracy_world_religions": 0.8421052631578947,
|
| 104 |
+
"mmlu_eval_accuracy_professional_medicine": 0.7096774193548387,
|
| 105 |
+
"mmlu_eval_accuracy_management": 0.9090909090909091,
|
| 106 |
+
"mmlu_eval_accuracy_marketing": 0.8,
|
| 107 |
+
"mmlu_eval_accuracy_college_physics": 0.36363636363636365,
|
| 108 |
+
"mmlu_eval_accuracy_professional_law": 0.4294117647058823,
|
| 109 |
+
"mmlu_eval_accuracy_college_mathematics": 0.36363636363636365,
|
| 110 |
+
"mmlu_eval_accuracy_high_school_psychology": 0.8333333333333334,
|
| 111 |
+
"mmlu_eval_accuracy_moral_disputes": 0.5789473684210527,
|
| 112 |
+
"mmlu_eval_accuracy_professional_accounting": 0.45161290322580644,
|
| 113 |
+
"mmlu_eval_accuracy_conceptual_physics": 0.4230769230769231,
|
| 114 |
+
"mmlu_eval_accuracy_high_school_chemistry": 0.36363636363636365,
|
| 115 |
+
"mmlu_eval_accuracy_nutrition": 0.7272727272727273,
|
| 116 |
+
"mmlu_eval_accuracy_high_school_geography": 0.7272727272727273,
|
| 117 |
+
"mmlu_eval_accuracy_high_school_statistics": 0.43478260869565216,
|
| 118 |
+
"mmlu_eval_accuracy_prehistory": 0.5714285714285714,
|
| 119 |
+
"mmlu_eval_accuracy_public_relations": 0.5833333333333334,
|
| 120 |
+
"mmlu_eval_accuracy_jurisprudence": 0.5454545454545454,
|
| 121 |
+
"mmlu_eval_accuracy_moral_scenarios": 0.4,
|
| 122 |
+
"mmlu_eval_accuracy_sociology": 0.8181818181818182,
|
| 123 |
+
"mmlu_eval_accuracy_college_biology": 0.5,
|
| 124 |
+
"mmlu_eval_accuracy_human_aging": 0.6521739130434783,
|
| 125 |
+
"mmlu_eval_accuracy_abstract_algebra": 0.36363636363636365,
|
| 126 |
+
"mmlu_eval_accuracy_high_school_computer_science": 0.6666666666666666,
|
| 127 |
+
"mmlu_eval_accuracy_electrical_engineering": 0.3125,
|
| 128 |
+
"mmlu_eval_accuracy_medical_genetics": 0.8181818181818182,
|
| 129 |
+
"mmlu_eval_accuracy_clinical_knowledge": 0.4827586206896552,
|
| 130 |
+
"mmlu_eval_accuracy_high_school_macroeconomics": 0.5813953488372093,
|
| 131 |
+
"mmlu_eval_accuracy_college_medicine": 0.5,
|
| 132 |
+
"mmlu_eval_accuracy_high_school_world_history": 0.6923076923076923,
|
| 133 |
+
"mmlu_eval_accuracy_high_school_mathematics": 0.3448275862068966,
|
| 134 |
+
"mmlu_eval_accuracy_international_law": 0.9230769230769231,
|
| 135 |
+
"mmlu_eval_accuracy_miscellaneous": 0.7558139534883721,
|
| 136 |
+
"mmlu_eval_accuracy_human_sexuality": 0.4166666666666667,
|
| 137 |
+
"mmlu_eval_accuracy_professional_psychology": 0.5942028985507246,
|
| 138 |
+
"mmlu_eval_accuracy_econometrics": 0.4166666666666667,
|
| 139 |
+
"mmlu_eval_accuracy_high_school_microeconomics": 0.5384615384615384,
|
| 140 |
+
"mmlu_eval_accuracy_us_foreign_policy": 0.9090909090909091,
|
| 141 |
+
"mmlu_eval_accuracy_machine_learning": 0.45454545454545453,
|
| 142 |
+
"mmlu_eval_accuracy_high_school_biology": 0.53125,
|
| 143 |
+
"mmlu_eval_accuracy_formal_logic": 0.14285714285714285,
|
| 144 |
+
"mmlu_eval_accuracy_high_school_us_history": 0.8636363636363636,
|
| 145 |
+
"mmlu_eval_accuracy_college_computer_science": 0.36363636363636365,
|
| 146 |
+
"mmlu_eval_accuracy": 0.5696901987706997,
|
| 147 |
+
"epoch": 3.05}
|
| 148 |
+
```
|
| 149 |
+
|
| 150 |
+
## License and intended use
|
| 151 |
+
|
| 152 |
+
This lazy-lora adapter is based on [Meta's LLaMA1-33b, huggyllama/llama-30b](https://huggingface.co/huggyllama/llama-30b), and using the [oasst1 dataset](https://huggingface.co/datasets/OpenAssistant/oasst1), following [Guanaco](https://huggingface.co/timdettmers/guanaco-65b).
|
| 153 |
+
|
| 154 |
+
lazy lora adapter weights are available under LLAMA-2 license. Note the use of the lazy lora adapter weights, requires access to the LLaMA model weighs. Lazy lora is based on LLaMA and therefore should be used according to the LLaMA license.
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
## Risks and Biases
|
| 158 |
+
|
| 159 |
+
The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. The model was trained on various public datasets; it is possible that this model could generate lewd, biased, or otherwise offensive outputs.
|
| 160 |
+
|
adapter_config.json
ADDED
|
@@ -0,0 +1,455 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"base_model_name_or_path": "huggyllama/llama-30b",
|
| 3 |
+
"bias": "none",
|
| 4 |
+
"fan_in_fan_out": false,
|
| 5 |
+
"inference_mode": true,
|
| 6 |
+
"init_lazy_lora_weights": true,
|
| 7 |
+
"is_r_by_svd": true,
|
| 8 |
+
"is_r_reuse": true,
|
| 9 |
+
"lazy_lora_alpha": 16.0,
|
| 10 |
+
"lazy_lora_dropout": 0.05,
|
| 11 |
+
"lazy_pre_adapter_type": "none",
|
| 12 |
+
"lazy_pre_lora_alpha": 0.1,
|
| 13 |
+
"modules_to_save": null,
|
| 14 |
+
"num_attention_heads": 52,
|
| 15 |
+
"num_layers": 60,
|
| 16 |
+
"num_transformer_submodules": 1,
|
| 17 |
+
"num_virtual_tokens": null,
|
| 18 |
+
"peft_type": "LAZY_LORA",
|
| 19 |
+
"prefix_tuning_config": null,
|
| 20 |
+
"prompt_tuning_config": null,
|
| 21 |
+
"r": 64,
|
| 22 |
+
"r_by_module_dict": {
|
| 23 |
+
"model.layers.0.mlp.down_proj": 55,
|
| 24 |
+
"model.layers.0.mlp.gate_proj": 28,
|
| 25 |
+
"model.layers.0.mlp.up_proj": 29,
|
| 26 |
+
"model.layers.0.self_attn.k_proj": 36,
|
| 27 |
+
"model.layers.0.self_attn.o_proj": 29,
|
| 28 |
+
"model.layers.0.self_attn.q_proj": 39,
|
| 29 |
+
"model.layers.0.self_attn.v_proj": 28,
|
| 30 |
+
"model.layers.1.mlp.down_proj": 61,
|
| 31 |
+
"model.layers.1.mlp.gate_proj": 49,
|
| 32 |
+
"model.layers.1.mlp.up_proj": 50,
|
| 33 |
+
"model.layers.1.self_attn.k_proj": 60,
|
| 34 |
+
"model.layers.1.self_attn.o_proj": 36,
|
| 35 |
+
"model.layers.1.self_attn.q_proj": 59,
|
| 36 |
+
"model.layers.1.self_attn.v_proj": 33,
|
| 37 |
+
"model.layers.10.mlp.down_proj": 63,
|
| 38 |
+
"model.layers.10.mlp.gate_proj": 64,
|
| 39 |
+
"model.layers.10.mlp.up_proj": 64,
|
| 40 |
+
"model.layers.10.self_attn.k_proj": 79,
|
| 41 |
+
"model.layers.10.self_attn.o_proj": 58,
|
| 42 |
+
"model.layers.10.self_attn.q_proj": 78,
|
| 43 |
+
"model.layers.10.self_attn.v_proj": 58,
|
| 44 |
+
"model.layers.11.mlp.down_proj": 63,
|
| 45 |
+
"model.layers.11.mlp.gate_proj": 64,
|
| 46 |
+
"model.layers.11.mlp.up_proj": 65,
|
| 47 |
+
"model.layers.11.self_attn.k_proj": 77,
|
| 48 |
+
"model.layers.11.self_attn.o_proj": 59,
|
| 49 |
+
"model.layers.11.self_attn.q_proj": 75,
|
| 50 |
+
"model.layers.11.self_attn.v_proj": 59,
|
| 51 |
+
"model.layers.12.mlp.down_proj": 62,
|
| 52 |
+
"model.layers.12.mlp.gate_proj": 64,
|
| 53 |
+
"model.layers.12.mlp.up_proj": 65,
|
| 54 |
+
"model.layers.12.self_attn.k_proj": 77,
|
| 55 |
+
"model.layers.12.self_attn.o_proj": 59,
|
| 56 |
+
"model.layers.12.self_attn.q_proj": 76,
|
| 57 |
+
"model.layers.12.self_attn.v_proj": 59,
|
| 58 |
+
"model.layers.13.mlp.down_proj": 63,
|
| 59 |
+
"model.layers.13.mlp.gate_proj": 63,
|
| 60 |
+
"model.layers.13.mlp.up_proj": 65,
|
| 61 |
+
"model.layers.13.self_attn.k_proj": 76,
|
| 62 |
+
"model.layers.13.self_attn.o_proj": 58,
|
| 63 |
+
"model.layers.13.self_attn.q_proj": 75,
|
| 64 |
+
"model.layers.13.self_attn.v_proj": 58,
|
| 65 |
+
"model.layers.14.mlp.down_proj": 63,
|
| 66 |
+
"model.layers.14.mlp.gate_proj": 63,
|
| 67 |
+
"model.layers.14.mlp.up_proj": 65,
|
| 68 |
+
"model.layers.14.self_attn.k_proj": 76,
|
| 69 |
+
"model.layers.14.self_attn.o_proj": 58,
|
| 70 |
+
"model.layers.14.self_attn.q_proj": 74,
|
| 71 |
+
"model.layers.14.self_attn.v_proj": 59,
|
| 72 |
+
"model.layers.15.mlp.down_proj": 63,
|
| 73 |
+
"model.layers.15.mlp.gate_proj": 63,
|
| 74 |
+
"model.layers.15.mlp.up_proj": 65,
|
| 75 |
+
"model.layers.15.self_attn.k_proj": 77,
|
| 76 |
+
"model.layers.15.self_attn.o_proj": 61,
|
| 77 |
+
"model.layers.15.self_attn.q_proj": 76,
|
| 78 |
+
"model.layers.15.self_attn.v_proj": 61,
|
| 79 |
+
"model.layers.16.mlp.down_proj": 63,
|
| 80 |
+
"model.layers.16.mlp.gate_proj": 63,
|
| 81 |
+
"model.layers.16.mlp.up_proj": 65,
|
| 82 |
+
"model.layers.16.self_attn.k_proj": 77,
|
| 83 |
+
"model.layers.16.self_attn.o_proj": 61,
|
| 84 |
+
"model.layers.16.self_attn.q_proj": 76,
|
| 85 |
+
"model.layers.16.self_attn.v_proj": 62,
|
| 86 |
+
"model.layers.17.mlp.down_proj": 63,
|
| 87 |
+
"model.layers.17.mlp.gate_proj": 63,
|
| 88 |
+
"model.layers.17.mlp.up_proj": 66,
|
| 89 |
+
"model.layers.17.self_attn.k_proj": 75,
|
| 90 |
+
"model.layers.17.self_attn.o_proj": 61,
|
| 91 |
+
"model.layers.17.self_attn.q_proj": 74,
|
| 92 |
+
"model.layers.17.self_attn.v_proj": 62,
|
| 93 |
+
"model.layers.18.mlp.down_proj": 64,
|
| 94 |
+
"model.layers.18.mlp.gate_proj": 63,
|
| 95 |
+
"model.layers.18.mlp.up_proj": 66,
|
| 96 |
+
"model.layers.18.self_attn.k_proj": 76,
|
| 97 |
+
"model.layers.18.self_attn.o_proj": 61,
|
| 98 |
+
"model.layers.18.self_attn.q_proj": 74,
|
| 99 |
+
"model.layers.18.self_attn.v_proj": 62,
|
| 100 |
+
"model.layers.19.mlp.down_proj": 64,
|
| 101 |
+
"model.layers.19.mlp.gate_proj": 63,
|
| 102 |
+
"model.layers.19.mlp.up_proj": 66,
|
| 103 |
+
"model.layers.19.self_attn.k_proj": 74,
|
| 104 |
+
"model.layers.19.self_attn.o_proj": 63,
|
| 105 |
+
"model.layers.19.self_attn.q_proj": 73,
|
| 106 |
+
"model.layers.19.self_attn.v_proj": 64,
|
| 107 |
+
"model.layers.2.mlp.down_proj": 62,
|
| 108 |
+
"model.layers.2.mlp.gate_proj": 58,
|
| 109 |
+
"model.layers.2.mlp.up_proj": 58,
|
| 110 |
+
"model.layers.2.self_attn.k_proj": 75,
|
| 111 |
+
"model.layers.2.self_attn.o_proj": 48,
|
| 112 |
+
"model.layers.2.self_attn.q_proj": 75,
|
| 113 |
+
"model.layers.2.self_attn.v_proj": 44,
|
| 114 |
+
"model.layers.20.mlp.down_proj": 64,
|
| 115 |
+
"model.layers.20.mlp.gate_proj": 63,
|
| 116 |
+
"model.layers.20.mlp.up_proj": 66,
|
| 117 |
+
"model.layers.20.self_attn.k_proj": 70,
|
| 118 |
+
"model.layers.20.self_attn.o_proj": 63,
|
| 119 |
+
"model.layers.20.self_attn.q_proj": 70,
|
| 120 |
+
"model.layers.20.self_attn.v_proj": 64,
|
| 121 |
+
"model.layers.21.mlp.down_proj": 64,
|
| 122 |
+
"model.layers.21.mlp.gate_proj": 63,
|
| 123 |
+
"model.layers.21.mlp.up_proj": 66,
|
| 124 |
+
"model.layers.21.self_attn.k_proj": 72,
|
| 125 |
+
"model.layers.21.self_attn.o_proj": 64,
|
| 126 |
+
"model.layers.21.self_attn.q_proj": 71,
|
| 127 |
+
"model.layers.21.self_attn.v_proj": 64,
|
| 128 |
+
"model.layers.22.mlp.down_proj": 65,
|
| 129 |
+
"model.layers.22.mlp.gate_proj": 63,
|
| 130 |
+
"model.layers.22.mlp.up_proj": 65,
|
| 131 |
+
"model.layers.22.self_attn.k_proj": 67,
|
| 132 |
+
"model.layers.22.self_attn.o_proj": 66,
|
| 133 |
+
"model.layers.22.self_attn.q_proj": 66,
|
| 134 |
+
"model.layers.22.self_attn.v_proj": 66,
|
| 135 |
+
"model.layers.23.mlp.down_proj": 65,
|
| 136 |
+
"model.layers.23.mlp.gate_proj": 64,
|
| 137 |
+
"model.layers.23.mlp.up_proj": 66,
|
| 138 |
+
"model.layers.23.self_attn.k_proj": 68,
|
| 139 |
+
"model.layers.23.self_attn.o_proj": 65,
|
| 140 |
+
"model.layers.23.self_attn.q_proj": 67,
|
| 141 |
+
"model.layers.23.self_attn.v_proj": 65,
|
| 142 |
+
"model.layers.24.mlp.down_proj": 65,
|
| 143 |
+
"model.layers.24.mlp.gate_proj": 64,
|
| 144 |
+
"model.layers.24.mlp.up_proj": 65,
|
| 145 |
+
"model.layers.24.self_attn.k_proj": 69,
|
| 146 |
+
"model.layers.24.self_attn.o_proj": 66,
|
| 147 |
+
"model.layers.24.self_attn.q_proj": 69,
|
| 148 |
+
"model.layers.24.self_attn.v_proj": 67,
|
| 149 |
+
"model.layers.25.mlp.down_proj": 65,
|
| 150 |
+
"model.layers.25.mlp.gate_proj": 64,
|
| 151 |
+
"model.layers.25.mlp.up_proj": 65,
|
| 152 |
+
"model.layers.25.self_attn.k_proj": 72,
|
| 153 |
+
"model.layers.25.self_attn.o_proj": 66,
|
| 154 |
+
"model.layers.25.self_attn.q_proj": 71,
|
| 155 |
+
"model.layers.25.self_attn.v_proj": 66,
|
| 156 |
+
"model.layers.26.mlp.down_proj": 65,
|
| 157 |
+
"model.layers.26.mlp.gate_proj": 64,
|
| 158 |
+
"model.layers.26.mlp.up_proj": 65,
|
| 159 |
+
"model.layers.26.self_attn.k_proj": 67,
|
| 160 |
+
"model.layers.26.self_attn.o_proj": 68,
|
| 161 |
+
"model.layers.26.self_attn.q_proj": 67,
|
| 162 |
+
"model.layers.26.self_attn.v_proj": 67,
|
| 163 |
+
"model.layers.27.mlp.down_proj": 65,
|
| 164 |
+
"model.layers.27.mlp.gate_proj": 64,
|
| 165 |
+
"model.layers.27.mlp.up_proj": 64,
|
| 166 |
+
"model.layers.27.self_attn.k_proj": 62,
|
| 167 |
+
"model.layers.27.self_attn.o_proj": 67,
|
| 168 |
+
"model.layers.27.self_attn.q_proj": 62,
|
| 169 |
+
"model.layers.27.self_attn.v_proj": 67,
|
| 170 |
+
"model.layers.28.mlp.down_proj": 65,
|
| 171 |
+
"model.layers.28.mlp.gate_proj": 65,
|
| 172 |
+
"model.layers.28.mlp.up_proj": 65,
|
| 173 |
+
"model.layers.28.self_attn.k_proj": 64,
|
| 174 |
+
"model.layers.28.self_attn.o_proj": 69,
|
| 175 |
+
"model.layers.28.self_attn.q_proj": 64,
|
| 176 |
+
"model.layers.28.self_attn.v_proj": 69,
|
| 177 |
+
"model.layers.29.mlp.down_proj": 65,
|
| 178 |
+
"model.layers.29.mlp.gate_proj": 65,
|
| 179 |
+
"model.layers.29.mlp.up_proj": 64,
|
| 180 |
+
"model.layers.29.self_attn.k_proj": 60,
|
| 181 |
+
"model.layers.29.self_attn.o_proj": 69,
|
| 182 |
+
"model.layers.29.self_attn.q_proj": 61,
|
| 183 |
+
"model.layers.29.self_attn.v_proj": 69,
|
| 184 |
+
"model.layers.3.mlp.down_proj": 62,
|
| 185 |
+
"model.layers.3.mlp.gate_proj": 61,
|
| 186 |
+
"model.layers.3.mlp.up_proj": 60,
|
| 187 |
+
"model.layers.3.self_attn.k_proj": 80,
|
| 188 |
+
"model.layers.3.self_attn.o_proj": 48,
|
| 189 |
+
"model.layers.3.self_attn.q_proj": 78,
|
| 190 |
+
"model.layers.3.self_attn.v_proj": 46,
|
| 191 |
+
"model.layers.30.mlp.down_proj": 65,
|
| 192 |
+
"model.layers.30.mlp.gate_proj": 65,
|
| 193 |
+
"model.layers.30.mlp.up_proj": 64,
|
| 194 |
+
"model.layers.30.self_attn.k_proj": 62,
|
| 195 |
+
"model.layers.30.self_attn.o_proj": 69,
|
| 196 |
+
"model.layers.30.self_attn.q_proj": 62,
|
| 197 |
+
"model.layers.30.self_attn.v_proj": 69,
|
| 198 |
+
"model.layers.31.mlp.down_proj": 65,
|
| 199 |
+
"model.layers.31.mlp.gate_proj": 65,
|
| 200 |
+
"model.layers.31.mlp.up_proj": 64,
|
| 201 |
+
"model.layers.31.self_attn.k_proj": 60,
|
| 202 |
+
"model.layers.31.self_attn.o_proj": 69,
|
| 203 |
+
"model.layers.31.self_attn.q_proj": 61,
|
| 204 |
+
"model.layers.31.self_attn.v_proj": 68,
|
| 205 |
+
"model.layers.32.mlp.down_proj": 65,
|
| 206 |
+
"model.layers.32.mlp.gate_proj": 66,
|
| 207 |
+
"model.layers.32.mlp.up_proj": 64,
|
| 208 |
+
"model.layers.32.self_attn.k_proj": 59,
|
| 209 |
+
"model.layers.32.self_attn.o_proj": 69,
|
| 210 |
+
"model.layers.32.self_attn.q_proj": 60,
|
| 211 |
+
"model.layers.32.self_attn.v_proj": 69,
|
| 212 |
+
"model.layers.33.mlp.down_proj": 65,
|
| 213 |
+
"model.layers.33.mlp.gate_proj": 66,
|
| 214 |
+
"model.layers.33.mlp.up_proj": 64,
|
| 215 |
+
"model.layers.33.self_attn.k_proj": 65,
|
| 216 |
+
"model.layers.33.self_attn.o_proj": 70,
|
| 217 |
+
"model.layers.33.self_attn.q_proj": 65,
|
| 218 |
+
"model.layers.33.self_attn.v_proj": 69,
|
| 219 |
+
"model.layers.34.mlp.down_proj": 65,
|
| 220 |
+
"model.layers.34.mlp.gate_proj": 66,
|
| 221 |
+
"model.layers.34.mlp.up_proj": 64,
|
| 222 |
+
"model.layers.34.self_attn.k_proj": 63,
|
| 223 |
+
"model.layers.34.self_attn.o_proj": 69,
|
| 224 |
+
"model.layers.34.self_attn.q_proj": 63,
|
| 225 |
+
"model.layers.34.self_attn.v_proj": 69,
|
| 226 |
+
"model.layers.35.mlp.down_proj": 65,
|
| 227 |
+
"model.layers.35.mlp.gate_proj": 66,
|
| 228 |
+
"model.layers.35.mlp.up_proj": 64,
|
| 229 |
+
"model.layers.35.self_attn.k_proj": 61,
|
| 230 |
+
"model.layers.35.self_attn.o_proj": 69,
|
| 231 |
+
"model.layers.35.self_attn.q_proj": 61,
|
| 232 |
+
"model.layers.35.self_attn.v_proj": 69,
|
| 233 |
+
"model.layers.36.mlp.down_proj": 65,
|
| 234 |
+
"model.layers.36.mlp.gate_proj": 66,
|
| 235 |
+
"model.layers.36.mlp.up_proj": 64,
|
| 236 |
+
"model.layers.36.self_attn.k_proj": 63,
|
| 237 |
+
"model.layers.36.self_attn.o_proj": 70,
|
| 238 |
+
"model.layers.36.self_attn.q_proj": 63,
|
| 239 |
+
"model.layers.36.self_attn.v_proj": 70,
|
| 240 |
+
"model.layers.37.mlp.down_proj": 65,
|
| 241 |
+
"model.layers.37.mlp.gate_proj": 66,
|
| 242 |
+
"model.layers.37.mlp.up_proj": 64,
|
| 243 |
+
"model.layers.37.self_attn.k_proj": 59,
|
| 244 |
+
"model.layers.37.self_attn.o_proj": 70,
|
| 245 |
+
"model.layers.37.self_attn.q_proj": 60,
|
| 246 |
+
"model.layers.37.self_attn.v_proj": 70,
|
| 247 |
+
"model.layers.38.mlp.down_proj": 65,
|
| 248 |
+
"model.layers.38.mlp.gate_proj": 66,
|
| 249 |
+
"model.layers.38.mlp.up_proj": 64,
|
| 250 |
+
"model.layers.38.self_attn.k_proj": 57,
|
| 251 |
+
"model.layers.38.self_attn.o_proj": 71,
|
| 252 |
+
"model.layers.38.self_attn.q_proj": 58,
|
| 253 |
+
"model.layers.38.self_attn.v_proj": 71,
|
| 254 |
+
"model.layers.39.mlp.down_proj": 65,
|
| 255 |
+
"model.layers.39.mlp.gate_proj": 66,
|
| 256 |
+
"model.layers.39.mlp.up_proj": 64,
|
| 257 |
+
"model.layers.39.self_attn.k_proj": 57,
|
| 258 |
+
"model.layers.39.self_attn.o_proj": 70,
|
| 259 |
+
"model.layers.39.self_attn.q_proj": 58,
|
| 260 |
+
"model.layers.39.self_attn.v_proj": 70,
|
| 261 |
+
"model.layers.4.mlp.down_proj": 62,
|
| 262 |
+
"model.layers.4.mlp.gate_proj": 63,
|
| 263 |
+
"model.layers.4.mlp.up_proj": 62,
|
| 264 |
+
"model.layers.4.self_attn.k_proj": 77,
|
| 265 |
+
"model.layers.4.self_attn.o_proj": 53,
|
| 266 |
+
"model.layers.4.self_attn.q_proj": 76,
|
| 267 |
+
"model.layers.4.self_attn.v_proj": 51,
|
| 268 |
+
"model.layers.40.mlp.down_proj": 65,
|
| 269 |
+
"model.layers.40.mlp.gate_proj": 67,
|
| 270 |
+
"model.layers.40.mlp.up_proj": 65,
|
| 271 |
+
"model.layers.40.self_attn.k_proj": 57,
|
| 272 |
+
"model.layers.40.self_attn.o_proj": 72,
|
| 273 |
+
"model.layers.40.self_attn.q_proj": 57,
|
| 274 |
+
"model.layers.40.self_attn.v_proj": 71,
|
| 275 |
+
"model.layers.41.mlp.down_proj": 65,
|
| 276 |
+
"model.layers.41.mlp.gate_proj": 66,
|
| 277 |
+
"model.layers.41.mlp.up_proj": 65,
|
| 278 |
+
"model.layers.41.self_attn.k_proj": 54,
|
| 279 |
+
"model.layers.41.self_attn.o_proj": 71,
|
| 280 |
+
"model.layers.41.self_attn.q_proj": 55,
|
| 281 |
+
"model.layers.41.self_attn.v_proj": 71,
|
| 282 |
+
"model.layers.42.mlp.down_proj": 65,
|
| 283 |
+
"model.layers.42.mlp.gate_proj": 66,
|
| 284 |
+
"model.layers.42.mlp.up_proj": 65,
|
| 285 |
+
"model.layers.42.self_attn.k_proj": 52,
|
| 286 |
+
"model.layers.42.self_attn.o_proj": 71,
|
| 287 |
+
"model.layers.42.self_attn.q_proj": 53,
|
| 288 |
+
"model.layers.42.self_attn.v_proj": 71,
|
| 289 |
+
"model.layers.43.mlp.down_proj": 65,
|
| 290 |
+
"model.layers.43.mlp.gate_proj": 67,
|
| 291 |
+
"model.layers.43.mlp.up_proj": 65,
|
| 292 |
+
"model.layers.43.self_attn.k_proj": 58,
|
| 293 |
+
"model.layers.43.self_attn.o_proj": 71,
|
| 294 |
+
"model.layers.43.self_attn.q_proj": 58,
|
| 295 |
+
"model.layers.43.self_attn.v_proj": 71,
|
| 296 |
+
"model.layers.44.mlp.down_proj": 65,
|
| 297 |
+
"model.layers.44.mlp.gate_proj": 67,
|
| 298 |
+
"model.layers.44.mlp.up_proj": 65,
|
| 299 |
+
"model.layers.44.self_attn.k_proj": 55,
|
| 300 |
+
"model.layers.44.self_attn.o_proj": 71,
|
| 301 |
+
"model.layers.44.self_attn.q_proj": 56,
|
| 302 |
+
"model.layers.44.self_attn.v_proj": 71,
|
| 303 |
+
"model.layers.45.mlp.down_proj": 65,
|
| 304 |
+
"model.layers.45.mlp.gate_proj": 67,
|
| 305 |
+
"model.layers.45.mlp.up_proj": 65,
|
| 306 |
+
"model.layers.45.self_attn.k_proj": 55,
|
| 307 |
+
"model.layers.45.self_attn.o_proj": 71,
|
| 308 |
+
"model.layers.45.self_attn.q_proj": 56,
|
| 309 |
+
"model.layers.45.self_attn.v_proj": 71,
|
| 310 |
+
"model.layers.46.mlp.down_proj": 65,
|
| 311 |
+
"model.layers.46.mlp.gate_proj": 67,
|
| 312 |
+
"model.layers.46.mlp.up_proj": 65,
|
| 313 |
+
"model.layers.46.self_attn.k_proj": 50,
|
| 314 |
+
"model.layers.46.self_attn.o_proj": 68,
|
| 315 |
+
"model.layers.46.self_attn.q_proj": 52,
|
| 316 |
+
"model.layers.46.self_attn.v_proj": 69,
|
| 317 |
+
"model.layers.47.mlp.down_proj": 65,
|
| 318 |
+
"model.layers.47.mlp.gate_proj": 67,
|
| 319 |
+
"model.layers.47.mlp.up_proj": 65,
|
| 320 |
+
"model.layers.47.self_attn.k_proj": 50,
|
| 321 |
+
"model.layers.47.self_attn.o_proj": 70,
|
| 322 |
+
"model.layers.47.self_attn.q_proj": 52,
|
| 323 |
+
"model.layers.47.self_attn.v_proj": 71,
|
| 324 |
+
"model.layers.48.mlp.down_proj": 65,
|
| 325 |
+
"model.layers.48.mlp.gate_proj": 67,
|
| 326 |
+
"model.layers.48.mlp.up_proj": 65,
|
| 327 |
+
"model.layers.48.self_attn.k_proj": 53,
|
| 328 |
+
"model.layers.48.self_attn.o_proj": 71,
|
| 329 |
+
"model.layers.48.self_attn.q_proj": 55,
|
| 330 |
+
"model.layers.48.self_attn.v_proj": 71,
|
| 331 |
+
"model.layers.49.mlp.down_proj": 65,
|
| 332 |
+
"model.layers.49.mlp.gate_proj": 67,
|
| 333 |
+
"model.layers.49.mlp.up_proj": 66,
|
| 334 |
+
"model.layers.49.self_attn.k_proj": 57,
|
| 335 |
+
"model.layers.49.self_attn.o_proj": 73,
|
| 336 |
+
"model.layers.49.self_attn.q_proj": 58,
|
| 337 |
+
"model.layers.49.self_attn.v_proj": 74,
|
| 338 |
+
"model.layers.5.mlp.down_proj": 62,
|
| 339 |
+
"model.layers.5.mlp.gate_proj": 64,
|
| 340 |
+
"model.layers.5.mlp.up_proj": 63,
|
| 341 |
+
"model.layers.5.self_attn.k_proj": 76,
|
| 342 |
+
"model.layers.5.self_attn.o_proj": 53,
|
| 343 |
+
"model.layers.5.self_attn.q_proj": 75,
|
| 344 |
+
"model.layers.5.self_attn.v_proj": 52,
|
| 345 |
+
"model.layers.50.mlp.down_proj": 65,
|
| 346 |
+
"model.layers.50.mlp.gate_proj": 67,
|
| 347 |
+
"model.layers.50.mlp.up_proj": 66,
|
| 348 |
+
"model.layers.50.self_attn.k_proj": 56,
|
| 349 |
+
"model.layers.50.self_attn.o_proj": 72,
|
| 350 |
+
"model.layers.50.self_attn.q_proj": 57,
|
| 351 |
+
"model.layers.50.self_attn.v_proj": 72,
|
| 352 |
+
"model.layers.51.mlp.down_proj": 65,
|
| 353 |
+
"model.layers.51.mlp.gate_proj": 67,
|
| 354 |
+
"model.layers.51.mlp.up_proj": 66,
|
| 355 |
+
"model.layers.51.self_attn.k_proj": 57,
|
| 356 |
+
"model.layers.51.self_attn.o_proj": 70,
|
| 357 |
+
"model.layers.51.self_attn.q_proj": 58,
|
| 358 |
+
"model.layers.51.self_attn.v_proj": 71,
|
| 359 |
+
"model.layers.52.mlp.down_proj": 65,
|
| 360 |
+
"model.layers.52.mlp.gate_proj": 66,
|
| 361 |
+
"model.layers.52.mlp.up_proj": 66,
|
| 362 |
+
"model.layers.52.self_attn.k_proj": 54,
|
| 363 |
+
"model.layers.52.self_attn.o_proj": 70,
|
| 364 |
+
"model.layers.52.self_attn.q_proj": 55,
|
| 365 |
+
"model.layers.52.self_attn.v_proj": 70,
|
| 366 |
+
"model.layers.53.mlp.down_proj": 65,
|
| 367 |
+
"model.layers.53.mlp.gate_proj": 66,
|
| 368 |
+
"model.layers.53.mlp.up_proj": 66,
|
| 369 |
+
"model.layers.53.self_attn.k_proj": 54,
|
| 370 |
+
"model.layers.53.self_attn.o_proj": 68,
|
| 371 |
+
"model.layers.53.self_attn.q_proj": 56,
|
| 372 |
+
"model.layers.53.self_attn.v_proj": 69,
|
| 373 |
+
"model.layers.54.mlp.down_proj": 66,
|
| 374 |
+
"model.layers.54.mlp.gate_proj": 66,
|
| 375 |
+
"model.layers.54.mlp.up_proj": 67,
|
| 376 |
+
"model.layers.54.self_attn.k_proj": 55,
|
| 377 |
+
"model.layers.54.self_attn.o_proj": 70,
|
| 378 |
+
"model.layers.54.self_attn.q_proj": 56,
|
| 379 |
+
"model.layers.54.self_attn.v_proj": 71,
|
| 380 |
+
"model.layers.55.mlp.down_proj": 66,
|
| 381 |
+
"model.layers.55.mlp.gate_proj": 66,
|
| 382 |
+
"model.layers.55.mlp.up_proj": 67,
|
| 383 |
+
"model.layers.55.self_attn.k_proj": 56,
|
| 384 |
+
"model.layers.55.self_attn.o_proj": 70,
|
| 385 |
+
"model.layers.55.self_attn.q_proj": 57,
|
| 386 |
+
"model.layers.55.self_attn.v_proj": 70,
|
| 387 |
+
"model.layers.56.mlp.down_proj": 65,
|
| 388 |
+
"model.layers.56.mlp.gate_proj": 66,
|
| 389 |
+
"model.layers.56.mlp.up_proj": 67,
|
| 390 |
+
"model.layers.56.self_attn.k_proj": 53,
|
| 391 |
+
"model.layers.56.self_attn.o_proj": 73,
|
| 392 |
+
"model.layers.56.self_attn.q_proj": 54,
|
| 393 |
+
"model.layers.56.self_attn.v_proj": 74,
|
| 394 |
+
"model.layers.57.mlp.down_proj": 66,
|
| 395 |
+
"model.layers.57.mlp.gate_proj": 66,
|
| 396 |
+
"model.layers.57.mlp.up_proj": 67,
|
| 397 |
+
"model.layers.57.self_attn.k_proj": 54,
|
| 398 |
+
"model.layers.57.self_attn.o_proj": 68,
|
| 399 |
+
"model.layers.57.self_attn.q_proj": 55,
|
| 400 |
+
"model.layers.57.self_attn.v_proj": 69,
|
| 401 |
+
"model.layers.58.mlp.down_proj": 65,
|
| 402 |
+
"model.layers.58.mlp.gate_proj": 67,
|
| 403 |
+
"model.layers.58.mlp.up_proj": 67,
|
| 404 |
+
"model.layers.58.self_attn.k_proj": 49,
|
| 405 |
+
"model.layers.58.self_attn.o_proj": 63,
|
| 406 |
+
"model.layers.58.self_attn.q_proj": 50,
|
| 407 |
+
"model.layers.58.self_attn.v_proj": 65,
|
| 408 |
+
"model.layers.59.mlp.down_proj": 65,
|
| 409 |
+
"model.layers.59.mlp.gate_proj": 68,
|
| 410 |
+
"model.layers.59.mlp.up_proj": 68,
|
| 411 |
+
"model.layers.59.self_attn.k_proj": 53,
|
| 412 |
+
"model.layers.59.self_attn.o_proj": 57,
|
| 413 |
+
"model.layers.59.self_attn.q_proj": 53,
|
| 414 |
+
"model.layers.59.self_attn.v_proj": 60,
|
| 415 |
+
"model.layers.6.mlp.down_proj": 62,
|
| 416 |
+
"model.layers.6.mlp.gate_proj": 64,
|
| 417 |
+
"model.layers.6.mlp.up_proj": 63,
|
| 418 |
+
"model.layers.6.self_attn.k_proj": 76,
|
| 419 |
+
"model.layers.6.self_attn.o_proj": 54,
|
| 420 |
+
"model.layers.6.self_attn.q_proj": 75,
|
| 421 |
+
"model.layers.6.self_attn.v_proj": 53,
|
| 422 |
+
"model.layers.7.mlp.down_proj": 62,
|
| 423 |
+
"model.layers.7.mlp.gate_proj": 64,
|
| 424 |
+
"model.layers.7.mlp.up_proj": 64,
|
| 425 |
+
"model.layers.7.self_attn.k_proj": 78,
|
| 426 |
+
"model.layers.7.self_attn.o_proj": 56,
|
| 427 |
+
"model.layers.7.self_attn.q_proj": 77,
|
| 428 |
+
"model.layers.7.self_attn.v_proj": 55,
|
| 429 |
+
"model.layers.8.mlp.down_proj": 62,
|
| 430 |
+
"model.layers.8.mlp.gate_proj": 64,
|
| 431 |
+
"model.layers.8.mlp.up_proj": 64,
|
| 432 |
+
"model.layers.8.self_attn.k_proj": 80,
|
| 433 |
+
"model.layers.8.self_attn.o_proj": 58,
|
| 434 |
+
"model.layers.8.self_attn.q_proj": 78,
|
| 435 |
+
"model.layers.8.self_attn.v_proj": 58,
|
| 436 |
+
"model.layers.9.mlp.down_proj": 62,
|
| 437 |
+
"model.layers.9.mlp.gate_proj": 64,
|
| 438 |
+
"model.layers.9.mlp.up_proj": 64,
|
| 439 |
+
"model.layers.9.self_attn.k_proj": 80,
|
| 440 |
+
"model.layers.9.self_attn.o_proj": 58,
|
| 441 |
+
"model.layers.9.self_attn.q_proj": 78,
|
| 442 |
+
"model.layers.9.self_attn.v_proj": 58
|
| 443 |
+
},
|
| 444 |
+
"target_modules": [
|
| 445 |
+
"down_proj",
|
| 446 |
+
"gate_proj",
|
| 447 |
+
"k_proj",
|
| 448 |
+
"o_proj",
|
| 449 |
+
"v_proj",
|
| 450 |
+
"q_proj",
|
| 451 |
+
"up_proj"
|
| 452 |
+
],
|
| 453 |
+
"task_type": "CAUSAL_LM",
|
| 454 |
+
"token_dim": 6656
|
| 455 |
+
}
|
adapter_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9c4df3d633f53083955784dfd9a84108eb5bdb14769a991e6428b2962184008f
|
| 3 |
+
size 975880621
|
usage.py
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import sys
|
| 2 |
+
sys.path.insert(1, '/workspace/asr/peft/src')
|
| 3 |
+
# TODO set this path to the lazy-lora source code path, or you can install it from source code:
|
| 4 |
+
# TODO, please install lazylora for usage:
|
| 5 |
+
# git clone [email protected]:Xianchao-Wu/peft.git
|
| 6 |
+
# cd peft
|
| 7 |
+
# python setup.py install
|
| 8 |
+
|
| 9 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
|
| 10 |
+
from peft import PeftModel, PeftConfig
|
| 11 |
+
import os
|
| 12 |
+
import torch
|
| 13 |
+
|
| 14 |
+
#import ipdb; ipdb.set_trace()
|
| 15 |
+
cache_dir="/workspace/asr/peft/qlora"
|
| 16 |
+
# TODO set this cache_dir to the path where you stored (or, want to store) llama1-33b (huggyllama/llama-30b) model
|
| 17 |
+
|
| 18 |
+
lazylora_dir=os.getcwd() # the path that contains 'adapter_config.json' and 'adapter_model.bin'
|
| 19 |
+
|
| 20 |
+
config = PeftConfig.from_pretrained(lazylora_dir)
|
| 21 |
+
|
| 22 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 23 |
+
config.base_model_name_or_path,
|
| 24 |
+
cache_dir=cache_dir,
|
| 25 |
+
use_auth_token=True
|
| 26 |
+
)
|
| 27 |
+
|
| 28 |
+
bnb_config = BitsAndBytesConfig(
|
| 29 |
+
load_in_4bit=True,
|
| 30 |
+
bnb_4bit_use_double_quant=True,
|
| 31 |
+
bnb_4bit_quant_type='nf4',
|
| 32 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
| 33 |
+
)
|
| 34 |
+
|
| 35 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 36 |
+
config.base_model_name_or_path,
|
| 37 |
+
quantization_config=bnb_config,
|
| 38 |
+
device_map="auto",
|
| 39 |
+
cache_dir=cache_dir,
|
| 40 |
+
use_auth_token=True
|
| 41 |
+
)
|
| 42 |
+
#model.print_trainable_parameters()
|
| 43 |
+
print(sum(p.numel() for p in model.parameters()))
|
| 44 |
+
# 16,477,866,496 -> half-size of 33B due to 4-bit loading
|
| 45 |
+
|
| 46 |
+
model = PeftModel.from_pretrained(model, lazylora_dir)
|
| 47 |
+
print('after adding lazy lora parameters:')
|
| 48 |
+
model.print_trainable_parameters()
|
| 49 |
+
# trainable params: 0 || all params: 16,965,645,824 || trainable%: 0.0
|
| 50 |
+
|
| 51 |
+
|