File size: 7,066 Bytes
5963b4e 0c79aa9 5963b4e 0c79aa9 5963b4e 0c79aa9 a630053 0c79aa9 bc99d8d 0c79aa9 bc99d8d 0c79aa9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 |
---
base_model: NousResearch/Meta-Llama-3-70B-Instruct
model_type: llama
pipeline_tag: text-generation
quantized_by: Compressa
license: other
license_name: llama3
license_link: https://llama.meta.com/llama3/license
tags:
- llama3
- omniquant
- gptq
- triton
---
# Llama 3 70B Instruct β OmniQuant
Based on [Llama 3 70B Instruct](https://huggingface.co/NousResearch/Meta-Llama-3-70B-Instruct).
Quantized with [OmniQuant](https://github.com/OpenGVLab/OmniQuant).
## Evaluation
### PPL (β)
| | wiki |
| ------------- | ---- |
| FP | 5,33 |
| **Quantized** | 5,90 |
### Accuracy on English Benchmarks, % (β)
| | piqa | arc_easy | arc_challenge | boolq | hellaswag | winogrande | mmlu_humanities | mmlu_social_sciences | mmlu_stem | mmlu_other |
| ------------- | ---- | -------- | ------------- | ----- | --------- | ---------- | --------------- | -------------------- | --------- | ---------- |
| FP | 81,5 | 86,2 | 61,9 | 87,4 | 63,7 | 75,8 | 78,7 | 84,4 | 71,1 | 80,2 |
| **Quantized** | 80,7 | 85,8 | 61,4 | 87,0 | 62,7 | 73,0 | 75,5 | 81,0 | 68,6 | 77,9 |
### Accuracy on Russian Benchmarks, % (β)
| | danetqa | terra | rwsd | muserc | rucos | lidirus | parus | rcb | russe | rucola |
| ------------- | ------- | ----- | ---- | ------ | ----- | ------- | ----- | ---- | ----- | ------ |
| FP | 88,9 | 88,6 | 75,5 | 81,8 | 82,4 | 70,7 | 77,0 | 35,0 | 63,1 | 34,7 |
| **Quantized** | 86,6 | 81,8 | 71,6 | 75,6 | 69,5 | 60,3 | 64,0 | 26,8 | 63,1 | 32,5 |
### Summary
| | Avg acc diff on Eng, % (β) | Avg acc diff on Rus, % (β) | Occupied disk space, % (β) |
| ------------- | -------------------------- | -------------------------- | -------------------------- |
| FP | 0 | 0 | 100 |
| **Quantized** | \-1,7 | \-6,6 | 28,2 |
## Examples
### Imports and Model Loading
<details>
<summary>Expand</summary>
```python
import gc
import auto_gptq.nn_modules.qlinear.qlinear_cuda as qlinear_cuda
import auto_gptq.nn_modules.qlinear.qlinear_triton as qlinear_triton
import torch
from accelerate import (
init_empty_weights,
infer_auto_device_map,
load_checkpoint_in_model,
)
from tqdm import tqdm
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
pipeline,
)
def get_named_linears(model):
return {
name: module for name, module in model.named_modules()
if isinstance(module, torch.nn.Linear)
}
def set_module(model, name, module):
parent = model
levels = name.split('.')
for i in range(len(levels) - 1):
cur_name = levels[i]
if cur_name.isdigit():
parent = parent[int(cur_name)]
else:
parent = getattr(parent, cur_name)
setattr(parent, levels[-1], module)
def load_model(model_path):
# Based on: https://github.com/OpenGVLab/OmniQuant/blob/main/runing_quantized_mixtral_7bx8.ipynb
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
if not hasattr(config, 'quantization_config'):
raise AttributeError(
f'No quantization info found in model config "{model_path}"'
f' (`quantization_config` section is missing).'
)
wbits = config.quantization_config['bits']
group_size = config.quantization_config['group_size']
# We are going to init an ordinary model and then manually replace all Linears with QuantLinears
del config.quantization_config
with init_empty_weights():
model = AutoModelForCausalLM.from_config(config=config, torch_dtype=torch.float16, trust_remote_code=True)
layers = model.model.layers
for i in tqdm(range(len(layers))):
layer = layers[i]
named_linears = get_named_linears(layer)
for name, module in named_linears.items():
params = (
wbits, group_size,
module.in_features, module.out_features,
module.bias is not None
)
if wbits in [2, 4]:
q_linear = qlinear_triton.QuantLinear(*params)
elif wbits == 3:
q_linear = qlinear_cuda.QuantLinear(*params)
else:
raise NotImplementedError("Only 2, 3 and 4 bits are supported.")
q_linear.to(next(layer.parameters()).device)
set_module(layer, name, q_linear)
torch.cuda.empty_cache()
gc.collect()
model.tie_weights()
device_map = infer_auto_device_map(model)
print("Loading pre-computed quantized weights...")
load_checkpoint_in_model(
model, checkpoint=model_path,
device_map=device_map, offload_state_dict=True,
)
print("Model loaded successfully!")
return model
```
</details>
### Inference
```python
model_path = "compressa-ai/Llama-3-70B-Instruct-OmniQuant"
model = load_model(model_path).cuda()
tokenizer = AutoTokenizer.from_pretrained(
model_path, use_fast=False, trust_remote_code=True
)
# Llama 3 "specifics"
# https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/discussions/4
terminators = [
tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
system_message = "You are a friendly chatbot who responds as if you are the Sandy Cheeks squirrel from the SpongeBob SquarePants cartoon."
user_message = "Do squirrels communicate with birds?"
messages = [
{"role": "system", "content": system_message},
{"role": "user", "content": user_message},
]
prompt = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {k: v.cuda() for k, v in inputs.items()}
outputs = model.generate(
**inputs, max_new_tokens=512,
do_sample=True, temperature=0.7, top_p=0.95,
eos_token_id=terminators,
)
response = tokenizer.decode(outputs[0])
continuation = response.removeprefix(prompt).removesuffix(tokenizer.eos_token)
print(f'Prompt:\n{prompt}')
print(f'Continuation:\n{continuation}\n')
```
### Inference Using Pipeline
```python
pipe = pipeline(
"text-generation",
model=model, tokenizer=tokenizer,
eos_token_id=terminators,
max_new_tokens=512, do_sample=True,
temperature=0.7, top_p=0.95,
device=0,
)
prompt = pipe.tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
outputs = pipe(prompt)
response = outputs[0]["generated_text"]
continuation = response.removeprefix(prompt)
print(f'Prompt:\n{prompt}')
print(f'Continuation:\n{continuation}\n')
```
|