File size: 5,161 Bytes
8e97a2b fd5b301 760ebb4 fd5b301 760ebb4 fd5b301 760ebb4 fd5b301 8e97a2b |
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 |
---
license: other
language:
- zh
- en
tags:
- chatglm
- glm-4v
- quantization
- auto-gptq
- 4bit
---
## Usage
This model is quantized using [AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ) for [THUDM/glm-4v-9b](https://huggingface.co/THUDM/glm-4v-9b).
Use pip install AutoGPTQ (required)
(The quantization script will be released later)
```bash
pip install auto-gptq
```
Since the original auto-gptq library does not support the quantization of chatglm models, manual import (hack) is required.
```python
from auto_gptq.modeling._base import BaseGPTQForCausalLM
from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP
class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
layer_type = ["GLMBlock", "TransformerLayer", "GLU"]
layers_block_names = ["transformer.encoder.layers",
"transformer.vision.transformer.layers",
"transformer.vision.linear_proj"]
outside_layer_modules = ["transformer.output_layer"]
inside_layer_modules = [
["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
]
GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
```
The complete model import code is as follows:
### Load model
```python
import os
import json
import random
import time
import torch
import datasets
from transformers import AutoTokenizer, AutoModelForCausalLM
from auto_gptq import AutoGPTQForCausalLM
from auto_gptq.modeling._base import BaseGPTQForCausalLM
from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP
class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
layer_type = ["GLMBlock", "TransformerLayer", "GLU"]
layers_block_names = ["transformer.encoder.layers",
"transformer.vision.transformer.layers",
"transformer.vision.linear_proj"]
outside_layer_modules = ["transformer.output_layer"]
inside_layer_modules = [
["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
]
GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
device = 'cuda:0'
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq'
trust_remote_code = True
tokenizer = AutoTokenizer.from_pretrained(
quantized_model_dir,
trust_remote_code=trust_remote_code,
)
model = AutoGPTQForCausalLM.from_quantized(
quantized_model_dir,
device=device,
trust_remote_code=trust_remote_code,
torch_dtype=torch.float16,
use_cache=True,
inject_fused_mlp=True,
inject_fused_attention=True,
)
```
You can also load the model using HuggingFace Transformers, but it will slow down inference.
```python
import os
import json
import random
import time
import torch
import datasets
from transformers import AutoTokenizer, AutoModelForCausalLM
device = 'cuda:0'
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq-4bit'
trust_remote_code = True
tokenizer = AutoTokenizer.from_pretrained(
quantized_model_dir,
trust_remote_code=trust_remote_code,
)
model = AutoModelForCausalLM.from_pretrained(
quantized_model_dir,
torch_dtype=torch.float16,
device_map="auto",
low_cpu_mem_usage=True,
trust_remote_code=trust_remote_code,
use_cache=True
).eval()
```
### inference test
Load the CogVLM-SFT-311K-subset-gptq dataset as test data, which is a dataset for quantization.
```python
dataset = datasets.load_dataset('alexwww94/CogVLM-SFT-311K-subset-gptq')
for example in dataset['single']:
# prompt = "为什么马会被围栏限制在一个区域内？"
prompt = json.loads(example['labels_zh'])['conversations'][0]
answer = json.loads(example['labels_zh'])['conversations'][1]
image = example['image']
print(f"prompt: {prompt}")
print("-" * 42)
print(f"golden: {answer}")
print("-" * 42)
start = time.time()
prompt.update({'image': image})
inputs = tokenizer.apply_chat_template([prompt],
add_generation_prompt=True, tokenize=True, return_tensors="pt",
return_dict=True, dtyp=torch.bfloat16) # chat mode
inputs = inputs.to(device)
inputs['images'] = inputs['images'].half()
gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
with torch.inference_mode():
outputs = model.generate(**inputs, **gen_kwargs)
outputs = outputs[:, inputs['input_ids'].shape[1]:]
generated_text = tokenizer.decode(outputs[0]).split('<|endoftext|>')[0]
end = time.time()
print(f"quant: {generated_text}")
num_new_tokens = len(tokenizer(generated_text)["input_ids"])
print(f"generate {num_new_tokens} tokens using {end-start: .4f}s, {num_new_tokens / (end - start)} tokens/s.")
print("=" * 42)
# break
```
### metrics
(to be released later) |