Update README.md
Browse files
README.md
CHANGED
@@ -1,7 +1,38 @@
|
|
1 |
## Usage
|
2 |
This model is quantized using [AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ) for [THUDM/glm-4v-9b](https://huggingface.co/THUDM/glm-4v-9b).
|
3 |
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
### Load model
|
7 |
```python
|
@@ -16,6 +47,27 @@ import datasets
|
|
16 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
17 |
from auto_gptq import AutoGPTQForCausalLM
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
device = 'cuda:0'
|
20 |
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq'
|
21 |
trust_remote_code = True
|
@@ -39,6 +91,25 @@ model = AutoGPTQForCausalLM.from_quantized(
|
|
39 |
You can also load the model using HuggingFace Transformers, but it will slow down inference.
|
40 |
|
41 |
```python
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
model = AutoModelForCausalLM.from_pretrained(
|
43 |
quantized_model_dir,
|
44 |
torch_dtype=torch.float16,
|
|
|
1 |
## Usage
|
2 |
This model is quantized using [AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ) for [THUDM/glm-4v-9b](https://huggingface.co/THUDM/glm-4v-9b).
|
3 |
|
4 |
+
Use pip install AutoGPTQ (required)
|
5 |
+
|
6 |
+
(The quantization script will be released later)
|
7 |
+
|
8 |
+
```bash
|
9 |
+
pip install auto-gptq
|
10 |
+
```
|
11 |
+
|
12 |
+
Since the original auto-gptq library does not support the quantization of chatglm models, manual import (hack) is required.
|
13 |
+
```python
|
14 |
+
from auto_gptq.modeling._base import BaseGPTQForCausalLM
|
15 |
+
from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP
|
16 |
+
|
17 |
+
class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
|
18 |
+
layer_type = ["GLMBlock", "TransformerLayer", "GLU"]
|
19 |
+
|
20 |
+
layers_block_names = ["transformer.encoder.layers",
|
21 |
+
"transformer.vision.transformer.layers",
|
22 |
+
"transformer.vision.linear_proj"]
|
23 |
+
|
24 |
+
outside_layer_modules = ["transformer.output_layer"]
|
25 |
+
|
26 |
+
inside_layer_modules = [
|
27 |
+
["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
|
28 |
+
["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
|
29 |
+
["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
|
30 |
+
]
|
31 |
+
|
32 |
+
GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
|
33 |
+
```
|
34 |
+
|
35 |
+
The complete model import code is as follows:
|
36 |
|
37 |
### Load model
|
38 |
```python
|
|
|
47 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
48 |
from auto_gptq import AutoGPTQForCausalLM
|
49 |
|
50 |
+
|
51 |
+
from auto_gptq.modeling._base import BaseGPTQForCausalLM
|
52 |
+
from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP
|
53 |
+
|
54 |
+
class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
|
55 |
+
layer_type = ["GLMBlock", "TransformerLayer", "GLU"]
|
56 |
+
|
57 |
+
layers_block_names = ["transformer.encoder.layers",
|
58 |
+
"transformer.vision.transformer.layers",
|
59 |
+
"transformer.vision.linear_proj"]
|
60 |
+
|
61 |
+
outside_layer_modules = ["transformer.output_layer"]
|
62 |
+
|
63 |
+
inside_layer_modules = [
|
64 |
+
["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
|
65 |
+
["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
|
66 |
+
["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
|
67 |
+
]
|
68 |
+
|
69 |
+
GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
|
70 |
+
|
71 |
device = 'cuda:0'
|
72 |
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq'
|
73 |
trust_remote_code = True
|
|
|
91 |
You can also load the model using HuggingFace Transformers, but it will slow down inference.
|
92 |
|
93 |
```python
|
94 |
+
import os
|
95 |
+
|
96 |
+
import json
|
97 |
+
import random
|
98 |
+
import time
|
99 |
+
|
100 |
+
import torch
|
101 |
+
import datasets
|
102 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
103 |
+
|
104 |
+
device = 'cuda:0'
|
105 |
+
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq-4bit'
|
106 |
+
trust_remote_code = True
|
107 |
+
|
108 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
109 |
+
quantized_model_dir,
|
110 |
+
trust_remote_code=trust_remote_code,
|
111 |
+
)
|
112 |
+
|
113 |
model = AutoModelForCausalLM.from_pretrained(
|
114 |
quantized_model_dir,
|
115 |
torch_dtype=torch.float16,
|