Update README.md
Browse files
README.md
CHANGED
@@ -81,4 +81,43 @@ with torch.inference_mode():
|
|
81 |
outputs = model.generate(**inputs, max_new_tokens=512)
|
82 |
result_text = tokenizer.batch_decode(outputs)[0]
|
83 |
print(result_text)
|
84 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
outputs = model.generate(**inputs, max_new_tokens=512)
|
82 |
result_text = tokenizer.batch_decode(outputs)[0]
|
83 |
print(result_text)
|
84 |
+
```
|
85 |
+
|
86 |
+
# Load model in 4bits
|
87 |
+
```python
|
88 |
+
|
89 |
+
import torch
|
90 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, BitsAndBytesConfig
|
91 |
+
|
92 |
+
# quantization_config = BitsAndBytesConfig(llm_int8_enable_fp32_cpu_offload=True)
|
93 |
+
bnb_config = BitsAndBytesConfig(
|
94 |
+
llm_int8_enable_fp32_cpu_offload=True,
|
95 |
+
load_in_4bit=True,
|
96 |
+
bnb_4bit_use_double_quant=True,
|
97 |
+
bnb_4bit_quant_type="nf4",
|
98 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
99 |
+
)
|
100 |
+
# control model memory allocation between devices for low GPU resource (0,cpu)
|
101 |
+
device_map = {
|
102 |
+
"transformer.word_embeddings": 0,
|
103 |
+
"transformer.word_embeddings_layernorm": 0,
|
104 |
+
"lm_head": 0,
|
105 |
+
"transformer.h": 0,
|
106 |
+
"transformer.ln_f": 0,
|
107 |
+
"model.embed_tokens": 0,
|
108 |
+
"model.layers":0,
|
109 |
+
"model.norm":0
|
110 |
+
}
|
111 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
112 |
+
|
113 |
+
# model use for inference
|
114 |
+
model_id="mychen76/mistral7b_ocr_to_json_v1"
|
115 |
+
model = AutoModelForCausalLM.from_pretrained(
|
116 |
+
model_id,
|
117 |
+
trust_remote_code=True,
|
118 |
+
torch_dtype=torch.float16,
|
119 |
+
quantization_config=bnb_config,
|
120 |
+
device_map=device_map)
|
121 |
+
# tokenizer
|
122 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
|
123 |
+
```
|