Update README.md
Browse files
README.md
CHANGED
@@ -12,19 +12,85 @@ This model is fine-tuned with LLaMA with 8 Nvidia RTX 1080Ti GPUs and enhanced w
|
|
12 |
### Here is how to use it with texts in HuggingFace
|
13 |
```python
|
14 |
# A list of special tokens the model was trained with
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
```
|
|
|
12 |
### Here is how to use it with texts in HuggingFace
|
13 |
```python
|
14 |
# A list of special tokens the model was trained with
|
15 |
+
|
16 |
+
from transformers import LlamaTokenizer, AutoModelForCausalLM
|
17 |
+
tokenizer = LlamaTokenizer.from_pretrained("Fan21/Llama-mt-lora")
|
18 |
+
BASE_MODEL = "Fan21/Llama-mt-lora"
|
19 |
+
if torch.cuda.is_available():
|
20 |
+
device = "cuda"
|
21 |
+
else:
|
22 |
+
device = "cpu"
|
23 |
+
if device == "cuda":
|
24 |
+
model = LlamaForCausalLM.from_pretrained(
|
25 |
+
BASE_MODEL,
|
26 |
+
load_in_8bit=False,
|
27 |
+
torch_dtype=torch.float16,
|
28 |
+
device_map="auto",
|
29 |
+
)
|
30 |
+
|
31 |
+
elif device == "mps":
|
32 |
+
model = LlamaForCausalLM.from_pretrained(
|
33 |
+
BASE_MODEL,
|
34 |
+
device_map={"": device},
|
35 |
+
torch_dtype=torch.float16,
|
36 |
+
)
|
37 |
+
|
38 |
+
else:
|
39 |
+
model = LlamaForCausalLM.from_pretrained(
|
40 |
+
BASE_MODEL, device_map={"": device}, low_cpu_mem_usage=True
|
41 |
+
)
|
42 |
+
|
43 |
+
def generate_prompt(instruction, input=None):
|
44 |
+
if input:
|
45 |
+
return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
|
46 |
+
### Instruction:
|
47 |
+
{instruction}
|
48 |
+
### Input:
|
49 |
+
{input}
|
50 |
+
### Response:"""
|
51 |
+
else:
|
52 |
+
return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
|
53 |
+
### Instruction:
|
54 |
+
{instruction}
|
55 |
+
### Response:"""
|
56 |
+
|
57 |
+
if device != "cpu":
|
58 |
+
model.half()
|
59 |
+
model.eval()
|
60 |
+
if torch.__version__ >= "2":
|
61 |
+
model = torch.compile(model)
|
62 |
+
|
63 |
+
|
64 |
+
def evaluate(
|
65 |
+
instruction,
|
66 |
+
input=None,
|
67 |
+
temperature=0.1,
|
68 |
+
top_p=0.75,
|
69 |
+
top_k=40,
|
70 |
+
num_beams=4,
|
71 |
+
max_new_tokens=128,
|
72 |
+
**kwargs,
|
73 |
+
):
|
74 |
+
prompt = generate_prompt(instruction, input)
|
75 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
76 |
+
input_ids = inputs["input_ids"].to(device)
|
77 |
+
generation_config = GenerationConfig(
|
78 |
+
temperature=temperature,
|
79 |
+
top_p=top_p,
|
80 |
+
top_k=top_k,
|
81 |
+
num_beams=num_beams,
|
82 |
+
**kwargs,
|
83 |
+
)
|
84 |
+
with torch.no_grad():
|
85 |
+
generation_output = model.generate(
|
86 |
+
input_ids=input_ids,
|
87 |
+
generation_config=generation_config,
|
88 |
+
return_dict_in_generate=True,
|
89 |
+
output_scores=True,
|
90 |
+
max_new_tokens=max_new_tokens,
|
91 |
+
)
|
92 |
+
s = generation_output.sequences[0]
|
93 |
+
output = tokenizer.decode(s)
|
94 |
+
return output.split("### Response:")[1].strip()
|
95 |
+
|
96 |
```
|