Spaces:
Runtime error
Runtime error
ashioyajotham
commited on
Commit
•
fde41fa
1
Parent(s):
7139589
Update app.py
Browse files
app.py
CHANGED
@@ -1,177 +1,19 @@
|
|
1 |
-
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
dataset = load_dataset(dataset_name, split="train")
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
# Defining the name of the Falcon model
|
16 |
-
model_name = "ybelkada/falcon-7b-sharded-bf16"
|
17 |
-
|
18 |
-
# Configuring the BitsAndBytes quantization
|
19 |
-
bnb_config = BitsAndBytesConfig(
|
20 |
-
load_in_4bit=True,
|
21 |
-
bnb_4bit_quant_type="nf4",
|
22 |
-
bnb_4bit_compute_dtype=torch.float16,
|
23 |
-
)
|
24 |
-
|
25 |
-
# Loading the Falcon model with quantization configuration
|
26 |
-
model = FalconForCausalLM.from_pretrained(
|
27 |
-
model_name,
|
28 |
-
quantization_config=bnb_config,
|
29 |
-
trust_remote_code=True
|
30 |
-
)
|
31 |
-
|
32 |
-
# Disabling cache usage in the model configuration
|
33 |
-
model.config.use_cache = False
|
34 |
-
|
35 |
-
# Load the tokenizer for the Falcon 7B model with remote code trust
|
36 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
37 |
-
|
38 |
-
# Set the padding token to be the same as the end-of-sequence token
|
39 |
-
tokenizer.pad_token = tokenizer.eos_token
|
40 |
-
|
41 |
-
# Import the necessary module for LoRA configuration
|
42 |
-
from peft import LoraConfig
|
43 |
-
|
44 |
-
# Define the parameters for LoRA configuration
|
45 |
-
lora_alpha = 16
|
46 |
-
lora_dropout = 0.1
|
47 |
-
lora_r = 64
|
48 |
-
|
49 |
-
# Create the LoRA configuration object
|
50 |
-
peft_config = LoraConfig(
|
51 |
-
lora_alpha=lora_alpha,
|
52 |
-
lora_dropout=lora_dropout,
|
53 |
-
r=lora_r,
|
54 |
-
bias="none",
|
55 |
-
task_type="CAUSAL_LM",
|
56 |
-
target_modules=[
|
57 |
-
"query_key_value",
|
58 |
-
"dense",
|
59 |
-
"dense_h_to_4h",
|
60 |
-
"dense_4h_to_h",
|
61 |
-
]
|
62 |
-
)
|
63 |
-
|
64 |
-
from transformers import TrainingArguments
|
65 |
-
# Define the directory to save training results
|
66 |
-
output_dir = "./results"
|
67 |
-
|
68 |
-
# Set the batch size per device during training
|
69 |
-
per_device_train_batch_size = 4
|
70 |
-
|
71 |
-
# Number of steps to accumulate gradients before updating the model
|
72 |
-
gradient_accumulation_steps = 4
|
73 |
-
|
74 |
-
# Choose the optimizer type (e.g., "paged_adamw_32bit")
|
75 |
-
optim = "paged_adamw_32bit"
|
76 |
-
|
77 |
-
# Interval to save model checkpoints (every 10 steps)
|
78 |
-
save_steps = 10
|
79 |
-
|
80 |
-
# Interval to log training metrics (every 10 steps)
|
81 |
-
logging_steps = 10
|
82 |
-
|
83 |
-
# Learning rate for optimization
|
84 |
-
learning_rate = 2e-4
|
85 |
-
|
86 |
-
# Maximum gradient norm for gradient clipping
|
87 |
-
max_grad_norm = 0.3
|
88 |
-
|
89 |
-
# Maximum number of training steps
|
90 |
-
max_steps = 50
|
91 |
-
|
92 |
-
# Warmup ratio for learning rate scheduling
|
93 |
-
warmup_ratio = 0.03
|
94 |
-
|
95 |
-
# Type of learning rate scheduler (e.g., "constant")
|
96 |
-
lr_scheduler_type = "constant"
|
97 |
-
|
98 |
-
# Create a TrainingArguments object to configure the training process
|
99 |
-
training_arguments = TrainingArguments(
|
100 |
-
output_dir=output_dir,
|
101 |
-
per_device_train_batch_size=per_device_train_batch_size,
|
102 |
-
gradient_accumulation_steps=gradient_accumulation_steps,
|
103 |
-
optim=optim,
|
104 |
-
save_steps=save_steps,
|
105 |
-
logging_steps=logging_steps,
|
106 |
-
learning_rate=learning_rate,
|
107 |
-
fp16=True, # Use mixed precision training (16-bit)
|
108 |
-
max_grad_norm=max_grad_norm,
|
109 |
-
max_steps=max_steps,
|
110 |
-
warmup_ratio=warmup_ratio,
|
111 |
-
group_by_length=True,
|
112 |
-
lr_scheduler_type=lr_scheduler_type,
|
113 |
-
)
|
114 |
-
|
115 |
-
|
116 |
-
dataset = dataset.map(lambda x: {"text": x["input"]+x["output"]})
|
117 |
-
|
118 |
-
# Import the SFTTrainer from the TRL library
|
119 |
-
from trl import SFTTrainer
|
120 |
-
|
121 |
-
# Set the maximum sequence length
|
122 |
-
max_seq_length = 512
|
123 |
-
|
124 |
-
# Create a trainer instance using SFTTrainer
|
125 |
-
trainer = SFTTrainer(
|
126 |
-
model=model,
|
127 |
-
train_dataset=dataset,
|
128 |
-
peft_config=peft_config,
|
129 |
-
dataset_text_field="text",
|
130 |
-
max_seq_length=max_seq_length,
|
131 |
-
tokenizer=tokenizer,
|
132 |
-
args=training_arguments,
|
133 |
)
|
134 |
|
|
|
135 |
|
136 |
-
# Iterate through the named modules of the trainer's model
|
137 |
-
for name, module in trainer.model.named_modules():
|
138 |
-
|
139 |
-
# Check if the name contains "norm"
|
140 |
-
if "norm" in name:
|
141 |
-
# Convert the module to use torch.float32 data type
|
142 |
-
module = module.to(torch.float32)
|
143 |
-
|
144 |
-
trainer.train()
|
145 |
-
|
146 |
-
|
147 |
-
prompt = "Generate a python script to add prime numbers between one and ten"
|
148 |
-
|
149 |
-
inputs = tokenizer.encode(prompt, return_tensors='pt')
|
150 |
-
|
151 |
-
outputs = model.generate(inputs, max_length=100, temperature = .7, do_sample=True)
|
152 |
-
|
153 |
-
completion = tokenizer.decode(outputs[0])
|
154 |
-
|
155 |
-
print(completion)
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
161 |
-
|
162 |
-
checkpoint_name= model
|
163 |
-
model = AutoModelForCausalLM.from_pretrained(checkpoint_name)
|
164 |
-
tokenizer = AutoTokenizer.from_pretrained(checkpoint_name)
|
165 |
-
|
166 |
-
prompt = "Create a gradio application that help to convert temperature in celcius into temperature in Fahrenheit"
|
167 |
-
inputs = tokenizer(f"Question: {prompt}\n\nAnswer: ", return_tensors="pt")
|
168 |
-
|
169 |
-
outputs = model.generate(
|
170 |
-
inputs["input_ids"],
|
171 |
-
temperature=0.2,
|
172 |
-
top_p=0.95,
|
173 |
-
max_new_tokens=200
|
174 |
-
)
|
175 |
-
|
176 |
-
input_len=len(inputs["input_ids"])
|
177 |
-
print(tokenizer.decode(outputs[0][input_len:]))
|
|
|
1 |
+
fine_tuned_model = "ashioyajotham/results"
|
2 |
|
3 |
+
def generate_code(input_text):
|
4 |
+
# Use your fine-tuned model to generate code here
|
5 |
+
generated_code = fine_tuned_model.generate(input_text)
|
6 |
+
return generated_code
|
7 |
|
8 |
+
import gradio as gr
|
|
|
9 |
|
10 |
+
iface = gr.Interface(
|
11 |
+
fn=generate_code,
|
12 |
+
inputs="text",
|
13 |
+
outputs="text",
|
14 |
+
title="Code Generation App",
|
15 |
+
description="Generate code from text input."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
)
|
17 |
|
18 |
+
iface.launch()
|
19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|