Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -29,88 +29,90 @@ def greet(traindata_,output_repo):
|
|
29 |
# new_model = "Gopal2002/zehpyr-gemma-dpo-finetune"
|
30 |
new_model = output_repo
|
31 |
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
model = AutoModelForCausalLM.from_pretrained(
|
39 |
-
model_name,
|
40 |
-
torch_dtype=torch.float16,
|
41 |
-
load_in_4bit=True
|
42 |
-
)
|
43 |
-
model.config.use_cache = False
|
44 |
-
|
45 |
-
# Reference model
|
46 |
-
ref_model = AutoModelForCausalLM.from_pretrained(
|
47 |
-
model_name,
|
48 |
-
torch_dtype=torch.float16,
|
49 |
-
load_in_4bit=True
|
50 |
-
)
|
51 |
-
|
52 |
-
# specify how to quantize the model
|
53 |
-
quantization_config = BitsAndBytesConfig(
|
54 |
-
load_in_4bit=True,
|
55 |
-
bnb_4bit_quant_type="nf4",
|
56 |
-
bnb_4bit_compute_dtype=torch.bfloat16,
|
57 |
-
)
|
58 |
-
device_map = {"": torch.cuda.current_device()} if torch.cuda.is_available() else None
|
59 |
-
|
60 |
-
# Step 1: load the base model (Mistral-7B in our case) in 4-bit
|
61 |
-
model_kwargs = dict(
|
62 |
-
# attn_implementation="flash_attention_2", # set this to True if your GPU supports it (Flash Attention drastically speeds up model computations)
|
63 |
-
torch_dtype="auto",
|
64 |
-
use_cache=False, # set to False as we're going to use gradient checkpointing
|
65 |
-
device_map=device_map,
|
66 |
-
quantization_config=quantization_config,
|
67 |
-
)
|
68 |
-
model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)
|
69 |
-
|
70 |
-
# Training arguments
|
71 |
-
peft_config = LoraConfig(
|
72 |
-
r=16,
|
73 |
-
lora_alpha=16,
|
74 |
-
lora_dropout=0.05,
|
75 |
-
bias="none",
|
76 |
-
task_type="CAUSAL_LM",
|
77 |
-
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
|
78 |
-
)
|
79 |
-
training_args = TrainingArguments(
|
80 |
-
per_device_train_batch_size=4,
|
81 |
-
gradient_accumulation_steps=4,
|
82 |
-
gradient_checkpointing=True,
|
83 |
-
learning_rate=5e-5,
|
84 |
-
lr_scheduler_type="cosine",
|
85 |
-
max_steps=200,
|
86 |
-
save_strategy="no",
|
87 |
-
logging_steps=1,
|
88 |
-
output_dir=new_model,
|
89 |
-
optim="paged_adamw_32bit",
|
90 |
-
warmup_steps=100,
|
91 |
-
bf16=True,
|
92 |
-
report_to="wandb",
|
93 |
-
)
|
94 |
-
|
95 |
-
#load the dataset
|
96 |
-
dataset = load_dataset(traindata_, split='train')
|
97 |
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
|
116 |
with gr.Blocks() as demo:
|
|
|
29 |
# new_model = "Gopal2002/zehpyr-gemma-dpo-finetune"
|
30 |
new_model = output_repo
|
31 |
|
32 |
+
try:
|
33 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
34 |
+
tokenizer.pad_token = tokenizer.eos_token
|
35 |
+
tokenizer.padding_side = "left"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
+
|
38 |
+
model = AutoModelForCausalLM.from_pretrained(
|
39 |
+
model_name,
|
40 |
+
torch_dtype=torch.float16,
|
41 |
+
load_in_4bit=True
|
42 |
+
)
|
43 |
+
model.config.use_cache = False
|
44 |
+
|
45 |
+
# Reference model
|
46 |
+
ref_model = AutoModelForCausalLM.from_pretrained(
|
47 |
+
model_name,
|
48 |
+
torch_dtype=torch.float16,
|
49 |
+
load_in_4bit=True
|
50 |
+
)
|
51 |
+
|
52 |
+
# specify how to quantize the model
|
53 |
+
quantization_config = BitsAndBytesConfig(
|
54 |
+
load_in_4bit=True,
|
55 |
+
bnb_4bit_quant_type="nf4",
|
56 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
57 |
+
)
|
58 |
+
device_map = {"": torch.cuda.current_device()} if torch.cuda.is_available() else None
|
59 |
+
|
60 |
+
# Step 1: load the base model (Mistral-7B in our case) in 4-bit
|
61 |
+
model_kwargs = dict(
|
62 |
+
# attn_implementation="flash_attention_2", # set this to True if your GPU supports it (Flash Attention drastically speeds up model computations)
|
63 |
+
torch_dtype="auto",
|
64 |
+
use_cache=False, # set to False as we're going to use gradient checkpointing
|
65 |
+
device_map=device_map,
|
66 |
+
quantization_config=quantization_config,
|
67 |
+
)
|
68 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)
|
69 |
+
|
70 |
+
# Training arguments
|
71 |
+
peft_config = LoraConfig(
|
72 |
+
r=16,
|
73 |
+
lora_alpha=16,
|
74 |
+
lora_dropout=0.05,
|
75 |
+
bias="none",
|
76 |
+
task_type="CAUSAL_LM",
|
77 |
+
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
|
78 |
+
)
|
79 |
+
training_args = TrainingArguments(
|
80 |
+
per_device_train_batch_size=4,
|
81 |
+
gradient_accumulation_steps=4,
|
82 |
+
gradient_checkpointing=True,
|
83 |
+
learning_rate=5e-5,
|
84 |
+
lr_scheduler_type="cosine",
|
85 |
+
max_steps=200,
|
86 |
+
save_strategy="no",
|
87 |
+
logging_steps=1,
|
88 |
+
output_dir=new_model,
|
89 |
+
optim="paged_adamw_32bit",
|
90 |
+
warmup_steps=100,
|
91 |
+
bf16=True,
|
92 |
+
report_to="wandb",
|
93 |
+
)
|
94 |
+
|
95 |
+
#load the dataset
|
96 |
+
dataset = load_dataset(traindata_, split='train')
|
97 |
+
|
98 |
+
# dataset = load_dataset('Gopal2002/zephyr-gemma-finetune-dpo', split='train')
|
99 |
+
|
100 |
+
# Create DPO trainer
|
101 |
+
dpo_trainer = DPOTrainer(
|
102 |
+
model,
|
103 |
+
ref_model=None,
|
104 |
+
args=training_args,
|
105 |
+
train_dataset=dataset,
|
106 |
+
tokenizer=tokenizer,
|
107 |
+
peft_config=peft_config,
|
108 |
+
beta=0.1,
|
109 |
+
max_prompt_length=2048,
|
110 |
+
max_length=1536,
|
111 |
+
)
|
112 |
+
dpo_trainer.train()
|
113 |
+
return "Training Done"
|
114 |
+
except MyException as e:
|
115 |
+
return str(e)
|
116 |
|
117 |
|
118 |
with gr.Blocks() as demo:
|