Upload dpo-train.py with huggingface_hub
Browse files- dpo-train.py +122 -0
dpo-train.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from unsloth import PatchDPOTrainer # This line is from the DPO Zephyr example ******
|
2 |
+
PatchDPOTrainer()
|
3 |
+
from huggingface_hub import HfApi
|
4 |
+
from huggingface_hub import create_repo
|
5 |
+
from unsloth import FastLanguageModel
|
6 |
+
import torch
|
7 |
+
from datasets import load_dataset
|
8 |
+
import random
|
9 |
+
|
10 |
+
max_seq_length = 4096 # Choose any! We auto support RoPE Scaling internally!
|
11 |
+
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
12 |
+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
|
13 |
+
repo_name = "dpo-v1-Nemo"
|
14 |
+
# do wandb stuff
|
15 |
+
import wandb
|
16 |
+
import random
|
17 |
+
wandb.init(
|
18 |
+
project="huggingface",
|
19 |
+
name= repo_name,)
|
20 |
+
|
21 |
+
|
22 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
23 |
+
model_name = "ijic062/Nemo-v1.1",
|
24 |
+
max_seq_length = max_seq_length,
|
25 |
+
dtype = dtype,
|
26 |
+
load_in_4bit = load_in_4bit,
|
27 |
+
token = "", # use one if using gated models like meta-llama/Llama-2-7b-hf
|
28 |
+
)
|
29 |
+
|
30 |
+
########################################################################################################
|
31 |
+
|
32 |
+
model = FastLanguageModel.get_peft_model(
|
33 |
+
|
34 |
+
model,
|
35 |
+
r = 64, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
|
36 |
+
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
|
37 |
+
"gate_proj", "up_proj", "down_proj",],
|
38 |
+
lora_alpha = 16,
|
39 |
+
lora_dropout = 0, # Supports any, but = 0 is optimized
|
40 |
+
bias = "none", # Supports any, but = "none" is optimized
|
41 |
+
# [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
|
42 |
+
use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
|
43 |
+
random_state = 3407,
|
44 |
+
use_rslora = False, # We support rank stabilized LoRA
|
45 |
+
loftq_config = None, # And LoftQ
|
46 |
+
|
47 |
+
)
|
48 |
+
|
49 |
+
######################################################################################################### ***
|
50 |
+
|
51 |
+
dataset = load_dataset(
|
52 |
+
"Chaser-cz/dpo-nice-prompt"
|
53 |
+
)
|
54 |
+
|
55 |
+
train_dataset = dataset['train'].shuffle(seed=random.randint(1, 9999))
|
56 |
+
|
57 |
+
# Shuffles data and take a small portion
|
58 |
+
# test_dataset = dataset['test_prefs']
|
59 |
+
|
60 |
+
column_names = list(dataset["train"].features)
|
61 |
+
print(f"This is column names: {column_names}")
|
62 |
+
|
63 |
+
import pprint
|
64 |
+
row = train_dataset[9]
|
65 |
+
pprint.pprint(row["prompt"])
|
66 |
+
pprint.pprint(row["chosen"])
|
67 |
+
pprint.pprint(row["rejected"])
|
68 |
+
##########################################################################################################
|
69 |
+
|
70 |
+
from unsloth import PatchDPOTrainer
|
71 |
+
PatchDPOTrainer()
|
72 |
+
from trl import DPOTrainer
|
73 |
+
from transformers import TrainingArguments
|
74 |
+
from unsloth import is_bfloat16_supported
|
75 |
+
|
76 |
+
dpo_trainer = DPOTrainer(
|
77 |
+
model = model,
|
78 |
+
beta = 0.5,
|
79 |
+
tokenizer = tokenizer,
|
80 |
+
max_length = 1024,
|
81 |
+
max_prompt_length = 512,
|
82 |
+
train_dataset = train_dataset,
|
83 |
+
ref_model = None,
|
84 |
+
# dataset_text_field = "text",
|
85 |
+
# max_seq_length = max_seq_length,
|
86 |
+
# dataset_num_proc = 2,
|
87 |
+
# packing = False, # Can make training 5x faster for short sequences.
|
88 |
+
args = TrainingArguments(
|
89 |
+
# loss_type = "sigmoid",
|
90 |
+
per_device_train_batch_size = 2,
|
91 |
+
gradient_accumulation_steps = 32,
|
92 |
+
gradient_checkpointing= True,
|
93 |
+
warmup_steps = 5,
|
94 |
+
#num_train_epochs = 3,
|
95 |
+
max_steps = 1000,
|
96 |
+
learning_rate = 2.5e-4,
|
97 |
+
fp16 = not is_bfloat16_supported(),
|
98 |
+
bf16 = is_bfloat16_supported(),
|
99 |
+
logging_steps = 1,
|
100 |
+
optim = "adamw_8bit",
|
101 |
+
weight_decay = 0.07,
|
102 |
+
lr_scheduler_type = "cosine",
|
103 |
+
seed = 3407,
|
104 |
+
output_dir = "outputs/dpo-out-13b",
|
105 |
+
save_strategy = "steps",
|
106 |
+
save_steps = 500,
|
107 |
+
),
|
108 |
+
)
|
109 |
+
|
110 |
+
dpo_trainer.train()
|
111 |
+
|
112 |
+
########################################################################################################### ***
|
113 |
+
model.save_pretrained_merged("outputs/dpo-out-13b/merged", tokenizer, save_method = "merged_16bit")
|
114 |
+
api = HfApi()
|
115 |
+
create_repo(f"jic062/{repo_name}", repo_type="model",private=True,token="")
|
116 |
+
api.upload_folder(
|
117 |
+
folder_path="outputs/dpo-out-13b/merged",
|
118 |
+
repo_id=f"jic062/{repo_name}",
|
119 |
+
repo_type="model",
|
120 |
+
)
|
121 |
+
wandb.finish()
|
122 |
+
|