Crystalcareai
commited on
Create train.py
Browse files
train.py
ADDED
@@ -0,0 +1,142 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
3 |
+
import random
|
4 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
+
from datasets import load_dataset
|
6 |
+
from transformers import TrainingArguments
|
7 |
+
from trl import SFTTrainer
|
8 |
+
from peft import LoraConfig
|
9 |
+
|
10 |
+
import time
|
11 |
+
random_seed = 42
|
12 |
+
torch.manual_seed(random_seed)
|
13 |
+
random.seed(random_seed)
|
14 |
+
|
15 |
+
dataset = load_dataset("HuggingFaceH4/deita-10k-v0-sft", split="train_sft")
|
16 |
+
|
17 |
+
n_ahead_talk_global = 4
|
18 |
+
n_passes_global = 2
|
19 |
+
n_ahead_global = 12
|
20 |
+
n_examples = 1_000
|
21 |
+
full_batch_size = 8
|
22 |
+
eval_and_logging_steps = 2
|
23 |
+
save_steps = 100
|
24 |
+
|
25 |
+
|
26 |
+
def model_init(params):
|
27 |
+
original = False
|
28 |
+
if params is None:
|
29 |
+
params = {}
|
30 |
+
else:
|
31 |
+
params = params.params
|
32 |
+
# save params to file
|
33 |
+
n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
|
34 |
+
n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
|
35 |
+
n_passes = params.get("n_passes", n_passes_global if not original else 1)
|
36 |
+
gumbel_temperature = params.get("gumbel_temperature", 1)
|
37 |
+
use_start_thought_token = params.get("use_start_thought_token", True)
|
38 |
+
use_end_thought_token = params.get("use_end_thought_token", True)
|
39 |
+
include_policy_loss = params.get("include_policy_loss", True)
|
40 |
+
gumbel_detach = params.get("gumbel_detach", True)
|
41 |
+
merged_talk_heads = params.get("merged_talk_heads", True)
|
42 |
+
gradient_accumulation_steps = params.get("gradient_accumulation_steps", global_gradient_accumulation_steps)
|
43 |
+
residual_think_head = params.get("residual_think_head", False)
|
44 |
+
optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)
|
45 |
+
|
46 |
+
model_id = "Crystalcareai/Quiet-Star-Custom"
|
47 |
+
tokenizer_id = "Crystalcareai/Quiet-Star-Custom"
|
48 |
+
print("Loading model")
|
49 |
+
model = AutoModelForCausalLM.from_pretrained(
|
50 |
+
model_id,
|
51 |
+
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
|
52 |
+
max_thoughts=n_ahead + n_ahead_talk + 1,
|
53 |
+
merged_talk_heads=merged_talk_heads,
|
54 |
+
merged_lm_and_talk_heads=False,
|
55 |
+
merged_lm_and_think_heads=True,
|
56 |
+
use_concat_talk_head=True,
|
57 |
+
use_shallow_think=True,
|
58 |
+
use_shallow_talk=False,
|
59 |
+
use_complex_think_head=False,
|
60 |
+
use_complex_talk_head=True,
|
61 |
+
use_weighted_talk_head=True,
|
62 |
+
trust_remote_code=True,
|
63 |
+
load_in_4bit=True,
|
64 |
+
)
|
65 |
+
print("Loaded model")
|
66 |
+
|
67 |
+
tokenizer = AutoTokenizer.from_pretrained(tokenizer_id,padding=False,truncation=True)
|
68 |
+
tokenizer.pad_token_id = tokenizer.eos_token_id
|
69 |
+
|
70 |
+
special_tokens_to_add = []
|
71 |
+
if model.use_start_thought_token:
|
72 |
+
special_tokens_to_add.append("<|startthought|>")
|
73 |
+
if model.use_end_thought_token:
|
74 |
+
special_tokens_to_add.append("<|endthought|>")
|
75 |
+
if special_tokens_to_add:
|
76 |
+
tokenizer.add_special_tokens({"additional_special_tokens": special_tokens_to_add})
|
77 |
+
model.resize_token_embeddings(len(tokenizer))
|
78 |
+
model.tokenizer = tokenizer
|
79 |
+
model.gumbel_detach = gumbel_detach
|
80 |
+
model.include_policy_loss = include_policy_loss
|
81 |
+
model.use_end_thought_token = use_end_thought_token
|
82 |
+
model.use_start_thought_token = use_start_thought_token
|
83 |
+
model.n_ahead = n_ahead
|
84 |
+
model.n_ahead_talk = n_ahead_talk
|
85 |
+
model.n_passes = n_passes
|
86 |
+
model.n_tokens_print = gradient_accumulation_steps
|
87 |
+
model.gradient_accumulation_steps = gradient_accumulation_steps
|
88 |
+
model.residual_think_head = residual_think_head
|
89 |
+
model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
|
90 |
+
model.gumbel_temperature = gumbel_temperature
|
91 |
+
model.original_mode = original
|
92 |
+
model.config_params = params
|
93 |
+
model.run_start = int(time.time())
|
94 |
+
model.kill_after = 100
|
95 |
+
model.train()
|
96 |
+
return model
|
97 |
+
|
98 |
+
|
99 |
+
batch_size = full_batch_size // n_passes_global
|
100 |
+
global_gradient_accumulation_steps = full_batch_size // batch_size
|
101 |
+
run_id = int(time.time())
|
102 |
+
training_args = TrainingArguments(
|
103 |
+
output_dir="./out",
|
104 |
+
num_train_epochs=3,
|
105 |
+
per_device_train_batch_size=1,
|
106 |
+
gradient_checkpointing=False,
|
107 |
+
optim="adamw_bnb_8bit",
|
108 |
+
logging_steps=2,
|
109 |
+
save_strategy="steps",
|
110 |
+
save_steps=300,
|
111 |
+
|
112 |
+
bf16=True,
|
113 |
+
tf32=True,
|
114 |
+
learning_rate=2e-4,
|
115 |
+
max_grad_norm=0.3,
|
116 |
+
warmup_ratio=0.00,
|
117 |
+
lr_scheduler_type="constant",
|
118 |
+
push_to_hub=False,
|
119 |
+
)
|
120 |
+
|
121 |
+
peft_config = LoraConfig(
|
122 |
+
lora_alpha=16,
|
123 |
+
lora_dropout=0.05,
|
124 |
+
r=32,
|
125 |
+
bias="none",
|
126 |
+
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj","gate_proj", "up_proj", "down_proj",],
|
127 |
+
task_type="CAUSAL_LM",
|
128 |
+
use_dora=False, # Enable Dora method
|
129 |
+
)
|
130 |
+
|
131 |
+
model = model_init(None) # Initialize the model
|
132 |
+
tokenizer = model.tokenizer
|
133 |
+
|
134 |
+
trainer = SFTTrainer(
|
135 |
+
args=training_args,
|
136 |
+
train_dataset=dataset,
|
137 |
+
model=model,
|
138 |
+
peft_config=peft_config,
|
139 |
+
tokenizer=tokenizer,
|
140 |
+
)
|
141 |
+
|
142 |
+
trainer.train()
|