File size: 6,232 Bytes
de26508
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig, BitsAndBytesConfig,AutoConfig
import time
import torch
torch.backends.cuda.matmul.allow_tf32 = True
import random
from datasets import load_dataset
from transformers import TrainingArguments
from trl import SFTTrainer
from peft import LoraConfig
# from accelerate import infer_auto_device_map, init_empty_weights, dispatch_model
from torch.nn import CrossEntropyLoss
torch.autograd.set_detect_anomaly(True)
random_seed = 42
torch.manual_seed(random_seed)
random.seed(random_seed)
# Set the device for each process
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# torch.cuda.set_device(device)



n_ahead_talk_global = 4
n_passes_global = 2
n_ahead_global = 8
n_examples = 0

def model_init(params):
    original = False
    if params is None:
        params = {}
    else:
        params = params.params
    # save params to file
    n_ahead = params.get("n_ahead", n_ahead_global if not original else 1)
    n_ahead_talk = params.get("n_ahead_talk", n_ahead_talk_global if not original else 1)
    n_passes = params.get("n_passes", n_passes_global if not original else 1)
    gumbel_temperature = params.get("gumbel_temperature", 1)
    use_start_thought_token = params.get("use_start_thought_token", True)
    use_end_thought_token = params.get("use_end_thought_token", True)
    include_policy_loss = params.get("include_policy_loss", True)
    gumbel_detach = params.get("gumbel_detach", True)
    merged_talk_heads = params.get("merged_talk_heads", True)
    residual_think_head = params.get("residual_think_head", False)
    optimize_lm_head_only_at_start = params.get("optimize_lm_head_only_at_start", False)

    model_id = "LeroyDyer/SpydazWeb_AGI_MistralStar"
    tokenizer_id = model_id
    print("Loading model")

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        max_thoughts=n_ahead + n_ahead_talk + 1,
        merged_talk_heads=merged_talk_heads,
        merged_lm_and_talk_heads=False,
        merged_lm_and_think_heads=True,
        use_concat_talk_head=True,
        use_shallow_think=True,
        use_shallow_talk=False,
        use_complex_think_head=False,
        use_complex_talk_head=True,
        use_weighted_talk_head=True,
        trust_remote_code=True,
        device_map="auto",
    )
    print("Loaded model")

    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, truncation=True, padding_side="right")
    tokenizer.pad_token_id = tokenizer.eos_token_id



    model.gumbel_detach = gumbel_detach
    model.include_policy_loss = include_policy_loss
    model.use_end_thought_token = use_end_thought_token
    model.use_start_thought_token = use_start_thought_token
    model.n_ahead = n_ahead
    model.n_ahead_talk = n_ahead_talk
    model.n_passes = n_passes
    model.residual_think_head = residual_think_head
    model.optimize_lm_head_only_at_start = optimize_lm_head_only_at_start
    model.gumbel_temperature = gumbel_temperature
    model.original_mode = original
    model.config_params = params
    return model
    
model,tokenizer = model_init(None)


## TRAINING : 

peft_config = LoraConfig(
          r = 128, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj","lm_head", "embed_tokens"],
    lora_alpha = 32,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none", 
    use_dora=True,
)

from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline, AutoConfig
from datasets import load_dataset
from transformers import TrainingArguments
from trl import SFTTrainer
from peft import LoraConfig

## DATA
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""
EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
def formatting_prompts_func(examples):
    instructions = examples["instruction"]
    inputs       = examples["input"]
    outputs      = examples["output"]
    texts = []
    for instruction, input, output in zip(instructions, inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts, }
pass
dataset = load_dataset("gate369/Alpaca-Star", split = "train[:2000]")
dataset = dataset.shuffle(seed=3704)
dataset = dataset.map(formatting_prompts_func, batched = True,)
## TRAIN

max_seq_length = 1024
training_args = TrainingArguments(
    output_dir="./out",
    num_train_epochs=3,
    per_device_train_batch_size=1,
    gradient_checkpointing=False,
    gradient_accumulation_steps=8,
    optim="lion_32bit",
    logging_steps=1,
    save_strategy="steps",
    max_steps=1000,
    bf16=True,
    tf32=False,
    learning_rate=6e-05,
    max_grad_norm=0.3,
    warmup_ratio=0.06,
    lr_scheduler_type="cosine",
    push_to_hub=False,

)
trainer = SFTTrainer(
    args=training_args,
    train_dataset=dataset,
    model=model,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length,
    dataset_text_field="text",
    peft_config=peft_config,
)
trainer.train()

## SAVE
tokenizer.save_pretrained("SFTTrainerModel")
model.save_pretrained("SFTTrainerModel")


import os
import huggingface_hub
from huggingface_hub import notebook_login
from huggingface_hub import create_repo, HfApi
from huggingface_hub import hf_hub_download
from huggingface_hub import create_repo, HfApi
from huggingface_hub import snapshot_download

MODEL_NAME = "_Spydaz_Web_AI_MistralStar"
Folderinput = "SFTTrainerModel"
WRITE_TOKEN = ""
username = "LeroyDyer"
huggingface_hub.login(WRITE_TOKEN)
api = HfApi(token=WRITE_TOKEN)
# Create empty repo
api.create_repo(
    repo_id = f"{username}/{MODEL_NAME}",
    repo_type="model",
    exist_ok=True,
)

api.upload_folder(
    repo_id = f"{username}/{MODEL_NAME}",
    folder_path = Folderinput
)