Spaces:
Build error
Build error
File size: 18,054 Bytes
e7ee027 c90d519 d7cdbfb e7ee027 d7cdbfb e7ee027 d7cdbfb e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 bf9d89b e7ee027 d7cdbfb e7ee027 3644ff3 e7ee027 d7cdbfb e7ee027 bf9d89b e7ee027 3644ff3 e7ee027 bf9d89b e7ee027 3644ff3 e7ee027 bf9d89b e7ee027 3644ff3 e7ee027 bf9d89b e7ee027 3644ff3 e7ee027 3644ff3 e7ee027 c90d519 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 |
import os
import argparse
import random
import torch
import gradio as gr
import transformers
from datasets import Dataset
from transformers import LlamaForCausalLM, LlamaTokenizer, GenerationConfig
from peft import prepare_model_for_int8_training, LoraConfig, get_peft_model, PeftModel
model = None
tokenizer = None
peft_model = None
def random_hyphenated_word():
word_list = ['apple', 'banana', 'cherry', 'date', 'elderberry', 'fig']
word1 = random.choice(word_list)
word2 = random.choice(word_list)
return word1 + '-' + word2
def maybe_load_models():
global model
global tokenizer
if model is None:
model = LlamaForCausalLM.from_pretrained(
"decapoda-research/llama-7b-hf",
load_in_8bit=True,
torch_dtype=torch.float16,
device_map="auto",
)
if tokenizer is None:
tokenizer = LlamaTokenizer.from_pretrained(
"decapoda-research/llama-7b-hf",
)
def reset_models():
global model
global tokenizer
del model
del tokenizer
model = None
tokenizer = None
def generate_text(
model_name,
text,
temperature,
top_p,
top_k,
repeat_penalty,
max_new_tokens,
progress=gr.Progress(track_tqdm=True)
):
global model
global tokenizer
maybe_load_models()
tokenizer.pad_token_id = 0
if model_name and model_name != "None":
model = PeftModel.from_pretrained(
model, model_name,
torch_dtype=torch.float16
)
inputs = tokenizer(text, return_tensors="pt")
input_ids = inputs["input_ids"].to(model.device)
# llama_config = transformers.LlamaConfig()
# print(llama_config)
stopping_criteria_list = transformers.StoppingCriteriaList()
generation_config = GenerationConfig(
# Whether to use greedy decoding. If set to False,
do_sample=True,
# Controls the 'temperature' of the softmax distribution during sampling.
# Higher values (e.g., 1.0) make the model generate more diverse and random outputs,
# while lower values (e.g., 0.1) make it more deterministic and
# focused on the highest probability tokens.
temperature=temperature,
# Sets the nucleus sampling threshold. In nucleus sampling,
# only the tokens whose cumulative probability exceeds 'top_p' are considered
# for sampling. This technique helps to reduce the number of low probability
# tokens considered during sampling, which can lead to more diverse and coherent outputs.
top_p=top_p,
# Sets the number of top tokens to consider during sampling.
# In top-k sampling, only the 'top_k' tokens with the highest probabilities
# are considered for sampling. This method can lead to more focused and coherent
# outputs by reducing the impact of low probability tokens.
top_k=top_k,
# Applies a penalty to the probability of tokens that have already been generated,
# discouraging the model from repeating the same words or phrases. The penalty is
# applied by dividing the token probability by a factor based on the number of times
# the token has appeared in the generated text.
repeat_penalty=repeat_penalty,
# Limits the maximum number of tokens generated in a single iteration.
# This can be useful to control the length of generated text, especially in tasks
# like text summarization or translation, where the output should not be excessively long.
max_new_tokens=max_new_tokens,
# typical_p=1,
# stopping_criteria=stopping_criteria_list,
# eos_token_id=llama_config.eos_token_id,
# pad_token_id=llama_config.eos_token_id
)
with torch.no_grad():
generation_output = model.generate(
input_ids=input_ids,
attention_mask=torch.ones_like(input_ids),
generation_config=generation_config,
# return_dict_in_generate=True,
# output_scores=True,
# eos_token_id=[tokenizer.eos_token_id],
use_cache=True,
)[0].cuda()
output_text = tokenizer.decode(generation_output)
return output_text.strip()
def tokenize_and_train(
training_text,
max_seq_length,
micro_batch_size,
gradient_accumulation_steps,
epochs,
learning_rate,
lora_r,
lora_alpha,
lora_dropout,
model_name,
progress=gr.Progress(track_tqdm=True)
):
global model
global tokenizer
reset_models()
maybe_load_models()
tokenizer.pad_token_id = 0
paragraphs = training_text.split("\n\n\n")
print("Number of samples: " + str(len(paragraphs)))
def tokenize(item):
result = tokenizer(
item["text"],
truncation=True,
max_length=max_seq_length,
padding="max_length",
)
return {
"input_ids": result["input_ids"][:-1],
"attention_mask": result["attention_mask"][:-1],
}
def to_dict(text):
return {"text": text}
paragraphs = [to_dict(x) for x in paragraphs]
data = Dataset.from_list(paragraphs)
data = data.shuffle().map(lambda x: tokenize(x))
model = prepare_model_for_int8_training(model)
model = get_peft_model(model, LoraConfig(
r=lora_r,
lora_alpha=lora_alpha,
target_modules=["q_proj", "v_proj"],
lora_dropout=lora_dropout,
bias="none",
task_type="CAUSAL_LM",
))
output_dir = f"lora-{model_name}"
print("Training...")
training_args = transformers.TrainingArguments(
# Set the batch size for training on each device (GPU, CPU, or TPU).
per_device_train_batch_size=micro_batch_size,
# Number of steps for gradient accumulation. This is useful when the total
# batch size is too large to fit in GPU memory. The effective batch size
# will be the product of 'per_device_train_batch_size' and 'gradient_accumulation_steps'.
gradient_accumulation_steps=gradient_accumulation_steps,
# Number of warmup steps for the learning rate scheduler. During these steps,
# the learning rate increases linearly from 0 to its initial value. Warmup helps
# to reduce the risk of very large gradients at the beginning of training,
# which could destabilize the model.
# warmup_steps=100,
# The total number of training steps. The training process will end once this
# number is reached, even if not all the training epochs are completed.
# max_steps=1500,
# The total number of epochs (complete passes through the training data)
# to perform during the training process.
num_train_epochs=epochs,
# The initial learning rate to be used during training.
learning_rate=learning_rate,
# Enables mixed precision training using 16-bit floating point numbers (FP16).
# This can speed up training and reduce GPU memory consumption without
# sacrificing too much model accuracy.
fp16=True,
# The frequency (in terms of steps) of logging training metrics and statistics
# like loss, learning rate, etc. In this case, it logs after every 20 steps.
logging_steps=20,
# The output directory where the trained model, checkpoints,
# and other training artifacts will be saved.
output_dir=output_dir,
# The maximum number of checkpoints to keep. When this limit is reached,
# the oldest checkpoint will be deleted to save a new one. In this case,
# a maximum of 3 checkpoints will be kept.
save_total_limit=3,
)
trainer = transformers.Trainer(
# The pre-trained model that you want to fine-tune or train from scratch.
# 'model' should be an instance of a Hugging Face Transformer model, such as BERT, GPT-2, T5, etc.
model=model,
# The dataset to be used for training. 'data' should be a PyTorch Dataset or
# a compatible format, containing the input samples and labels or masks (if required).
train_dataset=data,
# The TrainingArguments instance created earlier, which contains various
# hyperparameters and configurations for the training process.
args=training_args,
# A callable that takes a batch of samples and returns a batch of inputs for the model.
# This is used to prepare the input samples for training by batching, padding, and possibly masking.
data_collator=transformers.DataCollatorForLanguageModeling(
tokenizer,
# Whether to use masked language modeling (MLM) during training.
# MLM is a training technique used in models like BERT, where some tokens in the
# input are replaced by a mask token, and the model tries to predict the
# original tokens. In this case, MLM is set to False, indicating that it will not be used.
mlm=False,
),
)
result = trainer.train(resume_from_checkpoint=False)
model.save_pretrained(output_dir)
reset_models()
return result
with gr.Blocks(
css="#refresh-button { max-width: 32px }",
title="Simple LLaMA Finetuner") as demo:
with gr.Tab("Finetuning"):
with gr.Column():
training_text = gr.Textbox(lines=12, label="Training Data", info="Each sequence must be separated by a double newline")
max_seq_length = gr.Slider(
minimum=1, maximum=4096, value=512,
label="Max Sequence Length",
info="The maximum length of each sample text sequence. Sequences longer than this will be truncated."
)
with gr.Row():
with gr.Column():
micro_batch_size = gr.Slider(
minimum=1, maximum=100, value=1,
label="Micro Batch Size",
info="The number of examples in each mini-batch for gradient computation. A smaller micro_batch_size reduces memory usage but may increase training time."
)
gradient_accumulation_steps = gr.Slider(
minimum=1, maximum=10, value=1,
label="Gradient Accumulation Steps",
info="The number of steps to accumulate gradients before updating model parameters. This can be used to simulate a larger effective batch size without increasing memory usage."
)
epochs = gr.Slider(
minimum=1, maximum=100, value=1,
label="Epochs",
info="The number of times to iterate over the entire training dataset. A larger number of epochs may improve model performance but also increase the risk of overfitting.")
learning_rate = gr.Slider(
minimum=0.00001, maximum=0.01, value=3e-4,
label="Learning Rate",
info="The initial learning rate for the optimizer. A higher learning rate may speed up convergence but also cause instability or divergence. A lower learning rate may require more steps to reach optimal performance but also avoid overshooting or oscillating around local minima."
)
with gr.Column():
lora_r = gr.Slider(
minimum=1, maximum=16, value=8,
label="LoRA R",
info="The rank parameter for LoRA, which controls the dimensionality of the rank decomposition matrices. A larger lora_r increases the expressiveness and flexibility of LoRA but also increases the number of trainable parameters and memory usage."
)
lora_alpha = gr.Slider(
minimum=1, maximum=128, value=16,
label="LoRA Alpha",
info="The scaling parameter for LoRA, which controls how much LoRA affects the original pre-trained model weights. A larger lora_alpha amplifies the impact of LoRA but may also distort or override the pre-trained knowledge."
)
lora_dropout = gr.Slider(
minimum=0, maximum=1, value=0.01,
label="LoRA Dropout",
info="The dropout probability for LoRA, which controls the fraction of LoRA parameters that are set to zero during training. A larger lora_dropout increases the regularization effect of LoRA but also increases the risk of underfitting."
)
with gr.Column():
model_name = gr.Textbox(
lines=1, label="LoRA Model Name", value=random_hyphenated_word()
)
with gr.Row():
train_btn = gr.Button(
"Train", variant="primary", label="Train",
)
abort_button = gr.Button(
"Abort", label="Abort",
)
output_text = gr.Text("Training Status")
train_progress = train_btn.click(
fn=tokenize_and_train,
inputs=[
training_text,
max_seq_length,
micro_batch_size,
gradient_accumulation_steps,
epochs,
learning_rate,
lora_r,
lora_alpha,
lora_dropout,
model_name
],
outputs=output_text
)
abort_button.click(None, None, None, cancels=[train_progress])
with gr.Tab("Inference"):
with gr.Row():
with gr.Column():
with gr.Row():
lora_model = gr.Dropdown(
label="LoRA Model",
)
refresh_models_list = gr.Button(
"Reload Models",
elem_id="refresh-button"
)
inference_text = gr.Textbox(lines=7, label="Input Text")
inference_output = gr.Textbox(lines=12, label="Output Text")
with gr.Row():
with gr.Column():
# temperature, top_p, top_k, repeat_penalty, max_new_tokens
temperature = gr.Slider(
minimum=0, maximum=1.99, value=0.7, step=0.01,
label="Temperature",
info="Controls the 'temperature' of the softmax distribution during sampling. Higher values (e.g., 1.0) make the model generate more diverse and random outputs, while lower values (e.g., 0.1) make it more deterministic and focused on the highest probability tokens."
)
top_p = gr.Slider(
minimum=0, maximum=1, value=0.2, step=0.01,
label="Top P",
info="Sets the nucleus sampling threshold. In nucleus sampling, only the tokens whose cumulative probability exceeds 'top_p' are considered for sampling. This technique helps to reduce the number of low probability tokens considered during sampling, which can lead to more diverse and coherent outputs."
)
top_k = gr.Slider(
minimum=0, maximum=200, value=50, step=1,
label="Top K",
info="Sets the number of top tokens to consider during sampling. In top-k sampling, only the 'top_k' tokens with the highest probabilities are considered for sampling. This method can lead to more focused and coherent outputs by reducing the impact of low probability tokens."
)
repeat_penalty = gr.Slider(
minimum=0, maximum=1.5, value=0.8, step=0.01,
label="Repeat Penalty",
info="Applies a penalty to the probability of tokens that have already been generated, discouraging the model from repeating the same words or phrases. The penalty is applied by dividing the token probability by a factor based on the number of times the token has appeared in the generated text."
)
max_new_tokens = gr.Slider(
minimum=0, maximum=4096, value=50, step=1,
label="Max New Tokens",
info="Limits the maximum number of tokens generated in a single iteration."
)
with gr.Column():
with gr.Row():
generate_btn = gr.Button(
"Generate", variant="primary", label="Generate",
)
inference_abort_button = gr.Button(
"Abort", label="Abort",
)
inference_progress = generate_btn.click(
fn=generate_text,
inputs=[
lora_model,
inference_text,
temperature,
top_p,
top_k,
repeat_penalty,
max_new_tokens
],
outputs=inference_output,
)
lora_model.change(
fn=reset_models
)
def update_models_list():
return gr.Dropdown.update(choices=["None"] + [
d for d in os.listdir() if os.path.isdir(d) and d.startswith('lora-')
], value="None")
refresh_models_list.click(
update_models_list,
inputs=None,
outputs=lora_model,
)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Simple LLaMA Finetuner")
parser.add_argument("-s", "--share", action="store_true", help="Enable sharing of the Gradio interface")
args = parser.parse_args()
demo.queue().launch(share=args.share) |