text-generation-webui/modules/training.py

208 lines
12 KiB
Python
Raw Normal View History

2023-03-27 17:25:08 +00:00
import sys, torch, json, threading, time
2023-03-25 19:08:26 +00:00
from pathlib import Path
import gradio as gr
from datasets import load_dataset
import transformers
from modules import ui, shared
from peft import prepare_model_for_int8_training, LoraConfig, get_peft_model, get_peft_model_state_dict
2023-03-27 17:43:01 +00:00
WANT_INTERRUPT = False
2023-03-27 17:25:08 +00:00
CURRENT_STEPS = 0
MAX_STEPS = 0
2023-03-27 17:43:01 +00:00
CURRENT_GRADIENT_ACCUM = 1
2023-03-27 17:25:08 +00:00
2023-03-25 19:08:26 +00:00
def get_json_dataset(path: str):
def get_set():
return ['None'] + sorted(set(map(lambda x : '.'.join(str(x.name).split('.')[:-1]), Path(path).glob('*.json'))), key=str.lower)
return get_set
def create_train_interface():
with gr.Tab('Train LoRA', elem_id='lora-train-tab'):
loraName = gr.Textbox(label="Name", info="The name of your new LoRA file")
with gr.Row():
# TODO: Implement multi-device support.
microBatchSize = gr.Slider(label='Micro Batch Size', value=4, minimum=1, maximum=128, step=1, info='Per-device batch size (NOTE: multiple devices not yet implemented). Increasing this will increase VRAM usage.')
batchSize = gr.Slider(label='Batch Size', value=128, minimum=1, maximum=1024, step=4, info='Global batch size. The two batch sizes together determine gradient accumulation (gradientAccum = batch / microBatch). Higher gradient accum values lead to better quality training.')
with gr.Row():
epochs = gr.Number(label='Epochs', value=1, info='Number of times every entry in the dataset should be fed into training. So 1 means feed each item in once, 5 means feed it in five times, etc.')
learningRate = gr.Textbox(label='Learning Rate', value='3e-4', info='Learning rate, in scientific notation. 3e-4 is a good starting base point. 1e-2 is extremely high, 1e-6 is extremely low.')
2023-03-25 19:08:26 +00:00
# TODO: What is the actual maximum rank? Likely distinct per model. This might be better to somehow be on a log scale.
loraRank = gr.Slider(label='LoRA Rank', value=8, minimum=1, maximum=1024, step=4, info='LoRA Rank, or dimension count. Higher values produce a larger file with better control over the model\'s content. Smaller values produce a smaller file with less overall control. Small values like 4 or 8 are great for stylistic guidance, high values like 128 or 256 are good for teaching content upgrades. Higher ranks also require higher VRAM.')
loraAlpha = gr.Slider(label='LoRA Alpha', value=16, minimum=1, maximum=2048, step=4, info='LoRA Alpha. This divided by the rank becomes the scaling of the LoRA. Higher means stronger. A good standard value is twice your Rank.')
2023-03-25 19:48:35 +00:00
# TODO: Better explain what this does, in terms of real world effect especially.
2023-03-25 19:08:26 +00:00
loraDropout = gr.Slider(label='LoRA Dropout', minimum=0.0, maximum=1.0, step=0.025, value=0.05, info='Percentage probability for dropout of LoRA layers.')
cutoffLen = gr.Slider(label='Cutoff Length', minimum=1,maximum=2048, value=256, step=32, info='Cutoff length for text input. Essentially, how long of a line of text to feed in at a time. Higher values require drastically more VRAM.')
with gr.Row():
datasetFunction = get_json_dataset('training/datasets')
2023-03-25 19:48:35 +00:00
dataset = gr.Dropdown(choices=datasetFunction(), value='None', label='Dataset', info='The dataset file to use for training.')
2023-03-25 19:08:26 +00:00
ui.create_refresh_button(dataset, lambda : None, lambda : {'choices': datasetFunction()}, 'refresh-button')
2023-03-25 19:48:35 +00:00
evalDataset = gr.Dropdown(choices=datasetFunction(), value='None', label='Evaluation Dataset', info='The dataset file used to evaluate the model after training.')
2023-03-25 19:08:26 +00:00
ui.create_refresh_button(evalDataset, lambda : None, lambda : {'choices': datasetFunction()}, 'refresh-button')
formatsFunction = get_json_dataset('training/formats')
2023-03-25 19:48:35 +00:00
format = gr.Dropdown(choices=formatsFunction(), value='None', label='Data Format', info='The format file used to decide how to format the dataset input.')
2023-03-25 19:08:26 +00:00
ui.create_refresh_button(format, lambda : None, lambda : {'choices': formatsFunction()}, 'refresh-button')
2023-03-27 17:43:01 +00:00
with gr.Row():
startButton = gr.Button("Start LoRA Training")
stopButton = gr.Button("Interrupt")
2023-03-25 19:08:26 +00:00
output = gr.Markdown(value="(...)")
2023-03-27 17:43:01 +00:00
startEvent = startButton.click(do_train, [loraName, microBatchSize, batchSize, epochs, learningRate, loraRank, loraAlpha, loraDropout, cutoffLen, dataset, evalDataset, format], [output])
stopButton.click(doInterrupt, [], [], cancels=[], queue=False)
def doInterrupt():
global WANT_INTERRUPT
WANT_INTERRUPT = True
2023-03-25 19:08:26 +00:00
2023-03-27 17:25:08 +00:00
class Callbacks(transformers.TrainerCallback):
def on_step_begin(self, args: transformers.TrainingArguments, state: transformers.TrainerState, control: transformers.TrainerControl, **kwargs):
global CURRENT_STEPS, MAX_STEPS
2023-03-27 17:43:01 +00:00
CURRENT_STEPS = state.global_step * CURRENT_GRADIENT_ACCUM
MAX_STEPS = state.max_steps * CURRENT_GRADIENT_ACCUM
if WANT_INTERRUPT:
control.should_epoch_stop = True
control.should_training_stop = True
def on_substep_end(self, args: transformers.TrainingArguments, state: transformers.TrainerState, control: transformers.TrainerControl, **kwargs):
global CURRENT_STEPS
CURRENT_STEPS += 1
if WANT_INTERRUPT:
control.should_epoch_stop = True
control.should_training_stop = True
2023-03-27 17:25:08 +00:00
2023-03-25 19:08:26 +00:00
def cleanPath(basePath: str, path: str):
""""Strips unusual symbols and forcibly builds a path as relative to the intended directory."""
# TODO: Probably could do with a security audit to guarantee there's no ways this can be bypassed to target an unwanted path.
# Or swap it to a strict whitelist of [a-zA-Z_0-9]
path = path.replace('\\', '/').replace('..', '_')
if basePath is None:
return path
return f'{Path(basePath).absolute()}/{path}'
def do_train(loraName: str, microBatchSize: int, batchSize: int, epochs: int, learningRate: float, loraRank: int, loraAlpha: int, loraDropout: float, cutoffLen: int, dataset: str, evalDataset: str, format: str):
2023-03-27 17:43:01 +00:00
global WANT_INTERRUPT, CURRENT_STEPS, MAX_STEPS, CURRENT_GRADIENT_ACCUM
WANT_INTERRUPT = False
2023-03-27 17:25:08 +00:00
CURRENT_STEPS = 0
MAX_STEPS = 0
yield "Prepping..."
2023-03-27 17:25:08 +00:00
# == Input validation / processing ==
2023-03-25 19:08:26 +00:00
# TODO: --lora-dir PR once pulled will need to be applied here
loraName = f"loras/{cleanPath(None, loraName)}"
if dataset is None:
return "**Missing dataset choice input, cannot continue.**"
if format is None:
return "**Missing format choice input, cannot continue.**"
gradientAccumulationSteps = batchSize // microBatchSize
2023-03-27 17:43:01 +00:00
CURRENT_GRADIENT_ACCUM = gradientAccumulationSteps
2023-03-25 19:08:26 +00:00
actualLR = float(learningRate)
2023-03-25 19:57:36 +00:00
shared.tokenizer.pad_token = 0
shared.tokenizer.padding_side = "left"
2023-03-27 17:25:08 +00:00
# == Prep the dataset, format, etc ==
2023-03-25 19:08:26 +00:00
with open(cleanPath('training/formats', f'{format}.json'), 'r') as formatFile:
formatData: dict[str, str] = json.load(formatFile)
def tokenize(prompt):
2023-03-25 19:57:36 +00:00
result = shared.tokenizer(prompt, truncation=True, max_length=cutoffLen + 1, padding="max_length")
2023-03-25 19:08:26 +00:00
return {
"input_ids": result["input_ids"][:-1],
"attention_mask": result["attention_mask"][:-1],
}
def generate_prompt(data_point: dict[str, str]):
for options, data in formatData.items():
if set(options.split(',')) == set(x[0] for x in data_point.items() if len(x[1].strip()) > 0):
2023-03-25 19:08:26 +00:00
for key, val in data_point.items():
data = data.replace(f'%{key}%', val)
return data
raise RuntimeError(f'Data-point "{data_point}" has no keyset match within format "{list(formatData.keys())}"')
def generate_and_tokenize_prompt(data_point):
prompt = generate_prompt(data_point)
return tokenize(prompt)
print("Loading datasets...")
2023-03-25 19:08:26 +00:00
data = load_dataset("json", data_files=cleanPath('training/datasets', f'{dataset}.json'))
train_data = data['train'].shuffle().map(generate_and_tokenize_prompt)
if evalDataset == 'None':
evalData = None
else:
evalData = load_dataset("json", data_files=cleanPath('training/datasets', f'{evalDataset}.json'))
evalData = evalData['train'].shuffle().map(generate_and_tokenize_prompt)
2023-03-27 17:25:08 +00:00
# == Start prepping the model itself ==
2023-03-25 19:57:36 +00:00
if not hasattr(shared.model, 'lm_head') or hasattr(shared.model.lm_head, 'weight'):
print("Getting model ready...")
2023-03-25 19:57:36 +00:00
prepare_model_for_int8_training(shared.model)
print("Prepping for training...")
2023-03-25 19:08:26 +00:00
config = LoraConfig(
r=loraRank,
lora_alpha=loraAlpha,
# TODO: Should target_modules be configurable?
target_modules=[ "q_proj", "v_proj" ],
lora_dropout=loraDropout,
bias="none",
task_type="CAUSAL_LM"
)
2023-03-25 19:57:36 +00:00
loraModel = get_peft_model(shared.model, config)
2023-03-25 19:08:26 +00:00
trainer = transformers.Trainer(
2023-03-25 19:57:36 +00:00
model=loraModel,
2023-03-25 19:08:26 +00:00
train_dataset=train_data,
eval_dataset=evalData,
args=transformers.TrainingArguments(
per_device_train_batch_size=microBatchSize,
gradient_accumulation_steps=gradientAccumulationSteps,
# TODO: Should more of these be configurable? Probably.
warmup_steps=100,
num_train_epochs=epochs,
learning_rate=actualLR,
fp16=True,
logging_steps=20,
evaluation_strategy="steps" if evalData is not None else "no",
save_strategy="steps",
eval_steps=200 if evalData is not None else None,
save_steps=200,
output_dir=loraName,
save_total_limit=3,
load_best_model_at_end=True if evalData is not None else False,
# TODO: Enable multi-device support
ddp_find_unused_parameters=None
2023-03-25 19:08:26 +00:00
),
2023-03-25 19:57:36 +00:00
data_collator=transformers.DataCollatorForLanguageModeling(shared.tokenizer, mlm=False),
2023-03-27 17:25:08 +00:00
callbacks=list([Callbacks()])
2023-03-25 19:08:26 +00:00
)
2023-03-25 19:57:36 +00:00
loraModel.config.use_cache = False
old_state_dict = loraModel.state_dict
loraModel.state_dict = (
2023-03-25 19:08:26 +00:00
lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
2023-03-25 19:57:36 +00:00
).__get__(loraModel, type(loraModel))
2023-03-25 19:08:26 +00:00
if torch.__version__ >= "2" and sys.platform != "win32":
2023-03-25 19:57:36 +00:00
loraModel = torch.compile(loraModel)
2023-03-27 17:25:08 +00:00
# == Main run and monitor loop ==
# TODO: save/load checkpoints to resume from?
print("Starting training...")
2023-03-27 17:25:08 +00:00
yield "Starting..."
def threadedRun():
trainer.train()
thread = threading.Thread(target=threadedRun)
thread.start()
lastStep = 0
startTime = time.perf_counter()
while thread.is_alive():
time.sleep(0.5)
2023-03-27 17:43:01 +00:00
if WANT_INTERRUPT:
yield "Interrupting, please wait... *(Run will stop after the current training step completes.)*"
elif CURRENT_STEPS != lastStep:
2023-03-27 17:25:08 +00:00
lastStep = CURRENT_STEPS
timeElapsed = time.perf_counter() - startTime
if timeElapsed <= 0:
timerInfo = ""
2023-03-27 17:57:27 +00:00
totalTimeEstimate = 999
2023-03-27 17:25:08 +00:00
else:
its = CURRENT_STEPS / timeElapsed
if its > 1:
timerInfo = f"`{its:.2f}` it/s"
else:
timerInfo = f"`{1.0/its:.2f}` s/it"
2023-03-27 17:57:27 +00:00
totalTimeEstimate = (1.0/its) * (MAX_STEPS)
yield f"Running... **{CURRENT_STEPS}** / **{MAX_STEPS}** ... {timerInfo}, `{timeElapsed:.0f}`/`{totalTimeEstimate:.0f}` seconds"
print("Training complete, saving...")
2023-03-25 19:57:36 +00:00
loraModel.save_pretrained(loraName)
2023-03-27 17:43:01 +00:00
if WANT_INTERRUPT:
print("Training interrupted.")
yield f"Interrupted. Incomplete LoRA saved to `{loraName}`"
else:
print("Training complete!")
yield f"Done! LoRA saved to `{loraName}`"