ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2025-12-01 13:33:35 +00:00
commit fbb98c5c5c
47 changed files with 2646 additions and 2646 deletions
Download patch file Download diff file Expand all files Collapse all files

254
unsloth-cli.py
View file

@ -47,17 +47,17 @@ def run(args):
# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = args.model_name,
max_seq_length = args.max_seq_length,
dtype = args.dtype,
load_in_4bit = args.load_in_4bit,
model_name=args.model_name,
max_seq_length=args.max_seq_length,
dtype=args.dtype,
load_in_4bit=args.load_in_4bit,
)
# Configure PEFT model
model = FastLanguageModel.get_peft_model(
model,
r = args.r,
target_modules = [
r=args.r,
target_modules=[
"q_proj",
"k_proj",
"v_proj",
@ -66,13 +66,13 @@ def run(args):
"up_proj",
"down_proj",
],
lora_alpha = args.lora_alpha,
lora_dropout = args.lora_dropout,
bias = args.bias,
use_gradient_checkpointing = args.use_gradient_checkpointing,
random_state = args.random_state,
use_rslora = args.use_rslora,
loftq_config = args.loftq_config,
lora_alpha=args.lora_alpha,
lora_dropout=args.lora_dropout,
bias=args.bias,
use_gradient_checkpointing=args.use_gradient_checkpointing,
random_state=args.random_state,
use_rslora=args.use_rslora,
loftq_config=args.loftq_config,
)
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.
@ -102,40 +102,40 @@ def run(args):
if use_modelscope:
from modelscope import MsDataset
dataset = MsDataset.load(args.dataset, split = "train")
dataset = MsDataset.load(args.dataset, split="train")
else:
# Load and format dataset
dataset = load_dataset(args.dataset, split = "train")
dataset = dataset.map(formatting_prompts_func, batched = True)
dataset = load_dataset(args.dataset, split="train")
dataset = dataset.map(formatting_prompts_func, batched=True)
print("Data is formatted and ready!")
# Configure training arguments
training_args = SFTConfig(
per_device_train_batch_size = args.per_device_train_batch_size,
gradient_accumulation_steps = args.gradient_accumulation_steps,
warmup_steps = args.warmup_steps,
max_steps = args.max_steps,
learning_rate = args.learning_rate,
fp16 = not is_bfloat16_supported(),
bf16 = is_bfloat16_supported(),
logging_steps = args.logging_steps,
optim = args.optim,
weight_decay = args.weight_decay,
lr_scheduler_type = args.lr_scheduler_type,
seed = args.seed,
output_dir = args.output_dir,
report_to = args.report_to,
max_length = args.max_seq_length,
dataset_num_proc = 2,
packing = False,
per_device_train_batch_size=args.per_device_train_batch_size,
gradient_accumulation_steps=args.gradient_accumulation_steps,
warmup_steps=args.warmup_steps,
max_steps=args.max_steps,
learning_rate=args.learning_rate,
fp16=not is_bfloat16_supported(),
bf16=is_bfloat16_supported(),
logging_steps=args.logging_steps,
optim=args.optim,
weight_decay=args.weight_decay,
lr_scheduler_type=args.lr_scheduler_type,
seed=args.seed,
output_dir=args.output_dir,
report_to=args.report_to,
max_length=args.max_seq_length,
dataset_num_proc=2,
packing=False,
)
# Initialize trainer
trainer = SFTTrainer(
model = model,
processing_class = tokenizer,
train_dataset = dataset,
args = training_args,
model=model,
processing_class=tokenizer,
train_dataset=dataset,
args=training_args,
)
# Train model
@ -153,24 +153,24 @@ def run(args):
model.save_pretrained_gguf(
args.save_path,
tokenizer,
quantization_method = quantization_method,
quantization_method=quantization_method,
)
if args.push_model:
model.push_to_hub_gguf(
hub_path = args.hub_path,
hub_token = args.hub_token,
quantization_method = quantization_method,
hub_path=args.hub_path,
hub_token=args.hub_token,
quantization_method=quantization_method,
)
else:
print(f"Saving model with quantization method: {args.quantization}")
model.save_pretrained_gguf(
args.save_path, tokenizer, quantization_method = args.quantization
args.save_path, tokenizer, quantization_method=args.quantization
)
if args.push_model:
model.push_to_hub_gguf(
hub_path = args.hub_path,
hub_token = args.hub_token,
quantization_method = quantization_method,
hub_path=args.hub_path,
hub_token=args.hub_token,
quantization_method=quantization_method,
)
else:
model.save_pretrained_merged(args.save_path, tokenizer, args.save_method)
@ -183,38 +183,38 @@ def run(args):
if __name__ == "__main__":
# Define argument parser
parser = argparse.ArgumentParser(
description = "🦥 Fine-tune your llm faster using unsloth!"
description="🦥 Fine-tune your llm faster using unsloth!"
)
model_group = parser.add_argument_group("🤖 Model Options")
model_group.add_argument(
"--model_name",
type = str,
default = "unsloth/llama-3-8b",
help = "Model name to load",
type=str,
default="unsloth/llama-3-8b",
help="Model name to load",
)
model_group.add_argument(
"--max_seq_length",
type = int,
default = 2048,
help = "Maximum sequence length, default is 2048. We auto support RoPE Scaling internally!",
type=int,
default=2048,
help="Maximum sequence length, default is 2048. We auto support RoPE Scaling internally!",
)
model_group.add_argument(
"--dtype",
type = str,
default = None,
help = "Data type for model (None for auto detection)",
type=str,
default=None,
help="Data type for model (None for auto detection)",
)
model_group.add_argument(
"--load_in_4bit",
action = "store_true",
help = "Use 4bit quantization to reduce memory usage",
action="store_true",
help="Use 4bit quantization to reduce memory usage",
)
model_group.add_argument(
"--dataset",
type = str,
default = "yahma/alpaca-cleaned",
help = "Huggingface dataset to use for training",
type=str,
default="yahma/alpaca-cleaned",
help="Huggingface dataset to use for training",
)
lora_group = parser.add_argument_group(
@ -222,101 +222,101 @@ if __name__ == "__main__":
)
lora_group.add_argument(
"--r",
type = int,
default = 16,
help = "Rank for Lora model, default is 16. (common values: 8, 16, 32, 64, 128)",
type=int,
default=16,
help="Rank for Lora model, default is 16. (common values: 8, 16, 32, 64, 128)",
)
lora_group.add_argument(
"--lora_alpha",
type = int,
default = 16,
help = "LoRA alpha parameter, default is 16. (common values: 8, 16, 32, 64, 128)",
type=int,
default=16,
help="LoRA alpha parameter, default is 16. (common values: 8, 16, 32, 64, 128)",
)
lora_group.add_argument(
"--lora_dropout",
type = float,
default = 0.0,
help = "LoRA dropout rate, default is 0.0 which is optimized.",
type=float,
default=0.0,
help="LoRA dropout rate, default is 0.0 which is optimized.",
)
lora_group.add_argument(
"--bias", type = str, default = "none", help = "Bias setting for LoRA"
"--bias", type=str, default="none", help="Bias setting for LoRA"
)
lora_group.add_argument(
"--use_gradient_checkpointing",
type = str,
default = "unsloth",
help = "Use gradient checkpointing",
type=str,
default="unsloth",
help="Use gradient checkpointing",
)
lora_group.add_argument(
"--random_state",
type = int,
default = 3407,
help = "Random state for reproducibility, default is 3407.",
type=int,
default=3407,
help="Random state for reproducibility, default is 3407.",
)
lora_group.add_argument(
"--use_rslora", action = "store_true", help = "Use rank stabilized LoRA"
"--use_rslora", action="store_true", help="Use rank stabilized LoRA"
)
lora_group.add_argument(
"--loftq_config", type = str, default = None, help = "Configuration for LoftQ"
"--loftq_config", type=str, default=None, help="Configuration for LoftQ"
)
training_group = parser.add_argument_group("🎓 Training Options")
training_group.add_argument(
"--per_device_train_batch_size",
type = int,
default = 2,
help = "Batch size per device during training, default is 2.",
type=int,
default=2,
help="Batch size per device during training, default is 2.",
)
training_group.add_argument(
"--gradient_accumulation_steps",
type = int,
default = 4,
help = "Number of gradient accumulation steps, default is 4.",
type=int,
default=4,
help="Number of gradient accumulation steps, default is 4.",
)
training_group.add_argument(
"--warmup_steps",
type = int,
default = 5,
help = "Number of warmup steps, default is 5.",
type=int,
default=5,
help="Number of warmup steps, default is 5.",
)
training_group.add_argument(
"--max_steps", type = int, default = 400, help = "Maximum number of training steps."
"--max_steps", type=int, default=400, help="Maximum number of training steps."
)
training_group.add_argument(
"--learning_rate",
type = float,
default = 2e-4,
help = "Learning rate, default is 2e-4.",
type=float,
default=2e-4,
help="Learning rate, default is 2e-4.",
)
training_group.add_argument(
"--optim", type = str, default = "adamw_8bit", help = "Optimizer type."
"--optim", type=str, default="adamw_8bit", help="Optimizer type."
)
training_group.add_argument(
"--weight_decay",
type = float,
default = 0.01,
help = "Weight decay, default is 0.01.",
type=float,
default=0.01,
help="Weight decay, default is 0.01.",
)
training_group.add_argument(
"--lr_scheduler_type",
type = str,
default = "linear",
help = "Learning rate scheduler type, default is 'linear'.",
type=str,
default="linear",
help="Learning rate scheduler type, default is 'linear'.",
)
training_group.add_argument(
"--seed",
type = int,
default = 3407,
help = "Seed for reproducibility, default is 3407.",
type=int,
default=3407,
help="Seed for reproducibility, default is 3407.",
)
# Report/Logging arguments
report_group = parser.add_argument_group("📊 Report Options")
report_group.add_argument(
"--report_to",
type = str,
default = "tensorboard",
choices = [
type=str,
default="tensorboard",
choices=[
"azure_ml",
"clearml",
"codecarbon",
@ -331,62 +331,62 @@ if __name__ == "__main__":
"all",
"none",
],
help = "The list of integrations to report the results and logs to. Supported platforms are: \n\t\t 'azure_ml', 'clearml', 'codecarbon', 'comet_ml', 'dagshub', 'dvclive', 'flyte', 'mlflow', 'neptune', 'tensorboard', and 'wandb'. Use 'all' to report to all integrations installed, 'none' for no integrations.",
help="The list of integrations to report the results and logs to. Supported platforms are: \n\t\t 'azure_ml', 'clearml', 'codecarbon', 'comet_ml', 'dagshub', 'dvclive', 'flyte', 'mlflow', 'neptune', 'tensorboard', and 'wandb'. Use 'all' to report to all integrations installed, 'none' for no integrations.",
)
report_group.add_argument(
"--logging_steps", type = int, default = 1, help = "Logging steps, default is 1"
"--logging_steps", type=int, default=1, help="Logging steps, default is 1"
)
# Saving and pushing arguments
save_group = parser.add_argument_group("💾 Save Model Options")
save_group.add_argument(
"--output_dir", type = str, default = "outputs", help = "Output directory"
"--output_dir", type=str, default="outputs", help="Output directory"
)
save_group.add_argument(
"--save_model", action = "store_true", help = "Save the model after training"
"--save_model", action="store_true", help="Save the model after training"
)
save_group.add_argument(
"--save_method",
type = str,
default = "merged_16bit",
choices = ["merged_16bit", "merged_4bit", "lora"],
help = "Save method for the model, default is 'merged_16bit'",
type=str,
default="merged_16bit",
choices=["merged_16bit", "merged_4bit", "lora"],
help="Save method for the model, default is 'merged_16bit'",
)
save_group.add_argument(
"--save_gguf",
action = "store_true",
help = "Convert the model to GGUF after training",
action="store_true",
help="Convert the model to GGUF after training",
)
save_group.add_argument(
"--save_path", type = str, default = "model", help = "Path to save the model"
"--save_path", type=str, default="model", help="Path to save the model"
)
save_group.add_argument(
"--quantization",
type = str,
default = "q8_0",
nargs = "+",
help = "Quantization method for saving the model. common values ('f16', 'q4_k_m', 'q8_0'), Check our wiki for all quantization methods https://github.com/unslothai/unsloth/wiki#saving-to-gguf ",
type=str,
default="q8_0",
nargs="+",
help="Quantization method for saving the model. common values ('f16', 'q4_k_m', 'q8_0'), Check our wiki for all quantization methods https://github.com/unslothai/unsloth/wiki#saving-to-gguf ",
)
push_group = parser.add_argument_group("🚀 Push Model Options")
push_group.add_argument(
"--push_model",
action = "store_true",
help = "Push the model to Hugging Face hub after training",
action="store_true",
help="Push the model to Hugging Face hub after training",
)
push_group.add_argument(
"--push_gguf",
action = "store_true",
help = "Push the model as GGUF to Hugging Face hub after training",
action="store_true",
help="Push the model as GGUF to Hugging Face hub after training",
)
push_group.add_argument(
"--hub_path",
type = str,
default = "hf/model",
help = "Path on Hugging Face hub to push the model",
type=str,
default="hf/model",
help="Path on Hugging Face hub to push the model",
)
push_group.add_argument(
"--hub_token", type = str, help = "Token for pushing the model to Hugging Face hub"
"--hub_token", type=str, help="Token for pushing the model to Hugging Face hub"
)
args = parser.parse_args()