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

moved utils, dataset_utilsand datasets, updated the startTraining pydantic model

Browse source
This commit is contained in:
sshah229 2026-02-01 16:49:42 -07:00
commit c042223a7a
11 changed files with 37 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

0
backend/utils/datasets/alpaca_unsloth.json → backend/assets/datasets/alpaca_unsloth.json
View file

4
backend/backend/__init__.py
View file

@ -13,8 +13,8 @@ from .training import TrainingBackend, get_training_backend, create_training_han
from .model_config import is_vision_model, ModelConfig, scan_trained_loras
# Utilities
from .path_utils import normalize_path, is_local_path, is_model_cached
from .utils import without_hf_auth, format_error_message, get_gpu_memory_info, search_hf_models
from .dataset_utils import format_and_template_dataset
from utils.utils import without_hf_auth, format_error_message, get_gpu_memory_info, search_hf_models
from utils.datasets.dataset_utils import format_and_template_dataset
__all__ = [
# Inference

6
backend/backend/inference.py
View file

@ -11,7 +11,7 @@ import torch
from typing import Optional, Generator, Tuple
from .model_config import ModelConfig, get_base_model_from_lora
from .path_utils import is_model_cached
from .utils import format_error_message, log_gpu_memory
from utils.utils import format_error_message, log_gpu_memory
from io import StringIO
import logging
@ -489,7 +489,7 @@ class InferenceBackend:
# Step 1: Apply get_chat_template if model is in mapper
try:
from backend.dataset_utils import MODEL_TO_TEMPLATE_MAPPER, get_tokenizer_chat_template
from utils.datasets.dataset_utils import MODEL_TO_TEMPLATE_MAPPER, get_tokenizer_chat_template
model_name_lower = self.active_model_name.lower()
@ -954,7 +954,7 @@ class InferenceBackend:
}
try:
from backend.dataset_utils import MODEL_TO_TEMPLATE_MAPPER
from utils.datasets.dataset_utils import MODEL_TO_TEMPLATE_MAPPER
#Try exact match first
model_name_lower = model_name.lower()
if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:

2
backend/backend/model_config.py
View file

@ -5,7 +5,7 @@ from transformers import AutoConfig
from dataclasses import dataclass
from typing import Optional, Dict, Any
from .path_utils import normalize_path, is_local_path, is_model_cached
from .utils import without_hf_auth
from utils.utils import without_hf_auth
import logging
from pathlib import Path
from typing import List, Tuple

18
backend/backend/trainer.py
View file

@ -20,8 +20,8 @@ from datasets import Dataset, load_dataset
# Add the parent directory to sys.path to import unsloth modules
#sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from .model_config import is_vision_model
from .dataset_utils import format_and_template_dataset
from .dataset_utils import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
from utils.datasets.dataset_utils import format_and_template_dataset
from utils.datasets.dataset_utils import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
from trl import SFTTrainer, SFTConfig
# Import Unsloth trainers
@ -313,15 +313,23 @@ class UnslothTrainer:
# Load local datasets
all_data = []
for dataset_file in local_datasets:
file_path = os.path.join("datasets", dataset_file)
if file_path.endswith('.json'):
# dataset_file may already be an absolute path from routes/training.py
if os.path.isabs(dataset_file):
file_path = dataset_file
else:
# Fallback: try relative to assets/datasets
script_dir = Path(__file__).parent.parent
assets_datasets_dir = script_dir / "assets" / "datasets"
file_path = assets_datasets_dir / dataset_file
if str(file_path).endswith('.json'):
with open(file_path, 'r', encoding='utf-8') as f:
data = json.load(f)
if isinstance(data, list):
all_data.extend(data)
else:
all_data.append(data)
elif file_path.endswith('.csv'):
elif str(file_path).endswith('.csv'):
df = pd.read_csv(file_path)
all_data.extend(df.to_dict('records'))

11
backend/backend/training.py
View file

@ -66,6 +66,8 @@ class TrainingBackend:
weight_decay: float,
random_seed: int,
packing: bool,
optim: str,
lr_scheduler_type: str,
# LoRA parameters
use_lora: bool, # Should be derived from training_type
@ -89,9 +91,7 @@ class TrainingBackend:
wandb_token: str,
wandb_project: str,
enable_tensorboard: bool,
tensorboard_dir: str,
optim: str = "adamw_8bit",
lr_scheduler_type: str = "linear") -> Generator[Tuple, None, None]:
tensorboard_dir: str) -> Generator[Tuple, None, None]:
"""
Start training - yields UI updates as generator.
@ -601,12 +601,13 @@ def create_training_handlers(train_components: Dict[str, Any]) -> Dict[str, Any]
hf_dataset, local_datasets, format_type,
num_epochs, learning_rate, batch_size, gradient_accumulation_steps,
warmup_steps, warmup_ratio, max_steps, save_steps, weight_decay, random_seed, packing,
optim, lr_scheduler_type,
use_lora, lora_r, lora_alpha, lora_dropout, target_modules,
gradient_checkpointing, use_rslora, use_loftq, train_on_completions,
finetune_vision_layers, finetune_language_layers,
finetune_attention_modules, finetune_mlp_modules,
enable_wandb, wandb_token, wandb_project,
enable_tensorboard, tensorboard_dir, optim, lr_scheduler_type) = args
enable_tensorboard, tensorboard_dir) = args
# Start training with correctly named parameters - this is a generator
for update_tuple in backend.start_training(
@ -629,6 +630,8 @@ def create_training_handlers(train_components: Dict[str, Any]) -> Dict[str, Any]
weight_decay=weight_decay,
random_seed=random_seed,
packing=packing,
optim=optim,
lr_scheduler_type=lr_scheduler_type,
use_lora=use_lora,
lora_r=lora_r,
lora_alpha=lora_alpha,

4
backend/models/training.py
View file

@ -31,6 +31,8 @@ class TrainingStartRequest(BaseModel):
weight_decay: float = Field(0.01, description="Weight decay")
random_seed: int = Field(42, description="Random seed")
packing: bool = Field(False, description="Enable sequence packing")
optim: str = Field("adamw_8bit", description="Optimizer")
lr_scheduler_type: str = Field("linear", description="Learning rate scheduler type")
# LoRA parameters
use_lora: bool = Field(True, description="Use LoRA (derived from training_type)")
@ -55,8 +57,6 @@ class TrainingStartRequest(BaseModel):
wandb_project: Optional[str] = Field(None, description="W&B project name")
enable_tensorboard: bool = Field(False, description="Enable TensorBoard logging")
tensorboard_dir: Optional[str] = Field(None, description="TensorBoard directory")
optim: str = Field("adamw_8bit", description="Optimizer")
lr_scheduler_type: str = Field("linear", description="Learning rate scheduler type")
class TrainingStartResponse(BaseModel):

4
backend/routes/models.py
View file

@ -14,7 +14,7 @@ if str(backend_path) not in sys.path:
# Import backend functions
try:
from backend.utils import search_hf_models
from utils.utils import search_hf_models
from backend.model_config import (
scan_trained_loras,
load_model_defaults,
@ -28,7 +28,7 @@ except ImportError:
parent_backend = backend_path.parent / "backend"
if str(parent_backend) not in sys.path:
sys.path.insert(0, str(parent_backend))
from backend.utils import search_hf_models
from utils.utils import search_hf_models
from backend.model_config import (
scan_trained_loras,
load_model_defaults,

14
backend/routes/training.py
View file

@ -73,7 +73,7 @@ async def start_training(request: TrainingStartRequest):
validated_datasets = []
# Get the backend directory (where this file is located)
backend_dir = Path(__file__).parent.parent
utils_datasets_dir = backend_dir / "utils" / "datasets"
assets_datasets_dir = backend_dir / "assets" / "datasets"
for dataset_path in request.local_datasets:
dataset_file = Path(dataset_path)
@ -83,11 +83,11 @@ async def start_training(request: TrainingStartRequest):
# First try: relative to current working directory
candidate = Path.cwd() / dataset_path
if not candidate.exists():
# Second try: relative to utils/datasets folder
candidate = utils_datasets_dir / dataset_path
# Second try: relative to assets/datasets folder
candidate = assets_datasets_dir / dataset_path
if not candidate.exists():
# Third try: just the filename in utils/datasets
candidate = utils_datasets_dir / dataset_file.name
# Third try: just the filename in assets/datasets
candidate = assets_datasets_dir / dataset_file.name
dataset_file = candidate
if not dataset_file.exists():
@ -118,6 +118,8 @@ async def start_training(request: TrainingStartRequest):
"weight_decay": request.weight_decay,
"random_seed": request.random_seed,
"packing": request.packing,
"optim": request.optim,
"lr_scheduler_type": request.lr_scheduler_type,
"use_lora": request.use_lora,
"lora_r": request.lora_r,
"lora_alpha": request.lora_alpha,
@ -136,8 +138,6 @@ async def start_training(request: TrainingStartRequest):
"wandb_project": request.wandb_project or "",
"enable_tensorboard": request.enable_tensorboard,
"tensorboard_dir": request.tensorboard_dir or "",
"optim": request.optim,
"lr_scheduler_type": request.lr_scheduler_type,
}
# Generate job ID

0
backend/backend/dataset_utils.py → backend/utils/datasets/dataset_utils.py
View file

0
backend/backend/utils.py → backend/utils/utils.py
View file