Finetune / appf.py
eBlessings's picture
Rename app.py to appf.py
e2c36c9 verified
#!/usr/bin/env python3
"""
app.py – Quranic Data Training Pipeline Endpoint for ZeroGPU Spaces
--------------------------------------------------------------------
This script integrates a full Quranic data processing and training pipeline
into a Gradio interface endpoint. It is optimized for CPU/GPU-based training
on Hugging Face ZeroGPU (using the Gradio SDK) and uses chunked incremental
training, memory management, and gradient checkpointing to efficiently update
Google's Gemma-2-2b model with Quranic data.
Requirements:
- Transformers (>=4.42.0)
- Gradio (>=5.12.0)
- PyTorch (==2.2.2)
- psutil (==5.9.5)
- Accelerate (>=0.26.0)
- Hugging Face PRO subscription with ZeroGPU enabled (ensure your HF token is set as an environment variable HF_TOKEN)
- Ubuntu CPU/Linux with access to ZeroGPU hardware via Spaces
- Input data files placed in the project root.
- Sufficient storage in "working_directory"
Author: [M-Saddam Hussain]
Date: March 2025
Data References: [Tanzil.net, IslamSource, QuranicCorpus]
"""
import json
import logging
import os
import traceback
import gc
import time
import psutil
import math
import shutil
from datetime import datetime
from typing import Dict, List, Optional
from dataclasses import dataclass, asdict
import torch
# Limit PyTorch threads for CPU stability.
torch.set_num_threads(8)
from torch.utils.data import Dataset
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
TrainingArguments,
Trainer,
DataCollatorForLanguageModeling,
__version__ as transformers_version
)
from threading import Lock
import gradio as gr
import spaces
# Check for minimum required Transformers version for custom model support
MIN_TRANSFORMERS_VERSION = "4.42.0"
if tuple(map(int, transformers_version.split("."))) < tuple(map(int, MIN_TRANSFORMERS_VERSION.split("."))):
logging.warning(f"Transformers version {transformers_version} detected. Please upgrade to at least {MIN_TRANSFORMERS_VERSION} for proper support of the 'gemma2' architecture.")
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
handlers=[
logging.FileHandler('pipeline.log'),
logging.StreamHandler()
]
)
logger = logging.getLogger(__name__)
def manage_memory(threshold_percent: int = 90, min_available_mb: int = 500, sleep_duration: int = 10):
"""
Check memory usage; if usage is high or available memory is low,
force garbage collection and sleep briefly.
"""
vm = psutil.virtual_memory()
used_percent = vm.percent
available_mb = vm.available / (1024 * 1024)
logger.info(f"Memory usage: {used_percent}% used, {available_mb:.2f} MB available")
if used_percent > threshold_percent or available_mb < min_available_mb:
logger.warning("High memory usage detected, forcing garbage collection and sleeping...")
gc.collect()
time.sleep(sleep_duration)
def manage_gpu_resources(sleep_duration: int = 5):
"""
Checks GPU memory and empties cache if necessary.
"""
if torch.cuda.is_available():
allocated = torch.cuda.memory_allocated() / (1024 * 1024)
cached = torch.cuda.memory_reserved() / (1024 * 1024)
logger.info(f"GPU Memory Allocated: {allocated:.2f} MB, Reserved: {cached:.2f} MB")
torch.cuda.empty_cache()
time.sleep(sleep_duration)
def zip_checkpoint(checkpoint_dir: str) -> str:
"""
Zips the checkpoint directory and returns the path to the zip file.
"""
zip_file = checkpoint_dir + ".zip"
# Remove existing zip if it exists
if os.path.exists(zip_file):
os.remove(zip_file)
shutil.make_archive(checkpoint_dir, 'zip', checkpoint_dir)
return os.path.basename(zip_file)
@dataclass
class WordAnalysis:
"""Structured representation of word-level analysis"""
arabic: str
translation: str
position: str
morphology: Dict
features: List[str]
root: str
location: str
metadata: Dict
@dataclass
class VerseData:
"""Structured representation of verse-level data"""
chapter: int
verse: int
arabic_text: str
translation: str
words: List[WordAnalysis]
metadata: Dict
class QuranicDataset(Dataset):
"""Custom dataset for Quranic text training."""
def __init__(self, processed_data: List[Dict], tokenizer):
self.examples = []
self.tokenizer = tokenizer
for verse_data in processed_data:
self.examples.extend(self._create_training_examples(verse_data))
def _create_training_examples(self, verse_data: Dict) -> List[Dict]:
examples = []
text_block = (
f"[VERSE {verse_data['chapter']}:{verse_data['verse']}]\n"
f"Arabic: {verse_data['arabic_text']}\n"
f"Translation: {verse_data['translation']}\n"
"Morphological Analysis:\n"
)
for word in verse_data['words']:
text_block += (
f"[WORD] {word['arabic']}\n"
f"Root: {word['root']}\n"
f"Features: {', '.join(word['features'])}\n"
)
examples.append(self._format_example(text_block))
return examples
def _format_example(self, text: str) -> Dict:
encodings = self.tokenizer(
text,
truncation=True,
max_length=64,
padding="max_length",
return_tensors="pt"
)
return {
"input_ids": encodings["input_ids"][0],
"attention_mask": encodings["attention_mask"][0]
}
def __len__(self):
return len(self.examples)
def __getitem__(self, idx):
return self.examples[idx]
class QuranicDataProcessor:
"""Processes Quranic data into structured training examples."""
def __init__(self, source_dir: str, output_dir: str):
self.source_dir = source_dir
self.output_dir = output_dir
self.morphological_data: Dict[str, Dict] = {}
self.word_by_word_data: Dict[str, List[str]] = {}
self.translation_data: Dict[str, str] = {}
self.processing_lock = Lock()
os.makedirs(output_dir, exist_ok=True)
os.makedirs(os.path.join(output_dir, 'json'), exist_ok=True)
os.makedirs(os.path.join(output_dir, 'txt'), exist_ok=True)
os.makedirs(os.path.join(output_dir, 'checkpoints'), exist_ok=True)
logger.info(f"Initialized processor with source dir: {source_dir}")
def load_source_files(self) -> bool:
"""Loads morphological, translation, and word-by-word data from project root."""
try:
logger.info("Loading morphological data...")
morph_path = os.path.join(self.source_dir, 'quranic-corpus-morphology-0.4.txt')
with open(morph_path, 'r', encoding='utf-8') as f:
next(f)
for line in f:
if line.strip() and not line.startswith('#'):
parts = line.strip().split('\t')
if len(parts) >= 4:
location = parts[0].strip('()')
self.morphological_data[location] = {
'form': parts[1],
'tag': parts[2],
'features': parts[3]
}
logger.info(f"Loaded {len(self.morphological_data)} morphological entries")
logger.info("Loading translation data...")
trans_path = os.path.join(self.source_dir, 'en.sample.quran-maududi.txt')
with open(trans_path, 'r', encoding='utf-8') as f:
next(f)
for line in f:
if line.strip():
parts = line.strip().split('|')
if len(parts) >= 3:
key = f"{parts[0]}:{parts[1]}"
self.translation_data[key] = parts[2].strip()
logger.info(f"Loaded {len(self.translation_data)} verse translations")
logger.info("Loading word-by-word data...")
word_path = os.path.join(self.source_dir, 'en.w4w.qurandev.txt')
with open(word_path, 'r', encoding='utf-8-sig') as f:
lines = [line.strip() for line in f if line.strip()]
sorted_keys = sorted(self.translation_data.keys(), key=lambda x: (int(x.split(':')[0]), int(x.split(':')[1])))
if len(lines) != len(sorted_keys):
logger.warning("Mismatch between word-by-word file and translation data")
for i, verse_key in enumerate(sorted_keys):
if i < len(lines):
words = [w.strip() for w in lines[i].split('|') if w.strip()]
self.word_by_word_data[verse_key] = words
logger.info(f"Loaded word-by-word data for {len(self.word_by_word_data)} verses")
return True
except Exception as e:
logger.error(f"Error loading source files: {str(e)}")
logger.error(traceback.format_exc())
return False
def process_verse(self, chapter: int, verse: int) -> Optional[VerseData]:
"""Processes a single verse into structured format."""
try:
verse_ref = f"{chapter}:{verse}"
logger.info(f"Processing verse {verse_ref}")
translation = self.translation_data.get(verse_ref)
if not translation:
logger.warning(f"No translation for verse {verse_ref}")
return None
verse_word_list = self.word_by_word_data.get(verse_ref, [])
if not verse_word_list:
logger.warning(f"No word-by-word data for verse {verse_ref}")
return None
verse_words: List[WordAnalysis] = []
arabic_text = ""
for pos in range(1, len(verse_word_list) + 1):
pattern = f"{chapter}:{verse}:{pos}:"
matching_entries = [data for loc, data in self.morphological_data.items() if loc.startswith(pattern)]
if not matching_entries:
logger.debug(f"No morphological data for {pattern}")
continue
combined_form = " ".join(entry['form'] for entry in matching_entries)
combined_features = []
root = ""
for entry in matching_entries:
features = entry['features'].split('|')
combined_features.extend(features)
if not root:
for f in features:
if 'ROOT:' in f:
root = f.split('ROOT:')[1]
break
word_translation = verse_word_list[pos - 1]
word = WordAnalysis(
arabic=combined_form,
translation=word_translation,
position=str(pos),
morphology=matching_entries[0],
features=combined_features,
root=root,
location=f"{chapter}:{verse}:{pos}",
metadata={}
)
verse_words.append(word)
arabic_text += f" {combined_form}"
verse_data = VerseData(
chapter=chapter,
verse=verse,
arabic_text=arabic_text.strip(),
translation=translation,
words=verse_words,
metadata={
"processed_timestamp": datetime.now().isoformat(),
"word_count": len(verse_words)
}
)
self._save_verse_data(verse_data)
return verse_data
except Exception as e:
logger.error(f"Error processing verse {chapter}:{verse}: {str(e)}")
logger.error(traceback.format_exc())
return None
def _save_verse_data(self, verse_data: VerseData):
"""Saves processed verse data as JSON and TXT."""
try:
verse_ref = f"{verse_data.chapter}:{verse_data.verse}"
json_path = os.path.join(self.output_dir, 'json', f'verse_{verse_ref.replace(":", "_")}.json')
with open(json_path, 'w', encoding='utf-8') as f:
json.dump(asdict(verse_data), f, ensure_ascii=False, indent=2)
txt_path = os.path.join(self.output_dir, 'txt', f'verse_{verse_ref.replace(":", "_")}.txt')
with open(txt_path, 'w', encoding='utf-8') as f:
f.write(f"=== Verse {verse_ref} ===\n\n")
f.write(f"Arabic Text:\n{verse_data.arabic_text}\n\n")
f.write(f"Translation:\n{verse_data.translation}\n\n")
f.write("Word Analysis:\n")
for i, word in enumerate(verse_data.words, 1):
f.write(f"\nWord {i}:\n")
f.write(f" Arabic: {word.arabic}\n")
f.write(f" Translation: {word.translation}\n")
f.write(f" Root: {word.root}\n")
f.write(" Features:\n")
for feature in word.features:
f.write(f" - {feature}\n")
f.write("\n")
logger.info(f"Saved verse data to {json_path} and {txt_path}")
except Exception as e:
logger.error(f"Error saving verse data: {str(e)}")
logger.error(traceback.format_exc())
class QuranicModelTrainer:
"""Trains the Gemma-2-2b model on Quranic data using chunked incremental updates."""
def __init__(self,
model_name: str = "google/gemma-2-2b",
processed_data_dir: str = "processed_data",
checkpoint_dir: str = "checkpoints"):
self.processed_data_dir = processed_data_dir
self.checkpoint_dir = checkpoint_dir
# Dynamically assign device based on GPU availability.
self.device = "cuda" if torch.cuda.is_available() else "cpu"
logger.info(f"Using device: {self.device}")
logger.info("Loading tokenizer and model...")
# Load tokenizer with additional special tokens and HF token from environment
self.tokenizer = AutoTokenizer.from_pretrained(
model_name,
token=os.environ.get("HF_TOKEN"),
additional_special_tokens=["[VERSE]", "[WORD]", "[ROOT]", "[FEATURES]"],
trust_remote_code=True
)
if self.tokenizer.pad_token is None:
self.tokenizer.add_special_tokens({"pad_token": "[PAD]"})
# Load model using eager attention for Gemma2 and low_cpu_mem_usage.
try:
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
token=os.environ.get("HF_TOKEN"),
torch_dtype=torch.float32,
low_cpu_mem_usage=True,
trust_remote_code=True,
attn_implementation="eager"
)
except Exception as e:
logger.error(f"Error loading model directly: {str(e)}")
logger.info("Attempting to load with fallback parameters...")
from transformers import AutoConfig
config = AutoConfig.from_pretrained(
model_name,
token=os.environ.get("HF_TOKEN"),
trust_remote_code=True
)
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
token=os.environ.get("HF_TOKEN"),
config=config,
torch_dtype=torch.float32,
low_cpu_mem_usage=True,
trust_remote_code=True,
revision="main",
attn_implementation="eager"
)
# Resize token embeddings to match tokenizer vocabulary size
self.model.resize_token_embeddings(len(self.tokenizer))
self.model.train()
self.model.config.use_cache = False
if hasattr(self.model, "gradient_checkpointing_enable"):
self.model.gradient_checkpointing_enable()
else:
logger.warning("Gradient checkpointing not available for this model")
def prepare_training_data(self, chapter_data: List[Dict]) -> Dataset:
"""Creates a QuranicDataset from processed chapter data."""
return QuranicDataset(chapter_data, self.tokenizer)
def train_chapter(self,
chapter_num: int,
processed_verses: List[Dict],
chunk_size: int = 5, # Reduced chunk size to help with memory
num_train_epochs: int = 5, # Lower epochs for testing
per_device_train_batch_size: int = 1,
learning_rate: float = 3e-5,
weight_decay: float = 0.01,
gradient_accumulation_steps: int = 32) -> bool:
"""
Splits chapter data into chunks and trains incrementally to reduce memory usage.
After each chunk, creates a downloadable zip of the checkpoint.
"""
try:
total_examples = len(processed_verses)
total_chunks = math.ceil(total_examples / chunk_size)
logger.info(f"Chapter {chapter_num}: {total_examples} examples, {total_chunks} chunks.")
for chunk_index in range(total_chunks):
chunk_data = processed_verses[chunk_index * chunk_size: (chunk_index + 1) * chunk_size]
dataset = self.prepare_training_data(chunk_data)
chunk_output_dir = os.path.join(self.checkpoint_dir, f"chapter_{chapter_num}", f"chunk_{chunk_index}")
os.makedirs(chunk_output_dir, exist_ok=True)
# Set use_cpu dynamically based on GPU availability.
training_args = TrainingArguments(
output_dir=chunk_output_dir,
overwrite_output_dir=True,
num_train_epochs=num_train_epochs,
per_device_train_batch_size=per_device_train_batch_size,
learning_rate=learning_rate,
weight_decay=weight_decay,
gradient_accumulation_steps=gradient_accumulation_steps,
fp16=False,
remove_unused_columns=False,
logging_steps=50,
report_to="none",
# Updated per deprecation: use eval_strategy instead of evaluation_strategy.
eval_strategy="no",
use_cpu=not torch.cuda.is_available(),
dataloader_num_workers=0,
dataloader_pin_memory=False
)
data_collator = DataCollatorForLanguageModeling(
tokenizer=self.tokenizer,
mlm=False
)
trainer = Trainer(
model=self.model,
args=training_args,
train_dataset=dataset,
processing_class=self.tokenizer, # Updated per deprecation notice.
data_collator=data_collator
)
logger.info(f"Training chunk {chunk_index+1}/{total_chunks} for Chapter {chapter_num}...")
trainer.train()
trainer.save_model(chunk_output_dir)
# Zip the checkpoint folder and generate a download link.
zip_filename = zip_checkpoint(chunk_output_dir)
base_url = os.environ.get("HF_SPACE_URL", "http://localhost")
download_link = f"{base_url}/file/{zip_filename}"
logger.info(f"Checkpoint chunk {chunk_index+1} download link: {download_link}")
# Save the download link into a text file within the checkpoint directory.
with open(os.path.join(chunk_output_dir, "download_link.txt"), "w") as f:
f.write(download_link)
del trainer, dataset
gc.collect()
manage_memory()
manage_gpu_resources()
logger.info(f"Completed training for Chapter {chapter_num}")
return True
except Exception as e:
logger.error(f"Error training chapter {chapter_num}: {str(e)}")
logger.error(traceback.format_exc())
return False
class QuranicPipeline:
"""Integrates data processing and incremental model training for all chapters."""
def __init__(self,
source_dir: str = ".",
working_dir: str = "working_directory",
start_chapter: int = 1,
end_chapter: int = 114):
self.source_dir = source_dir
self.working_dir = working_dir
self.start_chapter = start_chapter
self.end_chapter = end_chapter
self.setup_directories()
global logger
logger = logging.getLogger(__name__)
self.state = {
"last_processed_chapter": 0,
"last_trained_chapter": 0,
"current_state": "initialized",
"errors": [],
"start_time": datetime.now().isoformat()
}
self.load_state()
try:
logger.info("Initializing Quranic Data Processor...")
self.processor = QuranicDataProcessor(
source_dir=self.source_dir,
output_dir=os.path.join(self.working_dir, "processed_data")
)
logger.info("Initializing Quranic Model Trainer...")
self.trainer = QuranicModelTrainer(
model_name="google/gemma-2-2b",
processed_data_dir=os.path.join(self.working_dir, "processed_data"),
checkpoint_dir=os.path.join(self.working_dir, "checkpoints")
)
self.state["current_state"] = "ready"
self.save_state()
except Exception as e:
self.handle_error("Initialization failed", e)
raise
def setup_directories(self):
dirs = [
self.working_dir,
os.path.join(self.working_dir, "processed_data"),
os.path.join(self.working_dir, "checkpoints"),
os.path.join(self.working_dir, "logs"),
os.path.join(self.working_dir, "state")
]
for d in dirs:
os.makedirs(d, exist_ok=True)
def load_state(self):
state_file = os.path.join(self.working_dir, "state", "pipeline_state.json")
if os.path.exists(state_file):
try:
with open(state_file, 'r') as f:
saved_state = json.load(f)
self.state.update(saved_state)
logger.info(f"Loaded previous state: Last processed chapter {self.state.get('last_processed_chapter')}, "
f"last trained chapter {self.state.get('last_trained_chapter')}")
except Exception as e:
logger.warning(f"Could not load previous state: {str(e)}")
def save_state(self):
state_file = os.path.join(self.working_dir, "state", "pipeline_state.json")
with open(state_file, 'w') as f:
json.dump(self.state, f, indent=2)
def handle_error(self, context: str, error: Exception):
error_detail = {
"timestamp": datetime.now().isoformat(),
"context": context,
"error": str(error),
"traceback": traceback.format_exc()
}
self.state.setdefault("errors", []).append(error_detail)
logger.error(f"{context}: {str(error)}")
self.save_state()
def run_pipeline(self):
"""Runs processing and training for chapters sequentially, then saves the final model."""
logger.info("Starting pipeline execution")
try:
if not self.processor.load_source_files():
raise Exception("Failed to load source files")
for chapter in range(self.start_chapter, self.end_chapter + 1):
logger.info(f"=== Processing Chapter {chapter} ===")
processed_chapter_data = []
verse = 1
while True:
verse_data = self.processor.process_verse(chapter, verse)
if verse_data is None:
break
processed_chapter_data.append(asdict(verse_data))
verse += 1
if processed_chapter_data:
success = self.trainer.train_chapter(chapter, processed_chapter_data)
if not success:
logger.error(f"Training failed for Chapter {chapter}. Stopping pipeline.")
break
self.state["last_trained_chapter"] = chapter
self.save_state()
else:
logger.warning(f"No processed data for Chapter {chapter}")
self.state["last_processed_chapter"] = chapter
self.save_state()
manage_memory()
manage_gpu_resources()
logger.info("Pipeline execution completed")
# Save the final model and tokenizer after all training is complete.
final_model_dir = os.path.join(self.working_dir, "final_model")
os.makedirs(final_model_dir, exist_ok=True)
self.trainer.model.save_pretrained(final_model_dir)
self.trainer.tokenizer.save_pretrained(final_model_dir)
logger.info(f"Final model saved to {final_model_dir}")
except Exception as e:
self.handle_error("Pipeline execution failed", e)
raise
@spaces.GPU() # Request ZeroGPU hardware for the Space
def start_pipeline():
try:
logger.info("Starting Quranic Training Pipeline with Gemma-2-2b")
logger.info(f"PyTorch version: {torch.__version__}")
logger.info(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
logger.info(f"CUDA device count: {torch.cuda.device_count()}")
logger.info(f"CUDA device name: {torch.cuda.get_device_name(0)}")
if not os.environ.get("HF_TOKEN"):
logger.warning("HF_TOKEN environment variable not set. Model loading may fail.")
required_files = [
'quranic-corpus-morphology-0.4.txt',
'en.sample.quran-maududi.txt',
'en.w4w.qurandev.txt'
]
missing_files = [f for f in required_files if not os.path.exists(f)]
if missing_files:
return f"Missing required data files: {', '.join(missing_files)}"
pipeline = QuranicPipeline(
source_dir=".",
working_dir="working_directory",
start_chapter=1,
end_chapter=114
)
pipeline.run_pipeline()
return "Pipeline execution completed successfully."
except Exception as e:
error_msg = f"Pipeline execution failed: {str(e)}\n{traceback.format_exc()}"
logger.error(error_msg)
return error_msg
iface = gr.Interface(
fn=start_pipeline,
inputs=[],
outputs=gr.Textbox(label="Pipeline Status", lines=10),
title="Quranic Training Pipeline for Gemma-2-2b",
description="""This pipeline fine-tunes Google's Gemma-2-2b model on Quranic data.
Click 'Submit' to trigger the Quranic data processing and training pipeline on ZeroGPU.
Requirements:
- Transformers (>=4.42.0)
- Gradio (>=5.12.0)
- PyTorch (==2.2.2)
- psutil (==5.9.5)
- Accelerate (>=0.26.0)
The pipeline processes all 114 chapters of the Quran sequentially, with memory and GPU resource management optimizations for dynamic ZeroGPU environments.
Checkpoint download links are provided after every training chunk."""
)
if __name__ == "__main__":
iface.launch()