optimisation
This commit is contained in:
Alex
parent
aee2716a41
commit
70e4932cd0
@ -5,7 +5,7 @@ from transformers import (
|
||||
AutoTokenizer,
|
||||
AutoModelForCausalLM,
|
||||
TrainingArguments,
|
||||
BitsAndBytesConfig
|
||||
BitsAndBytesConfig,
|
||||
)
|
||||
from peft import (
|
||||
LoraConfig,
|
||||
@ -46,17 +46,27 @@ print(f"Pad token id: {tokenizer.pad_token_id}")
|
||||
print(f"Max sequence length: {tokenizer.model_max_length}")
|
||||
|
||||
# ----------------------------
|
||||
# [2/7] Model loading (QLoRA)
|
||||
# [2/7] Quantization config (QLoRA)
|
||||
# ----------------------------
|
||||
print(f"{80 * '_'}\n[2/7] Loading model in 4-bit mode (QLoRA)...")
|
||||
print(f"{80 * '_'}\n[2/7] Configuring 4-bit quantization (BitsAndBytes)...")
|
||||
bnb_config = BitsAndBytesConfig(
|
||||
load_in_4bit=True,
|
||||
bnb_4bit_quant_type="nf4",
|
||||
bnb_4bit_compute_dtype=torch.float16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
)
|
||||
|
||||
print("4-bit NF4 quantization configured.")
|
||||
|
||||
print("Loading model...")
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
MODEL_NAME,
|
||||
load_in_4bit=True,
|
||||
device_map="auto",
|
||||
quantization_config=bnb_config,
|
||||
dtype=torch.float16,
|
||||
trust_remote_code=True,
|
||||
)
|
||||
print("Model loaded.")
|
||||
print("Model loaded successfully.")
|
||||
|
||||
# ----------------------------
|
||||
# [3/7] Prepare model for k-bit training
|
||||
@ -82,10 +92,14 @@ lora_config = LoraConfig(
|
||||
bias="none",
|
||||
task_type="CAUSAL_LM",
|
||||
target_modules=[
|
||||
"q_proj", "k_proj", "v_proj", "o_proj",
|
||||
"gate_proj", "up_proj", "down_proj"
|
||||
"q_proj",
|
||||
"k_proj",
|
||||
"v_proj",
|
||||
"o_proj",
|
||||
"gate_proj",
|
||||
"up_proj",
|
||||
"down_proj",
|
||||
],
|
||||
|
||||
)
|
||||
|
||||
model = get_peft_model(model, lora_config)
|
||||
@ -97,10 +111,7 @@ print("LoRA adapters successfully attached.")
|
||||
# [5/7] Dataset loading & formatting
|
||||
# ----------------------------
|
||||
print(f"{80 * '_'}\n[5/7] Loading dataset from JSON file...")
|
||||
dataset = load_dataset(
|
||||
"json",
|
||||
data_files=DATA_FILE
|
||||
)
|
||||
dataset = load_dataset("json", data_files=DATA_FILE)
|
||||
|
||||
print(f"Dataset loaded with {len(dataset['train'])} samples.")
|
||||
|
||||
@ -123,7 +134,8 @@ dataset = dataset.map(
|
||||
)
|
||||
|
||||
print("Dataset formatting completed.")
|
||||
print(f"Example prompt:\n{dataset['train'][0]['text']}")
|
||||
print("Example prompt:\n")
|
||||
print(dataset["train"][0]["text"])
|
||||
|
||||
# ----------------------------
|
||||
# [6/7] Training arguments
|
||||
@ -147,7 +159,10 @@ training_args = TrainingArguments(
|
||||
print("Training arguments ready.")
|
||||
print(f"Output directory: {OUTPUT_DIR}")
|
||||
print(f"Epochs: {training_args.num_train_epochs}")
|
||||
print(f"Effective batch size: {training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps}")
|
||||
print(
|
||||
f"Effective batch size: "
|
||||
f"{training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps}"
|
||||
)
|
||||
|
||||
# ----------------------------
|
||||
# Trainer
|
||||
|
||||
@ -3,23 +3,23 @@ import requests
|
||||
import json
|
||||
from reportlab.lib.pagesizes import letter
|
||||
from reportlab.lib.units import inch
|
||||
from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer, Flowable
|
||||
from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer
|
||||
from reportlab.lib.styles import getSampleStyleSheet, ParagraphStyle
|
||||
from reportlab.lib.enums import TA_JUSTIFY
|
||||
from reportlab.pdfbase import pdfmetrics
|
||||
from reportlab.pdfbase.ttfonts import TTFont
|
||||
import os
|
||||
import os, time
|
||||
|
||||
# Configuration
|
||||
DEBUG = True
|
||||
PDF_PATH = "Traduction\TaniaBorecMemoir(Ukr).pdf"
|
||||
PDF_PATH = "Traduction/TaniaBorecMemoir(Ukr).pdf"
|
||||
OLLAMA_MODEL = "traductionUkrainienVersFrancais:latest"
|
||||
OLLAMA_URL = "http://localhost:11434/api/generate"
|
||||
TARGET_LANGUAGE = "français"
|
||||
CHECKPOINT_FILE = "Traduction\checkpoint.json"
|
||||
TEMP_OUTPUT_TXT = "Traduction\output_temp.txt"
|
||||
FINAL_OUTPUT_PDF = PDF_PATH.replace(".pdf",f"({TARGET_LANGUAGE.upper()[:2]})_V8.pdf")
|
||||
FINAL_OUTPUT_TXT = PDF_PATH.replace(".pdf",f"({TARGET_LANGUAGE.upper()[:2]})_V8.txt")
|
||||
CHECKPOINT_FILE = "Traduction/checkpoint.json"
|
||||
TEMP_OUTPUT_TXT = "Traduction/output_temp.txt"
|
||||
FINAL_OUTPUT_PDF = PDF_PATH.replace(".pdf",f"({TARGET_LANGUAGE.upper()[:2]})_V9.pdf")
|
||||
FINAL_OUTPUT_TXT = PDF_PATH.replace(".pdf",f"({TARGET_LANGUAGE.upper()[:2]})_V9.txt")
|
||||
|
||||
DEBUG = True
|
||||
|
||||
@ -341,6 +341,7 @@ def main():
|
||||
print(f"Batches manquants détectés : {missing_batches}")
|
||||
|
||||
# Traduction des paragraphes manquants
|
||||
temps_cumule = 0.0
|
||||
for i in missing_batches:
|
||||
batch = paragraphs[i:i + batch_size]
|
||||
paragraph_cumul = "\n".join(batch)
|
||||
@ -348,13 +349,24 @@ def main():
|
||||
print(f"{15 * '-'} Traduction des paragraphes manquants {i+1} à {min(i + batch_size, len(paragraphs))} / {len(paragraphs)}")
|
||||
|
||||
try:
|
||||
debut_chrono = time.time()
|
||||
result = send_to_ollama(paragraph_cumul)
|
||||
fin_chrono = time.time()
|
||||
temps_paragraphe = fin_chrono - debut_chrono
|
||||
temps_cumule += temps_paragraphe
|
||||
|
||||
# Conversion en minutes et secondes
|
||||
minutes_paragraphe, secondes_paragraphe = divmod(temps_paragraphe, 60)
|
||||
minutes_cumule, secondes_cumule = divmod(temps_cumule, 60)
|
||||
|
||||
print(f"{result}")
|
||||
results[str(i)] = result
|
||||
save_checkpoint(len(paragraphs), results) # Met à jour le dernier indice du batch
|
||||
save_temp_results(results)
|
||||
except Exception as e:
|
||||
print(f"Erreur lors de la traduction du paragraphe {i}: {e}")
|
||||
print(f" Temps de traduction : {int(minutes_paragraphe)} min {secondes_paragraphe:.2f} sec")
|
||||
print(f" Temps cumulé : {int(minutes_cumule)} min {secondes_cumule:.2f} sec")
|
||||
|
||||
# Traitement des paragraphes suivants
|
||||
for i in range(last_index + 1, len(paragraphs), batch_size):
|
||||
@ -364,7 +376,16 @@ def main():
|
||||
print(f"{15 * '-'} Traduction des paragraphes {i+1} à {min(i + batch_size, len(paragraphs))} / {len(paragraphs)}")
|
||||
|
||||
try:
|
||||
debut_chrono = time.time()
|
||||
result = send_to_ollama(paragraph_cumul)
|
||||
fin_chrono = time.time()
|
||||
temps_paragraphe = fin_chrono - debut_chrono
|
||||
temps_cumule += temps_paragraphe
|
||||
|
||||
# Conversion en minutes et secondes
|
||||
minutes_paragraphe, secondes_paragraphe = divmod(temps_paragraphe, 60)
|
||||
minutes_cumule, secondes_cumule = divmod(temps_cumule, 60)
|
||||
|
||||
print(f"{result}")
|
||||
results[str(i)] = result
|
||||
save_checkpoint(i + batch_size - 1, results)
|
||||
@ -372,6 +393,9 @@ def main():
|
||||
except Exception as e:
|
||||
print(f"Erreur : {e}")
|
||||
continue
|
||||
print(f" Temps de traduction : {int(minutes_paragraphe)} min {secondes_paragraphe:.2f} sec")
|
||||
print(f" Temps cumulé : {int(minutes_cumule)} min {secondes_cumule:.2f} sec")
|
||||
|
||||
|
||||
save_temp_results(results)
|
||||
create_pdf_from_results(results, FINAL_OUTPUT_PDF)
|
||||
|
||||
2
run.bat
2
run.bat
@ -17,6 +17,7 @@ REM Activer l'environnement virtuel Python
|
||||
call %VENV_PATH%\Scripts\activate.bat
|
||||
|
||||
REM Lancer la compilation du modèle LLM pour Ollama
|
||||
echo Compilation du modèle LLM pour Ollama
|
||||
ollama create traductionUkrainienVersFrancais -f .\Traduction\Modelfile
|
||||
|
||||
:: 1. Vérifie si le processus ollama.exe est en cours d'exécution
|
||||
@ -39,6 +40,7 @@ if %ERRORLEVEL% neq 0 (
|
||||
)
|
||||
|
||||
REM Exécuter le script principal
|
||||
echo Lancement du script principal de traduction
|
||||
python %MAIN_SCRIPT_PATH%
|
||||
|
||||
endlocal
|
||||
Loading…
x
Reference in New Issue
Block a user