Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,16 +3,16 @@ import torch
|
|
| 3 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 4 |
import os
|
| 5 |
|
| 6 |
-
# 🔹
|
| 7 |
hf_token = os.environ.get("HF_TOKEN")
|
| 8 |
-
|
| 9 |
if hf_token is None:
|
| 10 |
-
raise ValueError("⚠️ Le token Hugging Face (HF_TOKEN)
|
| 11 |
-
"
|
| 12 |
|
| 13 |
-
# 🔹
|
| 14 |
model_id = "SafaaAI/final_llm_darija_fr_tech"
|
| 15 |
|
|
|
|
| 16 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 17 |
model_id,
|
| 18 |
token=hf_token,
|
|
@@ -23,56 +23,51 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
| 23 |
model_id,
|
| 24 |
token=hf_token,
|
| 25 |
trust_remote_code=True,
|
| 26 |
-
device_map=
|
| 27 |
-
)
|
|
|
|
|
|
|
| 28 |
|
| 29 |
-
# 🔹 Fonction d
|
| 30 |
def chat_with_model(message, history):
|
| 31 |
history = history or []
|
| 32 |
-
full_prompt =
|
| 33 |
-
|
| 34 |
-
|
|
|
|
|
|
|
| 35 |
for user_message, bot_message in history:
|
| 36 |
-
full_prompt += f"
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
inputs = tokenizer(full_prompt, return_tensors="pt").to(model.device)
|
| 43 |
-
|
| 44 |
-
# 📝 Extraire explicitement input_ids et attention_mask
|
| 45 |
-
input_ids = inputs["input_ids"]
|
| 46 |
-
attention_mask = inputs["attention_mask"]
|
| 47 |
-
|
| 48 |
-
# Générer la réponse
|
| 49 |
with torch.no_grad():
|
| 50 |
output_ids = model.generate(
|
| 51 |
-
input_ids,
|
| 52 |
-
attention_mask=attention_mask,
|
| 53 |
-
max_new_tokens=
|
| 54 |
do_sample=True,
|
| 55 |
top_p=0.9,
|
| 56 |
-
temperature=0.7
|
|
|
|
| 57 |
)
|
| 58 |
|
| 59 |
-
# Décoder la sortie
|
| 60 |
response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
response = response[response_start_index + len("ASSISTANT:"):].strip()
|
| 66 |
-
|
| 67 |
history.append((message, response))
|
| 68 |
return history, history
|
| 69 |
|
| 70 |
# 🔹 Interface Gradio
|
| 71 |
with gr.Blocks() as demo:
|
| 72 |
-
gr.Markdown("##
|
| 73 |
-
|
| 74 |
-
chatbot = gr.Chatbot()
|
| 75 |
-
msg = gr.Textbox(label="Écris ton message ici")
|
| 76 |
clear = gr.Button("🧹 Effacer la conversation")
|
| 77 |
|
| 78 |
state = gr.State([])
|
|
@@ -80,6 +75,5 @@ with gr.Blocks() as demo:
|
|
| 80 |
msg.submit(chat_with_model, [msg, state], [chatbot, state])
|
| 81 |
clear.click(lambda: ([], []), None, [chatbot, state])
|
| 82 |
|
| 83 |
-
# 🔹 Lancer l'application
|
| 84 |
if __name__ == "__main__":
|
| 85 |
-
demo.launch()
|
|
|
|
| 3 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 4 |
import os
|
| 5 |
|
| 6 |
+
# 🔹 Récupération du token Hugging Face
|
| 7 |
hf_token = os.environ.get("HF_TOKEN")
|
|
|
|
| 8 |
if hf_token is None:
|
| 9 |
+
raise ValueError("⚠️ Le token Hugging Face (HF_TOKEN) est manquant. "
|
| 10 |
+
"Ajoute-le dans les secrets de ton Space.")
|
| 11 |
|
| 12 |
+
# 🔹 Identifiant du modèle
|
| 13 |
model_id = "SafaaAI/final_llm_darija_fr_tech"
|
| 14 |
|
| 15 |
+
# 🔹 Charger tokenizer et modèle en CPU
|
| 16 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 17 |
model_id,
|
| 18 |
token=hf_token,
|
|
|
|
| 23 |
model_id,
|
| 24 |
token=hf_token,
|
| 25 |
trust_remote_code=True,
|
| 26 |
+
device_map=None # pas d’auto GPU
|
| 27 |
+
).to("cpu") # forcer CPU
|
| 28 |
+
|
| 29 |
+
print("✅ Modèle chargé sur CPU")
|
| 30 |
|
| 31 |
+
# 🔹 Fonction d’inférence
|
| 32 |
def chat_with_model(message, history):
|
| 33 |
history = history or []
|
| 34 |
+
full_prompt = (
|
| 35 |
+
"A chat between a curious user and an AI assistant capable of "
|
| 36 |
+
"understanding Darija, French, and technical language.\n"
|
| 37 |
+
)
|
| 38 |
+
|
| 39 |
for user_message, bot_message in history:
|
| 40 |
+
full_prompt += f"USER: {user_message}\nASSISTANT: {bot_message}\n"
|
| 41 |
+
|
| 42 |
+
full_prompt += f"USER: {message}\nASSISTANT:"
|
| 43 |
+
|
| 44 |
+
inputs = tokenizer(full_prompt, return_tensors="pt")
|
| 45 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
with torch.no_grad():
|
| 47 |
output_ids = model.generate(
|
| 48 |
+
inputs["input_ids"],
|
| 49 |
+
attention_mask=inputs["attention_mask"],
|
| 50 |
+
max_new_tokens=100,
|
| 51 |
do_sample=True,
|
| 52 |
top_p=0.9,
|
| 53 |
+
temperature=0.7,
|
| 54 |
+
pad_token_id=tokenizer.eos_token_id
|
| 55 |
)
|
| 56 |
|
|
|
|
| 57 |
response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 58 |
+
|
| 59 |
+
if "ASSISTANT:" in response:
|
| 60 |
+
response = response.split("ASSISTANT:")[-1].strip()
|
| 61 |
+
|
|
|
|
|
|
|
| 62 |
history.append((message, response))
|
| 63 |
return history, history
|
| 64 |
|
| 65 |
# 🔹 Interface Gradio
|
| 66 |
with gr.Blocks() as demo:
|
| 67 |
+
gr.Markdown("## 🤖 Chatbot SafaaAI - LLM (Darija + Français + Technique)")
|
| 68 |
+
|
| 69 |
+
chatbot = gr.Chatbot(height=400)
|
| 70 |
+
msg = gr.Textbox(label="💬 Écris ton message ici", placeholder="Pose ta question...")
|
| 71 |
clear = gr.Button("🧹 Effacer la conversation")
|
| 72 |
|
| 73 |
state = gr.State([])
|
|
|
|
| 75 |
msg.submit(chat_with_model, [msg, state], [chatbot, state])
|
| 76 |
clear.click(lambda: ([], []), None, [chatbot, state])
|
| 77 |
|
|
|
|
| 78 |
if __name__ == "__main__":
|
| 79 |
+
demo.launch()
|