Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -29,31 +29,27 @@ if not os.path.exists(MODEL_PATH):
|
|
| 29 |
print(f"❌ Erreur téléchargement: {e}")
|
| 30 |
|
| 31 |
# -------------------------
|
| 32 |
-
# CONFIGURATION LLAMA.CPP OPTIMISÉE
|
| 33 |
# -------------------------
|
| 34 |
os.environ["LLAMA_CPP_LOG_LEVEL"] = "OFF"
|
| 35 |
warnings.filterwarnings("ignore")
|
| 36 |
|
| 37 |
-
print("⚡ Chargement du modèle avec llama.cpp
|
| 38 |
|
| 39 |
# Détection automatique du nombre de threads
|
| 40 |
import multiprocessing
|
| 41 |
cpu_count = multiprocessing.cpu_count()
|
| 42 |
-
n_threads = max(2, cpu_count - 1)
|
| 43 |
|
| 44 |
llm = Llama(
|
| 45 |
model_path=MODEL_PATH,
|
| 46 |
-
n_ctx=2048,
|
| 47 |
-
n_gpu_layers=0, #
|
| 48 |
-
n_threads=n_threads,
|
| 49 |
-
n_batch=512, # Batch adapté pour CPU
|
| 50 |
-
n_threads_batch=n_threads, # Même nombre de threads pour le batch
|
| 51 |
-
use_mlock=False, # Désactivé pour meilleures performances
|
| 52 |
-
vocab_only=False,
|
| 53 |
verbose=False
|
| 54 |
)
|
| 55 |
|
| 56 |
-
print(f"✅ Modèle chargé! Threads: {n_threads}
|
| 57 |
|
| 58 |
# -------------------------
|
| 59 |
# ÉTAT & SYNCHRONISATION
|
|
@@ -61,6 +57,7 @@ print(f"✅ Modèle chargé! Threads: {n_threads} | CPU: {cpu_count} cores")
|
|
| 61 |
lock = threading.Lock()
|
| 62 |
conversations = {"Conversation 1": []}
|
| 63 |
stop_generation = threading.Event()
|
|
|
|
| 64 |
|
| 65 |
# -------------------------
|
| 66 |
# FONCTIONS UTILITAIRES OPTIMISÉES
|
|
@@ -72,32 +69,28 @@ def get_conv_names():
|
|
| 72 |
with lock:
|
| 73 |
return list(conversations.keys())
|
| 74 |
|
| 75 |
-
# Cache pour éviter la reconstruction complète du prompt
|
| 76 |
-
prompt_cache = {}
|
| 77 |
def build_conversation_prompt(history, new_message):
|
| 78 |
-
"""Format de prompt
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
if cache_key in prompt_cache:
|
| 82 |
-
return prompt_cache[cache_key]
|
| 83 |
|
| 84 |
prompt = ""
|
| 85 |
|
| 86 |
-
# System prompt
|
| 87 |
-
if not
|
| 88 |
-
prompt += """
|
|
|
|
|
|
|
| 89 |
|
| 90 |
"""
|
|
|
|
| 91 |
|
| 92 |
-
#
|
| 93 |
-
|
| 94 |
-
for user_msg, assistant_msg in recent_history:
|
| 95 |
prompt += f"### Instruction:\n{user_msg}\n\n### Response:\n{assistant_msg}\n\n"
|
| 96 |
|
| 97 |
-
#
|
| 98 |
prompt += f"### Instruction:\n{new_message}\n\n### Response:\n"
|
| 99 |
|
| 100 |
-
prompt_cache[cache_key] = prompt
|
| 101 |
return prompt
|
| 102 |
|
| 103 |
def send_message_stream(user_message, displayed_history, current_chat_name):
|
|
@@ -121,24 +114,16 @@ def send_message_stream(user_message, displayed_history, current_chat_name):
|
|
| 121 |
formatted_prompt = build_conversation_prompt(local_hist[:-1], str(user_message))
|
| 122 |
partial = ""
|
| 123 |
|
| 124 |
-
# PARAMÈTRES OPTIMISÉS POUR CPU
|
| 125 |
-
last_update = time.time()
|
| 126 |
-
token_count = 0
|
| 127 |
-
min_tokens = 3 # Regroupement modéré pour CPU
|
| 128 |
-
max_delay = 0.3 # 300ms entre updates pour CPU
|
| 129 |
-
buffer = ""
|
| 130 |
-
|
| 131 |
try:
|
|
|
|
| 132 |
stream = llm.create_completion(
|
| 133 |
prompt=formatted_prompt,
|
| 134 |
stream=True,
|
| 135 |
-
max_tokens=
|
| 136 |
temperature=0.7,
|
| 137 |
-
top_p=0.
|
| 138 |
-
repeat_penalty=1.
|
| 139 |
-
stop=["### Instruction:", "### Response:", "
|
| 140 |
-
min_p=0.05, # Acceleration CPU
|
| 141 |
-
typical_p=0.95 # Acceleration CPU
|
| 142 |
)
|
| 143 |
|
| 144 |
for chunk in stream:
|
|
@@ -148,35 +133,11 @@ def send_message_stream(user_message, displayed_history, current_chat_name):
|
|
| 148 |
if "choices" in chunk and chunk["choices"]:
|
| 149 |
token = chunk["choices"][0].get("text", "")
|
| 150 |
if token:
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
time_since_update = current_time - last_update
|
| 157 |
-
|
| 158 |
-
should_update = (
|
| 159 |
-
token_count >= min_tokens or
|
| 160 |
-
time_since_update > max_delay or
|
| 161 |
-
token in [".", "!", "?", "\n", " "]
|
| 162 |
-
)
|
| 163 |
-
|
| 164 |
-
if should_update and buffer.strip():
|
| 165 |
-
partial += buffer
|
| 166 |
-
cleaned = clean_output(partial)
|
| 167 |
-
local_hist[-1] = (str(user_message), cleaned)
|
| 168 |
-
yield local_hist, ""
|
| 169 |
-
last_update = current_time
|
| 170 |
-
token_count = 0
|
| 171 |
-
buffer = ""
|
| 172 |
-
|
| 173 |
-
# Dernier flush du buffer
|
| 174 |
-
if buffer:
|
| 175 |
-
partial += buffer
|
| 176 |
-
if partial:
|
| 177 |
-
cleaned = clean_output(partial)
|
| 178 |
-
local_hist[-1] = (str(user_message), cleaned)
|
| 179 |
-
yield local_hist, ""
|
| 180 |
|
| 181 |
except Exception as e:
|
| 182 |
err_text = f"[Erreur: {e}]"
|
|
@@ -185,8 +146,7 @@ def send_message_stream(user_message, displayed_history, current_chat_name):
|
|
| 185 |
|
| 186 |
finally:
|
| 187 |
end_time = time.time()
|
| 188 |
-
|
| 189 |
-
print(f"⏱️ Temps de génération: {generation_time:.2f}s - {len(partial)} caractères")
|
| 190 |
with lock:
|
| 191 |
conversations[current_chat_name] = local_hist.copy()
|
| 192 |
yield local_hist, ""
|
|
@@ -215,12 +175,14 @@ def request_stop():
|
|
| 215 |
return "🛑 Arrêt demandé..."
|
| 216 |
|
| 217 |
def clear_chat():
|
|
|
|
| 218 |
with lock:
|
| 219 |
conversations["Conversation 1"] = []
|
|
|
|
| 220 |
return [], "Conversation 1"
|
| 221 |
|
| 222 |
# -------------------------
|
| 223 |
-
# INTERFACE GRADIO OPTIMISÉE
|
| 224 |
# -------------------------
|
| 225 |
css = """
|
| 226 |
:root {
|
|
@@ -401,26 +363,17 @@ css = """
|
|
| 401 |
background: #1e293b;
|
| 402 |
border-radius: 8px;
|
| 403 |
}
|
| 404 |
-
|
| 405 |
-
.cpu-warning {
|
| 406 |
-
color: #fbbf24;
|
| 407 |
-
background: #431407;
|
| 408 |
-
padding: 8px;
|
| 409 |
-
border-radius: 8px;
|
| 410 |
-
margin-top: 10px;
|
| 411 |
-
font-size: 12px;
|
| 412 |
-
}
|
| 413 |
"""
|
| 414 |
|
| 415 |
-
with gr.Blocks(css=css, title="Alisia Chat -
|
| 416 |
history_visible = gr.State(True)
|
| 417 |
current_chat = gr.State("Conversation 1")
|
| 418 |
|
| 419 |
with gr.Row(elem_id="topbar"):
|
| 420 |
menu_btn = gr.Button("☰", elem_classes="hamburger")
|
| 421 |
-
gr.Markdown("### 💬 Alisia <span class='alisia-badge'>
|
| 422 |
gr.HTML("<div style='flex:1'></div>")
|
| 423 |
-
gr.Markdown(f"<small style='color:#94a3b8'>CPU: {
|
| 424 |
|
| 425 |
with gr.Row():
|
| 426 |
with gr.Column(scale=1, visible=True, elem_id="leftcol") as left_column:
|
|
@@ -442,21 +395,12 @@ with gr.Blocks(css=css, title="Alisia Chat - Optimisé CPU", theme=gr.themes.Sof
|
|
| 442 |
elem_classes="clear-btn"
|
| 443 |
)
|
| 444 |
|
| 445 |
-
# Informations de performance CPU
|
| 446 |
gr.Markdown("""
|
| 447 |
<div class="perf-info">
|
| 448 |
-
<strong>
|
| 449 |
-
•
|
| 450 |
-
•
|
| 451 |
-
•
|
| 452 |
-
• Réponses: 384 tokens max
|
| 453 |
-
</div>
|
| 454 |
-
""".format(n_threads=n_threads, cpu_count=cpu_count))
|
| 455 |
-
|
| 456 |
-
gr.Markdown("""
|
| 457 |
-
<div class="cpu-warning">
|
| 458 |
-
⚠️ Mode CPU - Les performances peuvent varier<br>
|
| 459 |
-
selon la puissance de votre processeur
|
| 460 |
</div>
|
| 461 |
""")
|
| 462 |
|
|
@@ -556,13 +500,9 @@ with gr.Blocks(css=css, title="Alisia Chat - Optimisé CPU", theme=gr.themes.Sof
|
|
| 556 |
# LANCEMENT
|
| 557 |
# -------------------------
|
| 558 |
if __name__ == "__main__":
|
| 559 |
-
print("🚀 Lancement de l'interface optimisée
|
| 560 |
-
print(
|
| 561 |
-
print(f"
|
| 562 |
-
print(f" - Threads utilisés: {n_threads}")
|
| 563 |
-
print(f" - Contexte: 2048 tokens")
|
| 564 |
-
print(f" - Réponses limitées: 384 tokens")
|
| 565 |
-
print("⏱️ Patience - Le CPU peut être plus lent que le GPU")
|
| 566 |
|
| 567 |
demo.launch(
|
| 568 |
share=True,
|
|
|
|
| 29 |
print(f"❌ Erreur téléchargement: {e}")
|
| 30 |
|
| 31 |
# -------------------------
|
| 32 |
+
# CONFIGURATION LLAMA.CPP OPTIMISÉE
|
| 33 |
# -------------------------
|
| 34 |
os.environ["LLAMA_CPP_LOG_LEVEL"] = "OFF"
|
| 35 |
warnings.filterwarnings("ignore")
|
| 36 |
|
| 37 |
+
print("⚡ Chargement du modèle avec llama.cpp...")
|
| 38 |
|
| 39 |
# Détection automatique du nombre de threads
|
| 40 |
import multiprocessing
|
| 41 |
cpu_count = multiprocessing.cpu_count()
|
| 42 |
+
n_threads = max(2, cpu_count - 1)
|
| 43 |
|
| 44 |
llm = Llama(
|
| 45 |
model_path=MODEL_PATH,
|
| 46 |
+
n_ctx=2048,
|
| 47 |
+
n_gpu_layers=0, # CPU uniquement
|
| 48 |
+
n_threads=n_threads,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
verbose=False
|
| 50 |
)
|
| 51 |
|
| 52 |
+
print(f"✅ Modèle chargé! Threads: {n_threads}")
|
| 53 |
|
| 54 |
# -------------------------
|
| 55 |
# ÉTAT & SYNCHRONISATION
|
|
|
|
| 57 |
lock = threading.Lock()
|
| 58 |
conversations = {"Conversation 1": []}
|
| 59 |
stop_generation = threading.Event()
|
| 60 |
+
system_prompt_used = False # Pour suivre si le system prompt a été utilisé
|
| 61 |
|
| 62 |
# -------------------------
|
| 63 |
# FONCTIONS UTILITAIRES OPTIMISÉES
|
|
|
|
| 69 |
with lock:
|
| 70 |
return list(conversations.keys())
|
| 71 |
|
|
|
|
|
|
|
| 72 |
def build_conversation_prompt(history, new_message):
|
| 73 |
+
"""Format de prompt Alpaca avec system prompt UNIQUEMENT au début"""
|
| 74 |
+
global system_prompt_used
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
prompt = ""
|
| 77 |
|
| 78 |
+
# System prompt UNIQUEMENT si jamais utilisé auparavant
|
| 79 |
+
if not system_prompt_used:
|
| 80 |
+
prompt += """Your name is Alisia, you are created by the Alisia research team.
|
| 81 |
+
Below is an instruction that describes a task, paired with an input that provides further context.
|
| 82 |
+
Write a response that appropriately completes the request.
|
| 83 |
|
| 84 |
"""
|
| 85 |
+
system_prompt_used = True
|
| 86 |
|
| 87 |
+
# Ajouter tout l'historique de conversation
|
| 88 |
+
for user_msg, assistant_msg in history:
|
|
|
|
| 89 |
prompt += f"### Instruction:\n{user_msg}\n\n### Response:\n{assistant_msg}\n\n"
|
| 90 |
|
| 91 |
+
# Ajouter le nouveau message
|
| 92 |
prompt += f"### Instruction:\n{new_message}\n\n### Response:\n"
|
| 93 |
|
|
|
|
| 94 |
return prompt
|
| 95 |
|
| 96 |
def send_message_stream(user_message, displayed_history, current_chat_name):
|
|
|
|
| 114 |
formatted_prompt = build_conversation_prompt(local_hist[:-1], str(user_message))
|
| 115 |
partial = ""
|
| 116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
try:
|
| 118 |
+
# Utilisation directe du streaming sans buffering complexe
|
| 119 |
stream = llm.create_completion(
|
| 120 |
prompt=formatted_prompt,
|
| 121 |
stream=True,
|
| 122 |
+
max_tokens=1024,
|
| 123 |
temperature=0.7,
|
| 124 |
+
top_p=0.8,
|
| 125 |
+
repeat_penalty=1.05,
|
| 126 |
+
stop=["### Instruction:", "### Response:", "<|endoftext|>", "\n\n\n"]
|
|
|
|
|
|
|
| 127 |
)
|
| 128 |
|
| 129 |
for chunk in stream:
|
|
|
|
| 133 |
if "choices" in chunk and chunk["choices"]:
|
| 134 |
token = chunk["choices"][0].get("text", "")
|
| 135 |
if token:
|
| 136 |
+
partial += token
|
| 137 |
+
# Mise à jour immédiate pour une meilleure réactivité
|
| 138 |
+
cleaned = clean_output(partial)
|
| 139 |
+
local_hist[-1] = (str(user_message), cleaned)
|
| 140 |
+
yield local_hist, ""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
except Exception as e:
|
| 143 |
err_text = f"[Erreur: {e}]"
|
|
|
|
| 146 |
|
| 147 |
finally:
|
| 148 |
end_time = time.time()
|
| 149 |
+
print(f"⏱️ Génération: {end_time - start_time:.2f}s - {len(partial)} chars")
|
|
|
|
| 150 |
with lock:
|
| 151 |
conversations[current_chat_name] = local_hist.copy()
|
| 152 |
yield local_hist, ""
|
|
|
|
| 175 |
return "🛑 Arrêt demandé..."
|
| 176 |
|
| 177 |
def clear_chat():
|
| 178 |
+
global system_prompt_used
|
| 179 |
with lock:
|
| 180 |
conversations["Conversation 1"] = []
|
| 181 |
+
system_prompt_used = False # Réinitialiser pour le prochain chat
|
| 182 |
return [], "Conversation 1"
|
| 183 |
|
| 184 |
# -------------------------
|
| 185 |
+
# INTERFACE GRADIO OPTIMISÉE
|
| 186 |
# -------------------------
|
| 187 |
css = """
|
| 188 |
:root {
|
|
|
|
| 363 |
background: #1e293b;
|
| 364 |
border-radius: 8px;
|
| 365 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 366 |
"""
|
| 367 |
|
| 368 |
+
with gr.Blocks(css=css, title="Alisia Chat - Ultra Rapide", theme=gr.themes.Soft()) as demo:
|
| 369 |
history_visible = gr.State(True)
|
| 370 |
current_chat = gr.State("Conversation 1")
|
| 371 |
|
| 372 |
with gr.Row(elem_id="topbar"):
|
| 373 |
menu_btn = gr.Button("☰", elem_classes="hamburger")
|
| 374 |
+
gr.Markdown("### 💬 Alisia <span class='alisia-badge'>AI Assistant</span>", elem_id="title")
|
| 375 |
gr.HTML("<div style='flex:1'></div>")
|
| 376 |
+
gr.Markdown(f"<small style='color:#94a3b8'>CPU: {n_threads} threads • Mode Rapide</small>")
|
| 377 |
|
| 378 |
with gr.Row():
|
| 379 |
with gr.Column(scale=1, visible=True, elem_id="leftcol") as left_column:
|
|
|
|
| 395 |
elem_classes="clear-btn"
|
| 396 |
)
|
| 397 |
|
|
|
|
| 398 |
gr.Markdown("""
|
| 399 |
<div class="perf-info">
|
| 400 |
+
<strong>🚀 Mode Alpaca Optimisé</strong><br>
|
| 401 |
+
• System prompt unique<br>
|
| 402 |
+
• Streaming direct<br>
|
| 403 |
+
• Format Alpaca pur
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 404 |
</div>
|
| 405 |
""")
|
| 406 |
|
|
|
|
| 500 |
# LANCEMENT
|
| 501 |
# -------------------------
|
| 502 |
if __name__ == "__main__":
|
| 503 |
+
print("🚀 Lancement de l'interface optimisée...")
|
| 504 |
+
print("📋 Format Alpaca avec system prompt unique")
|
| 505 |
+
print(f"⚡ Threads CPU: {n_threads}")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 506 |
|
| 507 |
demo.launch(
|
| 508 |
share=True,
|