Update app.py
Browse files
app.py
CHANGED
|
@@ -14,12 +14,12 @@ import spaces
|
|
| 14 |
model = None
|
| 15 |
status_message = "Modello non ancora caricato"
|
| 16 |
MODEL_PATH = "TheBloke/Mistral-7B-Instruct-v0.2-GGUF"
|
| 17 |
-
MODEL_FILE = "mistral-7b-instruct-v0.2.Q4_K_M.gguf"
|
| 18 |
MODEL_TYPE = "mistral"
|
| 19 |
MAX_NEW_TOKENS = 2048
|
| 20 |
-
MODEL_LOCK = Lock()
|
| 21 |
|
| 22 |
-
#
|
| 23 |
class Message(BaseModel):
|
| 24 |
role: str
|
| 25 |
content: str
|
|
@@ -41,24 +41,16 @@ class CompletionResponse(BaseModel):
|
|
| 41 |
choices: List[Dict[str, Any]]
|
| 42 |
usage: Dict[str, int]
|
| 43 |
|
| 44 |
-
# Funzioni di utilità
|
| 45 |
def format_chat_prompt(messages: List[Message]) -> str:
|
| 46 |
-
"""Formatta i messaggi nel formato atteso da Mistral Instruct."""
|
| 47 |
conversation = []
|
| 48 |
-
|
| 49 |
for message in messages:
|
| 50 |
-
if message.role == "system":
|
| 51 |
-
# Inserisce il messaggio di sistema come istruzione iniziale
|
| 52 |
-
conversation.append(f"<s>[INST] {message.content} [/INST]</s>")
|
| 53 |
-
elif message.role == "user":
|
| 54 |
conversation.append(f"<s>[INST] {message.content} [/INST]</s>")
|
| 55 |
elif message.role == "assistant":
|
| 56 |
conversation.append(f"<s>{message.content}</s>")
|
| 57 |
-
|
| 58 |
return "".join(conversation)
|
| 59 |
|
| 60 |
def load_model():
|
| 61 |
-
"""Carica il modello Mistral quantizzato."""
|
| 62 |
global model, status_message
|
| 63 |
try:
|
| 64 |
status_message = "Caricamento modello in corso..."
|
|
@@ -67,7 +59,7 @@ def load_model():
|
|
| 67 |
model_file=MODEL_FILE,
|
| 68 |
model_type=MODEL_TYPE,
|
| 69 |
context_length=4096,
|
| 70 |
-
threads=4
|
| 71 |
)
|
| 72 |
status_message = "Modello caricato con successo"
|
| 73 |
return True
|
|
@@ -76,14 +68,11 @@ def load_model():
|
|
| 76 |
return False
|
| 77 |
|
| 78 |
def generate_response(prompt, temperature=0.7, top_p=0.95, max_tokens=MAX_NEW_TOKENS):
|
| 79 |
-
"""Genera una risposta dal modello."""
|
| 80 |
global model, status_message
|
| 81 |
-
|
| 82 |
if model is None:
|
| 83 |
if not load_model():
|
| 84 |
return status_message
|
| 85 |
-
|
| 86 |
-
with MODEL_LOCK: # Previene richieste parallele che potrebbero causare OOM
|
| 87 |
try:
|
| 88 |
result = model(
|
| 89 |
prompt,
|
|
@@ -103,75 +92,43 @@ def generate_with_timing(text, temp, max_tok):
|
|
| 103 |
end_time = time.time()
|
| 104 |
return result, f"{end_time - start_time:.2f} secondi"
|
| 105 |
|
| 106 |
-
# Creazione dell'interfaccia Gradio
|
| 107 |
def create_gradio_interface():
|
| 108 |
with gr.Blocks(title="Mistral API") as interface:
|
| 109 |
gr.Markdown("# Mistral-7B API Server")
|
| 110 |
-
|
| 111 |
with gr.Row():
|
| 112 |
with gr.Column():
|
| 113 |
status = gr.Textbox(value=lambda: status_message, label="Stato del modello", interactive=False)
|
| 114 |
load_button = gr.Button("Carica Modello")
|
| 115 |
load_button.click(load_model, inputs=[], outputs=[])
|
| 116 |
-
|
| 117 |
with gr.Row():
|
| 118 |
with gr.Column():
|
| 119 |
-
input_text = gr.Textbox(
|
| 120 |
-
lines=5,
|
| 121 |
-
label="Input",
|
| 122 |
-
placeholder="Inserisci il tuo messaggio qui..."
|
| 123 |
-
)
|
| 124 |
-
|
| 125 |
with gr.Row():
|
| 126 |
-
temp_slider = gr.Slider(
|
| 127 |
-
|
| 128 |
-
maximum=1.0,
|
| 129 |
-
value=0.7,
|
| 130 |
-
step=0.1,
|
| 131 |
-
label="Temperatura"
|
| 132 |
-
)
|
| 133 |
-
|
| 134 |
-
max_token_slider = gr.Slider(
|
| 135 |
-
minimum=100,
|
| 136 |
-
maximum=MAX_NEW_TOKENS,
|
| 137 |
-
value=1024,
|
| 138 |
-
step=100,
|
| 139 |
-
label="Max Token"
|
| 140 |
-
)
|
| 141 |
-
|
| 142 |
submit_button = gr.Button("Genera")
|
| 143 |
-
|
| 144 |
with gr.Column():
|
| 145 |
output_text = gr.Textbox(lines=12, label="Risposta del modello")
|
| 146 |
-
|
| 147 |
gen_time = gr.Textbox(label="Tempo di generazione", interactive=False)
|
| 148 |
-
|
| 149 |
submit_button.click(
|
| 150 |
generate_with_timing,
|
| 151 |
inputs=[input_text, temp_slider, max_token_slider],
|
| 152 |
outputs=[output_text, gen_time]
|
| 153 |
)
|
| 154 |
-
|
| 155 |
gr.Markdown("""
|
| 156 |
## API Endpoint
|
| 157 |
Questa applicazione espone un endpoint API compatibile con OpenAI:
|
| 158 |
- `/v1/chat/completions` - Per richieste di completamento chat
|
| 159 |
- `/status` - Per verificare lo stato del modello
|
| 160 |
-
|
| 161 |
-
L'endpoint è accessibile dall'URL di questo Hugging Face Space.
|
| 162 |
""")
|
| 163 |
-
|
| 164 |
return interface
|
| 165 |
|
| 166 |
-
# Decorator per richiedere la GPU dallo space
|
| 167 |
@spaces.GPU
|
| 168 |
def get_gpu():
|
| 169 |
return "GPU allocata con successo"
|
| 170 |
|
| 171 |
-
# Crea l'applicazione FastAPI
|
| 172 |
app = FastAPI()
|
| 173 |
|
| 174 |
-
# Configura CORS
|
| 175 |
app.add_middleware(
|
| 176 |
CORSMiddleware,
|
| 177 |
allow_origins=["*"],
|
|
@@ -180,14 +137,11 @@ app.add_middleware(
|
|
| 180 |
allow_headers=["*"],
|
| 181 |
)
|
| 182 |
|
| 183 |
-
# API endpoint compatibile con OpenAI
|
| 184 |
@app.post("/v1/chat/completions", response_model=CompletionResponse)
|
| 185 |
async def create_completion(request: CompletionRequest):
|
| 186 |
try:
|
| 187 |
prompt = format_chat_prompt(request.messages)
|
| 188 |
-
|
| 189 |
-
max_tokens = min(request.max_tokens, MAX_NEW_TOKENS) # Limita i token per evitare OOM
|
| 190 |
-
|
| 191 |
start_time = time.time()
|
| 192 |
completion_text = generate_response(
|
| 193 |
prompt,
|
|
@@ -196,11 +150,8 @@ async def create_completion(request: CompletionRequest):
|
|
| 196 |
max_tokens=max_tokens
|
| 197 |
)
|
| 198 |
end_time = time.time()
|
| 199 |
-
|
| 200 |
-
# Calcola il numero di token (approssimativo)
|
| 201 |
input_tokens = len(prompt.split())
|
| 202 |
output_tokens = len(completion_text.split())
|
| 203 |
-
|
| 204 |
response = {
|
| 205 |
"id": f"chatcmpl-{os.urandom(4).hex()}",
|
| 206 |
"object": "chat.completion",
|
|
@@ -222,31 +173,23 @@ async def create_completion(request: CompletionRequest):
|
|
| 222 |
"total_tokens": input_tokens + output_tokens,
|
| 223 |
}
|
| 224 |
}
|
| 225 |
-
|
| 226 |
return response
|
| 227 |
except Exception as e:
|
| 228 |
raise HTTPException(status_code=500, detail=str(e))
|
| 229 |
|
| 230 |
-
# API endpoint per verificare lo stato del modello
|
| 231 |
@app.get("/status")
|
| 232 |
async def get_status():
|
| 233 |
return {"status": status_message, "model": MODEL_PATH}
|
| 234 |
|
| 235 |
-
# Crea l'interfaccia Gradio
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
# Monta Gradio su FastAPI usando il metodo corretto per Gradio 4
|
| 239 |
-
app = gr.mount_gradio_app(app, demo, path="/gradio")
|
| 240 |
|
| 241 |
-
# Precarica il modello all'avvio (usando il nuovo metodo lifespan invece di on_event)
|
| 242 |
@app.on_event("startup")
|
| 243 |
async def startup_load_model():
|
| 244 |
-
# Assicurati che la GPU sia allocata prima di caricare il modello
|
| 245 |
get_gpu()
|
| 246 |
load_model()
|
| 247 |
|
| 248 |
-
# Per Hugging Face Spaces, assicurati che l'app sia esportata correttamente
|
| 249 |
if __name__ == "__main__":
|
| 250 |
import uvicorn
|
| 251 |
-
|
| 252 |
-
uvicorn.run(app, host="0.0.0.0", port=7860, log_level="info")
|
|
|
|
| 14 |
model = None
|
| 15 |
status_message = "Modello non ancora caricato"
|
| 16 |
MODEL_PATH = "TheBloke/Mistral-7B-Instruct-v0.2-GGUF"
|
| 17 |
+
MODEL_FILE = "mistral-7b-instruct-v0.2.Q4_K_M.gguf"
|
| 18 |
MODEL_TYPE = "mistral"
|
| 19 |
MAX_NEW_TOKENS = 2048
|
| 20 |
+
MODEL_LOCK = Lock()
|
| 21 |
|
| 22 |
+
# Pydantic models
|
| 23 |
class Message(BaseModel):
|
| 24 |
role: str
|
| 25 |
content: str
|
|
|
|
| 41 |
choices: List[Dict[str, Any]]
|
| 42 |
usage: Dict[str, int]
|
| 43 |
|
|
|
|
| 44 |
def format_chat_prompt(messages: List[Message]) -> str:
|
|
|
|
| 45 |
conversation = []
|
|
|
|
| 46 |
for message in messages:
|
| 47 |
+
if message.role == "system" or message.role == "user":
|
|
|
|
|
|
|
|
|
|
| 48 |
conversation.append(f"<s>[INST] {message.content} [/INST]</s>")
|
| 49 |
elif message.role == "assistant":
|
| 50 |
conversation.append(f"<s>{message.content}</s>")
|
|
|
|
| 51 |
return "".join(conversation)
|
| 52 |
|
| 53 |
def load_model():
|
|
|
|
| 54 |
global model, status_message
|
| 55 |
try:
|
| 56 |
status_message = "Caricamento modello in corso..."
|
|
|
|
| 59 |
model_file=MODEL_FILE,
|
| 60 |
model_type=MODEL_TYPE,
|
| 61 |
context_length=4096,
|
| 62 |
+
threads=4
|
| 63 |
)
|
| 64 |
status_message = "Modello caricato con successo"
|
| 65 |
return True
|
|
|
|
| 68 |
return False
|
| 69 |
|
| 70 |
def generate_response(prompt, temperature=0.7, top_p=0.95, max_tokens=MAX_NEW_TOKENS):
|
|
|
|
| 71 |
global model, status_message
|
|
|
|
| 72 |
if model is None:
|
| 73 |
if not load_model():
|
| 74 |
return status_message
|
| 75 |
+
with MODEL_LOCK:
|
|
|
|
| 76 |
try:
|
| 77 |
result = model(
|
| 78 |
prompt,
|
|
|
|
| 92 |
end_time = time.time()
|
| 93 |
return result, f"{end_time - start_time:.2f} secondi"
|
| 94 |
|
|
|
|
| 95 |
def create_gradio_interface():
|
| 96 |
with gr.Blocks(title="Mistral API") as interface:
|
| 97 |
gr.Markdown("# Mistral-7B API Server")
|
|
|
|
| 98 |
with gr.Row():
|
| 99 |
with gr.Column():
|
| 100 |
status = gr.Textbox(value=lambda: status_message, label="Stato del modello", interactive=False)
|
| 101 |
load_button = gr.Button("Carica Modello")
|
| 102 |
load_button.click(load_model, inputs=[], outputs=[])
|
|
|
|
| 103 |
with gr.Row():
|
| 104 |
with gr.Column():
|
| 105 |
+
input_text = gr.Textbox(lines=5, label="Input", placeholder="Inserisci il tuo messaggio qui...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
with gr.Row():
|
| 107 |
+
temp_slider = gr.Slider(0.1, 1.0, value=0.7, step=0.1, label="Temperatura")
|
| 108 |
+
max_token_slider = gr.Slider(100, MAX_NEW_TOKENS, value=1024, step=100, label="Max Token")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 109 |
submit_button = gr.Button("Genera")
|
|
|
|
| 110 |
with gr.Column():
|
| 111 |
output_text = gr.Textbox(lines=12, label="Risposta del modello")
|
|
|
|
| 112 |
gen_time = gr.Textbox(label="Tempo di generazione", interactive=False)
|
|
|
|
| 113 |
submit_button.click(
|
| 114 |
generate_with_timing,
|
| 115 |
inputs=[input_text, temp_slider, max_token_slider],
|
| 116 |
outputs=[output_text, gen_time]
|
| 117 |
)
|
|
|
|
| 118 |
gr.Markdown("""
|
| 119 |
## API Endpoint
|
| 120 |
Questa applicazione espone un endpoint API compatibile con OpenAI:
|
| 121 |
- `/v1/chat/completions` - Per richieste di completamento chat
|
| 122 |
- `/status` - Per verificare lo stato del modello
|
|
|
|
|
|
|
| 123 |
""")
|
|
|
|
| 124 |
return interface
|
| 125 |
|
|
|
|
| 126 |
@spaces.GPU
|
| 127 |
def get_gpu():
|
| 128 |
return "GPU allocata con successo"
|
| 129 |
|
|
|
|
| 130 |
app = FastAPI()
|
| 131 |
|
|
|
|
| 132 |
app.add_middleware(
|
| 133 |
CORSMiddleware,
|
| 134 |
allow_origins=["*"],
|
|
|
|
| 137 |
allow_headers=["*"],
|
| 138 |
)
|
| 139 |
|
|
|
|
| 140 |
@app.post("/v1/chat/completions", response_model=CompletionResponse)
|
| 141 |
async def create_completion(request: CompletionRequest):
|
| 142 |
try:
|
| 143 |
prompt = format_chat_prompt(request.messages)
|
| 144 |
+
max_tokens = min(request.max_tokens, MAX_NEW_TOKENS)
|
|
|
|
|
|
|
| 145 |
start_time = time.time()
|
| 146 |
completion_text = generate_response(
|
| 147 |
prompt,
|
|
|
|
| 150 |
max_tokens=max_tokens
|
| 151 |
)
|
| 152 |
end_time = time.time()
|
|
|
|
|
|
|
| 153 |
input_tokens = len(prompt.split())
|
| 154 |
output_tokens = len(completion_text.split())
|
|
|
|
| 155 |
response = {
|
| 156 |
"id": f"chatcmpl-{os.urandom(4).hex()}",
|
| 157 |
"object": "chat.completion",
|
|
|
|
| 173 |
"total_tokens": input_tokens + output_tokens,
|
| 174 |
}
|
| 175 |
}
|
|
|
|
| 176 |
return response
|
| 177 |
except Exception as e:
|
| 178 |
raise HTTPException(status_code=500, detail=str(e))
|
| 179 |
|
|
|
|
| 180 |
@app.get("/status")
|
| 181 |
async def get_status():
|
| 182 |
return {"status": status_message, "model": MODEL_PATH}
|
| 183 |
|
| 184 |
+
# Crea l'interfaccia Gradio e monta su un path dedicato per evitare errori statici
|
| 185 |
+
interface = create_gradio_interface()
|
| 186 |
+
app = gr.mount_gradio_app(app, interface, path="/gradio")
|
|
|
|
|
|
|
| 187 |
|
|
|
|
| 188 |
@app.on_event("startup")
|
| 189 |
async def startup_load_model():
|
|
|
|
| 190 |
get_gpu()
|
| 191 |
load_model()
|
| 192 |
|
|
|
|
| 193 |
if __name__ == "__main__":
|
| 194 |
import uvicorn
|
| 195 |
+
uvicorn.run(app, host="0.0.0.0", port=7860, log_level="info")
|
|
|