Raffaele Terribile
commited on
Modifica app per utilizzare un modello locale
Browse files
app.py
CHANGED
@@ -5,10 +5,27 @@ import requests
|
|
5 |
import inspect
|
6 |
import pandas as pd
|
7 |
|
8 |
-
|
9 |
-
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
from litellm import LiteLLM
|
|
|
|
|
12 |
|
13 |
# (Keep Constants as is)
|
14 |
# --- Constants ---
|
@@ -25,11 +42,63 @@ def invert_sentence(sentence: str) -> str:
|
|
25 |
"""
|
26 |
return sentence[::-1]
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
# --- First Agent Definition ---
|
29 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
30 |
class FirstAgent:
|
31 |
### First Agent is the first attempt to develop an agent for the course. ###
|
32 |
def __init__(self):
|
|
|
33 |
# # Usa un modello Hugging Face gratuito
|
34 |
# token = os.getenv(os.getenv("TOKEN_NAME"))
|
35 |
# os.environ["HF_TOKEN"] = token
|
@@ -37,52 +106,81 @@ class FirstAgent:
|
|
37 |
# token=token
|
38 |
# )
|
39 |
|
40 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
model = None
|
42 |
|
43 |
-
# Try 1: Modello locale
|
44 |
try:
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
model = pipeline(
|
49 |
-
task="text-generation",
|
50 |
-
tokenizer=tokenizer,
|
51 |
-
model=model
|
52 |
-
)
|
53 |
-
print(f"Using local {model_id} model")
|
54 |
except Exception as e:
|
55 |
-
print(f"
|
56 |
|
57 |
-
# Try 2: Modello remoto
|
58 |
try:
|
59 |
-
|
60 |
-
|
61 |
-
)
|
62 |
-
print("Using Groq remote model")
|
63 |
except Exception as ex:
|
64 |
-
print(f"
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
|
|
67 |
self.agent = CodeAgent(
|
68 |
model=model,
|
69 |
tools=[
|
70 |
-
# DuckDuckGoSearchTool(),
|
71 |
-
# GoogleSearchTool(),
|
72 |
WebSearchTool(),
|
73 |
PythonInterpreterTool(),
|
74 |
WikipediaSearchTool(),
|
75 |
-
VisitWebpageTool()
|
76 |
-
# FinalAnswerTool #,
|
77 |
-
# Tool(name="invert_sentence", func=invert_sentence, description="Inverts the order of characters in a sentence.")
|
78 |
]
|
79 |
)
|
80 |
-
print("FirstAgent
|
|
|
81 |
def __call__(self, question: str) -> str:
|
82 |
-
print(f"Agent
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
86 |
|
87 |
# --- Basic Agent Definition ---
|
88 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
@@ -198,8 +296,8 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
198 |
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
|
199 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
200 |
except Exception as e:
|
201 |
-
|
202 |
-
|
203 |
|
204 |
if not answers_payload:
|
205 |
print("Agent did not produce any answers to submit.")
|
@@ -307,4 +405,4 @@ if __name__ == "__main__":
|
|
307 |
print("-"*(60 + len(" App Starting ")) + "\n")
|
308 |
|
309 |
print("Launching Gradio Interface for Basic Agent Evaluation...")
|
310 |
-
demo.launch(debug=True, share=False)
|
|
|
5 |
import inspect
|
6 |
import pandas as pd
|
7 |
|
8 |
+
# =============================================================================
|
9 |
+
# MODIFICHE APPORTATE PER RISOLVERE L'ERRORE "generate" NON TROVATO:
|
10 |
+
#
|
11 |
+
# PROBLEMA ORIGINALE:
|
12 |
+
# - Il pipeline di Transformers non è direttamente compatibile con smolagents
|
13 |
+
# - CodeAgent si aspetta un'interfaccia specifica che pipeline non implementa
|
14 |
+
# - L'errore "generate" si verificava perché smolagents cercava metodi non presenti
|
15 |
+
#
|
16 |
+
# SOLUZIONE IMPLEMENTATA:
|
17 |
+
# - Creata classe SimpleLocalModel che fa da wrapper
|
18 |
+
# - Implementa l'interfaccia __call__() che smolagents si aspetta
|
19 |
+
# - Gestisce la conversione dei messaggi e la generazione delle risposte
|
20 |
+
# - Fallback multipli: locale -> remoto -> fisso
|
21 |
+
# =============================================================================
|
22 |
+
|
23 |
+
from smolagents import CodeAgent, InferenceClientModel, VisitWebpageTool, PythonInterpreterTool, WebSearchTool, WikipediaSearchTool, FinalAnswerTool, Tool, tool
|
24 |
+
# Importazioni per modelli locali (SOLUZIONE per errore "generate"):
|
25 |
+
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
|
26 |
from litellm import LiteLLM
|
27 |
+
import threading
|
28 |
+
import time
|
29 |
|
30 |
# (Keep Constants as is)
|
31 |
# --- Constants ---
|
|
|
42 |
"""
|
43 |
return sentence[::-1]
|
44 |
|
45 |
+
# Wrapper semplificato per modelli locali
|
46 |
+
# NUOVO APPROCCIO: Questa classe risolve il problema dell'errore "generate"
|
47 |
+
# creando un'interfaccia compatibile tra Transformers pipeline e smolagents
|
48 |
+
class SimpleLocalModel:
|
49 |
+
"""Wrapper semplice per modelli Transformers locali."""
|
50 |
+
|
51 |
+
def __init__(self, model_name="gpt2"):
|
52 |
+
self.model_name = model_name
|
53 |
+
self.pipeline = None
|
54 |
+
self._load_model()
|
55 |
+
|
56 |
+
def _load_model(self):
|
57 |
+
"""Carica il modello locale."""
|
58 |
+
try:
|
59 |
+
print(f"Caricamento modello locale: {self.model_name}")
|
60 |
+
self.pipeline = pipeline(
|
61 |
+
"text-generation",
|
62 |
+
model=self.model_name,
|
63 |
+
# device=-1, # Usa CPU
|
64 |
+
return_full_text=False # Restituisce solo il testo generato
|
65 |
+
)
|
66 |
+
print(f"✅ Modello {self.model_name} caricato")
|
67 |
+
except Exception as e:
|
68 |
+
print(f"❌ Errore caricamento modello: {e}")
|
69 |
+
raise
|
70 |
+
|
71 |
+
def __call__(self, messages, **kwargs):
|
72 |
+
"""Genera risposta compatibile con smolagents."""
|
73 |
+
try:
|
74 |
+
# Estrai il prompt
|
75 |
+
if isinstance(messages, list) and messages:
|
76 |
+
prompt = messages[-1].get("content", "") if isinstance(messages[-1], dict) else str(messages[-1])
|
77 |
+
else:
|
78 |
+
prompt = str(messages)
|
79 |
+
|
80 |
+
if not prompt.strip():
|
81 |
+
return "Mi dispiace, non ho ricevuto una domanda."
|
82 |
+
|
83 |
+
# Genera risposta
|
84 |
+
result = self.pipeline(prompt, max_new_tokens=100, do_sample=True, temperature=0.7)
|
85 |
+
|
86 |
+
if result and len(result) > 0:
|
87 |
+
answer = result[0].get("generated_text", "").strip()
|
88 |
+
return answer if answer else "Non sono riuscito a generare una risposta."
|
89 |
+
else:
|
90 |
+
return "Errore nella generazione della risposta."
|
91 |
+
|
92 |
+
except Exception as e:
|
93 |
+
print(f"Errore generazione: {e}")
|
94 |
+
return f"Errore: {str(e)}"
|
95 |
+
|
96 |
# --- First Agent Definition ---
|
97 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
98 |
class FirstAgent:
|
99 |
### First Agent is the first attempt to develop an agent for the course. ###
|
100 |
def __init__(self):
|
101 |
+
# CODICE ORIGINALE COMMENTATO (che causava l'errore "generate"):
|
102 |
# # Usa un modello Hugging Face gratuito
|
103 |
# token = os.getenv(os.getenv("TOKEN_NAME"))
|
104 |
# os.environ["HF_TOKEN"] = token
|
|
|
106 |
# token=token
|
107 |
# )
|
108 |
|
109 |
+
# CODICE ORIGINALE COMMENTATO (approccio con pipeline non compatibile):
|
110 |
+
# # Configurazione con fallback multipli
|
111 |
+
# model = None
|
112 |
+
# # Try 1: Modello locale via Transformers
|
113 |
+
# try:
|
114 |
+
# model_id = "microsoft/Phi-4-mini-reasoning"
|
115 |
+
# tokenizer = AutoTokenizer.from_pretrained(model_id)
|
116 |
+
# model = AutoModelForCausalLM.from_pretrained(model_id) # ~500MB
|
117 |
+
# model = pipeline(
|
118 |
+
# task="text-generation",
|
119 |
+
# tokenizer=tokenizer,
|
120 |
+
# model=model
|
121 |
+
# )
|
122 |
+
# print(f"Using local {model_id} model")
|
123 |
+
# except Exception as e:
|
124 |
+
# print(f"Local model failed: {e}")
|
125 |
+
# # Try 2: Modello remoto gratuito
|
126 |
+
# try:
|
127 |
+
# model = LiteLLM(
|
128 |
+
# model_id="groq/mixtral-8x7b-32768" # Gratuito con registrazione
|
129 |
+
# )
|
130 |
+
# print("Using Groq remote model")
|
131 |
+
# except Exception as ex:
|
132 |
+
# print(f"Remote model failed: {ex}")
|
133 |
+
# raise Exception("No working model configuration found")
|
134 |
+
|
135 |
+
# NUOVO CODICE FUNZIONANTE:
|
136 |
+
# Configurazione con fallback per modelli locali
|
137 |
model = None
|
138 |
|
139 |
+
# Try 1: Modello locale semplificato
|
140 |
try:
|
141 |
+
print("🔄 Tentativo 1: Modello locale GPT-2")
|
142 |
+
model = SimpleLocalModel("microsoft/Phi-4-mini-reasoning")
|
143 |
+
print("✅ Usando modello locale GPT-2")
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
except Exception as e:
|
145 |
+
print(f"❌ Modello locale fallito: {e}")
|
146 |
|
147 |
+
# Try 2: Modello remoto (se disponibile)
|
148 |
try:
|
149 |
+
print("🔄 Tentativo 2: Modello remoto Groq")
|
150 |
+
model = LiteLLM(model="groq/mixtral-8x7b-32768")
|
151 |
+
print("✅ Usando modello remoto Groq")
|
|
|
152 |
except Exception as ex:
|
153 |
+
print(f"❌ Modello remoto fallito: {ex}")
|
154 |
+
|
155 |
+
# Try 3: Fallback finale - risposta fissa
|
156 |
+
class FallbackModel:
|
157 |
+
def __call__(self, messages, **kwargs):
|
158 |
+
return "Sono un agente semplificato. Il modello AI non è disponibile al momento."
|
159 |
+
|
160 |
+
model = FallbackModel()
|
161 |
+
print("⚠️ Usando modello di fallback")
|
162 |
|
163 |
+
# Inizializza l'agente
|
164 |
self.agent = CodeAgent(
|
165 |
model=model,
|
166 |
tools=[
|
|
|
|
|
167 |
WebSearchTool(),
|
168 |
PythonInterpreterTool(),
|
169 |
WikipediaSearchTool(),
|
170 |
+
VisitWebpageTool()
|
|
|
|
|
171 |
]
|
172 |
)
|
173 |
+
print("FirstAgent inizializzato.")
|
174 |
+
|
175 |
def __call__(self, question: str) -> str:
|
176 |
+
print(f"Agent ricevuto domanda (primi 50 char): {question[:50]}...")
|
177 |
+
try:
|
178 |
+
answer = self.agent.run(question)
|
179 |
+
print(f"Agent restituisce risposta: {str(answer)[:100]}...")
|
180 |
+
return str(answer)
|
181 |
+
except Exception as e:
|
182 |
+
print(f"Errore nell'agente: {e}")
|
183 |
+
return f"Errore nell'agente: {str(e)}"
|
184 |
|
185 |
# --- Basic Agent Definition ---
|
186 |
# ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
|
|
|
296 |
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
|
297 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
|
298 |
except Exception as e:
|
299 |
+
print(f"Error running agent on task {task_id}: {e}")
|
300 |
+
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
301 |
|
302 |
if not answers_payload:
|
303 |
print("Agent did not produce any answers to submit.")
|
|
|
405 |
print("-"*(60 + len(" App Starting ")) + "\n")
|
406 |
|
407 |
print("Launching Gradio Interface for Basic Agent Evaluation...")
|
408 |
+
demo.launch(debug=True, share=False)
|