Update app.py
Browse files
app.py
CHANGED
@@ -27,7 +27,33 @@ model_info = {
|
|
27 |
"Pythia": {"task": "text-generation", "description": "Pythia language model"},
|
28 |
"T5-Small": {"task": "text2text-generation", "description": "Small T5 model", "hf_model_name": "t5-small"},
|
29 |
"GPT-Neo": {"task": "text-generation", "description": "GPT-Neo model"},
|
30 |
-
"Distil-GPT-2": {"task": "text-generation", "description": "Distilled GPT-2 model"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
}
|
32 |
|
33 |
# --- Penyimpanan Model Global (untuk Lazy Loading) ---
|
@@ -99,7 +125,7 @@ def list_available_models():
|
|
99 |
def predict_with_model(model_id):
|
100 |
"""
|
101 |
Endpoint utama untuk prediksi model.
|
102 |
-
Menerima 'inputs' (teks) dan 'parameters' (dictionary) opsional.
|
103 |
"""
|
104 |
logger.info(f"Menerima permintaan untuk model: {model_id}")
|
105 |
if model_id not in model_info:
|
@@ -111,49 +137,51 @@ def predict_with_model(model_id):
|
|
111 |
model_task = model_info[model_id]["task"]
|
112 |
|
113 |
data = request.json
|
114 |
-
|
115 |
-
|
|
|
116 |
|
117 |
-
if not
|
118 |
-
return jsonify({"error": "Input 'inputs' tidak boleh kosong."}), 400
|
119 |
|
120 |
-
logger.info(f"Inferensi: Model='{model_id}', Task='{model_task}',
|
121 |
|
122 |
result = []
|
123 |
# --- Penanganan Parameter dan Inferensi berdasarkan Tipe Tugas ---
|
124 |
if model_task == "text-generation":
|
125 |
-
# Default parameters for text-generation
|
126 |
gen_params = {
|
127 |
-
"max_new_tokens": parameters.get("max_new_tokens", 150),
|
128 |
"temperature": parameters.get("temperature", 0.7),
|
129 |
"do_sample": parameters.get("do_sample", True),
|
130 |
"return_full_text": parameters.get("return_full_text", False), # Sangat penting untuk chatbot
|
131 |
"num_return_sequences": parameters.get("num_return_sequences", 1),
|
132 |
"top_k": parameters.get("top_k", 50),
|
133 |
"top_p": parameters.get("top_p", 0.95),
|
134 |
-
"repetition_penalty": parameters.get("repetition_penalty", 1.2),
|
135 |
}
|
136 |
-
|
|
|
137 |
|
138 |
elif model_task == "fill-mask":
|
139 |
mask_params = {
|
140 |
"top_k": parameters.get("top_k", 5)
|
141 |
}
|
142 |
-
|
|
|
|
|
143 |
|
144 |
-
elif model_task == "text2text-generation":
|
145 |
t2t_params = {
|
146 |
"max_new_tokens": parameters.get("max_new_tokens", 150),
|
147 |
"temperature": parameters.get("temperature", 0.7),
|
148 |
"do_sample": parameters.get("do_sample", True),
|
149 |
}
|
150 |
-
result = model_pipeline(
|
151 |
|
152 |
else:
|
153 |
-
|
154 |
-
result = model_pipeline(inputs, **parameters)
|
155 |
|
156 |
-
# --- Konsistensi Format Output ---
|
157 |
response_output = {}
|
158 |
if model_task == "text-generation" or model_task == "text2text-generation":
|
159 |
if result and len(result) > 0 and 'generated_text' in result[0]:
|
@@ -166,22 +194,18 @@ def predict_with_model(model_id):
|
|
166 |
for p in result
|
167 |
]
|
168 |
else:
|
169 |
-
# Untuk jenis tugas lain, kembalikan hasil mentah
|
170 |
response_output = result
|
171 |
|
172 |
logger.info(f"Inferensi berhasil untuk '{model_id}'. Output singkat: '{str(response_output)[:200]}'")
|
173 |
-
return jsonify({"model": model_id, "inputs":
|
174 |
|
175 |
except ValueError as ve:
|
176 |
-
# Error yang berasal dari get_model_pipeline atau validasi input
|
177 |
logger.error(f"Validasi atau konfigurasi error untuk model '{model_id}': {str(ve)}")
|
178 |
return jsonify({"error": str(ve), "message": "Kesalahan konfigurasi atau input model."}), 400
|
179 |
except RuntimeError as re:
|
180 |
-
# Error saat memuat model
|
181 |
logger.error(f"Error runtime saat memuat model '{model_id}': {str(re)}")
|
182 |
-
return jsonify({"error": str(re), "message": "Model gagal dimuat."}), 503
|
183 |
except Exception as e:
|
184 |
-
# Catch all other unexpected errors during prediction
|
185 |
logger.error(f"Terjadi kesalahan tak terduga saat memprediksi dengan model '{model_id}': {str(e)}", exc_info=True)
|
186 |
return jsonify({"error": str(e), "message": "Terjadi kesalahan internal server."}), 500
|
187 |
|
|
|
27 |
"Pythia": {"task": "text-generation", "description": "Pythia language model"},
|
28 |
"T5-Small": {"task": "text2text-generation", "description": "Small T5 model", "hf_model_name": "t5-small"},
|
29 |
"GPT-Neo": {"task": "text-generation", "description": "GPT-Neo model"},
|
30 |
+
"Distil-GPT-2": {"task": "text-generation", "description": "Distilled GPT-2 model"},
|
31 |
+
# --- MODEL EXTERNAL ---
|
32 |
+
"Gemma-2B-IT": { # ID yang Anda inginkan di API Anda
|
33 |
+
"task": "text-generation",
|
34 |
+
"description": "Google's Gemma 2B Instruct model",
|
35 |
+
"hf_model_name": "google/gemma-2b-it"
|
36 |
+
},
|
37 |
+
"Mistral-7B-Instruct": {
|
38 |
+
"task": "text-generation",
|
39 |
+
"description": "Mistral AI's Mistral 7B Instruct model",
|
40 |
+
"hf_model_name": "mistralai/Mistral-7B-Instruct-v0.3",
|
41 |
+
}
|
42 |
+
"Qwen3-4B-RPG": {
|
43 |
+
"task": "text-generation",
|
44 |
+
"description": "Chun121's Qwen 4B RPG Roleplay model (Uncensored)",
|
45 |
+
"hf_model_name": "Chun121/qwen3-4B-rpg-roleplay"
|
46 |
+
},
|
47 |
+
"Llama-3.2-Uncensored-3B": {
|
48 |
+
"task": "text-generation",
|
49 |
+
"description": "Dhirajlochib's Llama 3.2 Uncensored 3B",
|
50 |
+
"hf_model_name": "dhirajlochib/llama-3.2-unsensored-3b"
|
51 |
+
},
|
52 |
+
"TinyLLama-NSFW-Chatbot": {
|
53 |
+
"task": "text-generation",
|
54 |
+
"description": "BilalRahib's TinyLLama NSFW Chatbot",
|
55 |
+
"hf_model_name": "bilalRahib/TinyLLama-NSFW-Chatbot"
|
56 |
+
}
|
57 |
}
|
58 |
|
59 |
# --- Penyimpanan Model Global (untuk Lazy Loading) ---
|
|
|
125 |
def predict_with_model(model_id):
|
126 |
"""
|
127 |
Endpoint utama untuk prediksi model.
|
128 |
+
Menerima 'inputs' (teks pra-diformat) dan 'parameters' (dictionary) opsional.
|
129 |
"""
|
130 |
logger.info(f"Menerima permintaan untuk model: {model_id}")
|
131 |
if model_id not in model_info:
|
|
|
137 |
model_task = model_info[model_id]["task"]
|
138 |
|
139 |
data = request.json
|
140 |
+
# Input sekarang diharapkan sebagai fullPromptString dari frontend
|
141 |
+
full_prompt_string_from_frontend = data.get('inputs', '')
|
142 |
+
parameters = data.get('parameters', {})
|
143 |
|
144 |
+
if not full_prompt_string_from_frontend:
|
145 |
+
return jsonify({"error": "Input 'inputs' (full prompt string) tidak boleh kosong."}), 400
|
146 |
|
147 |
+
logger.info(f"Inferensi: Model='{model_id}', Task='{model_task}', Full Prompt='{full_prompt_string_from_frontend[:200]}...', Params='{parameters}'")
|
148 |
|
149 |
result = []
|
150 |
# --- Penanganan Parameter dan Inferensi berdasarkan Tipe Tugas ---
|
151 |
if model_task == "text-generation":
|
|
|
152 |
gen_params = {
|
153 |
+
"max_new_tokens": parameters.get("max_new_tokens", 150),
|
154 |
"temperature": parameters.get("temperature", 0.7),
|
155 |
"do_sample": parameters.get("do_sample", True),
|
156 |
"return_full_text": parameters.get("return_full_text", False), # Sangat penting untuk chatbot
|
157 |
"num_return_sequences": parameters.get("num_return_sequences", 1),
|
158 |
"top_k": parameters.get("top_k", 50),
|
159 |
"top_p": parameters.get("top_p", 0.95),
|
160 |
+
"repetition_penalty": parameters.get("repetition_penalty", 1.2),
|
161 |
}
|
162 |
+
# Langsung berikan full_prompt_string_from_frontend ke pipeline
|
163 |
+
result = model_pipeline(full_prompt_string_from_frontend, **gen_params)
|
164 |
|
165 |
elif model_task == "fill-mask":
|
166 |
mask_params = {
|
167 |
"top_k": parameters.get("top_k", 5)
|
168 |
}
|
169 |
+
# Untuk fill-mask, input harus string biasa, bukan prompt yang kompleks
|
170 |
+
# Anda perlu memastikan frontend tidak mengirim prompt kompleks ke fill-mask model
|
171 |
+
result = model_pipeline(full_prompt_string_from_frontend, **mask_params)
|
172 |
|
173 |
+
elif model_task == "text2text-generation":
|
174 |
t2t_params = {
|
175 |
"max_new_tokens": parameters.get("max_new_tokens", 150),
|
176 |
"temperature": parameters.get("temperature", 0.7),
|
177 |
"do_sample": parameters.get("do_sample", True),
|
178 |
}
|
179 |
+
result = model_pipeline(full_prompt_string_from_frontend, **t2t_params)
|
180 |
|
181 |
else:
|
182 |
+
result = model_pipeline(full_prompt_string_from_frontend, **parameters)
|
|
|
183 |
|
184 |
+
# --- Konsistensi Format Output (tidak berubah dari update sebelumnya) ---
|
185 |
response_output = {}
|
186 |
if model_task == "text-generation" or model_task == "text2text-generation":
|
187 |
if result and len(result) > 0 and 'generated_text' in result[0]:
|
|
|
194 |
for p in result
|
195 |
]
|
196 |
else:
|
|
|
197 |
response_output = result
|
198 |
|
199 |
logger.info(f"Inferensi berhasil untuk '{model_id}'. Output singkat: '{str(response_output)[:200]}'")
|
200 |
+
return jsonify({"model": model_id, "inputs": full_prompt_string_from_frontend, "outputs": response_output})
|
201 |
|
202 |
except ValueError as ve:
|
|
|
203 |
logger.error(f"Validasi atau konfigurasi error untuk model '{model_id}': {str(ve)}")
|
204 |
return jsonify({"error": str(ve), "message": "Kesalahan konfigurasi atau input model."}), 400
|
205 |
except RuntimeError as re:
|
|
|
206 |
logger.error(f"Error runtime saat memuat model '{model_id}': {str(re)}")
|
207 |
+
return jsonify({"error": str(re), "message": "Model gagal dimuat."}), 503
|
208 |
except Exception as e:
|
|
|
209 |
logger.error(f"Terjadi kesalahan tak terduga saat memprediksi dengan model '{model_id}': {str(e)}", exc_info=True)
|
210 |
return jsonify({"error": str(e), "message": "Terjadi kesalahan internal server."}), 500
|
211 |
|