SatyaDev13 commited on
Commit
e498f01
·
verified ·
1 Parent(s): cac9948

Upload app.py

Browse files
Files changed (1) hide show
  1. app.py +36 -92
app.py CHANGED
@@ -22,7 +22,7 @@ logger = logging.getLogger("ZeroGPU-App")
22
 
23
  MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
24
  HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
25
- DEFAULT_SYSTEM_PROMPT = "Anda adalah asisten AI yang membantu."
26
 
27
  logger.info(f"Memuat tokenizer: {MODEL_ID}")
28
  tokenizer = AutoTokenizer.from_pretrained(
@@ -49,21 +49,33 @@ model = AutoModelForCausalLM.from_pretrained(
49
  model.eval()
50
 
51
 
52
- @spaces.GPU
53
- def generate_reply(messages: List[dict], max_tokens: int, temperature: float) -> str:
54
- """Memproses seluruh rangkaian riwayat pesan langsung menggunakan chat template model."""
 
 
55
  if hasattr(tokenizer, "apply_chat_template"):
56
  try:
57
- text = tokenizer.apply_chat_template(
58
  messages, tokenize=False, add_generation_prompt=True
59
  )
60
  except Exception as e:
61
- logger.error(f"Gagal menerapkan template chat: {e}")
62
- text = "\n".join([f"{m['role']}: {m['content']}" for m in messages]) + "\nassistant:"
63
- else:
64
- text = "\n".join([f"{m['role']}: {m['content']}" for m in messages]) + "\nassistant:"
 
 
 
 
65
 
66
- inputs = tokenizer(text, return_tensors="pt")
 
 
 
 
 
 
67
  device = next(model.parameters()).device
68
  inputs = {k: v.to(device) for k, v in inputs.items()}
69
 
@@ -82,25 +94,19 @@ def generate_reply(messages: List[dict], max_tokens: int, temperature: float) ->
82
  return tokenizer.decode(new_tokens, skip_special_tokens=True).strip()
83
 
84
 
85
- # ============================================================
86
- # Skema Data Baru untuk Mendukung Struktur Percakapan
87
- # ============================================================
88
  class ChatMessage(BaseModel):
89
- role: str # "system", "user", atau "assistant"
90
  content: str
91
 
92
-
93
  class GenerateRequest(BaseModel):
94
- messages: List[ChatMessage] = Field(..., description="Daftar seluruh riwayat obrolan (Array of Objects)")
95
  max_tokens: int = Field(default=1024, ge=1, le=4096)
96
  temperature: float = Field(default=0.7, ge=0.0, le=2.0)
97
 
98
-
99
  class GenerateResponse(BaseModel):
100
  response: str
101
  model: str = MODEL_ID
102
 
103
-
104
  app = Server(title="ZeroGPU Chat API")
105
 
106
  app.add_middleware(
@@ -110,58 +116,26 @@ app.add_middleware(
110
  allow_headers=["*"],
111
  )
112
 
113
-
114
- @app.get("/api/health")
115
- def health():
116
- return {
117
- "status": "ok",
118
- "model": MODEL_ID,
119
- "cuda_available": torch.cuda.is_available(),
120
- }
121
-
122
-
123
  @app.post("/api/generate", response_model=GenerateResponse)
124
  def api_generate_post(body: GenerateRequest):
125
- """
126
- POST /api/generate — Menerima array pesan lengkap untuk mempertahankan konteks obrolan.
127
-
128
- Format JSON Request:
129
- {
130
- "messages": [
131
- {"role": "system", "content": "Kamu adalah robot pintar."},
132
- {"role": "user", "content": "Halo!"},
133
- {"role": "assistant", "content": "Hai! Ada yang bisa kubantu?"},
134
- {"role": "user", "content": "Siapa nama kamu tadi?"}
135
- ],
136
- "max_tokens": 512,
137
- "temperature": 0.7
138
- }
139
- """
140
  if not body.messages:
141
- return {"response": "Error: Parameter 'messages' tidak boleh kosong.", "model": MODEL_ID}
142
 
143
- # Konversi data Pydantic menjadi format dict standar Python
144
  formatted_messages = [{"role": m.role, "content": m.content} for m in body.messages]
 
 
 
145
 
146
- text = generate_reply(
147
- messages=formatted_messages,
 
148
  max_tokens=body.max_tokens,
149
  temperature=body.temperature,
150
  )
151
  return GenerateResponse(response=text)
152
 
153
-
154
  @app.get("/api/generate", response_model=GenerateResponse)
155
- def api_generate_get(
156
- prompt: str,
157
- system_prompt: str = DEFAULT_SYSTEM_PROMPT,
158
- max_tokens: int = 1024,
159
- temperature: float = 0.7,
160
- ):
161
- """
162
- GET /api/generate — Mode sekali jawab (Single-turn fallback) agar tetap
163
- bisa digunakan untuk pengetesan instan via URL browser/curl.
164
- """
165
  prompt = (prompt or "").strip()
166
  if not prompt:
167
  return {"response": "Error: prompt tidak boleh kosong.", "model": MODEL_ID}
@@ -170,18 +144,11 @@ def api_generate_get(
170
  {"role": "system", "content": system_prompt},
171
  {"role": "user", "content": prompt}
172
  ]
173
-
174
- text = generate_reply(
175
- messages=fallback_messages,
176
- max_tokens=max_tokens,
177
- temperature=temperature,
178
- )
179
  return GenerateResponse(response=text)
180
 
181
-
182
- # ============================================================
183
- # UI Baru Menggunakan Tampilan Chat Box Interaktif
184
- # ============================================================
185
  @app.get("/", response_class=HTMLResponse)
186
  async def homepage():
187
  return """
@@ -207,34 +174,23 @@ async def homepage():
207
  </head>
208
  <body>
209
  <h1>ZeroGPU Chat API (Multi-Turn)</h1>
210
- <p>UI Web ini otomatis menyimpan memori riwayat obrolan lokal Anda di browser.</p>
211
-
212
  <div id="chat-box"></div>
213
-
214
  <label>Pesan Anda</label>
215
  <textarea id="prompt" rows="3" placeholder="Ketik pesan lalu tekan Kirim..."></textarea>
216
-
217
  <label>System Prompt (Mengubah ini akan mereset obrolan)</label>
218
  <input id="system_prompt" value="Anda adalah asisten AI yang membantu." onchange="resetChat()">
219
-
220
  <div class="actions">
221
  <button id="btn" onclick="kirim()">Kirim</button>
222
  <button style="background:#ff4d4d;" onclick="resetChat()">Reset Obrolan</button>
223
  </div>
224
-
225
  <script>
226
  let chatHistory = [];
227
-
228
  function initChat() {
229
  const sysPrompt = document.getElementById('system_prompt').value;
230
  chatHistory = [{ role: "system", content: sysPrompt }];
231
  renderChat();
232
  }
233
-
234
- function resetChat() {
235
- initChat();
236
- }
237
-
238
  function renderChat() {
239
  const box = document.getElementById('chat-box');
240
  box.innerHTML = '';
@@ -246,32 +202,25 @@ function renderChat() {
246
  });
247
  box.scrollTop = box.scrollHeight;
248
  }
249
-
250
  function escapeHTML(str) {
251
  return str.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
252
  }
253
-
254
  async function kirim() {
255
  const promptInput = document.getElementById('prompt');
256
  const btn = document.getElementById('btn');
257
  const userText = promptInput.value.trim();
258
-
259
  if (!userText) return;
260
  if (chatHistory.length === 0) initChat();
261
-
262
  chatHistory.push({ role: "user", content: userText });
263
  renderChat();
264
-
265
  promptInput.value = "";
266
  btn.disabled = true;
267
-
268
  const box = document.getElementById('chat-box');
269
  const loadingDiv = document.createElement('div');
270
  loadingDiv.className = 'msg assistant';
271
  loadingDiv.textContent = 'Sedang mengetik jawaban...';
272
  box.appendChild(loadingDiv);
273
  box.scrollTop = box.scrollHeight;
274
-
275
  try {
276
  const res = await fetch('/api/generate', {
277
  method: 'POST',
@@ -282,10 +231,8 @@ async function kirim() {
282
  temperature: 0.7
283
  })
284
  });
285
-
286
  const data = await res.json();
287
  box.removeChild(loadingDiv);
288
-
289
  chatHistory.push({ role: "assistant", content: data.response });
290
  renderChat();
291
  } catch (e) {
@@ -300,15 +247,12 @@ async function kirim() {
300
  btn.disabled = false;
301
  }
302
  }
303
-
304
- // Inisialisasi awal saat halaman dimuat
305
  initChat();
306
  </script>
307
  </body>
308
  </html>
309
  """
310
 
311
-
312
  if __name__ == "__main__":
313
  app.launch(show_error=True)
314
 
 
22
 
23
  MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
24
  HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
25
+ DEFAULT_SYSTEM_PROMPT = "Anda adalah asisten AI yang cerdas, praktis, dan membantu."
26
 
27
  logger.info(f"Memuat tokenizer: {MODEL_ID}")
28
  tokenizer = AutoTokenizer.from_pretrained(
 
49
  model.eval()
50
 
51
 
52
+ # ============================================================
53
+ # PEMBARUAN: Format teks dikerjakan di CPU, BUKAN di GPU
54
+ # ============================================================
55
+ def build_prompt_string(messages: List[dict]) -> str:
56
+ """Mengubah array obrolan menjadi string rapi dengan format ChatML/Llama-3"""
57
  if hasattr(tokenizer, "apply_chat_template"):
58
  try:
59
+ return tokenizer.apply_chat_template(
60
  messages, tokenize=False, add_generation_prompt=True
61
  )
62
  except Exception as e:
63
+ logger.error(f"Peringatan: Gagal memuat template asli, beralih ke ChatML manual. Error: {e}")
64
+
65
+ # Fallback manual menggunakan format ChatML (Standar untuk Hermes 3)
66
+ prompt = ""
67
+ for m in messages:
68
+ prompt += f"<|im_start|>{m['role']}\n{m['content']}<|im_end|>\n"
69
+ prompt += "<|im_start|>assistant\n"
70
+ return prompt
71
 
72
+
73
+ # ============================================================
74
+ # FUNGSI GPU HANYA MENERIMA STRING MENTAH
75
+ # ============================================================
76
+ @spaces.GPU
77
+ def generate_reply_gpu(prompt_text: str, max_tokens: int, temperature: float) -> str:
78
+ inputs = tokenizer(prompt_text, return_tensors="pt")
79
  device = next(model.parameters()).device
80
  inputs = {k: v.to(device) for k, v in inputs.items()}
81
 
 
94
  return tokenizer.decode(new_tokens, skip_special_tokens=True).strip()
95
 
96
 
 
 
 
97
  class ChatMessage(BaseModel):
98
+ role: str
99
  content: str
100
 
 
101
  class GenerateRequest(BaseModel):
102
+ messages: List[ChatMessage] = Field(...)
103
  max_tokens: int = Field(default=1024, ge=1, le=4096)
104
  temperature: float = Field(default=0.7, ge=0.0, le=2.0)
105
 
 
106
  class GenerateResponse(BaseModel):
107
  response: str
108
  model: str = MODEL_ID
109
 
 
110
  app = Server(title="ZeroGPU Chat API")
111
 
112
  app.add_middleware(
 
116
  allow_headers=["*"],
117
  )
118
 
 
 
 
 
 
 
 
 
 
 
119
  @app.post("/api/generate", response_model=GenerateResponse)
120
  def api_generate_post(body: GenerateRequest):
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
121
  if not body.messages:
122
+ return {"response": "Error: messages kosong.", "model": MODEL_ID}
123
 
 
124
  formatted_messages = [{"role": m.role, "content": m.content} for m in body.messages]
125
+
126
+ # 1. Bangun format teks di CPU agar aman dari isolasi GPU
127
+ prompt_text = build_prompt_string(formatted_messages)
128
 
129
+ # 2. Kirim teks utuh ke GPU
130
+ text = generate_reply_gpu(
131
+ prompt_text=prompt_text,
132
  max_tokens=body.max_tokens,
133
  temperature=body.temperature,
134
  )
135
  return GenerateResponse(response=text)
136
 
 
137
  @app.get("/api/generate", response_model=GenerateResponse)
138
+ def api_generate_get(prompt: str, system_prompt: str = DEFAULT_SYSTEM_PROMPT, max_tokens: int = 1024, temperature: float = 0.7):
 
 
 
 
 
 
 
 
 
139
  prompt = (prompt or "").strip()
140
  if not prompt:
141
  return {"response": "Error: prompt tidak boleh kosong.", "model": MODEL_ID}
 
144
  {"role": "system", "content": system_prompt},
145
  {"role": "user", "content": prompt}
146
  ]
147
+ prompt_text = build_prompt_string(fallback_messages)
148
+ text = generate_reply_gpu(prompt_text=prompt_text, max_tokens=max_tokens, temperature=temperature)
 
 
 
 
149
  return GenerateResponse(response=text)
150
 
151
+ # UI Bawaan Tetap Dipertahankan
 
 
 
152
  @app.get("/", response_class=HTMLResponse)
153
  async def homepage():
154
  return """
 
174
  </head>
175
  <body>
176
  <h1>ZeroGPU Chat API (Multi-Turn)</h1>
 
 
177
  <div id="chat-box"></div>
 
178
  <label>Pesan Anda</label>
179
  <textarea id="prompt" rows="3" placeholder="Ketik pesan lalu tekan Kirim..."></textarea>
 
180
  <label>System Prompt (Mengubah ini akan mereset obrolan)</label>
181
  <input id="system_prompt" value="Anda adalah asisten AI yang membantu." onchange="resetChat()">
 
182
  <div class="actions">
183
  <button id="btn" onclick="kirim()">Kirim</button>
184
  <button style="background:#ff4d4d;" onclick="resetChat()">Reset Obrolan</button>
185
  </div>
 
186
  <script>
187
  let chatHistory = [];
 
188
  function initChat() {
189
  const sysPrompt = document.getElementById('system_prompt').value;
190
  chatHistory = [{ role: "system", content: sysPrompt }];
191
  renderChat();
192
  }
193
+ function resetChat() { initChat(); }
 
 
 
 
194
  function renderChat() {
195
  const box = document.getElementById('chat-box');
196
  box.innerHTML = '';
 
202
  });
203
  box.scrollTop = box.scrollHeight;
204
  }
 
205
  function escapeHTML(str) {
206
  return str.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
207
  }
 
208
  async function kirim() {
209
  const promptInput = document.getElementById('prompt');
210
  const btn = document.getElementById('btn');
211
  const userText = promptInput.value.trim();
 
212
  if (!userText) return;
213
  if (chatHistory.length === 0) initChat();
 
214
  chatHistory.push({ role: "user", content: userText });
215
  renderChat();
 
216
  promptInput.value = "";
217
  btn.disabled = true;
 
218
  const box = document.getElementById('chat-box');
219
  const loadingDiv = document.createElement('div');
220
  loadingDiv.className = 'msg assistant';
221
  loadingDiv.textContent = 'Sedang mengetik jawaban...';
222
  box.appendChild(loadingDiv);
223
  box.scrollTop = box.scrollHeight;
 
224
  try {
225
  const res = await fetch('/api/generate', {
226
  method: 'POST',
 
231
  temperature: 0.7
232
  })
233
  });
 
234
  const data = await res.json();
235
  box.removeChild(loadingDiv);
 
236
  chatHistory.push({ role: "assistant", content: data.response });
237
  renderChat();
238
  } catch (e) {
 
247
  btn.disabled = false;
248
  }
249
  }
 
 
250
  initChat();
251
  </script>
252
  </body>
253
  </html>
254
  """
255
 
 
256
  if __name__ == "__main__":
257
  app.launch(show_error=True)
258