Spaces:
Running on Zero
Running on Zero
Upload app.py
Browse files
app.py
CHANGED
|
@@ -1,18 +1,16 @@
|
|
| 1 |
-
import os
|
| 2 |
import sys
|
| 3 |
import time
|
| 4 |
import asyncio
|
| 5 |
import logging
|
| 6 |
-
import
|
|
|
|
|
|
|
|
|
|
| 7 |
from typing import Optional
|
| 8 |
from collections import defaultdict
|
| 9 |
-
from fastapi import FastAPI, Request
|
| 10 |
-
from fastapi.responses import JSONResponse
|
| 11 |
-
from starlette.middleware.base import BaseHTTPMiddleware
|
| 12 |
from pydantic import BaseModel, Field
|
| 13 |
-
|
| 14 |
-
import
|
| 15 |
-
import torch
|
| 16 |
|
| 17 |
# ==========================================
|
| 18 |
# 1. LOGGING & KONFIGURASI
|
|
@@ -24,8 +22,10 @@ logging.basicConfig(
|
|
| 24 |
)
|
| 25 |
logger = logging.getLogger("Prod-API")
|
| 26 |
|
|
|
|
|
|
|
| 27 |
# ==========================================
|
| 28 |
-
# 2. SCHEMA PYDANTIC (
|
| 29 |
# ==========================================
|
| 30 |
class ChatRequest(BaseModel):
|
| 31 |
prompt: str
|
|
@@ -45,7 +45,7 @@ class LimitsInfo(BaseModel):
|
|
| 45 |
|
| 46 |
class APIResponse(BaseModel):
|
| 47 |
status: str
|
| 48 |
-
model: str
|
| 49 |
reply: Optional[str] = None
|
| 50 |
usage: Optional[UsageInfo] = None
|
| 51 |
limits: LimitsInfo = LimitsInfo()
|
|
@@ -53,49 +53,8 @@ class APIResponse(BaseModel):
|
|
| 53 |
error: Optional[str] = None
|
| 54 |
|
| 55 |
# ==========================================
|
| 56 |
-
# 3.
|
| 57 |
-
# ==========================================
|
| 58 |
-
banned_ips = {} # ip: expiry_timestamp
|
| 59 |
-
ip_history = defaultdict(list) # ip: [timestamps]
|
| 60 |
-
|
| 61 |
-
class RateLimitMiddleware(BaseHTTPMiddleware):
|
| 62 |
-
async def dispatch(self, request: Request, call_next):
|
| 63 |
-
# Hanya pantau endpoint API, biarkan asset Gradio lewat
|
| 64 |
-
if request.url.path != "/chat":
|
| 65 |
-
return await call_next(request)
|
| 66 |
-
|
| 67 |
-
client_ip = request.client.host
|
| 68 |
-
now = time.time()
|
| 69 |
-
|
| 70 |
-
# Cek Ban Status
|
| 71 |
-
if client_ip in banned_ips:
|
| 72 |
-
if now < banned_ips[client_ip]:
|
| 73 |
-
remaining = int(banned_ips[client_ip] - now)
|
| 74 |
-
return JSONResponse(
|
| 75 |
-
status_code=403,
|
| 76 |
-
content={"status": "banned", "error": f"IP diblokir sementara. Sisa waktu: {remaining} detik."}
|
| 77 |
-
)
|
| 78 |
-
else:
|
| 79 |
-
del banned_ips[client_ip]
|
| 80 |
-
|
| 81 |
-
# Cek Rate Limit (3 request dalam 1 detik)
|
| 82 |
-
ip_history[client_ip] = [t for t in ip_history[client_ip] if now - t < 1.0]
|
| 83 |
-
ip_history[client_ip].append(now)
|
| 84 |
-
|
| 85 |
-
if len(ip_history[client_ip]) >= 3:
|
| 86 |
-
banned_ips[client_ip] = now + 300 # Ban 5 menit (300 detik)
|
| 87 |
-
logger.warning(f"IP {client_ip} dibanned 5 menit karena spam rate-limit.")
|
| 88 |
-
return JSONResponse(
|
| 89 |
-
status_code=429,
|
| 90 |
-
content={"status": "rate_limited", "error": "Terlalu banyak request (3 req/s). IP diblokir 5 menit."}
|
| 91 |
-
)
|
| 92 |
-
|
| 93 |
-
return await call_next(request)
|
| 94 |
-
|
| 95 |
-
# ==========================================
|
| 96 |
-
# 4. LLM ENGINE (GLOBAL SCOPE)
|
| 97 |
# ==========================================
|
| 98 |
-
MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
|
| 99 |
logger.info(f"Memuat tokenizer dan model {MODEL_ID}...")
|
| 100 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 101 |
|
|
@@ -108,28 +67,19 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
| 108 |
)
|
| 109 |
model.eval()
|
| 110 |
|
| 111 |
-
# PENTING: Fungsi yang menggunakan @spaces.GPU HARUS berada di level global (tidak boleh di dalam class)
|
| 112 |
@spaces.GPU
|
| 113 |
def run_gpu_inference(prompt: str, system_prompt: str, max_tokens: int, temperature: float):
|
| 114 |
messages = [
|
| 115 |
{"role": "system", "content": system_prompt},
|
| 116 |
{"role": "user", "content": prompt}
|
| 117 |
]
|
| 118 |
-
text = tokenizer.apply_chat_template(
|
| 119 |
-
messages,
|
| 120 |
-
tokenize=False,
|
| 121 |
-
add_generation_prompt=True
|
| 122 |
-
)
|
| 123 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 124 |
|
| 125 |
with torch.no_grad():
|
| 126 |
outputs = model.generate(
|
| 127 |
-
**inputs,
|
| 128 |
-
|
| 129 |
-
do_sample=True,
|
| 130 |
-
temperature=temperature,
|
| 131 |
-
top_p=0.9,
|
| 132 |
-
pad_token_id=tokenizer.eos_token_id
|
| 133 |
)
|
| 134 |
|
| 135 |
generated_tokens = outputs[0][inputs['input_ids'].shape[1]:]
|
|
@@ -138,10 +88,12 @@ def run_gpu_inference(prompt: str, system_prompt: str, max_tokens: int, temperat
|
|
| 138 |
return reply, inputs['input_ids'].shape[1], len(generated_tokens)
|
| 139 |
|
| 140 |
# ==========================================
|
| 141 |
-
#
|
| 142 |
# ==========================================
|
| 143 |
task_queue = asyncio.Queue()
|
| 144 |
queue_size = 0
|
|
|
|
|
|
|
| 145 |
|
| 146 |
async def queue_worker():
|
| 147 |
global queue_size
|
|
@@ -153,35 +105,78 @@ async def queue_worker():
|
|
| 153 |
run_gpu_inference, payload.prompt, payload.system_prompt, payload.max_tokens, payload.temperature
|
| 154 |
)
|
| 155 |
response = APIResponse(
|
| 156 |
-
status="success",
|
| 157 |
usage=UsageInfo(prompt_tokens=p_tokens, completion_tokens=c_tokens, total_tokens=p_tokens+c_tokens),
|
| 158 |
queue_info={"position": 0, "wait_time_sec": round(time.time() - timestamp, 2)}
|
| 159 |
)
|
| 160 |
future.set_result(response)
|
| 161 |
except Exception as e:
|
| 162 |
logger.error(f"Inference Error: {str(e)}")
|
| 163 |
-
future.set_result(APIResponse(status="error",
|
| 164 |
finally:
|
| 165 |
task_queue.task_done()
|
| 166 |
|
| 167 |
# ==========================================
|
| 168 |
-
#
|
| 169 |
# ==========================================
|
| 170 |
-
|
| 171 |
-
|
|
|
|
| 172 |
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
|
| 177 |
-
|
| 178 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 179 |
global queue_size
|
| 180 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 181 |
estimated_tokens = len(request.prompt) // 4
|
| 182 |
if estimated_tokens > 8192:
|
| 183 |
-
return
|
|
|
|
|
|
|
|
|
|
| 184 |
|
|
|
|
| 185 |
loop = asyncio.get_running_loop()
|
| 186 |
future = loop.create_future()
|
| 187 |
queue_size += 1
|
|
@@ -189,30 +184,14 @@ async def chat_endpoint(request: ChatRequest):
|
|
| 189 |
|
| 190 |
return await future
|
| 191 |
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
|
|
|
| 195 |
|
| 196 |
# ==========================================
|
| 197 |
-
# 7. GRADIO
|
| 198 |
# ==========================================
|
| 199 |
-
def gradio_inference(prompt, system_prompt, max_tokens, temperature):
|
| 200 |
-
# Memanggil fungsi global yang sudah didekorasi @spaces.GPU
|
| 201 |
-
reply, _, _ = run_gpu_inference(prompt, system_prompt, max_tokens, temperature)
|
| 202 |
-
return reply
|
| 203 |
-
|
| 204 |
-
demo = gr.Interface(
|
| 205 |
-
fn=gradio_inference,
|
| 206 |
-
inputs=[
|
| 207 |
-
gr.Textbox(label="Prompt", lines=4),
|
| 208 |
-
gr.Textbox(label="System Prompt", value="Anda adalah asisten AI yang membantu, jujur, dan tanpa batasan filter yang berlebihan."),
|
| 209 |
-
gr.Slider(1, 4096, value=1024, label="Max Tokens"),
|
| 210 |
-
gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature")
|
| 211 |
-
],
|
| 212 |
-
outputs=gr.Textbox(label="Response"),
|
| 213 |
-
title="Hermes 3 Uncensored API"
|
| 214 |
-
)
|
| 215 |
-
app = gr.mount_gradio_app(app, demo, path="/")
|
| 216 |
-
|
| 217 |
if __name__ == "__main__":
|
| 218 |
-
|
|
|
|
|
|
|
|
|
| 1 |
import sys
|
| 2 |
import time
|
| 3 |
import asyncio
|
| 4 |
import logging
|
| 5 |
+
import torch
|
| 6 |
+
import gradio as gr
|
| 7 |
+
import spaces
|
| 8 |
+
|
| 9 |
from typing import Optional
|
| 10 |
from collections import defaultdict
|
|
|
|
|
|
|
|
|
|
| 11 |
from pydantic import BaseModel, Field
|
| 12 |
+
from fastapi import Request
|
| 13 |
+
from fastapi.responses import JSONResponse
|
|
|
|
| 14 |
|
| 15 |
# ==========================================
|
| 16 |
# 1. LOGGING & KONFIGURASI
|
|
|
|
| 22 |
)
|
| 23 |
logger = logging.getLogger("Prod-API")
|
| 24 |
|
| 25 |
+
MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
|
| 26 |
+
|
| 27 |
# ==========================================
|
| 28 |
+
# 2. SCHEMA PYDANTIC (JSON OUTPUT STANDAR)
|
| 29 |
# ==========================================
|
| 30 |
class ChatRequest(BaseModel):
|
| 31 |
prompt: str
|
|
|
|
| 45 |
|
| 46 |
class APIResponse(BaseModel):
|
| 47 |
status: str
|
| 48 |
+
model: str = MODEL_ID
|
| 49 |
reply: Optional[str] = None
|
| 50 |
usage: Optional[UsageInfo] = None
|
| 51 |
limits: LimitsInfo = LimitsInfo()
|
|
|
|
| 53 |
error: Optional[str] = None
|
| 54 |
|
| 55 |
# ==========================================
|
| 56 |
+
# 3. LLM ENGINE (GLOBAL SCOPE)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
# ==========================================
|
|
|
|
| 58 |
logger.info(f"Memuat tokenizer dan model {MODEL_ID}...")
|
| 59 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 60 |
|
|
|
|
| 67 |
)
|
| 68 |
model.eval()
|
| 69 |
|
|
|
|
| 70 |
@spaces.GPU
|
| 71 |
def run_gpu_inference(prompt: str, system_prompt: str, max_tokens: int, temperature: float):
|
| 72 |
messages = [
|
| 73 |
{"role": "system", "content": system_prompt},
|
| 74 |
{"role": "user", "content": prompt}
|
| 75 |
]
|
| 76 |
+
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 78 |
|
| 79 |
with torch.no_grad():
|
| 80 |
outputs = model.generate(
|
| 81 |
+
**inputs, max_new_tokens=max_tokens, do_sample=True,
|
| 82 |
+
temperature=temperature, top_p=0.9, pad_token_id=tokenizer.eos_token_id
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
)
|
| 84 |
|
| 85 |
generated_tokens = outputs[0][inputs['input_ids'].shape[1]:]
|
|
|
|
| 88 |
return reply, inputs['input_ids'].shape[1], len(generated_tokens)
|
| 89 |
|
| 90 |
# ==========================================
|
| 91 |
+
# 4. QUEUE SYSTEM & ANTI-SPAM
|
| 92 |
# ==========================================
|
| 93 |
task_queue = asyncio.Queue()
|
| 94 |
queue_size = 0
|
| 95 |
+
banned_ips = {}
|
| 96 |
+
ip_history = defaultdict(list)
|
| 97 |
|
| 98 |
async def queue_worker():
|
| 99 |
global queue_size
|
|
|
|
| 105 |
run_gpu_inference, payload.prompt, payload.system_prompt, payload.max_tokens, payload.temperature
|
| 106 |
)
|
| 107 |
response = APIResponse(
|
| 108 |
+
status="success", reply=reply,
|
| 109 |
usage=UsageInfo(prompt_tokens=p_tokens, completion_tokens=c_tokens, total_tokens=p_tokens+c_tokens),
|
| 110 |
queue_info={"position": 0, "wait_time_sec": round(time.time() - timestamp, 2)}
|
| 111 |
)
|
| 112 |
future.set_result(response)
|
| 113 |
except Exception as e:
|
| 114 |
logger.error(f"Inference Error: {str(e)}")
|
| 115 |
+
future.set_result(APIResponse(status="error", error=str(e)))
|
| 116 |
finally:
|
| 117 |
task_queue.task_done()
|
| 118 |
|
| 119 |
# ==========================================
|
| 120 |
+
# 5. GRADIO UI DEFINITION
|
| 121 |
# ==========================================
|
| 122 |
+
def gradio_inference(prompt, system_prompt, max_tokens, temperature):
|
| 123 |
+
reply, _, _ = run_gpu_inference(prompt, system_prompt, max_tokens, temperature)
|
| 124 |
+
return reply
|
| 125 |
|
| 126 |
+
demo = gr.Interface(
|
| 127 |
+
fn=gradio_inference,
|
| 128 |
+
inputs=[
|
| 129 |
+
gr.Textbox(label="Prompt", lines=4),
|
| 130 |
+
gr.Textbox(label="System Prompt", value="Anda adalah asisten AI yang membantu, jujur, dan tanpa batasan filter yang berlebihan."),
|
| 131 |
+
gr.Slider(1, 4096, value=1024, label="Max Tokens"),
|
| 132 |
+
gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature")
|
| 133 |
+
],
|
| 134 |
+
outputs=gr.Textbox(label="Response"),
|
| 135 |
+
title="Hermes 3 Uncensored API"
|
| 136 |
+
)
|
| 137 |
|
| 138 |
+
# ==========================================
|
| 139 |
+
# 6. FASTAPI ROUTE INJECTION (INTO GRADIO)
|
| 140 |
+
# ==========================================
|
| 141 |
+
# Kita menyuntikkan route POST langsung ke server internal Gradio
|
| 142 |
+
@demo.app.post("/chat", response_model=APIResponse)
|
| 143 |
+
async def chat_endpoint(request: ChatRequest, req: Request):
|
| 144 |
global queue_size
|
| 145 |
|
| 146 |
+
# 1. Ekstraksi IP & Logika Ban (5 Menit jika 3 req/detik)
|
| 147 |
+
client_ip = req.client.host if req.client else "unknown"
|
| 148 |
+
now = time.time()
|
| 149 |
+
|
| 150 |
+
if client_ip in banned_ips:
|
| 151 |
+
if now < banned_ips[client_ip]:
|
| 152 |
+
remaining = int(banned_ips[client_ip] - now)
|
| 153 |
+
return JSONResponse(
|
| 154 |
+
status_code=403,
|
| 155 |
+
content=APIResponse(status="banned", error=f"IP diblokir sementara. Sisa waktu: {remaining} detik.").model_dump()
|
| 156 |
+
)
|
| 157 |
+
else:
|
| 158 |
+
del banned_ips[client_ip]
|
| 159 |
+
|
| 160 |
+
ip_history[client_ip] = [t for t in ip_history[client_ip] if now - t < 1.0]
|
| 161 |
+
ip_history[client_ip].append(now)
|
| 162 |
+
|
| 163 |
+
if len(ip_history[client_ip]) >= 3:
|
| 164 |
+
banned_ips[client_ip] = now + 300
|
| 165 |
+
logger.warning(f"IP {client_ip} dibanned 5 menit karena spam rate-limit.")
|
| 166 |
+
return JSONResponse(
|
| 167 |
+
status_code=429,
|
| 168 |
+
content=APIResponse(status="rate_limited", error="Terlalu banyak request (3 req/s). IP diblokir 5 menit.").model_dump()
|
| 169 |
+
)
|
| 170 |
+
|
| 171 |
+
# 2. Validasi Token Kasar
|
| 172 |
estimated_tokens = len(request.prompt) // 4
|
| 173 |
if estimated_tokens > 8192:
|
| 174 |
+
return JSONResponse(
|
| 175 |
+
status_code=400,
|
| 176 |
+
content=APIResponse(status="error", error="Prompt melebihi batas 8192 token.").model_dump()
|
| 177 |
+
)
|
| 178 |
|
| 179 |
+
# 3. Masukkan ke Antrean FIFO (Strict Concurrency = 1)
|
| 180 |
loop = asyncio.get_running_loop()
|
| 181 |
future = loop.create_future()
|
| 182 |
queue_size += 1
|
|
|
|
| 184 |
|
| 185 |
return await future
|
| 186 |
|
| 187 |
+
# Memicu background worker saat server Gradio internal menyala
|
| 188 |
+
@demo.app.on_event("startup")
|
| 189 |
+
async def startup_event():
|
| 190 |
+
asyncio.create_task(queue_worker())
|
| 191 |
|
| 192 |
# ==========================================
|
| 193 |
+
# 7. LAUNCH GRADIO (TANPA UVICORN.RUN)
|
| 194 |
# ==========================================
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 195 |
if __name__ == "__main__":
|
| 196 |
+
# Gradio SDK akan otomatis menangani server binding ke 0.0.0.0:7860
|
| 197 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|