Update app.py
Browse files
app.py
CHANGED
|
@@ -1,11 +1,74 @@
|
|
| 1 |
-
import
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
-
|
| 4 |
-
output = "/tmp/video.mp4" # HF Spaces usually allow writing in /tmp
|
| 5 |
|
| 6 |
-
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from fastapi import FastAPI, HTTPException
|
| 3 |
+
from pydantic import BaseModel
|
| 4 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 5 |
|
| 6 |
+
MODEL_ID = "Qwen/Qwen2.5-1.5B-Instruct"
|
|
|
|
| 7 |
|
| 8 |
+
app = FastAPI(
|
| 9 |
+
title="General AI REST API",
|
| 10 |
+
version="1.0.0"
|
| 11 |
+
)
|
| 12 |
+
|
| 13 |
+
# Load once at startup (critical for speed)
|
| 14 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
| 15 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 16 |
+
MODEL_ID,
|
| 17 |
+
device_map="cpu",
|
| 18 |
+
torch_dtype=torch.float32
|
| 19 |
+
)
|
| 20 |
+
model.eval()
|
| 21 |
+
|
| 22 |
+
# ----------- Schemas -----------
|
| 23 |
+
|
| 24 |
+
class ChatRequest(BaseModel):
|
| 25 |
+
prompt: str
|
| 26 |
+
max_tokens: int = 120
|
| 27 |
+
temperature: float = 0.35
|
| 28 |
+
top_p: float = 0.9
|
| 29 |
+
|
| 30 |
+
class ChatResponse(BaseModel):
|
| 31 |
+
reply: str
|
| 32 |
+
|
| 33 |
+
# ----------- Endpoints -----------
|
| 34 |
+
|
| 35 |
+
@app.get("/health")
|
| 36 |
+
def health():
|
| 37 |
+
return {"status": "ok"}
|
| 38 |
+
|
| 39 |
+
@app.post("/chat", response_model=ChatResponse)
|
| 40 |
+
def chat(req: ChatRequest):
|
| 41 |
+
if not req.prompt or len(req.prompt.strip()) == 0:
|
| 42 |
+
raise HTTPException(status_code=400, detail="Prompt is empty")
|
| 43 |
+
|
| 44 |
+
# Hard safety limits (CPU protection)
|
| 45 |
+
prompt = req.prompt[:500]
|
| 46 |
+
max_tokens = min(req.max_tokens, 150)
|
| 47 |
+
|
| 48 |
+
messages = [
|
| 49 |
+
{"role": "system", "content": "You are a concise, intelligent assistant. Answer clearly and briefly."},
|
| 50 |
+
{"role": "user", "content": prompt}
|
| 51 |
+
]
|
| 52 |
+
|
| 53 |
+
input_ids = tokenizer.apply_chat_template(
|
| 54 |
+
messages,
|
| 55 |
+
tokenize=True,
|
| 56 |
+
return_tensors="pt"
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
with torch.no_grad():
|
| 60 |
+
output = model.generate(
|
| 61 |
+
input_ids,
|
| 62 |
+
max_new_tokens=max_tokens,
|
| 63 |
+
temperature=req.temperature,
|
| 64 |
+
top_p=req.top_p,
|
| 65 |
+
repetition_penalty=1.1,
|
| 66 |
+
do_sample=True
|
| 67 |
+
)
|
| 68 |
+
|
| 69 |
+
reply = tokenizer.decode(
|
| 70 |
+
output[0][input_ids.shape[-1]:],
|
| 71 |
+
skip_special_tokens=True
|
| 72 |
+
).strip()
|
| 73 |
+
|
| 74 |
+
return {"reply": reply}
|