Spaces:
Sleeping
Sleeping
1.0
Browse files
app.py
CHANGED
|
@@ -2,114 +2,36 @@ from flask import Flask, request, jsonify
|
|
| 2 |
from flask_cors import CORS
|
| 3 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 4 |
import torch
|
| 5 |
-
import os
|
| 6 |
-
|
| 7 |
-
# --- DNS FIX ---
|
| 8 |
-
try:
|
| 9 |
-
with open("/etc/resolv.conf", "w") as f:
|
| 10 |
-
f.write("nameserver 8.8.8.8\nnameserver 8.8.4.4\n")
|
| 11 |
-
print("✅ DNS gesetzt.")
|
| 12 |
-
except Exception as e:
|
| 13 |
-
print(f"⚠️ DNS Fix skipped: {e}")
|
| 14 |
|
| 15 |
app = Flask(__name__)
|
| 16 |
CORS(app, origins="*")
|
| 17 |
|
| 18 |
-
print("🚀 Starte Nano-AI...")
|
| 19 |
-
|
| 20 |
-
# MODELL - Qwen2.5 ist stabiler auf HF Spaces
|
| 21 |
MODEL_NAME = "Qwen/Qwen2.5-2B-Instruct"
|
| 22 |
-
|
| 23 |
tokenizer = None
|
| 24 |
model = None
|
| 25 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 26 |
|
| 27 |
-
print(f"💻 Hardware: {device.upper()}")
|
| 28 |
-
|
| 29 |
try:
|
| 30 |
-
print(f"⏳ Lade: {MODEL_NAME} ...")
|
| 31 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
)
|
| 40 |
-
print("✅ Geladen auf GPU (float16)")
|
| 41 |
-
else:
|
| 42 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 43 |
-
MODEL_NAME,
|
| 44 |
-
torch_dtype=torch.float32,
|
| 45 |
-
device_map="auto",
|
| 46 |
-
low_cpu_mem_usage=True
|
| 47 |
-
)
|
| 48 |
-
print("✅ Geladen auf CPU (float32)")
|
| 49 |
-
|
| 50 |
-
print("🟢 Nano-AI ist BEREIT!")
|
| 51 |
except Exception as e:
|
| 52 |
-
print(f"❌
|
| 53 |
-
model = None
|
| 54 |
|
| 55 |
@app.route('/api/generate', methods=['POST'])
|
| 56 |
def generate():
|
| 57 |
if model is None:
|
| 58 |
-
return jsonify({"text": "⚠️ Modell lädt noch.
|
| 59 |
-
|
| 60 |
data = request.json
|
| 61 |
user_prompt = data.get('prompt', '')
|
| 62 |
|
| 63 |
if not user_prompt:
|
| 64 |
return jsonify({"text": "❌ Bitte stelle eine Frage."}), 400
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
full_prompt = (
|
| 68 |
-
f"<|im_start|>system\nDu bist Nano-AI, ein hilfreicher Assistent.<|im_end|>\n"
|
| 69 |
-
f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
|
| 70 |
-
f"<|im_start|>assistant\n"
|
| 71 |
-
)
|
| 72 |
-
|
| 73 |
-
try:
|
| 74 |
-
input_ids = tokenizer.encode(full_prompt, return_tensors="pt").to(model.device)
|
| 75 |
-
|
| 76 |
-
outputs = model.generate(
|
| 77 |
-
input_ids,
|
| 78 |
-
max_new_tokens=256,
|
| 79 |
-
temperature=0.7,
|
| 80 |
-
do_sample=True,
|
| 81 |
-
pad_token_id=tokenizer.eos_token_id,
|
| 82 |
-
repetition_penalty=1.1,
|
| 83 |
-
top_p=0.9
|
| 84 |
-
)
|
| 85 |
-
|
| 86 |
-
full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 87 |
-
|
| 88 |
-
# ANTWORT EXTRAHIEREN
|
| 89 |
-
if "<|im_start|>assistant\n" in full_response:
|
| 90 |
-
answer = full_response.split("<|im_start|>assistant\n")[-1].strip()
|
| 91 |
-
else:
|
| 92 |
-
answer = full_response.replace(full_prompt, "").strip()
|
| 93 |
-
|
| 94 |
-
# Aufräumen
|
| 95 |
-
if "<|im_start|>" in answer:
|
| 96 |
-
answer = answer.split("<|im_start|>")[0].strip()
|
| 97 |
-
if "<|im_end|>" in answer:
|
| 98 |
-
answer = answer.split("<|im_end|>")[0].strip()
|
| 99 |
-
|
| 100 |
-
if not answer or len(answer) < 2:
|
| 101 |
-
answer = "Ich habe deine Frage verstanden, kann aber gerade keine detaillierte Antwort geben. Versuche es mit einer anderen Frage!"
|
| 102 |
-
|
| 103 |
-
print(f"✅ Antwort generiert: {len(answer)} Zeichen")
|
| 104 |
-
return jsonify({"text": answer})
|
| 105 |
-
|
| 106 |
-
except Exception as e:
|
| 107 |
-
print(f"❌ Error: {e}")
|
| 108 |
-
return jsonify({"text": f"⚠️ Fehler: {str(e)[:100]}"}), 500
|
| 109 |
-
|
| 110 |
-
@app.route('/health', methods=['GET'])
|
| 111 |
-
def health():
|
| 112 |
-
return jsonify({"status": "ok", "model_loaded": model is not None})
|
| 113 |
-
|
| 114 |
-
if __name__ == '__main__':
|
| 115 |
-
app.run(host='0.0.0.0', port=7860)
|
|
|
|
| 2 |
from flask_cors import CORS
|
| 3 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 4 |
import torch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
app = Flask(__name__)
|
| 7 |
CORS(app, origins="*")
|
| 8 |
|
|
|
|
|
|
|
|
|
|
| 9 |
MODEL_NAME = "Qwen/Qwen2.5-2B-Instruct"
|
|
|
|
| 10 |
tokenizer = None
|
| 11 |
model = None
|
| 12 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 13 |
|
|
|
|
|
|
|
| 14 |
try:
|
|
|
|
| 15 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
| 16 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 17 |
+
MODEL_NAME,
|
| 18 |
+
torch_dtype=torch.float16 if device == "cuda" else torch.float32,
|
| 19 |
+
device_map="auto",
|
| 20 |
+
low_cpu_mem_usage=True
|
| 21 |
+
)
|
| 22 |
+
print("🟢 Model loaded!")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
except Exception as e:
|
| 24 |
+
print(f"❌ Error: {e}")
|
|
|
|
| 25 |
|
| 26 |
@app.route('/api/generate', methods=['POST'])
|
| 27 |
def generate():
|
| 28 |
if model is None:
|
| 29 |
+
return jsonify({"text": "⚠️ Modell lädt noch..."}), 503
|
| 30 |
+
|
| 31 |
data = request.json
|
| 32 |
user_prompt = data.get('prompt', '')
|
| 33 |
|
| 34 |
if not user_prompt:
|
| 35 |
return jsonify({"text": "❌ Bitte stelle eine Frage."}), 400
|
| 36 |
+
|
| 37 |
+
full_prompt = f"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|