app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Modellname für die kleinere Variante
model_name = "Qwen/Qwen2.5-Math-1.5B-Instruct"

# Überprüfen, ob eine GPU verfügbar ist
device = "cuda" if torch.cuda.is_available() else "cpu"

# Modell und Tokenizer laden
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",  # Modell auf verfügbare Geräte verteilen
    low_cpu_mem_usage=True,  # Versucht, den Speicherverbrauch zu reduzieren
    trust_remote_code=True,
    torch_dtype=torch.float16  # Reduziert den Speicherverbrauch
).to(device).eval()

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

# Eingabe für das Gesprächsmodell erstellen
chat = [
    {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{}."},
    {"role": "user", "content": "Janet’s ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?"}
]

# Vorbereitung des Eingabeformats
conversation_str = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=False)

# Tokenisierung der Eingabe
input_ids = tokenizer.encode(conversation_str, return_tensors="pt", add_special_tokens=False).to(device)

# Inferenz durchführen
with torch.no_grad():
    outputs = model.generate(input_ids=input_ids, max_length=256, num_return_sequences=1)

# Ausgabe anzeigen
print(tokenizer.decode(outputs[0], skip_special_tokens=True))