app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "mistralai/Devstral-Small-2505"

# Use slow tokenizer to avoid sentencepiece errors
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)

# Model loading with dtype and device fallback
device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto" if torch.cuda.is_available() else None
)

def chat(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        do_sample=True,
        top_p=0.9,
        eos_token_id=tokenizer.eos_token_id
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

gr.Interface(
    fn=chat,
    inputs=gr.Textbox(lines=2, placeholder="Talk with Devstral..."),
    outputs="text",
    title="Devstral-Small Chat",
    theme="compact"
).launch()