app.py

import spaces
import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from transformers import TextIteratorStreamer
from threading import Thread

import gradio as gr

text_generator = None
is_hugging_face = True
model_id = "AXCXEPT/phi-4-deepseek-R1K-RL-EZO"
model_id = "AXCXEPT/phi-4-open-R1-Distill-EZOv1"

huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
huggingface_token = None
device = "auto" # torch.device("cuda" if torch.cuda.is_available() else "cpu")
device = "cuda"
dtype = torch.bfloat16
dtype = torch.float16

if not huggingface_token:
    pass
    print("no HUGGINGFACE_TOKEN if you need set secret ")
    #raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")
        

    
    
    



tokenizer = AutoTokenizer.from_pretrained(model_id, token=huggingface_token)

print(model_id,device,dtype)
histories = []
#model = None



if not is_hugging_face:
    model = AutoModelForCausalLM.from_pretrained(
        model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
    )
    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer,torch_dtype=dtype,device_map=device,stream=True ) #pipeline has not to(device)
    
    if next(model.parameters()).is_cuda:
        print("The model is on a GPU")
    else:
        print("The model is on a CPU")

    #print(f"text_generator.device='{text_generator.device}")
    if str(text_generator.device).strip() == 'cuda':
        print("The pipeline is using a GPU")
    else:
        print("The pipeline is using a CPU")

print("initialized")


def generate_text(messages):
    if is_hugging_face:#need everytime initialize for ZeroGPU
        model = AutoModelForCausalLM.from_pretrained(
                model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
            )
        model.to(device)
    question = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    question = tokenizer(question, return_tensors="pt").to(device)

    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True)
    generation_kwargs = dict(question, streamer=streamer, max_new_tokens=200)
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    
    generated_output = ""
    thread.start()
    for new_text in streamer:
        generated_output += new_text
        yield generated_output            
generate_text.zerogpu = True
    
       

@spaces.GPU(duration=60)
def call_generate_text(message, history):
   # history.append({"role": "user", "content": message})
    #print(message)
    #print(history)
   
    messages = history+[{"role":"user","content":message}]
    try:
        
        for text in generate_text(messages):
            yield text
    except RuntimeError  as e:
        print(f"An unexpected error occurred: {e}")
        yield ""

demo = gr.ChatInterface(call_generate_text,type="messages")

if __name__ == "__main__":
    demo.launch()