app.py

import gradio as gr
import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextIteratorStreamer,
)
import os
from threading import Thread
import spaces
import time
import subprocess

print("\n=== Environment Setup ===")

if torch.cuda.is_available():
    print(f"GPU detected: {torch.cuda.get_device_name(0)}")
    try:
        subprocess.run(
            "pip install flash-attn --no-build-isolation",
            shell=True,
            check=True,
        )
        print("✅ flash-attn installed successfully")
    except subprocess.CalledProcessError as e:
        print("⚠️ flash-attn installation failed:", e)
else:
    print("⚙️ CPU detected — skipping flash-attn installation")
    # Disable flash-attn references safely
    os.environ["DISABLE_FLASH_ATTN"] = "1"
    os.environ["FLASH_ATTENTION_SKIP_CUDA_BUILD"] = "TRUE"
try:
    from transformers.utils import import_utils
    if "flash_attn" not in import_utils.PACKAGE_DISTRIBUTION_MAPPING:
        import_utils.PACKAGE_DISTRIBUTION_MAPPING["flash_attn"] = "flash-attn"
except Exception as e:
    print("⚠️ Patch skipped:", e)

    
token = os.environ["HF_TOKEN"]


model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-4",
    token=token,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16
)
tok = AutoTokenizer.from_pretrained("microsoft/phi-4", token=token)
terminators = [
    tok.eos_token_id,
]

if torch.cuda.is_available():
    device = torch.device("cuda")
    print(f"Using GPU: {torch.cuda.get_device_name(device)}")
else:
    device = torch.device("cpu")
    print("Using CPU")

model = model.to(device)
# Dispatch Errors


@spaces.GPU(duration=60)
def chat(message, history, temperature, do_sample, max_tokens):
    chat = []
    for item in history:
        chat.append({"role": "user", "content": item[0]})
        if item[1] is not None:
            chat.append({"role": "assistant", "content": item[1]})
    chat.append({"role": "user", "content": message})
    messages = tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
    model_inputs = tok([messages], return_tensors="pt").to(device)
    streamer = TextIteratorStreamer(
        tok, timeout=20.0, skip_prompt=True, skip_special_tokens=True
    )
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=temperature,
        eos_token_id=terminators,
    )

    if temperature == 0:
        generate_kwargs["do_sample"] = False

    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_text = ""
    for new_text in streamer:
        partial_text += new_text
        yield partial_text

    yield partial_text


demo = gr.ChatInterface(
    fn=chat,
    examples=[["Write me a poem about Machine Learning."]],
    # multimodal=False,
    additional_inputs_accordion=gr.Accordion(
        label="⚙️ Parameters", open=False, render=False
    ),
    additional_inputs=[
        gr.Slider(
            minimum=0, maximum=1, step=0.1, value=0.9, label="Temperature", render=False
        ),
        gr.Checkbox(label="Sampling", value=True),
        gr.Slider(
            minimum=128,
            maximum=4096,
            step=1,
            value=512,
            label="Max new tokens",
            render=False,
        ),
    ],
    stop_btn="Stop Generation",
    title="Chat With LLMs",
    description="Now Running [microsoft/phi-4](https://huggingface.co/microsoft/phi-4)",
)
demo.launch()