KernelPilot/KernelPilot-V1-Server2

1import os, tempfile, time2import gradio as gr3from tool.testv3 import run_autotune_pipeline4 5# ---------- Core callback ----------6def generate_kernel(text_input, n_iters, progress=gr.Progress()):7    """8    text_input : string from textbox (NL description or base CUDA code)9    file_input : gr.File upload object (or None)10    Returns   : (kernel_code_str, downloadable_file_path)11    """12    progress((0, n_iters), desc="Initializing...")13    # 1) Select input source14 15    if not text_input.strip():16        return "⚠️ Please paste a description or baseline CUDA code.", "", None17        18    td = tempfile.mkdtemp(prefix="auto_")19    src_path = os.path.join(td, f"input_{int(time.time())}.txt")20    with open(src_path, "w") as f:21        f.write(text_input)22 23    best_code = ""24    for info in run_autotune_pipeline(src_path, n_iters):25        # 1) update progress bar (if iteration known)26        if info["iteration"] is not None:27            # print(f"Iteration {info['iteration']} / {n_iters}: {info['message']}")28            progress((info["iteration"], n_iters), desc=info["message"])29 30        # 3) kernel output only when we get new code31        if info["code"]:32            best_code = info["code"]33 34 35    # last yield enables the download button36    return best_code37 38 39# ---------- Gradio UI ----------40with gr.Blocks(title="KernelPilot", theme=gr.themes.Soft(text_size="lg", font=[41        "system-ui",42        "-apple-system",43        "BlinkMacSystemFont",44        "Segoe UI",45        "Roboto",46        "Helvetica Neue",47        "Arial",48        "Noto Sans",49        "sans-serif"50    ])) as demo:51    gr.Markdown(52        """# 🚀 KernelPilot  53Enter a natural‑language description,  54then click **Generate** to obtain the kernel function."""55    )56 57    with gr.Row():58        txt_input = gr.Textbox(59            label="📝 Input",60            lines=10,61            placeholder="Describe the kernel",62            scale=363        )64        level = gr.Number(65            label="Optimization Level",66            minimum=1,67            maximum=5,68            value=2,69            step=1,70            scale=171        )72 73 74    gen_btn = gr.Button("⚡ Generate")75 76    kernel_output = gr.Code(77        label="🎯 Tuned CUDA Kernel",78        language="cpp"79    )80 81    gen_btn.click(82        fn=generate_kernel,83        inputs=[txt_input, level],84        outputs=[kernel_output],85        queue=True,               # keeps requests queued86        show_progress=True,  # show progress bar87        show_progress_on=kernel_output  # update log box with progress88    )89 90if __name__ == "__main__":91    demo.queue(default_concurrency_limit=1, max_size=50)92    demo.launch(server_name="0.0.0.0", server_port=7860)