app.py

"""Python to Rust / C++ performance translator.



An LLM ports a Python program to a compiled language; both are then executed

under the same sandbox so the speedup and - just as importantly - the

correctness of the translation can be measured rather than assumed.



Run locally:   python app.py

Deployed:      see Dockerfile (Hugging Face Spaces, Docker SDK)

"""

from __future__ import annotations

import html
import os
import re
import uuid

# Gradio 6 defaults to server-side rendering, which fronts the Python app with
# a Node proxy. On Spaces that proxy shuts down immediately after startup and
# takes the app with it ("Stopping Node.js server..." then RUNTIME_ERROR).
# Nothing here needs SSR - it is a single interactive page, not a crawlable
# site - so turn it off before gradio is imported and reads the setting.
#
# Assigned, not setdefault: Spaces sets this variable itself, so setdefault
# silently kept the platform's value and the proxy stayed on.
os.environ["GRADIO_SSR_MODE"] = "false"

import gradio as gr  # noqa: E402
from dotenv import load_dotenv  # noqa: E402

# override=False so real environment variables win over a stray .env file.
# On a host, secrets arrive as env vars and must not be shadowed by a file.
load_dotenv(override=False)

# Hugging Face Spaces sets SPACE_ID on every Space. Detecting it beats relying
# on a manually-set variable: the Gradio SDK has no Dockerfile to carry ENV, so
# a forgotten setting would silently run the public demo in local mode -
# frontier models ungated against the owner's key, no rate limit, and the
# on-disk-secret check disabled. This must happen before providers/sandbox are
# imported, since both read the flag at import time.
ON_SPACES = bool(os.getenv("SPACE_ID"))
if ON_SPACES:
    os.environ["PUBLIC_DEPLOYMENT"] = "1"

    # ZeroGPU refuses to start a Space that declares no GPU function, failing
    # with "No @spaces.GPU function detected during startup". This app is
    # CPU-only by nature - it needs compilers, not a GPU - and the honest fix
    # would be CPU Basic hardware, but downgrading an existing ZeroGPU Space
    # requires a PRO subscription.
    #
    # Declaring a function that is never wired to any event satisfies the check
    # at no cost: ZeroGPU allocates hardware on invocation, so one that is never
    # called consumes no GPU time. Guarded by ImportError because the `spaces`
    # package only exists on the platform, not in local development.
    try:
        import spaces

        @spaces.GPU(duration=1)
        def _zerogpu_startup_marker():  # never called; presence is the point
            return None

    except ImportError:
        pass

import providers  # noqa: E402  - imported after load_dotenv so it sees the keys
import sandbox  # noqa: E402
from examples import DEFAULT_EXAMPLE, EXAMPLES  # noqa: E402
from styles import CSS  # noqa: E402
from system_info import retrieve_system_info  # noqa: E402

sandbox.assert_safe_to_deploy()

REPO_URL = "https://github.com/Yugjohri/python-to-rust-cpp"

SYSTEM_INFO = retrieve_system_info()
TOOLCHAIN = sandbox.toolchain_status()
COMPILABLE = sandbox.available_languages()


# --------------------------------------------------------------------------
# prompting
# --------------------------------------------------------------------------


def build_system_prompt(language_key: str) -> str:
    lang = sandbox.LANGUAGES[language_key]
    integer_note = (
        "Rust integers are fixed width and overflow panics in debug but wraps in "
        "release - pick widths (i64/i128/u64) that cannot overflow for these inputs."
        if language_key == "rust" else
        "C++ integers are fixed width and signed overflow is undefined behaviour - "
        "pick widths (long long / __int128) that cannot overflow for these inputs."
    )
    return f"""You convert Python programs into high-performance {lang.display}.



Rules:

- Respond with {lang.display} source only. No prose, no markdown fences.

- A single self-contained file using only the standard library.

- Output must be byte-for-byte identical to the Python program's, including

  number formatting and wording. Format floats with the same precision.

- Preserve the algorithm. Do not substitute an asymptotically better one - the

  point is to measure what the same work costs in a compiled language.

- {integer_note}

- Python integers are arbitrary precision; account for that where it matters.

- Time the computation the same way the Python does and print it identically.

"""


def build_user_prompt(python_code: str, language_key: str) -> str:
    lang = sandbox.LANGUAGES[language_key]
    cpu = SYSTEM_INFO.get("cpu", {})
    compiler = lang.find_compiler() or lang.compilers[0]
    flags = " ".join(
        sandbox.build_compile_command(lang, "CC", f"main.{lang.extension}", "program")[1:]
    )
    # The exact toolchain version matters. Without it models reach for
    # extensions like unsigned __int128, which GCC supports on 64-bit Linux
    # but MinGW does not - a compile failure the model could have avoided.
    version = TOOLCHAIN.get(language_key, "") or "version unknown"
    os_name = SYSTEM_INFO.get("os", {}).get("system", "")

    constraints = ""
    if language_key == "cpp":
        if sandbox.cpp_has_int128():
            constraints = "- `__int128` IS available here if you need a wider accumulator.\n"
        else:
            constraints = (
                "- `__int128` / `unsigned __int128` are NOT available with this "
                "compiler. Using them will fail to compile. Use `long long` / "
                "`unsigned long long`, or restructure to avoid needing 128 bits.\n"
            )

    return f"""Port this Python program to {lang.display}.



Target machine: {cpu.get('brand', 'unknown CPU')}, {cpu.get('cores_logical', '?')} logical cores, {os_name}.

Compiler present: {version}

It will be compiled with exactly:

    {os.path.basename(compiler)} {flags}



Hard constraints for this toolchain:

{constraints}- Use only what that compiler version supports on this platform.



Respond with {lang.display} source only.



```python

{python_code}

```

"""


# --------------------------------------------------------------------------
# helpers
# --------------------------------------------------------------------------

FENCE = re.compile(r"^\s*```[a-zA-Z+]*\s*\n(.*?)\n?\s*```\s*$", re.DOTALL)
TIME_LINE = re.compile(r"^\s*execution time:.*$", re.IGNORECASE | re.MULTILINE)


def strip_fences(text: str) -> str:
    """Remove a surrounding markdown code fence if the model added one."""
    match = FENCE.match(text.strip())
    if match:
        return match.group(1)
    for token in ("```rust", "```rs", "```cpp", "```c++", "```"):
        text = text.replace(token, "")
    return text.strip()


def extract_seconds(output: str) -> float | None:
    """Pull the self-reported execution time out of a program's output."""
    match = re.search(r"execution time:\s*([0-9.]+)", output or "", re.IGNORECASE)
    if match:
        try:
            return float(match.group(1))
        except ValueError:
            return None
    return None


def comparable(output: str) -> str:
    """Output with the timing line removed, for correctness comparison.



    Timings legitimately differ between runs; everything else must not.

    """
    return TIME_LINE.sub("", output or "").strip()


def format_duration(seconds: float) -> tuple[str, str]:
    """Split the magnitude from the unit.



    Returned separately so the UI can right-align the digits in their own

    column and left-align the unit in another. Formatting them as one string

    and right-aligning that makes the decimal points jitter between rows,

    because "s" and "ms" are different widths.

    """
    if seconds >= 1:
        return f"{seconds:.3f}", "s"
    if seconds >= 1e-3:
        return f"{seconds * 1e3:.2f}", "ms"
    return f"{seconds * 1e6:.0f}", "us"


def idle_verdict(language_key: str = sandbox.DEFAULT_LANGUAGE) -> str:
    lang = sandbox.LANGUAGES[language_key]
    return (
        f'<div class="verdict"><p class="idle">Run the Python, port it, then run '
        f"the {lang.display} - the speedup and a correctness check appear here.</p></div>"
    )


def render_verdict(py_out: str, target_out: str, language_key: str) -> str:
    """Build the hero panel: speedup, timing bars, correctness badge."""
    lang = sandbox.LANGUAGES[language_key]
    py_seconds = extract_seconds(py_out)
    target_seconds = extract_seconds(target_out)

    if py_seconds is None or target_seconds is None or target_seconds <= 0:
        return idle_verdict(language_key)

    ratio = py_seconds / target_seconds
    speedup = f"{ratio:,.0f}x" if ratio >= 10 else f"{ratio:.1f}x"

    widest = max(py_seconds, target_seconds)
    py_width = max(2.0, py_seconds / widest * 100)
    target_width = max(2.0, target_seconds / widest * 100)

    py_body, target_body = comparable(py_out), comparable(target_out)
    if not py_body or not target_body:
        badge = '<span class="badge unknown">Run both to verify output</span>'
    elif py_body == target_body:
        badge = '<span class="badge match">Outputs identical</span>'
    else:
        badge = '<span class="badge mismatch">Outputs differ - translation is wrong</span>'

    py_num, py_unit = format_duration(py_seconds)
    tg_num, tg_unit = format_duration(target_seconds)

    return f"""

<div class="verdict">

  <div class="verdict-grid">

    <div class="speedup-cell">

      <div class="speedup">{speedup}</div>

      <div class="speedup-label">faster in {html.escape(lang.display)}</div>

    </div>

    <div class="timings">

      <div class="timing-row py">

        <span class="name">Python</span>

        <span class="bar-track"><span class="bar" style="width:{py_width:.1f}%"></span></span>

        <span class="num">{py_num}</span><span class="unit">{py_unit}</span>

      </div>

      <div class="timing-row target">

        <span class="name">{html.escape(lang.display)}</span>

        <span class="bar-track"><span class="bar" style="width:{target_width:.1f}%"></span></span>

        <span class="num">{tg_num}</span><span class="unit">{tg_unit}</span>

      </div>

    </div>

    <div class="badge-cell">{badge}</div>

  </div>

</div>

"""


# --------------------------------------------------------------------------
# event handlers
# --------------------------------------------------------------------------


def on_run_python(code: str, target_out: str, language_key: str):
    result = sandbox.run_python(code)
    text = result.display
    return text, render_verdict(text, target_out, language_key)


def on_run_target(code: str, py_out: str, language_key: str):
    result, _ = sandbox.run_compiled(language_key, code)
    text = result.display
    return text, render_verdict(py_out, text, language_key)


def on_port(model_id: str, python_code: str, language_key: str,

            user_key: str, session_id: str):
    """Translate the Python using the selected model."""
    lang = sandbox.LANGUAGES[language_key]
    if not python_code.strip():
        return "// Nothing to port - paste some Python on the left first.", ""

    user_key = (user_key or "").strip()

    # Only the host's own key is rate limited; a visitor's key costs us nothing.
    if not user_key and providers.PUBLIC:
        allowed, message = providers.limiter.check(session_id)
        if not allowed:
            return f"// {message}", _quota_html(session_id)

    try:
        client, used_host_key = providers.build_client(model_id, user_key)
    except providers.NoKeyError as e:
        return f"// {e}", _quota_html(session_id)

    model = providers.BY_ID[model_id]
    kwargs = {"reasoning_effort": "high"} if model.reasoning else {}

    try:
        response = client.chat.completions.create(
            model=model.id,
            messages=[
                {"role": "system", "content": build_system_prompt(language_key)},
                {"role": "user", "content": build_user_prompt(python_code, language_key)},
            ],
            **kwargs,
        )
    except Exception as e:
        # scrub() because provider SDKs sometimes echo the key back in errors
        detail = providers.scrub(f"{type(e).__name__}: {e}")
        return (f"// Request to {model.display} failed.\n// {detail}",
                _quota_html(session_id))

    if used_host_key and providers.PUBLIC:
        providers.limiter.record(session_id)

    reply = response.choices[0].message.content or ""
    return strip_fences(reply), _quota_html(session_id)


def on_change_language(language_key: str):
    """Relabel everything that names the target language, and clear stale output."""
    lang = sandbox.LANGUAGES[language_key]
    ready = language_key in COMPILABLE
    return (
        gr.update(label=f"{lang.display} translation", language=lang.highlight, value=""),
        gr.update(label=f"{lang.display} output", value=""),
        gr.update(value=f"Port to {lang.display}"),
        gr.update(
            value=f"Run {lang.display}" if ready else f"Run {lang.display} (no compiler)",
            interactive=ready,
        ),
        idle_verdict(language_key),
    )


def on_pick_example(name: str):
    return EXAMPLES.get(name, "")


def _quota_html(session_id: str) -> str:
    if not providers.PUBLIC or providers.limiter.limit <= 0:
        return ""
    left = providers.limiter.remaining(session_id)
    return (
        f'<p class="key-note">Free demo conversions left this hour: '
        f"<strong>{left}</strong>. Use your own key below for unlimited access.</p>"
    )


def new_session() -> str:
    return uuid.uuid4().hex


# --------------------------------------------------------------------------
# static chrome
# --------------------------------------------------------------------------

START_LANG = sandbox.DEFAULT_LANGUAGE if sandbox.DEFAULT_LANGUAGE in COMPILABLE \
    else (COMPILABLE[0] if COMPILABLE else sandbox.DEFAULT_LANGUAGE)
START = sandbox.LANGUAGES[START_LANG]

LANGUAGE_CHOICES = [
    (f"{lang.display}" if key in COMPILABLE else f"{lang.display}  (no compiler here)", key)
    for key, lang in sandbox.LANGUAGES.items()
]

MASTHEAD = """

<div class="masthead">

  <h1><span class="py">Python</span><span class="arrow">&#8594;</span><span class="rs">Rust &amp; C++</span></h1>

  <p class="tagline">LLM translation, measured and verified</p>

  <p>An LLM ports a Python program to Rust or C++. Both are compiled and executed

     in a sandbox on this machine, so the speedup is measured rather than

     estimated &mdash; and the outputs are compared, because a fast translation

     that changes the answer is a broken one.</p>

  <div class="links">

    <a href="%s" target="_blank" rel="noopener">Source on GitHub</a>

    <a href="#how">How it works</a>

  </div>

</div>

""" % REPO_URL


def build_footer() -> str:
    installed = [f"{sandbox.LANGUAGES[k].display}: {html.escape(TOOLCHAIN[k])}"
                 for k in COMPILABLE]
    missing = [sandbox.LANGUAGES[k].display for k in sandbox.LANGUAGES
               if k not in COMPILABLE]
    missing_note = (
        f"<br><strong>Note:</strong> no compiler found for {', '.join(missing)}, "
        f"so translations to it can be generated but not executed here."
        if missing else ""
    )
    return f"""

<div class="footer" id="how">

  <p><strong>How it works.</strong> Your Python goes to the selected model with a

  prompt that pins the exact compile flags and forbids algorithmic substitution.

  The returned source is compiled with full optimisation and run. Both programs

  self-report their execution time, and the non-timing output of each is compared

  to verify the translation actually preserved the answer.</p>

  <p><strong>Sandboxing.</strong> Submitted code never runs in the web server's

  process. Each run gets a fresh subprocess with an allowlisted environment

  containing no API keys, a scrubbed working directory, a wall-clock timeout, and

  &mdash; on Linux &mdash; CPU, memory, and process-count limits.

  Isolation here: {html.escape(TOOLCHAIN.get('isolation', 'unknown'))}.</p>

  <p><strong>Keys.</strong> Frontier models never run on this instance's key. A key

  you paste is used for that single request and is never logged, stored, or written

  to disk.{missing_note}</p>

  <p>{html.escape(TOOLCHAIN.get('python', ''))}{(' &middot; ' + ' &middot; '.join(installed)) if installed else ''}</p>

</div>

"""


# --------------------------------------------------------------------------
# interface
# --------------------------------------------------------------------------


def build_ui() -> gr.Blocks:
    with gr.Blocks(title="Python → Rust & C++",
                   analytics_enabled=False) as ui:
        session = gr.State(value=new_session)

        # Inject the stylesheet into the page rather than passing css= to
        # launch(). On Spaces the platform imports this module and calls
        # launch() itself, so any styling passed there is silently dropped -
        # the app would deploy looking completely unstyled.
        gr.HTML(f"<style>{CSS}</style>")

        gr.HTML(MASTHEAD)
        verdict = gr.HTML(idle_verdict(START_LANG))

        with gr.Row(equal_height=True, elem_classes=["picker"]):
            example = gr.Dropdown(
                choices=list(EXAMPLES.keys()), value=DEFAULT_EXAMPLE,
                label="Example program", scale=5,
            )
            language = gr.Dropdown(
                choices=LANGUAGE_CHOICES, value=START_LANG,
                label="Translate to", scale=3,
            )
            model = gr.Dropdown(
                choices=providers.dropdown_choices(), value=providers.default_model(),
                label="Model", scale=5,
            )

        gr.Markdown(providers.availability_summary(), elem_classes=["status-line"])

        with gr.Row(equal_height=True):
            with gr.Column(scale=6):
                python_box = gr.Code(
                    label="Python source", value=EXAMPLES[DEFAULT_EXAMPLE],
                    language="python", lines=24, elem_classes=["pane"],
                )
            with gr.Column(scale=6):
                target_box = gr.Code(
                    label=f"{START.display} translation", value="",
                    language=START.highlight, lines=24, elem_classes=["pane"],
                )

        with gr.Row(elem_classes=["controls"]):
            run_py = gr.Button("Run Python", elem_classes=["btn-run", "py"])
            convert = gr.Button(f"Port to {START.display}", variant="primary",
                                elem_classes=["btn-convert"])
            run_target = gr.Button(
                f"Run {START.display}" if START_LANG in COMPILABLE
                else f"Run {START.display} (no compiler)",
                interactive=START_LANG in COMPILABLE,
                elem_classes=["btn-run", "rust"],
            )

        with gr.Row(equal_height=True):
            with gr.Column(scale=6):
                py_out = gr.Textbox(label="Python output", lines=8,
                                    elem_classes=["out-box", "py-out"],
                                    buttons=["copy"])
            with gr.Column(scale=6):
                target_out = gr.Textbox(label=f"{START.display} output", lines=8,
                                        elem_classes=["out-box", "rust-out"],
                                        buttons=["copy"])

        quota = gr.HTML(_quota_html(""))

        with gr.Accordion("Use your own API key (optional)", open=False):
            gr.Markdown(
                "Needed for the frontier models marked **(need your own key)**, "
                "which never run on this instance's account. Also removes the "
                "rate limit.\n\n"
                "**Worth knowing first:** on this benchmark GPT-5 Nano scored 62x "
                "and GPT-5 scored 63x. The unmarked models need no key and give "
                "you effectively the same answer.\n\n"
                "A key you paste is sent over HTTPS, used for that single request, "
                "then discarded - never logged, stored, or written to disk. It is "
                "not saved in your browser either, so it clears on reload.\n\n"
                "The code is open source if you would rather "
                f"[read it]({REPO_URL}/blob/main/providers.py) or run this locally.",
                elem_classes=["key-note"],
            )
            user_key = gr.Textbox(label="API key", placeholder="sk-...",
                                  type="password", elem_classes=["keybox"])

        gr.HTML(build_footer())

        # ---- wiring ----
        example.change(fn=on_pick_example, inputs=[example], outputs=[python_box])

        language.change(
            fn=on_change_language,
            inputs=[language],
            outputs=[target_box, target_out, convert, run_target, verdict],
        )

        run_py.click(
            fn=on_run_python,
            inputs=[python_box, target_out, language],
            outputs=[py_out, verdict],
        )
        convert.click(
            fn=on_port,
            inputs=[model, python_box, language, user_key, session],
            outputs=[target_box, quota],
        )
        run_target.click(
            fn=on_run_target,
            inputs=[target_box, py_out, language],
            outputs=[target_out, verdict],
        )

    return ui


# Module-level so Hugging Face Spaces can find it. The Gradio SDK imports this
# file and serves the top-level Blocks object - it never runs the __main__
# block, so an app that only builds its UI in there deploys as a blank Space.
demo = build_ui()


if __name__ == "__main__":
    demo.launch(
        theme=gr.themes.Base(),
        # Spaces route traffic to the container, so it must bind all interfaces.
        # Locally, stay on loopback so the dev server is not exposed to the LAN.
        server_name=os.getenv("SERVER_NAME", "0.0.0.0" if ON_SPACES else "127.0.0.1"),
        server_port=int(os.getenv("PORT", "7860")),
    )