Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

3.4 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0

"""Progress and durable results for the Studio transcription screen."""

import asyncio
import collections
import contextlib
import json
import time

from fastapi import HTTPException

from core.inference import transcript_gallery
from loggers import get_logger

logger = get_logger(__name__)


async def stream_transcript(transcribe, title: str):
    loop = asyncio.get_running_loop()
    # Text progress is latest-wins; phase changes are never replaced, so "loading" is never lost.
    updates = collections.deque()
    ready = asyncio.Event()
    closed = False
    last_update = 0.0

    def publish(update):
        if closed:
            return
        if updates and not updates[-1].get("phase"):
            updates.pop()
        updates.append({"type": "progress", **update})
        ready.set()

    async def next_update():
        while not updates:
            ready.clear()
            await ready.wait()
        return updates.popleft()

    def on_progress(update):
        nonlocal last_update
        if closed:
            return
        now = time.monotonic()
        # A phase change ("loading", "downloading_aligner", "transcribing") is never throttled away.
        if update.get("phase") or now - last_update >= 0.25:
            last_update = now
            loop.call_soon_threadsafe(publish, update)

    async def run():
        result = await transcribe(on_progress)
        record = None
        if result.get("text", "").strip():
            try:
                record = await asyncio.to_thread(transcript_gallery.save, result, title)
            except Exception as exc:
                logger.warning("Could not save transcript to the gallery (%s)", type(exc).__name__)
        return {"type": "complete", **result, "record": record}

    task = asyncio.create_task(run())
    pending = asyncio.create_task(next_update())
    try:
        yield json.dumps({"type": "progress", "text": ""}) + "\n"
        while not task.done():
            done, _ = await asyncio.wait(
                {task, pending}, timeout = 5, return_when = asyncio.FIRST_COMPLETED
            )
            if pending in done:
                yield json.dumps(pending.result()) + "\n"
                pending = asyncio.create_task(next_update())
            elif not done:
                yield json.dumps({"type": "heartbeat"}) + "\n"
        # Stop the reader first, or an update it takes while we are suspended is never sent.
        if not pending.done():
            pending.cancel()
            with contextlib.suppress(asyncio.CancelledError):
                await pending
        if not pending.cancelled():
            yield json.dumps(pending.result()) + "\n"
        while updates:
            yield json.dumps(updates.popleft()) + "\n"
        yield json.dumps(task.result()) + "\n"
    except HTTPException as exc:
        yield json.dumps({"type": "error", "message": exc.detail}) + "\n"
    except Exception as exc:
        logger.warning("Transcription stream failed (%s)", type(exc).__name__)
        yield json.dumps({"type": "error", "message": "Transcription failed. Try again."}) + "\n"
    finally:
        closed = True
        pending.cancel()
        task.cancel()
        with contextlib.suppress(asyncio.CancelledError):
            await pending
        with contextlib.suppress(asyncio.CancelledError, Exception):
            await task