INNER CODE UNIT · Python

_warmup_primary_engine

raullenchai/Rapid-MLX · vllm_mlx/server.py:624

async def _warmup_primary_engine(engine: BaseEngine) -> None:
    """Compile the configured engine on its model-owning worker."""

    import time as _time

    logger.info("Warming up (compiling Metal shaders)...")
    warmup_start = _time.monotonic()
    try:
        is_hybrid = _detect_hybrid_for_warmup(engine)
        if not is_hybrid:
            # `generate_warmup()` synchronously waits for the model-owning MLX
            # executor. During eager boot there is no traffic yet, but on a
            # demand load that wait must leave the event loop free for health
            # probes and other coalesced requests.
            await asyncio.to_thread(engine.generate_warmup)
        else:
            logger.info(
                "Hybrid model: running full request warmup "

View source record →

📰 Research Paper
Loading…
⏳ Fetching content…