vllm-project/vllm · error · RuntimeError

CUDA is required for this proof.

Error message

CUDA is required for this proof.

What it means

This verification script measures exponential/Gumbel race precision and requires a real CUDA accelerator. The first guard fails when torch.accelerator.is_available() is false — torch sees no usable accelerator device in the current process.

Source

Thrown at tools/gumbel_precision/prove_exponential_race_precision.py:116

            remaining -= batch
        torch.accelerator.synchronize()
        elapsed = time.perf_counter() - start
        print(f"{dtype}: tail_hits={hits} elapsed={elapsed:.2f}s")


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--lower-tail-samples", type=int, default=200_000_000)
    parser.add_argument("--lower-tail-chunk-size", type=int, default=10_000_000)
    parser.add_argument("--race-trials", type=int, default=100_000)
    parser.add_argument("--race-tail-tokens", type=int, default=262_144)
    parser.add_argument("--race-gap", type=float, default=20.5)
    parser.add_argument("--race-chunk-trials", type=int, default=64)
    parser.add_argument("--seed", type=int, default=2026)
    args = parser.parse_args()

    if not torch.accelerator.is_available():
        raise RuntimeError("CUDA is required for this proof.")

    device = torch.accelerator.current_accelerator()
    if device.type != "cuda":
        raise RuntimeError("CUDA is required for this proof.")

    print(f"torch={torch.__version__} cuda={torch.version.cuda}")
    print(f"device={device}")
    measure_exponential_lower_tail(
        device=device,
        samples=args.lower_tail_samples,
        chunk_size=args.lower_tail_chunk_size,
        seed=args.seed,
    )
    run_many_tail_race(
        device=device,
        trials=args.race_trials,
        num_tail_tokens=args.race_tail_tokens,
        gap=args.race_gap,

View on GitHub (pinned to c794754062)

Solutions

1. Run the script on a machine with a working NVIDIA GPU and a CUDA-enabled torch build.
2. If using Docker, start the container with `--gpus all` and a CUDA base image.
3. Unset or correct CUDA_VISIBLE_DEVICES and verify `python -c "import torch; print(torch.accelerator.is_available())"` returns True first.

import torch
if not torch.accelerator.is_available():
    raise SystemExit("No accelerator visible; fix GPU passthrough/driver before running the proof")

When it happens

Trigger: Running tools/gumbel_precision/prove_exponential_race_precision.py on a CPU-only host, in a container without GPU passthrough, or with CUDA_VISIBLE_DEVICES set to an empty/invalid value.

Common situations: Running the numerical proof on a dev box without a GPU; docker run without --gpus; CUDA driver mismatch making the device invisible to torch.

Related errors

• managed Python headless engine exited unexpectedly with stat
• Numerics check failed for case {case}:\n{e}
• Could not determine Python executable. Please provide it man
• Failed to find the NIXL wheel after building it.
• Failed to find the repaired NIXL wheel.

AI-assisted analysis of vllm-project/vllm@c794754062 (2026-08-14). Data as JSON: /api/errors/12ee1bee51a76ef2. Report an issue: GitHub.