sgl-project/sglang · warning

Triton is not supported on current platform, roll back to CP

Error message

Triton is not supported on current platform, roll back to CPU.

What it means

Warning from the FLA (flash-linear-attention) Triton utility layer: no CUDA/GPU device is visible to Triton on this platform, so device-query helpers (multiprocessor count, available device, max shared memory) fall back to CPU values. It signals GPU-less execution of kernels that normally require Triton on CUDA.

Source

Thrown at python/sglang/kernels/ops/attention/fla/utils.py:223

                },
            )

        return wrapper

    return decorator


def checkpoint(fn):
    def wrapper(*args, **kwargs):
        return torch.utils.checkpoint.checkpoint(fn, *args, **kwargs)

    return wrapper


def _cpu_device_warning():
    import warnings

    warnings.warn(
        ("Triton is not supported on current platform, roll back to CPU."), stacklevel=1
    )


@lru_cache(maxsize=None)
def get_multiprocessor_count(tensor_idx: int = 0) -> int:
    try:
        return triton.runtime.driver.active.utils.get_device_properties(tensor_idx)[
            "multiprocessor_count"
        ]
    except BaseException:
        _cpu_device_warning()
        return -1


@lru_cache(maxsize=None)
def get_available_device() -> str:
    try:

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure a CUDA GPU is available and visible (nvidia-smi works, CUDA_VISIBLE_DEVICES not emptied)
2. Install a CUDA-enabled PyTorch/Triton build (e.g. reinstall torch with cu121+ wheels)
3. If CPU execution is intended, accept the warning — kernels fall back but performance/features are limited
4. In CI, gate GPU-dependent tests with a torch.cuda.is_available() skip

Example fix

# before
CUDA_VISIBLE_DEVICES= python -m pytest tests/  # no GPU visible
# after
python -m pytest tests/  # GPU visible; or skip:
pytest.mark.skipif(not torch.cuda.is_available(), reason="needs GPU")

import torch
if not torch.cuda.is_available() or torch.cuda.device_count() == 0:
    pytest.skip("FLA Triton kernels need a visible CUDA GPU")

def has_triton_gpu() -> bool:
    import torch
    return torch.cuda.is_available() and torch.cuda.device_count() > 0

When it happens

Trigger: Calling get_multiprocessor_count / get_available_device / get_all_max_shared_mem (directly or via FLA/gated-delta-net kernels) on a machine with no GPU, with CUDA hidden (CUDA_VISIBLE_DEVICES=''), or on a platform where Triton lacks GPU support (some ROCm/Windows/CPU-only builds).

Common situations: Running unit tests or imports on CPU-only CI boxes, setting CUDA_VISIBLE_DEVICES empty by mistake, using a PyTorch build without CUDA inside containers, or running linear-attention models on unsupported platforms.

Related errors

• This layer norm doesn't support feature dim >= 64KB.
• SANA-WM Triton GDN backend unavailable: {reason}
• {self._op_label()}: no triton backend
• Cannot find NVIDIA Math-DX (cuBLASDx) headers. Install the `
• sparse_attn_v4_paged_decode expects fp16/bf16 q, got {q.dtyp

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/725607771563a6e2. Report an issue: GitHub.