sgl-project/sglang · error · KeyError

No '{backend.value}' backend registered for op {op!r}

Error message

No '{backend.value}' backend registered for op {op!r}

What it means

Raised by select_kernel when the op IS registered but none of its registered KernelSpecs use the explicitly requested backend. This happens when the caller forces backend=<Backend.X> but that backend was never registered for this op (e.g. asking for a triton kernel when only a CUDA one exists).

Source

Thrown at python/sglang/kernels/selector.py:65

        current device*; selects which one. Otherwise optional.

    Raises
    ------
    KeyError
        If ``op`` is unknown, or if ``backend`` is requested but not registered.
    ValueError
        If ``op`` has multiple device-eligible backends and ``backend`` is not
        given, or if none are eligible on this platform.
    """
    specs = registry.get(op)
    if not specs:
        raise KeyError(f"No kernels registered for op {op!r}")

    if backend is not None:
        for spec in specs:
            if spec.backend == backend:
                return spec
        raise KeyError(f"No '{backend.value}' backend registered for op {op!r}")

    if len(specs) == 1:
        return specs[0]

    # Multiple backends: hard-filter by device eligibility.
    platform = _platform()
    eligible = [s for s in specs if s.is_available(platform)]
    if len(eligible) == 1:
        return eligible[0]
    if not eligible:
        raise ValueError(
            f"op {op!r} has no backend usable on device {platform.device.value!r} "
            f"(registered: {[s.backend.value for s in specs]})"
        )
    raise ValueError(
        f"op {op!r} has multiple backends usable on device "
        f"{platform.device.value!r} ({[s.backend.value for s in eligible]}); "
        f"pass backend=... to choose one"

View on GitHub (pinned to 0132848349)

Solutions

1. Query the registered backends for the op (e.g. [s.backend for s in registry.get(op, [])]) and pass one of those, or omit backend= to let device-eligibility auto-select.
2. If you need that backend, verify the sglang version supports it for this op and the optional dependency is installed, then register/upgrade accordingly.
3. Guard backend overrides behind a capability check instead of assuming availability.

Example fix

# before
spec = select_kernel(op, backend=Backend.TRITON)

# after
specs = registry.get(op, [])
if any(s.backend == Backend.TRITON for s in specs):
    spec = select_kernel(op, backend=Backend.TRITON)
else:
    spec = select_kernel(op)  # auto-select eligible backend

from sglang.kernels.registry import registry

def backend_available(op, backend) -> bool:
    return any(s.backend == backend for s in registry.get(op, []))

spec = select_kernel(op, backend=b) if backend_available(op, b) else select_kernel(op)

try:
    spec = select_kernel(op, backend=backend)
except KeyError:
    spec = select_kernel(op)  # fall back to auto-selection

When it happens

Trigger: select_kernel(op, backend=Backend.TRITON) where the op's specs list only contains, say, Backend.SGL_KERNEL / Backend.FLASHINFER / Backend.TORCH. Also triggered deliberately in test_unknown_op_or_backend_raises.

Common situations: Hard-coding a backend for portability or benchmarking when that op was never implemented in that backend; version drift where a backend implementation was removed or renamed between sglang releases.

Related errors

• No kernels registered for op {op!r}
• op {op!r} has multiple backends usable on device {platform.d
• op {op!r} has no backend usable on device {platform.device.v
• KernelSpec.target must be 'module:attr', got {self.target!r}
• LoRA with name {lora_name} does not exist. Loaded LoRAs: {se

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8bb11e8bfbbb18fb. Report an issue: GitHub.