sgl-project/sglang · error · ValueError

Petit is not installed. Please install it with `pip install

Error message

Petit is not installed. Please install it with `pip install petit-kernel`.

What it means

This ValueError is raised by the stub implementation of prepare_nvfp4_layer_for_petit that exists when the optional petit-kernel package is not installed. petit_utils.py guards all Petit NVFP4 code paths behind an import check and provides no-op/failing fallbacks, so calling the prepare function without the dependency immediately fails. It exists to make Petit an optional acceleration dependency.

Source

Thrown at python/sglang/srt/layers/quantization/petit_utils.py:18

from typing import Optional

import torch

try:
    from petit_kernel import mul_nvfp4_a16, process_nvfp4_scales, repack_nvfp4
except ImportError:

    def _check_petit_nvfp4_supported(
        quant_method: str, group_size: Optional[int]
    ) -> tuple[bool, Optional[str]]:
        return (
            False,
            "Petit is not installed. Please install it with `pip install petit-kernel`.",
        )

    def prepare_nvfp4_layer_for_petit(layer: torch.nn.Module) -> None:
        raise ValueError(
            "Petit is not installed. Please install it with `pip install petit-kernel`."
        )

    def apply_petit_nvfp4_linear(
        input: torch.Tensor,
        weight: torch.Tensor,
        weight_scale: torch.Tensor,
        weight_scale_2: torch.Tensor,
        size_n: int,
        size_k: int,
        bias: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        raise ValueError(
            "Petit is not installed. Please install it with `pip install petit-kernel`."
        )


def _check_petit_nvfp4_supported(

View on GitHub (pinned to 0132848349)

Solutions

1. pip install petit-kernel into the same environment that runs sglang, then retry.
2. Verify with `python -c "import petit_kernel"` using the exact interpreter/venv used by the sglang server.
3. If Petit is not intended, disable the Petit path in your quantization/server args so prepare_nvfp4_layer_for_petit is never invoked.
4. Check for GPU/architecture requirements of petit-kernel if the install itself fails, and install a build matching your CUDA/torch version.

Example fix

# before
sglang server launched with petit NVFP4 path enabled; no petit-kernel installed -> ValueError

# after
pip install petit-kernel
python -c "import petit_kernel; print(petit_kernel.__version__)"

import importlib.util
if importlib.util.find_spec("petit_kernel") is None:
    raise RuntimeError("petit-kernel not installed; run: pip install petit-kernel or disable Petit")
prepare_nvfp4_layer_for_petit(layer)

def petit_available() -> bool:
    return importlib.util.find_spec("petit_kernel") is not None

When it happens

Trigger: Calling prepare_nvfp4_layer_for_petit(layer) (directly or via a quantization config that enables Petit for NVFP4 layers) in an environment where `import petit_kernel` failed, i.e. petit-kernel is not pip-installed in the active Python/venv.

Common situations: Enabling Petit NVFP4 acceleration in sglang without installing petit-kernel; running in a container or venv missing the optional dependency; version upgrades where the dependency became optional; wrong Python environment (installed into a different interpreter than the one running the server).

Related errors

• {error_msg}
• Online MXFP4 requantization from compressed-tensors NVFP4 ch
• MIXED_PRECISION checkpoint has no NVFP4 layers to requantize
• Type must match: {self.a_dtype} != {self.b_dtype}
• nvfp4_gemm_swiglu_nvfp4_quant currently supports NVFP4 input

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6fd4f81adc515a1d. Report an issue: GitHub.