sgl-project/sglang · error · ValueError

Sampling factor should be a multiply of 2!

Error message

Sampling factor should be a multiply of 2!

What it means

The ConvSubsampling module only supports power-of-two subsampling factors (its internal logic takes log2 of the factor to build strided conv layers). Any subsampling_factor not divisible by 2 raises this ValueError in __init__.

Source

Thrown at python/sglang/srt/models/phi4mm_utils.py:1098

    def __init__(
        self,
        feat_in,
        feat_out,
        subsampling_factor=4,
        subsampling="dw_striding",
        conv_channels=256,
        subsampling_conv_chunking_factor=1,
        activation=nn.ReLU(),  # noqa: B008
        is_causal=False,
    ):
        super().__init__()
        self._subsampling = subsampling
        self._conv_channels = conv_channels
        self._feat_in = feat_in
        self._feat_out = feat_out

        if subsampling_factor % 2 != 0:
            raise ValueError("Sampling factor should be a multiply of 2!")
        self._sampling_num = int(math.log(subsampling_factor, 2))
        self.subsampling_factor = subsampling_factor
        self.is_causal = is_causal
        self.subsampling_causal_cond = subsampling in (
            "dw_striding",
            "striding",
            "striding_conv1d",
        )

        if (
            subsampling_conv_chunking_factor != -1
            and subsampling_conv_chunking_factor != 1
            and subsampling_conv_chunking_factor % 2 != 0
        ):
            raise ValueError(
                "subsampling_conv_chunking_factor should be -1, 1, or a " "power of 2"
            )
        self.subsampling_conv_chunking_factor = subsampling_conv_chunking_factor

View on GitHub (pinned to 0132848349)

Solutions

1. Set subsampling_factor to a power of two (2, 4, 8)
2. Adjust the upstream feature hop/chunking so the required factor becomes even
3. Verify the config value comes from the original checkpoint rather than a hand edit

Example fix

# before
ConvSubsampling(subsampling_factor=3, ...)  # ValueError
# after
ConvSubsampling(subsampling_factor=4, ...)

assert subsampling_factor > 0 and subsampling_factor % 2 == 0, "subsampling_factor must be a power of two"

def is_power_of_two(n: int) -> bool:
    return n > 0 and (n & (n - 1)) == 0

When it happens

Trigger: Constructing ConvSubsampling with subsampling_factor of 1, 3, 5, or any odd value; subsampling_factor is often derived from the ratio of audio feature rate to LLM token rate.

Common situations: Configuring a new audio frontend where input_feat_ratio/model chunk size yields an odd subsampling factor; editing configs to experiment with compression rates.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• subsampling_conv_chunking_factor should be -1, 1, or a power
• audio_out_channels must be divisible by tp_size for TP-shard
• projection_cls = {projection_cls}, not implemented
• audio_projection_mode = {audio_projection_mode} not implemen
• Unsupported activation type {self.glu_act}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2ebfe276b84b68f6. Report an issue: GitHub.