sgl-project/sglang · error · ValueError

unsupported input for wan_rmsnorm_silu

Error message

unsupported input for wan_rmsnorm_silu

What it means

wan_rmsnorm_silu raises when its support predicate can_use_wan_rmsnorm_silu returns False. The Triton kernel requires: CUDA 5D non-empty x, dtype fp16/bf16/fp32, 0 < channels <= 1024, channels_last_3d contiguous with stride(1)==1, no grad tracking, and gamma/bias CUDA on the same device with matching dtype (or fp32) and numel == channels.

Source

Thrown at python/sglang/kernels/ops/diffusion/norm/wan_rmsnorm_silu_triton.py:159

        and x.stride(1) == 1
        and _affine_supported(x, gamma)
        and (bias is None or _affine_supported(x, bias))
    )


def wan_rmsnorm_silu(
    x: torch.Tensor,
    gamma: torch.Tensor,
    bias: torch.Tensor | None = None,
    rms_scale: float | None = None,
    eps: float = 1e-12,
) -> torch.Tensor:
    """Fused ``SiLU(F.normalize(x, dim=1) * rms_scale * gamma + bias)``.

    Guard with :func:`can_use_wan_rmsnorm_silu`.
    """
    if not can_use_wan_rmsnorm_silu(x, gamma, bias):
        raise ValueError("unsupported input for wan_rmsnorm_silu")

    channels = x.shape[1]
    gamma = gamma.reshape(channels).contiguous()
    has_bias = bias is not None
    bias = gamma if bias is None else bias.reshape(channels).contiguous()
    if rms_scale is None:
        rms_scale = channels**0.5
    return _triton_wan_rmsnorm_silu_cuda(
        x, gamma, bias, float(rms_scale), eps, has_bias
    )


__all__ = ["can_use_wan_rmsnorm_silu", "wan_rmsnorm_silu"]

View on GitHub (pinned to 0132848349)

Solutions

1. Call can_use_wan_rmsnorm_silu(x, gamma, bias) and fall back to eager WanRMS_norm+SiLU when False
2. Convert x to channels_last_3d (x = x.to(memory_format=torch.channels_last_3d)) and ensure stride(1)==1
3. Wrap inference in torch.no_grad()/torch.inference_mode()
4. Verify gamma/bias are CUDA, same device, dtype equal to x or fp32, and numel == x.shape[1]

Example fix

# before
y = wan_rmsnorm_silu(x, gamma, bias)
# after
if can_use_wan_rmsnorm_silu(x, gamma, bias):
    y = wan_rmsnorm_silu(x, gamma, bias)
else:
    y = torch.nn.functional.silu(torch.nn.functional.normalize(x, dim=1) * scale * gamma.reshape(C) + (bias.reshape(C) if bias is not None else 0))

from sglang.kernels.ops.diffusion.norm.wan_rmsnorm_silu_triton import can_use_wan_rmsnorm_silu
if can_use_wan_rmsnorm_silu(x, gamma, bias):
    y = wan_rmsnorm_silu(x, gamma, bias, rms_scale, eps)
else:
    y = eager_wan_rms_silu(x, gamma, bias, rms_scale, eps)

can_use_wan_rmsnorm_silu(x, gamma, bias)  # the library's own predicate is the type guard

try:
    y = wan_rmsnorm_silu(x, gamma, bias)
except ValueError:
    y = eager_wan_rms_silu(x, gamma, bias)  # WanRMS_norm + SiLU in eager torch

When it happens

Trigger: Calling wan_rmsnorm_silu with a CPU tensor, non-channels-last 5D input, requires_grad enabled, channels > 1024, empty tensor, or gamma/bias of wrong dtype/size/device. The docstring explicitly says to guard with can_use_wan_rmsnorm_silu first.

Common situations: Running the Wan VAE decoder without channels_last_3d memory format; forgetting torch.no_grad() during inference wrappers; gamma kept on a different GPU in TP setups; channel counts above 1024 in a modified VAE.

Related errors

• unsupported input for Sana fused bias-SiLU
• num_token_non_padded must be a single-element tensor, got sh
• num_token_non_padded must be an integer tensor, got {num_tok
• v_cache must be provided
• q can only be None when only_qv=True

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2bcec2ae86c8389e. Report an issue: GitHub.