sgl-project/sglang · error · RuntimeError

unsupported input for causal Conv3D cat/pad CUDA

Error message

unsupported input for causal Conv3D cat/pad CUDA

What it means

The CUDA fused path for causal Conv3D cat/pad only runs on 5D, contiguous CUDA tensors whose padding configuration is accepted by can_use_fused_causal_conv3d_cat_pad_cuda. Any deviation (CPU tensor, wrong rank, non-contiguous memory, or an unsupported padding layout) triggers this guard.

Source

Thrown at python/sglang/kernels/ops/diffusion/layout/causal_conv3d_cat_pad_jit.py:115

def fused_causal_conv3d_cat_pad_cuda(
    x: torch.Tensor,
    cache_x: torch.Tensor,
    padding: list[int] | tuple[int, ...],
) -> torch.Tensor:
    if x.dtype not in _SUPPORTED_DTYPES:
        raise RuntimeError(f"unsupported dtype for causal Conv3D cat/pad: {x.dtype}")
    if not torch.compiler.is_compiling():
        if (
            not x.is_cuda
            or not cache_x.is_cuda
            or x.dim() != 5
            or cache_x.dim() != 5
            or not x.is_contiguous()
            or not cache_x.is_contiguous()
            or not can_use_fused_causal_conv3d_cat_pad_cuda(x, cache_x, padding)
        ):
            raise RuntimeError("unsupported input for causal Conv3D cat/pad CUDA")
    return _causal_conv3d_cat_pad_custom_op(x, cache_x, *padding)


def can_use_fused_causal_conv3d_cat_pad_cuda(
    x: torch.Tensor,
    cache_x: torch.Tensor,
    padding: list[int] | tuple[int, ...],
) -> bool:
    if x.dtype not in _SUPPORTED_DTYPES:
        return False
    pad_w_left, pad_w_right, pad_h_top, pad_h_bottom, pad_d_left, pad_d_right = padding
    cache_t = cache_x.shape[2]
    depth_left = pad_d_left - cache_t
    if depth_left < 0 or pad_d_right != 0:
        return False
    out_numel = (
        x.shape[0]
        * x.shape[1]

View on GitHub (pinned to 0132848349)

Solutions

1. Call .contiguous() on x and cache_x before invoking
2. Verify x.dim() == 5 and both tensors are on CUDA
3. Pre-check with can_use_fused_causal_conv3d_cat_pad_cuda(x, cache_x, padding) and use the fused_causal_conv3d_cat_pad dispatcher or eager fallback otherwise
4. Adjust padding to a supported configuration

Example fix

# before
y = fused_causal_conv3d_cat_pad_cuda(x[::1], cache, padding)
# after
x = x.contiguous(); cache = cache.contiguous()
y = fused_causal_conv3d_cat_pad_cuda(x, cache, padding) if can_use_fused_causal_conv3d_cat_pad_cuda(x, cache, padding) else eager_path(x, cache, padding)

from sglang.kernels.ops.diffusion.layout.causal_conv3d_cat_pad_jit import can_use_fused_causal_conv3d_cat_pad_cuda
x = x.contiguous(); cache_x = cache_x.contiguous()
assert x.is_cuda and x.dim() == 5
use_fused = can_use_fused_causal_conv3d_cat_pad_cuda(x, cache_x, padding)

def usable(x, c, p) -> bool:
    return x.is_cuda and c.is_cuda and x.dim() == 5 and c.dim() == 5 and x.is_contiguous() and c.is_contiguous() and can_use_fused_causal_conv3d_cat_pad_cuda(x, c, p)

When it happens

Trigger: Calling fused_causal_conv3d_cat_pad_cuda with non-CUDA tensors, tensors whose dim() != 5, non-contiguous x or cache_x (e.g. sliced/permuted views), or padding values the fused kernel cannot handle (checked by can_use_fused_causal_conv3d_cat_pad_cuda).

Common situations: Passing a permuted/sliced activation from a prior layer without calling .contiguous(); running the diffusion pipeline with padding configs only the Triton/eager fallback supports; accidental CPU tensors in tests.

Related errors

• q must be contiguous
• kv must be contiguous
• unsupported dtype for causal Conv3D cat/pad: {x.dtype}
• unsupported input for usp_merge_heads CUDA
• unsupported input for modulate_scale_shift CUDA

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1b6c468c3f230896. Report an issue: GitHub.