sgl-project/sglang · error · ValueError

SWIZZLE_128B_BASE32B is invalid for Major-K

Error message

SWIZZLE_128B_BASE32B is invalid for Major-K

What it means

make_smem_desc_base rejects SWIZZLE_128B_BASE32B layouts on the Major-K path: the base32B sub-atom encoding is only defined for Major-MN operands in the UMMA shared-memory descriptor, so a Major.K layout combined with Swizzle<2,5,2> is invalid.

Source

Thrown at python/sglang/kernels/ops/attention/flash_attn/cute/mma_sm100_desc.py:273

                "Not a canonical UMMA_MN Layout: Expected profile failure."
            )
        stride_00 = canonical_layout.stride[0][0]
        if layout_type is not LayoutType.SWIZZLE_NONE and stride_00 != 1:
            raise ValueError("Not a canonical UMMA_MN Layout: Expected stride failure.")
        stride_10 = canonical_layout.stride[1][0]
        if stride_10 != swizzle_atom_mn_size:
            raise ValueError("Not a canonical UMMA_MN Layout: Expected stride failure.")
        stride_01, stride_11 = (
            canonical_layout.stride[0][1],
            canonical_layout.stride[1][1],
        )
        if layout_type is LayoutType.SWIZZLE_NONE:
            stride_byte_offset, leading_byte_offset = stride_01, stride_11
        else:
            stride_byte_offset, leading_byte_offset = stride_11, stride_01
    else:
        if layout_type == LayoutType.SWIZZLE_128B_BASE32B:
            raise ValueError("SWIZZLE_128B_BASE32B is invalid for Major-K")
        if not cute.size(layout.shape[0]) % 8 == 0:
            raise ValueError(
                "Not a canonical UMMA_K Layout: Expected MN-size multiple of 8."
            )
        canonical_layout = cute.logical_divide(layout, (8, 2))
        if not cute.is_congruent(canonical_layout, ((1, 1), (1, 1))):
            raise ValueError("Not a canonical UMMA_K Layout: Expected profile failure.")
        stride_00 = canonical_layout.stride[0][0]
        if stride_00 != swizzle_atom_mn_size:
            raise ValueError("Not a canonical UMMA_K Layout: Expected stride failure.")
        stride_10 = canonical_layout.stride[1][0]
        if layout_type is not LayoutType.SWIZZLE_NONE and stride_10 != 1:
            raise ValueError("Not a canonical UMMA_K Layout: Expected stride failure.")
        stride_01 = canonical_layout.stride[0][1]
        stride_byte_offset, leading_byte_offset = stride_01, stride_10

    # ------------------------------------------------------------------ pack
    desc = 0

View on GitHub (pinned to 0132848349)

Solutions

1. Give the K-major operand a standard swizzle (SWIZZLE_32B/64B/128B or none) instead of the base32B atom.
2. Or keep the base32B layout but tag the operand Major.MN if it truly is MN-major.
3. Re-derive the layout from the operand's actual memory layout rather than sharing one layout across operands.

if major is Major.K:
    assert (swizzle.num_base, swizzle.num_bits, swizzle.num_shift) != (5, 2, 2), \
        "SWIZZLE_128B_BASE32B not allowed for Major-K"

When it happens

Trigger: Calling gemm_ptx*/smem_desc_base_from_tensor with major=Major.K while the layout's swizzle resolves to LayoutType.SWIZZLE_128B_BASE32B.

Common situations: Reusing an MN-major 128B-base32B tile layout for the K operand without changing its swizzle; copy-paste between A/B operand construction paths.

Related errors

• Unexpected swizzle shift – want S==3 for M==4
• Only Swizzle<2,5,2> supported for 128B_BASE32B
• Unsupported swizzle triple for UMMA smem descriptor
• Not a canonical UMMA_MN Layout: Expected profile failure.
• Not a canonical UMMA_MN Layout: Expected stride failure.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5490aa68f9f6bff4. Report an issue: GitHub.