sgl-project/sglang · error · ValueError

Not a canonical UMMA_K Layout: Expected stride failure.

Error message

Not a canonical UMMA_K Layout: Expected stride failure.

What it means

On the Major-K path, the atom-level stride along the MN direction (stride_00 after the (8,2) divide) must equal swizzle_atom_mn_size. This raise fires when that stride doesn't match, meaning atoms are not packed at the expected spacing.

Source

Thrown at python/sglang/kernels/ops/attention/flash_attn/cute/mma_sm100_desc.py:283

            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
    # leading_byte_offset_  [16:30)
    desc |= (leading_byte_offset & 0x3FFF) << 16
    # stride_byte_offset_   [32:46)
    desc |= (stride_byte_offset & 0x3FFF) << 32
    # version_             [46:48)
    desc |= (VERSION & 0x3) << 46
    # base_offset_         [49:52)
    desc |= (BASE_OFFSET & 0x7) << 49
    # lbo_mode_            [52:53)
    desc |= (LBO_MODE & 0x1) << 52

View on GitHub (pinned to 0132848349)

Solutions

1. Remove inter-atom padding; the descriptor requires tight atom packing.
2. Rebuild via cute.tile_to_shape with a standard atom so strides come out canonical.
3. Use the un-swizzled (SWIZZLE_NONE) path if you control the layout and need non-standard packing — but prefer canonical atoms.

canon = cute.logical_divide(layout, (8, 2))
assert canon.stride[0][0] == swizzle_atom_mn_size

When it happens

Trigger: Major.K layout where stride_00 != swizzle_atom_mn_size — e.g. atoms spaced with padding or an intermediate stride level between element and atom tiers.

Common situations: Bank-conflict padding inserted into K-major smem tiles; layouts composed with an extra hierarchy level from zipped/divide modes.

Related errors

• Not a canonical UMMA_MN Layout: Expected stride failure.
• Not a canonical UMMA_MN Layout: Expected profile failure.
• SWIZZLE_128B_BASE32B is invalid for Major-K
• Not a canonical UMMA_K Layout: Expected MN-size multiple of
• Not a canonical UMMA_K Layout: Expected profile failure.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/f427dfd615cfa4d2. Report an issue: GitHub.