sgl-project/sglang · error · ValueError

`mixed_qkv` must be contiguous in the last dim.

Error message

`mixed_qkv` must be contiguous in the last dim.

What it means

The replaySSM decode kernel requires mixed_qkv to be contiguous in its last dimension (stride(-1) == 1) so the Triton kernel can coalesced-load each token's packed row. Non-unit last stride (e.g. from a transposed or sliced view) is rejected.

Source

Thrown at python/sglang/kernels/ops/attention/fla/fused_recurrent_linear_replayssm.py:480

        ``dt_bias``=[HV], ``g_cache``=[num_slots, HV, L].
      * ``is_kda=True`` (KDA): per-K-channel gate.  ``a``=[B, HV, K],
        ``dt_bias``=[HV, K], ``g_cache``=[num_slots, HV, L, K].
    ``A_log`` is [HV] (per-head scalar) for both.

    Same call surface as the packed decode plus the three ring caches
    (``d_cache`` / ``k_cache`` / ``g_cache``) and the per-decode-row
    ``write_pos`` cursor.  ``initial_state`` is both the checkpoint read (h0)
    and the (flush-only) checkpoint write (ht), in place.

    Allocates nothing persistent: the caller owns the ring tensors and is
    responsible for advancing / resetting ``write_pos`` (e.g. ``(write_pos+1) %
    L`` after each step).  This is a STANDALONE kernel; the memory-pool / cache
    integration is a later phase.
    """
    if mixed_qkv.ndim != 2:
        raise ValueError(f"`mixed_qkv` must be 2D (got ndim={mixed_qkv.ndim}).")
    if mixed_qkv.stride(-1) != 1:
        raise ValueError("`mixed_qkv` must be contiguous in the last dim.")
    if b.ndim != 2:
        raise ValueError(f"`b` must be 2D (got b.ndim={b.ndim}).")
    if A_log.ndim != 1:
        raise ValueError("`A_log` must be a 1D tensor.")
    if initial_state.ndim != 4:
        raise ValueError(f"`initial_state` must be 4D (got ndim={initial_state.ndim}).")
    if not out.is_contiguous():
        raise ValueError("`out` must be contiguous.")
    if write_pos.ndim != 1 or write_pos.dtype != torch.int32:
        raise ValueError("`write_pos` must be a 1D int32 tensor.")
    if force_flush is not None and (
        force_flush.ndim != 1 or force_flush.dtype != torch.int32
    ):
        raise ValueError("`force_flush` must be a 1D int32 tensor or None.")

    B = mixed_qkv.shape[0]
    num_state_slots, HV, V, K = initial_state.shape
    qkv_dim = mixed_qkv.shape[1]

View on GitHub (pinned to 0132848349)

Solutions

1. Call .contiguous() on mixed_qkv before the kernel (or ensure the projection writes a contiguous tensor)
2. If slicing a packed buffer, copy the slice into a fresh contiguous tensor
3. Verify with mixed_qkv.stride(-1) == 1 in debug builds

Example fix

// before
qkv = buf[:, :, :D]  # non-unit stride
fused_recurrent_linear_replayssm_decode(qkv, ...)
// after
qkv = buf[:, :, :D].contiguous()
fused_recurrent_linear_replayssm_decode(qkv, ...)

if mixed_qkv.stride(-1) != 1:
    mixed_qkv = mixed_qkv.contiguous()

def last_dim_contiguous(t: torch.Tensor) -> bool:
    return t.stride(-1) == 1

When it happens

Trigger: Passing mixed_qkv produced by a transpose, narrow/slice of a larger buffer, or expand that leaves last-dim stride != 1.

Common situations: Slicing qkv out of a fused projection tensor with padding; using tensors from a ring buffer with non-standard strides in CUDA-graph replay tests.

Related errors

• `out` must be contiguous.
• `mixed_qkv` must be 2D (got ndim={mixed_qkv.ndim}).
• `b` must be 2D (got b.ndim={b.ndim}).
• `A_log` must be a 1D tensor.
• `initial_state` must be 4D (got ndim={initial_state.ndim}).

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1a50f59ff25fbe8c. Report an issue: GitHub.