sgl-project/sglang · error · ValueError

TGV cute_ext tactic {tactic} out of range [0, {len(_TGV_CUTE

Error message

TGV cute_ext tactic {tactic} out of range [0, {len(_TGV_CUTE_EXT_TACTIC_CONFIGS)}).

What it means

TGV tactic ids index into _TGV_CUTE_EXT_TACTIC_CONFIGS; a tactic >= len(configs) (or an explicitly non-negative out-of-range id) is invalid. Negative tactics are replaced by the default tactic.

Source

Thrown at python/sglang/kernels/ops/gemm/cutedsl_bf16_gemm.py:1256

    # bias is per-output-feature in PyTorch terms. After the A↔B swap the
    # kernel's M-axis maps to PyTorch's N-axis, which is exactly what the
    # bias indexes — so this is a direct (0, 1, 0) broadcast.
    L = c_swap.shape[0]
    M_ce = c_swap.shape[1]  # == PyTorch N
    N_ce = c_swap.shape[2]  # == PyTorch M
    bias_3d = bias_pt.as_strided(size=(L, M_ce, N_ce), stride=(0, 1, 0))
    bias_ = from_dlpack(bias_3d.detach(), assumed_align=2).mark_layout_dynamic(
        leading_dim=1
    )
    return a_, b_, c_, bias_, layout


def _resolve_tactic(tactic: int) -> Tuple[int, int, int, bool]:
    """Return (cta_m, cta_n, num_ab_stage, use_2cta) for the given tactic id."""
    if tactic < 0:
        tactic = _TGV_CUTE_EXT_DEFAULT_TACTIC
    if tactic >= len(_TGV_CUTE_EXT_TACTIC_CONFIGS):
        raise ValueError(
            f"TGV cute_ext tactic {tactic} out of range [0, {len(_TGV_CUTE_EXT_TACTIC_CONFIGS)})."
        )
    return _TGV_CUTE_EXT_TACTIC_CONFIGS[tactic]


def _run_tgv(
    a: torch.Tensor,
    b: torch.Tensor,
    bias: Optional[torch.Tensor],
    out: torch.Tensor,
    pdl: bool,
    tactic: int,
) -> torch.Tensor:
    """Dispatch one TGV GEMM: a (M, K) @ b (K, N) -> out (M, N), bias (N,)."""
    cta_m, cta_n, num_ab_stage, use_2cta = _resolve_tactic(tactic)
    has_bias = bias is not None

    a_, b_, c_, bias_, (a_leading, b_leading, c_leading) = _to_cute_swap(

View on GitHub (pinned to 0132848349)

Solutions

1. Clamp tactic to the valid range: 0 <= tactic < len(_TGV_CUTE_EXT_TACTIC_CONFIGS).
2. Pass -1 (or omit) to use the default tactic.
3. Regenerate tactic lists after upgrading the kernel module.

Example fix

// before
_run_tgv(x, w, bias, tactic=999)
// after
_run_tgv(x, w, bias, tactic=-1)  # default tactic

assert -1 <= tactic < len(_TGV_CUTE_EXT_TACTIC_CONFIGS)

When it happens

Trigger: Calling _run_tgv with a user-supplied tactic integer larger than the number of hardcoded tactic configs.

Common situations: Porting tactic ids from a newer/older kernel version where the config table changed size, or autotuning sweeps that assume more tactics exist.

Related errors

• TGV cute_ext backend supports {list(_TORCH_TO_CUTLASS_DTYPE)
• TGV cute_ext output supports {list(_TORCH_TO_CUTLASS_OUT_DTY
• cta_n={cta_n} invalid for use_2cta={use_2cta}: bf16 K-major
• cutedsl_bf16_gemm requires an SM10x GPU
• dsv3_fused_a_gemm requires SM90 (Hopper) or later

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/51e154b93d83b4ba. Report an issue: GitHub.