sgl-project/sglang · error · ValueError

attn_sink must be a CUDA tensor

Error message

attn_sink must be a CUDA tensor

What it means

The optional attn_sink argument to sparse_mla_q8kv8_prefill_fwd must live on CUDA memory. The kernel dereferences the sink pointer on the GPU, so a CPU tensor would cause illegal memory access; the Python validation rejects it up front.

Source

Thrown at python/sglang/kernels/ops/attention/sparse_mla_q8kv8_prefill_sm90.py:413

                "topk_length values must satisfy " f"0 <= topk_length <= topk ({topk})"
            )

    if d_v != 512:
        raise ValueError(
            f"sparse_mla_q8kv8_prefill_fwd only supports d_v=512, got {d_v}"
        )

    if attn_sink is not None and topk_length is None:
        raise ValueError("attn_sink requires topk_length to be provided as well")

    if attn_sink is not None:
        if attn_sink.shape != (h_q,) or attn_sink.dtype != torch.float32:
            raise ValueError(
                f"attn_sink must be float32 with shape ({h_q},), got "
                f"{tuple(attn_sink.shape)}/{attn_sink.dtype}"
            )
        if not attn_sink.is_cuda:
            raise ValueError("attn_sink must be a CUDA tensor")
        if attn_sink.device != device:
            raise ValueError(
                f"attn_sink must be on q's device {device}, got {attn_sink.device}"
            )
        if not attn_sink.is_contiguous():
            raise ValueError("attn_sink must be contiguous")

    for name, scale in (("q_scale", q_scale), ("kv_scale", kv_scale)):
        if not isinstance(scale, torch.Tensor):
            raise ValueError(f"{name} must be a torch.Tensor")
        if not scale.is_cuda:
            raise ValueError(f"{name} must be a CUDA tensor")
        if scale.device != device:
            raise ValueError(
                f"{name} must be on q's device {device}, got {scale.device}"
            )
        if scale.dtype != torch.float32:
            raise ValueError(f"{name} must be float32, got {scale.dtype}")

View on GitHub (pinned to 0132848349)

Solutions

  1. Move the sink to the compute device: attn_sink = attn_sink.to(q.device, non_blocking=True)
  2. Allocate the sink directly on q.device when constructing it

Example fix

// before
attn_sink = torch.zeros(h_q, dtype=torch.float32)  # CPU
// after
attn_sink = torch.zeros(h_q, dtype=torch.float32, device=q.device)
Defensive patterns

Strategy: validation

Validate before calling

assert attn_sink.is_cuda, 'attn_sink must be on CUDA'

Type guard

def sink_on_cuda(t: torch.Tensor) -> bool:
    return t.is_cuda

Prevention

When it happens

Trigger: Calling sparse_mla_q8kv8_prefill_fwd(..., attn_sink=cpu_tensor) where cpu_tensor.is_cuda is False, e.g. a tensor created without a device= argument or loaded from checkpoint on CPU.

Common situations: Loading sink values from a safetensors checkpoint (which defaults to CPU) and passing them straight to the kernel without a .to(q.device) transfer.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/50c38797f41c65ae. Report an issue: GitHub.