xai-org/x-algorithm · error · ValueError

{name} must have dtype torch.int32

Error message

{name} must have dtype torch.int32

What it means

Raised by _check_and_expand_metadata_tensor (called from normalize_block_sparse_tensors) when an auxiliary metadata tensor (e.g. a seqlen or schedule tensor) is provided but its dtype is not torch.int32. The kernels read these as 32-bit integers, so float/int64 tensors are rejected.

Source

Thrown at phoenix/xrex/cutedsl/ranker_fa4/block_sparsity.py:267

        expected_index_shape = (*expected_index_shape[:3], idx.shape[3])
    expanded_idx = _expand_sparsity_tensor(
        idx, expected_index_shape, f"{name}_block_idx", context, hint
    )
    return expanded_cnt, expanded_idx


def _check_and_expand_metadata_tensor(
    name: str,
    tensor: torch.Tensor | None,
    expected_shape: Tuple[int, ...],
    context: str | None,
    hint: str | Callable[[], str] | None,
    device: torch.device,
) -> torch.Tensor | None:
    if tensor is None:
        return None
    if tensor.dtype != torch.int32:
        raise ValueError(f"{name} must have dtype torch.int32")
    if tensor.device != device:
        raise ValueError(f"{name} must be on the same device as block sparse tensors")
    if not tensor.is_cuda:
        raise ValueError(f"{name} must live on CUDA")
    return _expand_sparsity_tensor(tensor, expected_shape, name, context, hint)


def get_block_sparse_expected_shapes(
    batch_size: int,
    num_head: int,
    seqlen_q: int,
    seqlen_k: int,
    m_block_size: int,
    n_block_size: int,
    q_stage: int,
) -> Tuple[Tuple[int, int, int], Tuple[int, int, int, int]]:
    m_block_size_effective = q_stage * m_block_size
    expected_m_blocks = ceildiv(seqlen_q, m_block_size_effective)

View on GitHub (pinned to 24c60942c5)

Solutions

1. Cast the tensor: t = t.to(torch.int32)
2. Create it with dtype=torch.int32 from the start
3. Add a pre-call dtype assert

Example fix

# before
seqlens = torch.tensor([128, 256, 64])  # int64

# after
seqlens = torch.tensor([128, 256, 64], dtype=torch.int32, device='cuda')

assert meta is None or meta.dtype == torch.int32, f"expected int32, got {meta.dtype}"

def is_int32(t: torch.Tensor) -> bool:
    return t.dtype == torch.int32

When it happens

Trigger: Passing a metadata tensor created with default dtype (float32), torch.long/int64 (e.g. from torch.tensor([...])), or torch.int16 to the block-sparse normalization path.

Common situations: Building metadata with torch.tensor(list) which defaults to int64; converting from numpy int64 arrays; mixing dtypes when metadata comes from a different component than the block tensors.

Related errors

• {name}_block_cnt and {name}_block_idx must be on the same de
• {name}_block tensors must live on CUDA
• {name} must be on the same device as block sparse tensors
• {name} must live on CUDA
• mask_block_cnt and mask_block_idx must be provided for block

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/8ea87d8494338d32. Report an issue: GitHub.