hankcs/HanLP · error · TypeError

Does not support dtype " + str(dtype)

Error message

Does not support dtype " + str(dtype)

What it means

tiny_value_of_dtype handles torch.float, torch.double, and torch.half but no other floating dtype. With torch.bfloat16 (or an exotic float dtype), neither branch matches and it raises TypeError naming the dtype.

Source

Thrown at hanlp/components/parsers/ud/udify_util.py:298

    """Returns a moderately tiny value for a given PyTorch data type that is used to avoid numerical
    issues such as division by zero.
    This is different from `info_value_of_dtype(dtype).tiny` because it causes some NaN bugs.
    Only supports floating point dtypes.

    Args:
      dtype: torch.dtype: 

    Returns:

    """
    if not dtype.is_floating_point:
        raise TypeError("Only supports floating point dtypes.")
    if dtype == torch.float or dtype == torch.double:
        return 1e-13
    elif dtype == torch.half:
        return 1e-4
    else:
        raise TypeError("Does not support dtype " + str(dtype))


def combine_initial_dims_to_1d_or_2d(tensor: torch.Tensor) -> torch.Tensor:
    """Given a (possibly higher order) tensor of ids with shape
    (d1, ..., dn, sequence_length)

    Args:
      tensor: torch.Tensor: 

    Returns:
      If original tensor is 1-d or 2-d, return it as is.

    """
    if tensor.dim() <= 2:
        return tensor
    else:
        return tensor.view(-1, tensor.size(-1))

View on GitHub (pinned to ddb1299bdd)

Solutions

1. Cast logits to float32 before the loss: logits.float()
2. Disable bf16 / use fp16 AMP so loss computation stays in half or float
3. Patch tiny_value_of_dtype to add a bfloat16 branch (e.g. return 1e-4) if you control the vendored copy

Example fix

# before
loss = sequence_cross_entropy_with_logits(model(input_ids).logits, targets, weights)  # bf16
# after
loss = sequence_cross_entropy_with_logits(model(input_ids).logits.float(), targets, weights)

if logits.dtype == torch.bfloat16:
    logits = logits.float()  # tiny_value_of_dtype doesn't support bf16

def is_supported_float_dtype(t: torch.Tensor) -> bool:
    return t.dtype in (torch.float32, torch.float64, torch.float16)

try:
    loss = sequence_cross_entropy_with_logits(logits, ...)
except TypeError as e:
    if 'Does not support dtype' in str(e):
        loss = sequence_cross_entropy_with_logits(logits.float(), ...)
    else:
        raise

When it happens

Trigger: Calling sequence_cross_entropy_with_logits with logits in bfloat16 (common when loading models with torch_dtype=torch.bfloat16 or bf16 AMP training), which propagates to tiny_value_of_dtype via norm.dtype.

Common situations: Training udify-style parsers under bf16 mixed precision or on Ampere+ GPUs where bfloat16 is the default; upgrading PyTorch and switching model dtype for memory savings.

Related errors

• Got average f{average}, expected one of None, 'token', or 'b
• alpha must be float, list of float, or torch.FloatTensor, {}
• Only supports floating point dtypes.
• activation must be callable: type={}
• DataParallel not supported when CRF is used

AI-assisted analysis of hankcs/HanLP@ddb1299bdd (2026-08-27). Data as JSON: /api/errors/fc7893d5c8d91d1e. Report an issue: GitHub.