hpcaitech/Open-Sora · error · ValueError

Unsupported dtype {dtype}

Error message

Unsupported dtype {dtype}

What it means

This ValueError is raised by to_torch_dtype in opensora/utils/misc.py when a string dtype is not one of the mapped keys: "fp32", "fp16", "half", "bf16". The string branch of the function looks the name up in dtype_mapping and raises before any conversion happens. It exists to reject misspelled or unsupported precision names early, typically when parsing training config.

Source

Thrown at opensora/utils/misc.py:247

        dtype (str | torch.dtype): The input dtype.

    Returns:
        torch.dtype: The converted dtype.
    """
    if isinstance(dtype, torch.dtype):
        return dtype
    elif isinstance(dtype, str):
        dtype_mapping = {
            "float64": torch.float64,
            "float32": torch.float32,
            "float16": torch.float16,
            "fp32": torch.float32,
            "fp16": torch.float16,
            "half": torch.float16,
            "bf16": torch.bfloat16,
        }
        if dtype not in dtype_mapping:
            raise ValueError(f"Unsupported dtype {dtype}")
        dtype = dtype_mapping[dtype]
        return dtype
    else:
        raise ValueError(f"Unsupported dtype {dtype}")


# ======================================================
# Profile
# ======================================================


class Timer:
    def __init__(self, name, log=False, barrier=False, coordinator: DistCoordinator | None = None):
        self.name = name
        self.start_time = None
        self.end_time = None
        self.log = log
        self.barrier = barrier

View on GitHub (pinned to 7ad6a96a13)

Solutions

1. Use one of the supported strings: "fp32", "fp16", "half", or "bf16"
2. If you have a torch.dtype or the string "auto"-style value, pass the actual torch.dtype instead of a string
3. Normalize config inputs before calling: strip whitespace and lowercase the string

Example fix

# before
dtype = to_torch_dtype("float16")

# after
dtype = to_torch_dtype("fp16")

SUPPORTED_DTYPES = {"fp32", "fp16", "half", "bf16"}
dtype = cfg.get("dtype", "bf16")
assert dtype in SUPPORTED_DTYPES, f"dtype must be one of {SUPPORTED_DTYPES}, got {dtype!r}"

def is_supported_dtype_str(dtype: str) -> bool:
    return isinstance(dtype, str) and dtype in {"fp32", "fp16", "half", "bf16"}

try:
    dtype = to_torch_dtype(cfg["dtype"])
except ValueError:
    dtype = torch.bfloat16  # safe default with a logged warning

When it happens

Trigger: Calling to_torch_dtype("float16"), to_torch_dtype("fp8"), to_torch_dtype("float"), or any string other than fp32/fp16/half/bf16. Non-string inputs take the other branch (error 44). Called from main when translating a config dtype string into a torch.dtype.

Common situations: Writing "float16" (the torch spelling) instead of "fp16" in a YAML/JSON training config; upgrading/downgrading OpenSora versions where accepted dtype names differ; setting fp8 or tf32 names that this mapping never supported.

Related errors

• Unknown optimizer: {optimizer_name}
• Unknown plugin {plugin}
• resize(mode={mode}) not implemented.
• dtype: {dtype}
• Invalid logging level: {level}

AI-assisted analysis of hpcaitech/Open-Sora@7ad6a96a13 (2026-08-28). Data as JSON: /api/errors/3dbcfb590dfe8206. Report an issue: GitHub.