jax-ml/jax · error · NotImplementedError

TPU generation is not supported

Error message

TPU generation is not supported

What it means

TpuInfo.get_sublane_tiling was called for a TPU generation (other than 5/5e/6/7/8 handled above) whose sublane tiling rules are unknown to this JAX version.

Source

Thrown at jax/_src/tpu_info.py:294

  def get_sublane_tiling(self, dtype: dtypes.DType) -> int:
    """Returns the sublane tiling for the given itemsize.

    Note that this is a heurustic and depends on the settings of the XLA flags.
    """
    bitwidth = dtypes.itemsize_bits(dtype)
    if self.generation < 7:
      # Caveat: before TPU7x, by default XLA does not use large 2nd minor tiling
      # but it can be enabled by setting the flag
      # xla_tpu_enable_large_2nd_minor_layout_for_x16.
      if bitwidth == 16 or bitwidth == 32:
        return self.num_sublanes
      else:
        # Large 2nd minor tiling is enabled for other types.
        return self.num_sublanes * (32 // bitwidth)
    # XLA allows large 2nd minor tiling by default starting with TPU7x.
    if self.generation == 7 or self.generation == 8:
      return self.num_sublanes * (32 // bitwidth)
    raise NotImplementedError("TPU generation is not supported")


def is_tpu_device() -> bool:
  return chip_version_from_device_kind(get_device_kind()) is not None


registry: dict[str, Callable[[], TpuInfo]] = {}


def _get_tpu_info_impl(chip_version: ChipVersion, num_cores: int) -> TpuInfo:
  """Returns the TPU hardware info for the given chip version and core count.

  Note that all information is *per-TensorCore* so you would need to multiply by
  `num_cores` to obtain the total for the chip.

  Args:
    chip_version: The TPU chip version.
    num_cores: The number of TensorCores per chip for this configuration. This

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Upgrade JAX/jaxlib to a version supporting your TPU generation
2. Pin workloads to supported generations

from jax._src.tpu_info import TpuInfo
assert info.generation in (5,6,7,8)

try:
    tiling = info.get_sublane_tiling(bw)
except NotImplementedError:
    tiling = None  # compute fallback

When it happens

Trigger: Running on a newer TPU generation (e.g. TPU9+) with an older JAX that lacks tiling rules, or a corrupted generation parse.

Common situations: Cloud TPU images with newer chips than the installed JAX supports.

Related errors

• Pallas TPU requires a recent libtpu version (at least 0.0.46
• stop_gradient only works on valid JAX arrays, but input argu
• Unknown GPU platform for __dlpack__: {platform_version}
• __dlpack__ device only supported for TPU pinned host memory
• `buffer_callback` not supported on {platform} backend.

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/a28ed4f3daf94a3b. Report an issue: GitHub.