jax-ml/jax · error · ValueError

Specified input which requires a copy since the source data

Error message

Specified input which requires a copy since the source data buffer {se[i:]} However copy=False. Set copy=True or copy=None to perform the requested operation.

What it means

from_dlpack with copy=False attempted zero-copy import of a buffer whose data pointer is not aligned as required by XLA; a copy is required to fix alignment, but the user forbade it. The XlaRuntimeError is re-raised as a ValueError with the alignment detail.

Source

Thrown at jax/_src/dlpack.py:286

    stream = None
  else:
    try:
      stream = dlpack_device.get_stream_for_external_ready_events()
    except _jax.JaxRuntimeError as err:
      if "UNIMPLEMENTED" in str(err):
        stream = None
      else:
        raise
  dlpack = external_array.__dlpack__(stream=stream)

  try:
    arr = _jax.dlpack_managed_tensor_to_buffer(
      dlpack, dlpack_device, stream, copy, int(dl_device_type))
  except xla_client.XlaRuntimeError as e:
    se = str(e)
    if "is not aligned to" in se:
      i = se.index("is not aligned to")
      raise ValueError(
        "Specified input which requires a copy since the source data "
        f"buffer {se[i:]} However copy=False. Set copy=True or "
        "copy=None to perform the requested operation."
      )
    else:
      raise
  # TODO(phawkins): when we are ready to support x64 arrays in
  # non-x64 mode, change the semantics to not canonicalize here.
  arr = jnp.asarray(arr, dtype=dtypes.canonicalize_dtype(arr.dtype))
  if copy:
    # copy was already handled by dlpack_managed_tensor_to_buffer.
    copy = None
  return _place_array(arr, device, dlpack_device, copy)

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

Use copy=True or copy=None (default) to let JAX realign
Copy/clone the tensor in the source framework to get an aligned base buffer
Avoid byte-offset views when passing buffers to JAX

Example fix

# before
jax.dlpack.from_dlpack(t[3:], copy=False)

# after
jax.dlpack.from_dlpack(t[3:].clone(), copy=False)  # or copy=True

Defensive patterns

Strategy: try-catch

Validate before calling

# avoid byte-offset views destined for zero-copy interop
t = t if t.data_ptr() % 64 == 0 else t.clone()  # torch example

Try / catch

try:
    jax.dlpack.from_dlpack(t, copy=False)
except ValueError as e:
    if 'not aligned' not in str(e) and 'copy=False' not in str(e): raise
    jax.dlpack.from_dlpack(t, copy=True)

Prevention

Use copy=None by default
Clone sliced tensors before zero-copy interop

When it happens

Trigger: jax.dlpack.from_dlpack(t, copy=False) where the producer's underlying buffer is misaligned (e.g. a torch tensor slice offset by a non-aligned number of bytes, or custom C-allocated buffers).

Common situations: Slicing tensors then interopping with JAX; producers that allocate unaligned storage; views with odd byte offsets (e.g. uint8 views into larger buffers).

Related errors

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/43681e5f7d0b4370. Report an issue: GitHub.