jax-ml/jax · error · nb::value_error

Buffer.__dlpack__ with copy=True is not supported.

Error message

Buffer.__dlpack__ with copy=True is not supported.

What it means

The FFI Buffer's __dlpack__ exports the buffer zero-copy; copy=True is not implemented and is explicitly rejected so callers don't assume a copied tensor.

Source

Thrown at jaxlib/ffi.cc:461

          throw nb::value_error(
              "dtype parameter is not supported by Buffer.__array__.");
        }
        if (!copy.is_none() && nb::cast<bool>(copy)) {
          throw nb::value_error(
              "Buffer.__array__ with copy=True is not supported.");
        }
        return xla::ValueOrThrow(self.NumpyArray());
      },
      nb::arg("dtype") = nb::none(), nb::arg("copy") = nb::none());
  buffer.def_prop_ro(
      "__cuda_array_interface__",
      xla::ValueOrThrowWrapper(&PyFfiAnyBuffer::CudaArrayInterface));
  buffer.def(
      "__dlpack__",
      [](PyFfiAnyBuffer self, nb::object stream, nb::object max_version,
         nb::object dl_device, nb::object copy) {
        if (!copy.is_none() && nb::cast<bool>(copy)) {
          throw nb::value_error(
              "Buffer.__dlpack__ with copy=True is not supported.");
        }

        // Fall back on the non-versioned API if unsupported by the requested
        // max_version.
        nb::tuple max_version_tuple;
        int64_t max_version_major;
        if (!nb::try_cast<nb::tuple>(max_version, max_version_tuple) ||
            max_version_tuple.size() < 2 ||
            !nb::try_cast<int64_t>(max_version_tuple[0], max_version_major) ||
            max_version_major < 1) {
          return xla::ValueOrThrow(self.DLPack());
        }

        // TODO(danfm): Handle other optional inputs.
        return xla::ValueOrThrow(self.DLPackVersioned());
      },
      nb::arg("stream") = nb::none(), nb::arg("max_version") = nb::none(),

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Omit copy (get the zero-copy export) and let the consumer copy if needed
2. Patch the consumer to pass copy=False/None
3. Ensure the buffer outlives the consumer tensor to keep zero-copy safe

Example fix

# before
consumer.from_dlpack(buf.__dlpack__(copy=True))
# after
consumer.from_dlpack(buf.__dlpack__())

capsule = buf.__dlpack__()  # no copy kwarg

When it happens

Trigger: buffer.__dlpack__(copy=True) on an xla.ffi Buffer, typically from torch.from_dlpack or generic dlpack consumers that pass copy under dlpack>=1.0 protocol.

Common situations: Interop code that unconditionally passes copy=True to __dlpack__ on NumPy 2 / DLPack 1.0 consumers.

Related errors

• Specified {device=} which requires a copy since the source d
• Specified {device=} which requires a copy since the source d
• Buffer.__array__ with copy=True is not supported.
• to_dlpack can only pack a dlpack tensor from an array on a s
• __dlpack__ only supported for unsharded arrays.

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/82e83ef83f9989aa. Report an issue: GitHub.