This issue comes from a Codex global scan of deepmodeling/deepmd-kit at commit 73de44b1f94471b2e3bdb6b11f57b34d7bc791bb.

Problem

The JAX C++ API exposes has_default_fparam(), but it does not load or materialize the default frame parameter values before constructing C API tensors.

The JAX export wrapper preserves both the flag and the default values:

• deepmd-kit/deepmd/jax/jax2tf/tfmodel.py

  Lines 72 to 80 in 73de44b

  	if hasattr(self.model, "has_default_fparam"):
  	# No attrs before v3.1.2
  	self._has_default_fparam = self.model.has_default_fparam().numpy().item()
  	else:
  	self._has_default_fparam = False
  	if hasattr(self.model, "get_default_fparam"):
  	self.default_fparam = self.model.get_default_fparam().numpy().tolist()
  	else:
  	self.default_fparam = None

• deepmd-kit/deepmd/jax/jax2tf/tfmodel.py

  Lines 345 to 351 in 73de44b

  	def has_default_fparam(self) -> bool:
  	"""Check whether the model has default frame parameters."""
  	return self._has_default_fparam
  	def get_default_fparam(self) -> list[float] | None:
  	"""Get the default frame parameters."""
  	return self.default_fparam

DeepPotJAX reads only the boolean:

• deepmd-kit/source/api_cc/src/DeepPotJAX.cc

  Lines 335 to 340 in 73de44b

  	try {
  	has_default_fparam_ = get_scalar<bool>(ctx, "has_default_fparam",
  	func_vector, device, status);
  	} catch (tf_function_not_found& e) {
  	has_default_fparam_ = false;
  	}

The header stores no default fparam vector, only the flag:

• deepmd-kit/source/api_cc/include/DeepPotJAX.h

  Lines 101 to 103 in 73de44b

  	bool has_default_fparam() const {
  	assert(inited);
  	return has_default_fparam_;

• deepmd-kit/source/api_cc/include/DeepPotJAX.h

  Lines 201 to 203 in 73de44b

  	bool do_message_passing;
  	// has default fparam
  	bool has_default_fparam_;

During inference, the direct path copies the caller-provided fparam vector and always shapes it as {nframes, dfparam}:

• deepmd-kit/source/api_cc/src/DeepPotJAX.cc

  Lines 397 to 426 in 73de44b

  	std::vector<double> coord_double(coord.begin(), coord.end());
  	std::vector<double> box_double(box.begin(), box.end());
  	std::vector<double> fparam_double(fparam.begin(), fparam.end());
  	std::vector<double> aparam_double(aparam.begin(), aparam.end());
  	TFE_Op* op;
  	if (atomic) {
  	op = get_func_op(ctx, "call_with_atomic_virial", func_vector, device,
  	status);
  	} else {
  	op = get_func_op(ctx, "call_without_atomic_virial", func_vector, device,
  	status);
  	}
  	std::vector<TFE_TensorHandle*> input_list(5);
  	std::vector<TF_Tensor*> data_tensor(5);
  	// coord
  	std::vector<int64_t> coord_shape = {nframes, nloc_real, 3};
  	input_list[0] =
  	add_input(op, coord_double, coord_shape, data_tensor[0], status);
  	// atype
  	std::vector<int64_t> atype_shape = {nframes, nloc_real};
  	input_list[1] = add_input(op, atype, atype_shape, data_tensor[1], status);
  	// box
  	int box_size = box_double.size() > 0 ? 3 : 0;
  	std::vector<int64_t> box_shape = {nframes, box_size, box_size};
  	input_list[2] = add_input(op, box_double, box_shape, data_tensor[2], status);
  	// fparam
  	std::vector<int64_t> fparam_shape = {nframes, dfparam};
  	input_list[3] =
  	add_input(op, fparam_double, fparam_shape, data_tensor[3], status);

The neighbor-list path has the same behavior:

• deepmd-kit/source/api_cc/src/DeepPotJAX.cc

  Lines 546 to 548 in 73de44b

  	std::vector<double> coord_double(coord.begin(), coord.end());
  	std::vector<double> fparam_double(fparam.begin(), fparam.end());
  	std::vector<double> aparam_double(aparam.begin(), aparam.end());

• deepmd-kit/source/api_cc/src/DeepPotJAX.cc

  Lines 622 to 625 in 73de44b

  	// fparam
  	std::vector<int64_t> fparam_shape = {nframes, dfparam};
  	input_list[4] =
  	add_input(op, fparam_double, fparam_shape, data_tensor[4], status);

create_tensor() allocates bytes from data.size(), not from the requested TensorFlow shape:

• deepmd-kit/source/api_cc/src/DeepPotJAX.cc

  Lines 176 to 183 in 73de44b

  	template <typename T>
  	inline TF_Tensor* create_tensor(const std::vector<T>& data,
  	const std::vector<int64_t>& shape) {
  	TF_Tensor* tensor =
  	TF_AllocateTensor(get_data_tensor_type(data), shape.data(), shape.size(),
  	data.size() * sizeof(T));
  	memcpy(TF_TensorData(tensor), data.data(), TF_TensorByteSize(tensor));
  	return tensor;

Therefore, when the model has a default fparam and the caller omits fparam, the C++ API reports that a default exists but still tries to build a non-empty {nframes, dfparam} tensor from an empty vector.

Impact

JAX models with default frame parameters can fail or send an invalid empty-buffer tensor through the C++ API unless callers explicitly provide fparam, even though the API advertises default fparam support.

Suggested fix

Load get_default_fparam during JAX model initialization, store it in DeepPotJAX, and materialize/tile it when the caller passes an empty fparam and has_default_fparam_ is true. Also validate create_tensor() inputs so the byte count matches the requested shape before allocating the TensorFlow tensor.