Context

OpenVINO component responsible for support of JAX/Flax models is called as JAX Frontend (JAX FE). JAX FE converts a JAX/Flax model represented by ClosedJAXpr graph object with operations from jax.lax opset to OpenVINO IR containing operations from OpenVINO opset.

In order to infer JAX/Flax models containing jax.lax.scatter operation by OpenVINO, JAX FE needs to be extended with this operation support.

What needs to be done?

For jax.lax.scatter operation support, you need to implement the corresponding loader into JAX FE op directory and to register it into the dictionary of Loaders. One loader is responsible for conversion (or decomposition) of one type of JAX operation.

Here is an example of loader implementation for jax.lax.reshape operation:

OutputVector translate_reshape(const NodeContext& context) {
    num_inputs_check(context, 1, 1);
    Output<Node> input = context.get_input(0);
    auto new_sizes = context.const_named_param<std::vector<int64_t>>("new_sizes");
    if (context.has_param("dimensions")) {
        auto dimensions = context.const_named_param<std::vector<int64_t>>("dimensions");
        // transpose the input first.
        auto permutation_node = std::make_shared<v0::Constant>(element::i64, Shape{dimensions.size()}, dimensions);
        input = std::make_shared<v1::Transpose>(input, permutation_node);
    }

    auto new_shape_node = std::make_shared<v0::Constant>(element::i64, Shape{new_sizes.size()}, new_sizes);
    Output<Node> res = std::make_shared<v1::Reshape>(input, new_shape_node, false);
    return {res};
};

In this example, translate_reshape expresses jax.lax.reshape using OpenVINO opset. Since jax.lax.reshape performs transposition and tensor reshaping according to JAX documentation, the resulted decomposition contains OpenVINO Transpose and Reshape operations. For Transpose and Reshape nodes, this conversion parses constant parameters dimensions to permute input tensor and new_size that is the target shape of the result.

Once you are done with implementation of the translator, you need to implement the corresponding layer tests test_scatter.py and put it into layer_tests/jax_tests directory. Example how to run some layer test:

export TEST_DEVICE=CPU
export JAX_TRACE_MODE=JAXPR
export 
cd openvino/tests/layer_tests/jax_tests
pytest test_reshape.py

Example Pull Requests

• [JAX FE] Support concatenate, integer_pow, reduce ops #26288
• [JAX FE] Support operations for Vision Transformer model #26254

Resources

• What is OpenVINO?
• How to Build OpenVINO
• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers

Contact points

• @openvinotoolkit/openvino-jax-frontend-maintainers
• @rkazants in GitHub and Discord

Ticket

No response