jax2onnx 0.8.0

export JAX to ONNX

jax2onnx 🌟

jax2onnx converts your JAX, Flax(nnx), Equinox functions directly into the ONNX format.

✨ Key Features

• Simple API
  Easily convert JAX callables—including Flax (NNX) and Equinox models—into ONNX format using to_onnx(...).

• Model structure preserved
  With @onnx_function, submodules appear as named functions in the ONNX graph (e.g. in Netron). Useful for readability and reuse.

• Dynamic input support
  Use abstract dimensions like 'B' or pass scalars as runtime inputs. Models stay flexible without retracing.

• Plugin-based extensibility
  Add support for new primitives by writing small, local plugins.

• Netron-friendly outputs
  All generated ONNX graphs include shape/type annotations and are structured for clear visualization.

---

🚀 Quickstart

Convert your JAX callable to ONNX in just a few lines:

import onnx
from flax import nnx
from jax2onnx import to_onnx

# Define a simple MLP (from Flax docs)
class MLP(nnx.Module):
    def __init__(self, din, dmid, dout, *, rngs): 
        self.linear1 = nnx.Linear(din, dmid, rngs=rngs)
        self.dropout = nnx.Dropout(rate=0.1, rngs=rngs)
        self.bn = nnx.BatchNorm(dmid, rngs=rngs)
        self.linear2 = nnx.Linear(dmid, dout, rngs=rngs) 
    def __call__(self, x): 
        x = nnx.gelu(self.dropout(self.bn(self.linear1(x))))
        return self.linear2(x)

# Instantiate model
my_callable = MLP(din=30, dmid=20, dout=10, rngs=nnx.Rngs(0))

# Convert to ONNX
onnx_model = to_onnx(my_callable, [("B", 30)])

# Save the model
onnx.save_model(onnx_model, "my_callable.onnx")

🔎 See it visualized: my_callable.onnx

---

🧠 ONNX Functions — Minimal Example

ONNX functions help encapsulate reusable subgraphs. Simply use the @onnx_function decorator to make your callable an ONNX function. Just an @onnx_function decorator to make your callable an ONNX function

from onnx import save_model
from flax import nnx
from jax2onnx import onnx_function, to_onnx

# just an @onnx_function decorator to make your callable an ONNX function
@onnx_function
class MLPBlock(nnx.Module):
  def __init__(self, dim, *, rngs):
    self.linear1 = nnx.Linear(dim, dim, rngs=rngs)
    self.linear2 = nnx.Linear(dim, dim, rngs=rngs)
    self.batchnorm = nnx.BatchNorm(dim, rngs=rngs)
  def __call__(self, x):
    return nnx.gelu(self.linear2(self.batchnorm(nnx.gelu(self.linear1(x)))))

# Use it inside another module
class MyModel(nnx.Module):
  def __init__(self, dim, *, rngs):
    self.block1 = MLPBlock(dim, rngs=rngs)
    self.block2 = MLPBlock(dim, rngs=rngs)
  def __call__(self, x):
    return self.block2(self.block1(x))

callable = MyModel(256, rngs=nnx.Rngs(0))
model = to_onnx(callable, [(100, 256)])
save_model(model, "docs/onnx/model_with_function.onnx")

🔎 See it visualized: model_with_function.onnx

---

📅 Roadmap and Releases

Planned Versions

• Ongoing

  • Expanding coverage of JAX, Flax (NNX) and Equinox components
  • Enhancing support for physics-based simulations

• Upcoming

  • Advanced ONNX function support, including function reuse, optimized internal graph structure, and improved variable naming for clarity and readability
  • Integrating onnx-ir as a backend to improve ONNX model construction, memory efficiency, and performance

Current Productive Version

• 0.8.0 (initial Equinox support, PyPI):

  • Added initial support for Equinox components: eqx.dropout, eqx.layer_norm, and eqx.linear. Introduced an Equinox MlpExample showcasing Linear + Dropout composition.
  • Stabilized SSA and shape handling across lax.scan and lax.fori_loop bodies, preventing dtype leaks.
  • Improved dtype propagation in lax.gather, lax.concatenate.
  • Added plugin support for lax.pad.

Past Versions

• 0.7.5 fixes tests for functions without arguments, adds support for lax.bitwise_not, lax.clamp, lax.ge, jnp.clip, lax.rev, and enhances support for nnx.dot_product_attention, nnx.conv, nnx.batch_norm, lax.mul, lax.reduce_max, lax.scan, lax.slice, lax.while_loop, nn.gelu, jnp.arange, jnp.cumsum, jnp.select, jnp.where, and jnp.concatenate.
• 0.7.4 adds support for lax.cumsum and jnp.cumsum, and improves lax.scatter.
• 0.7.3 improves polymorphism handling for transformers.
• 0.7.2 adds support for jnp.split, lax.split, lax.logistic, includes an example for nnx.GRUCell, and improves lax.scatter and lax.while_loop.
• 0.7.1 fixes a numeric equivalence bug in the test system, and adds support for core.custom_jvp_generic, eqx.identity, jnp.select, jnp.stack, jnp.unstack, lax.select, plus multiple nn.* activations (identity, celu, elu, gelu, relu, leaky_relu, mish, selu, sigmoid, soft-sign, softmax, truncated_normal).
• 0.7.0 introduces a GPT-2 example based on nanoGPT with ONNX function support and attention masking, adds support for jnp.concatenate, jnp.take, nnx.Embed, and starts hosting ONNX models on Hugging Face.
• 0.6.5 improves support for nnx.batch_norm, nnx.group_norm, nnx.layer_norm, nnx.rms_norm, lax.broadcast_in_dim, lax.cond, lax.fori_loop, lax.integer_pow, lax.scan, lax.scatter, lax.scatter_add, lax.scatter_mul, and lax.while_loop; and adds support for lax.and, lax.rem, and lax.remat2.
• 0.6.4: Improved support for lax.scatter_mul.
• 0.6.3: Double precision fixes for lax.fori_loop and lax.while_loop. Fixed bugs in lax.scan and jnp.where.
• 0.6.2: Fixed bugs in nnx.conv and lax.reshape; added new primitive jnp.prod.
• 0.6.1: Improved support for lax.cond and lax.select_n; added new primitives (lax.reduce_and, lax.reduce_or, lax.reduce_prod, lax.reduce_xor); and introduced new examples for jnp.select and jnp.sort.
• 0.6.0: Introduced the enable_double_precision parameter (default: False) to support physics simulations, and enhanced handling of lax.scatter.
• 0.5.2: Add support for additional primitives: jnp.where, jnp.arange, jnp.linspace.
• 0.5.1: Add support for subgraph using primitives: lax.while_loop, lax.cond, lax.fori_loop, lax.scan.
• 0.5.0: Improved dynamic batch dimension handling by leveraging shape polymorphism for more robust and flexible model export. Added support for jnp.sign, jnp.abs, jnp.iota primitives.
• 0.4.4: Added support for lax.cos, lax.cosh, lax.sin, lax.sinh and lax.scatter primitives.
• 0.4.3: Fixed a bug in the validation of JAX callable outputs against their ONNX counterparts. This fix exposed previously hidden failing tests, which are now fixed.
• 0.4.2: Cleanup and fixes to the basic ONNX function release.
• 0.4.1 (ONNX functions): Introducing simple ONNX function support. Making use of ONNX functions is easy for the user: just a @onnx_function decorator making a callable an ONNX function. Each @onnx_function decorator creates a new ONNX function instance on the call graph.
• 0.3.2: relaxed the minimum Python version to 3.10.
• 0.3.0: Streamlined the plugin system with automatic registration and simplified integration of custom primitives.
• 0.2.0 (First PyPI Release): Rebased the implementation on jaxpr, improving usability and adding low-level lax components.
• 0.1.0 (Initial Approach, Not Released to PyPI): Produced ONNX exports for some nnx components and nnx-based examples, including a VisualTransformer.

---

❓ Troubleshooting

If conversion doesn't work out of the box, it could be due to:

• Non-dynamic function references:
  JAXPR-based conversion requires function references to be resolved dynamically at call-time.
  Solution: Wrap your function call inside a lambda to enforce dynamic resolution:

  my_dynamic_callable_function = lambda x: original_function(x)
  

• Unsupported primitives:
  The callable may use a primitive not yet or not fully supported by jax2onnx.
  Solution: Write a plugin to handle the unsupported function (this is straightforward!).

---

🧩 Supported JAX/ONNX Components

JAX Component	ONNX Components	Testcases	Since
[core.custom_jvp_generic](Generic passthrough for custom JVP calls)	CustomJvp	custom_jvp_square ✅ custom_jvp_square_f64 ✅	v0.7.1
core.dim_as_value	Cast Gather Reshape Shape	dim_as_value_dynamic ✅ dim_as_value ✅	v0.5.0
eqx.dropout	Dropout Not	eqx_dropout_inference_mode ✅ eqx_dropout_inference_mode_f64 ✅ eqx_dropout_training_mode ✅ eqx_dropout_training_mode_f64 ✅ eqx_dropout_dynamic_inference ✅ eqx_dropout_dynamic_inference_f64 ✅ eqx_dropout_batched_inference_dynamic ✅ eqx_dropout_batched_inference_dynamic_f64 ✅ eqx_dropout_batched_inference ✅ eqx_dropout_batched_inference_f64 ✅	v0.8.0
eqx.identity	Identity	eqx_identity_static ✅ eqx_identity_static_f64 ✅ eqx_identity_symbolic_batch_dynamic ✅ eqx_identity_symbolic_batch_dynamic_f64 ✅ eqx_identity_symbolic_batch ✅ eqx_identity_symbolic_batch_f64 ✅	v0.7.1
eqx.layer_norm	LayerNormalization	layer_norm ✅ layer_norm_f64 ✅ layer_norm_multiaxis ✅ layer_norm_multiaxis_f64 ✅ batched_layer_norm_dynamic ✅ batched_layer_norm_dynamic_f64 ✅ batched_layer_norm ✅ batched_layer_norm_f64 ✅ layer_norm_no_bias_no_scale ✅ layer_norm_no_bias_no_scale_f64 ✅	v0.8.0
eqx.linear	Gemm Reshape	eqx_linear_symbolic_batch_dynamic ✅ eqx_linear_symbolic_batch_dynamic_f64 ✅ eqx_linear_symbolic_batch ✅ eqx_linear_symbolic_batch_f64 ✅ eqx_linear_high_rank ✅ eqx_linear_high_rank_f64 ✅ eqx_linear_vector ✅ eqx_linear_vector_f64 ✅	v0.8.0
jnp.add	Add	add ✅ add_f64 ✅	v0.1.0
jnp.arange	Range	arange_data_dependent_indices ✅ arange_stop_only_concrete_input_val ✅ arange_stop_only_concrete_input_val_f64 ✅ arange_start_stop_concrete_input_val ✅ arange_start_stop_concrete_input_val_f64 ✅ arange_start_stop_step_concrete_input_val ✅ arange_start_stop_step_concrete_input_val_f64 ✅ arange_float_concrete_input_val ✅ arange_float_concrete_input_val_f64 ✅ arange_static_stop_only_int ✅ arange_static_stop_only_int_f64 ✅ arange_static_stop_only_float ✅ arange_static_stop_only_float_f64 ✅ arange_static_start_stop_int ✅ arange_static_start_stop_int_f64 ✅ arange_static_start_stop_step_int ✅ arange_static_start_stop_step_int_f64 ✅ arange_static_empty_result_pos_step ✅ arange_static_empty_result_pos_step_f64 ✅ arange_static_empty_result_neg_step ✅ arange_static_empty_result_neg_step_f64 ✅ arange_static_negative_step ✅ arange_static_negative_step_f64 ✅ arange_static_float_step_explicit_dtype ✅ arange_static_float_step_explicit_dtype_f64 ✅ arange_static_float_step_inferred_dtype ✅ arange_static_float_step_inferred_dtype_f64 ✅ arange_static_stop_zero ✅ arange_static_stop_zero_f64 ✅ arange_static_start_equals_stop ✅ arange_static_start_equals_stop_f64 ✅ arange_static_large_numbers_int ✅ arange_static_large_numbers_int_f64 ✅	v0.5.2
jnp.clip	Max Min	clip_i32_scalar_bounds ✅ clip_i32_scalar_bounds_f64 ✅ clip_f32_scalar_bounds_no_upcast_f64_mode ✅ clip_only_upper ✅ clip_only_upper_f64 ✅ clip_only_lower ✅ clip_only_lower_f64 ✅ clip_broadcast_bounds ✅	v0.7.5
jnp.concatenate	Cast Concat	concatenate_basic ✅ concatenate_basic_f64 ✅ concatenate_mixed_dtypes ✅ concatenate_mixed_dtypes_f64 ✅ concatenate_with_explicit_dtype ✅ concatenate_with_explicit_dtype_f64 ✅ concatenate_with_explicit_dtype_casts_inputs ✅ concatenate_abstract_middle_dim_dynamic ✅ concatenate_abstract_middle_dim_dynamic_f64 ✅ concatenate_abstract_middle_dim ✅ concatenate_abstract_middle_dim_f64 ✅ concatenate_tile_and_symbolic_dynamic ✅ concatenate_tile_and_symbolic_dynamic_f64 ✅ concatenate_tile_and_symbolic ✅ concatenate_tile_and_symbolic_f64 ✅	v0.2.0
jnp.cumsum	CumSum	cumsum_axis2_i32 ✅ cumsum_axis2_i32_f64 ✅ cumsum_axis2_reverse_i32 ✅ cumsum_axis2_reverse_i32_f64 ✅	v0.7.4
jnp.einsum	Einsum	einsum_vector_dot ✅ einsum_vector_dot_f64 ✅ einsum_matrix_vector ✅ einsum_matrix_vector_f64 ✅ einsum_matrix_matrix_dynamic ✅ einsum_matrix_matrix_dynamic_f64 ✅ einsum_matrix_matrix ✅ einsum_matrix_matrix_f64 ✅ einsum_transpose ✅ einsum_transpose_f64 ✅ einsum_batch_transpose_dynamic ✅ einsum_batch_transpose_dynamic_f64 ✅ einsum_batch_transpose ✅ einsum_batch_transpose_f64 ✅ einsum_diag ✅ einsum_diag_f64 ✅ einsum_sum_reduce ✅ einsum_sum_reduce_f64 ✅ einsum_multi_operand ✅ einsum_multi_operand_f64 ✅ einsum_attention_logits_orig_dynamic ✅ einsum_attention_logits_orig_dynamic_f64 ✅ einsum_attention_logits_orig ✅ einsum_attention_logits_orig_f64 ✅ einsum_attention_output_orig_dynamic ✅ einsum_attention_output_orig_dynamic_f64 ✅ einsum_attention_output_orig ✅ einsum_attention_output_orig_f64 ✅ einsum_attention_logits_batched_dynamic ✅ einsum_attention_logits_batched_dynamic_f64 ✅ einsum_attention_logits_batched ✅ einsum_attention_logits_batched_f64 ✅ einsum_attention_output_batched_dynamic ✅ einsum_attention_output_batched_dynamic_f64 ✅ einsum_attention_output_batched ✅ einsum_attention_output_batched_f64 ✅ einsum_ellipsis_rank_mismatch ✅ einsum_ellipsis_rank_mismatch_f64 ✅ einsum_attention_logits_batched_rank_mismatch ✅ einsum_attention_logits_batched_rank_mismatch_f64 ✅	v0.1.0
jnp.linspace	Constant	linspace_static_basic ✅ linspace_static_basic_f64 ✅ linspace_static_endpoint_false ✅ linspace_static_endpoint_false_f64 ✅ linspace_static_num_1 ✅ linspace_static_num_1_f64 ✅ linspace_static_num_0 ✅ linspace_static_num_0_f64 ✅ linspace_static_int_inputs_default_dtype ✅ linspace_static_int_inputs_default_dtype_f64 ✅	v0.5.2
jnp.matmul	MatMul	matmul_2d ✅ matmul_2d_f64 ✅ matmul_1d_2d ✅ matmul_1d_2d_f64 ✅ matmul_2d_1d ✅ matmul_2d_1d_f64 ✅ matmul_dynamic_dynamic ✅ matmul_dynamic_dynamic_f64 ✅ matmul_dynamic ✅ matmul_dynamic_f64 ✅ matmul_dynamic_a_dynamic ✅ matmul_dynamic_a_dynamic_f64 ✅ matmul_dynamic_a ✅ matmul_dynamic_a_f64 ✅ matmul_1d ✅ matmul_1d_f64 ✅ matmul_3d ✅ matmul_3d_f64 ✅	v0.1.0
jnp.prod	ReduceProd	basic_prod ✅ basic_prod_f64 ✅ prod_with_axis ✅ prod_with_axis_f64 ✅ prod_with_keepdims ✅ prod_with_keepdims_f64 ✅	v0.6.2
jnp.reshape	Reshape	reshape_1 ✅ reshape_1_f64 ✅ reshape_2 ✅ reshape_2_f64 ✅ reshape_3 ✅ reshape_3_f64 ✅ reshape_4_dynamic ✅ reshape_4_dynamic_f64 ✅ reshape_4 ✅ reshape_4_f64 ✅ reshape_to_scalar ✅ reshape_to_scalar_f64 ✅ reshape_from_scalar ✅ reshape_from_scalar_f64 ✅ reshape_cnn_dynamic ✅ reshape_cnn_dynamic_f64 ✅ reshape_cnn ✅ reshape_cnn_f64 ✅ reshape_valid_flatten_trailing ✅ reshape_valid_flatten_trailing_f64 ✅ reshape_with_target_shape_from_symbolic_dim_computation ✅ reshape_with_target_shape_from_symbolic_dim_computation_f64 ✅	v0.1.0
jnp.select	Where	select_simple ✅ select_simple_f64 ✅ select_broadcast ✅ select_broadcast_f64 ✅ select_gpt2_attention_mask_dynamic ✅ select_gpt2_attention_mask_dynamic_f64 ✅ select_gpt2_attention_mask ✅ select_gpt2_attention_mask_f64 ✅	v0.7.1
jnp.shape	Shape	shape_basic ✅ shape_basic_f64 ✅ shape_dynamic_dynamic ✅ shape_dynamic_dynamic_f64 ✅ shape_dynamic ✅ shape_dynamic_f64 ✅	0.4.0
jnp.sort	TopK	sort_1d ✅ sort_1d_f64 ✅ sort_2d_axis0_dynamic ✅ sort_2d_axis0_dynamic_f64 ✅ sort_2d_axis0 ✅ sort_2d_axis0_f64 ✅	v0.5.2
jnp.split	Split	split_by_sections ✅ split_by_sections_f64 ✅ split_by_indices ✅ split_by_indices_f64 ✅ split_by_indices_symbolic_dynamic ✅ split_by_indices_symbolic_dynamic_f64 ✅ split_by_indices_symbolic ✅ split_by_indices_symbolic_f64 ✅	v0.7.2
jnp.squeeze	Squeeze	squeeze_single_dim ✅ squeeze_single_dim_f64 ✅ squeeze_multiple_dims ✅ squeeze_multiple_dims_f64 ✅ squeeze_vit_output ✅ squeeze_vit_output_f64 ✅ squeeze_dynamic_batch_dynamic ✅ squeeze_dynamic_batch_dynamic_f64 ✅ squeeze_dynamic_batch ✅ squeeze_dynamic_batch_f64 ✅ squeeze_all_dims ✅ squeeze_all_dims_f64 ✅ squeeze_negative_axis ✅ squeeze_negative_axis_f64 ✅ squeeze_negative_axis_tuple ✅ squeeze_negative_axis_tuple_f64 ✅ squeeze_dynamic_and_negative_axis_dynamic ✅ squeeze_dynamic_and_negative_axis_dynamic_f64 ✅ squeeze_dynamic_and_negative_axis ✅ squeeze_dynamic_and_negative_axis_f64 ✅	v0.1.0
jnp.stack	Concat Unsqueeze	stack_axis_0 ✅ stack_axis_0_f64 ✅ stack_axis_1 ✅ stack_axis_1_f64 ✅ stack_negative_axis ✅ stack_negative_axis_f64 ✅ stack_scalars ✅ stack_scalars_f64 ✅	v0.7.1
jnp.take	Gather	take_data_dependent_indices ✅	v0.7.0
jnp.tile	Tile	tile_repeats ✅ tile_repeats_f64 ✅ tile_a ✅ tile_a_f64 ✅ tile_b ✅ tile_b_f64 ✅ tile_c ✅ tile_c_f64 ✅ tile_d ✅ tile_d_f64 ✅ tile_dynamic_input_static ✅ tile_dynamic_input_static_f64 ✅ tile_dynamic_input_dynamic ✅ tile_dynamic_input_dynamic_f64 ✅ tile_dynamic_input ✅ tile_dynamic_input_f64 ✅ tile_pad ✅ tile_pad_f64 ✅ tile_with_symbolic_repeats_static ✅ tile_with_symbolic_repeats_static_f64 ✅ tile_with_symbolic_repeats_dynamic ✅ tile_with_symbolic_repeats_dynamic_f64 ✅ tile_with_symbolic_repeats ✅ tile_with_symbolic_repeats_f64 ✅ tile_param_symbolic_dynamic ✅ tile_param_symbolic_dynamic_f64 ✅ tile_param_symbolic ✅ tile_param_symbolic_f64 ✅	v0.1.0
jnp.transpose	Transpose	transpose_basic ✅ transpose_basic_f64 ✅ transpose_reverse ✅ transpose_reverse_f64 ✅ transpose_4d_dynamic ✅ transpose_4d_dynamic_f64 ✅ transpose_4d ✅ transpose_4d_f64 ✅ transpose_square_matrix ✅ transpose_square_matrix_f64 ✅ transpose_high_dim ✅ transpose_high_dim_f64 ✅ transpose_no_axes ✅ transpose_no_axes_f64 ✅ transpose_3d_dynamic ✅ transpose_3d_dynamic_f64 ✅ transpose_3d ✅ transpose_3d_f64 ✅	v0.1.0
jnp.unstack	Split Squeeze	unstack_axis_0 ✅ unstack_axis_0_f64 ✅ unstack_axis_1 ✅ unstack_axis_1_f64 ✅ unstack_negative_axis ✅ unstack_negative_axis_f64 ✅	v0.7.1
jnp.where	Where	where_gpt_mask_scores_literal_else_dynamic ✅ where_gpt_mask_scores_literal_else_dynamic_f64 ✅ where_gpt_mask_scores_literal_else ✅ where_gpt_mask_scores_literal_else_f64 ✅ where_simple ✅ where_simple_f64 ✅ where_broadcast ✅ where_broadcast_f64 ✅ where_multidim_condition_scalar_branches_broadcast ✅ where_multidim_condition_scalar_branches_broadcast_f64 ✅ where_multidim_condition_scalar_branches_broadcast ✅ where_multidim_condition_scalar_branches_broadcast_f64 ✅ where_A ✅ where_A_f64 ✅ where_B ✅ where_B_f64 ✅ where_gpt_mask_scores_scalar_else_dynamic ✅ where_gpt_mask_scores_scalar_else_dynamic_f64 ✅ where_gpt_mask_scores_scalar_else ✅ where_gpt_mask_scores_scalar_else_f64 ✅ where_int_condition_cast ✅ where_int_condition_cast_f64 ✅ where_literal_else_pyfloat ✅ where_literal_else_pyfloat_f64 ✅ where_jax_int_literals_broadcast_f64_mode ✅ where_dtype_mismatch_f64_vs_i32_promote ✅ where_simple ✅ where_simple_f64 ✅	v0.5.2
lax.abs	Abs	abs ✅ abs_f64 ✅	v0.5.0
lax.add	Add	add ✅ add_f64 ✅	v0.2.0
lax.and	And BitwiseAnd	and_bool ✅ and_bool_f64 ✅ and_int ✅ and_int_f64 ✅	v0.6.5
lax.argmax	ArgMax	argmax_float_axis0 ✅ argmax_float_axis0_f64 ✅ argmax_float_axis1 ✅ argmax_float_axis1_f64 ✅ argmax_boolean_input_axis0_specific_values ✅ argmax_boolean_input_axis0_specific_values_f64 ✅ argmax_boolean_input_axis1_specific_values ✅ argmax_boolean_input_axis1_specific_values_f64 ✅ argmax_boolean_random_input_axis0 ✅ argmax_boolean_random_input_axis0_f64 ✅	v0.2.0
lax.argmin	ArgMin	argmin_test1 ✅ argmin_test1_f64 ✅ argmin_test2 ✅ argmin_test2_f64 ✅	v0.2.0
lax.bitwise_not	BitwiseNot	bitwise_not_bool ✅ bitwise_not_bool_f64 ✅ bitwise_not_i32 ✅ bitwise_not_i32_f64 ✅	v0.7.5
lax.broadcast_in_dim	Expand Identity Reshape	broadcast_in_dim ✅ broadcast_in_dim_f64 ✅ broadcast_in_dim_2d_to_3d ✅ broadcast_in_dim_2d_to_3d_f64 ✅ broadcast_in_dim_scalar ✅ broadcast_in_dim_scalar_f64 ✅ broadcast_in_dim_batch_dynamic ✅ broadcast_in_dim_batch_dynamic_f64 ✅ broadcast_in_dim_batch ✅ broadcast_in_dim_batch_f64 ✅ broadcast_in_dim_dynamic_B_dynamic ✅ broadcast_in_dim_dynamic_B_dynamic_f64 ✅ broadcast_in_dim_dynamic_B ✅ broadcast_in_dim_dynamic_B_f64 ✅	v0.2.0
lax.clamp	Max Min	clamp_i32_scalar_bounds ✅ clamp_i32_scalar_bounds_f64 ✅ clamp_scalar_float_bounds_match_x ✅ clamp_scalar_float_bounds_match_x_f64 ✅ clamp_vector_bounds_match ✅ clamp_pyint_bounds_promote_to_x_dtype ✅ clamp_pyint_bounds_promote_to_x_dtype_f64 ✅	v0.7.5
lax.concatenate	Cast Concat	concatenate ✅ concatenate_f64 ✅ concatenate_axis1_dynamic ✅ concatenate_axis1_dynamic_f64 ✅ concatenate_axis1 ✅ concatenate_axis1_f64 ✅ concatenate_axis0 ✅ concatenate_axis0_f64 ✅ concatenate_3d ✅ concatenate_3d_f64 ✅ concatenate_internal_int32_then_cast_to_f32_zeroarg ✅	v0.2.0
lax.cond	If	cond_scalar ✅ cond_scalar_f64 ✅ cond_multiple_operands_in_tuple ✅ cond_multiple_operands_in_tuple_f64 ✅ cond_my_new_complex_scenario ✅ cond_my_new_complex_scenario_f64 ✅ cond_nested_conditional ✅ cond_nested_conditional_f64 ✅ cond_variables ✅ cond_variables_f64 ✅ cond_internal_constant_f64 ✅ cond_passthrough_identity ✅ cond_passthrough_identity_f64 ✅ cond_with_scatter ✅ cond_with_scatter_f64 ✅	v0.5.1
lax.conv	Conv	conv ✅ conv2 ✅ conv_general_dilated_nhwc_output ✅	v0.2.0
lax.convert_element_type	Cast	convert_element_type ✅ convert_element_type_f64 ✅	v0.2.0
[lax.copy](Handles the JAX primitive lax.copy_p. Note: jax.lax.copy API is removed.)	Identity	copy_float32_array ✅ copy_int64_scalar ✅	<your_current_version>
lax.cos	Cos	cos ✅ cos_f64 ✅	v0.4.4
lax.cosh	Cosh	cosh ✅ cosh_f64 ✅	v0.4.4
lax.cumsum	CumSum	cumsum_i32_axis2 ✅ cumsum_i32_axis2_f64 ✅ cumsum_f32_axism1_reverse ✅ cumsum_f32_axism1_reverse_f64 ✅	v0.7.4
lax.device_put	Identity	device_put_array ✅ device_put_array_f64 ✅ device_put_scalar ✅ device_put_scalar_f64 ✅	v0.4.0
lax.div	Div	div ✅ div_f64 ✅	v0.2.0
lax.dot_general	MatMul	dot_general ✅ dot_general_f64 ✅ dot_general_lhs1_rhs1 ✅ dot_general_lhs1_rhs1_f64 ✅	v0.2.0
lax.dynamic_slice	Slice	dynamic_slice_test1 ✅ dynamic_slice_test1_f64 ✅ dynamic_slice_2d ✅ dynamic_slice_2d_f64 ✅ dynamic_slice_3d ✅ dynamic_slice_3d_f64 ✅ dynamic_slice_vit_like_dynamic ✅ dynamic_slice_vit_like_dynamic_f64 ✅ dynamic_slice_vit_like ✅ dynamic_slice_vit_like_f64 ✅	v0.1.0
lax.eq	Equal	eq ✅ eq_f64 ✅	v0.2.0
lax.exp	Exp	exp ✅ exp_f64 ✅	v0.2.0
lax.fori_loop	Loop	fori_loop_counter ✅ fori_loop_counter_f64 ✅ fori_loop_zero ✅ fori_loop_zero_f64 ✅ fori_loop_vector ✅ fori_loop_vector_f64 ✅ fori_loop_example ✅ fori_loop_example_f64 ✅ fori_loop_test ✅ fori_loop_test_f64 ➖	v0.5.1
lax.gather	GatherND	gather_trig_where_pipeline_f64_indices_i64 ✅ gather_trig_where_pipeline_f64_indices_i32 ✅ gather_f64_data_i64_indices_output_is_f64 ✅ gather_f64_data_i32_indices_cast_and_output_is_f64 ✅ gather_static ✅ gather_static_f64 ✅ gather_dynamic_batch_simple_index_dynamic ✅ gather_dynamic_batch_simple_index_dynamic_f64 ✅ gather_dynamic_batch_simple_index ✅ gather_dynamic_batch_simple_index_f64 ✅	v0.2.0
lax.greater_equal	GreaterOrEqual	greater_equal ✅ greater_equal_f64 ✅	v0.7.5
lax.gt	Greater	gt ✅ gt_f64 ✅	v0.2.0
lax.integer_pow	Pow	integer_pow ✅ integer_pow_f64 ✅	v0.2.0
lax.iota	Range	iota_int32 ✅ iota_int32_f64 ✅ iota_float32 ✅ iota_float32_f64 ✅ broadcasted_iota ✅ broadcasted_iota_f64 ✅	v0.5.0
lax.log	Log	log ✅ log_f64 ✅	v0.2.0
lax.logistic	Sigmoid	lax_logistic_basic ✅ lax_logistic_basic_f64 ✅	v0.7.2
lax.lt	Less	lt ✅ lt_f64 ✅	v0.2.0
lax.max	Max	max ✅ max_f64 ✅	v0.2.0
lax.min	Min	min_test1 ✅ min_test1_f64 ✅	v0.1.0
lax.mul	Mul	mul_test1 ✅ mul_test1_f64 ✅ mul_test2 ✅ mul_test2_f64 ✅ mul_pyfloat_promotes_to_array_dtype_f64 ✅ mul_scalar_broadcast_promote_to_f64 ✅	v0.1.0
lax.ne	Equal Not	ne ✅ ne_f64 ✅	v0.2.0
lax.neg	Neg	neg ✅ neg_f64 ✅	v0.2.0
lax.pad	Pad	pad_const_1d ✅ pad_const_1d_f64 ✅ pad_const_2d ✅ pad_const_2d_f64 ✅ pad_const_2d_cval ✅ pad_const_2d_cval_f64 ✅ pad_inside_scan_smoke_f64 ✅ pad_inside_nested_scan_smoke_f64 ✅	v0.8.0
lax.reduce_and	Cast ReduceMin	reduce_and_all_true ✅ reduce_and_all_true_f64 ✅ reduce_and_one_false ✅ reduce_and_one_false_f64 ✅ reduce_and_keepdims ✅ reduce_and_keepdims_f64 ✅	v0.6.1
lax.reduce_max	ReduceMax	reduce_max ✅ reduce_max_f64 ✅ reduce_max_allaxes ✅ reduce_max_allaxes_f64 ✅ reduce_max_keepdims ✅ reduce_max_keepdims_f64 ✅ reduce_max_axes_input ✅ reduce_max_axes_input_f64 ✅	v0.2.0
lax.reduce_min	ReduceMin	reduce_min ✅ reduce_min_f64 ✅ reduce_min_allaxes ✅ reduce_min_allaxes_f64 ✅ reduce_min_keepdims ✅ reduce_min_keepdims_f64 ✅	v0.2.0
lax.reduce_or	Cast ReduceMax	reduce_or_all_false ✅ reduce_or_all_false_f64 ✅ reduce_or_one_true ✅ reduce_or_one_true_f64 ✅ reduce_or_keepdims ✅ reduce_or_keepdims_f64 ✅	v0.6.1
lax.reduce_prod	ReduceProd	reduce_prod ✅ reduce_prod_f64 ✅ reduce_prod_allaxes ✅ reduce_prod_allaxes_f64 ✅ reduce_prod_keepdims ✅ reduce_prod_keepdims_f64 ✅ reduce_prod_dtype_f64 ✅ reduce_prod_dtype ✅	v0.6.1
lax.reduce_sum	ReduceSum	reduce_sum ✅ reduce_sum_f64 ✅ reduce_sum_allaxes ✅ reduce_sum_allaxes_f64 ✅ reduce_sum_keepdims ✅ reduce_sum_keepdims_f64 ✅ reduce_sum_dtype_f64 ✅ reduce_sum_dtype ✅	v0.2.0
lax.reduce_xor	Cast Mod ReduceSum	reduce_xor_all_false ✅ reduce_xor_all_false_f64 ✅ reduce_xor_one_true ✅ reduce_xor_one_true_f64 ✅ reduce_xor_two_true ✅ reduce_xor_two_true_f64 ✅ reduce_xor_keepdims ✅ reduce_xor_keepdims_f64 ✅	v0.6.1
lax.rem	Div Mod	rem_int ✅ rem_int_f64 ✅ rem_float ✅ rem_float_f64 ✅ rem_int_neg ✅ rem_int_neg_f64 ✅ rem_float_neg ✅ rem_float_neg_f64 ✅	v0.6.5
lax.reshape	Reshape	reshape ✅ reshape_f64 ✅ reshape_valid_squeeze_middle_dim_from_problematic_source ✅ reshape_valid_squeeze_middle_dim_from_problematic_source_f64 ✅ reshape_valid_flatten_trailing ✅ reshape_valid_flatten_trailing_f64 ✅ reshape_with_target_shape_from_symbolic_dim_computation ✅ reshape_with_target_shape_from_symbolic_dim_computation_f64 ✅ reshape_with_inferred_dimension_from_input_dynamic_dynamic ✅ reshape_with_inferred_dimension_from_input_dynamic_dynamic_f64 ✅ reshape_with_inferred_dimension_from_input_dynamic ✅ reshape_with_inferred_dimension_from_input_dynamic_f64 ✅ reshape_with_inferred_dimension_from_input ✅ reshape_with_inferred_dimension_from_input_f64 ✅ reshape_merge_symbolic_with_static_and_check_name_dynamic ✅ reshape_merge_symbolic_with_static_and_check_name ✅	v0.2.0
lax.rev	Flip	rev_vector ✅ rev_vector_f64 ✅ rev_matrix_axes01 ✅ rev_matrix_axes01_f64 ✅	v0.7.5
lax.scan	Scan	scan_cumsum ✅ scan_cumsum_f64 ✅ scan_carry_only ✅ scan_carry_only_f64 ✅ scan_multiple_sequences ✅ scan_multiple_sequences_f64 ✅ scan_multiple_carry ✅ scan_multiple_carry_f64 ✅ scan_matrix_carry_multidim_xs ✅ scan_matrix_carry_multidim_xs_f64 ✅ scan_no_xs ✅ scan_no_xs_f64 ✅ scan_fn ✅ scan_fn_f64 ✅ scan_jit_no_xs ✅ scan_jit_no_xs_f64 ✅ scan_captured_scalar ✅ scan_captured_scalar_f64 ✅ scan_rank0_sequence_vectorized ✅ scan_rank0_sequence_vectorized_f64 ✅ scan_two_diff_lengths ✅ scan_two_diff_lengths_f64 ✅ scan_two_diff_lengths ✅ scan_two_diff_lengths_f64 ✅ scan_nested_len_mismatch ✅ scan_nested_len_mismatch_f64 ✅ scan_two_diff_lengths_broadcast ✅ scan_two_diff_lengths_broadcast_f64 ✅ scan_two_diff_lengths_with_broadcast ✅ scan_two_diff_lengths_f64 ✅ scan_captured_scalar_with_xs ✅ scan_captured_vector_with_xs_f64 ✅	v0.5.1
lax.scatter	ScatterND	scatter_set_axis0 ✅ scatter_set_axis0_f64 ✅ scatter_set_middle ✅ scatter_set_middle_f64 ✅ scatter_correct_axis_determination ✅ scatter_correct_axis_determination_f64 ✅ scatter_updates_slice_needed_axis0 ✅ scatter_updates_slice_needed_axis0_f64 ✅ scatter_from_user_warning_shapes_valid_jax ✅ scatter_from_user_warning_shapes_valid_jax_f64 ✅ scatter_user_error_scenario_precise ✅ scatter_user_error_scenario_precise_f64 ✅ scatter_window_update_f64 ✅ scatter_window_update_depth3_shapes_ok ✅ scatter_static_slice_set_f64 ✅ scatter_depth2_fp64_type_mismatch ✅ scatter_simple_2d_window_out_of_bounds ✅ scatter_clip_2d_window_at_edge ✅ scatter_depth2_mixed_dtypes_fp_mismatch_f64 ✅ scatter_depth2_mixed_dtypes_fp_mismatch ✅	v0.4.4
lax.scatter_add	ScatterND	scatter_add_simple_1d ✅ scatter_add_simple_1d_f64 ✅ scatter_add_window_2d_operand_1d_indices ✅ scatter_add_window_2d_operand_1d_indices_f64 ✅ scatter_add_batch_updates_1d_operand ✅ scatter_add_batch_updates_1d_operand_f64 ✅ scatter_add_mismatched_window_dims_from_user_report ✅ scatter_add_mismatched_window_dims_from_user_report2 ✅ scatter_add_mismatched_window_dims_from_user_report3 ✅ scatter_add_fluids_pattern_updates_5_4_1_1 ✅ scatter_add_in_cond_float64 ✅ scatter_add_fp64_dtype_mismatch ✅ scatter_add_depth2_depth2_helper_regression ✅ scatter_depth2_fp64_type_mismatch ✅	v0.5.3
lax.scatter_max	ScatterND	scatter_max_simple_1d ✅ scatter_max_simple_1d_f64 ✅ scatter_max_window_2d_operand_1d_indices ✅ scatter_max_window_2d_operand_1d_indices_f64 ✅ scatter_max_batch_updates_1d_operand ✅ scatter_max_batch_updates_1d_operand_f64 ✅ scatter_max_fp64_dtype_path_check ✅ scatter_max_depth2_helper_regression_fp64 ✅	v0.7.5
lax.scatter_min	ScatterND	scatter_min_simple_1d ✅ scatter_min_simple_1d_f64 ✅ scatter_min_window_2d_operand_1d_indices ✅ scatter_min_window_2d_operand_1d_indices_f64 ✅ scatter_min_batch_updates_1d_operand ✅ scatter_min_batch_updates_1d_operand_f64 ✅ scatter_min_fp64_dtype_path_check ✅ scatter_min_depth2_helper_regression_fp64 ✅	v0.7.5
lax.scatter_mul	ScatterND	scatter_mul_simple_1d ✅ scatter_mul_simple_1d_f64 ✅ scatter_mul_window_2d_operand_1d_indices ✅ scatter_mul_window_2d_operand_1d_indices_f64 ✅ scatter_mul_batch_updates_1d_operand ✅ scatter_mul_batch_updates_1d_operand_f64 ✅ scatter_mul_mismatched_window_dims_from_user_report ✅ scatter_mul_mismatched_window_dims_from_user_report2 ✅ scatter_mul_mismatched_window_dims_from_user_report3 ✅ scatter_mul_fluids_pattern_updates_5_4_1_1 ✅ scatter_mul_in_cond_float64 ✅	v0.6.4
lax.select	Where	select_simple ✅ select_simple_f64 ✅ select_mask_scores_tensor_else_dynamic ✅ select_mask_scores_tensor_else_dynamic_f64 ✅ select_mask_scores_tensor_else ✅ select_mask_scores_tensor_else_f64 ✅	v0.7.1
lax.select_n	Where	select_n_bool_predicate_two_cases_float ✅ select_n_bool_predicate_two_cases_float_f64 ✅ select_n_bool_predicate_two_cases_int ✅ select_n_bool_predicate_two_cases_int_f64 ✅ select_n_bool_predicate_scalar_broadcast ✅ select_n_bool_predicate_scalar_broadcast_f64 ✅ select_n_int_indices_three_cases ✅ select_n_int_indices_three_cases_f64 ✅ select_n_int_indices_four_cases ✅ select_n_int_indices_four_cases_f64 ✅	v0.2.0
lax.sign	Sign	sign ✅ sign_f64 ✅	v0.5.0
lax.sin	Sin	sin ✅ sin_f64 ✅	v0.4.4
lax.sinh	Sinh	sinh ✅ sinh_f64 ✅	v0.4.4
lax.slice	Slice	slice_test1 ✅ slice_test1_f64 ✅ slice_3d_none_strides ✅ slice_3d_none_strides_f64 ✅ slice_scan_axis_drop ✅ slice_scan_axis_drop_f64 ✅	v0.1.0
lax.sort	TopK	sort_1d ✅ sort_1d_f64 ✅ sort_2d ✅ sort_2d_f64 ✅	v0.2.0
lax.split	Split	lax_split_equal_parts ✅ lax_split_equal_parts_f64 ✅ lax_split_unequal_parts ✅ lax_split_unequal_parts_f64 ✅	v0.7.2
lax.sqrt	Sqrt	sqrt ✅ sqrt_f64 ✅	v0.2.0
lax.square	Mul	square ✅ square_f64 ✅	v0.2.0
lax.squeeze	Squeeze	lax_squeeze_specific_axis_0 ✅ lax_squeeze_specific_axis_0_f64 ✅ lax_squeeze_multiple_axes ✅ lax_squeeze_multiple_axes_f64 ✅ lax_squeeze_no_op_empty_dims ✅ lax_squeeze_no_op_empty_dims_f64 ✅ lax_squeeze_problem_case_input_squeeze_only_axis_0 ✅ lax_squeeze_problem_case_input_squeeze_only_axis_0_f64 ✅ lax_squeeze_problem_case_input_squeeze_axes_0_2 ✅ lax_squeeze_problem_case_input_squeeze_axes_0_2_f64 ✅ lax_squeeze_problem_case_input_squeeze_all_dims_explicitly ✅ lax_squeeze_problem_case_input_squeeze_all_dims_explicitly_f64 ✅	v0.2.0
lax.stop_gradient	Identity	stop_gradient ✅ stop_gradient_f64 ✅	v0.2.0
lax.sub	Sub	sub_test1 ✅ sub_test1_f64 ✅ sub_test2 ✅ sub_test2_f64 ✅	v0.1.0
lax.tanh	Tanh	tanh ✅ tanh_f64 ✅	v0.2.0
lax.transpose	Transpose	transpose_basic ✅ transpose_basic_f64 ✅	v0.2.0
lax.while_loop	Loop	while_loop_counter ✅ while_loop_counter_f64 ✅ while_loop_vector ✅ while_loop_vector_f64 ✅ while_loop_f64 ✅ while_loop_multi_state_f32 ✅ while_loop_multi_state_f64 ✅ while_loop_with_closure ✅ while_loop_with_closure_f64 ✅ while_loop_basic ✅ while_loop_two_state ✅ while_loop_captured_tracer ✅ while_loop_with_scalar_state ✅ while_loop_renamed_passthrough ✅ while_loop_closure_topo ✅ while_loop_mixed_rank ✅ while_loop_tracer_passthrough ✅ while_loop_no_loop_output_reused_as_input ✅ while_loop_4d_and_scalar_state ✅ while_loop_4d_and_scalar_state_f64 ✅ while_loop_cnn_scalar_state_bug ✅ while_loop_cnn_scalar_state_bug_f64 ✅ while_loop_nnx_repro ✅ while_loop_nnx_repro_f64 ✅	v0.5.1
nn.celu	Celu	jaxnn_celu ✅ jaxnn_celu_f64 ✅ jaxnn_celu_1 ✅ jaxnn_celu_1_f64 ✅	v0.7.1
nn.dot_product_attention	Add Cast MatMul Mul Not Softmax Transpose Where	dpa_basic ✅ dpa_basic_f64 ✅ dpa_positional_bias_mask ✅ dpa_positional_bias_mask_f64 ✅ dpa_diff_heads_embed ✅ dpa_diff_heads_embed_f64 ✅ dpa_batch4_seq16 ✅ dpa_batch4_seq16_f64 ✅ dpa_float64 ✅ dpa_float64_f64 ✅ dpa_heads1_embed4 ✅ dpa_heads1_embed4_f64 ✅ dpa_heads8_embed8 ✅ dpa_heads8_embed8_f64 ✅ dpa_batch1_seq2 ✅ dpa_batch1_seq2_f64 ✅ dpa_batch8_seq4 ✅ dpa_batch8_seq4_f64 ✅ dpa_axis1 ✅ dpa_axis1_f64 ✅ dpa_with_tensor_mask ✅ dpa_with_tensor_mask_f64 ✅ dpa_tiny_mask_all_valid ✅ dpa_tiny_mask_all_valid_f64 ✅ dpa_tiny_mask_mixed ✅ dpa_tiny_mask_mixed_f64 ✅ dpa_one_false ✅ dpa_one_false_f64 ✅ dpa_mostly_false ✅ dpa_mostly_false_f64 ✅ dpa_with_causal_mask ✅ dpa_with_causal_mask_f64 ✅ dpa_with_padding_mask ✅ dpa_with_padding_mask_f64 ✅ dpa_with_local_window_mask ✅ dpa_with_local_window_mask_f64 ✅ dpa_mask_none ✅	v0.1.0
nn.elu	Elu	jaxnn_elu ✅ jaxnn_elu_f64 ✅ jaxnn_elu_1 ✅ jaxnn_elu_1_f64 ✅	v0.7.1
nn.gelu	Gelu	jaxnn_gelu ✅ jaxnn_gelu_f64 ✅ jaxnn_gelu_1 ✅ jaxnn_gelu_1_f64 ✅ jaxnn_gelu_approx ✅ jaxnn_gelu_approx_f64 ✅	v0.7.1
nn.identity	Identity	jaxnn_identity ✅ jaxnn_identity_f64 ✅ jaxnn_identity_1 ✅ jaxnn_identity_1_f64 ✅	v0.7.1
nn.leaky_relu	LeakyRelu	jaxnn_leaky_relu ✅ jaxnn_leaky_relu_f64 ✅ jaxnn_leaky_relu_1 ✅ jaxnn_leaky_relu_1_f64 ✅	v0.7.1
nn.mish	Mish	jaxnn_mish ✅ jaxnn_mish_f64 ✅ jaxnn_mish_1 ✅ jaxnn_mish_1_f64 ✅	v0.7.1
nn.relu	Relu	jaxnn_relu ✅ jaxnn_relu_f64 ✅ jaxnn_relu_1 ✅ jaxnn_relu_1_f64 ✅	v0.7.1
nn.selu	Selu	jaxnn_selu ✅ jaxnn_selu_f64 ✅ jaxnn_selu_1 ✅ jaxnn_selu_1_f64 ✅	v0.7.1
nn.sigmoid	Sigmoid	jaxnn_sigmoid ✅ jaxnn_sigmoid_f64 ✅ jaxnn_sigmoid_1 ✅ jaxnn_sigmoid_1_f64 ✅	v0.7.1
nn.soft_sign	Softsign	jaxnn_soft_sign ✅ jaxnn_soft_sign_f64 ✅ jaxnn_soft_sign_1 ✅ jaxnn_soft_sign_1_f64 ✅	v0.7.1
nn.softmax	Softmax	softmax ✅ softmax_f64 ✅ softmax_2d ✅ softmax_2d_f64 ✅ softmax_3d ✅ softmax_3d_f64 ✅	v0.1.0
nn.softplus	Softplus	jaxnn_softplus ✅ jaxnn_softplus_f64 ✅ jaxnn_softplus_1 ✅ jaxnn_softplus_1_f64 ✅	v0.7.1
nn.truncated_normal	➖	initializer ✅ random_truncated_normal_positional ✅ flax_dense_like_init ✅	v0.7.1
nnx.avg_pool	AveragePool Transpose	avg_pool_dynamic ✅ avg_pool ✅ avg_pool_same_padding_dynamic ✅ avg_pool_same_padding ✅ avg_pool_default_padding_dynamic ✅ avg_pool_default_padding ✅ avg_pool_stride1_dynamic ✅ avg_pool_stride1 ✅ avg_pool_win3x3_stride2_dynamic ✅ avg_pool_win3x3_stride2 ✅ avg_pool_stride_none_dynamic ✅ avg_pool_stride_none ✅ avg_pool_count_include_pad_false_dynamic ✅ avg_pool_count_include_pad_false ✅	v0.1.0
nnx.batch_norm	BatchNormalization	batch_norm_no_bias_no_scale_dynamic ✅ batch_norm_no_bias_no_scale ✅ batch_norm_bias_no_scale_dynamic ✅ batch_norm_bias_no_scale ✅ batch_norm_no_bias_scale_dynamic ✅ batch_norm_no_bias_scale ✅ batch_norm_bias_scale_dynamic ✅ batch_norm_bias_scale ✅ batch_norm_3d_dynamic ✅ batch_norm_3d ✅ batch_norm_4d_dynamic ✅ batch_norm_4d ✅ batch_norm_4d_no_bias_no_scale_dynamic ✅ batch_norm_4d_no_bias_no_scale ✅	v0.1.0
nnx.conv	Conv Transpose	conv_basic_bias_dynamic ✅ conv_basic_bias ✅ conv_basic_bias_2 ✅ conv_basic_bias_3 ✅ conv_stride2_bias ✅ conv_no_bias_dynamic ✅ conv_no_bias ✅ conv_valid_padding ✅ conv_stride1 ✅ conv_stride2 ✅ conv_different_kernel ✅ conv_float64 ✅ conv_single_batch ✅ conv_large_batch ✅ conv_1d ✅ conv_1d_more_1d_inputs ✅ conv_1d_more_2d_inputs ✅ conv_1d_large_kernel ✅ conv_1d_dilation ✅ conv_1d_stride_dilation ✅ conv_2d_asymmetric_kernel ✅ conv_2d_asymmetric_stride ✅ conv_2d_asymmetric_dilation ✅ conv_2d_large_dilation ✅ conv_2d_large_stride ✅ conv_2d_mixed_params ✅ conv_3d_basic ✅ conv_3d_stride ✅ conv_3d_asymmetric ✅ conv_3d_dilation ✅ conv_2d_small_input ✅ conv_2d_many_channels ✅ conv_1d_wide_input ✅ conv_2d_kernel_1x1 ✅ conv_1d_kernel_1 ✅ conv_2d_group_conv ✅ conv_1d_group_conv_more_dims ✅ conv_2d_depthwise ✅ conv_1d_complex_on_4d ✅ conv_2d_complex_on_5d ✅ conv_1d_high_dilation_on_3d ✅ conv_2d_group_stride_dilation ✅ conv_1d_group_on_higher_dim ✅ conv_1d_same_padding_on_3d ✅ conv_2d_same_padding_mixed_dilation ✅ conv_1d_large_kernel_on_4d ✅ conv_2d_asymmetric_on_5d ✅ conv_3d_group_complex ✅ conv_1d_unit_group_on_multi_dim ✅	v0.1.0
nnx.dot_product_attention	Cast Div Einsum Gather Shape Softmax Sqrt	dpa_basic ✅ dpa_basic_f64 ✅ dpa_with_tensor_mask ✅ dpa_with_bias ✅ dpa_with_causal_mask ✅ dpa_with_causal_mask_f64 ✅ dpa_with_mask_and_bias ✅	v0.1.0
nnx.dropout	Dropout	dropout_init_params_dynamic ✅ dropout_init_params_dynamic_f64 ✅ dropout_init_params ✅ dropout_init_params_f64 ✅ dropout_call_params_dynamic ✅ dropout_call_params_dynamic_f64 ✅ dropout_call_params ✅ dropout_call_params_f64 ✅	v0.1.0
nnx.einsum	Add Einsum	einsum_module_with_bias ✅ einsum_module_with_bias_f64 ✅ einsum_module_no_bias ✅ einsum_module_no_bias_f64 ✅	v0.4.2
nnx.elu	Elu	elu ✅	v0.1.0
nnx.embed	Gather	token_embedding_dynamic ✅ token_embedding_dynamic_f64 ✅ token_embedding ✅ token_embedding_f64 ✅ positional_embedding_dynamic ✅ positional_embedding_dynamic_f64 ✅ positional_embedding ✅ positional_embedding_f64 ✅	v0.7.0
nnx.gelu	Gelu	gelu ✅ gelu_1 ✅ gelu_2 ✅ gelu_2_f64 ✅ gelu_3_dynamic ✅ gelu_3_dynamic_f64 ✅ gelu_3 ✅ gelu_3_f64 ✅	v0.1.0
nnx.group_norm	GroupNormalization	group_norm ✅ group_norm_no_bias_no_scale_dynamic ✅ group_norm_no_bias_no_scale ✅ group_norm_bias_no_scale_dynamic ✅ group_norm_bias_no_scale ✅ group_norm_no_bias_scale_dynamic ✅ group_norm_no_bias_scale ✅ group_norm_bias_scale_dynamic ✅ group_norm_bias_scale ✅	v0.3.0
nnx.layer_norm	LayerNormalization	layer_norm_dynamic ✅ layer_norm ✅ layer_norm_no_bias_no_scale_dynamic ✅ layer_norm_no_bias_no_scale ✅ layer_norm_bias_no_scale_dynamic ✅ layer_norm_bias_no_scale ✅ layer_norm_no_bias_scale_dynamic ✅ layer_norm_no_bias_scale ✅ layer_norm_bias_scale_dynamic ✅ layer_norm_bias_scale ✅ layer_norm_multiaxis_dynamic ✅ layer_norm_multiaxis ✅ layer_norm_symbolic_batch_dynamic ✅ layer_norm_symbolic_batch ✅ layer_norm_negative_axis_no_div_dynamic ✅ layer_norm_negative_axis_no_div ✅	v0.1.0
nnx.leaky_relu	LeakyRelu	leaky_relu ✅	v0.1.0
nnx.linear	Gemm Reshape	linear_symbolic_batch_dynamic ✅ linear_symbolic_batch_dynamic_f64 ✅ linear_symbolic_batch ✅ linear_symbolic_batch_f64 ✅ linear_high_rank_dynamic ✅ linear_high_rank_dynamic_f64 ✅ linear_high_rank ✅ linear_high_rank_f64 ✅ linear_no_bias_dynamic ✅ linear_no_bias_dynamic_f64 ✅ linear_no_bias ✅ linear_no_bias_f64 ✅ linear_high_rank_no_bias_dynamic ✅ linear_high_rank_no_bias_dynamic_f64 ✅ linear_high_rank_no_bias ✅ linear_high_rank_no_bias_f64 ✅ linear_merge_symbolic_dim_dynamic ✅	v0.1.0
nnx.linear_general	Gemm Reshape	linear_general_merge_symbolic_dim_dynamic ✅ linear_general_dynamic ✅ linear_general ✅ linear_general_2 ✅ linear_general_3 ✅ linear_general_4 ✅ linear_general_abstract_eval_axes ✅ linear_general_abstract_eval_axes_pair ✅ dynamic_batch_and_feature_dims_dynamic ✅	v0.1.0
nnx.log_softmax	LogSoftmax	log_softmax ✅ log_softmax_f64 ✅	v0.1.0
nnx.max_pool	MaxPool Transpose	max_pool ✅ max_pool_same_padding ✅	v0.1.0
nnx.relu	Relu	relu_1d ✅ relu_1d_f64 ✅ relu_4d_dynamic ✅ relu_4d_dynamic_f64 ✅ relu_4d ✅ relu_4d_f64 ✅	v0.1.0
nnx.rms_norm	RMSNormalization	rms_norm_basic ✅ rms_norm_use_scale_false ✅ rms_norm_4d_dynamic_dynamic ✅ rms_norm_4d_dynamic ✅ rms_norm_4d_dynamic_no_scale_dynamic ✅ rms_norm_4d_dynamic_no_scale ✅	v0.3.0
nnx.sigmoid	Sigmoid	sigmoid ✅ sigmoid_f64 ✅	v0.1.0
nnx.softmax	Softmax	softmax_dynamic ✅ softmax_dynamic_f64 ✅ softmax ✅ softmax_f64 ✅	v0.1.0
nnx.softplus	Softplus	softplus ✅	v0.1.0
nnx.tanh	Tanh	tanh ✅ tanh_f64 ✅	v0.1.0

Legend:
✅ = Passed
❌ = Failed
➖ = No testcase yet

---

🎯 Examples

Component	Description	Testcases	Since
MlpExample	A simple MLP example using Equinox.	mlp_training_mode ✅ mlp_training_mode_f64 ✅ mlp_inference_mode ✅ mlp_inference_mode_f64 ✅ mlp_batched_training_mode_dynamic ✅ mlp_batched_training_mode_dynamic_f64 ✅ mlp_batched_training_mode ✅ mlp_batched_training_mode_f64 ✅	v0.8.0
SimpleLinearExample	A simple linear layer example using Equinox.	simple_linear_dynamic ✅ simple_linear_dynamic_f64 ✅ simple_linear ✅ simple_linear_f64 ✅ nn_linear_dynamic ✅ nn_linear_dynamic_f64 ✅ nn_linear ✅ nn_linear_f64 ✅	v0.7.1
GPT	A simple GPT model that reuses nnx.MultiHeadAttention.	gpt_dynamic ✅ gpt ✅	v0.7.0
GPT_Attention	A multi-head attention layer.	gpt_attention ✅	v0.7.1
GPT_CausalSelfAttention	A causal self-attention module.	causal_self_attention_dynamic ✅ causal_self_attention ✅	v0.7.0
GPT_Embeddings	Combines token and position embeddings with dropout.	gpt_embeddings_dynamic ✅ gpt_embeddings ✅	v0.7.0
GPT_Head	The head of the GPT model.	gpt_head_dynamic ✅ gpt_head ✅	v0.7.0
GPT_MLP	An MLP block with GELU activation from nanoGPT.	gpt_mlp_dynamic ✅ gpt_mlp ✅	v0.7.0
GPT_PositionEmbedding	A positional embedding layer using nnx.Embed.	position_embedding ✅	v0.7.0
GPT_TokenEmbedding	A token embedding layer using nnx.Embed.	token_embedding_dynamic ✅ token_embedding ✅	v0.7.0
GPT_TransformerBlock	A transformer block combining attention and MLP.	gpt_block_dynamic ✅ gpt_block ✅	v0.7.0
GPT_TransformerStack	A stack of transformer blocks.	transformer_stack_dynamic ✅ transformer_stack ✅	v0.7.0
broadcast_add	Simple dynamic broadcast + add	broadcast_add_dynamic_dynamic ✅ broadcast_add_dynamic_dynamic_f64 ✅ broadcast_add_dynamic ✅ broadcast_add_dynamic_f64 ✅	v0.7.0
cfl_timestep	Tests the CFL condition timestep calculation.	cfl_timestep_f64 ✅	v0.6.5
weno_reconstruction	Tests the complex arithmetic pattern found in WENO schemes.	weno_reconstruction_f64 ✅	v0.6.5
fori_loop_test	fori_loop_test: Demonstrates jax.lax.fori_loop with a simple loop.	fori_loop_test ✅ fori_loop_test_f64 ➖	v0.6.3
issue18_abs	Test jnp.abs from issue 18	abs_fn ✅ abs_fn_f64 ✅	v0.6.3
issue18_arange	Test arange from issue 18	arange_fn ✅ arange_fn_f64 ✅	v0.6.3
issue18_fori_loop	Test fori_loop from issue 18	fori_loop_fn ✅ fori_loop_fn_f64 ✅	v0.6.3
issue18_linspace	Test linspace from issue 18	linspace_fn ✅ linspace_fn_f64 ✅	v0.6.3
issue18_scan	Test scan from issue 18 (no xs)	scan_fn ✅ scan_fn_f64 ✅	v0.6.3
issue18_sign	Test jnp.sign from issue 18	sign_fn ✅ sign_fn_f64 ✅	v0.6.3
issue18_where	Test where from issue 18	where_fn ✅ where_fn_f64 ✅	v0.6.3
issue18_while_loop	Test while_loop from issue 18	while_loop_fn ✅	v0.6.3
select_test	select_test: Demonstrates jnp.select with a dynamic condition based on an input array.	select_test_all_options ➖ select_test_scalar_select_option_0 ➖ select_test_scalar_select_option_1 ➖ select_test_scalar_select_option_2 ➖ select_test_default_case ➖	v0.6.1
sort_test	sort_test: Demonstrates jnp.sort on slices of an input array.	sort_test_basic ➖	v0.6.1
cond_scatter_add_mul	Tests scatter_add/mul inside jnp.where branches	cond_scatter_add_mul_f64_a ✅ cond_scatter_add_mul_f64_b ✅	v0.6.4
cond_scatter_repro	Reproduces a bug where lax.cond subgraphs do not inherit parent initializers.	cond_scatter_repro_f64 ✅	v0.6.4
remat2	Tests a simple case of jax.checkpoint (also known as jax.remat2).	checkpoint_scalar_f32 ✅ checkpoint_scalar_f32_f64 ✅	v0.6.5
scatter_window	Window-scatter (H×W patch) with implicit batch (depth-3 path). Exercises GatherScatterMode.FILL_OR_DROP and double precision. Regression of a prior conversion failure.	scatter_window_update_f64_example ✅	v0.7.4
AutoEncoder	A simple autoencoder example.	simple_autoencoder ✅ simple_autoencoder_f64 ✅	v0.2.0
CNN	A simple convolutional neural network (CNN).	simple_cnn_explicit_dimensions ✅ simple_cnn_dynamic ✅ simple_cnn ✅	v0.1.0
ForiLoop	fori_loop example	fori_loop_counter ✅ fori_loop_counter_f64 ✅	v0.5.1
GRUCell	Vanilla gated-recurrent-unit cell from Flax/nnx. There is no 1-to-1 ONNX operator, so the converter decomposes it into MatMul, Add, Sigmoid, Tanh, etc.	gru_cell_basic ✅	v0.7.2
MLP	A simple Multi-Layer Perceptron (MLP) with BatchNorm, Dropout, and GELU activation.	simple_mlp_dynamic ✅ simple_mlp_dynamic_f64 ✅ simple_mlp ✅ simple_mlp_f64 ✅ simple_mlp_with_call_params_dynamic ✅ simple_mlp_with_call_params_dynamic_f64 ✅ simple_mlp_with_call_params ✅ simple_mlp_with_call_params_f64 ✅	v0.1.0
MultiHeadAttention	This is a multi-head attention module implemented by Flax/nnx that has no ONNX correspondent on the same granularity.	multihead_attention_nn_dynamic ✅ multihead_attention_nn ✅ multihead_attention_nnx_dynamic ✅ multihead_attention_nnx ✅ multihead_attention_2_nnx_dynamic ✅ multihead_attention_2_nnx ✅	v0.2.0
SequentialReLU	Two ReLU activations chained with nnx.Sequential (no parameters).	sequential_double_relu ✅ sequential_double_relu_f64 ✅	v0.7.1
SequentialWithResidual	Tests nnx.Sequential nested inside a module with a residual connection.	sequential_nested_with_residual ✅	v0.7.1
TransformerDecoderWithSequential	A single-layer Transformer decoder built with nnx primitives (MHA, LayerNorm, Feed-Forward, Dropout).	tiny_decoder_with_sequential_dynamic ✅ tiny_decoder_with_sequential ✅ tiny_decoder_with_sequential_and_full_dynamic_shapes_dynamic ✅	v0.7.1
TransformerDecoderWithoutSequential	A single-layer Transformer decoder built with nnx primitives (MHA, LayerNorm, Feed-Forward, Dropout).	tiny_decoder_without_sequential_dynamic ✅ tiny_decoder_without_sequential ✅	v0.7.1
onnx_functions_000	one function on an outer layer.	000_one_function_on_outer_layer_dynamic ✅ 000_one_function_on_outer_layer ✅	v0.4.0
onnx_functions_001	one function on an inner layer.	001_one_function_inner_dynamic ✅ 001_one_function_inner ✅	v0.4.0
onnx_functions_002	two nested functions.	002_two_nested_functions_dynamic ✅ 002_two_nested_functions ✅	v0.4.0
onnx_functions_003	two nested functions.	003_two_simple_nested_functions_dynamic ✅ 003_two_simple_nested_functions ✅	v0.4.0
onnx_functions_004	nested function plus component	004_nested_function_plus_component_dynamic ✅ 004_nested_function_plus_component ✅	v0.4.0
onnx_functions_005	nested function plus more components	005_nested_function_plus_component_dynamic ✅ 005_nested_function_plus_component ✅	v0.4.0
onnx_functions_006	one function on an outer layer.	006_one_function_outer_dynamic ✅ 006_one_function_outer ✅	v0.4.0
onnx_functions_007	transformer block with nested mlp block with call parameter	007_transformer_block_dynamic ✅ 007_transformer_block ✅	v0.4.0
onnx_functions_008	transformer block with nested mlp block no call parameter	008_transformer_block_dynamic ✅ 008_transformer_block ✅	v0.4.0
onnx_functions_009	transformer block using decorator on class and function	009_transformer_block_dynamic ✅ 009_transformer_block ✅	v0.4.0
onnx_functions_010	transformer stack	010_transformer_stack_dynamic ✅ 010_transformer_stack ✅	v0.4.0
onnx_functions_012	Vision Transformer (ViT)	012_vit_conv_embedding_dynamic ✅ 012_vit_conv_embedding ✅	v0.4.0
onnx_functions_013	Vision Transformer (ViT)	013_vit_conv_embedding_with_call_params_dynamic ✅ 013_vit_conv_embedding_with_call_params ✅ 013_vit_conv_embedding_with_internal_call_params_dynamic ✅ 013_vit_conv_embedding_with_internal_call_params ✅	v0.4.0
onnx_functions_014	one function on an outer layer.	014_one_function_with_input_param_with_default_value ✅ 014_one_function_without_input_param_with_default_value_dynamic ✅ 014_one_function_without_input_param_with_default_value ✅	v0.4.0
onnx_functions_015	one function on an outer layer.	015_one_function_with_input_param_without_default_value_dynamic ✅ 015_one_function_with_input_param_without_default_value ✅	v0.4.0
onnx_functions_016	nested function plus more components	016_internal_function_with_input_param_with_default_value_dynamic ✅ 016_internal_function_with_input_param_with_default_value ✅	v0.4.0
ClassificationHead	Classification head for Vision Transformer	classification_head_dynamic ✅ classification_head ✅	v0.4.0
ClassificationHeadFlatten	Classification head for Vision Transformer	classification_head_flat_dynamic ✅ classification_head_flat ✅	v0.4.0
ConcatClsToken	Concatenate CLS token to the input embedding	concat_cls_token_dynamic ✅ concat_cls_token ✅	v0.4.0
ConcatClsTokenFlatten	Concatenate CLS token to the input embedding	concat_cls_token_flat_dynamic ✅ concat_cls_token_flat ✅	v0.4.0
ConvEmbedding	Convolutional Token Embedding for MNIST with hierarchical downsampling.	mnist_conv_embedding_dynamic ✅ mnist_conv_embedding ✅	v0.1.0
ConvEmbeddingFlatten	Convolutional Token Embedding for MNIST with hierarchical downsampling.	mnist_conv_embedding_flat_dynamic ✅ mnist_conv_embedding_flat ✅	v0.1.0
FeedForward	MLP in Transformer	feed_forward_dynamic ✅ feed_forward ✅	v0.1.0
FeedForwardFlatten	MLP in Transformer	feed_forward_flat_dynamic ✅ feed_forward_flat ✅	v0.1.0
GetToken	Get the CLS token from the input embedding	get_token_dynamic ✅ get_token ✅	v0.4.0
GetTokenFlatten	Get the CLS token from the input embedding	get_token_flat_dynamic ✅ get_token_flat ✅	v0.4.0
PatchEmbedding	Cutting the image into patches and linearly embedding them.	patch_embedding_dynamic ✅ patch_embedding ✅	v0.1.0
PatchEmbeddingFlatten	Cutting the image into patches and linearly embedding them.	patch_embedding_flat_dynamic ✅ patch_embedding_flat ✅	v0.1.0
PositionalEmbedding	Add positional embedding to the input embedding	positional_embedding_dynamic ✅ positional_embedding ✅	v0.4.0
PositionalEmbeddingFlatten	Add positional embedding to the input embedding	positional_embedding_flat_dynamic ✅ positional_embedding_flat ✅	v0.4.0
TransformerBlock	Transformer from 'Attention Is All You Need.'	transformer_block_dynamic ✅ transformer_block ✅	v0.1.0
TransformerBlockFlatten	Transformer from 'Attention Is All You Need.'	transformer_block_flat_dynamic ✅ transformer_block_flat ✅	v0.1.0
TransformerStack	Stack of Transformer blocks	transformer_stack_dynamic ✅ transformer_stack ✅	v0.1.0
TransformerStackFlatten	Stack of Transformer blocks	transformer_stack_flat_dynamic ✅ transformer_stack_flat ✅	v0.1.0
VisionTransformer	A Vision Transformer (ViT) model for MNIST with configurable embedding type.	vit_conv_embedding_dynamic ✅ vit_conv_embedding ✅ vit_patch_embedding ✅	v0.2.0
VisionTransformerFlatten	A Vision Transformer (ViT) model for MNIST with configurable embedding type.	vit_conv_embedding_flat_dynamic ✅ vit_conv_embedding_flat ✅ vit_patch_embedding_flat_dynamic ✅ vit_patch_embedding_flat ✅	v0.2.0

---

📌 Dependencies

Versions of Major Dependencies:

Library	Versions
JAX	0.6.2
Flax	0.11.1
onnx	1.18.0
onnxruntime	1.22.1

Note: For more details, check pyproject.toml.

---

⚠️ Limitations

• Currently not all JAX/Flax components are supported (you can easily help expand this coverage!).
• Function references need dynamic resolution at call-time.
• ONNX graph composition is done in-memory before saving to disk, potentially causing memory issues with very large models.

---

🤝 How to Contribute

We warmly welcome contributions!

How you can help:

• Add a plugin: Extend jax2onnx by writing a simple Python file in jax2onnx/plugins. a custom primitive or an example.
• Bug fixes & improvements: PRs and issues are always welcome.

---

💾 Installation

Install from PyPI:

pip install jax2onnx  

---

📜 License

This project is licensed under the Apache License, Version 2.0. See LICENSE for details.

---

🌟 Special Thanks

Special thanks for example contributions to @burakssen, @Cadynum, @clementpoiret and @PVirie

Special thanks for plugin contributions to @burakssen, @clementpoiret and @Clouder0

Special thanks to tumaer/JAXFLUIDS for contributing valuable insights rooted in physics simulation use cases.

Special thanks to @lutzroeder for making shapes internal to ONNX function visible in his great Netron viewer.

• ONNX: Function value_info support #1447

Special thanks to the community members involved in:

• Flax Feature Request #4430
• JAX Feature Request #26430

A huge thanks especially to @limarta, whose elegant jaxpr-to-ONNX demonstration significantly inspired this project.

---

Happy converting! 🎉