[Relax][Frontend][TFLite] Add REDUCE_WINDOW support

[THINKER-ONLY]

Summary

Add Relax TFLite frontend support for the builtin REDUCE_WINDOW operator.
This covers the ordinary TFLite op only, not STABLEHLO_REDUCE_WINDOW.

The converter parses ReduceWindowOptions from BuiltinOptions2, validates
the static window attributes, and lowers supported reduce functions through
topi.sliding_window plus Relax reductions.

Supported modes:

• ADD
• MUL
• MINIMUM
• MAXIMUM
• ALL
• ANY

Empty output shapes are handled directly with relax.op.zeros. Quantized
REDUCE_WINDOW, dynamic window attributes, and unsupported reduce functions
remain rejected with explicit errors.

Testing

• python -m py_compile python/tvm/relax/frontend/tflite/tflite_frontend.py tests/python/relax/test_frontend_tflite.py
• python -m pytest tests/python/relax/test_frontend_tflite.py -k reduce_window -q -p no:tvm.testing.plugin
• python -m pytest tests/python/relax/test_frontend_tflite.py -k "reduce_window or reduction_ops" -q -p no:tvm.testing.plugin
• conda run -n test python -m ruff check python/tvm/relax/frontend/tflite/tflite_frontend.py tests/python/relax/test_frontend_tflite.py

Related

Related to #19519.

[gemini-code-assist]

Code Review

This pull request adds support for the REDUCE_WINDOW operator in the TFLite Relax frontend, along with comprehensive unit tests covering various reduction modes and error cases. The reviewer identified a critical issue where the call to topi.sliding_window in both the frontend implementation and the test helpers passes an incorrect number of arguments and incorrect argument types, which would lead to runtime failures. Correcting this call to use the proper 5-argument signature with explicit shape expressions for padding is recommended.

[gemini-code-assist]

The topi.sliding_window packed function expects 5 arguments: data, pad_before, pad_after, window_shape, and strides. The current call only passes 4 arguments, omitting pad_after. Additionally, pad_before is passed as 0 (an integer) instead of a shape expression of zeros of length rank (e.g., relax.ShapeExpr([0] * rank)). This will cause a runtime failure when executing the compiled model.

Suggested change

	windowed = relax.op.call_dps_packed(
	"topi.sliding_window",
	(
	data,
	0,
	relax.ShapeExpr(dilated_window_shape),
	relax.ShapeExpr(window_strides),
	),
	out_sinfo=relax.TensorStructInfo(output_shape + dilated_window_shape, input_dtype),
	)
	windowed = relax.op.call_dps_packed(
	"topi.sliding_window",
	(
	data,
	relax.ShapeExpr([0] * rank),
	relax.ShapeExpr([0] * rank),
	relax.ShapeExpr(dilated_window_shape),
	relax.ShapeExpr(window_strides),
	),
	out_sinfo=relax.TensorStructInfo(output_shape + dilated_window_shape, input_dtype),
	)

[gemini-code-assist]

Update the expected topi.sliding_window call in the test helper to match the corrected 5-argument signature with explicit pad_before and pad_after shape expressions.

            windowed = bb.emit(
                relax.op.call_dps_packed(
                    "topi.sliding_window",
                    (
                        x,
                        relax.ShapeExpr([0] * rank),
                        relax.ShapeExpr([0] * rank),
                        relax.ShapeExpr(dilated_window_shape),
                        relax.ShapeExpr(window_strides),
                    ),
                    out_sinfo=relax.TensorStructInfo(
                        output_shape + tuple(dilated_window_shape), dtype
                    ),
                )
            )

[gemini-code-assist]

Update the expected topi.sliding_window call in the boolean test helper to match the corrected 5-argument signature with explicit pad_before and pad_after shape expressions.

Suggested change

	windowed = bb.emit(
	relax.op.call_dps_packed(
	"topi.sliding_window",
	(
	x,
	0,
	relax.ShapeExpr(dilated_window_shape),
	relax.ShapeExpr(window_strides),
	),
	out_sinfo=relax.TensorStructInfo(
	output_shape + tuple(dilated_window_shape), "bool"
	),
	)
	)
	windowed = bb.emit(
	relax.op.call_dps_packed(
	"topi.sliding_window",
	(
	x,
	relax.ShapeExpr([0] * rank),
	relax.ShapeExpr([0] * rank),
	relax.ShapeExpr(dilated_window_shape),
	relax.ShapeExpr(window_strides),
	),
	out_sinfo=relax.TensorStructInfo(
	output_shape + tuple(dilated_window_shape), "bool"
	),
	)
	)

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds support for the TFLite REDUCE_WINDOW operator in the Relax TFLite frontend and introduces corresponding unit tests.

Changes:

• Register and implement REDUCE_WINDOW conversion in the Relax TFLite frontend.
• Add FlatBuffers test-model builders for REDUCE_WINDOW and expected Relax IR constructors.
• Add coverage for supported modes and key failure cases (unsupported function, rank mismatch, invalid stride, shape mismatch, empty output).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
tests/python/relax/test_frontend_tflite.py	Adds schema bindings, model builder helpers, and tests for REDUCE_WINDOW.
python/tvm/relax/frontend/tflite/tflite_frontend.py	Implements REDUCE_WINDOW conversion and wires it into the operator dispatch table.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tlopex]

Thanks for adding this converter. I think this currently rejects valid TFLite REDUCE_WINDOW models because the init value is required to be rank-0 here.

The TFLite kernel tests build REDUCE_WINDOW with the init tensor as a one-element tensor with shape [1], not shape []. For example, the official test helper uses AddConstInput(kTensorType, {init_value_data_}, {1}). The op semantics only need a single init value, so accepting only len(shape) == 0 here is too strict and will reject models that the TFLite runtime accepts.

I think this should check that the init tensor has exactly one element instead of requiring rank 0, and then reshape/squeeze it to a scalar if needed before combining it with the reduced window result.

The current tests miss this because _build_reduce_window_model constructs the init tensor with shape []. Please add coverage for the [1] init tensor form, ideally with a numeric test matching TFLite behavior.

[tlopex]

LGTM