[Relax][Frontend][TFLite] Fix dynamic FILL/SPLIT_V partial implementations by Aharrypotter · Pull Request #19433 · apache/tvm

Aharrypotter · 2026-04-23T14:25:39Z

This PR fixes partial TFLite frontend support for dynamic FILL and SPLIT_V.

Key changes:

Allow FILL to accept runtime dims tensors by converting them with
relax.op.tensor_to_shape before calling relax.op.full.
Allow SPLIT_V to accept runtime size_splits tensors by decomposing the
op into cumsum and dynamic_strided_slice.
Use the TFLite output tuple arity as the source of truth for dynamic
SPLIT_V, instead of relying on size_splits static shape information.
Add TFLite frontend regression tests covering dynamic FILL import and
dynamic SPLIT_V import/compile behavior.

This addresses the FILL and SPLIT_V items under the "Fix partial
implementations" section of #19412.

Validation:

python -m ruff check 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 "split_v_dynamic or fill_dynamic_dims" -v

Result:

All checks passed

gemini-code-assist

Code Review

This pull request enhances the TFLite frontend by enabling support for dynamic dimensions in the Fill operator and dynamic split sizes in the SPLIT_V operator. The implementation utilizes tensor_to_shape for dynamic Fill operations and decomposes dynamic SPLIT_V operations into a series of dynamic strided slices. Reviewer feedback recommends renaming the variables num_splits and rank to num_outputs and input_rank, respectively, to enhance code clarity.

gemini-code-assist · 2026-04-23T14:28:22Z

+            padded_cumsum = relax.op.concat([zero, cumsum], axis=0)
+            # TFLite fixes the tuple arity in the graph, even when the split
+            # sizes themselves are supplied at runtime.
+            num_splits = len(output_tensors)


The variable num_splits is used but it is better to use len(output_tensors) directly if it is only used once, or rename it to num_outputs to be more descriptive of what it represents in the context of TFLite output tensors.

gemini-code-assist · 2026-04-23T14:28:22Z

+
+            # end_base is the full input shape; only split_axis changes per slice.
+            end_base = relax.op.shape_to_tensor(relax.op.shape_of(in_expr))
+            begin_base = relax.const(np.zeros((rank,), dtype="int64"), "int64")


The variable rank is used to create begin_base and strides. It is clearer to define this as input_rank to avoid ambiguity with other potential rank definitions.

Thanks, these are naming-only suggestions. Since the variables are local and the current intent is already documented by comments, I’d prefer to keep the diff minimal.

tlopex

FILL frontend looks good, but dynamic FILL still breaks at legalize.
The issue is in relax.full legalize (legalize_ops/create.py): it assumes call.args[0] is a static ShapeExpr and passes it directly to topi.full. For dynamic FILL, this can be a tensor_to_shape(...) call, so legalize fails before tvm.compile.

Could you have a fix for _full to handle dynamic shape args just in this PR dynamic_strided_slice in legalize_ops/index.py should be a good template: keep the current static ShapeExpr path, and for dynamic ShapeStructInfo, match-cast to fresh symbolic vars and pass those vars to topi.full.

Then we can update test_fill_dynamic_dims to actually run the verify test

tlopex · 2026-04-23T23:38:16Z

+    assert "R.scatter_elements" in ir
+
+    # Regression guard: compile to catch the dynamic split lowering path.
+    tvm.compile(mod, tvm.target.Target("llvm"))


Is this really needed? I meantvm.complie here

You're right, adding tvm.compile(...) to test_split_v_dynamic was over-testing on my side.

That test is meant to cover the frontend lowering change for dynamic SPLIT_V, namely preserving fixed tuple arity and lowering to the expected Relax ops. Those
properties are already directly captured by the IR assertions, so I removed the extra tvm.compile(...) / verify(...) from that test.

For dynamic FILL, I kept the compile check intentionally. In that case, the imported Relax IR was already shape-aware, while the actual remaining issue was that
dynamic-shape relax.full still failed in the legalize/compile path. So I updated the legalization for dynamic relax.full, and kept tvm.compile(...) in
test_fill_dynamic_dims to cover that specific regression.

…ze_splits in SPLIT_V This PR fixes partial implementations for FILL and SPLIT_V operators. FILL: - Remove OpNotImplemented when dims is non-constant - Use relax.op.tensor_to_shape to convert dynamic dims tensor to shape - Pass the shape expression to relax.op.full SPLIT_V: - Remove OpNotImplemented when size_splits is non-constant - Decompose dynamic SPLIT_V into multiple dynamic_strided_slice ops - Use cumsum to compute split offsets and scatter_elements to build begin/end tensors for each slice Tests: - Add test_fill_dynamic_dims - Add test_split_v_dynamic

…c FILL

Aharrypotter · 2026-04-28T15:32:06Z

Addressed both review points:

added dynamic-shape legalization for relax.full and updated test_fill_dynamic_dims to compile/verify
removed the unnecessary compile/verify path from test_split_v_dynamic

One extra fix was needed in addition to _full: LegalizeOps was filtering out dynamic relax.full before the legalization hook could run, so I also relaxed the op’s shape gate to make the dynamic legalization path reachable. After that, dynamic FILL legalizes and compiles correctly.

Aharrypotter · 2026-04-28T17:11:26Z

The imported Relax IR was not well formed before legalization. I fixed the frontend to bind runtime fill dims to fresh shape vars before calling relax.full

Aharrypotter · 2026-04-28T17:21:25Z

Please take a look @tlopex

… `STRIDED_SLICE/SPLIT_V` tests (#19468) ## Summary This PR continues the TFLite frontend work tracked in #18971 for `STRIDED_SLICE` and `SPLIT_V`. Since the dynamic `FILL` / `SPLIT_V` partial-implementation work has already been handled separately in #19433, this PR focuses on the remaining pieces in this branch: - fixing negative-stride `STRIDED_SLICE` conversion in the TFLite frontend - adding regression coverage for `STRIDED_SLICE` and static `SPLIT_V` Relates to #18971. ## Changes 1. **`STRIDED_SLICE` negative-stride fix** - Update the TFLite frontend `convert_strided_slice` handling of `end_mask` when `stride < 0`. - Use an exclusive lower bound compatible with Relax slicing semantics so reverse slices like `x[::-1]` include index `0` correctly. 2. **TFLite frontend test coverage** - Add `test_strided_slice_stride` to cover non-unit stride handling. - Add `test_strided_slice_negative_stride` to cover reverse slicing with negative strides. - Add `test_split_v_static` to cover static `SPLIT_V` conversion. 3. **Scope clarification** - Keep this PR focused on the remaining `STRIDED_SLICE` / `SPLIT_V` work from #18971. - Exclude the dynamic `FILL` / `SPLIT_V` changes that are already addressed in #19433. ## Testing ```bash python -m pytest -n 1 tests/python/relax/test_frontend_tflite.py::test_split_v_static -q python -m pytest -n 1 tests/python/relax/test_frontend_tflite.py::test_strided_slice_stride -q python -m pytest -n 1 tests/python/relax/test_frontend_tflite.py::test_strided_slice_negative_stride -q ``` ## Result - The added STRIDED_SLICE and SPLIT_V tests passed locally. - The negative-stride STRIDED_SLICE path now matches Relax slicing semantics.

gemini-code-assist Bot reviewed Apr 23, 2026

View reviewed changes

tlopex requested changes Apr 23, 2026

View reviewed changes

Aharrypotter added 2 commits April 28, 2026 22:27

[Relax][Frontend][TFLite] Fix SPLIT_V tuple arity and legalize dynami…

34093e8

…c FILL

Aharrypotter force-pushed the fix-partial-fill-splitv branch from c043efc to 34093e8 Compare April 28, 2026 15:08

[Relax][Frontend][TFLite] Fix dynamic FILL shape binding

0152656

tlopex approved these changes Apr 28, 2026

View reviewed changes

tlopex merged commit ccd81f2 into apache:main Apr 28, 2026
9 checks passed

Aharrypotter mentioned this pull request Apr 29, 2026

[Relax][Frontend][TFLite] Fix STRIDED_SLICE negative stride and add STRIDED_SLICE/SPLIT_V tests #19468

Merged

Aharrypotter mentioned this pull request May 1, 2026

[Relax][Frontend][TFLite] Add RANDOM_UNIFORM, RANDOM_STANDARD_NORMAL, and MULTINOMIAL #19473

Merged

ysh329 mentioned this pull request May 6, 2026

[Release] v0.24.0 Release Candidate Notes #19513

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Relax][Frontend][TFLite] Fix dynamic FILL/SPLIT_V partial implementations#19433

[Relax][Frontend][TFLite] Fix dynamic FILL/SPLIT_V partial implementations#19433
tlopex merged 3 commits into
apache:mainfrom
Aharrypotter:fix-partial-fill-splitv

Aharrypotter commented Apr 23, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Apr 23, 2026

Uh oh!

gemini-code-assist Bot Apr 23, 2026

Uh oh!

Aharrypotter Apr 23, 2026

Uh oh!

tlopex left a comment

Uh oh!

tlopex Apr 23, 2026 •

edited

Loading

Uh oh!

Aharrypotter Apr 28, 2026 •

edited

Loading

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

Aharrypotter commented Apr 23, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Apr 23, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Apr 23, 2026

Choose a reason for hiding this comment

Uh oh!

Aharrypotter Apr 23, 2026

Choose a reason for hiding this comment

Uh oh!

tlopex left a comment

Choose a reason for hiding this comment

Uh oh!

tlopex Apr 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Aharrypotter Apr 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Aharrypotter commented Apr 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

tlopex Apr 23, 2026 •

edited

Loading

Aharrypotter Apr 28, 2026 •

edited

Loading