[Tracking Issue][TFLite] Remaining builtin operator coverage beyond #19412

[tlopex]

The Relax TFLite frontend has made good progress recently:

• [Tracking Issue][TFLite] Expand unit test coverage for supported non-quantized operators #18971 tracks unit test coverage for already-supported non-quantized operators.
• [Tracking Issue][TFlite] Expand Missing / Partial Operator Coverage #19412 tracks a first batch of missing or partial operators that are good-first-issue or medium-sized frontend tasks.

This issue tracks the remaining larger TFLite builtin operator coverage that is not already well-covered by those two issues.

The goal is to make the remaining gaps visible, split them by complexity, and let contributors pick coherent operator families instead of one-off unsupported ops.

Scope

This issue covers remaining TFLite builtin operators that fall into larger semantic groups. This issue does not track quantized model support as a whole; that should be tracked separately.

A. Sequence / recurrent model operators

• RNN
• LSTM
• SVDF
• UNIDIRECTIONAL_SEQUENCE_RNN
• UNIDIRECTIONAL_SEQUENCE_LSTM
• BIDIRECTIONAL_SEQUENCE_RNN
• BIDIRECTIONAL_SEQUENCE_LSTM

B. 3D convolution family

• CONV_3D
• CONV_3D_TRANSPOSE

C. FFT / complex operators

• RFFT2D
• REAL
• IMAG
• COMPLEX_ABS

D. Legacy embedding / sparse embedding operators

• CONCAT_EMBEDDINGS
• EMBEDDING_LOOKUP_SPARSE
• HASHTABLE_LOOKUP
• LSH_PROJECTION
• SKIP_GRAM

E. Window reduction

• REDUCE_WINDOW

F. Control flow / multi-subgraph operators

• IF
• WHILE
• CALL
• CALL_ONCE

G. Resource / variable / hashtable operators

• VAR_HANDLE
• READ_VARIABLE
• ASSIGN_VARIABLE
• HASHTABLE
• HASHTABLE_FIND
• HASHTABLE_IMPORT
• HASHTABLE_SIZE

H. Operator markers

• DELEGATE
• PLACEHOLDER_FOR_GREATER_OP_CODES

I. StableHLO operators in TFLite

• STABLEHLO_ABS
• STABLEHLO_ADD
• STABLEHLO_AND
• STABLEHLO_BROADCAST_IN_DIM
• STABLEHLO_CBRT
• STABLEHLO_CLAMP
• STABLEHLO_COMPARE
• STABLEHLO_COMPOSITE
• STABLEHLO_CONCATENATE
• STABLEHLO_CONVERT
• STABLEHLO_CONVOLUTION
• STABLEHLO_COSINE
• STABLEHLO_CUSTOM_CALL
• STABLEHLO_DIVIDE
• STABLEHLO_DOT_GENERAL
• STABLEHLO_DYNAMIC_SLICE
• STABLEHLO_DYNAMIC_UPDATE_SLICE
• STABLEHLO_EXPONENTIAL
• STABLEHLO_FLOOR
• STABLEHLO_GATHER
• STABLEHLO_IOTA
• STABLEHLO_LOG
• STABLEHLO_LOGISTIC
• STABLEHLO_MAXIMUM
• STABLEHLO_MINIMUM
• STABLEHLO_MULTIPLY
• STABLEHLO_NEGATE
• STABLEHLO_OR
• STABLEHLO_PAD
• STABLEHLO_POWER
• STABLEHLO_REDUCE
• STABLEHLO_REDUCE_WINDOW
• STABLEHLO_REMAINDER
• STABLEHLO_RESHAPE
• STABLEHLO_RNG_BIT_GENERATOR
• STABLEHLO_RSQRT
• STABLEHLO_SCATTER
• STABLEHLO_SELECT
• STABLEHLO_SHIFT_LEFT
• STABLEHLO_SLICE
• STABLEHLO_SORT
• STABLEHLO_SUBTRACT
• STABLEHLO_TANH
• STABLEHLO_TRANSPOSE
• STABLEHLO_WHILE

---

Suggested contribution workflow

1. Claim an item by commenting on this issue before starting, so we avoid duplicated work.
2. Read a recent reference PR for handler + test style.
3. Add the handler, register in convert_map (alphabetical within its region).
4. Add a structural-equal test following the verify(TestClass, expected=None) pattern (test_frontend_tflite.py:56). Nightly-gated E2E activates automatically via CI_ENV_NIGHTLY.
5. Tag the PR title [Relax][Frontend][TFLite].

References

• [Tracking Issue][TFLite] Expand unit test coverage for supported non-quantized operators #18971: supported non-quantized operator test coverage
• [Tracking Issue][TFlite] Expand Missing / Partial Operator Coverage #19412: first batch of missing / partial TFLite operator coverage
• Relax TFLite frontend: python/tvm/relax/frontend/tflite/tflite_frontend.py
• TFLite frontend tests: tests/python/relax/test_frontend_tflite.py

cc @leandron

[weicheng-hsu]

Hi @tlopex, I’d like to pick up the TFLite 3D convolution family in the Relax TFLite frontend (tracking issue #19519).

I’ll follow the existing frontend style by adding the handler(s), registering them in convert_map, and adding structural-equal tests in tests/python/relax/test_frontend_tflite.py.

To keep the PRs reviewable, I plan to start with CONV_3D first, then follow up with CONV_3D_TRANSPOSE in a separate PR if this direction looks reasonable.

[Aharrypotter]

Hi @tlopex, I'd like to claim I. StableHLO operators in TFLite.

I'm very interested in StableHLO recently, especially because I'm currently learning JAX/XLA/StableHLO-related stacks. I'd like to take the full StableHLO-in-TFLite operator family and submit the implementation incrementally through multiple reviewable PRs.

I will start with a small subset first, add the corresponding Relax TFLite frontend handlers and structural-equal tests, then extend the coverage step by step.