[Tracking Issue][TFlite] Expand Missing / Partial Operator Coverage

[tlopex]

Background

Excited to see the huge progress the community has made on the Relax TFLite frontend recently — many operators and tests have landed in the last few weeks. To keep pushing this forward, this issue tracks the remaining operator gap so contributors can pick up well-scoped items in parallel.

A. Good-first-issues — drop-in mappings

Workflow: add a handler, wire into convert_map (alphabetical within its region), add one verify(TestClass, Expected) test in tests/python/relax/test_frontend_tflite.py (pattern at L55). All target Relax ops already exist — no new op work needed.

• CUMSUM → relax.op.cumsum
• UNIQUE → relax.op.unique
• SCATTER_ND → relax.op.scatter_nd
• BROADCAST_TO → relax.op.broadcast_to
• SIGN → relax.op.sign
• BITWISE_XOR → relax.op.bitwise_xor
• RIGHT_SHIFT → relax.op.right_shift
• BUCKETIZE → relax.op.bucketize
• EMBEDDING_LOOKUP → relax.op.take
• SELECT_V2 → relax.op.where
• RELU_0_TO_1 → relax.op.clip
• NON_MAX_SUPPRESSION_V4 → port from existing convert_nms_v5

B. Medium — may need new Relax op or non-trivial decomposition

Prefer composing with existing Relax ops over adding a new op. Only register a new relax.op.* when the semantics genuinely cannot be expressed by existing primitives (e.g. random generators, bit-level reinterpret). If a decomposition using existing ops is feasible, go with that — it's a smaller PR and avoids bloating the op set.

• REDUCE_ANY — commented at tflite_frontend.py:198; can be lowered to relax.op.max on a bool tensor, no new op needed
• REDUCE_ALL — same, lower to relax.op.min on bool
• RANK — emit relax.const(ndim) for static shapes; no op call needed
• ATAN2 — compose from atan + sign + where, or add relax.op.atan2 for a cleaner impl
• SEGMENT_SUM — reuse relax.op.scatter_nd with reduction="add" (TFLite requires sorted segment ids)
• UNSORTED_SEGMENT_SUM — scatter_nd(reduction="add")
• UNSORTED_SEGMENT_MAX — scatter_nd(reduction="max")
• UNSORTED_SEGMENT_MIN — scatter_nd(reduction="min")
• UNSORTED_SEGMENT_PROD — scatter_nd(reduction="mul")
• DYNAMIC_UPDATE_SLICE — reuse relax.op.slice_scatter
• DILATE — likely needs a new op (or decompose via pad + strided_slice)
• MULTINOMIAL — wrap existing relax.op.multinomial_from_uniform; blocked on RANDOM_UNIFORM
• RANDOM_STANDARD_NORMAL — needs a new random-generator op in Relax
• RANDOM_UNIFORM — same
• BITCAST — needs a new reinterpret/view op
• BROADCAST_ARGS — shape-level computation; consider static-shape-only path first

C. Fix partial implementations

These ops are already in convert_map but their handler raises OpNotImplemented on certain input patterns, making them unusable on real models that hit those patterns.

• RANGE — handler raises on dynamic scalar inputs (tflite_frontend.py:958); should lower to relax.op.arange.
• FILL — handler raises when dims is non-constant (L1729); should lower to relax.op.full with a dynamic shape
• SPLIT_V — handler raises when size_splits is non-constant (L2267); decompose into dynamic strided slices.
• NON_MAX_SUPPRESSION_V5 — handler raises on soft-NMS (L3546); needs soft_nms_sigma != 0 support
• FAKE_QUANT — handler raises on some narrow-range vector configs (L3913)

Contributor 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) pattern (test_frontend_tflite.py:55). Nightly-gated E2E activates automatically via CI_ENV_NIGHTLY.
5. Tag the PR title [Relax][Frontend][TFLite].

cc @leandron

[swjng]

Claiming B: REDUCE_ANY and REDUCE_ALL — will lower to relax.op.max / relax.op.min on a bool tensor as suggested, and bundle them into a single PR.

[Aharrypotter]

Hi @tlopex , I'd like to claim NON_MAX_SUPPRESSION_V5 under "Fix partial implementations" to add support for soft_nms_sigma != 0. I will open a PR with the implementation and tests soon.

[Dayuxiaoshui]

Hi @tlopex, I'd like to claim the following tasks:
A. Good-first-issues:
BITWISE_XOR → relax.op.bitwise_xor
RIGHT_SHIFT → relax.op.right_shift
RELU_0_TO_1 → relax.op.clip
B. Medium:
RANK — will emit relax.const for static shapes
UNSORTED_SEGMENT_MAX — will use relax.op.scatter_nd with reduction="max"
I'll open PRs for these implementations soon.

[Aharrypotter]

Hi @tlopex , I'd like to claim FILL and SPLIT_V under "Fix partial implementations" . I will open a PR with the implementation and tests soon.