StableHLO converter: missing ops RoPE, groupNorm, upsample/interpolate (gaps remaining after #666)

[michalharakal]

Summary

Issue #666 (closed COMPLETED 2026-06-06) tracked, among other things, "RoPE/groupNorm/upsample have no op." As of 0.31.2, three of those converter ops are still missing — a GraphNode with any of these operation names falls through NeuralNetOperationsConverter / ActivationOperationsConverter to the "Operation not supported" path and never lowers to StableHLO.

Evidence (0.31.2)

The combined supportedOperations across the HLO converters contains no entry for rope/rotary, groupNorm, or upsample/interpolate:

... batchNorm layerNorm rmsNorm silu swish embedding scaledDotProductAttention
    conv1d conv2d maxPool2d avgPool2d gather reshape transpose ... (no rope/groupNorm/upsample)

Downstream, the skainet-iree-conformance harness records these as ❌ "no converter op" in its capability matrix; RoPE/GroupNorm/upsample are the only remaining hard gaps now that LayerNorm/RMSNorm/BatchNorm/SiLU/Embedding all lower end-to-end and validate on IREE.

Missing ops

Op	Difficulty	Lowering sketch
groupNorm	low	reshape (N,C,*)→(N,G,M), per-group mean/var (reuse @reduce_mean/@reduce_variance as LayerNorm does), reshape back, per-channel affine. Addressed by the PR linked below.
upsample / interpolate (nearest)	medium	integer-factor nearest-neighbour via broadcast_in_dim + reshape; bilinear needs gather/weighting.
RoPE (rotary position embedding)	high	even/odd split (needs reshape — now fixed), per-position cos/sin tables, rotate-and-recombine. cos/sin should be materialised as constants (not externalised as un-fed args).

Proposed plan

1. groupNorm — implement now (sibling of the existing LayerNorm decomposition). PR incoming.
2. upsample (nearest) — follow-up.
3. RoPE — follow-up; the largest piece, needs a design note on the cos/sin source.

Filing this so the remaining gaps are tracked explicitly rather than under the now-closed #666.

[michalharakal]

Resolved — all three remaining converter gaps now lower to StableHLO and validate end-to-end on IREE (llvm-cpu) in the skainet-iree-conformance harness.

Gap	Status
groupNorm	✅ converter (#752) + real-stablehlo.reduce fix (#754), shipped in 0.32.1. Conformance op-suite: groupnorm: ok (25 stablehlo ops), validates vs numpy.
upsample / interpolate	✅ Nearest and Bilinear converter (#770 / #771), merged to develop and queued for 0.33.0 (release PR #775). Everything lowers to fixed reshape/broadcast_in_dim/dot_general — no custom_call. Conformance: upsample2d_nearest max_abs_err 0.0, upsample2d_bilinear 1.19e-7.
RoPE	✅ exportable via the NN-DSL decomposition — even/odd (or split-half) reshape + per-position cos/sin materialised as stablehlo.constant (not externalised), enabled by the reshape fix and the isRecording traceable path (#757). Conformance: op-suite rope: ok (9 ops) and the rope-attention model export ok(128 ops) · compile · run · validate ok. No dedicated "rope" converter op was needed.

Closing — the three gaps tracked here are done. Remaining nuance (tracked elsewhere): a few newly-differentiable ops (tril/indexSelect/unfold/convTranspose1d, #774) are trainable but not yet StableHLO-exportable.