OmarAzizi created an issue (apache/tvm#20124)

## Context

The Relax ONNX frontend currently implements ~160 operators from the ONNX spec. 
Comparing against the current operator list 
(https://onnx.ai/onnx/operators/index.html), a number of ops are not yet 
supported.

Before opening PRs, I'd like to check on priorities/approach with maintainers. 
I went through `python/tvm/relax/op/` and `python/tvm/topi/` to check what's 
actually available to build on, rather than guessing from op names alone, so 
the grouping below reflects what I found there.

## Missing operators

| Category | Operators |
|---|---|
| Activations / math | Celu, Swish, LpNormalization, LinearAttention, Det, 
Col2Im, TensorScatter |
| Casting / bit-level | CastLike, BitCast |
| Cropping / sequence | CenterCropPad, ReverseSequence, SequenceMap |
| Normalization | GroupNormalization |
| Recurrent networks | RNN, GRU, LSTM |
| Convolution | ConvInteger, DeformConv, CausalConvWithState |
| Quantization | QLinearConv, QLinearMatMul |
| Random / sampling | RandomNormal, RandomNormalLike, RandomUniform, 
RandomUniformLike, Multinomial, Bernoulli |
| Signal processing | DFT, STFT, BlackmanWindow, HammingWindow, HannWindow, 
MelWeightMatrix |
| Control flow | Loop, Scan |
| Loss functions | NegativeLogLikelihoodLoss, SoftmaxCrossEntropyLoss |
| String / text | StringConcat, StringSplit, StringNormalizer, RegexFullMatch, 
TfIdfVectorizer |
| Attention / misc | RotaryEmbedding, ImageDecoder |

## Implementation notes

These vary a lot in difficulty. Some already have a clear path based on what's 
in the codebase:

- **RNN, GRU, LSTM** - no fused recurrent op in Relax, but the torch and tflite 
Relax frontends already implement LSTM by unrolling the recurrence with 
existing ops; the ONNX converter could follow the same pattern.
- **CastLike, Swish** - Relax has direct equivalents (`astype`, `nn.silu`) that 
cover the common case (default `alpha=1.0` for Swish; standard dtype casts for 
CastLike). Non-default variants (float8 `round_mode`/`saturate` for CastLike, 
non-unit alpha for Swish) would need a small amount of extra composition.
- **CumProd, ReverseSequence** - Relax has ops covering the same purpose 
(`cumprod`, `reverse_sequence`), but attribute defaults/semantics differ from 
the ONNX spec (Relax's `cumprod` has no reverse mode; `reverse_sequence`'s axis 
defaults are swapped relative to ONNX's `time_axis`/`batch_axis`), so the 
converter needs explicit attribute mapping rather than a pass-through.

Others (quantized conv/matmul, signal processing, `Loop`/`Scan`) likely need 
new compute/schedule work or, in the case of `Loop`/`Scan`, graph-level 
control-flow support. I'm not certain of the right approach here and would 
appreciate maintainer input.

## Questions for maintainers

1. Is there an existing priority order for ONNX op coverage, or known user 
demand for any of the above from specific model families?
2. Are any of these considered out of scope or low priority for the Relax 
frontend (e.g. control flow, quantized ops, signal processing)?

## Proposal

I'd like to pick up work on some of these, starting with the ones that have a 
clear implementation path (CastLike, Swish, ReverseSequence, CumProd attribute 
mapping, then RNN/GRU/LSTM via unrolling). 

## Questions for the maintainers

- Is there an existing priority order for ONNX op coverage, or known user 
demand for any of the above from specific model families?
- Are any of these (e.g. control flow, quantized ops, signal processing) 
considered out of scope or low priority for the Relax frontend?

Happy to pick up items above, split into smaller PRs by category, and help 
review PRs from others in this area as it progresses.

### Triage
* needs-triage
* frontend:onnx
* status: RFC

-- 
Reply to this email directly or view it on GitHub:
https://github.com/apache/tvm/issues/20124
You are receiving this because you are subscribed to this thread.

Message ID: <apache/tvm/issues/[email protected]>

Reply via email to