Aharrypotter commented on issue #19412: URL: https://github.com/apache/tvm/issues/19412#issuecomment-4787683805
> Hi [@tlopex](https://github.com/tlopex), I’d like to help close out the remaining unmerged items in this tracking issue. > > I noticed a few of the unchecked items were claimed back in April, but I could not find corresponding merged PRs for them, and the issue has been quiet for several weeks. To avoid leaving the tracking issue stalled, I plan to work on the remaining gaps in one cleanup PR: > > * UNIQUE, SIGN, BITWISE_XOR, RIGHT_SHIFT, BUCKETIZE, RELU_0_TO_1 > * RANK, UNSORTED_SEGMENT_SUM, UNSORTED_SEGMENT_MAX > * RANGE dynamic scalar inputs and FAKE_QUANT narrow_range/vector coverage > * DYNAMIC_UPDATE_SLICE status cleanup, depending on the current TFLite schema support > > If anyone is already actively working on any of these, happy to coordinate or drop that part. Hi @tlopex — following up on my earlier comment. I went back through the remaining unchecked boxes; with the PRs below, all of them are now addressed. Quick status: **A. Good-first-issues — all landed & tested** - `CUMSUM` (#19434), `UNIQUE`, `SIGN`, `BITWISE_XOR`, `RIGHT_SHIFT`, `BUCKETIZE`, `RELU_0_TO_1` (all in #19813) **B. Medium — all landed & tested** - `RANK`, `UNSORTED_SEGMENT_SUM`, `UNSORTED_SEGMENT_MAX` (#19813); the other segment ops landed earlier in #19491 - `DYNAMIC_UPDATE_SLICE` — runtime (dynamic) start indices now supported in #19881 (builds the `scatter_nd` index grid in-graph with StableHLO clamping; the static path from #19587 is unchanged) **C. Fix partial implementations** - `FAKE_QUANT` — #19813 fixed the handler crashing on a non-existent `relax.op.const` call (switched to `relax.const`), which had broken the op on every invocation, and added `test_fake_quant_narrow_range_vector` as a regression test. The builtin FAKE_QUANT parameter space (scalar min/max, `narrow_range`, `num_bits`) is fully supported (the builtin op has no per-channel/vector min/max). - `RANGE` — runtime (dynamic) scalar bounds now supported in #19867, for both integer and float, ascending or descending. Both dynamic-input items fit cleanly in the frontend: `RANGE` via `tensor_to_shape` + `match_cast`, `DYNAMIC_UPDATE_SLICE` via `scatter_nd` with runtime indices. No new Relax op was needed. Suggest closing this issue once #19881 merges. Please let me know if I've mischaracterized anything. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
