The GitHub Actions job "CI" on tvm.git/fix/relax-torch-split-int-split-size has succeeded. Run started by GitHub user siyiweigeHEW (triggered by siyiweigeHEW).
Head commit for run: e907c6439c8a702ac291830682ae6b951d108806 / HuEnwei <[email protected]> [Relax][Frontend][PyTorch] Fix x.split(int) with a non-divisible split_size torch.split(x, s, dim) splits dim into chunks of size s, with the last chunk smaller when D % s != 0. The frontend `_split` converter (base_fx_graph_translator.py) converted the per-chunk size into a section count n_section = ceil(D / s) and passed it to relax.op.split's integer argument, whose semantics are "split into n_section equal sections" (each of size ceil(D / n_section)). Whenever ceil(D / ceil(D / s)) != s -- e.g. split_size > D/2 with a non-divisible D -- valid PyTorch models silently produced differently-shaped chunks (e.g. x.split(6) on a (10,) tensor yielded (5,), (5,) instead of (6,), (4,)). Convert the int per-chunk size to the cumulative cut positions [s, 2s, ..., (ceil(D/s) - 1)*s] instead, the same list form the list/tuple branch already passes to relax.op.split. Differential testing vs native PyTorch (fixed `_split` injected verbatim): the 10 previously-diverging cases (non-divisible split_size > D/2 across dim=0, dim=1 and negative dims) all now match torch shapes and values; the divisible-size baseline and the list/tuple (split_with_sizes) control group remain unchanged. Added regression test test_split_int_split_size covering the structural IR (R.split(input, indices_or_sections=[6], axis=0) for x.split(6) on a (10,) input) and numerical equivalence for non-divisible sizes/dims. The pre-existing test_split expected IR in tests/python/relax/test_frontend_from_fx.py asserted the old int-section form (indices_or_sections=3 for x.split(1, dim=1)); it is updated to the correct list form ([1, 2]). Co-Authored-By: Claude <[email protected]> Report URL: https://github.com/apache/tvm/actions/runs/33331841488 With regards, GitHub Actions via GitBox --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
