$ the-wire · showcase
HyperClovaX Vision lands in transformers, XPU test baselines reworked
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
transformers added HyperClovaX Vision support and re-baselined three non-deterministic Glm46V batch tests on XPU, while kernels-community fixed ten broken relative imports in its vendored fla source.
huggingface/transformers added HyperClovaX Vision, bringing the vision tower and processor into the model library [1]. On the test side, the three batched `Glm46VIntegrationTest` generation tests hardcoded CUDA output; they now account for XPU producing a different but valid completion. One of them, `test_small_model_integration_test_batch`, is genuinely non-deterministic on XPU, showing 2 distinct outputs across 4 runs with `use_deterministic_algorithms` off [2][3]. Separately, tests gained a `MemoryCleanupMixin` class to manage cleanup in test runs [4].
kernels-community fixed the vendored fla source, where ten relative imports in `fla/torch-ext/fla` used a dot count one or two levels too shallow. Inside an `__init__.py` the current package is the directory itself, one level deeper than for a plain module, so `from ...ops.common...` in `ops/common/backends/tilelang/__init__.py` resolved against `fla.ops.common` [5][6]. Verification ran against the published build `aa89c60` using the original repro: the `chunk_gated_delta_rule` backward now runs, and gradients match the repo's own `naive_chunk_gated_delta_rule` to 1.9e-3 relative in fp32 [5]. The same desk trimmed the smoke tests by using the chunk variation and dropping the recurrent versions [7][8].
huggingface/trl added support for vLLM 0.29.0 [9]. Its CI now cancels superseded runs on `main`: previously the `concurrency` fallback `github.event.pull_request.number || github.run_id` gave every push event a unique group, so pushes were never cancelled. It now uses `github.ref`, grouping push runs by branch, so a newer push to `main` cancels the run started by the previous merge [10][11]. trl also moved to skip tests on pushes that change no tested file. `tests.yml` filters `pull_request` events by path, but the `push:` trigger carries no filter, so docs-only changes still ran the full matrix once landed. From 2026-08-26 to 2026-09-10, 27 of the 114 commits that reached `main` would not have matched the `pull_request` filter, 10 touching nothing but `.md` files, at around 77 GPU runner-minutes per matrix [12].
Action items
- → Update trl to vLLM 0.29.0 if you pin vLLM in your training stack huggingface/trl [plan]
- → Rebuild kernels-community fla kernels at or after build aa89c60 to pick up the relative import fix huggingface/kernels-community [plan]
- → Re-check Glm46V batched generation on XPU against the new baselines huggingface/transformers [monitor]
References
- [1] add HyperClovaX Vision ↗ huggingface/transformers
- [2] Fix slow integration tests on XPU (#48611) ↗ huggingface/transformers
- [3] Fix slow integration tests on XPU ↗ huggingface/transformers
- [4] QA: Added a `MemoryCleanupMixin` class for tests ↗ huggingface/transformers
- [5] fla: fix relative import depths in the vendored source and add smoke tests ↗ huggingface/kernels-community
- [6] fla: fix relative import depths in the vendored source and add smoke tests (#1129) ↗ huggingface/kernels-community
- [7] fla: shorter smoke tests using the chunk variation ↗ huggingface/kernels-community
- [8] fla: shorter smoke tests using the chunk variation (#1146) ↗ huggingface/kernels-community
- [9] Add support for vLLM 0.29.0 (#7178) ↗ huggingface/trl
- [10] Cancel superseded runs on main ↗ huggingface/trl
- [11] Cancel superseded runs on main (#7172) ↗ huggingface/trl
- [12] Skip the tests on pushes that change no tested file ↗ huggingface/trl
