$ the-wire · showcase
serge PR triage tightens, TPU and multi-GPU test fixes land
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
serge now only writes the expectation updates maintainers actually merge and closes the PRs that never needed review, while transformers and TRL close test gaps on TPU and split-device training.
Send back expectation changes that are not on the CI runner's key huggingface/serge
Expectation updates now have to land on the CI runner's own ("cuda", 8) key. Of the serge PRs opened in Sept and Oct 2026, the three that touched only that entry merged cleanly, while the one that overwrote ("cuda", None) merged with a wrong value and the open queue rewrote XPU values with no XPU evidence and loosened tolerances.
Let the nightly triage close serge's stale task PRs huggingface/serge
A nightly triage hook can finally close serge's stale PRs itself, since the triage token could only read them. It targets the 19 of 47 open serge PRs on huggingface/transformers on 2026-10-11 that needed no review: 14 covering tests that already pass on main, three duplicating a newer PR, and two recording degenerate expectations.
ITF: dedupe serge PRs by test, retire stale and superseded ones huggingface/transformers-ci
Duplicate PRs get deduped by test instead of by fingerprint, because target_fingerprint hashes the group's label and GPU set, both of which shift night to night for an unchanged failure. Six groups had two open PRs, each move having minted a new fingerprint and a second PR.
ITF: parse the CI dataset's "(line N)" traces huggingface/transformers-ci
The ITF parser now reads the CI dataset's "(line N) AssertionError" traces, which terminal_exception() previously returned as (None, "") for all 376 integration failures on 2026-10-10. Because every group's terminal_exc collapsed to "other", crash groups were never split by exception type.
Count ep and fsdp in caching_allocator_warmup byte calculation huggingface/transformers
get_total_byte_count now counts ep and fsdp in the caching_allocator_warmup byte calculation. Since the TP and EP plans were decoupled, EP-sharded experts stopped appearing in model.tp_plan unless it happened to list them, so an EP-only layout divided sizes incorrectly.
Test DPO, KTO, Reward, GRPO and RLOO on a model split across devices huggingface/trl
TRL adds distributed tests for DPO, KTO, Reward, GRPO and RLOO on a model split across two GPUs with device_map, following the SFT coverage added for #7552. The Reward test splits a sequence classification model.
[tests] Some fixes for test suite failures on TPU huggingface/transformers
TPU test-suite failures get a batch of fixes: CUDA SDPA tolerances on TPU, full-precision float32 matmuls (enable_tf32(False) only sets the new fp32_precision API, while torch_tpu follows legacy float32_matmul_precision, which it sets to "medium"), and in-process subprocess tests on single-process devices.
