$ the-wire ยท showcase
MoE expert sharding gets a test, lerobot swaps vendor SDKs for rustypot
By RepoJournal ยท Filed ยท About Hugging Face ยท Composed from the cited sources ยท methodology
Transformers gained a regression test that catches MoE parallel plans silently leaving every expert on every rank, while lerobot replaced both vendor serial SDKs with a single Rust-backed bus.
๐จ MoE: shard the experts in every TP/EP plan, and test for it huggingface/transformers
A MoE parallel plan that misses the experts still produces correct outputs, so nothing fails when sharding quietly degrades to full replication; the new `test_moe_parallel_plans_shard_experts` in `TensorParallelTesterM` asserts the experts are actually split across ranks.
refactor(motors): one rustypot-backed SerialMotorsBus for Feetech and Dynamixel huggingface/lerobot
The Feetech and Dynamixel buses now run through rustypot (>= 1.13.1) behind one `SerialMotorsBus`, with the vendor control tables removed and framing, checksums and packet parsing moved into compiled code that releases the GIL. `feetech-servo-sdk` and `dynamixel-sdk` are no longer involved.
Measure GPU leaks per class and leave failing runs out huggingface/transformers-ci
`pytest_test_cuda_delta_bytes` was being read as a leak signal, but a failing test's traceback keeps its frames and the model in their locals alive past teardown; on 2026-10-17, 25 of the 40 largest deltas came from failing tests. Deltas are now measured per class with failing runs excluded, so the leak number means what it says.
Merge-queue history per failing test on the PR CI view huggingface/transformers-ci
The PR CI view now reports whether a failing test also fails in other PRs' merge-queue runs, via `GET /pr-ci-view/queue-history?pr=N&t=<nodeid>`, backed by a three-day Tempo search over `gh-readonly-queue/*` refs. It separates your breakage from a test that was already red on `main` plus other PRs.
Quantization cleanpup (#49548) huggingface/transformers
ArthurZucker's quantization cleanup unwinds method patching left in the weight converter, splitting serialization into a plain cast versus a `ConversionOps` step and moving TP plan updates out of bloated code. Separately, the docs desk added a Google TPU hardware guide and a sys.meta_path finder for legacy module aliases, and the CI desk fixed 31-character Tempo trace ids, landed a Tests catego...
