$ the-wire · showcase
XPU flash-attention speed fix, FSDP2 dtype cast for async RL
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
Two published-kernel regressions got fixed today: slow XPU flash-attention builds under DPC++ 2026.1 and a dtype mismatch in TRL's FSDP2-to-vLLM weight transfer.
Fix XPU flash-attention performance on DPC++ 2026.1 (#921) huggingface/kernels
Two independent problems in dep-sycl-tla.cmake made the published XPU kernels slow; the fix was measured end to end on flash-attn2 via nix build .#redistributable.torch214-xpu20261-x86_64-linux, using Intel Arc Pro B60, torch 2.14.0+xpu, fp16 and causal, against a sycl-tla v0.8 baseline with identical AOT targets across full 86-TU builds.
Cast FSDP2 weight-transfer payloads to the vLLM manifest dtype in AsyncGRPO and AsyncDistillation huggingface/trl
AsyncGRPO and AsyncDistillation now cast FSDP2 weight-transfer payloads to the vLLM manifest dtype. Review trimmed the fix to the cast alone (59dca9d4): since the manifest and the stream walk _iter_vllm_named_params on the same model, only the dtype can disagree, and once cast, "vLLM can no longer return while the sender is blocked" per the PR description.
nix-builder: bump nixpkgs and remove cacert overrides huggingface/kernels
The nix-builder bumps nixpkgs, which drops the cacert overrides and fixes macOS builds, while libnvshmem moves to a binary package to cut build times. The new patchelf that comes with it is broken against some binary packages, so an older patchelf is pinned where needed.
Fix port race between xdist workers in the OpenReward test fixture (#7886) huggingface/trl
The echo_env_url test fixture picked a free port, released it, then let the server subprocess rebind it, so two pytest-xdist workers could collide: the second server exited unnoticed while its health check passed against the first worker's server. The fixture now holds the socket it already bound.
