$ the-wire · showcase
TRL seeds before adapter init, docs links repaired
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
TRL's experimental trainers now honor args.seed when initializing LoRA adapters, and a sweep of dead links across Hugging Face blog posts, kernels tooling, and model cards landed alongside an automated refresh of the kernel ABI symbol files.
Seed before creating the model in experimental trainers huggingface/trl
These trainers built the PEFT model in __init__ before Trainer.__init__ called set_seed(args.seed), so adapter initialization drew on whatever RNG state happened to be present: two runs with the same seed started from different adapters. The fix calls set_seed(args.seed) first in AsyncGRPOTrainer, CPOTrainer, IWOPDTrainer and the other experimental trainers, matching what the main trainers alre...
Fix dead links in 2026 posts (#3540) huggingface/blog
Every 2026 post link that returned 404 or lacked a scheme now points at a verified target: Reachy Mini docs moved to docs/source on main, the kernels repo renamed builder/ to kernel-builder/, and the skills repo renamed hugging-face-model-trainer to huggingface-llm-trainer. Anyone following those guides from a saved bookmark or internal doc was hitting dead ends.
kernel-abi-check: refresh vendored ABI symbol files huggingface/kernels
The update_abi_symbols workflow re-vendored kernel-abi-check's symbol files against CPython's stable_abi.toml, auditwheel's manylinux-policy.json, and PyTorch's shim_function_versions.txt, which had drifted from upstream. The ABI check now validates against current symbols rather than stale ones.
Align tiny DeepSeek-V3-0528 config with deepseek-ai/DeepSeek-R1-0528 huggingface/trl
The tiny-DeepseekV3ForCausalLM-0528 generator config now mirrors deepseek-ai/DeepSeek-R1-0528, split across one commit for fields DeepseekV3Config models and another for the keys it does not. It is the sibling of the base-model alignment, and the two references differ on none of the affected fields.
Fix dead links in 2026 posts huggingface/blog
The long tail: the Inferenece Providers docs got an automated refresh, model-card user study images switched to the correct resolve path, the github-actions-and-pre-commit group moved doc-builder forward, and the GRPO continuous batching test under DataParallel is xfailed to unblock CI.
