$ cat huggingface/week/2026-11-02.log
the week in review · Nov 2 – Nov 8, 2026
LeRobot 0.7.0 rebuilds install, training and hardware layer
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Transformers also renamed gradient_checkpointing and dropped rope deltas this week, while TRL taught tool calling to GLM and Qwen3-Coder templates.
Release v0.7.0 huggingface/lerobot
The release leaps from v0.6.1 to 0.7.0 and rewrites installation, training, inference and the hardware layer, with several changes the notes flag as requiring migration. Check the breaking-change section against your pinned setup before upgrading.
Rename `gradient_checkpointing` to `activation_checkpointing` in the public API (#49248) huggingface/transformers
gradient_checkpointing is now activation_checkpointing in the public API, with the old names kept as deprecated aliases, so existing code keeps running but the deprecation clock is running. The rope delta fields are gone from ModelOutput, which breaks anything that reads them.
:rotating_light: Finally drop rope deltas from ModelOutput! huggingface/transformers
The rope delta entries that ModelOutput had been carrying are removed outright. Code that unpacked or indexed those fields needs updating rather than a rename.
fix(worker): stop the web app from blocking on its unread stdout pipe huggingface/dataset-viewer
Worker pods were restarting because mirakuru's TCPExecutor left the web app's stdout pipe unread, so /healthcheck and /metrics stopped answering after about 40 minutes; worker-medium saw roughly 1,700 restarts a day, mostly liveness probe failures. The fix stops the block at the source.
Recognize the GLM-4.7 and GLM-5 chat templates for tool calling huggingface/trl
TRL now recognizes the chat templates shipped by zai-org's GLM-4.7, GLM-4.7-Flash and GLM-5 through GLM-5.3 checkpoints, so add_response_schema and tool-calling GRPOTrainer work on them. The same treatment applied to Qwen2.5 base and Qwen3-Coder templates this week, and assistant_only_loss now works for Qwen2.5 base.
Support assistant-only loss for the Qwen2.5 base chat template huggingface/trl
assistant_only_loss=True previously had nothing to key on for the Qwen2.5 base checkpoints, since TRL does not store that template. It now works for Qwen/Qwen2.5-0.5B through the 14B variant and beyond, where you were otherwise masking by hand.
FEAT healthcheck: warn when a randomly initialized task head is not saved huggingface/peft
healthcheck now emits a TASK_HEAD_NOT_SAVED warning, usually triggered when task_type is unset and the randomly initialized head never lands in modules_to_save. That is a silently untrained head caught at save time instead of at eval time.
FIX reduce_intruder_dimension for fan_in_fan_out and embedding LoRA layers huggingface/peft
reduce_intruder_dimension rebuilt lora_A and lora_B without transposing the delta for fan_in_fan_out and nn.Embedding layers, so GPT-2 Conv1D adapters came back with wrong shapes. Alongside: workflow changes land across the stack, a fix for bfloat16 baselines on CPUs without avx512_bf16, a BPE loading speedup, and two modular diffusers pipelines.
$ ls huggingface/week/ # the briefings behind this review
Keep up with Hugging Face in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.