$ the-wire · showcase
Dataset Viewer pods were killing themselves over a full stdout pipe
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
The day's most consequential fix is an infrastructure one: dataset-viewer traced thousands of daily pod restarts to an unread subprocess pipe and an idle memory limit set below the actual idle footprint, while TRL closed out chat-template support for assistant-only loss and tool calling across four checkpoint families.
fix(worker): stop the web app from blocking on its unread stdout pipe huggingface/dataset-viewer
The worker's web app died every ~40 minutes because mirakuru's TCPExecutor defaults to stdout=subprocess.PIPE and nothing drained it, so uvicorn's per-request access log line (~65 bytes per probe) filled the 64 KiB buffer and the next write blocked, taking /healthcheck down with it. In prod, worker-medium was restarting about 1,700 times a day, mostly from liveness probe failures; stdout now go...
fix(chart): give the light workers room for a job above their idle memory (#3498) huggingface/dataset-viewer
The light workers' 1Gi limit sat below their actual idle footprint: executor, worker loop and web app together use ~975Mi before any job runs, so almost every job triggered an OOM kill, roughly 2,900 restarts a day. The chart now requests the idle usage and raises the limit to 1536Mi.
Support assistant-only loss for the Qwen2.5 base chat template huggingface/trl
assistant_only_loss=True now works for the chat template shipped by the Qwen2.5 base checkpoints (0.5B through 14B and up), which TRL does not store, so get_training_chat_template previously raised and SFTTrainer failed at init. The template is qwen2_5.jinja with a different default system prompt, so the loss mask needed its own handling.
Recognize the Qwen3-Coder chat templates for tool calling huggingface/trl
TRL now recognizes the templates shipped by Qwen3-Coder-30B-A3B-Instruct, Qwen3-Coder-480B-A35B-Instruct and Qwen3-Coder-Next, so add_response_schema and tool-calling GRPOTrainer stop raising Unrecognized chat template at init. They render tool calls in the same XML format as Qwen3.5.
Planner: greedy first draft, closest approved contract as example, real parameter-name patterns; docs vocabulary huggingface/transformers-mic
The planner moved to a greedy first draft at temperature 0 with retries at 0.7, and now seeds its example with the closest approved contract from the control and prototype fleets; on planner-eval (4 models x 3 drafts, no repairs) build pass went from 5/12 on main to 9/12 in round 1. The merge commit message reports 6/9 after fixes in round 2, with SmolVLM still running.
MIC: never compare or record bfloat16 on a CPU without native bfloat16 huggingface/transformers-ci
Elsewhere: bfloat16 cases on an x86 CPU lacking avx512_bf16/amx_bf16 now fail with "wrong hardware" and write no candidate, instead of being compared against old cpu-x86_64 baselines for llava, mistral3 and smolvlm; transformers-ci also synced its fixture schema so weights: "real" builds load (that setup error hit 21 runs in a day).
