$ the-wire · showcase
XPU rollout images, a tensordict 0.11 data fix, and a gated-judge guard
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl opens Intel Arc Pro B-series as a first-class rollout target while patching two Ascend-specific breakages, and open-instruct stops burning GPQA runs on a missing HF_TOKEN.
[hardware, docs] feat: add multi-stage Intel GPU (XPU) Docker image for vLLM 0.32.0 verl-project/verl
A new docker/xpu/Dockerfile.stable.vllm gives you a multi-stage image pairing vLLM 0.32.0+xpu rollout with FSDP/FSDP2 training on torch 2.14.0+xpu, with XPU support supplied by verl-hardware-plugin installed alongside verl. The transformers window now matches the 0.32.0 CUDA stack, so XPU and GPU runs sit on the same version set.
[data] fix: keep the first NonTensorData in concat_tensordict with tensordict >= 0.11 verl-project/verl
From tensordict 0.11, TensorDict.cat stacks differing NonTensorData into a NonTensorStack with one value per sample, so concat_tensordict now restores the documented behavior of keeping the first instance. If you are on tensordict 0.11 and relied on per-batch metadata surviving concatenation, check your data path before upgrading.
Require HF_TOKEN for gated GPQA evals; document eval batch timing allenai/open-instruct
gpqa_diamond and gpqa_main:cot pull Idavidrein/gpqa, which is gated on the Hub, and the earlier credential guard covered only WildGuard and Gemma judges, so a panel without HF_TOKEN passed planning and then died at GPQA after finishing its 7 math tasks. validate_judge_credentials now covers GPQA task names, with tests.
Fix MILES evaluator sandbox packaging and gated judge preflight allenai/open-instruct
The pinned olmo-eval wheel drops sandbox shell scripts, which made LiveCodeBench abort before scoring when it could not load monitor.sh. The image now uses an editable install, checks the resource at build time, and prints a bounded log tail to Beaker for failed evaluator subprocesses while keeping full result logs and the original exception.
[trainer] fix: preserve the token axis when packing position IDs (#8413) verl-project/verl
Qwen3.5-4B GRPO training on Ascend 910B failed in actor minibatch selection when unbinding a nested tensor; position IDs now keep the token axis through packing. Two older fixes in the same window: a fused-MoE weight transposition fix for DeepSeek V4 Flash with FSDP-Turbo and vLLM-Ascend on NPU, and TPU overlays that mount only post-training/data and tokenizers instead of listing every checkpoint.
Action items
- → Add HF_TOKEN to any GPQA evaluation panel or expect the run to fail at the first GPQA task allenai/open-instruct [plan]
