$ the-wire · showcase
Qwen3.5 GRPO lands on TPU v7x, and verl fixes a rollout stall that made weight syncs wait
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl shipped a Qwen3.5 recipe for TPU v7x and a rollout fix that shortens the pause before every weight sync, while open-instruct operators get direct control over Beaker job scheduling before the October scheduler cutover.
[hardware, recipe] feat: Qwen3.5 (incl. 35B-A3B MoE) GRPO recipe for TPU v7x, plus context parallelism and QKV fusion configs verl-project/verl
verl added a Qwen3.5 GRPO recipe for TPU v7x, defaulting to Qwen/Qwen3.5-35B-A3B with thinking mode enabled, and bumped the TPU GRPO Dockerfile to vLLM v0.32.0. The same PR sets --xla_tpu_bf16_emission_mode=PROMOTE_F32 across the existing TPU GRPO recipes and adds context-parallelism, QKV-fusion and VERL_TPU_TIMELINE knobs.
[rollout, vllm] fix: release the submission-gate barrier when a request lands in the engine, not on its first token verl-project/verl
The submission gate used to release only on a request's first output, so with len(batch) > max_num_seqs a pre-weight-sync pause waited for every queued request to be scheduled and prefilled. The barrier now releases when a request reaches the engine, which shortens the window between abort_all_requests() and pause_generation().
Add --min_runtime and --auto_resume to mason allenai/open-instruct
Beaker's allocation-based scheduler is live on Jupiter, Titan and Ceres and switches on Saturn, Neptune, Rhea and Prometheus on 2026-10-22. mason can now send minRuntime and autoResume directly instead of the legacy preemptible field, which only expressed min runtime 0 or the 8-hour maximum; the flags are opt-in and the spec is unchanged without them, and beaker-py is bumped to 2.6.0, the first...
Bound MILES producer budgets and expose opt-in judge CUDA graphs allenai/open-instruct
The one-node MILES example now defaults to 1,024 samples instead of 1,536, and the three-node example uses 1,536 explicitly instead of the omitted-setting heuristic of 19,456, with per-engine admission still at 512. Managed SGLang judges also accept judges.NAME.cuda_graphs=true to enable both decode and prefill CUDA graphs.
[rollout, hardware, recipe] fix: run the DeepSeek-V4 VeOmni recipe on AMD ROCm out of the box verl-project/verl
examples/grpo_trainer/run_deepseek_v4_veomni.sh now runs unchanged on AMD gfx942 and gfx950 clusters with no exports and no overrides, because FP8 vLLM rollouts on ROCm had been defaulting to DeepGEMM backends that have no ROCm kernels. The training parameters themselves are untouched.
Correct the macOS --no_auto_dataset_cache guidance in AGENTS.md (#1907) allenai/open-instruct
Elsewhere, AGENTS.md no longer tells every mason.py launch script to pass --no_auto_dataset_cache on macOS, since the import that caused the failure is gone; the flag still applies to two entrypoints, and the Codex repo-skill wiring and judge CUDA graph work round out a quiet day.
Action items
- → Bump beaker-py to 2.6.0 before passing --min_runtime or --auto_resume to mason; older releases cannot send the minRun... allenai/open-instruct [plan]
