$ cat alignment/week/2026-10-26.log
the week in review · Oct 26 – Nov 1, 2026
verl pins DeepEP for uv Megatron, Qwen3.5 TPU recipe
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
MILES operators got a long list of default changes this week: no startup eval on resume, native SGLang evals, bounded producer budgets.
[megatron, docker] fix: install DeepEP in uv training environments and restore NVSHMEM verl-project/verl
The uv `megatron` extra now installs `deep-ep==1.2.1` from the matching Torch index, so MoE training that enables the DeepEP dispatcher stops failing on a missing package; the CUDA 13 image also gets the system NVSHMEM at the path the prebuilt wheel expects. If you build Megatron runs from the uv extra or that image, a rebuild picks it up.
[rollout, vllm] fix: do not pass data_parallel_start_rank to the DP leader (vLLM >= 0.28 hybrid-LB) (#8234) verl-project/verl
Multi-node vLLM rollout replicas where `data_parallel_size > data_parallel_size_local` failed at engine start on vLLM 0.28 and later with the hybrid-LB behavior; the DP leader no longer receives `data_parallel_start_rank`. Operators pinned to older vLLM keep the previous path.
[hardware, recipe] feat: Qwen3.5 (incl. 35B-A3B MoE) GRPO recipe for TPU v7x, plus context parallelism and QKV fusion configs verl-project/verl
New `examples/tpu/grpo/run_qwen3_5_torchtitan.sh` defaults to the `Qwen/Qwen3.5-35B-A3B` MoE with thinking mode on, and the TPU GRPO Dockerfile moves to vLLM v0.32.0 with `--xla_tpu_bf16_emission_mode=PROMOTE_F32` set across the configs. Context parallelism and QKV fusion configs land alongside it.
[rollout, vllm] fix: release the submission-gate barrier when a request lands in the engine, not on its first token verl-project/verl
`generate()` used to decrement `_admitting` only on a request's first output, so a weight sync could sit waiting on the submission-gate barrier for a request already handed to the engine. The barrier now releases when the request lands in the engine, which shortens the stall between rollout and weight sync.
Skip MILES startup evaluation on nonzero-step resumes allenai/open-instruct
A MILES run with `skip_eval_before_train=false` restored from update 60 blocked further training on a fresh startup evaluation. Checkpoints past step zero now start rollout without one; expect a shorter resume path on long runs.
Use native SGLang for MILES background evaluations allenai/open-instruct
Background policy evaluations now use olmo-eval's native `sglang` provider instead of a shared engine, with `node.toml` running a separate one-GPU GSM8K, GSM8K-Platinum and IFBench panel every 50 optimizer updates, plus a final pass. Budget the extra GPU on the eval node.
Bound MILES producer budgets and expose opt-in judge CUDA graphs allenai/open-instruct
Unfinished rollout work is now bounded independently of reserved serving capacity: the one-node example uses 1,024 samples instead of 1,536, and three-node uses 1,536 instead of the 19,456 the omitted-setting heuristic produced, with per-engine admission still at 512. Opt-in judge CUDA graphs are exposed for those who want them.
Pin the compiled MILES base and layer runtime sources allenai/open-instruct
The compiled dependency base with FA2, FA3 and FA4 is pinned so application builds fetch locked sources instead of recompiling attention kernels; the earlier 8.06 GB install layer shrank to 2.77 MB as a result. Alongside: the CUDA 13 image fixes, the tensordict 0.11 `concat_tensordict` fix that keeps the first NonTensorData, and a gated-GPQA guard requiring HF_TOKEN where the judge guard only c...
$ ls alignment/week/ # the briefings behind this review
Keep up with Alignment & Post-training 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.