$ cat alignment/week/2026-10-05.log
the week in review · Oct 5 – Oct 11, 2026
verl adds TPU SFT via TorchTitan and ETP checkpoint conversion
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Both arrive behind the same push that moved the profiler and attention padding paths off CUDA-only assumptions.
[doc, hardware, worker, trainer] feat: add SFT support for TPU using TorchTitan engine (#8098) verl-project/verl
Supervised fine-tuning now runs on TPU through the TorchTitan engine, and the TPU-specific runtime workarounds, `ReduceOp.AVG`, `all_gather_object` on `tpu_dist` and the FSDP2 handling, live only in that engine rather than in verl core. If you operate on TPU, this is the first path to SFT without maintaining a fork of the core loops.
[megatron] feat: support TP and ETP in Bridge checkpoint conversion (#8166) verl-project/verl
Bridge checkpoint conversion now carries tensor parallelism and expert tensor parallelism, which is what stands between a Megatron TP/ETP checkpoint and loading it under a different parallel layout. Worth knowing if your checkpoints are split across ranks and you have been converting with degenerate layouts to work around it.
[hardware, perf, training_utils] feat: add a profiler_markers() platform hook and share the attention padding helpers (#7917) verl-project/verl
`PlatformBase` gains a `profiler_markers()` hook, so a non-CUDA backend can emit its own profiling ranges, and the attention padding helpers are now shared instead of living in CUDA-shaped code. Custom backends no longer need to intercept or patch the attention path to keep padding correct.
[data] fix: correct enable_thinking_default YAML value and loss mask mismatch (#8086) verl-project/verl
Two bugs in `MultiTurnSFTDataset` made tokenization and loss masking disagree whenever the chat template produced a different token count for thinking ON versus OFF: a wrong YAML default for `enable_thinking_default` and a loss mask mismatch. If you fine-tune on multi-turn data with a thinking mode, samples in that configuration were being trained against misaligned masks.
[megatron,worker] fix: offload actor parameters after disaggregated weight sync (#7631) verl-project/verl
In disaggregated mode with a non-naive checkpoint backend, Megatron actor parameters stayed resident on GPU after `update_weights()`. This moves them back to CPU, which is the difference between holding your memory budget and creeping into OOM over a long run.
[ckpt] feat: add veomni model merger verl-project/verl
A veomni model merger takes a DCP checkpoint and produces merged weights, so a training run you would otherwise have to re-export by hand is now a supported path. Relevant if your pipeline ends in veomni format and your downstream tooling expects a single set of weights.
[vllm, hardware] fix: vLLM-Ascend 0.23 sched_yield on ARM hosts and FIA causal mask for genuine decode (#8144) verl-project/verl
vLLM-Ascend 0.23 compatibility fixes land on `main`: `sched_yield` on ARM hosts and the FIA causal mask for genuine decode, adapted to the refactored `apply_npu_vllm_patches()` structure. Operators on that stack should check they are not carrying the patches locally.
[doc, ci] feat: add Ascend tutorial doc translation pipeline verl-project/verl
Alongside: a weekly workflow translating `docs/ascend_tutorial/zh/` into `en/`, the Ascend 950 line restated as 950PR and 950DT, nightly PPO CI moved back to A2 runners with the matching 910B image after host OOM failures on A3, and reduced Ascend workflow triggers, a shared cluster PyPI cache, and an explicit Qwen3.5 model path.
$ 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.