$ the-wire · showcase
TPU GRPO image fixed, Megatron optimizer offload leaves HDO state alone
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl repaired two runtime failures that bite at container build and phase switch time, while open-instruct added an opt-in You.com search backend and documented the chat template its released RL Zero checkpoints actually used.
[hardware] fix: pin tensordict and install orjson in the TPU GRPO image (#8292) verl-project/verl
An image built from examples/tpu/grpo/Dockerfile failed at bare `import verl` with `ModuleNotFoundError: No module named 'orjson'`, because verl/utils/tracking.py imports orjson and tensordict 0.15.2 dropped the requirement verl had been inheriting. The fix pins tensordict and installs orjson explicitly; if you build your own TPU image from an unpinned base, the same import failure is waiting f...
[megatron] fix: leave HybridDeviceOptimizer state untouched in optimizer offload (#8237) verl-project/verl
With Megatron's `optimizer_cpu_offload=True` the phase-switch helper `offload_megatron_optimizer` used to move HybridDeviceOptimizer state to CPU, including the GPU-side state a fused CUDA optimizer owns. It now leaves HDO state untouched, so long runs that switch between training and inference phases stop hitting an optimizer that cannot operate on CPU tensors.
Add You.com search tool (keyed Search API) allenai/open-instruct
`you_search` now registers as an opt-in backend alongside Serper, calling `POST https://ydc-index.io/v1/search` with `X-API-Key` set from `YDC_API_KEY` and reusing Serper's `query` schema and title/snippet/source formatting, per the keyed-only scope in the linked issue. Missing credentials fail at construction; no existing tool defaults or dependencies change.
Document which chat template the released RL Zero checkpoints used (#1929) allenai/open-instruct
The RL Zero Code and IF scripts on main select the simple rlzero templates, but the released checkpoints were trained with `olmo_thinker` at d928a7c and ship that template on the Hub. The olmo3 README and both scripts now say so, so reproduction runs match the published weights.
Make tracked files satisfy the trailing-whitespace and end-of-file hooks (#1909) allenai/open-instruct
Also landed: trailing-whitespace and end-of-file hook conformance across tracked files, and a requirements.txt export regenerated so `datasets==4.8.5` matches uv.lock and the project's `datasets>=4.8.4,<5` constraint, which only matters if you install from the exported file.
