$ the-wire · showcase
verl lands 397B GRPO on TPU v7x and a DeepSeek-V4-Pro recipe
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl shipped two recipes that put frontier-scale RL within reach of existing TPU and H100/GB200 clusters, and open-instruct flipped its olmo-core SFT document-boundary flags on by default.
[hardware, recipe] feat: Qwen3.5-397B-A17B full-parameter GRPO recipe on TPU v7x (8K context, multi-replica rollout) verl-project/verl
Full-parameter Qwen3.5-397B-A17B MoE GRPO (8K thinking context, 512 + 7,680) now runs on TPU v7x via examples/tpu/grpo/run_qwen3_5_torchtitan.sh on top of verl-hardware-plugin 0bc8bbac. The PR states plainly that this enables 397B RL on TPU and that performance still needs optimization, so treat it as a working starting point rather than a tuned one.
[megatron, vllm, doc] feat: DeepSeek-V4-Pro GRPO example with H100 and GB200 profiles verl-project/verl
new example plus a "Running DeepSeek-V4-Pro at scale" doc section covers the 61-layer, 384-routed-expert, ~1.6T-parameter Pro model with H100 and GB200 profiles; the 8-node Flash example only covers the 40-layer model. The authors note the first run otherwise costs a lot of node-hours in trial and error, which is exactly the cost this example removes.
olmo-core SFT: document boundaries from row metadata by default allenai/open-instruct
In olmo-core SFT, --row_aligned_parts and --document_boundaries_from_metadata now default to true, so document boundaries come from per-row metadata without any flag in your launch command. If your data pipeline expects the old packing behavior, opt out explicitly with --no_document_boundaries_from_metadata.
Integrate bounded structured-output dataset rewards allenai/open-instruct
MILES gains opt-in verifiers for JSON, YAML, TOML, XML and CSV outputs, with explicit schema and coercion rules and bounded validation that rejects ambiguous parses and unsafe references. Existing JSONSchemaVerifier behavior and existing configurations are unchanged, so nothing moves until you enable the new formats.
Limit Beaker CI to the merge queue and preserve CPU regression coverage allenai/open-instruct
The long tail: open-instruct moved Beaker GPU tests out of the PR-push path into the merge queue, adding structured-output (JSON, YAML, TOML, XML, CSV) and instruction/calendar reward adapters for MILES along the way, and verl's Ascend docs now require Python >=3.12 for SGLang and no longer double-register Qwen3-ASR.
Action items
- → Pass --no_document_boundaries_from_metadata to olmo-core SFT jobs that depend on the old packing boundaries allenai/open-instruct [plan]
