A packing fix for reinforcement-learning workloads
The included GRPO change, described in pull request 18685, deals with verification that compares packed sequences with separate rows. Its authors explain that small numerical changes in mixture-of-experts routing can create sparse large differences, causing a batch to fall back even without a document-boundary error.
The change adds a bounded MoE noise check and installs variable-length attention for packed GRPO rows. The description says cross-document leakage checks remain active. It is not an instruction to remove correctness tests: dense models retain the earlier rule, and failed checks still trigger fallback.
Benchmark scope matters more than a headline multiplier
The change description reports median step times of 10.099 to 9.410 seconds for Qwen3-30B-A3B on an A100-40GB, and 7.511 to 5.890 seconds for Qwen3-1.7B on an L4. These are author-run, short GRPO experiments with four-bit models and rank-16 LoRA, not independent results for every training recipe.
The same description reports that an optional MoE PrefixGrouper configuration was slower than plain packing in that test. Faster attention therefore should not be read as a guaranteed improvement in every configuration or in final model quality. We have not reproduced these measurements.
Stable tool lists are a separate inference change
The release notes say a chat keeps its MCP tool list fixed so that subsequent messages process only new material. The configured tool-output length also applies to MCP results across llama.cpp, vLLM, Ollama and custom connections. This concerns the running conversation, rather than an update to model weights.
The practical implication is to check both retained context and tool-result completeness. Our proposed evaluation records repeated prompt processing, truncation, tool failures and the correctness of the final answer. A shorter processed prompt alone does not establish that an agent finished its job.
Compatibility and release boundaries
The notes also list PEFT 0.22 compatibility and a fix for mixed-length Gemma-4 LoRA batches with older PyTorch versions. Those are specific compatibility statements, not evidence that every combination of training packages works. Preserve the exact environment when comparing results.
This article groups these changes under one beta-release event. Inkling Mini integration is not presented as a second model launch, and the release’s separate image and video claims are not transferred to language-model training. We inspected public documentation and repository changes without installing or executing the released software.

Discussion.
Your perspective matters. Sign in, pick a nickname and choose your cyber identity.
Loading discussion…
Your account name, email and photo stay off your comments.