Hugging Face TRL 1.16.0 brings training to Trainium and speeds up shared rollouts

Hugging Face’s TRL 1.16.0 adds AWS Trainium support and tree-based packing for asynchronous GRPO training, with the project reporting substantial speed and memory gains in its tests. It also adds support for AMD GPUs in asynchronous trainers, widening the hardware choices for AI training teams.

Hugging Face Watch analysis

What happened

The TRL 1.16.0 release notes describe changes across hardware support, training efficiency and developer workflows. On Trainium2, the project says a 20-step Qwen3-0.6B supervised fine-tuning run took 130 seconds with the default padding setting, compared with 260 seconds without it. The release also adds a guide for training on Trainium.

For AsyncGRPO, tree packing processes shared prompt tokens once rather than forwarding them separately for each rollout. In one Qwen3-1.7B test, the release reports that a shared-prompt workload with a 3.37 packing ratio took 2,755 milliseconds instead of 8,179, while peak memory fell from 35.2 GiB to 13.6 GiB. Results vary with the packing pattern: without shared tokens, the notes report roughly a one per cent slowdown.

Our top picks

• AWS Trainium training
  TRL now supports training on Trainium; a Trainium2 test run took 130 seconds rather than 260 with the default padding setting.
• Tree packing for AsyncGRPO
  Shared prompt tokens are processed once, with the release reporting up to 3.10 times faster execution in a tested nested-fork workload.
• Lower memory in shared-rollout tests
  One reported Qwen3-1.7B configuration fell from 35.2 GiB to 13.6 GiB of peak memory.
• AMD support for async trainers
  Async GRPO and async distillation trainers now support AMD GPUs.
• Tool arguments typed from their schema
  Tool-call arguments can be parsed according to the tool’s JSON schema, rather than relying on the raw parser output.
• Support for Muse Spark Open
  The release adds model and test coverage, plus a LoRA SFT example with a specified multi-node setup and dependencies.

Why it matters

Training efficiency is not an abstract leaderboard sport when it determines how much work fits on available hardware. Trainium support gives teams another route beyond GPUs, while tree packing targets repeated prompt computation in multi-turn and multi-generation training. The reported gains are tied to particular hardware, models and workloads, so they are useful evidence, not a universal speed guarantee.

Our read

This is a substantial toolkit release, with consequential additions for teams training on Trainium, AMD or shared-prompt rollouts. The tree-packing results are the standout: less repeated work, less memory, and no packing switch to remember. Check the new dependency requirements and removals before upgrading; vLLM 0.21.0 support and the experimental SSD trainer are among the things dropped.

What to watch

• Whether teams reproduce the Trainium and tree-packing results on their own workloads.
• How the new AMD async-trainer support performs across different setups.
• Whether the dependency changes or removed features require migration work.

Discussion spark: Should AI training teams prioritise workload-specific efficiency gains like tree packing, or favour broader hardware and software compatibility even when it means leaving some performance on the table?

not affiliated with or endorsed by Hugging Face