huggingface/trl v1.16.0
🛠️ TRL now trains on AWS Trainium
ProofCode ↗
Key points
- The fused LM head from v1.15 is what made Neuron tractable; the remaining gaps are closed here.
- Neuron recompiles once per input shape, so an unpadded run stalls about a minute on every new sequence length. padtomultipleof now defaults to "auto": 128 on Neuron, None elsewhere, so nothing changes on GPU.
- Trainium2, Qwen3-0.6B SFT, 20 steps: 130 s with the default against 260 s without.
- No mask tensor: every layer except attention runs once per unique token, and only K/V is duplicated along shared prefixes.
Sources (1)
- [1]huggingface/trl v1.16.0GitHub: huggingface/trl · Oct 15, 06:00 PM
### 🛠️ TRL now trains on AWS Trainium
The fused LM head from v1.15 is what made Neuron tractable; the remaining gaps are closed here.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026REKI: Registering Knowledge in VLAs
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026Welcome FLUX.3 Image Klein in Diffusers
- Oct 15, 2026Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
- Oct 14, 2026NVIDIA/TensorRT-LLM v1.4.0rc0
- Oct 14, 2026mistralai/Timestral-209M