7-szer hosszabb szöveggel is taníthatók AI-modellek az új TRL-lel

TRL v1.15 is out, and it’s an absolute banger of a release for memory-efficient post-training. The main change: SFT, DPO, KTO, GRPO, RLOO and Distillation now use a fused LM head by default.Instead of materializing the huge [batch, seq, vocab] logits tensor, a Triton kernel computes the token-level quantities we actually need directly. The results are quite significant. On Gemma 3-1B with a 262k vocabulary, on the same GPU: DPO: 10k → 59k max sequence length KTO: 9k → 63k GRPO: 28k → 114k RLOO: 23k → 100k SFT: 20k → 107k Up to 6.9x longer sequences, with peak memory at 8k context reduced by 52-82%. Speed isn't negatively impacted: training is also up to ~11% faster. There’s more in the release too: selective activation checkpointing for SFT, assistant-only loss for vision datasets, better conversation logging, improvements to AsyncGRPO / AsyncDistillation, and a long list of fixes. I really like optimizations like this: the training API doesn’t need to become more complicated as the implementation underneath gets much better.