X · @huggingface
· Infrastructure
RT Sergio Paniego: continuous batching just landed in TRL for GRPO at 64 generations it runs faster and uses less VRAM than plain generate, no vLLM ne…
RT Sergio Paniegocontinuous batching just landed in TRL for GRPOat 64 generations it runs faster and uses less VRAM than plain generate, no vLLM neededhow it works and when to reach for it, belowSergio Paniego: http://x.com/i/article/2067886927936671744