prime-rl
A reinforcement learning training framework developed by Prime Intellect for post-training and RL fine-tuning of large models, with support for training MoE models.
Timeline (last 90 days)
In prime-rl, Qwen3 MoE training step speed improved by up to 34%.
prime-rl is an open-source reinforcement learning training framework from Prime Intellect, aimed at post-training and RL fine-tuning of large models.
The 34% step speed improvement was measured under a specific model (Qwen3 MoE) and a specific framework (prime-rl), and the gains will vary when transferred to other model architectures or parallelization strategies.
After integrating the trainable Mega MoE kernel into prime-rl, both the forward and backward sides of the RL loop use the same fused implementation, avoiding path-switching overhead between sampling and updating.
All Facts (4)
In prime-rl, Qwen3 MoE training step speed improved by up to 34%.
50%Unverifiedprime-rl is an open-source reinforcement learning training framework from Prime Intellect, aimed at post-training and RL fine-tuning of large models.
50%UnverifiedThe 34% step speed improvement was measured under a specific model (Qwen3 MoE) and a specific framework (prime-rl), and the gains will vary when transferred to other model architectures or parallelization strategies.
50%UnverifiedAfter integrating the trainable Mega MoE kernel into prime-rl, both the forward and backward sides of the RL loop use the same fused implementation, avoiding path-switching overhead between sampling and updating.
50%