From 231de4c29f9403628e37f7cb2e2d70c0b3b189fb Mon Sep 17 00:00:00 2001 From: gaoziyuan Date: Thu, 20 Mar 2025 16:49:08 +0800 Subject: [PATCH] fix reference also use dp_actor bug --- verl/trainer/config/ppo_megatron_trainer.yaml | 1 + verl/trainer/config/ppo_trainer.yaml | 1 + 2 files changed, 2 insertions(+) diff --git a/verl/trainer/config/ppo_megatron_trainer.yaml b/verl/trainer/config/ppo_megatron_trainer.yaml index a9f9134b5ea..c51c814f06c 100644 --- a/verl/trainer/config/ppo_megatron_trainer.yaml +++ b/verl/trainer/config/ppo_megatron_trainer.yaml @@ -61,6 +61,7 @@ actor_rollout_ref: param_offload: False log_prob_micro_batch_size: null # will be deprecated, use log_prob_micro_batch_size_per_gpu log_prob_micro_batch_size_per_gpu: null + use_fused_kernels: True rollout: name: vllm temperature: 1.0 diff --git a/verl/trainer/config/ppo_trainer.yaml b/verl/trainer/config/ppo_trainer.yaml index 09e5e702f60..1765a629a1c 100644 --- a/verl/trainer/config/ppo_trainer.yaml +++ b/verl/trainer/config/ppo_trainer.yaml @@ -65,6 +65,7 @@ actor_rollout_ref: log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz} log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu} ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size} # sp size + use_fused_kernels: True rollout: name: vllm temperature: 1.0