Update scripts for engine workers.

Shangwei-Li · Shangwei-Li · commit 41b786c5757a · 2026-02-25T23:57:38.000+08:00
diff --git a/verl/experimental/fully_async_policy/shell/dapo_30b_a3b_base_math_fsdp.sh b/verl/experimental/fully_async_policy/shell/dapo_30b_a3b_base_math_fsdp.sh
@@ -103,7 +103,7 @@ ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     data.val_files="${TEST_FILE}" \
     data.prompt_key=prompt \
     data.truncation='left' \
-    actor_rollout_ref.actor.strategy=fsdp \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp \
     critic.strategy=fsdp \
     data.max_prompt_length=${max_prompt_length} \
     data.max_response_length=${max_response_length} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_async_retool.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_async_retool.sh
@@ -98,7 +98,7 @@ python3 -m verl.experimental.fully_async_policy.fully_async_main \
     actor_rollout_ref.actor.use_dynamic_bsz=True \
     actor_rollout_ref.actor.ppo_mini_batch_size=$ppo_mini_batch_size \
     actor_rollout_ref.actor.ppo_max_token_len_per_gpu=$actor_max_token_len_per_gpu \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.fsdp_config.fsdp_size=${fsdp_size} \
     actor_rollout_ref.actor.ulysses_sequence_parallel_size=$train_sp \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_16_16.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_16_16.sh
@@ -90,7 +90,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_32_32.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_32_32.sh
@@ -90,7 +90,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_12.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_12.sh
@@ -92,7 +92,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh
@@ -92,7 +92,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_64_64.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_64_64.sh
@@ -90,7 +90,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_64_64_mis.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_64_64_mis.sh
@@ -96,7 +96,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_8_8.sh b/verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_8_8.sh
@@ -90,7 +90,7 @@ python -m verl.experimental.fully_async_policy.fully_async_main \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_4_12.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_4_12.sh
@@ -73,7 +73,7 @@ python3 -m verl.experimental.one_step_off_policy.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_64_64.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_64_64.sh
@@ -75,7 +75,7 @@ python -m verl.experimental.one_step_off_policy.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_64_64_ris.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_64_64_ris.sh
@@ -85,7 +85,7 @@ python -m verl.experimental.one_step_off_policy.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_colocate.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_colocate.sh
@@ -68,7 +68,7 @@ python3 -m verl.trainer.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_sglang_4_12.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_sglang_4_12.sh
@@ -73,7 +73,7 @@ python3 -m verl.experimental.one_step_off_policy.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_sglang_colocate.sh b/verl/experimental/one_step_off_policy/shell/dapo_7b_math_fsdp2_sglang_colocate.sh
@@ -68,7 +68,7 @@ python3 -m verl.trainer.main_ppo \
     algorithm.adv_estimator=${adv_estimator} \
     algorithm.use_kl_in_reward=${use_kl_in_reward} \
     algorithm.kl_ctrl.kl_coef=${kl_coef} \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
     actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
diff --git a/verl/experimental/one_step_off_policy/shell/grpo_0.6b_gsm8k_fsdp2_2_6.sh b/verl/experimental/one_step_off_policy/shell/grpo_0.6b_gsm8k_fsdp2_2_6.sh
@@ -26,7 +26,7 @@ python3 -m verl.experimental.one_step_off_policy.main_ppo \
     data.max_response_length=1024 \
     data.filter_overlong_prompts=True \
     data.truncation='error' \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.actor.optim.lr=1e-6 \
diff --git a/verl/experimental/one_step_off_policy/shell/grpo_0.6b_gsm8k_fsdp2_sglang_2_6.sh b/verl/experimental/one_step_off_policy/shell/grpo_0.6b_gsm8k_fsdp2_sglang_2_6.sh
@@ -26,7 +26,7 @@ python3 -m verl.experimental.one_step_off_policy.main_ppo \
     data.max_response_length=1024 \
     data.filter_overlong_prompts=True \
     data.truncation='error' \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.actor.optim.lr=1e-6 \
diff --git a/verl/experimental/one_step_off_policy/shell/grpo_3b_gsm8k_fsdp2_2_6.sh b/verl/experimental/one_step_off_policy/shell/grpo_3b_gsm8k_fsdp2_2_6.sh
@@ -25,7 +25,7 @@ python3 -m verl.experimental.one_step_off_policy.main_ppo \
     data.max_response_length=1024 \
     data.filter_overlong_prompts=True \
     data.truncation='error' \
-    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.fsdp_config.strategy=fsdp2 \
     critic.strategy=fsdp2 \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.actor.optim.lr=1e-6 \
