increase max_num_seqs

binary-husky · binary-husky · commit 32de3833bdb7 · 2026-04-10T11:34:10.000+08:00
diff --git a/tests/bench/benchmark_appworld/benchmark_appworld.yaml b/tests/bench/benchmark_appworld/benchmark_appworld.yaml
@@ -28,7 +28,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096
diff --git a/tests/bench/benchmark_appworld/benchmark_appworld_oai_sdk.yaml b/tests/bench/benchmark_appworld/benchmark_appworld_oai_sdk.yaml
@@ -27,7 +27,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096
diff --git a/tests/bench/benchmark_appworldlora/benchmark_appworldlora.yaml b/tests/bench/benchmark_appworldlora/benchmark_appworldlora.yaml
@@ -28,7 +28,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096
diff --git a/tests/bench/benchmark_frozenlake/benchmark_frozenlake.yaml b/tests/bench/benchmark_frozenlake/benchmark_frozenlake.yaml
@@ -34,7 +34,7 @@ ajet:
 
     mode: async
 
-    max_num_seqs: 40
+    max_num_seqs: 64
 
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist: []
diff --git a/tests/bench/benchmark_frozenlakelora/benchmark_frozenlakelora.yaml b/tests/bench/benchmark_frozenlakelora/benchmark_frozenlakelora.yaml
@@ -34,7 +34,7 @@ ajet:
 
     mode: async
 
-    max_num_seqs: 40
+    max_num_seqs: 64
 
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist: []
diff --git a/tests/bench/benchmark_learn2ask/benchmark_learn2ask.yaml b/tests/bench/benchmark_learn2ask/benchmark_learn2ask.yaml
@@ -19,7 +19,7 @@ ajet:
     max_env_worker: 64
     num_repeat: 6
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     multi_turn:
       max_sample_per_task: 2
 
diff --git a/tests/bench/benchmark_learn2asklora/benchmark_learn2asklora.yaml b/tests/bench/benchmark_learn2asklora/benchmark_learn2asklora.yaml
@@ -19,7 +19,7 @@ ajet:
     max_env_worker: 64
     num_repeat: 6
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     multi_turn:
       max_sample_per_task: 2
 
diff --git a/tutorial/example_appworld/appworld.yaml b/tutorial/example_appworld/appworld.yaml
@@ -28,7 +28,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096
diff --git a/tutorial/example_appworld/appworld_oai_sdk.yaml b/tutorial/example_appworld/appworld_oai_sdk.yaml
@@ -27,7 +27,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096
diff --git a/tutorial/example_frozenlake/frozenlake_easy.yaml b/tutorial/example_frozenlake/frozenlake_easy.yaml
@@ -33,7 +33,7 @@ ajet:
 
     mode: async
 
-    max_num_seqs: 40
+    max_num_seqs: 64
 
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist: []
diff --git a/tutorial/example_frozenlake/frozenlake_hard.yaml b/tutorial/example_frozenlake/frozenlake_hard.yaml
@@ -33,7 +33,7 @@ ajet:
 
     mode: async
 
-    max_num_seqs: 40
+    max_num_seqs: 64
 
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist: []
diff --git a/tutorial/example_learn2ask/learn2ask.yaml b/tutorial/example_learn2ask/learn2ask.yaml
@@ -19,7 +19,7 @@ ajet:
     max_env_worker: 64
     num_repeat: 6
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     multi_turn:
       max_sample_per_task: 2
 
diff --git a/tutorial/example_math_agent/math_agent.yaml b/tutorial/example_math_agent/math_agent.yaml
@@ -29,7 +29,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: 0.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     multi_turn:
       max_sample_per_task: 2
     compute_madness_checklist:
diff --git a/tutorial/example_math_lora/math_agent.yaml b/tutorial/example_math_lora/math_agent.yaml
@@ -29,7 +29,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: 0.0
     tensor_model_parallel_size: 1
-    max_num_seqs: 40
+    max_num_seqs: 64
     multi_turn:
       max_sample_per_task: 2
     compute_madness_checklist:
diff --git a/tutorial/example_werewolves/werewolves.yaml b/tutorial/example_werewolves/werewolves.yaml
@@ -30,7 +30,7 @@ ajet:
     # the number of vllm engines, number of gpus for infer is `n_vllm_engine*tensor_model_parallel_size`, this argument is NOT effective when NOT using trinity
     n_vllm_engine: 8
 
-    max_num_seqs: 40
+    max_num_seqs: 64
 
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist:
diff --git a/tutorial/example_werewolves_swarm/werewolves.yaml b/tutorial/example_werewolves_swarm/werewolves.yaml
@@ -15,7 +15,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: 0.0
     tensor_model_parallel_size: 1
-    # max_num_seqs: 40
+    # max_num_seqs: 64
     # monitor LLM's abormal behaviors during rollout
     compute_madness_checklist:
       - "nonsense"
