increase max_num_seqs

Author: binary-husky

ajet/default_config/ajet_default.yaml

@@ -63,7 +63,7 @@ ajet:
     n_vllm_engine: 1
 
     # how many sequences are allowed to be processed in parallel by each vllm engine
-    max_num_seqs: 10
+    max_num_seqs: 64
 
     # the usage of infer engine, options: (vllm, sglang)
     name: vllm

ajet/default_config/verl/verl_default.yaml

@@ -270,7 +270,7 @@ actor_rollout_ref:
     pipeline_model_parallel_size: 1
     max_num_batched_tokens: 8192
     max_model_len: null
-    max_num_seqs: 10
+    max_num_seqs: 64
     enable_chunked_prefill: false
     enable_prefix_caching: false
     logprobs_mode: processed_logprobs

tests/bench/benchmark_appworld/benchmark_appworld_2nodes.yaml

@@ -28,7 +28,7 @@ ajet:
     num_repeat: 6
     agent_madness_reward: -1.0
     tensor_model_parallel_size: 4
-    max_num_seqs: 10
+    max_num_seqs: 64
     compute_madness_checklist:
       - "nonsense"
     max_response_length_in_one_turn: 4096

tests/bench/benchmark_countdown/benchmark_countdown.yaml

@@ -59,7 +59,7 @@ ajet:
     n_vllm_engine: 2
 
     # how many sequences are allowed to be processed in parallel by each vllm engine
-    max_num_seqs: 10
+    max_num_seqs: 64
 
     # the usage of infer engine, options: (vllm, sglang)
     name: vllm

tests/bench/benchmark_countdownlora/benchmark_countdownlora.yaml

@@ -59,7 +59,7 @@ ajet:
     n_vllm_engine: 2
 
     # how many sequences are allowed to be processed in parallel by each vllm engine
-    max_num_seqs: 10
+    max_num_seqs: 64
 
     # the usage of infer engine, options: (vllm, sglang)
     name: vllm

tests/bench/benchmark_math/benchmark_math.yaml

@@ -21,7 +21,7 @@ ajet:
     user_workflow: "tutorial.example_math_agent.math_agent->ExampleMathLearn" # ✨✨✨✨ 编写并选择Agent
     temperature: 0.9
     max_env_worker: 64
-    max_num_seqs: 10
+    max_num_seqs: 64
     num_repeat: 6
     agent_madness_reward: 0.0
     tensor_model_parallel_size: 1

tests/bench/benchmark_mathlora/benchmark_mathlora.yaml

@@ -21,7 +21,7 @@ ajet:
     user_workflow: "tutorial.example_math_agent.math_agent->ExampleMathLearn" # ✨✨✨✨ 编写并选择Agent
     temperature: 0.9
     max_env_worker: 64
-    max_num_seqs: 10
+    max_num_seqs: 64
     num_repeat: 6
     agent_madness_reward: 0.0
     tensor_model_parallel_size: 1

tutorial/example_countdown/countdown.yaml

@@ -59,7 +59,7 @@ ajet:
     n_vllm_engine: 2
 
     # how many sequences are allowed to be processed in parallel by each vllm engine
-    max_num_seqs: 10
+    max_num_seqs: 64
 
     # the usage of infer engine, options: (vllm, sglang)
     name: vllm