(EnvironmentWorker(train_env-0) pid=746905) 2026-03-19T17:41:53.502+08:00 INFO:rock_agent.py:382 [rock.sdk.sandbox.agent.rock_agent] [] [] -- [240498b70e464981a6cfc7165776dc31] Start to Initializing ModelService
(EnvironmentWorker(train_env-0) pid=746905) 2026-03-19T17:41:53.503+08:00 ERROR:rock_agent.py:224 [rock.sdk.sandbox.agent.rock_agent] [] [] -- [240498b70e464981a6cfc7165776dc31] Agent initialization failed - 'RuntimeEnvConfig' object has no attribute 'pip' (elapsed: 0.50s),

defaults:
  - ../config/traj_envs@_here_
  - ../config/deepspeed_zero@_here_
  - ../config/deepspeed_zero2@_here_
  - ../config/deepspeed_zero3@_here_
  - ../config/deepspeed_zero3_cpuoffload@_here_

hydra:
  run:
    dir: .
  output_subdir: null

exp_name: "agentic_rollout_swe"
seed: 42

logging_dir: ./output/logs
output_dir: ./output
model_name: ${exp_name}-${now:%Y%m%d_%H%M%S}
rollout_dump_dir: ./output/rollout_dump
system_envs:
  USE_MODELSCOPE: '1'


num_gpus_per_node: 8
rpc_timeout: 72000

max_steps: 10
save_steps: 10
logging_steps: 1
eval_steps: 2
resume_from_checkpoint: false

rollout_batch_size: 1
val_batch_size: 1
sequence_length: 65536

max_tokens_per_step: 4096

advantage_clip: 0.2
ppo_epochs: 1
adv_estimator: "step_reinforce"
batch_adjust_mode: "random_sample"
step_reward_gamma: 1.0

#pg_clip: 0.1
#dual_clip_loss: True
init_kl_coef: 0.0
whiten_advantages: true
entropy_loss_coef: 0
max_grad_norm: 1.0


pretrain: /var/model/Qwen2.5-7B-Instruct
reward_pretrain: /var/model/Qwen2.5-7B-Instruct

actor_train:
  model_args:
    flash_attn: fa2
    disable_gradient_checkpointing: false
    dtype: bf16
    model_type: ~
actor_infer:
  model_args:
    flash_attn: fa2
    disable_gradient_checkpointing: true
    dtype: bf16
  generating_args:
    max_new_tokens: ${max_tokens_per_step} # single-turn response length
    top_p: 1.0
    top_k: 50
    num_beams: 1
    temperature: 1.0
    num_return_sequences: 1
    stop_strings: ["</tool_call>","</tool_call>\n","\n</tool_call>\n","\n</function>"]
    include_stop_str_in_output: true
  data_args:
    template: qwen3_coder
  strategy_args:
    strategy_name: vllm
    strategy_config:
      gpu_memory_utilization: 0.8
      block_size: 16
      load_format: auto
      tensor_parallel_size: 1
  device_mapping: list(range(1,2))

reward_normalization:
  grouping: traj_group_id # tags(env_type)/traj_group_id(group)/batch(rollout_batch)... group_by reward/adv
  method: mean
  # norm_mean_type: batch
  # norm_std_type: group

train_env_manager:
  max_env_num_per_worker: 1
  num_env_groups: 1
  # under the same group, the env config and env seed are ensured to be equal
  group_size: 1
  tags: [swebench_native_verified]
  num_groups_partition: [1] # If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
  system_envs:
    # if you cannot get python env in rock due to connetion error, try to use this, may expire in the future
    ROCK_RTENV_PYTHON_V31114_INSTALL_CMD: '[ -f cpython31115.tar.gz ] && rm cpython31115.tar.gz; [ -d python ] && rm -rf python; wget -q -O cpython31115.tar.gz https://mirror.nju.edu.cn/github-release/astral-sh/python-build-standalone/20260303/cpython-3.11.15+20260303-x86_64-unknown-linux-gnu-install_only.tar.gz && tar -xzf cpython31115.tar.gz && mv python runtime-env'
val_env_manager:
  max_env_num_per_worker: 1
  num_env_groups: 1
  group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
  tags: [swebench_native_verified]
  num_groups_partition: [1] # TODO: If not set, all env names divide nums equally. Under the same group, the env config and env seed (prompt) are equal in each generation
  system_envs:
    # if you cannot get python env in rock due to connetion error, try to use this, may expire in the future
    ROCK_RTENV_PYTHON_V31114_INSTALL_CMD: '[ -f cpython31115.tar.gz ] && rm cpython31115.tar.gz; [ -d python ] && rm -rf python; wget -q -O cpython31115.tar.gz https://mirror.nju.edu.cn/github-release/astral-sh/python-build-standalone/20260303/cpython-3.11.15+20260303-x86_64-unknown-linux-gnu-install_only.tar.gz && tar -xzf cpython31115.tar.gz && mv python runtime-env'

max_actions_per_traj: 60
env_manager_cls: roll.pipeline.agentic.env_manager.agent_native_env_manager.AgentNativeStepEnvManager

agent_config_common:
  agent_type: swe-agent
  model_service_config: 
    type: "local"
    enabled: True
    port: 8081
  run_cmd: >
    sweagent run
    --problem_statement.type text
    --problem_statement.text <<PROMPT>>
    --env.repo.type local
    --env.repo.path "$(pwd)"
    --env.deployment.type local
    --agent.tools.execution_timeout 1000
    --agent.model.per_instance_cost_limit 0
    --agent.model.total_cost_limit 0
    --agent.model.name kimi/kimi-k2.5
    --agent.model.api_base "http://localhost:8081/v1"
    --agent.model.api_key sk-xxxx
  # replace <MODEL_NAME> <BASE_URL> <API_KEY>; keep others as-is
  # if repo at workspace root, change env.repo's config like: --env.repo.type preexisting --env.repo.repo_name testbed

  # Python Agent often meets python conflicts, so we need to avoid add RuntimeEnv to PATH in agent_run
  skip_wrap_run_cmd: true

  runtime_env_config:
    type: python
    version: "3.12"
    custom_install_cmd: git clone https://github.com/SWE-agent/SWE-agent.git && cd SWE-agent && pip install -e .

  project_path: "/tmp/testbed"

  # SWE-agent requires an existing repository, so we need to initialize one.
  post_init_cmds:
    - command: "mkdir -p /tmp/testbed && cd /tmp/testbed && git init && echo init >README.md && git add README.md && git -c user.email=test@example.com -c user.name='Test User' commit -m 'Initial commit'"
      timeout_seconds: 30

custom_envs:
  swebench_native_verified:
    env_type: "rock_tb_native_env"
    max_steps: ${max_actions_per_traj}
    max_tokens_per_step: ${max_tokens_per_step}
    env_manager_cls: ${env_manager_cls}
    agent_system_template: "agent_system_template placeholder"
    agent_template: "agent_template placeholder"
    env_config:
      dataset_name: /workspace/ROLL/data/swe_bench_verified_example.jsonl
      tools: ~
      max_steps: ${max_actions_per_traj}
      mode: "val"
      sandbox_base_url: http://rock:8080 # change to your own service address if needed
      user_id: "xxx"
      experiment_id: "test_tb_native"
      test_files: ["/terminal-bench-datasets/datasets/swebench-verified"]
      agent_config: ${agent_config_common}

def __init__(self, sandbox: Sandbox, runtime_env_config: PythonRuntimeEnvConfig) -> None:
    # 确保是 PythonRuntimeEnvConfig 类型
    if not isinstance(runtime_env_config, PythonRuntimeEnvConfig):
        runtime_env_config = PythonRuntimeEnvConfig.model_validate(runtime_env_config.model_dump())
    ...