Delete verl_0720_main/9b_sp1_16k_fused_lmhead_smoke.sh

verl_0720_main/9b_sp1_16k_fused_lmhead_smoke.sh

@@ -1,1111 +0,0 @@
-#!/bin/bash
-# Qwen3.5-9B Nanoclaw 多轮 GRPO — NPU SP1/16K fused LM-head 两步真实训练实验
-#
-# 关键修改：
-#   1) 入口改为 python3 -m verl.trainer.main_ppo，不再使用 recipe.grpo_mindspeed_mm.main_ppo；
-#   2) 不再使用 MM_CONFIG_FILE / MindSpeed-MM YAML；
-#   3) 显式 text-only：data.return_multi_modal_inputs=False；
-#   4) 数据输入改为 Nanoclaw base_tasks 目录，不再使用 Retool parquet/json；
-#   5) 保留 27B 作业验证过的 HCCL buffer 与端口范围，降低 HcclAllreduce socket/资源压力；
-#   6) smoke 默认关闭训练前/训练后验证和 checkpoint，只验证两次完整训练 step；
-#   7) 稳定版固定使用已经完成初始化与 rollout 验证的 FSDP1/offload 形状。
-#   8) 使用最新版 VERL V1/TransferQueue 和 nanoclaw_recipe，并按参考 YAML 的训练意图映射优化器与 KL；
-#   9) actor/ref 的全词表 entropy 使用 256-token 分块，避免 16K SP1 old-log-prob 阶段产生十几 GiB 临时张量。
-#  10) 启用 Qwen3.5 fused Torch LM-head，按 token 分块直接产生 log-prob/entropy，不保留完整 T×vocab logits；
-#  11) 默认仅跑 2 个完整训练 step，覆盖真实 rollout/reward/old-ref-log-prob/actor backward/optimizer/权重同步。
-
-set -x
-
-# ================= 独立稳定版：强制 Qwen3.5 NPU SP1 + 16K =================
-# 不允许外部作业环境把这些核心安全参数覆盖回 SP8/31K。
-export TRAIN_SP=1
-export QWEN35_FLA_BACKEND=disabled
-export MAX_PROMPT_LENGTH=8192
-export MAX_RESPONSE_LENGTH=8192
-export MAX_ASSISTANT_RESPONSE_LENGTH=8192
-export MAX_TOOL_RESPONSE_LENGTH=8192
-export ACTOR_MAX_TOKEN_LEN_PER_GPU=16384
-export LOG_PROB_MAX_TOKEN_LEN_PER_GPU=16384
-export ROLLOUT_MAX_NUM_BATCHED_TOKENS=16384
-export ACTOR_STRATEGY=fsdp
-export OFFLOAD=True
-# 当前 VERL 的 checkpointing 分支不会传递自定义 chunk size；稳定版明确关闭它，确保实际使用 256。
-export ENTROPY_FROM_LOGITS_WITH_CHUNKING=True
-export ENTROPY_FROM_LOGITS_CHUNK_SIZE=256
-export ENTROPY_CHECKPOINTING=False
-# 本实验只改变 LM-head 路径；保留已经走到 old-log-prob 的 FSDP1/SP1/offload 拓扑。
-export USE_FUSED_KERNELS=True
-export FUSED_KERNEL_BACKEND=torch
-export ENABLE_ACTIVATION_OFFLOAD=False
-# 缩小全局实验规模，但每条样本仍允许真实 16K，显存压力不会被短上下文掩盖。
-export TRAIN_BATCH_SIZE=8
-export PPO_MINI_BATCH_SIZE=8
-export N_RESP_PER_PROMPT=8
-export TOTAL_TRAINING_STEPS=2
-export VAL_BEFORE_TRAIN=False
-export TEST_FREQ=-1
-export SAVE_FREQ=-1
-export LOG_VAL_GENERATIONS=0
-
-
-npu-smi info || true
-pip install --upgrade pip
-pip uninstall -y moxing-framework || true
-
-# ================= 路径配置 =================
-SCRIPT_DIR=/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/verl_0720_nanoclaw/verl_0720_main
-if [ -f "${SCRIPT_DIR}/verl/requirements-npu.txt" ]; then
-    DEFAULT_WORK_DIR=${SCRIPT_DIR}/verl
-elif [ -f "${SCRIPT_DIR}/requirements-npu.txt" ]; then
-    DEFAULT_WORK_DIR=${SCRIPT_DIR}
-else
-    DEFAULT_WORK_DIR=${SCRIPT_DIR}/verl
-fi
-WORK_DIR=${WORK_DIR:-${DEFAULT_WORK_DIR}}
-INSTALL_DIR=${INSTALL_DIR:-/home/ma-user}
-BKGS=${BKGS:-/opt/huawei/dataset/zyr_yuyin/bkgs}
-chmod 755 "${INSTALL_DIR}"
-
-# Nanoclaw 自定义包已随 WORK_DIR 提供：nanoclaw_recipe。
-
-GCC_INSTALL_PREFIX=${GCC_INSTALL_PREFIX:-/home/ma-user/gcc-11.3.0}
-COMPILED_GCC_ARCHIVE_PATH=${COMPILED_GCC_ARCHIVE_PATH:-/opt/huawei/dataset/zyr_yuyin/bkgs/gcc-11.3.0-compiled-aarch64.tar.gz}
-
-echo "--> 正在从缓存恢复 GCC 11.3.0..."
-tar -xzf "${COMPILED_GCC_ARCHIVE_PATH}" -C /home/ma-user/
-export PATH=${GCC_INSTALL_PREFIX}/bin:${PATH}
-export LD_LIBRARY_PATH=${GCC_INSTALL_PREFIX}/lib64:${GCC_INSTALL_PREFIX}/lib:${LD_LIBRARY_PATH:-}
-export CC=${GCC_INSTALL_PREFIX}/bin/gcc
-export CXX=${GCC_INSTALL_PREFIX}/bin/g++
-echo "--> 验证 GCC 版本:"
-gcc --version
-
-cd "${BKGS}"
-cp jemalloc-5.3.0.tar.bz2 "${INSTALL_DIR}"
-
-VLLM_LATEST_PKGS=${VLLM_LATEST_PKGS:-/opt/huawei/dataset/zyr_yuyin/lyf/verl-05-12/verl_new_26_05_09/pkgs}
-rm -rf "${INSTALL_DIR}/vllm" "${INSTALL_DIR}/vllm-ascend"
-cp -r "${VLLM_LATEST_PKGS}/vllm" "${INSTALL_DIR}"
-cp -r "${VLLM_LATEST_PKGS}/vllm-ascend" "${INSTALL_DIR}"
-
-CANN_BKGS=${CANN_BKGS:-/opt/huawei/dataset/zyr_yuyin/bkgs/cann_0527}
-cp "${CANN_BKGS}/Ascend-cann-toolkit_9.0.0_linux-aarch64.run" "${INSTALL_DIR}"
-cp "${CANN_BKGS}/Ascend-cann-910b-ops_9.0.0_linux-aarch64.run" "${INSTALL_DIR}"
-cp "${CANN_BKGS}/Ascend-cann-nnal_9.0.0_linux-aarch64.run" "${INSTALL_DIR}"
-
-echo "################"
-echo "## set verl env"
-echo "################"
-
-cd "${INSTALL_DIR}"
-
-chmod +x Ascend-cann-toolkit_9.0.0_linux-aarch64.run
-bash Ascend-cann-toolkit_9.0.0_linux-aarch64.run --install --quiet
-source "${INSTALL_DIR}/Ascend/ascend-toolkit/set_env.sh"
-
-chmod +x Ascend-cann-910b-ops_9.0.0_linux-aarch64.run
-bash Ascend-cann-910b-ops_9.0.0_linux-aarch64.run --install --quiet
-
-chmod +x Ascend-cann-nnal_9.0.0_linux-aarch64.run
-bash Ascend-cann-nnal_9.0.0_linux-aarch64.run --install --quiet
-source "${INSTALL_DIR}/Ascend/nnal/atb/set_env.sh"
-
-export ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
-export LD_LIBRARY_PATH=/usr/local/Ascend/driver/lib64:/usr/local/Ascend/driver/lib64/common:${LD_LIBRARY_PATH:-}
-echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
-
-pip3 install torch==2.9.0
-pip3 install pyyaml setuptools
-pip3 install torch-npu==2.9.0
-pip3 install torchvision==0.24.0 torchaudio==2.9.0
-
-ASCEND_TOOLKIT_PYTHON_PATH=/home/ma-user/Ascend/ascend-toolkit/latest/python/site-packages
-export PYTHONPATH=${PYTHONPATH:-}:${INSTALL_DIR}:${ASCEND_TOOLKIT_PYTHON_PATH}
-pip install pybind11==2.13.6
-
-cd "${INSTALL_DIR}/vllm"
-VLLM_TARGET_DEVICE=empty pip install .
-
-cd "${INSTALL_DIR}/vllm-ascend"
-pip install -e .
-export VLLM_LOGGING_LEVEL=INFO
-
-cd "${INSTALL_DIR}"
-tar -xvf jemalloc-5.3.0.tar.bz2
-cd jemalloc-5.3.0
-./configure --prefix="${INSTALL_DIR}"
-make -j"$(nproc)"
-make install
-export LD_PRELOAD=${INSTALL_DIR}/lib/libjemalloc.so.2:${LD_PRELOAD:-}
-
-# # ================= 可选：安装 MindSpeed 栈 =================
-# # 纯 VERL engine 路线不依赖 MindSpeed-MM YAML。默认不安装，避免和新版 VERL engine 混淆。
-# INSTALL_MINDSPEED_STACK=${INSTALL_MINDSPEED_STACK:-0}
-# if [ "${INSTALL_MINDSPEED_STACK}" = "1" ]; then
-#     MindSpeed_PATH=${MindSpeed_PATH:-/opt/huawei/dataset/zyr_yuyin/lyf/verl-slow-stable}
-#     cd "${MindSpeed_PATH}"
-#     cd Megatron-LM && pip install -e . --no-deps && cd ..
-#     cd MindSpeed && pip install -e . --no-deps && cd ..
-#     cd MindSpeed-MM && mkdir -p logs data ckpt && pip install -e . --no-deps && cd ..
-#     pip install beartype bs4 diffusers==0.30.3 ftfy imageio-ffmpeg pandarallel pytest-mock
-# else
-#     echo "--> Skip MindSpeed/MindSpeed-MM installation for pure VERL engine route."
-# fi
-
-
-
-# ================= 安装 Triton-Ascend 3.2.1 =================
-# 1. 卸载 triton（增加 -y 自动确认）
-pip uninstall -y triton
-
-# 2. 卸载 triton-ascend（增加 -y 自动确认）
-pip uninstall -y triton-ascend
-pip install --no-cache-dir --force-reinstall triton==3.5.0
-pip install --no-deps /opt/huawei/dataset/zyr_yuyin/bkgs/triton_ascend-3.2.1-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl
-
-
-# ================= 安装新版 VERL =================
-cd "${WORK_DIR}"
-pip install -r requirements-npu.txt
-# NPU requirements 明确要求 numpy<2；editable 安装不能再次按 setup.py 把 NumPy升级到 2.x。
-python3 -m pip install -e . --no-deps
-pip install --upgrade 'urllib3==1.26.11'
-pip install loguru
-pip install tree_sitter==0.21.3
-pip install tree-sitter-java==0.21.0
-pip install tree-sitter-javascript==0.21.4
-
-ACL_PATH=/home/ma-user/Ascend/ascend-toolkit/latest/aarch64-linux/lib64
-export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${ACL_PATH}
-echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
-
-pip uninstall -y transformers || true
-pip install transformers==5.3.0
-pip install accelerate==1.13.0 mathruler
-pip install jsonargparse
-pip install deepdiff sympy html2text requests bs4 mpmath swanlab PandoraBox json_repair openai httpx
-
-# 稳定版只允许 SP=1：不安装普通 FLA，也不进入尚未完成 NPU 适配的 Ulysses CP。
-if [ "${TRAIN_SP}" != "1" ]; then
-    echo "ERROR: the standalone stable profile requires TRAIN_SP=1; got ${TRAIN_SP}." >&2
-    exit 2
-fi
-export NANOCLAW_REQUIRE_FLA=0
-echo "--> Stable SP1: flash-linear-attention is disabled; Qwen3.5 will not build an Ulysses CP context."
-
-# Transformers 5.x 会经 sklearn 间接导入 pandas/scipy。固定同一套 NumPy ABI，
-# 避免出现 "numpy.dtype size changed"。这些版本均支持 Python 3.11/aarch64。
-NUMPY_VERSION=${NUMPY_VERSION:-1.26.4}
-PANDAS_VERSION=${PANDAS_VERSION:-2.2.3}
-SCIPY_VERSION=${SCIPY_VERSION:-1.14.1}
-SKLEARN_VERSION=${SKLEARN_VERSION:-1.6.1}
-python3 -m pip install --no-cache-dir --force-reinstall \
-    "numpy==${NUMPY_VERSION}" \
-    "pandas==${PANDAS_VERSION}" \
-    "scipy==${SCIPY_VERSION}" \
-    "scikit-learn==${SKLEARN_VERSION}"
-
-python3 - <<'PY' || exit 2
-import numpy
-import pandas
-import scipy
-import sklearn
-import sys
-import transformers
-import vllm
-
-fla_version = "disabled-sp1"
-
-print(
-    "[python_stack_preflight] "
-    f"python={sys.executable} "
-    f"numpy={numpy.__version__} "
-    f"pandas={pandas.__version__} "
-    f"scipy={scipy.__version__} "
-    f"sklearn={sklearn.__version__} "
-    f"transformers={transformers.__version__} "
-    f"vllm={vllm.__version__} "
-    f"fla={fla_version}"
-)
-print(
-    "[python_stack_paths] "
-    f"numpy={numpy.__file__} "
-    f"pandas={pandas.__file__}"
-)
-PY
-pip list
-
-# ================= 检查 Nanoclaw recipe =================
-if [ ! -f "${WORK_DIR}/nanoclaw_recipe/nanoclaw.py" ]; then
-    echo "ERROR: Nanoclaw recipe not found: ${WORK_DIR}/nanoclaw_recipe/nanoclaw.py" >&2
-    exit 2
-fi
-test -f "${WORK_DIR}/nanoclaw_recipe/__init__.py"
-
-# Qwen3.5 MRoPE position_ids 是 3/4 轴张量。未应用此补丁时，NPU
-# FlashAttention 会把 seqLen 重复累计（例如 T=10131、sum(seqLen)=30393）。
-QWEN35_MONKEY_PATCH_FILE=${WORK_DIR}/verl/models/transformers/monkey_patch.py
-if [ ! -f "${QWEN35_MONKEY_PATCH_FILE}" ] || ! grep -q "def _normalize_fa_position_ids" "${QWEN35_MONKEY_PATCH_FILE}"; then
-    echo "ERROR: Qwen3.5 FlashAttention position_ids normalization patch is missing: ${QWEN35_MONKEY_PATCH_FILE}" >&2
-    echo "Upload the modified verl/ directory together with this standalone script." >&2
-    exit 2
-fi
-
-# Nanoclaw 会在首次模型调用前写入 rollout metadata。旧 ToolAgentLoop 因此
-# 未复制 min/max_global_steps，导致完整训练 step 后在 metrics 阶段把 None 转 int 崩溃。
-TOOL_AGENT_LOOP_FILE=${WORK_DIR}/verl/experimental/agent_loop/tool_agent_loop.py
-TRAINER_BASE_FILE=${WORK_DIR}/verl/trainer/ppo/v1/trainer_base.py
-if [ ! -f "${TOOL_AGENT_LOOP_FILE}" ] || ! grep -q "output_min_global_steps" "${TOOL_AGENT_LOOP_FILE}"; then
-    echo "ERROR: Nanoclaw rollout version metadata merge fix is missing: ${TOOL_AGENT_LOOP_FILE}" >&2
-    exit 2
-fi
-if [ ! -f "${TRAINER_BASE_FILE}" ] || ! grep -q "def resolve_model_version" "${TRAINER_BASE_FILE}"; then
-    echo "ERROR: PPO metrics None-version fallback fix is missing: ${TRAINER_BASE_FILE}" >&2
-    exit 2
-fi
-
-# fused Torch LM-head 必须同时具备 Qwen3.5 backend 和分块 autograd 实现。
-QWEN35_MODEL_PATCH_FILE=${WORK_DIR}/verl/models/transformers/qwen3_5.py
-FUSED_LINEAR_FILE=${WORK_DIR}/verl/utils/experimental/torch_functional.py
-if [ ! -f "${QWEN35_MODEL_PATCH_FILE}" ] || ! grep -q "def forward_with_torch_backend" "${QWEN35_MODEL_PATCH_FILE}"; then
-    echo "ERROR: Qwen3.5 fused Torch backend is missing: ${QWEN35_MODEL_PATCH_FILE}" >&2
-    exit 2
-fi
-if [ ! -f "${FUSED_LINEAR_FILE}" ] || ! grep -q "class FusedLinearForPPO" "${FUSED_LINEAR_FILE}"; then
-    echo "ERROR: chunked FusedLinearForPPO is missing: ${FUSED_LINEAR_FILE}" >&2
-    exit 2
-fi
-
-# ================= PLOG =================
-ma_vj_name=$(echo "${MA_VJ_NAME}" | sed 's:ma-job:modelarts-job:g')
-task_name=worker-${VC_TASK_INDEX}
-task_plog_path=${MA_LOG_DIR}/${ma_vj_name}/${task_name}
-mkdir -p "${task_plog_path}"
-export ASCEND_PROCESS_LOG_PATH=${task_plog_path}/${VC_TASK_INDEX}
-echo "plog path: ${ASCEND_PROCESS_LOG_PATH}"
-
-MASTER_ADDR=${MA_VJ_NAME}-${MA_TASK_NAME}-${VC_TASK_INDEX}.${MA_VJ_NAME}
-MASTER_PORT=${PORT}
-MA_CURRENT_INSTANCE_NAME=${MA_CURRENT_INSTANCE_NAME}
-
-cd "${WORK_DIR}"
-
-mkdir -p /cache/ray_tmp
-
-echo "Cleaning up old Ray processes..."
-ray stop --force || true
-sleep 5
-rm -rf /cache/ray_tmp/*
-pkill -9 -f raylet || true
-pkill -9 -f plasma_store || true
-pkill -9 -f gcs_server || true
-echo "Waiting 20s for NPU/Ray resources to be released..."
-npu-smi info || true
-sleep 20
-
-# ================= NPU / HCCL / Ray 环境 =================
-export NON_MEGATRON=true
-export MULTI_STREAM_MEMORY_REUSE=2
-export OMP_NUM_THREADS=1
-export PYTORCH_NPU_ALLOC_CONF=${PYTORCH_NPU_ALLOC_CONF:-max_split_size_mb:512}
-export VLLM_LOGGING_LEVEL=INFO
-export RAY_DEDUP_LOGS=0
-export HCCL_EXEC_TIMEOUT=${HCCL_EXEC_TIMEOUT:-3600}
-export HCCL_LOG_LEVEL=${HCCL_LOG_LEVEL:-WARN}
-export HCCL_CONNECT_TIMEOUT=${HCCL_CONNECT_TIMEOUT:-3600}
-export HCCL_EVENT_TIMEOUT=${HCCL_EVENT_TIMEOUT:-7200}
-export ACL_DEVICE_SYNC_TIMEOUT=${ACL_DEVICE_SYNC_TIMEOUT:-7200}
-export GLOO_SOCKET_TIMEOUT=${GLOO_SOCKET_TIMEOUT:-7200}
-
-# 关键：降低 HCCL buffer，增加 socket 端口范围，缓解 HcclAllreduce ra socket batch connect failed。
-export HCCL_BUFFSIZE=${HCCL_BUFFSIZE:-300}
-export P2P_HCCL_BUFFSIZE=${P2P_HCCL_BUFFSIZE:-64}
-export HCCL_HOST_SOCKET_PORT_RANGE=${HCCL_HOST_SOCKET_PORT_RANGE:-60000-60050}
-export HCCL_NPU_SOCKET_PORT_RANGE=${HCCL_NPU_SOCKET_PORT_RANGE:-61000-61050}
-
-export CUDA_DEVICE_MAX_CONNECTIONS=1
-export VLLM_ASCEND_ENABLE_NZ=${VLLM_ASCEND_ENABLE_NZ:-0}
-export HCCL_OP_EXPANSION_MODE=${HCCL_OP_EXPANSION_MODE:-AIV}
-export VLLM_ENGINE_ITERATION_TIMEOUT_S=${VLLM_ENGINE_ITERATION_TIMEOUT_S:-3600}
-export WANDB_MODE=${WANDB_MODE:-disabled}
-export PYTHONUNBUFFERED=1
-export TASK_QUEUE_ENABLE=${TASK_QUEUE_ENABLE:-1}
-export COMBINED_ENABLE=${COMBINED_ENABLE:-1}
-export TOKENIZERS_PARALLELISM=false
-export CLOSE_MATMUL_K_SHIFT=${CLOSE_MATMUL_K_SHIFT:-1}
-export ATB_MATMUL_SHUFFLE_K_ENABLE=${ATB_MATMUL_SHUFFLE_K_ENABLE:-0}
-export HCCL_DETERMINISTIC=${HCCL_DETERMINISTIC:-true}
-export VLLM_ENABLE_V1_MULTIPROCESSING=${VLLM_ENABLE_V1_MULTIPROCESSING:-0}
-export VLLM_USE_V1=${VLLM_USE_V1:-1}
-export ASCEND_GLOBAL_LOG_LEVEL=${ASCEND_GLOBAL_LOG_LEVEL:-3}
-export HYDRA_FULL_ERROR=1
-export RAY_gcs_server_rpc_server_thread_num=${RAY_gcs_server_rpc_server_thread_num:-32}
-export RAY_gcs_server_request_timeout_seconds=${RAY_gcs_server_request_timeout_seconds:-600}
-export RAY_timeout_ms=${RAY_timeout_ms:-600000}
-export RAY_worker_register_timeout_seconds=${RAY_worker_register_timeout_seconds:-600}
-export RAY_USAGE_STATS_ENABLED=0
-export VERL_REUSE_AGENT_LOOP=${VERL_REUSE_AGENT_LOOP:-1}
-
-ulimit -n 65536
-
-# Ray 不要覆盖 ASCEND_RT_VISIBLE_DEVICES；VERL 内部按 local_rank 选卡。
-export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
-
-# ================= 路径与数据配置 =================
-HDFS_ROOT=${HDFS_ROOT:-$PWD}
-DATA_ROOT=${DATA_ROOT:-/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt}
-
-# Nanoclaw 数据输入支持两种目录，优先推荐 0625 扁平格式：
-#   base_tasks/data_*/env_builder.py
-#   base_tasks/data_*/prompts.md
-#   base_tasks/data_*/workplace_verifier.py
-#   base_tasks/data_*/manifest.json
-# 也兼容旧格式：base_tasks/tasks/data_* + base_tasks/scripts|scrips/data_*。
-DEFAULT_NANOCLAW_BASE_TASKS=${DEFAULT_NANOCLAW_BASE_TASKS:-/opt/huawei/dataset/zyr_yuyin/lyf/datasets/nanoclawRLdata/0710_add1000agent_qwen3_7_max_v1/exported_new_data}
-train_base_tasks=${TRAIN_DATA_PATH:-${BASE_TASKS:-${DEFAULT_NANOCLAW_BASE_TASKS}}}
-val_base_tasks=${VAL_DATA_PATH:-${VAL_BASE_TASKS:-${train_base_tasks}}}
-train_files="['$train_base_tasks']"
-test_files="['$val_base_tasks']"
-
-if [ ! -d "${train_base_tasks}" ]; then
-    echo "ERROR: Nanoclaw TRAIN_DATA_PATH/BASE_TASKS directory not found: ${train_base_tasks}" >&2
-    exit 2
-fi
-if [ ! -d "${val_base_tasks}" ]; then
-    echo "ERROR: Nanoclaw VAL_DATA_PATH/VAL_BASE_TASKS directory not found: ${val_base_tasks}" >&2
-    exit 2
-fi
-
-model_path=${MODEL_PATH:-/opt/huawei/dataset/zyr_yuyin/models/Qwen/Qwen3___5-9B}
-verifier_model_path=${VERIFIER_MODEL_PATH:-/opt/huawei/dataset/zyr_yuyin/models/Qwen/Qwen3___5-9B}
-# 纯 VERL engine 路线：不要使用 MindSpeed-MM YAML。
-unset MM_CONFIG_FILE || true
-
-# Nanoclaw 工具配置
-tool_config_path=${TOOL_CONFIG_PATH:-nanoclaw_recipe/nanoclaw_tool_config.yaml}
-nanoclaw_task_glob=${NANOCLAW_TASK_GLOB:-data_*}
-nanoclaw_task_ids=${NANOCLAW_TASK_IDS:-}
-# 多机训练必须用所有节点都能访问的共享目录；不要用 /tmp，否则 reward worker 可能跨节点找不到 workspace。
-nanoclaw_temp_root=${NANOCLAW_TEMP_ROOT:-/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_workplace_qwen35_9b_sp1_16k_fused_lmhead_smoke}
-# 默认保留每个 step/data_sample 的目录，方便复盘每条 GRPO 采样；磁盘紧张时手动设 NANOCLAW_CLEANUP_WORKSPACES=True。
-nanoclaw_cleanup_workspaces=${NANOCLAW_CLEANUP_WORKSPACES:-False}
-nanoclaw_keep_failed_workspaces=${NANOCLAW_KEEP_FAILED_WORKSPACES:-False}
-nanoclaw_env_builder_timeout=${NANOCLAW_ENV_BUILDER_TIMEOUT:-120}
-nanoclaw_verifier_timeout=${NANOCLAW_VERIFIER_TIMEOUT:-3600}
-nanoclaw_reward_score_mode=${NANOCLAW_REWARD_SCORE_MODE:-ratio}
-nanoclaw_allow_bash=${NANOCLAW_ALLOW_BASH:-True}
-nanoclaw_max_steps=${NANOCLAW_MAX_STEPS:-}
-nanoclaw_require_final_answer=${NANOCLAW_REQUIRE_FINAL_ANSWER:-True}
-nanoclaw_final_answer_bonus_enable=${NANOCLAW_FINAL_ANSWER_BONUS_ENABLE:-False}
-nanoclaw_final_answer_bonus_score=${NANOCLAW_FINAL_ANSWER_BONUS_SCORE:-0.0}
-nanoclaw_turn_penalty_only_positive_score=${NANOCLAW_TURN_PENALTY_ONLY_POSITIVE_SCORE:-False}
-nanoclaw_assistant_turn_penalty=${NANOCLAW_ASSISTANT_TURN_PENALTY:-0.0}
-nanoclaw_duplicate_tool_call_penalty=${NANOCLAW_DUPLICATE_TOOL_CALL_PENALTY:-0.0}
-nanoclaw_repeated_response_penalty=${NANOCLAW_REPEATED_RESPONSE_PENALTY:-0.0}
-nanoclaw_repeated_response_min_chars=${NANOCLAW_REPEATED_RESPONSE_MIN_CHARS:-50}
-nanoclaw_repeated_response_min_consecutive_repeats=${NANOCLAW_REPEATED_RESPONSE_MIN_CONSECUTIVE_REPEATS:-5}
-nanoclaw_mask_looping_responses=${NANOCLAW_MASK_LOOPING_RESPONSES:-True}
-nanoclaw_mask_only_positive_advantage=${NANOCLAW_MASK_ONLY_POSITIVE_ADVANTAGE:-True}
-nanoclaw_mask_budget_exhausted_last_turn=${NANOCLAW_MASK_BUDGET_EXHAUSTED_LAST_TURN:-True}
-nanoclaw_mask_duplicate_tool_result_turns=${NANOCLAW_MASK_DUPLICATE_TOOL_RESULT_TURNS:-True}
-nanoclaw_mask_error_tool_result_turns=${NANOCLAW_MASK_ERROR_TOOL_RESULT_TURNS:-True}
-
-# verify_workplace.py 如需调用本地 OpenAI-compatible API，可用这些变量传入 reward。
-# 默认假设 5 机 40 卡：前 4 个节点加入 Ray 训练，第 5 个节点部署 verifier/vLLM API。
-verifier_api_node_rank=${VERIFIER_API_NODE_RANK:-4}
-verifier_api_port=${VERIFIER_API_PORT:-8000}
-verifier_api_host=${VERIFIER_API_HOST:-${MA_VJ_NAME}-${MA_TASK_NAME}-${verifier_api_node_rank}.${MA_VJ_NAME}}
-verifier_api_start_cmd=${VERIFIER_API_START_CMD:-}
-verifier_api_bind_host=${VERIFIER_API_BIND_HOST:-0.0.0.0}
-# 9B verifier 默认使用整台 8 卡节点：两份 TP4 副本由 vLLM 内置 DP 统一服务。
-# 如需单副本 TP8，可设置 VERIFIER_API_TP=8 VERIFIER_API_DP=1。
-verifier_api_tp=${VERIFIER_API_TP:-4}
-verifier_api_dp=${VERIFIER_API_DP:-2}
-verifier_api_devices=${VERIFIER_API_DEVICES:-0,1,2,3,4,5,6,7}
-verifier_api_distributed_executor_backend=${VERIFIER_API_DISTRIBUTED_EXECUTOR_BACKEND:-mp}
-verifier_api_max_model_len=${VERIFIER_API_MAX_MODEL_LEN:-32768}
-verifier_api_max_num_batched_tokens=${VERIFIER_API_MAX_NUM_BATCHED_TOKENS:-32768}
-verifier_api_max_num_seqs=${VERIFIER_API_MAX_NUM_SEQS:-160}
-verifier_api_gpu_memory_utilization=${VERIFIER_API_GPU_MEMORY_UTILIZATION:-0.70}
-verifier_api_enforce_eager=${VERIFIER_API_ENFORCE_EAGER:-0}
-verifier_api_enable_graph_mode=${VERIFIER_API_ENABLE_GRAPH_MODE:-1}
-verifier_api_enable_prefix_caching=${VERIFIER_API_ENABLE_PREFIX_CACHING:-0}
-verifier_api_startup_timeout=${VERIFIER_API_STARTUP_TIMEOUT:-1800}
-verifier_api_log=${VERIFIER_API_LOG:-logs/vllm-verifier-api.log}
-mock_api_base=${MOCK_API_BASE:-http://${verifier_api_host}:${verifier_api_port}/v1}
-mock_api_key=${MOCK_API_KEY:-dummy_key}
-mock_model_name=${MOCK_MODEL_NAME:-qwen3_5_9b_verifier}
-# verify_workplace.py 内部 OpenAI/httpx 单次请求超时；reward API 排队时宁可多等，不要轻易误判 0 分。
-mock_api_timeout=${MOCK_API_TIMEOUT:-1800}
-mock_api_connect_timeout=${MOCK_API_CONNECT_TIMEOUT:-300}
-# 强制 verifier/OpenAI judge 请求关闭 thinking，sitecustomize 会自动注入 extra_body.chat_template_kwargs.enable_thinking=False。
-nanoclaw_force_no_thinking=${NANOCLAW_FORCE_NO_THINKING:-1}
-nanoclaw_force_max_tokens=${NANOCLAW_FORCE_MAX_TOKENS:-50}
-# 默认控制台只打一行 reward 摘要；如需每项 details，设 NANOCLAW_REWARD_PRINT_DETAILS=1。
-nanoclaw_reward_print_details=${NANOCLAW_REWARD_PRINT_DETAILS:-0}
-# verifier API 是单独节点，默认低并发，避免 RewardLoopWorker 同时打爆 API 导致排队超时。
-reward_num_workers=${REWARD_NUM_WORKERS:-52}
-
-project_name=${PROJECT_NAME:-qwen3.5-9b_nanoclaw_grpo_verl_0720}
-experiment_name=${EXPERIMENT_NAME:-qwen3.5-9b_nanoclaw_grpo_sp1_16k_fused_lmhead_smoke}
-default_local_dir=${DEFAULT_LOCAL_DIR:-$DATA_ROOT/checkpoint/$experiment_name}
-start_time=$(date +%Y%m%d)_$(date +%H%M%S)
-mkdir -p logs "${default_local_dir}"
-
-# ================= 算法与并行参数 =================
-adv_estimator=grpo
-max_turns=${MAX_TURNS:-35}
-max_prompt_length=${MAX_PROMPT_LENGTH:-8192}
-max_response_length=${MAX_RESPONSE_LENGTH:-22768}
-max_assistant_response_length=${MAX_ASSISTANT_RESPONSE_LENGTH:-16384}
-max_tool_response_length=${MAX_TOOL_RESPONSE_LENGTH:-8192}
-max_model_len=$((max_prompt_length + max_response_length))
-
-# MindSpeed 配置仅作为训练意图参考；以下均使用最新版 VERL 的原生字段。
-actor_lr=${ACTOR_LR:-1e-6}
-actor_lr_scheduler_type=${ACTOR_LR_SCHEDULER_TYPE:-constant}
-actor_lr_warmup_steps_ratio=${ACTOR_LR_WARMUP_STEPS_RATIO:-0.0}
-actor_weight_decay=${ACTOR_WEIGHT_DECAY:-0.01}
-actor_adam_beta1=${ACTOR_ADAM_BETA1:-0.9}
-actor_adam_beta2=${ACTOR_ADAM_BETA2:-0.95}
-actor_clip_grad=${ACTOR_CLIP_GRAD:-1.0}
-actor_ppo_epochs=${ACTOR_PPO_EPOCHS:-1}
-actor_shuffle=${ACTOR_SHUFFLE:-False}
-actor_entropy_coeff=${ACTOR_ENTROPY_COEFF:-0.0}
-actor_clip_ratio_low=${ACTOR_CLIP_RATIO_LOW:-0.2}
-actor_clip_ratio_high=${ACTOR_CLIP_RATIO_HIGH:-0.2}
-# MindSpeed 配置没有 Dual-Clip PPO 对应项，保留该 27B 脚本原来的 C=10。
-actor_clip_ratio_c=${ACTOR_CLIP_RATIO_C:-10.0}
-
-# YAML 的 fixed init_kl_coef + low_var_kl 对应 VERL 的 reward-KL 路径。
-algorithm_gamma=${ALGORITHM_GAMMA:-1.0}
-algorithm_lam=${ALGORITHM_LAM:-0.95}
-use_kl_in_reward=${USE_KL_IN_REWARD:-True}
-kl_penalty=${KL_PENALTY:-low_var_kl}
-kl_ctrl_type=${KL_CTRL_TYPE:-fixed}
-kl_coef=${KL_COEF:-0.001}
-# 关闭 actor-KL，避免与 reward-KL 重复惩罚。
-actor_use_kl_loss=${ACTOR_USE_KL_LOSS:-False}
-actor_kl_loss_coef=${ACTOR_KL_LOSS_COEF:-0.001}
-actor_kl_loss_type=${ACTOR_KL_LOSS_TYPE:-low_var_kl}
-
-train_batch_size=${TRAIN_BATCH_SIZE:-64}
-ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE:-16}
-n_resp_per_prompt=${N_RESP_PER_PROMPT:-8}
-# 先压低验证，避免验证和训练稳定性混在一起。
-n_resp_per_prompt_val=${N_RESP_PER_PROMPT_VAL:-1}
-log_val_generations=${LOG_VAL_GENERATIONS:-10}
-
-infer_tp=${INFER_TP:-4}
-train_sp=${TRAIN_SP:-1}
-offload=${OFFLOAD:-True}
-
-# 稳定版固定沿用已经完成初始化与 rollout 验证的 FSDP1/offload 形状。
-actor_strategy=${ACTOR_STRATEGY:-fsdp}
-fsdp_size=${FSDP_SIZE:-}
-
-actor_pack=${ACTOR_PACK:-1}
-logprob_pack=${LOGPROB_PACK:-2}
-actor_max_token_len_per_gpu=${ACTOR_MAX_TOKEN_LEN_PER_GPU:-$(((max_model_len * actor_pack + train_sp - 1) / train_sp))}
-log_prob_max_token_len_per_gpu=${LOG_PROB_MAX_TOKEN_LEN_PER_GPU:-$(((max_model_len * logprob_pack + train_sp - 1) / train_sp))}
-entropy_from_logits_with_chunking=${ENTROPY_FROM_LOGITS_WITH_CHUNKING:-True}
-entropy_from_logits_chunk_size=${ENTROPY_FROM_LOGITS_CHUNK_SIZE:-256}
-entropy_checkpointing=${ENTROPY_CHECKPOINTING:-False}
-use_fused_kernels=${USE_FUSED_KERNELS:-True}
-fused_kernel_backend=${FUSED_KERNEL_BACKEND:-torch}
-enable_activation_offload=${ENABLE_ACTIVATION_OFFLOAD:-False}
-rollout_max_num_batched_tokens=${ROLLOUT_MAX_NUM_BATCHED_TOKENS:-32384}
-rollout_gpu_memory_utilization=${ROLLOUT_GPU_MEMORY_UTILIZATION:-0.60}
-update_weights_bucket_mb=${UPDATE_WEIGHTS_BUCKET_MB:-8192}
-
-# Qwen 官方推荐：Instruct/non-thinking reasoning tasks
-rollout_temperature=${ROLLOUT_TEMPERATURE:-0.6}
-rollout_top_p=${ROLLOUT_TOP_P:-0.95}
-rollout_top_k=${ROLLOUT_TOP_K:-20}
-rollout_min_p=${ROLLOUT_MIN_P:-0.0}
-rollout_presence_penalty=${ROLLOUT_PRESENCE_PENALTY:-0.0}
-rollout_frequency_penalty=${ROLLOUT_FREQUENCY_PENALTY:-0.0}
-rollout_repetition_penalty=${ROLLOUT_REPETITION_PENALTY:-1.0}
-
-echo "DEBUG: max_response_length=${max_response_length}, max_assistant_response_length=${max_assistant_response_length}, max_model_len=${max_model_len}"
-echo "DEBUG: max_turns=${max_turns}"
-echo "DEBUG: max_tool_response_length=${max_tool_response_length}"
-echo "DEBUG: entropy_chunking=${entropy_from_logits_with_chunking}, entropy_chunk_size=${entropy_from_logits_chunk_size}, entropy_checkpointing=${entropy_checkpointing}"
-echo "DEBUG: fused_lmhead=${use_fused_kernels}, fused_backend=${fused_kernel_backend}, activation_offload=${enable_activation_offload}"
-echo "DEBUG: train_batch_size=${train_batch_size}, ppo_mini_batch_size=${ppo_mini_batch_size}, n=${n_resp_per_prompt}"
-echo "DEBUG: train_sp=${train_sp}, infer_tp=${infer_tp}, actor_strategy=${actor_strategy}, fsdp_size=${fsdp_size:-<default>}"
-echo "DEBUG: Qwen3.5 Ulysses FLA required=${NANOCLAW_REQUIRE_FLA}, backend=${QWEN35_FLA_BACKEND} (TRAIN_SP=${train_sp})"
-echo "DEBUG: actor_max_token_len_per_gpu=${actor_max_token_len_per_gpu}, log_prob_max_token_len_per_gpu=${log_prob_max_token_len_per_gpu}"
-echo "DEBUG: rollout sampling temperature=${rollout_temperature}, top_p=${rollout_top_p}, top_k=${rollout_top_k}, min_p=${rollout_min_p}, presence_penalty=${rollout_presence_penalty}, frequency_penalty=${rollout_frequency_penalty}, repetition_penalty=${rollout_repetition_penalty}"
-echo "DEBUG: optimizer lr=${actor_lr}, scheduler=${actor_lr_scheduler_type}, warmup_ratio=${actor_lr_warmup_steps_ratio}, weight_decay=${actor_weight_decay}, betas=(${actor_adam_beta1},${actor_adam_beta2}), clip_grad=${actor_clip_grad}, ppo_epochs=${actor_ppo_epochs}, shuffle=${actor_shuffle}"
-echo "DEBUG: KL use_in_reward=${use_kl_in_reward}, penalty=${kl_penalty}, ctrl=${kl_ctrl_type}, coef=${kl_coef}, actor_kl=${actor_use_kl_loss}"
-echo "DEBUG: HCCL_BUFFSIZE=${HCCL_BUFFSIZE}, HCCL_HOST_SOCKET_PORT_RANGE=${HCCL_HOST_SOCKET_PORT_RANGE}, HCCL_NPU_SOCKET_PORT_RANGE=${HCCL_NPU_SOCKET_PORT_RANGE}"
-
-val_before_train=${VAL_BEFORE_TRAIN:-False}
-trainer_use_v1=${TRAINER_USE_V1:-True}
-test_freq=${TEST_FREQ:-5000}
-save_freq=${SAVE_FREQ:-2}
-total_training_steps=${TOTAL_TRAINING_STEPS:-2}
-
-# ================= 分布式 =================
-export TOTAL_NNODES=${TOTAL_NNODES:-5}
-export TRAIN_NNODES=${TRAIN_NNODES:-4}
-export NNODES=${NNODES:-${TRAIN_NNODES}}
-export NODE_RANK=${VC_TASK_INDEX}
-export NPUS_PER_NODE=${NPUS_PER_NODE:-8}
-export WORLD_SIZE=$((NPUS_PER_NODE * NNODES))
-
-export MASTER_ADDR=${MA_VJ_NAME}-${MA_TASK_NAME}-0.${MA_VJ_NAME}
-export MASTER_PORT=${MASTER_PORT:-6167}
-export DASHBOARD_PORT=${DASHBOARD_PORT:-8191}
-export RAY_PORT=${RAY_PORT:-6167}
-
-readonly SOCKET_IFNAME=${SOCKET_IFNAME:-eth0}
-export HCCL_SOCKET_IFNAME=${HCCL_SOCKET_IFNAME:-${SOCKET_IFNAME}}
-export GLOO_SOCKET_IFNAME=${GLOO_SOCKET_IFNAME:-${SOCKET_IFNAME}}
-export CURRENT_IP=$(ifconfig ${SOCKET_IFNAME} | grep -Eo 'inet (addr:)?([0-9]{1,3}\.){3}[0-9]{1,3}' | awk '{print $NF}')
-export RAY_NODE_IP=${MA_CURRENT_IP:-${CURRENT_IP}}
-
-export ASCEND_RT_VISIBLE_DEVICES=${ASCEND_RT_VISIBLE_DEVICES:-$(seq -s, 0 $((NPUS_PER_NODE - 1)))}
-
-cat <<EOF
-DEBUG: MASTER_ADDR=${MASTER_ADDR}
-DEBUG: MASTER_PORT=${MASTER_PORT}
-DEBUG: RAY_PORT=${RAY_PORT}
-DEBUG: MA_CURRENT_IP=${MA_CURRENT_IP}
-DEBUG: CURRENT_IP=${CURRENT_IP}
-DEBUG: RAY_NODE_IP=${RAY_NODE_IP}
-DEBUG: ASCEND_RT_VISIBLE_DEVICES=${ASCEND_RT_VISIBLE_DEVICES}
-DEBUG: HCCL_SOCKET_IFNAME=${HCCL_SOCKET_IFNAME}
-DEBUG: GLOO_SOCKET_IFNAME=${GLOO_SOCKET_IFNAME}
-DEBUG: TOTAL_NNODES=${TOTAL_NNODES}
-DEBUG: TRAIN_NNODES=${TRAIN_NNODES}
-DEBUG: VERIFIER_API_NODE_RANK=${verifier_api_node_rank}
-DEBUG: MOCK_API_BASE=${mock_api_base}
-DEBUG: MOCK_MODEL_NAME=${mock_model_name}
-DEBUG: MOCK_API_TIMEOUT=${mock_api_timeout}
-DEBUG: NANOCLAW_FORCE_NO_THINKING=${nanoclaw_force_no_thinking}
-DEBUG: NANOCLAW_FORCE_MAX_TOKENS=${nanoclaw_force_max_tokens}
-DEBUG: NANOCLAW_REWARD_PRINT_DETAILS=${nanoclaw_reward_print_details}
-DEBUG: NANOCLAW_REQUIRE_FINAL_ANSWER=${nanoclaw_require_final_answer}
-DEBUG: NANOCLAW_FINAL_ANSWER_BONUS_ENABLE=${nanoclaw_final_answer_bonus_enable}
-DEBUG: NANOCLAW_FINAL_ANSWER_BONUS_SCORE=${nanoclaw_final_answer_bonus_score}
-DEBUG: NANOCLAW_TURN_PENALTY_ONLY_POSITIVE_SCORE=${nanoclaw_turn_penalty_only_positive_score}
-DEBUG: NANOCLAW_ASSISTANT_TURN_PENALTY=${nanoclaw_assistant_turn_penalty}
-DEBUG: NANOCLAW_DUPLICATE_TOOL_CALL_PENALTY=${nanoclaw_duplicate_tool_call_penalty}
-DEBUG: NANOCLAW_REPEATED_RESPONSE_PENALTY=${nanoclaw_repeated_response_penalty}
-DEBUG: NANOCLAW_REPEATED_RESPONSE_MIN_CHARS=${nanoclaw_repeated_response_min_chars}
-DEBUG: NANOCLAW_REPEATED_RESPONSE_MIN_CONSECUTIVE_REPEATS=${nanoclaw_repeated_response_min_consecutive_repeats}
-DEBUG: NANOCLAW_MASK_LOOPING_RESPONSES=${nanoclaw_mask_looping_responses}
-DEBUG: NANOCLAW_MASK_ONLY_POSITIVE_ADVANTAGE=${nanoclaw_mask_only_positive_advantage}
-DEBUG: NANOCLAW_MASK_BUDGET_EXHAUSTED_LAST_TURN=${nanoclaw_mask_budget_exhausted_last_turn}
-DEBUG: NANOCLAW_MASK_DUPLICATE_TOOL_RESULT_TURNS=${nanoclaw_mask_duplicate_tool_result_turns}
-DEBUG: NANOCLAW_MASK_ERROR_TOOL_RESULT_TURNS=${nanoclaw_mask_error_tool_result_turns}
-DEBUG: NANOCLAW_LOOPING_RESPONSE_MIN_CHARS=${nanoclaw_looping_response_min_chars}
-DEBUG: NANOCLAW_LOOPING_RESPONSE_MIN_CONSECUTIVE_REPEATS=${nanoclaw_looping_response_min_consecutive_repeats}
-DEBUG: VERIFIER_API_TP=${verifier_api_tp}
-DEBUG: VERIFIER_API_DP=${verifier_api_dp}
-DEBUG: VERIFIER_API_DEVICES=${verifier_api_devices}
-DEBUG: VERIFIER_API_DISTRIBUTED_EXECUTOR_BACKEND=${verifier_api_distributed_executor_backend}
-DEBUG: VERIFIER_API_MAX_NUM_SEQS=${verifier_api_max_num_seqs}
-DEBUG: VERIFIER_API_ENFORCE_EAGER=${verifier_api_enforce_eager}
-DEBUG: VERIFIER_API_ENABLE_GRAPH_MODE=${verifier_api_enable_graph_mode}
-DEBUG: REWARD_NUM_WORKERS=${reward_num_workers}
-EOF
-
-if [ "${NODE_RANK}" = "${verifier_api_node_rank}" ]; then
-    echo "--> [Verifier API Node] This node is reserved for vLLM/OpenAI-compatible verifier API."
-    echo "--> [Verifier API Node] API base: ${mock_api_base}"
-    export VLLM_ENABLE_GRAPH_MODE=${verifier_api_enable_graph_mode}
-    mkdir -p "$(dirname "${verifier_api_log}")"
-    if [ -n "${verifier_api_start_cmd}" ]; then
-        echo "--> [Verifier API Node] Running VERIFIER_API_START_CMD..."
-        bash -lc "${verifier_api_start_cmd}" &
-        verifier_api_pid=$!
-    else
-        echo "--> [Verifier API Node] Starting default vLLM verifier API..."
-        verifier_api_device_count=$(awk -F',' '{print NF}' <<<"${verifier_api_devices}")
-        verifier_api_expected_device_count=$((verifier_api_tp * verifier_api_dp))
-        if [ "${verifier_api_device_count}" -ne "${verifier_api_expected_device_count}" ]; then
-            echo "ERROR: verifier TP*DP=${verifier_api_tp}*${verifier_api_dp}=${verifier_api_expected_device_count}, but VERIFIER_API_DEVICES=${verifier_api_devices} contains ${verifier_api_device_count} devices." >&2
-            exit 2
-        fi
-        export ASCEND_RT_VISIBLE_DEVICES=${verifier_api_devices}
-        verifier_api_args=(
-            --model "${verifier_model_path}"
-            --tokenizer "${verifier_model_path}"
-            --host "${verifier_api_bind_host}"
-            --port "${verifier_api_port}"
-            --served-model-name "${mock_model_name}"
-            --tensor-parallel-size "${verifier_api_tp}"
-            --data-parallel-size "${verifier_api_dp}"
-            --distributed-executor-backend "${verifier_api_distributed_executor_backend}"
-            --dtype bfloat16
-            --max-model-len "${verifier_api_max_model_len}"
-            --max-num-batched-tokens "${verifier_api_max_num_batched_tokens}"
-            --max-num-seqs "${verifier_api_max_num_seqs}"
-            --gpu-memory-utilization "${verifier_api_gpu_memory_utilization}"
-            --trust-remote-code
-        )
-        if [ "${verifier_api_enforce_eager}" = "1" ] || [ "${verifier_api_enforce_eager}" = "true" ] || [ "${verifier_api_enforce_eager}" = "True" ]; then
-            verifier_api_args+=(--enforce-eager)
-        fi
-        if [ "${verifier_api_enable_prefix_caching}" = "1" ] || [ "${verifier_api_enable_prefix_caching}" = "true" ] || [ "${verifier_api_enable_prefix_caching}" = "True" ]; then
-            verifier_api_args+=(--enable-prefix-caching)
-        fi
-        echo "--> [Verifier API Node] Command: python3 -m vllm.entrypoints.openai.api_server ${verifier_api_args[*]}"
-        python3 -m vllm.entrypoints.openai.api_server "${verifier_api_args[@]}" >"${verifier_api_log}" 2>&1 &
-        verifier_api_pid=$!
-    fi
-
-    echo "--> [Verifier API Node] vLLM API pid=${verifier_api_pid}, log=${verifier_api_log}"
-    echo "--> [Verifier API Node] Waiting for ${mock_api_base}/models ..."
-    python3 - "${mock_api_base}/models" "${verifier_api_startup_timeout}" "${verifier_api_log}" "${verifier_api_pid}" <<'PY'
-import os
-import sys
-import time
-import urllib.request
-from pathlib import Path
-
-url = sys.argv[1]
-timeout = float(sys.argv[2])
-log_path = Path(sys.argv[3])
-pid = int(sys.argv[4]) if len(sys.argv) > 4 and sys.argv[4] else None
-started = time.time()
-last_error = None
-while time.time() - started < timeout:
-    if pid is not None:
-        try:
-            os.kill(pid, 0)
-        except OSError:
-            print(f"ERROR: verifier API process exited early: pid={pid}", file=sys.stderr)
-            if log_path.is_file():
-                print("\n".join(log_path.read_text(encoding="utf-8", errors="replace").splitlines()[-120:]), file=sys.stderr)
-            sys.exit(1)
-    try:
-        with urllib.request.urlopen(url, timeout=5) as response:
-            if 200 <= response.status < 300:
-                print(f"READY: {url}", file=sys.stderr)
-                sys.exit(0)
-    except Exception as exc:
-        last_error = exc
-    time.sleep(5)
-print(f"ERROR: timed out waiting for {url}; last_error={last_error}", file=sys.stderr)
-if log_path.is_file():
-    print("\n".join(log_path.read_text(encoding="utf-8", errors="replace").splitlines()[-120:]), file=sys.stderr)
-sys.exit(1)
-PY
-    verifier_readiness_rc=$?
-    if [ "${verifier_readiness_rc}" -ne 0 ]; then
-        echo "ERROR: verifier API readiness check failed with rc=${verifier_readiness_rc}." >&2
-        if kill -0 "${verifier_api_pid}" 2>/dev/null; then
-            kill "${verifier_api_pid}" 2>/dev/null || true
-        fi
-        wait "${verifier_api_pid}" 2>/dev/null || true
-        exit "${verifier_readiness_rc}"
-    fi
-
-    echo "--> [Verifier API Node] Ready. Keeping node alive."
-    wait "${verifier_api_pid}"
-    verifier_api_rc=$?
-    if [ "${verifier_api_rc}" -ne 0 ]; then
-        echo "ERROR: verifier API exited with rc=${verifier_api_rc}; log=${verifier_api_log}" >&2
-    fi
-    exit "${verifier_api_rc}"
-fi
-
-export TMPDIR=/cache/ray_tmp
-export HCCL_ASYNC_ERROR_HANDLING=${HCCL_ASYNC_ERROR_HANDLING:-0}
-
-wait_for_ray_npu_resources() {
-    expected_npu=$1
-    timeout_seconds=${2:-900}
-    begin_ts=$(date +%s)
-
-    while true; do
-        total_npu=$(python3 - <<'PY' 2>/dev/null
-import ray
-
-try:
-    ray.init(address="auto", ignore_reinit_error=True, logging_level="ERROR")
-    print(int(ray.cluster_resources().get("NPU", 0)))
-    ray.shutdown()
-except Exception:
-    print(0)
-PY
-)
-        total_npu=${total_npu:-0}
-        now_ts=$(date +%s)
-        elapsed=$((now_ts - begin_ts))
-
-        echo "Ray NPU resources: ${total_npu}/${expected_npu}, elapsed=${elapsed}s"
-        ray status || true
-
-        if [ "${total_npu}" -ge "${expected_npu}" ]; then
-            echo "Ray cluster is ready: ${total_npu}/${expected_npu} NPU resources registered."
-            break
-        fi
-
-        if [ "${elapsed}" -ge "${timeout_seconds}" ]; then
-            echo "ERROR: Timed out waiting for Ray NPU resources: ${total_npu}/${expected_npu}" >&2
-            return 1
-        fi
-
-        sleep 5
-    done
-}
-
-wait_for_verifier_api() {
-    api_url="${mock_api_base}/models"
-    timeout_seconds=${VERIFIER_API_CLIENT_WAIT_TIMEOUT:-1800}
-    begin_ts=$(date +%s)
-    last_diag_ts=0
-    while true; do
-        verifier_check_output=$(python3 - "${api_url}" <<'PY' 2>&1
-import socket
-import sys
-import urllib.parse
-import urllib.request
-
-url = sys.argv[1]
-parsed = urllib.parse.urlparse(url)
-host = parsed.hostname
-port = parsed.port or (443 if parsed.scheme == "https" else 80)
-print(f"check url={url} host={host} port={port}")
-try:
-    infos = socket.getaddrinfo(host, port, type=socket.SOCK_STREAM)
-    print("dns=" + ",".join(sorted({item[4][0] for item in infos})))
-except Exception as exc:
-    print(f"dns_error={type(exc).__name__}: {exc}")
-    raise SystemExit(1)
-try:
-    with socket.create_connection((host, port), timeout=5):
-        print("tcp=ok")
-except Exception as exc:
-    print(f"tcp_error={type(exc).__name__}: {exc}")
-    raise SystemExit(1)
-try:
-    with urllib.request.urlopen(url, timeout=10) as response:
-        print(f"http_status={response.status}")
-        raise SystemExit(0 if 200 <= response.status < 300 else 1)
-except Exception as exc:
-    print(f"http_error={type(exc).__name__}: {exc}")
-    raise SystemExit(1)
-PY
-)
-        check_rc=$?
-        if [ "${check_rc}" = "0" ]; then
-            echo "Verifier API is ready: ${api_url}"
-            echo "${verifier_check_output}"
-            break
-        fi
-        now_ts=$(date +%s)
-        elapsed=$((now_ts - begin_ts))
-        echo "Waiting for verifier API: ${api_url}, elapsed=${elapsed}s"
-        if [ $((now_ts - last_diag_ts)) -ge 60 ]; then
-            last_diag_ts=${now_ts}
-            echo "--- verifier API check diagnostics ---"
-            echo "${verifier_check_output}"
-            echo "--- expected verifier node: rank=${verifier_api_node_rank}, host=${verifier_api_host}, port=${verifier_api_port} ---"
-            echo "--- check verifier node log: ${verifier_api_log} ---"
-            echo "--------------------------------------"
-        fi
-        if [ "${elapsed}" -ge "${timeout_seconds}" ]; then
-            echo "ERROR: Timed out waiting for verifier API: ${api_url}" >&2
-            echo "Last verifier API diagnostics:" >&2
-            echo "${verifier_check_output}" >&2
-            return 1
-        fi
-        sleep 10
-    done
-}
-
-# ================= Nanoclaw workspace 根目录 =================
-mkdir -p "${nanoclaw_temp_root}"
-if ! touch "${nanoclaw_temp_root}/.nanoclaw_write_test_${NODE_RANK}" 2>/dev/null; then
-    echo "ERROR: Cannot write NANOCLAW_TEMP_ROOT: ${nanoclaw_temp_root}" >&2
-    exit 2
-fi
-rm -f "${nanoclaw_temp_root}/.nanoclaw_write_test_${NODE_RANK}" || true
-if [[ "${nanoclaw_temp_root}" == /tmp/* ]]; then
-    echo "WARNING: NANOCLAW_TEMP_ROOT is under /tmp. Multi-node reward workers may not see rollout workspaces." >&2
-    echo "WARNING: Prefer a shared path, e.g. ${DATA_ROOT}/nanoclaw_workspaces" >&2
-fi
-echo "DEBUG: Nanoclaw train_base_tasks=${train_base_tasks}"
-echo "DEBUG: Nanoclaw val_base_tasks=${val_base_tasks}"
-echo "DEBUG: Nanoclaw task_glob=${nanoclaw_task_glob}, task_ids=${nanoclaw_task_ids:-<all>}"
-echo "DEBUG: Nanoclaw temp_root=${nanoclaw_temp_root}, cleanup=${nanoclaw_cleanup_workspaces}, keep_failed=${nanoclaw_keep_failed_workspaces}"
-
-# ================= 生成 Ray runtime env =================
-RUNTIME_ENV_FILE=${WORK_DIR}/verl_engine_runtime_env.generated.yaml
-cat > "${RUNTIME_ENV_FILE}" <<YAML
-working_dir: ./
-excludes: ["/.git/", "/logs/", "/checkpoint/"]
-env_vars:
-  TORCH_NCCL_AVOID_RECORD_STREAMS: "1"
-  CUDA_DEVICE_MAX_CONNECTIONS: "1"
-  HCCL_HOST_SOCKET_PORT_RANGE: "${HCCL_HOST_SOCKET_PORT_RANGE}"
-  HCCL_NPU_SOCKET_PORT_RANGE: "${HCCL_NPU_SOCKET_PORT_RANGE}"
-  HCCL_CONNECT_TIMEOUT: "${HCCL_CONNECT_TIMEOUT}"
-  HCCL_EXEC_TIMEOUT: "${HCCL_EXEC_TIMEOUT}"
-  HCCL_EVENT_TIMEOUT: "${HCCL_EVENT_TIMEOUT}"
-  HCCL_LOG_LEVEL: "${HCCL_LOG_LEVEL}"
-  HCCL_BUFFSIZE: "${HCCL_BUFFSIZE}"
-  P2P_HCCL_BUFFSIZE: "${P2P_HCCL_BUFFSIZE}"
-  VLLM_USE_V1: "${VLLM_USE_V1}"
-  VLLM_ENABLE_GRAPH_MODE: "${verifier_api_enable_graph_mode}"
-  VLLM_ASCEND_ENABLE_NZ: "${VLLM_ASCEND_ENABLE_NZ}"
-  VLLM_ENABLE_V1_MULTIPROCESSING: "${VLLM_ENABLE_V1_MULTIPROCESSING}"
-  VLLM_ENGINE_ITERATION_TIMEOUT_S: "${VLLM_ENGINE_ITERATION_TIMEOUT_S}"
-  RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES: "${RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES}"
-  TOKENIZERS_PARALLELISM: "false"
-  HYDRA_FULL_ERROR: "1"
-  PYTHONUNBUFFERED: "1"
-  RAY_DEDUP_LOGS: "0"
-  WANDB_MODE: "${WANDB_MODE}"
-  MOCK_API_BASE: "${mock_api_base}"
-  MOCK_API_KEY: "${mock_api_key}"
-  MOCK_MODEL_NAME: "${mock_model_name}"
-  MOCK_API_TIMEOUT: "${mock_api_timeout}"
-  MOCK_API_CONNECT_TIMEOUT: "${mock_api_connect_timeout}"
-  NANOCLAW_FORCE_NO_THINKING: "${nanoclaw_force_no_thinking}"
-  NANOCLAW_FORCE_MAX_TOKENS: "${nanoclaw_force_max_tokens}"
-  NANOCLAW_REWARD_PRINT_DETAILS: "${nanoclaw_reward_print_details}"
-  NANOCLAW_REQUIRE_FINAL_ANSWER: "${nanoclaw_require_final_answer}"
-  NANOCLAW_FINAL_ANSWER_BONUS_ENABLE: "${nanoclaw_final_answer_bonus_enable}"
-  NANOCLAW_FINAL_ANSWER_BONUS_SCORE: "${nanoclaw_final_answer_bonus_score}"
-  NANOCLAW_TURN_PENALTY_ONLY_POSITIVE_SCORE: "${nanoclaw_turn_penalty_only_positive_score}"
-  NANOCLAW_ASSISTANT_TURN_PENALTY: "${nanoclaw_assistant_turn_penalty}"
-  NANOCLAW_DUPLICATE_TOOL_CALL_PENALTY: "${nanoclaw_duplicate_tool_call_penalty}"
-  NANOCLAW_REPEATED_RESPONSE_PENALTY: "${nanoclaw_repeated_response_penalty}"
-  NANOCLAW_REPEATED_RESPONSE_MIN_CHARS: "${nanoclaw_repeated_response_min_chars}"
-  NANOCLAW_REPEATED_RESPONSE_MIN_CONSECUTIVE_REPEATS: "${nanoclaw_repeated_response_min_consecutive_repeats}"
-  NANOCLAW_MASK_LOOPING_RESPONSES: "${nanoclaw_mask_looping_responses}"
-  NANOCLAW_MASK_ONLY_POSITIVE_ADVANTAGE: "${nanoclaw_mask_only_positive_advantage}"
-  NANOCLAW_MASK_BUDGET_EXHAUSTED_LAST_TURN: "${nanoclaw_mask_budget_exhausted_last_turn}"
-  NANOCLAW_MASK_DUPLICATE_TOOL_RESULT_TURNS: "${nanoclaw_mask_duplicate_tool_result_turns}"
-  NANOCLAW_MASK_ERROR_TOOL_RESULT_TURNS: "${nanoclaw_mask_error_tool_result_turns}"
-  NANOCLAW_LOOPING_RESPONSE_MIN_CHARS: "${nanoclaw_looping_response_min_chars}"
-  NANOCLAW_LOOPING_RESPONSE_MIN_CONSECUTIVE_REPEATS: "${nanoclaw_looping_response_min_consecutive_repeats}"
-YAML
-
-# ================= 启动 Ray 多机集群 =================
-if [ "${NODE_RANK}" = "0" ]; then
-    echo "--> [Head Node] Starting Ray Head on ${CURRENT_IP}..."
-    ray start --head \
-        --node-ip-address=${RAY_NODE_IP} \
-        --port=${RAY_PORT} \
-        --dashboard-host=0.0.0.0 \
-        --dashboard-port=${DASHBOARD_PORT} \
-        --resources="{\"NPU\":${NPUS_PER_NODE}}" \
-        --disable-usage-stats \
-        --block &
-
-    sleep 10
-    wait_for_ray_npu_resources ${WORLD_SIZE} 900 || exit 1
-    wait_for_verifier_api || exit 1
-else
-    echo "--> [Worker Node] Starting Ray Worker, connecting to ${MASTER_ADDR}:${RAY_PORT}..."
-    sleep 20
-    ray start --address=${MASTER_ADDR}:${RAY_PORT} \
-        --node-ip-address=${RAY_NODE_IP} \
-        --resources="{\"NPU\":${NPUS_PER_NODE}}" \
-        --disable-usage-stats \
-        --block &
-    sleep 10
-fi
-
-# ================= 训练参数数组 =================
-training_args=(
-    python3 -m verl.trainer.main_ppo
-    +ray_kwargs.ray_init.address=auto
-    reward.num_workers=${reward_num_workers}
-    algorithm.adv_estimator=${adv_estimator}
-    algorithm.gamma=${algorithm_gamma}
-    algorithm.lam=${algorithm_lam}
-    algorithm.use_kl_in_reward=${use_kl_in_reward}
-    algorithm.kl_penalty=${kl_penalty}
-    algorithm.kl_ctrl.type=${kl_ctrl_type}
-    algorithm.kl_ctrl.kl_coef=${kl_coef}
-    algorithm.rollout_correction.bypass_mode=False
-    data.train_files="${train_files}"
-    data.val_files="${test_files}"
-    data.return_raw_chat=True
-    data.return_multi_modal_inputs=False
-    data.image_key=images
-    data.shuffle=True
-    data.train_batch_size=${train_batch_size}
-    data.max_prompt_length=${max_prompt_length}
-    data.max_response_length=${max_response_length}
-    data.filter_overlong_prompts=True
-    data.truncation=error
-    data.custom_cls.path=pkg://nanoclaw_recipe.nanoclaw
-    data.custom_cls.name=CustomRLHFDataset
-    "data.tool_config_path=${tool_config_path}"
-    "+data.nanoclaw_task_glob=${nanoclaw_task_glob}"
-    "+data.nanoclaw_temp_root=${nanoclaw_temp_root}"
-    "+data.nanoclaw_cleanup_workspaces=${nanoclaw_cleanup_workspaces}"
-    "+data.nanoclaw_keep_failed_workspaces=${nanoclaw_keep_failed_workspaces}"
-    "+data.nanoclaw_env_builder_timeout=${nanoclaw_env_builder_timeout}"
-    "+data.nanoclaw_verifier_timeout=${nanoclaw_verifier_timeout}"
-    "+data.nanoclaw_reward_score_mode=${nanoclaw_reward_score_mode}"
-    "+data.nanoclaw_allow_bash=${nanoclaw_allow_bash}"
-    +data.apply_chat_template_kwargs.enable_thinking=True
-    reward.custom_reward_function.path=pkg://nanoclaw_recipe.nanoclaw
-    reward.custom_reward_function.name=compute_score
-    "+reward.custom_reward_function.reward_kwargs.cleanup_workspaces=${nanoclaw_cleanup_workspaces}"
-    "+reward.custom_reward_function.reward_kwargs.keep_failed_workspaces=${nanoclaw_keep_failed_workspaces}"
-    "+reward.custom_reward_function.reward_kwargs.verifier_timeout=${nanoclaw_verifier_timeout}"
-    "+reward.custom_reward_function.reward_kwargs.reward_score_mode=${nanoclaw_reward_score_mode}"
-    "+reward.custom_reward_function.reward_kwargs.require_final_answer=${nanoclaw_require_final_answer}"
-    "+reward.custom_reward_function.reward_kwargs.final_answer_bonus_enable=${nanoclaw_final_answer_bonus_enable}"
-    "+reward.custom_reward_function.reward_kwargs.final_answer_bonus_score=${nanoclaw_final_answer_bonus_score}"
-    "+reward.custom_reward_function.reward_kwargs.turn_penalty_only_positive_score=${nanoclaw_turn_penalty_only_positive_score}"
-    "+reward.custom_reward_function.reward_kwargs.assistant_turn_penalty=${nanoclaw_assistant_turn_penalty}"
-    "+reward.custom_reward_function.reward_kwargs.duplicate_tool_call_penalty=${nanoclaw_duplicate_tool_call_penalty}"
-    "+reward.custom_reward_function.reward_kwargs.repeated_response_penalty=${nanoclaw_repeated_response_penalty}"
-    "+reward.custom_reward_function.reward_kwargs.repeated_response_min_chars=${nanoclaw_repeated_response_min_chars}"
-    "+reward.custom_reward_function.reward_kwargs.repeated_response_min_consecutive_repeats=${nanoclaw_repeated_response_min_consecutive_repeats}"
-    "+reward.custom_reward_function.reward_kwargs.mock_api_base=${mock_api_base}"
-    "+reward.custom_reward_function.reward_kwargs.mock_api_key=${mock_api_key}"
-    "+reward.custom_reward_function.reward_kwargs.mock_model_name=${mock_model_name}"
-    "+reward.custom_reward_function.reward_kwargs.mock_api_timeout=${mock_api_timeout}"
-    "+reward.custom_reward_function.reward_kwargs.mock_api_connect_timeout=${mock_api_connect_timeout}"
-    actor_rollout_ref.model.path=${model_path}
-    actor_rollout_ref.model.use_remove_padding=True
-    actor_rollout_ref.model.enable_gradient_checkpointing=True
-    actor_rollout_ref.model.enable_activation_offload=${enable_activation_offload}
-    actor_rollout_ref.model.use_fused_kernels=${use_fused_kernels}
-    actor_rollout_ref.model.fused_kernel_options.impl_backend=${fused_kernel_backend}
-    actor_rollout_ref.actor.strategy=${actor_strategy}
-    actor_rollout_ref.ref.strategy=${actor_strategy}
-    actor_rollout_ref.actor.use_kl_loss=${actor_use_kl_loss}
-    actor_rollout_ref.actor.kl_loss_coef=${actor_kl_loss_coef}
-    actor_rollout_ref.actor.kl_loss_type=${actor_kl_loss_type}
-    actor_rollout_ref.actor.clip_ratio_low=${actor_clip_ratio_low}
-    actor_rollout_ref.actor.clip_ratio_high=${actor_clip_ratio_high}
-    actor_rollout_ref.actor.clip_ratio_c=${actor_clip_ratio_c}
-    actor_rollout_ref.actor.entropy_coeff=${actor_entropy_coeff}
-    actor_rollout_ref.actor.ppo_epochs=${actor_ppo_epochs}
-    actor_rollout_ref.actor.shuffle=${actor_shuffle}
-    actor_rollout_ref.actor.optim.lr=${actor_lr}
-    actor_rollout_ref.actor.optim.lr_scheduler_type=${actor_lr_scheduler_type}
-    actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=${actor_lr_warmup_steps_ratio}
-    actor_rollout_ref.actor.optim.weight_decay=${actor_weight_decay}
-    "actor_rollout_ref.actor.optim.betas=[${actor_adam_beta1},${actor_adam_beta2}]"
-    actor_rollout_ref.actor.optim.clip_grad=${actor_clip_grad}
-    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.ulysses_sequence_parallel_size=${train_sp}
-    actor_rollout_ref.actor.entropy_from_logits_with_chunking=${entropy_from_logits_with_chunking}
-    actor_rollout_ref.actor.entropy_from_logits_chunk_size=${entropy_from_logits_chunk_size}
-    actor_rollout_ref.actor.entropy_checkpointing=${entropy_checkpointing}
-    actor_rollout_ref.actor.fsdp_config.param_offload=${offload}
-    actor_rollout_ref.actor.fsdp_config.optimizer_offload=${offload}
-    actor_rollout_ref.actor.fsdp_config.entropy_from_logits_with_chunking=${entropy_from_logits_with_chunking}
-    actor_rollout_ref.actor.fsdp_config.entropy_from_logits_chunk_size=${entropy_from_logits_chunk_size}
-    actor_rollout_ref.actor.fsdp_config.entropy_checkpointing=${entropy_checkpointing}
-    actor_rollout_ref.ref.fsdp_config.param_offload=${offload}
-    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=True
-    actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${log_prob_max_token_len_per_gpu}
-    actor_rollout_ref.ref.ulysses_sequence_parallel_size=${train_sp}
-    actor_rollout_ref.ref.entropy_from_logits_with_chunking=${entropy_from_logits_with_chunking}
-    actor_rollout_ref.ref.entropy_from_logits_chunk_size=${entropy_from_logits_chunk_size}
-    actor_rollout_ref.ref.entropy_checkpointing=${entropy_checkpointing}
-    actor_rollout_ref.ref.fsdp_config.entropy_from_logits_with_chunking=${entropy_from_logits_with_chunking}
-    actor_rollout_ref.ref.fsdp_config.entropy_from_logits_chunk_size=${entropy_from_logits_chunk_size}
-    actor_rollout_ref.ref.fsdp_config.entropy_checkpointing=${entropy_checkpointing}
-    actor_rollout_ref.rollout.name=vllm
-    actor_rollout_ref.rollout.mode=async
-    actor_rollout_ref.rollout.calculate_log_probs=True
-    actor_rollout_ref.rollout.temperature=${rollout_temperature}
-    actor_rollout_ref.rollout.top_p=${rollout_top_p}
-    actor_rollout_ref.rollout.top_k=${rollout_top_k}
-    actor_rollout_ref.rollout.min_p=${rollout_min_p}
-    actor_rollout_ref.rollout.presence_penalty=${rollout_presence_penalty}
-    actor_rollout_ref.rollout.frequency_penalty=${rollout_frequency_penalty}
-    actor_rollout_ref.rollout.repetition_penalty=${rollout_repetition_penalty}
-    actor_rollout_ref.rollout.tensor_model_parallel_size=${infer_tp}
-    actor_rollout_ref.rollout.checkpoint_engine.update_weights_bucket_megabytes=${update_weights_bucket_mb}
-    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=True
-    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${log_prob_max_token_len_per_gpu}
-    actor_rollout_ref.rollout.enable_chunked_prefill=True
-    actor_rollout_ref.rollout.max_num_batched_tokens=${rollout_max_num_batched_tokens}
-    actor_rollout_ref.rollout.free_cache_engine=True
-    actor_rollout_ref.rollout.enforce_eager=False
-    actor_rollout_ref.rollout.enable_prefix_caching=False
-    actor_rollout_ref.rollout.multi_turn.enable=True
-    actor_rollout_ref.rollout.multi_turn.max_user_turns=${max_turns}
-    actor_rollout_ref.rollout.multi_turn.max_assistant_turns=${max_turns}
-    actor_rollout_ref.rollout.multi_turn.max_assistant_response_length=${max_assistant_response_length}
-    "actor_rollout_ref.rollout.multi_turn.tool_config_path=${tool_config_path}"
-    actor_rollout_ref.rollout.multi_turn.format=qwen3_coder
-    "actor_rollout_ref.rollout.multi_turn.max_tool_response_length=${max_tool_response_length}"
-    actor_rollout_ref.rollout.gpu_memory_utilization=${rollout_gpu_memory_utilization}
-    actor_rollout_ref.rollout.n=${n_resp_per_prompt}
-    actor_rollout_ref.rollout.val_kwargs.temperature=${rollout_temperature}
-    actor_rollout_ref.rollout.val_kwargs.top_p=${rollout_top_p}
-    actor_rollout_ref.rollout.val_kwargs.top_k=${rollout_top_k}
-    actor_rollout_ref.rollout.val_kwargs.min_p=${rollout_min_p}
-    actor_rollout_ref.rollout.val_kwargs.presence_penalty=${rollout_presence_penalty}
-    actor_rollout_ref.rollout.val_kwargs.frequency_penalty=${rollout_frequency_penalty}
-    actor_rollout_ref.rollout.val_kwargs.repetition_penalty=${rollout_repetition_penalty}
-    actor_rollout_ref.rollout.val_kwargs.do_sample=True
-    actor_rollout_ref.rollout.val_kwargs.n=${n_resp_per_prompt_val}
-    actor_rollout_ref.actor.use_torch_compile=False
-    actor_rollout_ref.ref.use_torch_compile=False
-    actor_rollout_ref.actor.use_torch_compile=False
-    actor_rollout_ref.ref.use_torch_compile=False
-    actor_rollout_ref.actor.fsdp_config.use_torch_compile=False
-    actor_rollout_ref.ref.fsdp_config.use_torch_compile=False
-    critic.fsdp.use_torch_compile=False
-    trainer.use_v1=${trainer_use_v1}
-    trainer.v1.trainer_mode=sync
-    trainer.critic_warmup=0
-    trainer.balance_batch=True
-    trainer.logger=['console','tensorboard']
-    trainer.project_name=${project_name}
-    trainer.experiment_name=${experiment_name}
-    trainer.nnodes=${NNODES}
-    trainer.n_gpus_per_node=${NPUS_PER_NODE}
-    trainer.val_before_train=${val_before_train}
-    trainer.log_val_generations=${log_val_generations}
-    trainer.save_freq=${save_freq}
-    trainer.default_local_dir=${default_local_dir}
-    trainer.resume_mode=disable
-    trainer.test_freq=${test_freq}
-    trainer.total_epochs=10
-    trainer.total_training_steps=${total_training_steps}
-)
-
-if [ -n "${fsdp_size}" ]; then
-    training_args+=(
-        actor_rollout_ref.actor.fsdp_config.fsdp_size=${fsdp_size}
-        actor_rollout_ref.ref.fsdp_config.fsdp_size=${fsdp_size}
-    )
-fi
-
-if [ -n "${nanoclaw_task_ids}" ]; then
-    training_args+=("+data.nanoclaw_task_ids=${nanoclaw_task_ids}")
-fi
-
-if [ -n "${nanoclaw_max_steps}" ]; then
-    training_args+=("+data.nanoclaw_max_steps=${nanoclaw_max_steps}")
-fi
-
-# ================= 启动训练主进程：仅主节点执行 =================
-if [ "${NODE_RANK}" = "0" ]; then
-    echo "--> [Head Node] Starting VERL unified engine training..."
-    echo "DEBUG: runtime_env=${RUNTIME_ENV_FILE}"
-    echo "DEBUG: entrypoint=${training_args[*]}"
-
-    ray job submit \
-        --address="http://127.0.0.1:${DASHBOARD_PORT}" \
-        --runtime-env="${RUNTIME_ENV_FILE}" \
-        -- \
-        "${training_args[@]}" 2>&1 | tee "logs/qwen3.5-nanoclaw-grpo-verl-engine-${start_time}.log"
-else
-    echo "--> [Worker Node] Setup finished. Keeping node alive for Ray..."
-    tail -f /dev/null
-fi