inference_clawbenchpro.sh

#!/bin/bash
# Qwen3.5-27B half-turn checkpoint 在 ClawBenchPro 高质量子集上的训练同构推理。
#
# 唯一 rollout 链路：
#   CustomRLHFDataset
#     -> VERL LLMServerManager/vLLM
#     -> VERL AgentLoopManager
#     -> tool_agent / ToolAgentLoop
#     -> NanoclawWorkspaceTool
#
# 该脚本不实现第二套 Agent，不调用 verifier/reward。它复用训练的 system
# prompt、Qwen3-Coder XML 工具协议、9 个 workspace tools、完整多轮历史拼接、
# workspace 生命周期、response mask 与 trajectory 持久化。

set -x

SCRIPT_DIR=/opt/huawei/dataset/zyr_yuyin/lyf/datasets/testClawBenchPro/upload_clawbenchpro_base100_hard100_npu/v14/0708_new
BUNDLE_ROOT=/opt/huawei/dataset/zyr_yuyin/lyf/datasets/testClawBenchPro/upload_clawbenchpro_base100_hard100_npu

# ==============================================================================
# 直接在这里填写要推理的多个模型。每项是一个普通 Bash 字符串：
#   "唯一模型输出名|已合并 Hugging Face checkpoint 的绝对路径"
#
# 示例（删除行首 # 后改成实际路径）：
MODEL_CHECKPOINTS=(
    "qwen35_4b_base|/opt/huawei/dataset/zyr_yuyin/models/Qwen/Qwen3___5-4B"
    "qwen35_4b_step_2|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_2"
    "qwen35_4b_step_4|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_4"
    "qwen35_4b_step_6|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_6"
    "qwen35_4b_step_8|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_8"
    "qwen35_4b_step_10|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_10"
    "qwen35_4b_step_12|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_12"
    "qwen35_4b_step_14|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_14"
    "qwen35_4b_step_16|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_16"
    "qwen35_4b_step_18|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_18"
    "qwen35_4b_step_20|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_20"
    "qwen35_4b_step_22|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_22"
    "qwen35_4b_step_24|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_24"
    "qwen35_4b_step_26|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_26"
    "qwen35_4b_step_28|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_28"
    "qwen35_4b_step_30|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_30"
    "qwen35_4b_step_32|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_32"
    "qwen35_4b_step_34|/opt/huawei/dataset/zyr_yuyin/lyf/verl-nanoclaw-rl/nanoclawRL_temp_ckpt_hugging_face/qwen3.5-4b_31k/qwen3.5-4b_31k_step_34"
)
# ==============================================================================

# 多 checkpoint 输入，按以下优先级解析：
#   1. MODEL_PATH_LIST（换行分隔）及可选 MODEL_NAME_LIST；
#   2. 兼容旧用法的单个 MODEL_PATH / MODEL_NAME；
#   3. 上面的 MODEL_CHECKPOINTS 字符串数组（推荐日常使用）。
# 所有路径都必须是 vLLM 可直接加载的、已合并 Hugging Face checkpoint；
# 未合并的 VERL/FSDP shard 不能直接用于该推理入口。
export MODEL_INPUT_VALIDATE_ONLY=${MODEL_INPUT_VALIDATE_ONLY:-0}

declare -a INPUT_MODEL_PATHS=()
declare -a INPUT_MODEL_NAMES=()

if [ -n "${MODEL_PATH_LIST:-}" ]; then
    while IFS= read -r model_path; do
        model_path=${model_path%$'\r'}
        if [ -n "${model_path}" ]; then
            INPUT_MODEL_PATHS+=("${model_path}")
        fi
    done <<< "${MODEL_PATH_LIST}"
    if [ -n "${MODEL_NAME_LIST:-}" ]; then
        while IFS= read -r model_name; do
            model_name=${model_name%$'\r'}
            if [ -n "${model_name}" ]; then
                INPUT_MODEL_NAMES+=("${model_name}")
            fi
        done <<< "${MODEL_NAME_LIST}"
    fi
elif [ -n "${MODEL_PATH:-}" ]; then
    INPUT_MODEL_PATHS+=("${MODEL_PATH}")
    if [ -n "${MODEL_NAME:-}" ]; then
        INPUT_MODEL_NAMES+=("${MODEL_NAME}")
    fi
elif [ "${#MODEL_CHECKPOINTS[@]}" -gt 0 ]; then
    for model_spec in "${MODEL_CHECKPOINTS[@]}"; do
        if [[ "${model_spec}" != *"|"* ]]; then
            echo "ERROR: invalid MODEL_CHECKPOINTS item; expected \"model_name|/absolute/checkpoint/path\": ${model_spec}" >&2
            exit 2
        fi
        model_name=${model_spec%%|*}
        model_path=${model_spec#*|}
        if [ -z "${model_name}" ] || [ -z "${model_path}" ] || [[ "${model_path}" == *"|"* ]]; then
            echo "ERROR: invalid MODEL_CHECKPOINTS item; expected exactly one | delimiter: ${model_spec}" >&2
            exit 2
        fi
        INPUT_MODEL_NAMES+=("${model_name}")
        INPUT_MODEL_PATHS+=("${model_path}")
    done
fi

if [ "${#INPUT_MODEL_PATHS[@]}" -eq 0 ]; then
    echo "ERROR: no model checkpoints configured." >&2
    echo "Edit MODEL_CHECKPOINTS at the top of this script, set MODEL_PATH_LIST, or set legacy MODEL_PATH." >&2
    exit 2
fi
if [ "${#INPUT_MODEL_NAMES[@]}" -ne 0 ] && [ "${#INPUT_MODEL_NAMES[@]}" -ne "${#INPUT_MODEL_PATHS[@]}" ]; then
    echo "ERROR: model name count ${#INPUT_MODEL_NAMES[@]} does not match path count ${#INPUT_MODEL_PATHS[@]}." >&2
    exit 2
fi

declare -A INPUT_MODEL_NAME_SEEN=()
for model_index in "${!INPUT_MODEL_PATHS[@]}"; do
    model_path=${INPUT_MODEL_PATHS[model_index]}
    if [[ "${model_path}" != /* ]]; then
        echo "ERROR: model checkpoint path must be absolute: ${model_path}" >&2
        exit 2
    fi
    if [ ! -d "${model_path}" ]; then
        echo "ERROR: model checkpoint directory not found: ${model_path}" >&2
        exit 2
    fi
    if [ ! -f "${model_path}/config.json" ]; then
        echo "ERROR: merged Hugging Face config.json not found: ${model_path}/config.json" >&2
        exit 2
    fi
    if [ "${#INPUT_MODEL_NAMES[@]}" -gt 0 ]; then
        model_name=${INPUT_MODEL_NAMES[model_index]}
        if [[ ! "${model_name}" =~ ^[a-zA-Z0-9._-]+$ ]]; then
            echo "ERROR: model name may only contain letters, digits, dot, underscore and hyphen: ${model_name}" >&2
            exit 2
        fi
        if [ -n "${INPUT_MODEL_NAME_SEEN[${model_name}]:-}" ]; then
            echo "ERROR: duplicate configured model name: ${model_name}" >&2
            exit 2
        fi
        INPUT_MODEL_NAME_SEEN[${model_name}]=1
    else
        model_name='<auto>'
    fi
    echo "MODEL_INPUT[$model_index] name=${model_name} path=${model_path}"
done

printf -v MODEL_PATH_LIST_NORMALIZED '%s\n' "${INPUT_MODEL_PATHS[@]}"
export MODEL_PATH_LIST=${MODEL_PATH_LIST_NORMALIZED}
if [ "${#INPUT_MODEL_NAMES[@]}" -gt 0 ]; then
    printf -v MODEL_NAME_LIST_NORMALIZED '%s\n' "${INPUT_MODEL_NAMES[@]}"
    export MODEL_NAME_LIST=${MODEL_NAME_LIST_NORMALIZED}
else
    unset MODEL_NAME_LIST
fi
unset MODEL_PATH MODEL_NAME

echo "MODEL_INPUT_COUNT=${#INPUT_MODEL_PATHS[@]}"
if [ "${MODEL_INPUT_VALIDATE_ONLY}" = "1" ]; then
    echo "MODEL_INPUT_VALIDATE_ONLY=1: model configuration is valid; inference not started."
    exit 0
fi

# 必须把训练时修改过的整份 VERL v12 代码同步到此目录；不能只安装上游 VERL。
export WORK_DIR=${WORK_DIR:-${BUNDLE_ROOT}/verl}

# 已离线筛选并适配好的训练兼容数据：base 100 + hard 100。这里直接读取
# manifest-backed data_* bundle，不在推理节点重新扫描或适配完整 1000 题数据。
# 部署到共享存储后，可通过 BASE_TASKS 覆盖为共享目录中的副本。
export BASE_TASKS=${BASE_TASKS:-${BUNDLE_ROOT}/data/ClawBenchPro_base100_hard100_quality}

# v14/0710/inference.sh 只有在该变量非空时才会启动全量 ClawBenchPro 适配器。
# 专用入口固定使用上面的精选子集，避免意外退回 991/1000 题路径。
export CLAWBENCHPRO_ROOT=
export CLAWBENCHPRO_ADAPTED_ROOT=

# 输出严格为 OUTPUT_ROOT/<model_name>/step_1/<task_id>_sample_<n>/。
export OUTPUT_ROOT=${OUTPUT_ROOT:-/opt/huawei/dataset/zyr_yuyin/lyf/datasets/testClawBenchPro/output/qwen35_4b}
export OVERWRITE_OUTPUT=${OVERWRITE_OUTPUT:-True}
export CONTINUE_ON_MODEL_ERROR=${CONTINUE_ON_MODEL_ERROR:-1}
export MODEL_SWITCH_COOLDOWN=${MODEL_SWITCH_COOLDOWN:-20}
export MODEL_RESOURCE_RELEASE_TIMEOUT=${MODEL_RESOURCE_RELEASE_TIMEOUT:-600}

# 默认单机 8 NPU、TP=4，即 2 个 vLLM rollout replica。多机时所有节点提交
# 同一脚本，ModelArts 通过 VC_TASK_INDEX 区分 rank。
export INFER_NNODES=${INFER_NNODES:-1}
export NPUS_PER_NODE=${NPUS_PER_NODE:-8}
export INFER_TP=${INFER_TP:-4}

# 评测默认每题 1 条轨迹；如需复现训练时的 GRPO 采样数量可设为 8。
export N_RESP_PER_PROMPT=${N_RESP_PER_PROMPT:-1}
export PROMPT_BATCH_SIZE=${PROMPT_BATCH_SIZE:-64}
export AGENT_NUM_WORKERS=${AGENT_NUM_WORKERS:-128}
export CALCULATE_LOG_PROBS=${CALCULATE_LOG_PROBS:-False}

# 与 v14/0708_new/half_turn.sh 的 actor rollout 完全对齐。
export MAX_TURNS=${MAX_TURNS:-35}
export MAX_PROMPT_LENGTH=${MAX_PROMPT_LENGTH:-8192}
export MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-22768}
export MAX_ASSISTANT_RESPONSE_LENGTH=${MAX_ASSISTANT_RESPONSE_LENGTH:-16384}
export MAX_TOOL_RESPONSE_LENGTH=${MAX_TOOL_RESPONSE_LENGTH:-8192}
export ROLLOUT_MAX_NUM_BATCHED_TOKENS=${ROLLOUT_MAX_NUM_BATCHED_TOKENS:-16384}
export ROLLOUT_GPU_MEMORY_UTILIZATION=${ROLLOUT_GPU_MEMORY_UTILIZATION:-0.70}
export ROLLOUT_TEMPERATURE=${ROLLOUT_TEMPERATURE:-1.0}
export ROLLOUT_TOP_P=${ROLLOUT_TOP_P:-0.95}
export ROLLOUT_TOP_K=${ROLLOUT_TOP_K:-20}
export ROLLOUT_MIN_P=${ROLLOUT_MIN_P:-0.0}
export ROLLOUT_PRESENCE_PENALTY=${ROLLOUT_PRESENCE_PENALTY:-0.0}
export ROLLOUT_FREQUENCY_PENALTY=${ROLLOUT_FREQUENCY_PENALTY:-0.0}
export ROLLOUT_REPETITION_PENALTY=${ROLLOUT_REPETITION_PENALTY:-1.0}
export ROLLOUT_FREE_CACHE_ENGINE=${ROLLOUT_FREE_CACHE_ENGINE:-True}
export ROLLOUT_ENFORCE_EAGER=${ROLLOUT_ENFORCE_EAGER:-False}

# 与训练相同：thinking actor、Qwen3-Coder parser、同一 tool YAML、受限 bash、
# 不保存 workspace_before，并严格禁止输出路径静默追加 request-id 后缀。
export TOOL_CONFIG_PATH=${TOOL_CONFIG_PATH:-recipe/nanoclaw/nanoclaw_tool_config.yaml}
export NANOCLAW_MAX_STEPS=${NANOCLAW_MAX_STEPS:-}
export NANOCLAW_CLEANUP_WORKSPACES=False
export NANOCLAW_KEEP_FAILED_WORKSPACES=False
export NANOCLAW_ENV_BUILDER_TIMEOUT=${NANOCLAW_ENV_BUILDER_TIMEOUT:-120}
export NANOCLAW_ALLOW_BASH=True
export NANOCLAW_STRICT_RESULT_DIR=True
export NANOCLAW_SAVE_WORKSPACE_BEFORE=False

# 默认安装与训练一致的 GCC/CANN/torch-npu/vLLM/Triton/VERL 依赖。
export SETUP_ENVIRONMENT=${SETUP_ENVIRONMENT:-1}

exec bash "/opt/huawei/dataset/zyr_yuyin/lyf/datasets/testClawBenchPro/upload_clawbenchpro_base100_hard100_npu/v14/0710/inference.sh"