Skip to content

kali20gakki/msAgent / skills/accuracy/rl-prefill-consistency-dump

rl-prefill-consistency-dump

Verified

训推一致性(prefill)采集侧:为训练/推理两侧采集可比的 prefill dump(输入对齐 PROMPTS_ONLY → 参数对齐校验(dump 前前置:校验两侧 config 超参/结构/权重,不一致输出参数不一致点并对齐)→ msprobe 接入采集 → 采集验证 → 恢复代码)。按用户需求判断用法:① 仅采集 dump——用户只要 dump 数据,走采集主流程(0→7→9),直接产出两侧可比 dump;② 配合分析——用户要求找所有独立差异点/差异点隔离验证/配合 rl-consistency-dump-analysis 分析时,才按分析侧打桩指令用 msprobe `load` 覆盖边界二次采集(阶段 8)再恢复代码。差异分析、首差异定位、打桩指令、整体结论由 rl-consistency-dump-analysis 承担。触发词:训推一致、训推一致性、prefill 比对、msprobe dump/采集、精度比对数据采集、打桩二次采集、参数对齐校验、参数不一致点、mHC 配置不一致、权重不同步、清卡验环境后采数据;已采 dump 的分析转 rl-consistency-dump-analysis。

references/task_card.json

View on GitHub
{
  "model": "<模型名>(<权重路径,如 xxx 诊断权重/完整权重>)",
  "collection_scope": "确定性单样本 prefill (固定样本, seed 固定, TP=1, 关 shuffle/eval)",
  "evidence_tool": "msprobe PrecisionDebugger (level=mix)",
  "train_framework": "<训练框架,如 verl/slime/vime>",
  "train_framework_version": "<版本>",
  "train_backend": "<FSDP / Megatron / DeepSpeed / 自研>",
  "architecture_mode": "<显式 micro_batch 循环 (FSDP 型, 包循环体) / 连续前向 (Megatron 型, 插 forward 内部)>",
  "inference_engine": "<vLLM / SGLang / Transformers / 自研>",
  "inference_engine_version": "<版本>",
  "hardware": "<硬件后端,如 Ascend NPU> 卡 <M,N> (actor 1 + rollout 1)",
  "stop_condition": "采到两侧可比的 prefill dump (level=mix)",

  "stage3_onsite": {
    "crop_location": "<现场 grep 定位的 PROMPTS_ONLY 裁剪位置,以现场符号为准>",
    "instrument_location": "<现场定位的 log_prob 主前向接入位置;判断包循环体还是插 forward 内部>",
    "main_forward_gate": "<主前向是否被缓存跳过(类似 can_reuse_log_probs_in_loss),采不到数据先怀疑它>",
    "request_id": "<推理侧调度日志与训练 rollout 数据的对应字段;无现成日志就在接入处自己打>",
    "padding": "<去训练端 padding 的开关/做法,两侧输入 shape 对齐真实 prompt 长度 (P18)>",
    "infer_msprobe": "<引擎是否原生集成 msprobe(vLLM-ascend/SGLang 原生, 传 config 即启用);无原生集成按 new-framework.md §2 手写接入>"
  },

  "authorization": {
    "AUTH_STOP_RUNNING_SERVICES": false,
    "AUTH_EDIT_TARGET_CODE": true,
    "AUTH_RUN_DUMP_JOB": true,
    "AUTH_CLEAN_APPROVED_DIRS": true,
    "AUTH_RESTORE_CODE": true,
    "notes": "<按 SKILL 阶段 2 一次性问清并记录;停服务范围 / 允许改的代码 / 允许删的目录 / 采集后恢复,避免流程中反复打断>"
  },

  "environment": {
    "ssh_target": null,
    "container": "<容器标识,以现场为准(不预设容器形态,见框架线索卡对应说明)>",
    "conda_env": "<环境激活方式>",
    "workdir": "<工作目录,绝对路径>",
    "log_path": "<日志位置>",
    "artifact_root": "<会话/产物根目录>",
    "credentials_note": "凭据不写入本卡或报告"
  },

  "assets": {
    "model_path": "<权重路径>",
    "baseline_command": "<基线命令,绝对路径>",
    "train_msprobe_config": "<session>/train_dump_config.json (task=statistics, level=mix)",
    "infer_msprobe_config": "<session>/infer_dump_config.json (task=statistics, level=mix)",
    "train_dump_path": "<session>/dumps/train",
    "infer_dump_path": "<session>/dumps/infer"
  },

  "constraints": {
    "padding": "训练端 padding 必须去掉, 两侧输入 shape 对齐真实 prompt 长度 (P18)",
    "determinism": "seed 固定, 关 shuffle, 单样本小 batch (num-rollout 1 / rollout-batch-size 1 / n-samples-per-prompt 1 / rollout-max-response-len 4 / rollout-temperature 0 / global-batch-size 1), enforce_eager",
    "e2e_target": "response 段 E2E 指标(PROMPTS_ONLY 下为 0 值是'没算', 需补跑 PROMPTS_ONLY=0 标准 step 才有实测)",
    "prefill_logprob_diff": "训练循环内直接算并打印训练日志 (主指标, mean≈0 是训推一致)"
  }
}
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall