Skip to content

[DEV P0 AUDIT · DO NOT MERGE/CLOSE] Living correctness and real-hardware test ledger · master #61 - #75

Draft
hydrogen1222 wants to merge 1 commit into
mainfrom
audit/living-correctness-and-test-matrix-20261031
Draft

hydrogen1222 wants to merge 1 commit into
mainfrom
audit/living-correctness-and-test-matrix-20261031

Conversation

@hydrogen1222

@hydrogen1222 hydrogen1222 commented Oct 31, 2026 •

Copy link
Copy Markdown
Owner

状态指针 · 2026-11-18(Rev 10 第二阶段):基线 main = fb2935c730118bcaa86ede6154ac10c512f46677。多材料 / 全工作流 / 无人值守验收结果见 Rev 10 · 第二阶段执行结果;修复 PR #110(P1:numpy bool 未收敛被记为成功)、#111(P2:R10-F01 NPT 校验顺序)均 Draft、12/12 checks SUCCESS、未合并。未执行/受阻项(DFT 对照、NEP 原生 GPUMD、官方 T5/T6、Na₃PS₄ 电导率)在评论中逐条标注 BLOCKED/NOT_RUN/INCONCLUSIVE,未勾选任何计划复选框。

Rev 10 · 2026-11-17 · 冻结功能后的全覆盖验证与稳定化执行单(当前有效)

唯一执行入口:本 PR #75。 基线 main = 449ec9a(#105/#106/#107 已合并;GitHub 主分支 CI 14/14 success)。#61、#75 永久保持 Open + Draft,零代码变更,不得合并/关闭。本节是本轮待执行计划,不代表测试已经通过;以下 Rev 9 及更早的记录按原样保留,旧 SHA 与旧待办仅代表当时状态。测试结果回填本 PR 正文或讨论,修复代码走独立 PR → main。本轮禁止新增功能、后端、算法和未获授权的需求,包括但不限于 Fully flexible/Hybrid NPT、NEP 新运行模式、MLIP 微调。

R10-执行状态 · 2026-11-17

第一批/第二批已执行完毕,结果见下方评论Rev 10 · 执行结果(2026-11-17)(main @ 449ec9a4)。计划中的复选框未在本文勾选,避免把未执行项伪装成通过;确认缺陷的修复 PR:#108(#102/E8)、#109(R10-F02/F03)——两者已按 #108 → #109 合并,结果与 main 最新 SHA 见 合并结果;未修条目 F01/F04/F05 与 NOT_RUN/BLOCKED 清单保留在执行结果评论中。

R10-0. Agent 执行指令与验收原则

  • 先盘点后动手。 拉取最新 main、记录 commit/CI/环境;读 AGENTS.md、CONTRIBUTING.md、docs/development.md、docs/validation.md、本 PR 的 §4–10 及 [DEV MASTER · DO NOT MERGE/CLOSE] MLIP validation roadmap (#75) + confsearch living spec #61。确认所有已合并 PR、Issue fix(md): preserve MDRunner positional constructor compatibility after NPT (E8) #102 的现实状态和工作区;若 main 又更新,重新确定基线。
  • 一次性尽可能完成所有适用测试。 以旧 §4 的 S-01–S-09、§5 的 B/W/C/A/I/O 全项、§6 的异常输入族、§10 逐文件台账为既有主矩阵;本 Rev 10 补充近期 feat(md): ASE isotropic MTK NPT for UMA/MACE/GRACE/DPA #101/Task folder INCAR core: results/ default, config-before-structure, CONFSEARCH dispatch (R1 + R3.2/R3.3) #104–TUI new-task wizard + task-folder helper (R5) #107 未覆盖的项目和明确执行顺序。不得重复开平行审计总 PR;不得把旧清单未勾选项擅自视为已通过。用模块/接口/模型/设备/输入类别的覆盖映射找缺口;避免不可能完成的笛卡尔积,优先代表性组合与高风险成对组合。
  • 全部原始状态如实报告。 每一用例给 case ID、版本/SHA、原始期望、实际、命令、log/产物位置、PASS / FAIL / BLOCKED / UNSUPPORTED / NOT_RUN / INCONCLUSIVE;区分 core synthetic/mocked 与真实 MLIP,旧版证据不可冒充现版,skipped/xfail 不算通过。
  • 修复准则。 已复现缺陷先给 red test 或独立已知答案,再做最小 green fix;每个 fix 独立小 PR,写 DATA IMPACT,检查 main 全量回归和可观测结果,必要时中止当前受影响分支的验收。纯覆盖缺口先补测试,勿为测试而改变生产语义;不许删除测试、抬阈值、静默改变默认值/单位/任务 head/dtype/PBC、覆盖历史数据。
  • 算力边界。 现有环境、现有模型权重、廉价 CPU fixtures 和短时 GPU smoke 可优先开展;禁止擅自下载大型权重、提交昂贵 DFT、大规模超算/云 GPU、ns 级生产 MD、性能跑分风暴或破坏用户真实结果。此类工作列 BLOCKED + 具体资源清单,集中申请一次授权;不阻塞其余可执行测试。所有破坏性故障注入只能在临时目录。
  • 不引入功能。 发现“想要 Fully flexible MTK”等需求只记为限制/未来提案,不以本轮稳定化为借口实现。凡属模型物理精度问题,不直接认定为 Mliport 代码 Bug。

R10-1. 基线与全仓工程检查(E)

  • E01 建立 main HEAD、Python/OS/编译器/驱动/CUDA/ASE/torch 等实际清单,记录 3.10/3.11/3.12/3.13 GitHub Actions 必需 job 的结果、跳过项和 xfail;与最近已验证本地树对照。
  • E02 在干净 core dev 环境执行 pytest tests -q -ra;Ruff check 和 format --check;generate_docs.py --check;repository_hygiene;uv build mliport;mkdocs build --strict;检查依赖缺失是否如实 BLOCKED。不得把预期外 skipped 算 PASS。
  • E03 检查实际 pytest 收集数量、source→test 映射、跨配置测试盲区;新增必要的小而有判别力的参数化/性质/已知答案回归,不简单复制大量同义测试。比较 main 与 validation/science 旧证据对应 SHA 的差异,标记现版需要复验的路径。
  • E04 wheel/sdist/全新目录安装、console scripts、无源码树运行、包数据/许可证、CPU-only 环境、安装器 plan/doctor/launchers;按当前 registry 核对运行环境隔离、包版本/设备及模型查找。测试 Linux;具备可用 Windows 环境时补原生 Windows 路径/安装冒烟,否则 BLOCKED。

R10-2. #104–#107 任务文件夹、INCAR、models/ 与 TUI(U;本轮新增重点)

  • U01 同一真实小结构的 SP/OPT/短时 MD:分别经 CLI 参数、任务目录内 INCAR.mliport + mliport run、TUI New Task/Create & Run、Python API(有等价公开入口时)执行;比较 effective config、模型 identity、实际计算器、原子顺序、输出、status。只比较相同初始条件/确定性路径,随机过程必须固定 seed。
  • U02 路径优先级:本地声明目录同名文件优先;MLIPORT_ROOT 环境变量 vs 安装器 venv root;models/ 根目录与恰好一层子目录;GRACE SavedModel 目录;绝对路径和含分隔符的相对路径不被错误重定向;两个候选模型必须明确拒绝而非任取;不存在模型时展示已查目录;目录不存在/无权限/错误根定位 fail closed。
  • U03 非标准路径:空格、中文、Unicode、相对/绝对、软链接、只读、任务目录被移动、从其他 CWD 调用、Windows 路径(可用时);确认 OUTPUT_DIR = . 不异常、INCAR 相对 OUTPUT_DIR 相对声明任务目录解析且不影响历史非 INCAR 语义;不得意外写入源结构或模型目录。
  • U04 schema 统一选项表 → 模板 → 中英键参考文档三方一致;别名/覆盖优先级/inline 注释/负数/布尔/非法 key/冲突项/空值和 NaN/Inf;检查由模板产生的五类任务是否符合解析器,条件项(NPT、不同 thermostat、confsearch)只在适用时生效。
  • U05 TUI 新建向导:计算类型→结构→后端/模型,后端不同的选择列表、路径手输、目录重名、缺文件、模型选错、CIF/POSCAR/XYZ 逐字节复制、默认任务名、重复创建拒绝覆盖、REQUIRED 保留、NEB/confsearch 不偷偷指定参数、Create & Run 调用同一 CLI run 路径;Advanced 旧页面与 terminal 无交互环境不回归。
  • U06 检查任务生成/执行/失败/重试时 results/ 中 JSON、输入快照、模型路径与哈希、source SHA、运行目录、终态是否一致;并发同名任务/残留目录不默默覆盖;警惕 stale 产物污染后续分析。

R10-3. 数值核心、优化与所有 MD 系综(P;高风险)

  • P01 对现成的可用 calculator:E/F/stress 的形状、原子索引、符号、单位、Voigt 排列/压力正负号;中心有限差分校验原子力和应变应力(至少两种步长并估计数值噪声);参考 ASE/上游单点、解析势或可靠基准。含旋转、平移、晶格平移、原子重排、非正交晶胞、超胞复制(适用时)。
  • P02 固定胞/变胞 OPT:收敛、不收敛、异常、约束/固定原子、势能下降、最终力/应力、晶胞有效性和结果状态;不可把优化器中断写为成功;检查 stress 缺失的计算器明确拒绝相关变胞任务。
  • P03 NVE:Velocity Verlet 及速度/单位/原子质量,合适步长下的能量漂移与步长收敛,total energy 与 kinetic/potential 一致,COM、seed、预弛豫和生产相边界;不要在能量漂移上设置脱离步长/数值精度的死阈值。
  • P04 NVT:分别测试 Langevin、Bussi、Nosé–Hoover Chain 的参数、统计温度/DOF/COM、约束和不可用组合;判断温度分布需考虑自相关时间与有限样本,不以每个时刻恰为目标 T 为判据。
  • P05 NPT(现状仅 ASE IsotropicMTKNPT;UMA/MACE/GRACE/DPA,不包括 NEP):温度/压强及实际应力可用性,GPa↔eV/ų、压强含动能贡献、瞬时应力/体积波动,温度/压强链参数和 damping 单位,数值稳定/完整 provenance;与当前版本 ASE 直接调用结果或解析已知答案比对,不能只看“不报错”。
  • P06 非立方/掺杂晶胞专项:使用至少一个接近立方但 a≠b≠c、角度≠90° 的小型非正交盒子;各向同性 NPT 应保持初始晶胞形状(边长比、夹角)而缩放体积,不得偷偷强制 cubic;完整平均应力张量可能有残余偏应力,这是方法限制、不是必然程序故障。明确记录此限制;不开发柔性晶胞算法。
  • P07 NPT 拒绝链:零/负温度、非法 pressure/tdamp/pdamp/chain、非 3D PBC、约束及 COM 策略不满足、stress 不可用、NEP/GPUMD、NVE/NVT 携带 NPT-only 参数;尽可能在模型加载/产物写入前 fail closed。NPT 输出 raw/trajectory.traj 含逐帧晶胞,禁止写固定胞 XDATCAR,child manifest 标记 variable_cell=True、transport_eligible=False,analyze msd/transport 拒绝其轨迹;隐藏链变量不被虚称能精确续算。
  • P08 Issue fix(md): preserve MDRunner positional constructor compatibility after NPT (E8) #102,已知真实兼容性缺陷:证明 MDRunner 历史位置参数在插入 NPT 参数后发生绑定错位;用代表性调用测试覆盖 friction、Bussi/NHC、输出选项及其他旧参数。把 NPT 新增参数移至历史位置参数之后并设为 keyword-only(如果签名检查不出现额外兼容性风险),保持旧默认值与 NVE/NVT/NPT 数值/序列化契约;独立最小 fix PR,red→green 后关闭 fix(md): preserve MDRunner positional constructor compatibility after NPT (E8) #102。任何公有 API 限制同步 EN/ZH 文档。
  • P09 所有 MD:初始/生产/最终帧计数、步数/时间、写盘间隔、equilibration、不完整结果、重入、写盘中断/取消/磁盘失败;OUTCAR/md.csv/raw .traj/manifest 一致。长轨迹使用轻量 synthetic 数据做内存与写入压力测试,真实大规模/ns 模拟不在本轮默认预算。

R10-4. 全工作流与分析科学正确性(W/A)

  • W01 SP / OPT / batch:子任务级 status 与顶层汇总一致;已知失败不变成功;中途异常的 last terminal status 优先;负载模型/设备真实来源,不把输出格式伪装成 DFT 数据。
  • W02 NEB/CI-NEB:端点与原子对应、插值 PBC/MIC、周期跨界、约束、收敛/不收敛/错误与 saddle、能垒正反向和独立 ASE/解析势已知答案;中断→恢复后的几何/身份/status。NEB 恢复只是 geometry-level,不承诺同一优化器历史、同一鞍点或逐步数值轨迹。
  • W03 confsearch 两种入口、随机 seed/多副本、吸附位点与碰撞判据、结构分类和去重、Metropolis 与独立 RNG、计算预算、成功/失败统计、schema-2 checkpoint 完整几何与身份、畸形旧 schema-1 fail closed;随机中断每个持久化边界、恢复结果一致性和历史 provenance,真实小模型补证据。
  • W04 分析 MSD/transport:对解析/合成随机游走、定向漂移、各向异性扩散、跨 PBC 边界及稀疏采样歧义做 known answer;确认 Nernst–Einstein 与集体导电率不同、粒子数/电荷/体积/单位、kinisi 不确定度与有效拟合区间、不足样本 fail closed;NPT 或失败/不可信来源不得被当作 publication-grade。
  • W05 VACF/频谱、RDF/配位、结构分析、Arrhenius 拟合、GEMDAT(若已装):FFT vs 直接算法、频率/时间/温度/长度单位、取样/窗口/统计误差、边界条件和反例;对可解析的小体系建立独立数值答案,不能仅测试函数能返回数组。
  • W06 VASP-like 输出与外部读取:POSCAR/CONTCAR/OUTCAR/OSZICAR/XDATCAR 适用格式、atom order、cell/PBC、张量和数值精度、ASE/独立读取 roundtrip;失真的 PBC 信息必须警告或阻止错误物理分析。NEB/confsearch export 回读不改几何/固定掩码。
  • W07 真实小任务从目录生成→执行→分析→再次运行/续算闭环;用一套确实存在的可用 checkpoint/小体系(可选 LPSC、Na3PS4、Li2S),在同一模型与相同计算定义下核对每阶段身份;比较不同头或不同 DFT 参考零点的绝对能量禁止作为同一基准。

R10-5. 全后端与硬件(B;有条件尽可能实测)

  • B01 枚举目前文档支持的 MACE / GRACE / DPA / UMA / NEP-ASE / NEP-GPUMD 及 D3;为每个可获得且实际支持的 checkpoint×支持的设备/任务建立运行表;区分能加载、能完成 SP、E/F/stress 验证、OPT、MD、NPT(仅支持的 ASE 后端)、NEB/分析、precision/速度、科学精度。不要因某一个模型 PASS 就宣称整个家族 PASS。
  • B02 现成 CPU 与可用 V100:统一小结构,冷/热启动、同模型重复、E/F/stress FD、模型 hash/task/head/dtype、device 实际匹配、显存/内存占用、错误 head/task/device 拒绝、CPU/GPU 差异解释;需要不同 checkpoint 的结果不可假装直接比较绝对能量。可用其他 GPU 时加同一案例,不可用标 BLOCKED。
  • B03 NEP89 使用新版权重仅在 GPUMD 路径验证;Calorine 不支持该格式属于已知限制,不作为意外失败;分别验证文档声称的 NEP-ASE 与 GPUMD 能力,不能把 GPUMD 成功借给 Calorine。V100 tier-1 已有旧证据,应验证现 HEAD 无相关退化或明确记录继承依据。
  • B04 模型缓存与生命周期:不同结构/元素/胞、单点后更换结构、A→B→A、GPU OOM/回收、多个任务隔离;cache 开/关须 E/F/stress 一致。GPU 非确定性注明精度与重复范围。
  • B05 科学参考仅使用已有可信标签/known-answer:相同化学组成的相对能量、力、应力与 CP2K/ASE 合适对照;标出泛函、赝势、基组、未收敛参考等差异。特别是历史 Na3PS4 未收敛 DFT 力标签不得用于声称模型力精度达标。MLIP 的误差、迁移机制、输运外推不能被自动归罪于代码。
  • B06 真实性与资源表:精确记录权重 SHA256、runtime、GPU UUID/driver、torch/TF/ASE/GPUMD/Calorine 版本、任务 head/dtype、实际设备及失败日志;不可用的权重/平台明确 BLOCKED,并说明最小解除条件,不为了填满表格强行下载/折腾不兼容驱动。

R10-6. 可靠性、接口、安装与安全(R)

  • R01 故障注入:model exception、NaN/Inf、异常单点力/应力、磁盘满/权限不足、JSON 半写、SIGINT/SIGTERM、进程终止、重复运行和中途取消;确保不产生假成功、不吞掉真正异常、不丢失可恢复的原始文件、不用旧 manifest 冒充本次完成。
  • R02 任务队列:同时提交/领取/中止、GPU lease、PID stale、退出后的资源清理与重新领取、双调度器并发;只在临时目录和受控短任务测试,不触碰真实长任务。
  • R03 所有公共入口(CLI/API/TUI/INCAR/queue)同义参数实际解析与优先级、参数门禁、终态和输出一致;明确不支持的组合 fail closed;对预期错误提供可操作诊断信息。
  • R04 包装与兼容:独立虚拟环境、冷安装和轮子安装、backend 隔离、GPU/CPU 无静默回退、无打包遗漏;版本兼容矩阵以 install/compatibility.py 为准,不凭猜测扩张支持声明。确保不执行非预期 shell 命令或在路径异常时越界写入。
  • R05 性能回归仅针对重复可复现的固定基准,记录时间/显存、样本数、warmup、硬件;性能优化不能改变科学数值合同。除已复现的显著回归外,不开展无终点性能调优。

R10-7. 文档、证据与发布门禁(D/G)

  • D01 同步 docs/validation(.zh).md、models(.zh).md、installation(.zh).md、configuration(.zh).md、workflows/md(.zh).md、outputs(.zh).md、troubleshooting(.zh).md 与实际行为;文档两语言同义,自动生成段由 generate_docs.py 驱动。重点明确 Isotropic MTK 固定晶胞形状(非立方不必强制成 cubic)、NPT 不可直接用于固定胞输运、NEP89 GPUMD 格式局限、模型/DFT 数据精度与软件运行正确性分离;检查 Nosé–Hoover Chain 的表述不会造成误导。
  • D02 依据实际验证结果更新本 PR 的状态清单与 [DEV MASTER · DO NOT MERGE/CLOSE] MLIP validation roadmap (#75) + confsearch living spec #61 顶部的有效开发状态;保留历史记录及已有 raw evidence。校对 doc examples 的解析/执行;诊断错误与真正支持的选项相符。
  • D03 每项失败记录 root cause 与复现脚本、处理 PR/commit、测试前后对比和 DATA IMPACT(是否需要重新生成既有科学结果);旧模型证据带 stale/历史标签,不改写原始报告来假装现版重测。
  • G01 Release gate:最新 main 核验、所有必需 CI 通过、可执行 core tests 通过、P0 科学错误/数据损坏缺陷无未处理项、P1 有明确处理或显式风险接受、fix(md): preserve MDRunner positional constructor compatibility after NPT (E8) #102 已修复或有清晰兼容性警告、至少一个可用真实 MLIP 端到端闭环及可用后端的硬件证据;未运行场景列表完整,不能写“所有 GPU/所有模型通过”。
  • G02 当且仅当上述门禁成立,提出版本冻结/标签建议(暂不擅自发布、打 tag 或升级依赖);本轮完成后转入 bugfix-only 维护,新功能保持冻结。
  • G03 最终提交一份人类可读的中文验收报告在 [DEV P0 AUDIT · DO NOT MERGE/CLOSE] Living correctness and real-hardware test ledger · master #61 #75 讨论或修订状态正文:覆盖率(用例与组合的分母)、PASS/FAIL/BLOCKED/UNSUPPORTED/NOT_RUN、确认的缺陷及 fix PR、科学影响、未完成项和释放门禁结论。报告不要把 500 行流水账代替结论;完整日志和矩阵可附链接。标出“当前 HEAD 真正验收通过”的范围。

R10 执行顺序与停止条件

  1. 第一批并行: E 全仓基线/工程回归 + U 新工作流 + P01/P08 数值与 fix(md): preserve MDRunner positional constructor compatibility after NPT (E8) #102;优先发现阻塞真实使用的缺陷。
  2. 第二批: P 其余 MD/NPT + W 工作流/分析 + R 故障注入,首先用 synthetic/解析/ASE 独立参考,修复后复测。
  3. 第三批: B 可用真实后端/硬件 + W07 闭环与科学证据;时间和资源不足先做可判别的最小真实 smoke,再报告扩展测试成本。
  4. 第四批: D/G 文档、证据、缺陷 PR、版本冻结建议;所有功能仍保持冻结。

停止条件: 发现 P0 错误结果/静默数据破坏时暂停受影响工作流的信任声明并优先复现修复,其余独立测试继续;出现可破坏用户数据的操作或需要高成本硬件/下载时不执行并集中申请许可;无法复现只报 INCONCLUSIVE,不无休止重试。分阶段更新 #75,不等待全部工作完成后才首次报告。#61 和 #75 不添加代码、不合并关闭;独立修复 PR 是否 merge 仍由 owner 最终授权。


状态同步 · 2026-11-14(不改变 Revision 9 原始证据):本 #75 是长期科学正确性审计与验收台账,保持 Open + Draft、零代码变更,绝不合并/关闭;PR #61 是唯一长期开发总入口。此前两项 P0 的相关实现/定向验收已有阶段性结果:NEP 双运行时 #88、新版 NEP89 格式 #90、AUD-007 checkpoint 修复 #83 与 CPU 等价性测试 #82 均已合并;当前 main @ ec28ba3f39fe9aa865e99aebf9c4c2452c72e256 的 CI 全绿。NEP89 已在真实 V100/GPUMD 完成 tier 1 接口/数值一致性验收(§2b),但不是 Na₃PS₄ 精度验证,也不覆盖全部 GPU、长时 MD 或真实后端续算。AUD-004/005 和未签收的逐文件/物理验证继续保留为缺口,不得因 CI 全绿而标记完成。下一步拟先规划小规模 Na₃PS₄/CP2K benchmark(由 #61 统筹,未经 owner 授权不开展昂贵计算);DPA-SSE 暂缓。原有 113-file ledger、历史失败与 276 evidence records 保留;任何 skipped/blocked/xfail 不得写成 pass。


[DEV AUDIT · DO NOT MERGE / DO NOT CLOSE] Continuous correctness, reliability & exhaustive test campaign

Permanent, living audit PR. This PR is NEVER merged or closed during normal development.
Revision 9 — 2026-11-14. Original inspection baseline: 1ef9a1b; latest verified main: ec28ba3 (merge of #82; AUD-007 acceptance; NEP89 real V100/GPUMD tier 1).
Code is deliberately unchanged on this branch; it has one empty commit solely to host this PR. Keep the requirements and progress in THIS PR DESCRIPTION / review conversation. Fixes and test additions must be submitted in separate, small PRs targeting main, each linking "Tracking: #75". Do not commit transient audit reports, agent plans or scratchpads (AGENTS.md rules 8–10).
No new features in this campaign. A behavior-changing "fix" needs a demonstrated defect and owner approval where scientific semantics are involved.

0. 中文总纲(给项目负责人及执行 agent)

本 PR 只管理穷尽式代码审计、Bug 复现、科学/工程正确性测试和回归修复。不新增计算功能,不改训练/推理模型算法,不靠放宽阈值刷绿,不删除失败记录。任何性能改动也必须先有正确性证明,不允许静默改变默认任务、head、dtype、单位或 PBC。

分开评估:①静态代码审查;②隔离的 CPU/模拟 calculator 自动测试;③真实后端及硬件验证;④可复现的物理 known-answer / DFT 或 ASE 对照;⑤破坏性故障注入与结果持久化。CI 绿灯仅说明该 CI 矩阵通过,并不说明 GPU、科学外推可靠性及真实模型全覆盖。

覆盖四个推理后端 MACE / DPA / GRACE / UMA、D3、SP、opt、MD、NEB、confsearch、analysis、CLI/API/INCAR/TUI/queue、安装器及 VASP-like 输出。逐项填写证据和真实状态,不得虚构测试已经执行。

1. Binding rules (agents must follow)

  1. Read AGENTS.md, CONTRIBUTING.md, docs/development.md and relevant tests first. Existing scope rules override this tracker. If conflict: comment, propose correction, do not silently improvise.
  2. Never merge/close/mark ready this PR. Its branch stays code-free; update ONLY this description for accepted audit scope / checklists / status. No new planning Markdown in repository.
  3. Every implementation/fix/test PR targets main, quotes the audit ID(s) and links this tracker. First reproduce with a failing test (red); then smallest safe fix (green); run affected unit, integration, documentation and relevant science suites. Required DATA IMPACT in commit/PR: which scientific records need regeneration or "no scientific record invalidated."
  4. Evidence before verdict. Record baseline SHA, OS/CPU/GPU/driver, Python and dependencies, model filename+hash, backend/task/head/dtype/inference_mode, exact commands, fixture, expected/observed result, raw logs, exit code and reproduction repeat count. Record status in {not_run, blocked, pass, fail, unsupported, inconclusive}. Skipped ≠ pass; mocked calculator ≠ real MLIP; green tests ≠ scientific accuracy.
  5. Priority P0: possible wrong scientific answer / data loss / silently wrong identity/units/PBC. P1: wrong outcome, nonreproducibility, crash, bad resume, concurrency defect. P2: UX, portability, docs, performance regression. Risk without reproduction is a hypothesis, not an established bug.
  6. No feature creep, broad reformat, dependency churn, unreviewed compatibility upgrades or production behavior changes solely to get CI green. Never change numeric tolerance without a physical/numerical justification and independent reference.
  7. Respect user-owned HPC compute budgets: no expensive GPU or 8573C campaign, checkpoint download, or destructive test without explicit owner approval. Small CPU fixtures and CI are preferred first. Run failure injections inside disposable temp directories only.
  8. Any resolved item records fix PR, merge SHA, red/green test, hardware/backend coverage, science data impact, and residual untested risks. Preserve old scientific artifacts, indicate when they are stale; do not rewrite old results with newer identities.

2. Baseline and confidence boundaries

  • Observed main HEAD: 1ef9a1b, recent confsearch progress-log merge (feat(confsearch): per-cycle outcome, MD heartbeat and ETA in run.log #74).
  • Recursive tree: 485 entries, 113 package Python source files (~1.71 MB), 113 test Python files (~1.23 MB), 37 validation Python scripts (~0.71 MB). These counts are an inventory, not proof every line was executed or manually verified.
  • Root pyproject.toml is tooling only. Installable project is mliport/pyproject.toml. Core test environment and four real backend environments are distinct by design. Do NOT propose one universal backend runtime.
  • Reviewed architecture/documentation/CI plus focused source inspection of calculators, MD, NEB, confsearch, config, analysis, writers, install, queue, TUI. All 113 production Python files have now been fetched for second-pass static pattern/source inspection; detailed function-by-function signoff, source-to-test mapping and dynamic validation remain the agent's work below; do not tick the per-file ledger just because it appears in this inventory.
  • GitHub Actions for this SHA: tests, lint, package-build report success (2026-10-29). This does not establish real-model GPU coverage.
  • The published beta scientific campaign in validation/science/reports points to software SHA 5f8d91d; LGPS acceptance refers to 51d2a78. These are older than current main. Preserve evidence and identify changed paths; rerun affected tiers before asserting current-HEAD scientific validation.
  • Existing long-lived confsearch feature specification is [DEV MASTER · DO NOT MERGE/CLOSE] MLIP validation roadmap (#75) + confsearch living spec #61. Do not duplicate or silently replace its design decisions. Audit what is actually implemented; suspected spec mismatches should be raised there only after reproduction. This tracker is test/bug focused, not a backdoor feature scope.

Current execution ledger (Revision 9)

Work item State Evidence / follow-up
AUD-001 — MD trajectory step integrity Fixed and merged #76, main merge ad44f578; finite fractional/overflow and exact large integer/CSV parsing regressions; rare longdouble/other dtype questions remain outside scope
AUD-002 — XDATCAR PBC provenance Fixed and merged #77, main merge c7880a2; standard VASP XDATCAR preserved; matching MD provenance restores PBC; missing/untrusted PBC blocks bulk analysis. Branch CI green before merge; latest main CI tests still running at this update. Standalone-old-run analyses need provenance review.
AUD-003 — optimizer failure reported as non-converged Fixed and merged #78, main merge 47fa28f; error result/JSON/progress regressions
AUD-BATCH-001 — returned batch optimizer failure reported successful Fixed and merged #80, main merge 0d71f0a; per-result success/status/error/converged fields and summary totals now follow shared outcome classifier; PR and merge-to-main CI were green. Historical batch summaries may overstate success; check child results.
S-02 — MD pressure conversion known answers Partial test coverage merged #79, main merge 77a94afd; kinetic stress with nonzero momenta and other writer paths remain
AUD-004/005 Coverage gap / partly reduced Old scientific evidence still bound to old source SHAs; real GPU backend checks not established by core CI. Update 2026-11-14: the NEP backend now has a first real V100/GPUMD acceptance on current main (ec28ba3f) — see §2b — covering NEP89 weight loading, single-point E/F, force finite differences, 100 fs NVE and provenance. MACE / DPA / GRACE / UMA still have no real-model GPU evidence in this ledger.
AUD-005-NEP — NEP89 real V100/GPUMD acceptance Tier 1 pass; no defect found; no fix PR opened main ec28ba3f, model nep89_20250409.txt SHA256 518c7065…d82b85, GPUMD 5.9.1 (sm_70) SHA256 148259f2…d94bedb97, Tesla V100-SXM2-16GB GPU-f24116e9…. Native runtime loads format_version 1 weights and produces FD-consistent forces (max 4.2e-4 eV/Å), zero net force, bit-identical repeats and 100 fs NVE drift of 5.5e-6 eV/atom. ASE/Calorine runtime cannot parse this model layout (documented limitation — fails closed). Provenance records model/gpumd/GPU/source SHA256 and float32 precision note. Details, numbers and residual gaps in §2b.
AUD-006 — MD post-integration finalization Fixed and merged #81, main merge 3d6ca2e: synthetic CPU fault injection covers final-state validation, output/summary writing, manifest write failure and cancellation with reporting failure; the last OUTCAR status supersedes earlier trailers. PR head bf9d45ee 11/11 checks passed. Residual: manifest writes are non-atomic, cross-file atomicity unavailable, external OUTCAR consumers must use the last Status:. Main tests post-merge were in progress at update.
AUD-007a — NEB + confsearch resume contracts Complete — tests merged, acceptance passed #82, main merge ec28ba3f (head f7b0afb0, synced to main 4ddd54bb before merge). Adds CPU resume-equivalence tests (tests/mliport/mliport/confsearch/test_resume_equivalence.py, tests/mliport/mliport/neb/test_neb_resume_equivalence.py) and the corrected EN/ZH geometry-only resume wording (checkpointed band restored, optimiser state not replayed, convergence/saddle/barrier may differ). Review objection (same-saddle guarantee) resolved in d84b795f (EN) and 585da68a (ZH). Acceptance AC1–AC3 passed — see §2a. PR head CI green: tests (3.10–3.13), lint, docs, package-build, wheel-install-smoke×4, analysis-backends, installer-real-nep (deploy is main-only and skipped on PRs). Residual: synthetic CPU calculators only, no real-backend/GPU resume evidence.
AUD-007b — confsearch checkpoint corruption and output integrity Fixed and merged (schema rev 2) #83, main merge 2d0cbea5df7a2c62fe32b749e0ec73f238b6ba43: fail-closed on out-of-range unit, missing RNG streams and malformed JSON/schema; fixed-atom mask restored on every exported minimum; registry members now serialize and restore their own numbers/positions/cell/PBC/classification, which removes the representative-geometry substitution (3/21 frames, up to 0.19 Å, with retained per-member energies). Checkpoint layout bumped to schema/revision 2; schema-1 checkpoints are refused as non-resumable instead of fabricating lost geometry; historical files untouched for forensics. The previously strict xfail was removed and selected output plus minima/all_minima.extxyz are compared against an uninterrupted synthetic-CPU reference. Historical caveat (unchanged): confsearch runs resumed from schema-1 checkpoints before this fix may carry wrong member geometry; do not relabel such artifacts as corrected — regenerate from the original quench raw files if available.

Current CI baseline: latest main ec28ba3f (merge of #82) is green across tests (3.10–3.13), lint, docs, package-build, wheel-install-smoke×4, analysis-backends and installer-real-nep. Two pre-existing main failures were repaired separately in #92 (merge 4ddd54bb): 14 stale DPA torch==2.10.0 plan assertions, and uv venv --allow-existing being applied unconditionally. No real-model GPU campaign was run by the auditor.

Owner decisions: retain standard XDATCAR layout, preserve original ASE PBC mask in matching MD artifacts and refuse unsupported bulk-physics claims when source PBC is unverified. Never alter historical artifacts. Keep the two living tracker PRs #61 and #75 open. This ledger is a rolling summary; only mark per-file checkboxes after function-level audit evidence.

2a. AUD-007 acceptance record — 2026-11-14

Scope: verify the merged AUD-007 work (#82 tests/docs + #83 schema-2 repair) end to end on main 4ddd54bb, then merge #82 as ec28ba3f. Environment: Python 3.12.13, pytest 9.1.1, CPU-only .venv-dev; synthetic calculators (EggCrateCalculator, curved-MEP _CurvedMEP); no real MLIP backend, no GPU.

AC1 — confsearch resume equivalence (PASS). Interrupted after every durable unit boundary (1–3 units for 1 replica × 3 cycles; 1–4 units for 2 replicas × 2 cycles; 7 interruption points) and resumed, compared with an uninterrupted reference run of the same fixture:

Compared Result
search_results.json (all fields except model.model_path) identical
every minimum in minima/all_minima.extxyz — numbers, positions, cell, PBC, energy_eV, fmax_eV_A, move_mask identical (10 and 14 minima)
minima/registry.json — including each member's own positions_A (the #83 surface) identical (10 and 14 members)
rng_states (md/mc bit_generator.state) at every unit index, pre-crash and at resume identical

Negative controls (mutations applied to the resumed state; each detected): re-seeded replica RNGs (13 differences), 1e-6 Å perturbation of a restored member (positions differ), simulated #83 regression where members borrow the representative geometry (positions + registry differ). The committed tests assert selected-result equality and RNG restoration; the coordinate/registry comparison above is audit evidence from the acceptance run, not a committed assertion.

AC2 — NEB resume correctness (PASS). Resumed attempt's first checkpoint is byte-identical to the band the interrupted attempt checkpointed (band sha256 verified, then re-validated for atom count/indices/masses/PBC/cell/symbols); the interrupted attempt's step_* checkpoints and vasp_path* exports all survive; the resumed run reaches status=completed, converged=True, the same climbing image and the same saddle. Barrier equality is deliberately not asserted: the optimiser state is not replayed (geometry_resume_only: true in both the checkpoint and run_context.json), and the test only guards a loose 1e-3 eV bound (observed ≈2e-4 eV on this fixture).

AC3 — regression (PASS). Full suite on the merged head: 2077 passed, 48 skipped, 0 failed (pytest tests). For comparison the same suite on unmodified 327c44ad was 14 failed, 2039 passed; those failures were main-side defects (stale DPA torch==2.10.0 assertions; unconditional --allow-existing) and were repaired in #92, not attributed to #82. ruff check, ruff format --check, generate_docs.py --check, check_repository_hygiene.py and test_docs_contract.py clean.

DATA IMPACT: #82 is tests + documentation only; no checkpoint, trajectory or evidence file is touched, and no stored result is invalidated. #83's schema-2 change only affects newly written checkpoints; schema-1 checkpoints are refused rather than repaired.

Residual risk carried forward (not a pass): synthetic CPU calculators only — no real-backend (MACE/DPA/GRACE/UMA/NEP) checkpoint/resume run, no GPU, no interrupted multi-hour production run. test_resume_converges_to_the_same_saddle verifies the resume contract on a synthetic curved MEP, not model accuracy.

2b. NEP89 real V100/GPUMD acceptance — 2026-11-14 (tier 1)

Scope: first real-model, real-GPU acceptance of the NEP backend on current main. Priority order per the owner request: single-point energy → atomic forces → short NVE → output provenance. No source change, no fix PR (no defect reproduced). Not a model-accuracy campaign: this verifies the mliport NEP adapter (format handling, forces, energy conservation, provenance) against one published universal potential.

Identity (all hashes recorded)

Item Value
mliport source main ec28ba3f39fe9aa865e99aebf9c4c2452c72e256 (v2.0.0b4), editable install in .venv-nep
Runtime stack Python 3.12.13, numpy 2.5.4, ase 3.29.0, calorine 4.1 (ASE runtime only)
Model nep89_20250409.txt, 14,156,597 B, 976,331 parameters, SHA256 518c70656302ea9d1e95ad8da66bd2e73b31bb03f4134e2f9fd7b3b589d82b85; header nep 89 … / format_version 1 / model nep / zbl 1 2 / cutoff 6 5 274 158 / 89 element types / energy references all zero
GPUMD 5.9.1, /tmp/gpumd-src/src/gpumd, SHA256 148259f2c24fe17c12be09fd0bba02528b98988cd86bd553fa896d6d94bedb97, built -arch=sm_70 from source commit aa63ffdd60dc57c03b5641a9b29eb329514452c9 (CUDA runtime libraries supplied by the PyTorch wheel set of another environment, driver 580.173.02 / CUDA 13.0)
Device Tesla V100-SXM2-16GB, CC 7.0, 16,384 MiB, UUID GPU-f24116e9-bea2-e0a6-70a9-d0b50811856b; GPUMD stdout confirms Device name: Tesla V100-SXM2-16GB
Inputs Si16.vasp SHA256 810560d8f9a5564cf87e9ca397154b5332bf254f616408a4ffbe6bf22c461370; Cu8.vasp 1e7bc25b…f8be0; distorted Si16 0c948788…0f09c; mixed Si/O b9209861…04c94; per-run effective input native/model.xyz SHA256 7a4e971fb96a5afb917eab326f888bdf6197d7e2011cd4311578a847c05d957c

mliport doctor --engine nep --device cuda:0 passes (gpumd probe ✓). Warnings only for optional features missing from this environment (textual, kinisi, gemdat) — environment state, not a NEP finding.

1. Single-point energy (native NEP_RUNTIME=gpumd, V100, float32)

System E (eV) E/atom (eV) max abs F (eV/Å) max abs net F (eV/Å)
Si16 diamond a=5.43 −87.400204 −5.462513 5.0e−06 1.0e−08
Cu8 fcc a=3.615 −17.133877 −4.283469 2.0e−06 1.0e−08
Si16 distorted (σ=0.05 Å, seed 7) −86.560678 −5.410042 1.957 2.0e−08
Si/O mixed (synthetic geometry, element-mapping check only) −72.6309 — 6.234 —

Mechanical equilibrium is reproduced at the symmetry-preserving geometries (forces ~1e−6), while the distorted cell yields physical force magnitudes with a zero net force.

2. Atomic forces (verification)

  • Central finite differences on the distorted Si16 cell, h = 0.02 Å, 2 atoms × 3 axes: reported vs −dE/dx agreement within 4.2e−4 eV/Å (max relative 0.16 %), consistent with the float32 energy floor (ε·|E|/2h ≈ 2.6e−4 eV/Å at E ≈ −86.55 eV). Sign, magnitude and units (eV/Å) are correct.
  • Hydrostatic strain scan of Cu8 (scale 0.97 … 1.03): forces stay ~1e−6 eV/Å at every volume — the correct symmetry behaviour (hydrostatic strain produces stress, not forces) — while the energy varies smoothly: parabola minimum at scale 0.9838 → a₀(Cu) = 3.5565 Å (1.6 % below the experimental 3.615 Å; a universal-potential accuracy statement, not a mliport defect).
  • Repeatability: two independent runs of the same input are bit-identical in E and F (dE = 0, max|dF| = 0).

3. Short NVE (native, 200 steps × 0.5 fs = 100 fs, 16 Si atoms, velocity 300 seed 2026)

  • Total energy KE + PE drift over the run: +8.8e−5 eV = +5.5e−6 eV/atom (max−min 1.5e−4 eV = 9.1e−6 eV/atom; relative 1.0e−6) — consistent with float32 GPUMD inference.
  • thermo.out internals are self-consistent (T ↔ KE ratio 1.000 in all 20 dumped rows); mliport's normalized trajectory.extxyz reproduces the raw GPUMD potential energies in 20/20 frames.
  • Documented limitation confirmed, not a defect: the normalized trajectory holds no velocities (dump_xyz runs without vel), so VACF-type analyses remain unavailable; T/KE live in native/thermo.out.

4. Output provenance (verified in manifest.json / native/provenance.json)

Recorded and cross-checked: model SHA256 (nep.sha256), format version/model family/element list/cutoffs/ZBL/energy references, effective input model.xyz SHA256, GPUMD path + SHA256 + version, GPU UUID + requested device, numeric_precision: float32 with the explicit note that precision double only widens text output, mliport_source_sha, and the full run.in (velocity 300 seed 2026, time_step 0.5, ensemble nve, run 200, dump_thermo 10, dump_xyz 10 … force virial mass unwrapped_position). Raw GPUMD stdout.log/stderr.log, thermo.out and traj.xyz are retained next to the normalized outputs; run status completed, process return code 0.

Failure paths (classified, no fix PR)

Case Observed Classification
NEP_RUNTIME=ase (CPUNEP on cpu, GPUNEP on cuda:0) with NEP89 v1 exit 1, Failed to parse line 6 from …nep89_20250409.txt, no manifest, no success artifacts Backend implementation limitation, already documented in docs/backends/nep.md ("Calorine support for this newer model layout has not been verified; use the native runtime for its first real test"). Fails closed; mliport's own header parser and identity reporting still work.
Structure with an element outside the 89-element list (Am) exit 1 before GPUMD launch, message lists both the model's and the structure's elements Fail-closed chemistry validation, correct
Native runtime without --gpumd-command exit 1, explicit "requires NEP_GPUMD_COMMAND" Fail-closed, no silent fallback to the ASE runtime

Residual gaps (explicitly not passed)

  • Single V100, single GPUMD build; no second GPU family/architecture comparison.
  • Build provenance of the existing gpumd binary is partially unknown: the CUDA toolkit version used for that build was not recorded (the binary is sm_70, version 5.9.1, source commit pinned above).
  • No real-model accuracy campaign (cohesive energies, equations of state, defects, surfaces, MD vs DFT/NEP reference data) — this tier only verifies adapter correctness and internal consistency.
  • No stress validation beyond the symmetry observation; no long-run (ns) energy-conservation or performance measurement.
  • No native checkpoint/restart equivalence (explicitly not claimed by the docs).
  • Raw logs/artifacts are transient local files under /tmp/nep89-accept (not committed, per AGENTS.md rule 9); every number above is reproducible with the commands recorded in the issue conversation.

3. Source-backed audit leads — dispositions updated as reproduced

ID Evidence / source code Precise reproduction question Severity until reproduced
AUD-001 analysis/dataset.py original lossy astype(int) path Reproduced and fixed in #76 (ad44f578), with fractional, mixed large-int/float and exact CSV text regressions; rare nonstandard dtype limitations remain outside the tested contract P0 resolved (standard inputs)
AUD-002 Standard XDATCAR omits source ASE PBC Reproduced and fixed in #77 (c7880a2): warn on lossy export, persist PBC in MD metadata, restore if source identifies same artifact, deny bulk analysis eligibility without verified source PBC P0 resolved for supported metadata contract
AUD-003 runners/optimization.py original exception→not-converged reporting Reproduced and fixed in #78 (47fa28f) with status, event and error provenance regressions; cross-batch masking handled separately in #80 P1 resolved
AUD-004 scientific validation report's pinned commit is older than main; real backend runs are in validation/ rather than core CI Diff current HEAD against campaign SHA; map impacted scientific tests; create current-HEAD real model evidence where feasible; do not relabel old evidence P0 validation debt
AUD-005 .github/workflows/test.yml runs core unit tests + analysis integration, but does not install all four MLIP backends or execute their real checkpoint inference on GPUs Count mock-vs-real paths; test smoke with pinned checkpoint profiles in backend-specific environments on approved hardware. Explicitly mark unsupported/no-model cases; do not treat unit mocking as model validation P1 coverage gap
AUD-006 runners/md.py / runners/md_output.py: asynchronous writer and force-abort handling are intentionally fail-closed; there are multiple exception/finalization paths Inject disk-full, writer failure, KeyboardInterrupt, cancellation and model failure; prove requested frames aren't silently lost and manifest statuses are terminal; detect potential secondary exception masking P0 hypothesis
AUD-007 confsearch/checkpoint.py, engine.py, workflow.py and neb/io.py all carry resume and identity logic Reproduced and fixed in #83 (2d0cbea5) and pinned by #82 (ec28ba3f): interrupting at every durable unit boundary and resuming reproduces search_results.json, every minimum's coordinates and the per-unit md/mc bit-generator states; NEB restores the checkpointed band verbatim and keeps prior checkpoints/exports; corrupt schema-1 checkpoints are refused. Residual: synthetic CPU calculators only, no GPU/real-model resume P0 resolved (CPU contract)

When a lead is disproven, mark "not a bug" with reproduction and add a guard test if useful; do not keep stale defect claims.

4. Universal scientific invariants (apply across ALL public entry points)

  • S-01 Model identity: checkpoint SHA256, engine, task/head, dtype, inference mode, compile flags, model versions, dispersion identity are pinned and reported. No energy comparison across incompatible heads/tasks/reference zeroes.
  • S-02 Units: energy eV, force eV/Å, stress correct ASE Voigt order and sign, pressure conversion, time fs/ps, diffusivity Ų/fs → m²/s, conductivity S/m → S/cm/mS/cm. Independent hand-calculated known answers and finite difference.
  • S-03 E/F/stress consistency: numerical central differences across multiple displacement/strain scales; errors compared against measured numerical floor, not arbitrary lenient thresholds.
  • S-04 Physics and geometry: periodic/nonperiodic/partial-PBC contracts, slab vacuum, triclinic skew, cell handedness/singularity, wrapping/unwrapping, atom permutation and chemical identity.
  • S-05 Constrained atoms: raw vs constraint-applied forces never silently conflated; FixAtoms/FixCartesian/FixCom/FixSymmetry, mass/DOF and thermostat handling verified against ASE and analytical references.
  • S-06 Charge/spin: task-specific support, no silent default override, no unsupported molecular/solid task fallback.
  • S-07 Randomness: same seed/same versions/replay identity; different seeds; nondeterministic GPU tolerance reported rather than assuming bitwise identity.
  • S-08 Failure closed: invalid, unknown, NaN/Inf, zero/negative/out-of-range physical inputs, missing output/model, unsupported combinations rejected before committing misleading outputs.
  • S-09 Immutable science evidence: source hashes, exact resolved configuration, actual device, software SHA, run/attempt IDs and output manifests survive every successful/failed/cancelled run.

5. Exhaustive executable test matrix

5.1 Backend calculators, runtime and GPU

  • B-01 MACE real model on CPU and approved NVIDIA GPU; float32/64, energy/force/stress repeatability and actual-device reporting.
  • B-02 DPA multitask checkpoint: correct explicit head, absent/misspelled head refusal, task incompatibility, dtype, native precision and CUDA runtime handling.
  • B-03 GRACE TensorFlow: CPU/GPU initialization, periodic extractor, cache-off/cache-on identical E/F/stress at fresh, small displacement, cell strain, species/order change, PBC change, cutoff boundary.
  • B-04 UMA FAIRChem: omat/omol route restrictions, molecular charge/spin where supported, direct-force capabilities, stress and permitted inference/compile modes.
  • B-05 Never silently fall back from requested CUDA to CPU, or from one GPU to another; direct Python/API preinitialized frameworks vs isolated CLI behavior are explicit.
  • B-06 GPU architecture and driver/CUDA wheel selection: documented V100 baseline; other families explicitly supported/experimental/unsupported; never infer "verified" from theoretical compatibility.
  • B-07 D3: additive energy/force/stress for supported PBC, finite differences, damping/xc policy, device/dtype compatibility, model+D3 provenance, detect double counting and exception behavior.
  • B-08 Identity metamorphic tests: rigid translation, lattice image translation, atom permutation (with inverse mapping), rigid rotation including stress tensor, supercell replicate when model physics permits.
  • B-09 Cold/warm repeated inference, A→B→A cache invalidation, multiple sizes, OOM/fragmentation and resource cleanup; compare results not just runtime.

5.2 Single point, optimization, MD, NEB

  • W-01 SP consistent output energies, forces, stress, missing stress as unsupported; PBC and task validity gates.
  • W-02 Opt positions-only and cell relaxation; constrained and unconstrained, fixed-cell preservation, converged/non-converged/exception semantics, NaN and singular cell rejection.
  • W-03 MD NVE: ASE VelocityVerlet known answers, time-step drift scaling, energy accounting, momentum handling, deterministic seeded initialization.
  • W-04 MD NVT: Langevin/Bussi/NHC separately; thermal statistics, effective DOF, COM handling, constraints support/rejection matrix and correct thermostat reporting.
  • W-05 MD preservable positions/momenta vs newly initialized velocities; pre-relax incompatibility, deterministic restart, initial/final checkpoint and phase boundary semantics.
  • W-06 MD safety: raw vs projected force threshold, unsafe-frame checkpoint, abnormal outputs, empty/zero DOF, fast-collision and runaway model.
  • W-07 MD I/O stress: save interval, frame zero/end, equilibrate/production, raw .traj + md.csv + XDATCAR + OUTCAR coherence, 10^5-frame bounded memory, writer backpressure, interrupt/disk fail.
  • W-08 NEB atom identity, fixed atoms, endpoints, mapping/periodic image shifts, triclinic MIC, no accidental long image jumps, collision preflight, 3D/partial-PBC contracts.
  • W-09 NEB projected tangent/force and ASE reference, CI switching and convergence; endpoint energies; forward/reverse barriers; no invented saddle.
  • W-10 NEB gradient preflight: analytic quadratic, force-vs-energy inconsistency, noisy float32 floor, direct-force model rejection, false positive/negative controls.
  • W-11 NEB checkpoint corruption, restart identity, fixed atom changes, model revision, interrupted stage, stale lock, failed and resumed output manifests.
  • W-12 CI-NEB Hessian FD grid, one robust negative mode, tangent alignment, two-mode false saddle, fixed active DOF and displacement-size sensitivity.

5.3 Configuration search (#61; audit ONLY, no redesign)

  • C-01 Both prepared-start and automatic adsorbate placement modes, input hashes and index bounds, molecule/slab partition and chemical formula.
  • C-02 Surface +z geometry, vacuum and self-image checks, skewed 2D lattice, non-3D PBC refusal, mixed conformer identities.
  • C-03 Element-pair covalent radius contact distance; anchored vs unanchored placement; rejection attempt limits and complete rejection counters; fixed slab atoms.
  • C-04 MD anneal–quench counts, temperature schedule, snapshots, fixed/relaxed atom mask, per-cycle budget and no unrecorded extra MD/quench.
  • C-05 Reaction classification: intact/covalent chemisorbed/ionic contact/reacted; reference bonding and PBC topology; don't accidentally exclude legitimate chemisorption.
  • C-06 Oganov-Valle focused fingerprint vs independent ASE/known answer, cell rotation/permutation, periodic self-images near cutoff, near-zero norm and dedup threshold boundaries.
  • C-07 Metropolis acceptance, independent RNG streams, reproducibility, saturation counts, duplicate hit counts, ranking and tie resolution.
  • C-08 Different model/head/identity absolute energies never compared or merged; extension/resume only on allowed changes; failed attempts retained.
  • C-09 Crash during each unit, atomic checkpoint, partial file, stale lock, concurrent search, archive/merge repeat runs and all exported structures and metadata roundtrip.
  • C-10 DFT candidate export: POSCAR/CP2K-friendly files preserve coordinates, cell, species, constraints where supported; MLIP screening never claimed to be DFT reference.

5.4 Analysis, thermodynamics, transport

  • A-01 Import .traj / extxyz where supported / XDATCAR / mliport run; explicit vs inferred times, step/phase/frame order, missing/corrupt files, nonuniform stride, truncated runs.
  • A-02 Wrapped vs unwrapped positions, exact image vs MIC ambiguity, rotated/triclinic, high-velocity multi-crossing cannot be reported as certain from wrapped sparse samples.
  • A-03 MSD direct vs FFT synthetic random walk and anisotropic transport; stationary structure; sampling sufficiency and ballistic/plateau vs diffusive classification.
  • A-04 Drift correction none/mobile/framework/all, arithmetic vs mass weighted, overlapping species and constraint effects; no false diffusion from COM drift.
  • A-05 kinisi posterior variance, lag grid, insufficient windows, convergence diagnostics, tracer vs collective diffusion, Haven ratio and correlated/anticorrelated known answers.
  • A-06 Nernst–Einstein: ionic charge z, mobile particle number density and volume, correct SI conversion, independent conductivity benchmark and never infer experimental agreement.
  • A-07 GEMDAT site/jump counts, site origin, residency, transition rates and low-event conditions; real installed backend vs mocked output.
  • A-08 Arrhenius weighted/unweighted regression, uncertainties, two-point special case, duplicate T, extrapolation warnings and log-fit algebra known answer.
  • A-09 VACF direct vs FFT, one-sided spectrum, normalization, time units/Nyquist/aliases; RDF shell coordination, triclinic cutoff, RMSD/RMSF PBC and alignment.
  • A-10 Source contamination: incomplete/failed MD runs never become valid transport conclusions; provenance hash, repeated analysis attempts, overwrite prevention and corrupt-result rejection.
  • A-11 Array files and scientific JSON: no pickle, checksums, schema versioning, NaN/Inf handling, no fake zero for missing data.

5.5 Four user interfaces, queue, cancellation, installation

  • I-01 Equivalent effective settings across Python API / CLI / INCAR-style / TUI / queued command, including defaults, alias resolution and source precedence.
  • I-02 Unknown/misspelled/deprecated/conflicting keys fail closed; bool/int/float strictness; nonfinite numeric input; relative path resolution from config file location vs CWD.
  • I-03 Task options don't bleed between SP, opt, MD, NEB, confsearch and backend-specific options; exclusive model resource/settings preserved.
  • I-04 TUI: field dependency, disabled control not forwarded, preview, validation before launch, exit/cancel and resume, keyboard/scroll sizing and error message fidelity.
  • I-05 Queue: two schedulers simultaneously, claim/lease CAS and GPU UUID fairness, stale PID reuse, kill/reap crash window, pause/resume, device isolation and job-specific output trees.
  • I-06 Queue path security, malformed task, symlinks/escaping/Unicode/long paths, Windows/POSIX state file compatibility; do not execute untrusted shell fragments.
  • I-07 Installer: core-vs-backend isolation, Python 3.10–3.13 support where claimed, environment pin/lock reproducibility, CPU-only and GPU routes, offline/proxy and bad wheel failure.
  • I-08 Doctor actually tests requested backend/task/device with real minimal inference where feasible; "package importable" is not "scientifically validated".
  • I-09 Clean package installation from wheel/sdist with no source tree, include runtime JSON data/licenses, no archive import or transient files; docs and bilingual parity.
  • I-10 Signal handling: SIGINT/SIGTERM/KeyboardInterrupt/worker killed while writing; no phantom success, locked directory recovery and terminal statuses.

5.6 File formats, scientific reporting, durability

  • O-01 OUTCAR/OSZICAR/CONTCAR/XDATCAR grammar and independent ASE/VASP-compatible reader roundtrip, including atom symbol order, triclinic cells, high precision, optional stress/force.
  • O-02 VASP-shaped files never imply DFT/SCF; model provenance, total energy reference, forces (raw vs applied) and stress conventions explicit.
  • O-03 Fixed-cell XDATCAR contract and partial-PBC information-loss test (AUD-002).
  • O-04 Atomic write, fsync, locks, race, disk quota, unclean shutdown and retry on every report/checkpoint/output artifact.
  • O-05 Scientific evidence schemas and status classification (pass/fail/characterized/unsupported/not_run), no cherry-picked successful runs, immutable model/structure identity.
  • O-06 Documentation EN/ZH parity, CLI examples run, generated sections regeneration only, public docs accurately describe only implemented behavior.
  • O-07 Performance: cold/warm latency, atom-size scaling, MD step/s, peak VRAM/RAM, neighbor-cache equivalence before profiling; benchmark must pin environment and tolerances.

6. Cross-product / adversarial input families (mandatory where applicable)

Test all applicable combinations rather than only happy paths: zero/one/many atoms; small/large systems; identical/mixed elements; atom order permutations; zero/negative/singular/inverted/strained/strongly skewed cell; 3D/2D/1D/0D PBC; fixed/unfixed atoms; nonperiodic molecule; float32/64; negative/zero/NaN/Inf stress/force/energy/time; tiny/huge velocity; too-close atoms; 0/1/2 frames; irregular saved steps; failed/interrupted calculation; backend out-of-memory; wrong/corrupt/stale checkpoints; read-only directory/full disk; canceled job; simultaneous scheduler. Avoid generating scientifically unrealistic "pass" fixtures: use them to prove clean rejection, not to establish model quality.

7. Required evidence, status table and acceptance protocol

For EACH case create a stable short ID; update this tracker with:

Field Required
Case ID, code path, physical assertion, independent oracle
Inputs small fixture, exact script/command, version/model SHA, seed, task/head/dtype/device
Reproduction initial observed failure / explicit not_run; test fails before fix where appropriate
Result pass / fail / unsupported / blocked / inconclusive, machine/environment and log artifact
Fix independent PR number, commit SHA, data impact, regression test name
Residual what was NOT validated (real model, GPU, platform, reference accuracy, etc.)

Definition of done for an individual bug: reproducible red test, minimal green fix, upstream independent known answer for physical claims, affected matrix tests green, no record corruption, and reviewed DATA IMPACT. A suite skip must have a recorded explicit reason and cannot count as success.

Campaign acceptance does not mean "all models are accurate". It means the audited scope is documented, every production module accounted for, no unresolved P0 without explicit owner decision, all feasible P1 tested/fixed, CI green, real-model campaign repeated for current source where required, and residual limitations transparently reported.

8. Execution order: small independent PRs (no code on this branch)

  • Phase 0 — Baseline freeze. Snapshot HEAD, inventory actual tests/mocks/skips/xfails, current CI, scientific campaign provenance. Build precise source→test map; isolate changed paths since last GPU campaign.
  • Phase 1 — P0 numerical contracts. Units/forces/stress/PBC/model identity; triage AUD-001, AUD-002 and the real-data provenance gap. Include known-answer tests and correct format roundtrips.
  • Phase 2 — P0 evidence integrity & crash safety. MD/NEB/confsearch raw files, cancellation/crash/disk-full, resume, cross-task output locks, stale statuses, queue race.
  • Phase 3 — P1 workflow correctness. SP/opt/MD/NEB/confsearch error cases and physical invariants, including AUD-003 and AUD-006/007.
  • Phase 4 — P1 analysis & multi-entrypoint parity. Diffusion/conductivity uncertainty, sampling sufficiency, TUI/API/CLI/INCAR/queue consistency and edge cases.
  • Phase 5 — Backend runtime and platform verification. Separate approved MACE/DPA/GRACE/UMA environments; CPU and available V100 GPU; D3, cache, precision, CUDA and driver characterization. Regenerate or clearly supersede outdated evidence.
  • Phase 6 — Reliability/performance/regression/release gate. Package/wheel, docs, hardware resource leaks, repeatable performance, red/green audit ledger and final risk register.
  • Phase 7 — Continuous maintenance. Every subsequent feature/dependency/bug PR checks affected source→test cases and refreshes this PR's revision/status and any stale evidence. This phase never finishes; PR stays open.

9. Exact initial CI and local commands (do not conflate environments)

Core, from repository root, on a clean development Python:

uv venv .venv-dev
uv pip install --python .venv-dev/bin/python -e './mliport[dev]'
uv pip install --python .venv-dev/bin/python torch --index-url https://download.pytorch.org/whl/cpu
.venv-dev/bin/python -m pytest tests -q -ra
.venv-dev/bin/ruff check mliport/mliport tests scripts
.venv-dev/bin/ruff format --check mliport/mliport tests scripts
python scripts/check_repository_hygiene.py
python scripts/generate_docs.py --check
uv build mliport

Run the isolated analysis-backend CI environment per .github/workflows/test.yml (including kinisi and GEMDAT); run actual backend smoke and validation/science tiers only in correct isolated backends on approved hardware. Capture the complete commands, logs, model hashes and exact software commit. Use pytest reporting of skipped/xfail to prevent "green" hiding unexecuted scientific tests. Do not introduce new CI dependencies blindly; verify across Python 3.10/3.11/3.12/3.13.

10. Code-by-code review ledger (every box starts UNTICKED)

Agent must inspect each listed production Python file individually, assign one or more test IDs, inspect error paths as well as normal paths, compare to public interface and existing tests, and only then tick it with a linked evidence comment. A directory or large file being named here does NOT mean it was completely reviewed. Review/compare existing tests under tests/ and validation/ as part of the evidence mapping. Update this ledger when files are added/renamed on main.

(root) (23)

  • mliport/mliport/init.py
  • mliport/mliport/api.py
  • mliport/mliport/backend_selection.py
  • mliport/mliport/base_calculator.py
  • mliport/mliport/capabilities.py
  • mliport/mliport/cli.py
  • mliport/mliport/devices.py
  • mliport/mliport/doctor.py
  • mliport/mliport/engine.py
  • mliport/mliport/gpu_compat.py
  • mliport/mliport/gpu_setup.py
  • mliport/mliport/job_worker.py
  • mliport/mliport/jobs.py
  • mliport/mliport/logger.py
  • mliport/mliport/model_record.py
  • mliport/mliport/protocols.py
  • mliport/mliport/queue.py
  • mliport/mliport/queue_daemon.py
  • mliport/mliport/run_directory.py
  • mliport/mliport/run_status.py
  • mliport/mliport/timing.py
  • mliport/mliport/utils.py
  • mliport/mliport/validation.py

analysis (15)

  • mliport/mliport/analysis/init.py
  • mliport/mliport/analysis/arrhenius.py
  • mliport/mliport/analysis/dataset.py
  • mliport/mliport/analysis/drift.py
  • mliport/mliport/analysis/electrolyte.py
  • mliport/mliport/analysis/msd.py
  • mliport/mliport/analysis/plots.py
  • mliport/mliport/analysis/runner.py
  • mliport/mliport/analysis/schema.py
  • mliport/mliport/analysis/spectral.py
  • mliport/mliport/analysis/structure.py
  • mliport/mliport/analysis/thermo.py
  • mliport/mliport/analysis/transport.py
  • mliport/mliport/analysis/units.py
  • mliport/mliport/analysis/validation.py

calculators (7)

  • mliport/mliport/calculators/init.py
  • mliport/mliport/calculators/dispersion.py
  • mliport/mliport/calculators/dpa_calc.py
  • mliport/mliport/calculators/factory.py
  • mliport/mliport/calculators/grace_calc.py
  • mliport/mliport/calculators/mace_calc.py
  • mliport/mliport/calculators/uma.py

config (8)

  • mliport/mliport/config/init.py
  • mliport/mliport/config/aliases.py
  • mliport/mliport/config/defaults.py
  • mliport/mliport/config/incar.py
  • mliport/mliport/config/provenance.py
  • mliport/mliport/config/resolver.py
  • mliport/mliport/config/schema.py
  • mliport/mliport/config/settings.py

confsearch (18)

  • mliport/mliport/confsearch/init.py
  • mliport/mliport/confsearch/checkpoint.py
  • mliport/mliport/confsearch/engine.py
  • mliport/mliport/confsearch/errors.py
  • mliport/mliport/confsearch/fingerprint.py
  • mliport/mliport/confsearch/geometry.py
  • mliport/mliport/confsearch/graphs.py
  • mliport/mliport/confsearch/index_spec.py
  • mliport/mliport/confsearch/options.py
  • mliport/mliport/confsearch/placement.py
  • mliport/mliport/confsearch/prepare.py
  • mliport/mliport/confsearch/quench.py
  • mliport/mliport/confsearch/regions.py
  • mliport/mliport/confsearch/registry.py
  • mliport/mliport/confsearch/results.py
  • mliport/mliport/confsearch/revisions.py
  • mliport/mliport/confsearch/status.py
  • mliport/mliport/confsearch/workflow.py

data (1)

  • mliport/mliport/data/init.py

install (10)

  • mliport/mliport/install/init.py
  • mliport/mliport/install/main.py
  • mliport/mliport/install/cli.py
  • mliport/mliport/install/compatibility.py
  • mliport/mliport/install/docs.py
  • mliport/mliport/install/hardware.py
  • mliport/mliport/install/inventory.py
  • mliport/mliport/install/launcher.py
  • mliport/mliport/install/plan.py
  • mliport/mliport/install/sources.py

neb (8)

  • mliport/mliport/neb/init.py
  • mliport/mliport/neb/io.py
  • mliport/mliport/neb/preflight.py
  • mliport/mliport/neb/prepare.py
  • mliport/mliport/neb/results.py
  • mliport/mliport/neb/revisions.py
  • mliport/mliport/neb/schema.py
  • mliport/mliport/neb/workflow.py

runners (9)

  • mliport/mliport/runners/init.py
  • mliport/mliport/runners/base.py
  • mliport/mliport/runners/batch.py
  • mliport/mliport/runners/md.py
  • mliport/mliport/runners/md_output.py
  • mliport/mliport/runners/neb.py
  • mliport/mliport/runners/optimization.py
  • mliport/mliport/runners/safety.py
  • mliport/mliport/runners/singlepoint.py

tui (7)

  • mliport/mliport/tui/init.py
  • mliport/mliport/tui/analysis_screen.py
  • mliport/mliport/tui/app.py
  • mliport/mliport/tui/config_screen.py
  • mliport/mliport/tui/jobs_screen.py
  • mliport/mliport/tui/main_screen.py
  • mliport/mliport/tui/run_screen.py

writers (7)

  • mliport/mliport/writers/init.py
  • mliport/mliport/writers/contcar.py
  • mliport/mliport/writers/json_writer.py
  • mliport/mliport/writers/oszicar.py
  • mliport/mliport/writers/outcar.py
  • mliport/mliport/writers/trajectory.py
  • mliport/mliport/writers/xdatcar.py

11. Required progress update format

On every meaningful update, replace the revision/date and add one entry:

Rev N / YYYY-MM-DD / main SHA examined
Scope: source files read X/113 (refresh denominator), tests mapped Y/113, cases executed pass/fail/blocked/not_run
Evidence: links to separate fix/test PRs, known-answer data and CI run, hardware/runtime identity
New confirmed defects: IDs with one-sentence reproduction
Disproven hypotheses: IDs with the actual negative evidence
Scientific data invalidated: none / list of profile IDs and regeneration status
Remaining P0/P1 and next narrow independent PR

Don't use a checkbox alone as proof. New scientific problems are never "resolved" because a mock tests green. When an upstream version or feature PR changes a surface, reopen affected cases, and leave this audit PR open indefinitely.


12. Findings ledger — agent audit + GPU validation (2026-11-05)

Baseline examined: main @ 3d6ca2ea. Hardware: Tesla V100-SXM2-16GB, driver
580.173.02, CUDA 13.0. CPU core suite: 1694 passed / 33 skipped / 1 xfailed.
Real-model validation on CPU+GPU: MACE-OMAT-0 medium (d4b14be9…) and
GRACE-2L-OMAT-medium-base + fp64 (d497af4c… / 53b945ef…), both hash-verified
against model_manifest.json; 276 evidence records aggregated with no harness
violation (224 pass / 34 characterized / 14 fail / 4 unsupported).

Nothing below has been fixed unless the Status column says so. The venv path
issue is excluded by request.

12.1 Product-code findings — previous development round

ID Finding (one-sentence reproduction) Evidence Status
AUD-001 TrajectoryDataset.load cast the finite raw/md.csv step column with astype(int), so 0.5,1.5,2.5,3.5 loaded as 0,1,2,3; 2**53+1 was also rounded through float and a mixed [2**53+1, 2.0] sequence lost the integer identity. red→green tests in test_dataset_validation.py; PR #76 fixed, merged (#76, 2adfc2ef)
AUD-002 Standard XDATCAR has no PBC mask; pbc=(T,T,F) / (F,F,F) round-tripped to (T,T,T) and flipped the three_dimensional_pbc / eligible_for_rdf gates. test_xdatcar_convert.py; PR #77 fixed, merged (#77, rewritten by the owner into a PBC-provenance fix)
AUD-003 A crashed optimizer was swallowed and reported as converged=False → exit 2 (not_converged) instead of 1 (failed), contradicting docs/workflows/optimization.md. test_optimization_failure_outcome.py; PR #78 fixed, merged (#78)
AUD-006 (a) MDOutcarWriter.finalize() was first-call-wins, so the post-run guard's corrective finalize(status="failed") was a guaranteed no-op and a run that raised kept an OUTCAR saying completed. (b) The pre-existing terminal handlers called finalize()/_write_artifacts_manifest() unguarded, so a reporting OSError masked the real cause — a cancelled run with a full disk was classified FAILED (exit 1) instead of CANCELLED (130). review of PR #81; red tests test_md_manifest_write_failure_does_not_leave_a_completed_outcar, test_md_cancellation_survives_a_manifest_write_failure fixed, merged (#81, bf9d45ee)
batch outcome BatchRunner.run_from_files set success: True whenever _run_single returned, and _run_single dropped the optimization status/converged. flagged in review of #78; fixed separately fixed, merged (#80, 0d71f0a9)
S-02 The MD configurational_pressure_gpa / total_pressure_gpa conversion had no numeric known-answer test (only None-ness and inequality). known-answer tests with an independent SI constant; PR #79 fixed, merged (#79)
AUD-007 (confsearch) (a) A damaged checkpoint was resumed instead of refused: unit_index=10**6 made the unit loop skip everything and report a truncated search as completed (15 quenches/12 minima vs the reference's 21/15); an empty/partial rng_states silently re-seeded the replicas and diverged; a missing stream raised a raw KeyError. (b) A truncated checkpoint file raised a bare JSONDecodeError and the schema was never checked. (c) minima/all_minima.extxyz was not reproducible across a resume — the fixed-atom mask survived only on minima registered after the resume (15 of 21 frames lost it). tests/mliport/mliport/confsearch/test_resume_integrity.py (6 red→green, 1 xfail) fixed, PR #83 open draft
AUD-007 (NEB) NEB resume is geometry_resume_only, so the resumed optimiser restarts: the barrier is 0.9592721 eV uninterrupted vs 0.9590986 eV resumed (~2e-4 eV). Structure itself is restored byte-identically. tests/mliport/mliport/neb/test_neb_resume_equivalence.py; PR #82 documented limitation, PR #82 open draft

12.2 Validation-harness findings — this GPU round

ID Finding Evidence Status
H-01 validation/science/scripts/neb_suite.py::model_identity_check compares mliport's directory artifact hash against the manifest's single-file hash, so every GRACE t5h_saddle_hessian record is a false fail. mliport computes a8df2705… (= hash_model_path(dir), verified), the manifest pins d497af4c… (= saved_model.pb). The harness's own resolve_profile already handles directory artifacts via secondary_hashes, so the two disagree. The science in that record passed (ci_converged=true, min_eigenvalue = -2.04 eV/Ų, min_mode_overlap_vs_first = 0.99999998). validation-work/grace/t5/t5_saddle__t5h_saddle_hessian__grace.json open, not fixed
H-02 run_suite.py does not forward --subset to omat_eval.py, so run_suite.py --tiers t3 … dies with omat_eval.py: error: the following arguments are required: --subset. Calling the tier script directly works. val-mace-t3.log (first attempt) open, not fixed
H-03 T4 phonon records with a 2×2×2 supercell are reported as fail (robust imaginary modes) while the same fixture at 3×3×3 passes. This is a supercell-convergence artifact (see 12.3), but the status vocabulary has no way to say "converged at 3×3×3, not at 2×2×2" — the summary therefore reads as 12 failures. t4f_*_2x2x2_* vs t4f_*_3x3x3_* records, MACE + GRACE open, needs an owner call on status semantics

12.3 Characterised, NOT defects

Observation Why it is not a defect
T4 phonon imaginary modes on 2×2×2 only: Cu fcc -12.35 meV (68 modes), Si diamond -4.14 meV (6 modes); identical at δ = 0.005/0.01/0.02 Å, and 0 imaginary modes at 3×3×3 Amplitude-independent + supercell-dependent = long-range force constants under-converged in the small cell (classic metals artifact), i.e. a model+supercell characterisation. Also new evidence: the published campaign records phonons as not_run for every engine.
GRACE float32 t2_force_fd on MgO fail (2.6e-3 eV/Å) The analytic forces there are ~1e-7 eV/Å (near equilibrium), so the FD is at the float32 floor. The fp64 profile on the same fixture gives 7.1e-12 eV/Å → characterized. Precision characterisation.
T4 energetics = unsupported (2 records) No pinned isolated-atom reference energies — the harness refuses to fake them, as designed.
run_suite.py exit code 1 after an otherwise complete run It counts the T4 fail records; not a crash.

12.4 Blocked coverage (explicitly blocked, never silently skipped)

  • DPA-3.1-3M and UMA-s-1.2: no local pinned artifacts on this machine → T1–T8 for
    those engines were not run. Downloading checkpoints was not authorised.
  • test_grace_neighbor_cache_gpu.py (opt-in GPU gate): needs
    models/grace/GRACE-2L-SMAX-OMAT-large and local-data/LGPS222.vasp/LGPS222.vasp;
    neither exists → cannot run.
  • T7 analysis line: needs the kinisi/GEMDAT analysis environment; not configured.
  • Backend venvs are runtime-only (no pytest). The core suite therefore ran in
    .venv-dev; the real-backend coverage above comes from the harness, which is the
    project's designed path.

12.5 Environment / process observations

  • Installer inventory drift: doctor reports Installed environment differs from saved inventory: packages because torch-dftd was installed after the inventory was
    recorded. The inventory was deliberately not rewritten, since doing so would hide
    the drift.
  • mliport/utils.py is unreferenced dead code (all 8 public functions unused by
    production or tests); its check_structure_valid also disagrees with the real
    cell check in runners/base.py. Not removed (no unrelated refactors).
  • Pressure conversion is spelled inline in four places with two different literals
    (160.2177 vs 160.21766208, ~8e-9 relative). Not a correctness issue; a
    consolidation follow-up.
  • AUD-001 residual: a non-finite step column still falls back to
    info/arange(nframes) * stride instead of failing closed.
  • AUD-006 residual: a caller-supplied progress callback that raises on the terminal
    done event leaves artifacts.json = completed while the run raises (the manifest is
    correct about the MD; the callback is the caller's).
  • AUD-007 residual (needs an owner decision): the checkpoint stores only the
    representative geometry of each unique minimum, so after a resume every member of
    that minimum reports the representative's coordinates in all_minima.extxyz — 3 of 21
    frames, up to 0.19 Å off (energies identical). Fixing it is a checkpoint-layout
    change and therefore a CS_CHECKPOINT_SCHEMA_REVISION bump, which would make
    in-flight checkpoints non-resumable. Pinned as xfail(strict=True) in PR fix(confsearch): fail closed on a damaged checkpoint and keep the minima mask (AUD-007) #83.

12.6 Evidence locations

  • Validation records: ~/hdd500/dev/mliport-audit/validation-work/ (outside the
    repository; 276 records + summary.json).
  • Logs: ~/hdd500/dev/mliport-audit/logs/.
  • Scratch reproductions: ~/hdd500/dev/mliport-audit/repro/.
  • Nothing from this campaign was committed to the repository, and main was not
    modified.

Revision log


Rev 8 — 2026-11-06/07 · local full test campaign, H-01/H-02 repaired, NEP regression coverage

Appended, not replacing anything above. Baseline for this revision is
main @ 3d6ca2ea (clean detached worktree), origin/main @ 43c95804.
Implementation PRs: #87 (H-01 + H-02, open, not merged) and #88
(NEP dual runtime, open, not merged). Full round report:
#61 comment.

R8.1 Environment and baseline (per section 4.1 of the #61 acceptance matrix)

host        Linux vasp-v100, x86_64, 72 cores, 109 GiB RAM
GPU         Tesla V100-SXM2-16GB 16 GiB, uuid GPU-f24116e9-bea2-e0a6-70a9-d0b50811856b
driver      580.173.02 / CUDA 13.0 runtime, toolkit 12.9 (/opt/apps/cuda-12.9)
python      3.12.13
core venv   .venv-dev (editable, torch 2.14.1+cpu)
baseline    /tmp/mliport-baseline @ 3d6ca2ea in a pristine venv (/tmp/venv-base)
NEP venv    .venv-nep  (calorine 4.1, ase 3.29.0)
gpumd       5.9.1 source build, sm_70
            sha256 148259f2c24fe17c12be09fd0bba02528b98988cd86bd553fa896d6d94bedb97
NEP model   nep_BaZrO3.txt (GPUMD tests_pytest fixture)
            sha256 2a3a54346261d5c177f607d51d66f2f25c67c6e3d136ae94febab573929d3fb5
structure   BaZrO3-nat40-rattled.xyz (40 atoms)

Core suite, same command in both trees:

tree result
clean baseline 3d6ca2ea 1688 passed, 33 skipped, 0 failed
#87 (audit fixes) 1694 passed, 33 skipped, 0 failed (+6)
#88 (NEP) 1893 passed, 34 skipped, 0 failed (+205)

New skips explained: exactly one, test_nep_real_cpu.py, which skips in
.venv-dev because calorine is deliberately not installed there. It was
executed in .venv-nep: 1 passed. No xfail was added anywhere.

ruff check + ruff format --check (mliport, tests, scripts, validation):
clean. scripts/check_repository_hygiene.py: pass.
scripts/generate_docs.py --check: up to date. mkdocs build (en + zh): OK.
uv build mliport: wheel + sdist both contain calculators/nep_calc.py and
gpumd/*.

R8.2 H-01 — GRACE directory-artifact identity: reproduced and fixed (#87)

Confirmed with the unmodified historical record
validation-work/grace/t5/t5_saddle__t5h_saddle_hessian__grace.json:

status                                  = fail
metrics.model_identity.checks.model_sha256 = false
metrics.model_identity.record.model_sha256 = a8df2705…  (mliport tree hash)
record.model_sha256                        = d497af4c…  (manifest/archive id)
barrier_forward_sampled_eV = 0.7456, ci_converged = true,
active_physical_fmax_eV_A = 0.0295 ≤ 0.03, min_eigenvalue ≈ −2.041 eV/Ų

The physics criteria all passed; only the hash scheme differed. Fix: the
harness hashes the local artifact with mliport's own algorithm
(common.artifact_sha256, proven equal to
mliport.model_record.hash_model_path for a file and a directory by a new
test), compares like-for-like, and records manifest_model_sha256 separately.
resolve_profile's file-by-file secondary_hashes verification is unchanged,
and a genuinely different artifact still fails. No threshold relaxed.

R8.3 H-02 — --tiers t3 could never run: reproduced and fixed (#87)

run_suite.py launched omat_eval.py without its required --subset.
--subset is now forwarded and the tier refuses to start without it with an
actionable message. Two red→green tests cover forwarding and refusal.

R8.4 Two additional defects found this round (fixed in #88)

ID defect evidence fix
NEP-D01 the native GPUMD child never received CUDA_VISIBLE_DEVICES; the leased GPU UUID was recorded in provenance but not enforced, so a queued job could execute on a different GPU than its lease gpumd/runner.py::_execute_process passed only self.env export the leased UUID to the child; test asserts a fake binary actually observes it, and a real V100 run records actual_device_uuid == nvidia-smi
NEP-D02 NVT run.in provenance recorded thermostat: null while thermostat_mapping.thermostat was LANGEVIN real out-native-nvt/manifest.json top-level thermostat now set from the mapping; test_gpumd_input.py covers it

R8.5 Per-file ledger progress

The 113-file ledger above is not bulk-ticked. This revision adds function-level
evidence for:

  • mliport/calculators/factory.py, backend_selection.py,
    config/{schema,defaults,resolver,incar}.py, engine.py, api.py,
    cli.py, model_record.py, doctor.py, queue.py,
    tui/{config_screen,run_screen}.py, install/{cli,plan,compatibility}.py
    — each asserted by test_nep_config.py (49 tests) to expose the same backend
    set and the same runtime contract, so a future backend cannot be added to one
    entry point only.
  • mliport/runners/{base,md,optimization,singlepoint}.py — unchanged; NEP-ASE
    reuses them (real CPU sp/opt/nvt/nve/neb runs recorded in [DEV MASTER · DO NOT MERGE/CLOSE] MLIP validation roadmap (#75) + confsearch living spec #61).
  • validation/science/scripts/{common,engines,neb_suite,run_suite}.py — H-01/H-02.

Still not function-by-function signed off: the analysis package, confsearch,
writers, install/inventory, jobs/queue_daemon, tui screens beyond the config
form. No claim is made for them.

R8.6 Real-backend re-verification on the modified tree

backend records result
MACE mace_omat t1+t2 27 23 pass / 4 characterized / 0 fail; t1 energies bit-identical to the 2026-11-05 records
GRACE grace_omat t1+t2 54 46 pass / 8 characterized / 0 fail
NEP (13 acceptance cases) 13 13 pass on V100 (CPUNEP cpu, GPUNEP gpu, native sp/opt/nve/nvt, FD stress, invariance, fail-closed)
DPA 0 blocked — no checkpoint on this machine
UMA 0 blocked — no checkpoint; not downloaded (no authorisation)

No old evidence file was overwritten; the new records live in disposable
directories outside the repository.

R8.7 Still open — unchanged owner decisions

R8.8 DATA IMPACT

None. No stored result was modified. Historical GRACE saddle records written as
fail by H-01 are misclassified, not wrong physics; they must be re-run
with the corrected harness before being cited. No scientific record
invalidated, none rewritten.

Revision log entry

Documentation only in pull request body; no tree changes.

DATA IMPACT: no scientific record invalidated.

Copy link
Copy Markdown
Owner Author

Round-1 supervisory review — 2026-10-31 (main @ 1ef9a1b)

Reviewed the source diffs, PR bodies and GitHub CI check-runs for audit PRs #76–#79. All four have green applicable GitHub Actions on Python 3.10–3.13; none has been merged. These are code/CI reviews, not independent full GPU science reproductions.

PR Review outcome Open action
#76 / AUD-001 Confirmed fractional-step truncation and improved fail-closed guard; follow-up requested Guard precision loss above 2**53 for integer input / md.csv parsing; clarify int64 edge and nonfinite CSV fallback
#77 / AUD-002 Reproduction valid; blocked on owner choice Pick explicit warning+metadata vs writer fail-closed; replace provisional xfail with the decided acceptance contract
#78 / AUD-003 Fix improves failure classification and JSON evidence; follow-up requested Ensure failed optimizer never emits 'done' progress; verify repeated backend failure and separately track batch summary masking
#79 / S-02 Test-only, no blocking diff concern Safe to consider normal merge after current-head check; nonzero kinetic-stress and writer known answers remain untested

Review comments were posted on the four implementation PRs. No approval, merge, close or feature change occurred. Keep this PR open and Draft; update this tracker only with actual red/green evidence and merged fixes.

hydrogen1222 added a commit that referenced this pull request Nov 1, 2026
…cating

AUD-001 (tracker #75).  TrajectoryDataset.load cast the finite raw/md.csv
"step" column with astype(int), so steps 0.5, 1.5, 2.5, 3.5 loaded as
0, 1, 2, 3.  The fabricated indices were then re-exported as the step axis
(thermodynamic_diagnostics "step" column) and used to recompute
frame_stride_steps in analysis_view()/slice_frames().  A step beyond the
int64 range wrapped as well, with a numpy "invalid value encountered in
cast" RuntimeWarning.

Step indices are integers by contract: mliport writes int steps and
_uniform_integer_step already treats non-integral values as unknown.  Add
_validated_steps(), which requires finite, integral, in-range values, and
use it in both cast sites (TrajectoryDataset.load and
TrajectoryDataset.from_frames).  Invalid input now fails closed with a
clear ValueError instead of being silently reinterpreted.

Tests (red before the fix, green after):
  test_fractional_md_csv_steps_are_rejected_not_truncated
  test_out_of_range_md_csv_steps_are_rejected
  test_from_frames_rejects_fractional_steps
  test_integral_float_md_csv_steps_still_load (guard)

DATA IMPACT: no scientific record invalidated.  No stored trajectory is
rewritten; only the handling of an already-invalid step column changes,
from silent truncation to an explicit error.

Tracking: #75
hydrogen1222 added a commit that referenced this pull request Nov 1, 2026
…float

AUD-001 review follow-up (tracker #75, PR #76).

_validated_steps() began with np.asarray(values, dtype=float), so an
integer input such as steps=[2**53 + 1] silently became 2**53 before the
integrality/range checks could see it, and the md.csv step column was
parsed through float() for the same reason.  The symmetric abs() >= 2**63
bound also rejected the representable int64 minimum -2**63.

Dispatch on the input dtype instead:

  * integer input is never routed through float; only the asymmetric
    int64 range is enforced (so -2**63 is accepted, 2**63 is rejected);
  * float input must be finite, integral and within the exactly
    representable float range (|v| <= 2**53);
  * a CSV step column is parsed exactly from its text with Decimal, so
    "9007199254740993" and "9007199254740993.0" both keep their exact
    integer value.  A finite non-integral cell is rejected; blank,
    unparseable or non-finite cells keep the pre-existing "axis unknown"
    fallback (unchanged, tracked separately).

Regressions added (red before this commit, green after):
  test_from_frames_preserves_large_integer_steps_exactly
  test_from_frames_accepts_int64_min_and_rejects_overflow
  test_md_csv_huge_integer_step_text_is_parsed_exactly
  test_md_csv_float_literal_step_keeps_its_exact_integer_value
  test_md_csv_unparseable_step_makes_the_axis_unknown

DATA IMPACT: no scientific record invalidated.  No stored trajectory is
rewritten; only the handling of a step column changes.

Tracking: #75
hydrogen1222 added a commit that referenced this pull request Nov 1, 2026
AUD-003 (tracker #75).  OptimizationRunner.run caught every non-NaN
optimizer/calculator exception, logged it, re-evaluated energy/forces and
returned results with converged=False.  classify_result maps that to
NOT_CONVERGED (exit 2), so a backend crash was indistinguishable from an
honest "hit max_steps" run: exit 2 instead of 1, queue status
"not_converged" instead of "failed", and no record of the error outside
the log.  This contradicted docs/workflows/optimization.md, which promises
that "a backend or runtime error exits with 1, and partial results are
kept".

Keep the partial results, but record the failure: set results["status"] =
"failed" and results["error"] to the exception summary.  JsonWriter now
carries status/error into mliport_results.json for optimization runs
instead of dropping them.  The NaN/inf abort path still re-raises.

Tests (red before, green after):
  test_optimizer_exception_is_reported_as_failed_not_non_convergence
  test_optimizer_exception_is_recorded_in_results_json

DATA IMPACT: no scientific record invalidated.  No stored trajectory or
structure is rewritten; only the outcome classification and the
machine-readable record of an already-failed optimization change.

Tracking: #75
hydrogen1222 added a commit that referenced this pull request Nov 1, 2026
AUD-003 review follow-up (tracker #75, PR #78).

BaseRunner deferred only the ``done`` progress phase, and
OptimizationRunner emitted ``done`` ("Optimization not converged ...") even
after an optimizer exception, so a UI/progress consumer received a
success-looking terminal phase for a failed run.

* _emit_progress() now defers the ``error`` phase as well as ``done``, and
  execute() publishes the deferred terminal event with the timing attached,
  so the terminal phase reflects the outcome.  (Renamed the private
  _pending_done_event to _pending_terminal_event.)
* OptimizationRunner emits an ``error`` event when the optimizer crashed;
  an honest max_steps non-convergence still ends on ``done``.
* The post-crash energy/force re-evaluation is wrapped: if it also fails,
  the raised error names both causes and chains the second via __cause__
  instead of replacing the original crash.  CancellationRequested is
  excluded from that wrap so a cancel stays a cancel.

Tests (red before, green after):
  test_failed_optimization_publishes_error_not_done
  test_reevaluation_failure_keeps_the_original_crash
  test_cancellation_during_reevaluation_is_not_wrapped
guards:
  test_honest_non_convergence_still_publishes_done

DATA IMPACT: no scientific record invalidated.  Only the terminal progress
phase and the failure message of an already-failed optimization change.

Tracking: #75

Copy link
Copy Markdown
Owner Author

Review round 2 — 2026-11-01 · main @ 77a94af

Only GitHub source diffs and CI evidence were reviewed by the supervisor. No real MLIP GPU inference independently executed. No merge or status change performed by reviewer. Keep #75 open and update status/checkboxes only where accepted evidence is recorded.

hydrogen1222 added a commit that referenced this pull request Nov 1, 2026
…coercion

AUD-001 round-2 review follow-up (tracker #75, PR #76).

_validated_steps still began with np.asarray(values), so a heterogeneous
Python sequence such as [2**53 + 1, 2.0] was promoted to float64 and the
integer was silently rounded to 2**53 before any dtype check could see it
-- exactly the data-identity failure AUD-001 exists to prevent.

* An np.ndarray keeps the vectorised dtype path (_validated_step_array).
* Any other sequence is validated element by element via the new
  _step_index(), which checks Python int first (exact, never through
  float), then float (finite / integral / |v| <= 2**53), and rejects
  everything else.  The int64 array is only built afterwards, so numpy
  promotion can no longer round a value.

Secondary hardening: _parse_step_column now compares the Decimal against
the int64 bounds before int(integral).  This is a real robustness fix, not
just a micro-optimisation -- on the previous commit int(Decimal("1e999999999"))
tried to build a ~10**9-digit integer and the process hung.

Regressions (red before, green after):
  test_from_frames_preserves_mixed_integer_and_float_steps_exactly
  test_from_frames_rejects_invalid_mixed_sequences (4 parametrised cases)
  test_md_csv_astronomical_step_exponent_is_rejected

DATA IMPACT: no scientific record invalidated.  No stored trajectory is
rewritten; only the handling of a step axis changes.

Tracking: #75

Copy link
Copy Markdown
Owner Author

Round-3 review update — 2026-11-01, main @ 77a94af

No automatic merges, approvals, scientific data rewrites or physical GPU revalidation performed in this review. Keep #75 open as the audit tracker.

hydrogen1222 added a commit that referenced this pull request Nov 2, 2026
AUD-007 (tracker #75).  No production behaviour changes; test-only plus one
documentation sentence.

confsearch: the engine already checkpoints the per-replica md_rng/mc_rng
bit-generator states after every unit and restores them in _restore().  The
new test kills a run after every durable unit boundary (1-3 units for one
replica / three cycles, 1-4 for two replicas / two cycles) and asserts the
resumed run's search_results.json is identical to an uninterrupted run's,
apart from model.model_path (not identity).  Verified non-vacuous: skipping
the RNG restore or perturbing the restored geometry each make it fail.

NEB: the checkpoint stores the whole band plus band_sha256, and the loader
verifies the checksum and re-validates the band.  The new tests show the
resumed attempt starts from a byte-identical band, that the interrupted
attempt's checkpoints and exports survive, and that the resumed run converges
to the same climbing image and saddle.  NEB has no stochastic optimizer, so
there is no generator state to restore.

The checkpoint is geometry_resume_only, so a resume restarts the optimiser
instead of replaying its state: in the fixture the barrier is 0.9592721 eV
uninterrupted vs 0.9590986 eV resumed (~2e-4 eV).  That limitation was already
recorded in the checkpoint and run_context.json; document it in
docs/workflows/neb.md and its Chinese twin as well.

DATA IMPACT: no scientific record invalidated (test-only + docs).

Tracking: #75
hydrogen1222 added a commit that referenced this pull request Nov 2, 2026
… review)

Independent-review follow-up on PR #81.  Two defects found by fault injection;
neither touches the integrator, model inference, energy or forces.

1. The post-run guard could not correct the OUTCAR it had already written.
   MDOutcarWriter.finalize() returned early once _finished was set, so when
   _write_artifacts_manifest(status="completed") failed -- the exact boundary
   the guard exists for -- the corrective finalize(status="failed") was a
   silent no-op and the OUTCAR kept claiming "completed" for a run that
   raised.  finalize() now appends a corrective summary when the status
   differs (the earlier block is kept: evidence is never deleted) and the
   last Status: line is the authoritative one.

2. Best-effort reporting could mask the original exception.  The pre-existing
   terminal handlers (output drain, force abort, cancellation, Ctrl-C, generic
   failure) called finalize() and _write_artifacts_manifest() unguarded, so a
   reporting OSError replaced the real cause: a cancelled run with a full disk
   was classified FAILED (exit 1) instead of CANCELLED (exit 130).  The
   post-run guard's own best-effort pattern is now one publish_terminal()
   helper used by every terminal path, so reporting failures are suppressed
   and the original exception always propagates.

Regressions (red on 8eee6f2, green now):
  test_md_manifest_write_failure_does_not_leave_a_completed_outcar
  test_md_cancellation_survives_a_manifest_write_failure

DATA IMPACT: no scientific record invalidated.  No stored trajectory, md.csv
or energy/force value changes; only the terminal status trailer of a failed
MD OUTCAR and which exception propagates from a reporting failure.

Tracking: #75
@hydrogen1222 hydrogen1222 changed the title [DEV · DO NOT MERGE/CLOSE] Living code audit & exhaustive scientific test campaign [DEV P0 AUDIT · DO NOT MERGE/CLOSE] Living correctness and real-hardware test ledger · master #61 Nov 6, 2026
@hydrogen1222

Copy link
Copy Markdown
Owner Author

Model-support round · callable checks on the V100 host (2026-11-14)

Evidence for #61 model-support spec Rev 2 §5. Scope: "called correctly", not accuracy.
No benchmark, no DFT, no long MD.

Environment

Item Value
Baseline main@ec28ba3f (PR #82 merged)
Implementation PRs #94 (registry), #95 (probe), #96 (MACE gate), #97 (UMA runtime), #98 (DPA widening fix), #99 (status docs)
GPU Tesla V100-SXM2-16GB, CC 7.0, 16384 MiB, driver 580.173.02
GPU UUID (sha256) c0c37a4d8178c0b3b6f37d93b2c5f835972fbe611593bf16204793867b4f202a
MACE env mace-torch 0.3.16, torch 2.8.0+cu126, e3nn 0.4.4, .venv-mace
DPA env deepmd-kit 3.2.1.dev63+g0edeaef56 (source @0edeaef), torch 2.11.0+cu126, .venv-dpa
GRACE env tensorpotential 0.6.0, TensorFlow 2.20.0, .venv-grace
UMA env fairchem-core 2.25.0, torch 2.13.0+cu126 (arch list sm_50/60/70/75/80/86/90), .venv
mliport 2.0.0b4 (editable, this round's branches)

Structures:

  • examples/structures/li10gep2s12_primitive.vasp, 50 atoms, sha256
    0e0702be3a31778839f730430c6d7989118df316c4976a1a98f945c32fd3561a
    — used for the single-point smoke on every model (doctor --structure).
  • built-in probe fixture: 4-atom fcc Cu, a=3.65 Å, one atom displaced by
    (0.07, -0.03, 0.02) Å — used for the shared-calculator/FD/MD/NEB checks,
    because the probe refuses >32 atoms by design.

Commands (per model; CUDA_VISIBLE_DEVICES set to '' for --device cpu and
to the GPU UUID for --device cuda:0, as DPA and GRACE require):

mliport doctor --engine E --device D --model M --task bulk [--head H] [--dtype D] \
  --structure examples/structures/li10gep2s12_primitive.vasp
python scripts/validate_runtime.py E M report.json --device D --task bulk [--head H] \
  [--dtype D] --full --raw-dir raw/

--full runs single point, shared-calculator repeat, central-difference forces,
10 NVE steps and a five-image NEB on the Cu fixture; its endpoints are relaxed,
saddle_validation: not_performed.

Identity read-out (verbatim from the artifacts)

Label Artifact SHA-256 Read-out
mace_omat d4b14be9afa294eebdbe31a0280b26a0fa29715771e978cbfd3ca24e0d90307a 1 head default, r_max 6.0 Å, 89 elements, float64
mace_matpes_r2scan 8f147ecffa1d06a696e5648b13b81abda9c96fda9d0fb3faba3cd3fb27d9bba9 1 head default, r_max 6.0 Å, 89 elements, float64
mace_mh1_* ec00a2705854622fbbd898ccfb7701072fcd674709102d009fb919c1b8cc5dde 6 heads [matpes_r2scan, mp_pbe_refit_add, spice_wB97M, oc20_usemppbe, omol, omat_pbe], r_max 6.0 Å, 89 elements; embedding_specs absent (not an OMOL embedding model)
dpa_omat, dpa31_sse_* 86dd3a804d78ca5d203ebf98747e8f16dff9713ba8950097ceb760b161e19907 multitask, 32 branches incl. Domains_SSE_PBE, Domains_SSE_PBESol, Omat24; type_map 96 elements
dpa4_neo/air/plus fd7f34ae28f921201e4a0328ddf56179892752084f8363e7796038201471c989 / 7ac5fb696fc057229ceebc33f31f7a96b2f3616ac5b63a512acd1848f2d41a09 / 6820a320dc241a4002e6c257ad21983950df178b1f80271a6e6af199adb18567 singletask, type_map 96 elements, tag v20260805
grace_omat dir hash a8df2705e39a12bc6e1bc85b560a5329f42fbe0ca67f1952cd5f9ca218bc0a16 metadata: 89 elements, cutoff 6.0 Å, has_uq: false
grace_omat_fp64 dir hash e717cef1e41d8ba501f50101ddf783c3249ff8b8ddfbe59b61bc3c8cba2af066 same metadata, checkpoint variables float64
grace_3l_omat_large_ft_am dir hash fc88052f2bbf5908f949c1f6c74062c21337c8e2ed5ac2d9f5a184324350a956 89 elements, cutoff 6.0 Å
grace_2l_smax_omat_large dir hash 87e72b50a2eb151f00242a0ad7ab1253d507860f9da837cdc2f9010371b1490a 94 elements, cutoff 7.5 Å (element-dependent cutoff matrix)
uma_omat, uma_s_1p2p1* not acquired facebook/UMA is gated (401 without a token)

Directory hashes are sha256 over sorted relpath length + relpath + content;
source URLs and licenses: mace-foundations release assets (ASL), DPA-3.1-3M
(CC-BY-4.0), DPA4-OMat24 (CC-BY-NC-4.0), AMS-ICAMS-RUB/grace-foundation-models
(Academic Software License), facebook/UMA (FAIR non-commercial). Cards read
2026-11-14.

Results · MACE

Model device doctor SP (50 atoms) probe SP + FD 10 NVE 5-image NEB
mace_omat cpu, cuda:0 passed, E=-216.11142 eV, stress checked passed passed passed
mace_omat_float32 cpu, cuda:0 passed, E=-216.11142 eV passed passed passed
mace_matpes_r2scan cpu, cuda:0 passed, E=-359.92815 eV passed passed passed
mace_mh1_omat_pbe cpu, cuda:0 passed, E=-216.20645 eV passed passed passed
mace_mh1_matpes_r2scan cpu, cuda:0 passed, E=-359.50212 eV passed passed passed

Results · DPA

Model device doctor SP (50 atoms) probe SP + FD 10 NVE 5-image NEB
dpa_omat (Omat24) cpu, cuda:0 passed, E=-216.10157 eV passed passed passed
dpa31_sse_pbe cpu, cuda:0 passed, E=-216.22144 eV passed passed passed
dpa31_sse_pbesol cpu, cuda:0 passed, E=-162.85829 eV passed passed passed
dpa4_neo_omat cpu, cuda:0 passed, E=-216.23211 eV passed passed passed
dpa4_air_omat cpu, cuda:0 passed, E=-216.32644 eV passed passed passed
dpa4_plus_omat cpu, cuda:0 passed, E=-216.35980 eV passed passed passed

Results · GRACE

Model device doctor SP (50 atoms) probe SP + FD 10 NVE 5-image NEB
grace_omat cpu, cuda:0 passed, E=-216.15582 eV passed passed passed
grace_omat_fp64 cpu, cuda:0 passed, E=-216.15582 eV passed passed passed
grace_3l_omat_large_ft_am cpu, cuda:0 passed, E=-216.23825 eV passed passed passed
grace_2l_smax_omat_large cpu, cuda:0 passed, E=-216.28754 eV passed passed passed

Results · UMA

not_run. The pinned runtime was upgraded and the installer's closing
mliport doctor --engine uma --device auto passes on this host (fairchem-core
2.25.0, torch 2.13.0+cu126, real tensor dot product on cuda:0), but no UMA
checkpoint could be loaded: https://huggingface.co/facebook/UMA is gated
(HTTP 401 without a token, none configured on this host). No UMA number is
claimed.

CPU vs GPU agreement (§5.3)

Same model, same 50-atom LGPS cell, same task/head/dtype; forces compared
element-wise.

Model dtype reported |ΔE|/atom (eV) max |ΔF| (eV/Å)
mace_omat float64 1.1e-15 3.0e-14
mace_mh1_matpes_r2scan float64 0.0 1.3e-14
mace_matpes_r2scan float64 0.0 2.7e-14
mace_omat_float32 float32 6.1e-07 7.5e-06
grace_omat float32 1.9e-08 1.8e-06
grace_3l_omat_large_ft_am float32+float64 tensors 2.4e-08 8.6e-07
grace_2l_smax_omat_large float32 9.5e-08 1.2e-06
dpa_omat float32 3.8e-08 3.9e-06
dpa4_neo_omat float32 2.2e-08 3.0e-06

On the Cu probe fixture the same comparison gives |ΔE|/atom ≤ 5e-07 eV and
max Δ(max|F|) ≤ 3e-07 eV/Å for all 15 models. All values are inside the spec's
thresholds (float64: 1e-6 eV/atom, 1e-5 eV/Å; float32/model-defined: 1e-4,
1e-3), so nothing was flagged.

Failures found by these runs (verbatim) and their fixes

  1. DPA-3.1 could not load on the pinned DPA4 runtime.
    RuntimeError: DPA checkpoint tensor changed during loading: atomic_model.out_bias; refusing changed values, shapes or dtype.
    Cause: deepmd 3.2.1 builds the three statistics tensors (out_bias,
    out_std, bias_atom_e) as float64 while the distributed checkpoint stores
    float32; all 351 keys and every value are unchanged. Fixed by accepting only
    a strictly wider floating type with bit-identical values and recording the
    widening in the checkpoint migrations (still refusing narrowing, shape and
    value changes). Before the fix, DPA-3.1 only loaded on the older deepmd
    3.1.3, which cannot read the DPA4 checkpoints at all
    (RuntimeError: Unknown model type: dpa4.).
  2. UMA preset semantics changed upstream. fairchem 2.22+ made UMA Nitro the
    upstream default and removed turbo
    (ValueError: inference_settings='turbo' is no longer available...), which
    would have silently switched mliport's default to reduced precision and
    exposed Nitro on Ampere+. Fixed by mapping mliport's two names onto upstream
    batch (old default numbers) and multigpu (upstream's "turbo's former
    settings"), refusing any preset that sets nitro, and recording the
    effective preset/tf32/nitro flags in the model identity.
  3. Probe defects (fixed in the probe PR): GRACE refused the GPU-only
    gpu_memory_limit_mb on CPU; DPA and GRACE refuse to run in a process that
    never set CUDA_VISIBLE_DEVICES
    (DPA requires process-level device isolation before framework imports...);
    the five-image NEB failed its own endpoint check
    (Endpoint 0 fmax=0.481702 eV/A exceeds endpoint_fmax_eV_A=0.02) until the
    probe relaxed endpoints.
  4. Not a model defect: the first MACE doctor run reported
    Installer inventory mismatch: packages because .venv-mace predated its
    saved inventory. After re-running the installer for MACE the check is green;
    the model checks had already passed.

Residual gaps

  • UMA (3 profiles) is not_run: gated weights, no token. The runtime upgrade
    itself is verified only by the installer/doctor checks, not by a model call.
  • No FD/MD/NEB was run on the 50-atom LGPS cell: the probe caps the fixture at
    32 atoms. Those cells got single-point energy/force/stress only.
  • NEB bands were not saddle-validated (saddle_validation: not_performed);
    convergence is not a verified transition state.
  • Nothing here says anything about accuracy on sulfides, about stress
    correctness, or about long-MD stability.
  • Out of scope and untested: DPA4C .pt2, DPA-SSE, GRACE uncertainty heads,
    MACE-POLAR/MDP/OMOL physics (they are refused), UMA Nitro, NEP89 on CPU.

@hydrogen1222

Copy link
Copy Markdown
Owner Author

Model-support round · UMA callable checks on the V100 host (2026-11-14)

Follow-up to the model-support evidence comment,
covering §11 task 1: the UMA checkpoints the owner downloaded into models/,
checked on the #97 runtime (fairchem-core 2.25.0, torch 2.13.0+cu126).

Same host and conventions as the previous comment: Tesla V100-SXM2-16GB
(CC 7.0, driver 580.173.02), GPU UUID sha256
c0c37a4d8178c0b3b6f37d93b2c5f835972fbe611593bf16204793867b4f202a,
examples/structures/li10gep2s12_primitive.vasp (50 atoms, sha256
0e0702be…fd3561a) for the single point and the CPU/GPU deltas, the built-in
4-atom Cu cell for the --full probe.

Checkpoints

Label File SHA-256 (local) Task list read from the checkpoint Upstream
uma_omat uma-s-1p2.pt 45893aa31458da691f861cb0d8b857ef5935144d95bb2afb2912abd364012642 not readable: file is corrupt (see below) facebook/UMA checkpoints/uma-s-1p2.pt, 2333393167 bytes
uma_s_1p2p1_omat uma-s-1p2p1.pt b2673b85037b075674c25f55c34ffe1ff1e15db924be977b10a184765df0d5ce omat (UMATask: omol, omat, odac, oc20, oc25, omc) facebook/UMA checkpoints/uma-s-1p2p1.pt, 2333393615 bytes
uma_s_1p2p1_ft_r2scan_omat uma-s-1p2p1-ft-r2scan.pt c2fdf4f27323fae4f3a25b19bb0139e8bdaa3ba3688c3e83e5410b6aedc3bbab omat facebook/UMA checkpoints/uma-s-1p2p1-ft-r2scan.pt, 2330618438 bytes

All three local sizes equal the upstream tree listing byte-for-byte. License:
FAIR Chemistry License (non-commercial research); the repository is gated, so
this host cannot re-fetch (HTTP 401).

uma_omat (uma-s-1p2.pt) — blocked, artifact corrupt

Loading it fails on both devices, before any mliport code decides anything:

PytorchStreamReader failed reading file data/115: invalid header or archive is
corrupted. This is an internal miniz error. If you are seeing this error, there
is a high likelihood that your checkpoint file is corrupted. This can happen if
the checkpoint was not saved properly, was transferred incorrectly, or the file
was modified after saving.

Independent check of the zip container: zipfile.ZipFile(...).testzip() returns
the member inference_ckpt_stage2/data/114 (CRC mismatch), while
uma-s-1p2p1.pt and uma-s-1p2p1-ft-r2scan.pt pass testzip() over all 315 and
303 members. Size matches upstream, so the transfer was silently damaged.

Per spec §1 rule 9 this model is recorded and stopped; nothing is reshaped,
renamed or re-exported. To re-run it, re-download with an account that has
access to facebook/UMA and repeat the same commands. Status for this round:
not_run (blocked), not a fail of the adapter.

Check cpu cuda:0
mliport doctor single point (50 atoms) failed: checkpoint corrupt failed: checkpoint corrupt
validate_runtime.py --task omat --full failed: checkpoint corrupt failed: checkpoint corrupt

uma_s_1p2p1_omat — verified on both devices

Check cpu cuda:0
mliport doctor --task omat --structure li10gep2s12_primitive.vasp passed, E=-216.344373 eV, stress checked passed, E=-216.344373 eV
validate_runtime.py --task omat --full (Cu) passed: SP, shared calculator, FD, 10 NVE, 5-image NEB passed: same
recorded identity task: omat, inference_mode: default, inference_preset: batch, nitro_enabled: false, tf32_enabled: false, compile_enabled: false same, actual_device_type: cuda

LGPS CPU–GPU: |ΔE|/atom = 1.03e-08 eV, max |ΔF| = 3.40e-06 eV/Å.

Turbo single point (GPU, §11 task 1 second half):

Run Recorded identity Energy (LGPS)
mliport sp ... --inference-mode turbo inference_mode: turbo, inference_preset: multigpu, nitro_enabled: false, tf32_enabled: true, compile_enabled: true -216.344368 eV

So --inference-mode turbo reproduces upstream's "turbo's former settings"
(multigpu) and still cannot select Nitro.

uma_s_1p2p1_ft_r2scan_omat — callable, NEB refused by the physics gate

Check cpu cuda:0
mliport doctor --task omat --structure li10gep2s12_primitive.vasp passed, E=-417537.948266 eV, stress checked passed, E=-417537.948269 eV
probe single point / FD / 10 NVE passed passed
probe 5-image NEB refused (verbatim below) refused
recorded identity inference_preset: batch, nitro_enabled: false, tf32_enabled: false same

LGPS CPU–GPU: |ΔE|/atom = 4.57e-08 eV, max |ΔF| = 4.11e-06 eV/Å.

The NEB leg stops in mliport's own gradient preflight, which is the intended
fail-closed behaviour and not an adapter defect:

mliport.neb.schema.NEBGradientConsistencyError: NEB refused: the model's forces
are not verifiably the gradient of its energy (12/12 checks failed; worst:
image 4, force direction, an energy resolution too coarse to discriminate the
derivative (noise allowance 1.62 eV/A); use float64 or a smaller cell). NEB
optimizes with these forces, so the band and any barrier would be meaningless.
Use a conservative (autograd-force) model or precision.

Reason: the r²SCAN head carries total energies of order -1.8e5 eV for the tiny
Cu fixture, so float32 energy differences at a 1e-3 Å displacement are inside
the numerical noise. Single point, finite-difference forces and MD are fine; NEB
on that model/precision/cell is unsupported, and mliport says so instead of
producing a band.

Residual gaps

  • uma_omat is blocked by a corrupt local checkpoint; re-download needed.
  • UMA Nitro is never exercised by design (mliport refuses nitro presets);
    the multigpu preset was verified on Volta, where TF32 is a no-op.
  • The NEB inside the probe is not saddle-validated (saddle_validation: not_performed), and the LGPS cell gets single points only (the probe caps the
    fixture at 32 atoms).
  • No accuracy statement of any kind.

@hydrogen1222

Copy link
Copy Markdown
Owner Author

LGPS CPU–GPU agreement · complete table (2026-11-14)

Follow-up to the MACE/DPA/GRACE comment
and the UMA comment:
the six profiles that comment did not list (mace_mh1_omat_pbe, dpa31_sse_pbe,
dpa31_sse_pbesol, dpa4_air_omat, dpa4_plus_omat, grace_omat_fp64), plus
the two UMA profiles, measured the same way: same model, same
li10gep2s12_primitive.vasp (50 atoms, sha256 0e0702be…fd3561a), same
task/head/dtype, energy per atom and element-wise force difference.

Profile dtype reported |ΔE|/atom (eV) max |ΔF| (eV/Å)
mace_omat float64 1.14e-15 2.99e-14
mace_omat_float32 float32 6.10e-07 7.48e-06
mace_matpes_r2scan float64 0.0 2.71e-14
mace_mh1_omat_pbe float64 5.68e-16 1.24e-14
mace_mh1_matpes_r2scan float64 0.0 1.25e-14
grace_omat float32 1.91e-08 1.83e-06
grace_omat_fp64 float64 0.0 4.11e-14
grace_3l_omat_large_ft_am float32+float64 2.38e-08 8.61e-07
grace_2l_smax_omat_large float32 9.54e-08 1.23e-06
dpa_omat float32 3.76e-08 3.89e-06
dpa31_sse_pbe float32 4.78e-08 3.87e-06
dpa31_sse_pbesol float32 5.42e-08 3.83e-06
dpa4_neo_omat float32 2.18e-08 2.99e-06
dpa4_air_omat float32 2.02e-08 2.29e-06
dpa4_plus_omat float32 1.74e-08 2.03e-06
uma_s_1p2p1_omat model-defined 1.03e-08 3.40e-06
uma_s_1p2p1_ft_r2scan_omat model-defined 4.57e-08 4.11e-06
uma_omat — not measured (corrupt checkpoint) not measured

Measured maxima across every profile: 6.10e-07 eV/atom and
7.48e-06 eV/Å, both from mace_omat_float32. The spec's flagging thresholds
(float64: 1e-6 eV/atom, 1e-5 eV/Å; float32 or model-defined: 1e-4 eV/atom,
1e-3 eV/Å) were not reached by any profile, so nothing was flagged.

hydrogen1222 added a commit that referenced this pull request Nov 14, 2026
Tracking: #61 (model-support spec §6.5 + §11 tasks 1 and 5)

What changes:
- The status table leaves its initial state and carries what was measured on
  2026-11-14 on the owner's V100 in the installer-built environments: 17
  profiles `verified` on CPU and GPU (15 MACE/DPA/GRACE + the two UMA
  `uma-s-1p2p1` checkpoints), each with finite energy/force/stress on the
  50-atom LGPS cell, a matching recorded identity and the bounded probe
  (shared calculator, central differences, 10 NVE steps, five-image NEB).
- UMA rows filled from §11 task 1: `uma_s_1p2p1_omat` and
  `uma_s_1p2p1_ft_r2scan_omat` verified on both devices and recorded with
  `inference_preset: batch`, `nitro_enabled: false`, `tf32_enabled: false`;
  `--inference-mode turbo` records `multigpu` with `nitro_enabled: false`. The
  r2SCAN fine-tune's NEB leg is refused by the gradient preflight (energy
  resolution), which the table states instead of hiding.
- `uma_omat` stays `untested`: the local `uma-s-1p2.pt` fails the zip CRC
  (`PytorchStreamReader ... data/115`), the repository is gated and this host
  cannot re-fetch it.
- The CPU-GPU claim now quotes the measured maxima (6.1e-7 eV/atom, 7.5e-6
  eV/Angstrom, `mace_omat_float32`) instead of a rounded "1e-7/1e-5 or better".
- Every Evidence cell links the specific #75 comment (runs, deltas, UMA record,
  NEP89 record) instead of the pull request.

DATA IMPACT: no scientific record invalidated. Statuses describe calling the
models correctly; no accuracy claim is made and no stored campaign record
changed.
@hydrogen1222

Copy link
Copy Markdown
Owner Author

uma_omat (uma-s-1p2.pt) re-checked after the artifact was re-uploaded

Follow-up to the UMA record
and the delta table.
The owner re-downloaded models/uma-s-1p2.pt; the new copy is intact and this
model now completes the §5 checks.

Artifact

Item Value
SHA-256 ba5c0d912efa22dc238e5fb1b5b7f66ee2e68c48c1b95b7cfd7fe1da5938398b
Size 2333393167 bytes (matches checkpoints/uma-s-1p2.pt in the upstream tree)
Container zipfile.testzip() → clean over all 315 members
Upstream pin matches validation/science/model_manifest.json profile uma_omat (ba5c0d912efa…)
Task list omat (UMATask: omol, omat, odac, oc20, oc25, omc)

The previous copy had the same size but failed the CRC on
inference_ckpt_stage2/data/114, so it was a transfer corruption, not an
adapter problem.

Checks (Tesla V100-SXM2-16GB, driver 580.173.02, fairchem-core 2.25.0 / torch 2.13.0+cu126)

Check cpu cuda:0
mliport doctor --engine uma --task omat --structure li10gep2s12_primitive.vasp passed, E=-216.30181 eV, stress checked passed, E=-216.30181 eV
validate_runtime.py --task omat --full (Cu fixture) passed: SP, shared calculator, FD, 10 NVE, 5-image NEB passed: same
recorded identity task: omat, inference_mode: default, inference_preset: batch, nitro_enabled: false, tf32_enabled: false, compile_enabled: false same, actual_device_type: cuda

LGPS CPU–GPU: |ΔE|/atom = 3.60e-09 eV, max |ΔF| = 5.70e-06 eV/Å
(E = -216.30180917 eV on CPU, -216.30180899 eV on GPU).

NEB inside the probe is the usual bounded band with relaxed endpoints and
saddle_validation: not_performed; no accuracy claim is made.

Status

uma_omat: untested → verified on CPU and GPU. The previous "blocked"
note in the UMA comment stands as history for the corrupt artifact; this comment
replaces it for the status table.

hydrogen1222 added a commit that referenced this pull request Nov 14, 2026
Tracking: #61 (model-support spec §11 task 1, re-check)

The earlier `uma-s-1p2.pt` copy had the right size but failed the zip CRC; the
owner re-downloaded it and the new file is clean: 315 members pass
`testzip()`, and its SHA-256 `ba5c0d912efa…` matches the pinned `uma_omat`
profile in `validation/science/model_manifest.json`.

Measured on the same host (Tesla V100-SXM2-16GB, driver 580.173.02,
fairchem-core 2.25.0 / torch 2.13.0+cu126):
- `mliport doctor --task omat --structure li10gep2s12_primitive.vasp` passes on
  cpu and cuda:0 (E = -216.30181 eV, stress checked);
- `validate_runtime.py --task omat --full` passes on both devices (single point,
  shared calculator, central differences, 10 NVE steps, five-image NEB);
- recorded identity `inference_preset: batch`, `nitro_enabled: false`,
  `tf32_enabled: false`, `compile_enabled: false`;
- LGPS CPU-GPU: |dE|/atom 3.60e-09 eV, max |dF| 5.70e-06 eV/A.

The table now shows 18 profiles `verified` on CPU and GPU, the Chinese table is
translated column by column (dtype and evidence cells were still English), and
the Evidence cell points at the re-check record in #75.

DATA IMPACT: no scientific record invalidated. The previous "blocked" note
stays in #75 as history for the corrupt artifact.
hydrogen1222 added a commit that referenced this pull request Nov 14, 2026
Tracking: #61 (model-support spec §6.5 + §11 tasks 1 and 5)

Rebased onto `main` after #94, #96 and #97 merged, so this carries only the
status-table change.

What changes:
- The status table leaves its initial state and carries what was measured on
  2026-11-14 on the owner's V100 in the installer-built environments: 18
  profiles `verified` on CPU and GPU, each with finite energy/force/stress on
  the 50-atom LGPS cell, a matching recorded identity and the bounded probe
  (shared calculator, central differences, 10 NVE steps, five-image NEB).
- UMA rows from §11 task 1: `uma_omat`, `uma_s_1p2p1_omat` and
  `uma_s_1p2p1_ft_r2scan_omat` verified on both devices and recorded with
  `inference_preset: batch`, `nitro_enabled: false`, `tf32_enabled: false`;
  `--inference-mode turbo` records `multigpu` with `nitro_enabled: false`. The
  r2SCAN fine-tune's NEB leg is refused by the gradient preflight (energy
  resolution), which the table states instead of hiding. `uma_omat` needed the
  re-uploaded checkpoint: the first copy had the right size but failed the zip
  CRC; the new file matches the pinned manifest hash `ba5c0d912efa...`.
- The CPU-GPU claim quotes the measured maxima (6.1e-7 eV/atom, 7.5e-6
  eV/Angstrom, `mace_omat_float32`) instead of a rounded "1e-7/1e-5 or better".
- Every Evidence cell links the specific #75 comment (runs, deltas, UMA record,
  re-check, NEP89 record) instead of the pull request.
- The Chinese table is translated column by column; it previously carried the
  English dtype and evidence cells.

DATA IMPACT: no scientific record invalidated. Statuses describe calling the
models correctly; no accuracy claim is made and no stored campaign record
changed.
@hydrogen1222

Copy link
Copy Markdown
Owner Author

DPA-SSE reference values · original runtime (deepmd-kit 3.0.0b3 + torch 2.3.1)

Evidence for #61 DPA-SSE spec §2.1. These are the numbers a conversion or a
legacy runtime must reproduce.

Artifacts

Item Value
Original checkpoint SSE-VASP-dpa2.pt, sha256 04c34fe24c85af70872569ac39868991a78c7237572111c7e2e0e1c2a7056931, 120922466 bytes
#91-converted SSE-VASP-dpa2-inference.pt, sha256 8194e51f5050f763e7b5ce450ed12bc9fa5fb57882314d7ee756a8c3ddbb50f4, 41101267 bytes
Model page https://www.aissquare.com/models/detail?pageType=models&name=SSE-VASP-dpa2 — file SSE-VASP-dpa2.pt, 115.32 MB, LGPL-3.0
Dataset page https://www.aissquare.com/datasets/detail?pageType=datasets&name=Solid_State_Electrolyte&id=217 — 20240412-SSE.zip, 266.35 MB, CC BY 4.0

Runtime that loads the original file unmodified

deepmd-kit==3.0.0b3, torch==2.3.1+cpu, Python 3.12, CPU only.

  • 226 of 226 tensors present in both the checkpoint and the constructed model, matched by
    name and shape
    : 0 missing learned tensors, 0 unexpected tensors, 0 shape mismatches
    (the only extra model-state key is the non-tensor _extra_state).
  • DeepPot(...).get_rcut() = 9.0, get_type_map() = the 118-element map stored in the file.

Structures

Key File sha256 Composition
lgps_primitive examples/structures/li10gep2s12_primitive.vasp 0e0702be3a31778839f730430c6d7989118df316c4976a1a98f945c32fd3561a Li20Ge2P4S24, 50 atoms
li2s_conventional generated with ASE (antifluorite, a=5.72 Å) 4404e50348ca1d776a0da6c15b4368d6a209318192cd4cf050db6bfa4c89ce4b Li8S4, 12 atoms
licl_rocksalt generated with ASE (rocksalt, a=5.14 Å) 2a4bbb76c18d041504e72aff5f40c93077a0647351cd4662d714886baea2eefc Li4Cl4, 8 atoms

All three contain only elements from the model's training list; li2s_conventional is also
the ≤32-atom cell the §2.3 probe should use.

Reference values (float64, CPU, virial in eV, 3x3 row-major)

{
 "runtime": {
  "deepmd-kit": "3.0.0b3",
  "torch": "2.3.1+cpu",
  "python": "3.12",
  "precision": "default (=float64 in deepmd-pt)"
 },
 "lgps_primitive": {
  "formula": "Ge2Li20P4S24",
  "atoms": 50,
  "energy_eV": -226.088601521,
  "forces_eV_A": [
   [
    0.007015,
    -0.014847,
    -0.031144
   ],
   [
    0.022789,
    0.010903,
    0.028326
   ],
   [
    0.005196,
    0.007061,
    -0.030123
   ],
   [
    0.000593,
    -0.006614,
    0.019599
   ],
   [
    0.023264,
    0.025453,
    0.03796
   ],
   [
    -0.025446,
    -0.000844,
    0.023252
   ],
   [
    -0.000423,
    -0.03947,
    -0.004332
   ],
   [
    -0.012887,
    -0.01258,
    0.006172
   ],
   [
    0.001488,
    -0.005003,
    -0.01709
   ],
   [
    -0.011884,
    0.016481,
    -0.000751
   ],
   [
    0.021289,
    0.016951,
    0.005585
   ],
   [
    -0.043332,
    -0.015528,
    -0.061074
   ],
   [
    0.005363,
    0.004098,
    0.029064
   ],
   [
    0.021945,
    0.000978,
    -0.016849
   ],
   [
    -0.001881,
    -0.02592,
    -0.014794
   ],
   [
    0.005911,
    -0.01828,
    0.001968
   ],
   [
    0.00232,
    -0.033098,
    -0.038294
   ],
   [
    -0.034692,
    0.019484,
    0.015896
   ],
   [
    -0.003314,
    -0.029785,
    0.045681
   ],
   [
    -0.007582,
    0.001414,
    0.00927
   ],
   [
    -0.022498,
    0.080276,
    -0.081263
   ],
   [
    -0.001579,
    0.049822,
    -0.051992
   ],
   [
    0.091779,
    0.009556,
    -0.11985
   ],
   [
    0.040508,
    0.027901,
    -0.014845
   ],
   [
    0.037476,
    0.005179,
    0.035534
   ],
   [
    0.119105,
    -0.047866,
    -0.150389
   ],
   [
    0.004106,
    -0.216207,
    0.120095
   ],
   [
    -0.010645,
    0.070928,
    -0.046309
   ],
   [
    0.034549,
    0.001998,
    0.035677
   ],
   [
    0.018253,
    0.248033,
    0.118504
   ],
   [
    -0.070139,
    -0.02041,
    -0.054312
   ],
   [
    -0.120679,
    0.054476,
    -0.119427
   ],
   [
    -0.180178,
    0.000404,
    0.075166
   ],
   [
    0.215386,
    -0.008312,
    -0.050698
   ],
   [
    0.015891,
    -0.014186,
    0.079453
   ],
   [
    0.05336,
    -0.03095,
    -0.029273
   ],
   [
    -0.032547,
    0.139951,
    -0.043748
   ],
   [
    -0.018166,
    0.08947,
    -0.02624
   ],
   [
    0.131541,
    0.119264,
    0.189869
   ],
   [
    -0.00355,
    0.016581,
    -0.02696
   ],
   [
    0.019237,
    -0.214659,
    -0.091803
   ],
   [
    0.062376,
    -0.138945,
    0.09269
   ],
   [
    -0.158878,
    -0.007211,
    -0.026202
   ],
   [
    -0.072684,
    -0.016182,
    0.108558
   ],
   [
    -0.299584,
    -0.055233,
    -0.105917
   ],
   [
    0.121531,
    0.070928,
    -0.10962
   ],
   [
    -0.083528,
    -0.00748,
    0.120572
   ],
   [
    0.148012,
    0.021305,
    0.093327
   ],
   [
    0.010044,
    -0.112839,
    0.093083
   ],
   [
    -0.024231,
    -0.016445,
    -0.022003
   ]
  ],
  "virial_eV": [
   -8.377595,
   -0.329522,
   -0.260692,
   -0.329522,
   -9.159341,
   0.070161,
   -0.260692,
   0.070161,
   -7.469619
  ]
 },
 "li2s_conventional": {
  "formula": "Li8S4",
  "atoms": 12,
  "energy_eV": -49.468821625,
  "forces_eV_A": [
   [
    -0.0,
    -0.0,
    0.0
   ],
   [
    -0.0,
    0.0,
    0.0
   ],
   [
    -0.0,
    0.0,
    -0.0
   ],
   [
    0.0,
    -0.0,
    0.0
   ],
   [
    -0.0,
    -0.0,
    -0.0
   ],
   [
    -0.0,
    0.0,
    0.0
   ],
   [
    0.0,
    -0.0,
    0.0
   ],
   [
    0.0,
    0.0,
    -0.0
   ],
   [
    0.0,
    0.0,
    0.0
   ],
   [
    0.0,
    -0.0,
    -0.0
   ],
   [
    -0.0,
    0.0,
    -0.0
   ],
   [
    0.0,
    -0.0,
    0.0
   ]
  ],
  "virial_eV": [
   -1.997121,
   0.0,
   -0.0,
   0.0,
   -1.997121,
   0.0,
   -0.0,
   0.0,
   -1.997121
  ]
 },
 "licl_rocksalt": {
  "formula": "Cl4Li4",
  "atoms": 8,
  "energy_eV": -30.530082885,
  "forces_eV_A": [
   [
    -0.0,
    -0.0,
    -0.0
   ],
   [
    -0.0,
    0.0,
    0.0
   ],
   [
    -0.0,
    0.0,
    0.0
   ],
   [
    0.0,
    0.0,
    -0.0
   ],
   [
    0.0,
    -0.0,
    0.0
   ],
   [
    0.0,
    -0.0,
    -0.0
   ],
   [
    -0.0,
    0.0,
    -0.0
   ],
   [
    -0.0,
    -0.0,
    0.0
   ]
  ],
  "virial_eV": [
   -1.240451,
   0.0,
   0.0,
   0.0,
   -1.240451,
   -0.0,
   0.0,
   -0.0,
   -1.240451
  ]
 }
}

sha256 of the JSON payload above (as printed, canonical json.dumps(indent=1)):
8c4b9cc203f84329992bfd1a897cc77936ba1ca2480ec74e972328d92b7ffeb2

Forces are printed to 1e-6 eV/A, energies to 1e-9 eV, virials to 1e-6 eV: finer than the
§2.2 tolerances (|dE| <= 1e-6 eV/atom, max |dF| <= 1e-5 eV/A, max |dvirial| <= 1e-5 eV).
li2s_conventional and licl_rocksalt are perfect crystals, so their forces are ~1e-13 and
~1e-8 eV/A; they test the energy and virial path, while lgps_primitive carries the real
non-equilibrium forces (max |F| = 0.2996 eV/A).

@hydrogen1222

Copy link
Copy Markdown
Owner Author

DPA-SSE callable check (legacy runtime) · 2026-11-15

Evidence for #61 DPA-SSE spec §2.3 and §5, after the route decision in §5
(legacy runtime). References are in the
reference comment.

Environment

Item Value
Environment .venv-dpa-legacy, created by ./scripts/install_mliport.sh --engines dpa-legacy --device auto (7 steps, closing doctor green)
Runtime deepmd-kit==3.0.0b3, torch==2.3.1+cu121, Python 3.12.13
GPU Tesla V100-SXM2-16GB, CC 7.0, driver 580.173.02 (torch arch list sm_50/60/70/75/80/86/90)
Model file SSE-VASP-dpa2-inference.pt (the #91 trust-gated conversion), sha256 8194e51f5050f763e7b5ce450ed12bc9fa5fb57882314d7ee756a8c3ddbb50f4 — the original training checkpoint is never loaded
Probe cell li2s_rattled_sigma005_seed20261115.vasp, Li8S4, 12 atoms, sha256 e0e179f9d4e1267cb370e58fd1ed0fa4b986b8fcbbbfeb24090e4e0c96ff89b6 (Li2S conventional cell + Gaussian displacements σ = 0.05 Å, numpy.random.default_rng(20261115), shape (12, 3))

mliport against the references (§5.1 and §2.3)

mliport sp --dpa-runtime legacy --device cpu on every reference structure,
compared with the deepmd-kit 3.0.0b3 numbers:

Structure |ΔE|/atom (eV) max |ΔF| (eV/Å) max |Δvirial| (eV)
LGPS primitive (50 atoms) 0.0 0.0 1.1e-14
Li₂S conventional (12 atoms) 0.0 4.2e-16 1.2e-15
LiCl rocksalt (8 atoms) 0.0 8.0e-16 2.7e-16
Li₂S rattled (12 atoms) 0.0 6.7e-16 6.5e-16

Tolerances were |ΔE| ≤ 1e-6 eV/atom, max |ΔF| ≤ 1e-5 eV/Å, max |Δvirial| ≤ 1e-5 eV.

Identity read-out

mliport sp and doctor record:

  • dpa_runtime: legacy
  • deepmd_version: 3.0.0b3
  • model_precision: ['float64'] (the checkpoint's precision: default)
  • actual_device_type: cpu / cuda; on GPU actual_device_uuid_source: CUDA_VISIBLE_DEVICES — torch 2.3.1 does not expose
    device_properties().uuid, so the UUID is taken from the single-entry
    visibility variable the isolation contract requires, and its provenance is
    recorded instead of being presented as a runtime check.

Per-device checks

Check cpu cuda:0
mliport doctor --engine dpa-legacy --task bulk --structure li10gep2s12_primitive.vasp passed, E = -226.0886 eV passed, same
validate_runtime.py --dpa-runtime legacy --structure li2s_rattled… --full passed: SP, shared calculator, central differences, 10 NVE steps, 5-image NEB passed: same
CPU vs GPU on the rattled cell |ΔE|/atom 0.0, max |ΔF| 1.8e-15 eV/Å, max |Δstress| 2.3e-17 eV/ų —

Fail-closed behaviour verified by hand

  • DPA_RUNTIME=legacy in .venv-dpa → refused with the install command.
  • DPA_RUNTIME=current in .venv-dpa-legacy → refused, naming DPA_RUNTIME=legacy.
  • The original SSE-VASP-dpa2.pt in the legacy runtime → refused before any
    unpickling (does not load with torch.load(weights_only=True) … convert with scripts/convert_dpa_checkpoint.py).
  • A DPA4/DPA-3 descriptor type in the legacy runtime → refused, naming the
    current runtime.

Limits

  • The rattled cell is an element-coverage probe, not a chemistry or accuracy
    claim: DPA-SSE was trained on sulfide electrolytes (15 elements) and anything
    outside that is extrapolation.
  • The probe's five-image NEB is a bounded band with relaxed endpoints;
    saddle_validation: not_performed.
  • No accuracy benchmark, no long MD, no comparison against DFT.

hydrogen1222 added a commit that referenced this pull request Nov 15, 2026
Tracking: #61 (DPA-SSE spec §5), review of #100

1. The strict tensor check can no longer be skipped in the legacy runtime.
   `_validate_pt_checkpoint` returns quietly when the loader exposes no
   inspectable torch model; in `.venv-dpa-legacy` that verification is the only
   guard against a partial or reshaped load, so `DPA_RUNTIME=legacy` now raises
   unless `checkpoint_tensors_verified` is true. The flag is recorded in
   `info()`, in the model record and in `IDENTITY_KEYS`, and the real CPU and GPU
   runs report `checkpoint_tensors_verified: true` (see #75).
2. `dpa-legacy` no longer claims Blackwell: torch 2.3.1+cu121 carries kernels up
   to sm_90 only, so `blackwell.upstream_supported` is false (like `maxwell`).
3. `dpa_runtime_problems` requires exactly `3.0.0b3` (a local version suffix is
   allowed); b0 is too old for `use_tebd_bias` and b4+ already changed the
   repformer layout, so accepting any `3.0.0b*` was wrong.
4. The legacy runtime accepts only `.pt`: any other suffix is refused before
   loading, and an unreadable model definition is now an error instead of a
   silent return.

Tests: `test_dpa_runtime.py` grows to 20 cases (non-`.pt` refusal, unverified
load refusal, exact-pin matrix, Blackwell support flag); the real install/probe
still pass.

DATA IMPACT: no scientific record invalidated; the legacy runtime simply refuses
more unsafe inputs.
@hydrogen1222

Copy link
Copy Markdown
Owner Author

DPA-SSE · checkpoint_tensors_verified from the real runs (review follow-up)

Review of #100 asked for the recorded strict-tensor-verification value from the real CPU and
GPU runs. It is now part of the identity record (info() metadata, the model record and
IDENTITY_KEYS), so mliport sp and the probe report it.

Command (both devices, legacy runtime):

CUDA_VISIBLE_DEVICES={'' | <GPU UUID>} .venv-dpa-legacy/bin/mliport sp \
  li2s_rattled_sigma005_seed20261115.vasp \
  --model SSE-VASP-dpa2-inference.pt --model-type dpa --task bulk \
  --device {cpu | cuda:0} --dpa-runtime legacy
Field cpu cuda:0
dpa_runtime legacy legacy
deepmd_version 3.0.0b3 3.0.0b3
checkpoint_tensors_verified true true
checkpoint_migrations [] (nothing widened, renamed or reshaped) []
model_precision ['float64'] ['float64']
actual_device_type cpu cuda
actual_device_uuid_source — CUDA_VISIBLE_DEVICES
energy on the rattled Li₂S cell -49.37500995424997 eV -49.37500995424997 eV

The --full probe on the same cell still passes on both devices after the change (SP, shared
calculator, central differences, 10 NVE steps, five-image NEB; checkpoint_tensors_verified
is true in the probe's record as well).

Behaviour change requested by the review: in DPA_RUNTIME=legacy a load that cannot run the
strict tensor comparison now fails closed instead of continuing with
checkpoint_tensors_verified: false.

@hydrogen1222

Copy link
Copy Markdown
Owner Author

Rev 10 · 执行结果(第一批 + 第二批,部分第三批) — 2026-11-17 · main @ 449ec9a4

本节只记录本次实际执行的结果;Rev 10 计划中的复选框未在正文勾选,勾选与否以本评论为准(避免把未执行项伪装成通过)。所有证据在本地 /tmp/r10/(会话级),关键数字与命令写在这里;CI 结果均为对应提交的运行。

0. 基线与环境(E01)— PASS

项 实测值
main 449ec9a4(#105/#106/#107 已合并;合并后 main CI 13/13 success)
Python 3.12.13(core dev .venv-dev);仓库声称支持 3.10–3.13,本轮只在本地 3.12 全量跑,其余版本以 CI 矩阵为准
核心依赖 ase 3.29.0, numpy 2.5.4, scipy 1.18.1, torch 2.14.1+cpu, textual 8.2.8, ruff 0.5.1, mkdocs 1.6.1
后端环境 .venv(fairchem, torch 2.13+cu126) / .venv-mace(mace 0.3.16, torch 2.8+cu126, torch-dftd 0.5.3) / .venv-grace(tf 2.20.0, GPU 可见) / .venv-dpa(deepmd 3.2.1.dev63, torch 2.11+cu126) / .venv-dpa-legacy(deepmd 3.0.0b3, torch 2.3.1+cu121) / .venv-nep(calorine 4.1)
GPU Tesla V100-SXM2-16GB,UUID GPU-f24116e9-bea2-e0a6-70a9-d0b50811856b,driver 580.173.02,sm_70;mliport setup --json 识别一致
权重 未下载任何新权重。使用本机既有:DPA(Nano/Mini/Neo/Plus/Air/4.1c-air/DPA-3.1-3M/SSE-VASP-dpa2)、UMA(s-1p1, s-1p2, s-1p2p1, -ft-d4rev, -ft-r2scan)、MACE(mace-mh-1 多头)、GRACE(2L-SMAX-OMAT-medium SavedModel)、NEP89(nep89_20250409.txt)。hash 见 §6

1. E 批次

ID 状态 证据
E02 全工程检查 PASS pytest tests -q -ra → 2310 passed, 48 skipped(2356 collected);ruff@0.5.1 check/format --check 干净;generate_docs.py --check up to date;check_repository_hygiene.py passed;uv build --wheel --sdist + check_package_contents 2/2 PASS;mkdocs build --strict 通过
跳过项如实统计 PASS 48 skips 全部有原因:pymatgen×2、kinisi×13、scipp×12、fairchem/coralline/matscipy×3、MLIPORT_REAL_INSTALL=1 门控×13、MLIPORT_PACKAGE_ARTIFACTS×1、GRACE GPU 可选件×1、fairchem checkpoint×1(详见日志)。skips 不计入 PASS
E03 source→test 映射 PARTIAL 114 个模块比对:7 个模块在测试中没有名字引用;逐个核查后 queue_daemon(test_scheduler_daemon_processes_queue 真实起进程)、writers/oszicar(optimization 写 OSZICAR)、nep_model_format(nep/gpumd 间接)均有间接覆盖;confsearch/revisions、neb/revisions 由 checkpoint 测试间接覆盖;mliport/utils.py 无任何 importer、无任何测试(见 F04)
E04 打包/安装 PASS wheel 609 KB / sdist 545 KB;干净 venv(无源码树,cd /)安装后 mliport --version、--help、setup --json(识别 V100/Volta/sm_70/16384 MiB)、doctor(5 个可选特性 WARN 且给出 mliport[analysis-all] 指引,未导入后端)全部正常;CI 风格 --no-deps capability smoke 通过

2. U 批次(#104–#107,本轮重点)

ID 状态 证据
U01 四入口一致 PASS 同一结构(Si8 弛豫胞)+ 同一 DPA-nano/GPU:CLI 参数 E=-42.852352333;INCAR 任务夹 + mliport run -42.852353287;TUI 向导同一函数路径(create_task_folder+run_task_folder)-42.852346135(DEVICE 默认 cpu);Python API -42.852357102。相对偏差 ≤2e-7,差异在 float32 GPU/CPU 噪声内;公式/原子序/身份一致
U02 models/ 查找 PASS 真实用例:bare 文件名→<MLIPORT_ROOT>/models;一层子目录;GRACE SavedModel 目录按名解析成功;两个同名候选明确拒绝并列出两者;MLIPORT_ROOT 环境变量优先于 venv 派生;含分隔符/绝对路径不被重定向;找不到时列出全部已查位置 + models/ 实际内容(真实触发过一次,消息完整)。定向测试 19/19 通过
U03 非标准路径 PASS(含 2 项已修缺陷) 空格+中文路径 odd dir 中文/task one 正常运行;OUTPUT_DIR = . 真实写回任务夹(不再生成 results/);符号链接结构文件正常;从其他 CWD 调用见 F02;只读输出目录见 F03(两者已在 #109 修复)。未测:只读输入、磁盘配额、Windows 原生(BLOCKED)
U04 schema→模板→文档一致 PASS test_incar_templates.py(模板/schema/文档三方键集与顺序)+ test_template_roundtrip.py(5 类型×6 后端模板可被 run 接受)+ generate_docs --check;严格性实测:未知键 NAPT_TDAMP→Did you mean 'npt_tdamp'?;WRITE_OUTCAR = perhaps→布尔转换错误;TEMPERATURE = nan→"expects a finite float";TIMESTEP = 0→拒绝;OPT 里写 MD_ENSEMBLE→"not valid for calc_type='opt'";CS_MODE = MODE_B→"must be one of 'adsorbate'"(信息可操作)
U05 TUI 向导 PASS 创建任务夹:INCAR.mliport(MODEL_PATH 保留 bare 名,让 models/ 生效)+ POSCAR 逐字节副本;remaining_placeholders = ['HEAD','TASK'];未填 TASK 时 run_task_folder 返回 1 且报 Invalid TASK 'required'(这正是 #107 复审修复的行为:不允许一键启动未选定的 task/head);填好 TASK 后正常跑完
U06 结果/provenance PASS SP 布局与 docs/outputs.md 一致(artifacts.json 是 MD 专有);resolved_config.json 记录解析后的绝对模型路径(symlink 解析到真实目标);MD 有 artifacts.json(units/observables/hashes)+ raw/ + vasp/

3. P 批次(数值核心 / MD / NPT / #102)

ID 状态 证据
P01 E/F/stress 有限差分 PASS 先用 ASE EMT(Cu)标定 FD 装置本身:力 max err 1.3e-4 eV/Å(h=0.01)、应力 2.0e-5 eV/ų(h=0.005)。真实模型(Si8 + 0.06 Å 扰动 + 剪切胞,fmax≈2.1 eV/Å):DPA-nano 力 4.4e-4 / 应力 8.2e-6;MACE(omat_pbe) 1.5e-4 / 7.3e-6;GRACE-2L 2.6e-4 / 1.4e-5;UMA-s1p1(omat) 2.0e-4 / 5.5e-6。误差随 h 增大(截断)而非噪声,符号/Voigt 顺序/单位均正确。过程中发现并纠正了审计脚本自身两处错误(未按 fractional 坐标缩放应变、DOM 无真空),已用 EMT 独立复核
P02 OPT PASS 立方 Si8 与三斜胞均收敛(fmax 降到 1e-6 eV/Å 量级);cell_opt 把人为剪切的胞弛豫回立方(物理正确)
P03 NVE PASS(PARTIAL 统计) 初始温度精确 300.0 K(速度初始化);200 步内总能量 −42.541245→−42.541050 eV(漂移 ≈2e-5 eV/atom / 0.2 ps);短单次运行的瞬时 T 波动大(8 原子 24 DOF,σ_T/T≈29%),未做步长收敛速率研究(时间预算)
P04 NVT 三种恒温器 PARTIAL Langevin/Bussi/NHC 均正常完成并写出 md.csv/traj/OUTCAR;但 150 步(0.15 ps)从 150 K 起步不足以判定统计温度(297/244/362 K vs 目标 500 K 属瞬态),未做自相关/有效样本统计
P05 NPT(Isotropic MTK) PASS 弛豫立方胞:P=0 GPa → 平均总压 +0.010 ± 0.389 GPa,体积稳定(+0.15%);P=2 GPa → +2.078 ± 1.367 GPa,体积 −3.21%;GPa↔eV/ų 换算由 stress 列独立重算,误差 1e-8 GPa;动能压强贡献 ≈+0.17 GPa 且确实计入
P06 非立方胞 NPT 形状 PASS a≠b≠c、角度 86.67/91.72/87.24° 的 Si8 胞,12 帧内长度比与夹角漂移 = 0.000(机器精度),体积随压力变化(P=0 −1.68%,P=2 GPa −5.75%)→ 各向同性 MTK 只缩放体积,没有偷偷强制 cubic;残余偏应力属方法限制(已知)
P07 NPT 拒绝链与输出契约 PASS(拒绝链) / F01 NPT+BUSSI、NPT+T=0、NPT+TCHAIN=0、NPT+WRITE_TRAJECTORY=FALSE、NVE+PRESSURE_GPA、NPT+NEP 全部 rc=1;NPT 产物:无 XDATCAR、artifacts.json 中 variable_cell=True/transport_eligible=False;analyze validate 报 "Cell changes across frames… refuses MSD/transport",analyze msd/transport 直接拒绝并给出原因;固定胞 NVT 可正常 msd。缺陷 F01:这些拒绝发生在 results/ 已创建、模型已加载之后(见下)
P08 Issue #102 FAIL → 已修复(PR #108) 复现:MDRunner(calc, "NVT", 300, 1, 500, 10, "NHC", 0.5, …) 历史上第 8 个位置参数是 friction,NPT 参数插入后绑定成 pressure_gpa,bussi_tau→npt_tdamp、nhc_tdamp→npt_pdamp、save_interval→npt_tloop、output_dir→npt_ploop(后者直接 TypeError: npt_ploop must be an integer)。红测 3/4 失败 → 把 7 个 NPT 参数移到末尾并设为 keyword-only → 4/4 绿,全量 2314 passed。PR #108(CI 全绿、MERGEABLE/CLEAN,待 owner 授权合并)
P09 帧数/写盘一致性 PARTIAL 抽查 NVE/NVT/NPT:帧数 = steps/save_interval + 1,md.csv/trajectory.traj/OUTCAR/CONTCAR 一致;未做 1e5 帧内存上界、写盘背压、磁盘满

4. W 批次

ID 状态 证据
W03 confsearch PASS 真实 DPA-nano/GPU、mliport run 派发 CONFSEARCH:预算行 "2 replicas x 1 cycles, 160 MD steps, <= 4 quenches";每循环 anneal→quench、quench 收敛统计(2 valid)、去重(1 new unique / 1 重复被拒)、Metropolis 接受(+0.0046 eV)/拒绝(+0.0023 eV);产物 search_results.json(status=completed、schema/scientific_revision/ranking/exports)、minima/all_minima.extxyz+registry.json、checkpoints/、diagnostics/、raw/。失败路径同样实测:无真空 slab → 放置 1000 次全 height_not_solvable 并列出计数;反应性吸附质 → reacted/quench_not_converged 分类 + "No replica survived the start-quench stage" 明确失败
W07 闭环(任务夹→执行→分析) PARTIAL SP:任务夹→run→JSON;MD:任务夹→run→analyze msd 成功产出分析目录;NEB 闭环本轮未跑
W01/W02/W04/W05/W06 NOT_RUN batch/NEB 真实任务、kinisi/GEMDAT/scipp 分析(核心环境无这些包;CI analysis-backends job 为绿但不算本地通过)、VACF/RDF 已知答案:本轮未执行

5. R 批次

ID 状态 证据
R01 故障注入 PARTIAL + 2 修复 SIGINT 中止长 MD:日志停在 "next safe point"、artifacts.json status="cancelled"(不是 completed)、退出码非 0;只读输出目录 → 修复前是 PermissionError traceback,修复后 Error: Cannot write in output directory …(#109);模型异常/NaN 由既有测试覆盖(test_fixes.py),磁盘满/半写 JSON 未测
R02 队列 NOT_RUN 双调度器/lease/GC 竞争由 test_queue.py(64 用例,含真实 daemon 进程)覆盖,本轮未做活体并发实验
R03 入口一致性 PARTIAL + F05 见 U01;发现 confsearch run 子命令没有 --device(sp/opt/md/neb/batch 都有),用户只能靠 INCAR/settings 指定设备 → 记 F05(P2,未修,属 CLI 表面不一致)
R04 包装/隔离 PASS 见 E04;后端隔离、无静默 CPU 回退(DPA/GRACE 的 require_isolated_visibility 实测会拒绝父进程已初始化 CUDA 的情况,报错给出 UUID 用法)
R05 性能回归 NOT_RUN 仅记录了冷/热启动 wall time(DPA-nano SP ≈7.7 s 含模型加载、MD ≈40 steps/s 含 8 原子),未做受控基准对比

6. B 批次(真实后端矩阵,Si8 金刚石,E/F/stress 全开)

案例 结果 E/atom (eV) 设备 精度
DPA4C-Nano (cpu / cuda) PASS / PASS −5.3516 / −5.3516 cpu / cuda:0 float32
DPA4C-Mini / DPA4-Neo / DPA4-Plus / DPA4-Air / dpa-4.1c-air PASS(cpu) −5.3938 / −5.4291 / −5.4225 / −5.4244 / −5.3472 cpu float32
DPA-3.1-3M(--head Omat24) PASS −5.3893 cpu float32
SSE-VASP-dpa2(legacy runtime) PASS +0.2896 cpu float64
MACE mace-mh-1(--head omat_pbe) PASS(cpu/cuda) −5.4232 cpu, cuda:0 float64
GRACE-2L-SMAX-OMAT-medium PASS(cpu/cuda) −5.4187 cpu, cuda:0 float32
UMA s-1p1 (omat, cpu) / s-1p2 (cuda) / s-1p2p1 (cuda) PASS −5.4116 / −5.4165 / −5.4205 cpu, cuda:0 float32
NEP89 (nep89_20250409.txt, Calorine) FAIL(已知限制) – – Calorine 无法解析 NEP89 v1:Failed to parse line 6
NEP89 GPUMD runtime BLOCKED – – 本机无 gpumd 可执行文件;解除条件:安装 GPUMD 二进制(或 owner 指定路径)
UMA uma-s-1p2p1-ft-d4rev 用 omat 拒绝(正确) – – Invalid task_name 'omat'. Valid options are ['omol']
载荷子模型缺 --head(MACE 多头、DPA-3.1 多头) 拒绝(正确) – – 列全部可用 head,要求显式指定
  • B01/B02/B06:上表 + metadata.actual_device_type/uuid、model_precision、模型解析后绝对路径全部记录;CPU/GPU 同模型 E 差 ≤1.1e-4 eV(float32 可解释),MACE float64 差 1e-6 eV。
  • B03:GPUMD 路径 BLOCKED(见上),Calorine 限制复现(与文档一致,不算产品缺陷)。
  • B04(缓存 A→B→A、OOM 回收)NOT_RUN;B05(DFT 参考精度)BLOCKED(本机无可信 CP2K/DFT 标签,且按要求不得用 Na₃PS₄ 未收敛力标签)。

7. 本轮确认的缺陷 / 修复 PR

编号 级别 现象 状态
#102 / E8 P1(兼容性,静默改配置) MDRunner 位置参数错绑(见 P08) PR #108 待授权合并
R10-F03 P1(诊断) 只读输出目录 → Python traceback PR #109 待授权合并
R10-F02 P2(诊断) run --incar 跨目录时只说“没有结构”,不说查了哪里 PR #109(仅改消息;结构默认仍是 CWD 相对——若要改成 INCAR 目录锚定需 owner 决策)
R10-F01 P2(顺序) MD 参数/设备拒绝发生在 results/、resolved_config.json 与模型加载之后 未修,建议下一步:把交叉字段校验提前到 CalculationEngine._validate()
R10-F04 P3(死代码) mliport/utils.py(191 行、8 个函数)无 importer、无测试,但仍打进 wheel 未修,建议 owner 决定删除或补测试
R10-F05 P3(CLI 不一致) confsearch run 无 --device,与其他子命令不一致 未修(文档化或补齐需 owner 决策)

8. BLOCKED / NOT_RUN 清单(含解除条件)

  • BLOCKED:NEP89 GPUMD(需 gpumd 二进制);真实 GPU 家族扩展(仅 1×V100);Windows 原生路径/安装;可信 DFT 参考(CP2K/DFT 标签不在本轮预算,需 owner 授权算力)。
  • NOT_RUN:NEB 真实闭环(W02/W07)、batch 真实任务(W01)、kinisi/GEMDAT/scipp 分析(W04/W05;CI job 绿但本地未跑)、队列活体并发(R02)、性能基准(R05)、缓存 A→B→A / OOM(B04)、1e5 帧内存与磁盘满(P09/R01)、NVT 统计收敛与 NVE 步长速率研究(P03/P04)、文档内容级双语言复核(D01 仅自动检查通过)。

9. 版本冻结建议(G02)

建议:可以冻结功能(bugfix-only),但暂不打 tag,直到 #108、#109 合并且 main CI 全绿。

理由:Rev 10 覆盖的核心数值契约(E/F/stress 有限差分、单位/Voigt/符号、NPT 压力换算与形状保持、models/ 与任务夹、四入口一致、confsearch 端到端)在实际模型与 V100 上 未发现 P0 级科学错误或数据破坏;已知 P1 缺陷只有 #102(已修待合并)。剩余缺口集中在 NEB/batch 真实闭环、分析后端本地执行、NEP GPUMD 与统计收敛研究——都是可列明的残余限制,不是未知风险。

冻结后建议的后续顺序:合并 #108 → 合并 #109 → 补 NEB/batch 真实闭环证据 → 视算力许可补 DFT 参考 → 再考虑打 tag。

@hydrogen1222

Copy link
Copy Markdown
Owner Author

Rev 10 · 修复 PR 合并结果 — 2026-11-17

接上方 执行结果评论。只记录合并后的事实;该评论中的 NOT_RUN / BLOCKED / 残余风险全部保留不变,未因合并而勾销。

合并记录(顺序 #108 → #109)

步骤 PR 合并前核对 merge commit main 更新
1 #108(#102 / E8) head 6e2fb28f、MERGEABLE/CLEAN、12/12 checks pass 30615cc2 ✅ 已确认
2 #109(R10-F02/F03) head 02431d53、MERGEABLE/CLEAN、12/12 checks pass、diff = 4 文件 +94/−4(无意外改动) fb2935c7 ✅ 已确认

#109 合并前的追加修正(按要求,仅测试)

Windows 上 os.geteuid 不存在,原测试会在收集期抛 AttributeError 而不是跳过。修正限定在测试文件内:

  • test_run_directory.py 的只读目录用例加 @pytest.mark.skipif(os.name != "posix", ...),root 判定改为 getattr(os, "geteuid", lambda: -1)();
  • 提交 02431d53(仅 tests/mliport/mliport/test_run_directory.py),生产代码未改;
  • 更新后重跑 test_run_directory.py + test_cli.py(66 passed)、ruff check/format 干净、CI 12/12 通过后才合并。

本轮之后仍未完成 / 仍有风险(与上方评论一致,逐条保留)

编号 状态 说明
R10-F01 未修(P2) MD 参数/设备拒绝仍发生在 results/、resolved_config.json 与模型加载之后;建议把交叉字段校验提前到 CalculationEngine._validate()
R10-F04 未修(P3) mliport/utils.py(191 行、8 函数)无 importer、无测试,仍随 wheel 发布
R10-F05 未修(P3) confsearch run 无 --device,与其他子命令不一致
NOT_RUN 保留 NEB 真实闭环、batch 真实任务、kinisi/GEMDAT/scipp 本地执行、队列活体并发、性能基准、缓存 A→B→A/OOM、1e5 帧内存与磁盘满、NVT 统计收敛与 NVE 步长速率、文档内容级双语言复核
BLOCKED 保留 NEP89 GPUMD(无 gpumd 二进制)、多 GPU 家族(仅 1×V100)、Windows 原生路径、可信 DFT 参考(未获算力授权)

冻结建议(不变)

功能维持冻结(bugfix-only)。#102 这个已知 P1 兼容性缺陷已修复并合并,main CI 全绿,因此“先合并修复再谈冻结”的前置条件已满足;但 tag/发布建议仍等 NOT_RUN 清单中的 NEB/batch 真实闭环与分析后端本地执行补齐(或由 owner 明确接受这些残余风险)后再决定。

@hydrogen1222

hydrogen1222 commented Nov 17, 2026 •

Copy link
Copy Markdown
Owner Author

Rev 10 · 第二阶段执行结果(多材料 / 全工作流 / 无人值守)— 2026-11-18 · main @ fb2935c7

本节只记录本轮实际执行的结果。计划复选框不在正文勾选(避免把未执行项伪装成通过)。基线 = 任务书给定 fb2935c730118bcaa86ede6154ac10c512f46677,执行前重新 fetch 确认 HEAD..origin/main = 0。仓库工作区未被修改;两个修复走独立 worktree + 独立 PR。完整报告与 154 条用例矩阵在本节末尾。

0. 一句话结论

软件契约层面(单位/符号/Voigt/PBC/task-head/续算/失败语义)在本轮覆盖面内可用,可作为下一个 beta;材料精度与统计结论没有验证(本机无 DFT 对照标签,输运 10 ps 采样不足)。发现 1 个 P1、2 个 P2 缺陷,其中 P1 与 R10-F01 已修复成 PR,NEB resume 文档/行为冲突记录待裁定。

1. 环境与工程基线

项 实测
main SHA / 工作区 fb2935c7 / 干净;所有测试在仓库外 mliport_validation_runs/rev10_remaining_20261117_232213/
pytest tests -q 2316 passed, 48 skipped(exit 0);修复分支 worktree 上 2317 / 2322 passed
48 skips 44 个真实安装门控 + calorine/pymatgen/kinisi;其中 4 个(kinisi/pymatgen)已在具备依赖的环境实跑通过(33 passed)
ruff / docs / hygiene check 通过、format 通过、generate_docs.py --check up to date、hygiene passed
GPU V100-SXM2-16GB UUID GPU-f24116e9-…,单任务串行(flock),累计 GPU 墙钟 ≈3 h(harness 记录 1.72 h + 裸子进程链)
后端环境 .venv(fairchem 2.25.0) / .venv-mace(0.3.16) / .venv-dpa(deepmd 3.2.1.dev63) / .venv-dpa-legacy(3.0.0b3) / .venv-grace(tf 2.20.0) / .venv-nep(calorine 4.1),全部实测可加载

2. 覆盖(真实模型 × 真实材料)

  • 16 个本地 checkpoint 全部实测(15 个成功加载,NEP89 因格式限制单列):DPA(Nano/Mini/Neo/Air/Plus/4.1C-Air/3.1-3M/SSE-legacy)、MACE(mh-1 四 head、matpes-r2scan-ft)、GRACE(SMAX-medium、OMAT-base fp64)、UMA(s-1p1/s-1p2/s-1p2p1/两个 ft)、NEP89。
  • 7 种晶胞:Si(立方 / 菱方 60° 非正交)、Cu(fcc/超胞/空位)、Li₂S(反萤石)、Na₃PS₄(四方,含 2×2×2 128 原子与 2×1×2)、Li₆PS₅Cl(13/26 原子)、LGPS(50/200 原子,三斜)、LLZO(96 原子)、受控剪切 Na₃PS₄。
  • 结构全部来自仓库 fixture 或 Materials Project OPTIMADE 几何 pin(mp-1153/mp-28782/mp-985592/mp-942733/mp-696138),并记录 sha256;Na₃PS₄ 与 LGPS 的 pin 与仓库既有 fixture/examples 逐坐标一致。不使用 MP 能量。

3. 关键科学结果

项 数值 判读
上游直连对照 DPA 能量逐位相同(−42.81268901860108 eV)、MACE ΔE=7.1e-15 eV、应力差 <1e-9 mliport 未引入自己的换算
Si 变胞 OPT DPA4C-Nano a=5.3751 Å;mace-mh-1(omat_pbe) a=5.4674 Å PBE≈5.47 → MACE 对,Nano 偏小(模型性质)
Si 菱方原胞 OPT a=3.8003 Å、60° 保持 与 5.3751/√2=3.8008 自洽(晶胞算法正确)
Cu 空位 NEB 0.739337 / 0.739353 eV(普通);0.750910 / 0.750900 eV(CI-NEB) 正反一致 ≤1.6e-5 eV;落在文档 0.735–0.762 区间
NVE 能量守恒 Si64 3 ps:|ΔE|=1.2e-5 / 2.5e-4 / 1.3e-3 eV(0.25/0.5/1.0 fs);Na₃PS₄128:1.1e-3 / 4.5e-3 eV 漂移 ~1e-8…5e-7 eV/atom/ps;ASE VelocityVerlet 对照同量级 → 噪声主导,非积分器缺陷
NVE 复现 同种子逐帧差 2.7e-5 eV;save-interval 1 vs 10 差 2.7e-5 eV;换种子差 6.6e-4 eV 复现合同成立
NVT 300 K:294.5±6.1 / 293.7±8.2 / 294.8±8.1(Langevin/Bussi/NHC);600 K:587±7 / 580±18;800 K:784±18 三恒温器均在块 SEM 内到点
NPT(各向同性 MTK) 0 GPa:0.011±0.012(Si)、0.011±0.028(Na₃PS₄)、0.014±0.037(LGPS);2 GPa:2.054±0.023 / 1.987±0.035 / 1.996±0.028 轴比散布 ≤6.7e-16、角度散布 0.0° → 严格等比例、不强行立方;逐帧偏应力幅度 0.003–0.006 eV/ų(非立方体系的已知限制,晶胞形状未被改变)
NPT 数据边界 artifacts.json: variable_cell=true、transport_eligible=false,不写 XDATCAR 正确
confsearch 续算 参考 vs 同参重跑:7 极小、最大能量差 4.4e-4 eV;参考 vs 中断+续算:7 极小、1.2e-3 eV 续算偏差 ≈ 后端自身重复性 → 忠实
输运(Na₃PS₄ 10 ps) 600 K D=3.18e-11(CI 3.1e-12…7.0e-11)、σ_NE=17.5 mS/cm;700 K 9.63e-11;800 K 8.23e-11 sampling_insufficient:700→800 非单调;Arrhenius Ea=0.120±0.121 eV、R²=0.43
Arrhenius 已知答案 合成 Ea=0.3 → 反演 0.29999995±2.6e-8 eV、D₀=9.99999e-5、R²=0.9999999999 算法正确
GEMDAT 三温度 No jumps found(显式晶格 Na 位点) 事件不足,非软件错误

4. 本轮发现并修复的缺陷

R10-F06(新,P1)→ PR #110(Draft,12/12 checks SUCCESS,MERGEABLE/CLEAN)

DPA 返回 np.bool_(False),classify_result 用 converged is False 判断 → 未收敛的 OPT 被记为 completed:

接口 main 上实测 修复后
mliport opt(步数耗尽) 打印 Converged: No,exit 0 exit 2(文档承诺)
mliport batch --calc-type opt status: completed、success: true、not_converged: 0 not_converged、success: false
队列作业状态(按退出码派生) completed not_converged

红测试 2 个(batch 参数化 + classify_result),修复后 2317 passed。这是本轮唯一会悄悄污染批量/队列统计的缺陷。

R10-F01(旧账,P2)→ PR #111(Draft,12/12 checks SUCCESS)

NPT 的 4 个纯配置冲突(非 Langevin thermostat / --no-write-trajectory / 温度≤0 / 阻尼≤0)原先在 results/、resolved_config.json、模型加载之后才拒绝。现将校验前移到 resolver:

  • 修复前:4 个用例全部 results_dir=Y resolved=Y model_loaded=Y;
  • 修复后:全部 N/N/N,退出码与错误文本不变;
  • 合法 NVE / NVT-NHC / NPT 的 md.csv 与 main 逐字节相同(md5)。

5. 已确认但未修复(需 owner 裁定)

编号 级别 内容
NEB-RESUME-01 P2 docs/workflows/neb.md 承诺"resume 改 --max-steps 会被拒绝"。实测:同一 checkpoint 分别用相同选项、--max-steps 123、去掉 --climb resume,三者全部 exit 0 跑完;resolved_config.json 里 max_steps 仍是记录值 200(source 标 CLI)→ checkpoint 层覆盖了用户显式值,指纹恒等。科学数据不会被污染(协议保持一致),但静默忽略用户意图,违背"fail closed"。修复点在 cmd_neb resume 分支与 resolver 层优先级,属公共行为重设计,本轮按规则停止改码并记录。建议二选一:显式拒绝并列出冲突项,或文档改为"resume 一律沿用 checklist 记录值"。
R10-F07 P3 mliport analyze <单点运行目录> thermo 报 [Errno 21] Is a directory,真实原因是"该目录没有轨迹",信息误导。
R10-F08 P3 分析结果 id 不覆盖 raw/mliport_results.json;结果文件被截断时若同 id 分析已完成则直接复用(--force 后会在轨迹截断处硬失败)。
R10-F05 P3 confsearch run 无 --device(复现),只能经 settings/INCAR;confsearch resume 只接受计算器选项。本轮用 --settings 的 DEVICE=cuda:0 完成 confsearch 验收,建议补文档。
R10-F04 P3 mliport/utils.py(191 行、8 函数)无任何 importer,仍随 wheel 发布;按任务书不删除。
— 观察 官方 harness 6 条 fail 全是 float32 旋转/平移不变性在 ~1e-6 eV/Å 的偏差(阈值 = 10×实测重复性下限);保留 fail 原样,不改判。

6. 可靠性 / 故障注入 / 性能

  • 队列(隔离 MLIPORT_JOBS_DIR):max_concurrent=1 实测并发度恒为 1、作业区间零重叠;第二个调度器被拒;暂停/恢复生效;失败作业不中断整批并如实记 failed;SIGTERM 前台调度器 0.5 s 退出、运行中的 worker 独立跑完并写回状态、重启后队列排空、结束无残留调度器。
  • 故障注入:只读目录 / 已存在结果目录 / 并发写同一目录 → 全部拒绝且无产物;ENOSPC(定向注入 JsonWriter.write)→ OSError(28)、只留 OUTCAR、无 results JSON、无 success 声明、无临时残留;截断轨迹 → analyze 硬失败不写分析。
  • 缓存与显存:DPA A→B→A 能量差 2e-6 eV(float32 噪声级,无跨模型污染证据);加载失败后正常加载结果不变;8 次创建/销毁计算器显存恒为 17.039 MB;GRACE neighbor-cache ON/OFF ΔE=7.2e-7 eV;MACE dtype 正确上报。
  • 性能(官方 T8,pass):DPA-3.1-3M SP 暖 165.7/167.9/177.6 ms(32/128/512 原子),MD 5.94 步/s(128 原子);UMA-s-1p2 SP 103.7/105.2/205.0 ms,MD 9.50 步/s。冷启动 14.0 s(DPA@32 原子)vs 暖 0.17 s。

7. NOT_RUN / BLOCKED / INCONCLUSIVE(不得当成通过)

项 状态 原因
MLIP vs DFT 精度对照 BLOCKED 本机无已收敛且元数据完整的 DFT 标签;旧 Na₃PS₄ 力标签不可当真值
NEP89 原生 GPUMD BLOCKED GPUMD/src/gpumd 缺 libnvJitLink.so.12(其余 CUDA 库可从 uv 缓存补齐);不重新编译 GPUMD
NEP89 ASE/Calorine UNSUPPORTED 新版 nep.txt 格式 Calorine 解析失败(已知限制)
官方 T6(MD 套件) NOT_RUN(启动后中止) pinned DPA-3.1-3M 实测 ~4.6 步/s(128 原子),NVE 步长扫描即超预算;已用自建 MD 矩阵覆盖并记录速率
官方 T5(NEB 套件) NOT_RUN 8 条内置工作流耗时不可控;用直接 CLI NEB/CI-NEB + 中断续算覆盖
MACE 官方 pinned profile BLOCKED manifest 钉的 mace-omat-0-medium 本机不存在;mace-mh-1 是不同 artifact,不得替代
Na₃PS₄ 电导率/Ea INCONCLUSIVE 10 ps×3 温度,D 不确定度约 2×,700→800 非单调
GEMDAT 定量 INCONCLUSIVE 三温度零跳跃事件
LLZO / LPSC 深度工作流 部分 LLZO 只做 SP + E/F/stress(5 个模型,96 原子);LPSC 做 SP + 固定胞 OPT + 1500 K 短 NVT
TUI NOT_RUN(交互) 由 test_tui.py 覆盖(全绿)
500 原子以上长时 MD/NEB NOT_RUN 仅 512 原子 SP/短 MD 性能冒烟

8. 停留状态

9. 完整报告与用例矩阵

FINAL_REPORT.md(中文)

Mliport Rev 10 第二阶段综合验收 · 最终报告


1. 总结:Mliport 现在可靠到什么程度

结论分三层,不要混着读:

已经用真实模型 + 真实硬件验证的部分。 四个后端(UMA / MACE / DPA / GRACE)在本机 V100 上都能被 mliport 正确调用,15 个本地 checkpoint 全部加载成功并给出有限的 E/F/stress(NEP89 是第 16 个,因格式不兼容单列,见 §3);6 种材料(Si、Cu、Li₂S、Na₃PS₄、Li₆PS₅Cl、LGPS)加 LLZO 的 7 个晶胞上,SP、变胞/固定胞 OPT、batch、NEB/CI-NEB、confsearch、NVE/NVT/NPT、MSD/kinisi/GEMDAT/VACF/RDF/Arrhenius 都能真实跑通并留下可追溯产物。Cu 空位 NEB 得到 0.739 eV(普通 NEB)与 0.751 eV(CI-NEB),正反向一致到 2×10⁻⁵ eV,落在仓库文档记录的 0.735–0.762 eV 区间内。

发现的真实软件缺陷。 2 个,都不在"跑不通"的层面,而在"悄悄把失败写成成功"的层面:

明确没有证明的部分。 本机没有任何 MLIP 与 DFT 的同体系对照标签(Na₃PS₄ 的旧 CP2K 力标签未充分收敛,按既定规则不可用),因此没有做 MLIP-vs-DFT 精度验证;所有绝对能量只用于同一模型内部的相对比较。输运只完成"软件闭环",10–12 ps 的轨迹不足以给出可发表的 Na₃PS₄ 电导率(见 §8)。

一句话:软件层面的科学正确性(单位、符号、Voigt 顺序、PBC、task/head、续算、失败语义)在本轮覆盖面内是可靠的;模型精度与统计收敛没有、也不应该由本轮验收来背书。


2. 基线、环境与工程检查

项 实测
main SHA fb2935c730118bcaa86ede6154ac10c512f46677(= 任务书给定基线,git rev-list --count 前后均为 0)
工作区 干净;本轮所有测试在仓库外目录执行;修复走独立 worktree + 独立分支
pytest tests -q -ra(.venv-dev) 2316 passed, 48 skipped(7 分 40 秒,exit 0)
同一套测试在修复分支 worktree 2317 passed, 48 skipped(+1 为新增参数化用例)
ruff check / ruff format --check All checks passed / 275 files already formatted
check_repository_hygiene.py / generate_docs.py --check passed / up to date
48 个 skip 的性质 44 个 MLIPORT_REAL_INSTALL 等真实安装门控、calorine 缺失、pymatgen/kinisi 在 .venv-dev 缺失;其中 4 个(kinisi/pymatgen)已在具备依赖的环境中实跑并通过(test_science_analysis_known_answers.py:33 passed)
后端环境 .venv(fairchem 2.25.0, torch 2.13+cu126) / .venv-mace(mace 0.3.16) / .venv-dpa(deepmd 3.2.1.dev63) / .venv-dpa-legacy(3.0.0b3) / .venv-grace(tf 2.20.0) / .venv-nep(calorine 4.1) 全部实测可用且均以 editable 指向本仓库
GPU / CPU 分离 所有 GPU 用例经同一把 flock 串行;全程只有一个真实 GPU 任务

3. 模型与材料覆盖(T1:身份 + 加载 + E/F/stress)

16 个本地 checkpoint 全部实测(不是按文件名判断;其中 15 个成功加载,NEP89 见下)。每个模型在 6–7 个材料晶胞上求 E/F/stress,全部有限值;无一例 stress 缺失或静默跳过。

模型 类型/head task 加载 材料数 Si E/atom (eV) Na₃PS₄ E/atom LGPS E/atom
DPA4C-Nano dpa bulk LOADED 7 −5.3516 −4.0821 −4.2799
DPA4C-Mini dpa bulk LOADED 6 −5.3938 −4.0833 −4.3043
DPA4-Neo dpa bulk LOADED 6 −5.4291 −4.1149 −4.3246
DPA4-Air dpa bulk LOADED 6 −5.4244 −4.1151 −4.3265
DPA4-Plus dpa bulk LOADED 6 −5.4225 −4.1151 −4.3272
DPA-4.1C-Air dpa bulk LOADED 6 −5.3472 −4.1134 −4.3237
DPA-3.1-3M dpa head=Omat24 bulk LOADED 6 −5.3893 −4.0999 −4.3220
SSE-VASP-dpa2 dpa legacy 运行时 bulk LOADED 6 +0.2896 −2.6760 −4.5218
mace-mh-1 mace head=omat_pbe, float64 bulk LOADED 7 −5.4232 −4.1102 −4.3241
GRACE-2L-SMAX-OMAT-medium grace bulk LOADED 7 −5.4187 −4.1127 −4.3221
GRACE-2L-OMAT-base fp64 grace bulk LOADED 6 −5.4233 −4.1128 −4.3231
UMA-s-1p2 uma omat LOADED 7 −5.4165 −4.1151 −4.3260
UMA-s-1p1 uma omat LOADED 6 −5.4116 −4.1156 −4.3273
UMA-s-1p2p1 uma omat LOADED 6 −5.4205 −4.1147 −4.3269
UMA-s-1p2p1-ft-r2scan uma omat LOADED 2 −7904.61 – –
NEP89 nep ase/Calorine bulk 不支持 0 – – –

必须分开读的三点:

  1. 绝对能量不可跨模型/跨参考面比较。 DPA 家族内部 −5.35…−5.43 eV/atom 的差距是模型规模差异;SSE-VASP-dpa2(+0.29 eV/atom)与 uma-…-ft-r2scan(−7904.61 eV/atom)是不同参考面/带任意能量零点的 checkpoint,它们的力与应力仍然合理(r2scan 微调在 Si₈ 上 fmax=5.6×10⁻⁶ eV/Å,stress 有限),但绝对能量不得与 PBE/OMat 系模型混用。这是本轮实测证据,不是猜测。
  2. mace-mh-1 的 4 个 head 是 4 个不同参考面:omat_pbe −5.4232、mp_pbe_refit_add −5.3989、matpes_r2scan −8.7584、MACE-matpes-r2scan-omat-ft −8.7912 eV/atom。必须显式选 head,mliport 会拒绝不存在的 head(MACE head 'omat_pbe' is not available in this model. Available heads: ['default'])。
  3. NEP89 新版模型格式:Calorine 解析失败(Failed to parse line 6),这是已知且文档化的限制,不是新 bug。原生 gpumd 路径本轮 BLOCKED:本机 GPUMD/src/gpumd 存在,但动态链接缺少 libnvJitLink.so.12(libcublas.so.12/libcusolver.so.11/libcufft.so.11 可从 uv 缓存补齐,libnvJitLink.so.12 本机没有),按授权不得重新编译 GPUMD,因此只记录,不重试。

4. 单点、几何优化与批处理

4.1 单点输出契约与 CPU/GPU 一致性

同一结构、同一模型,CLI 在 cuda 与 cpu 上的差异:

模型 Si₈ ΔE (eV) Na₃PS₄ ΔE (eV) 说明
DPA4C-Nano (float32) 1.4×10⁻⁵ 1.6×10⁻⁵ float32 后端噪声
GRACE-2L (float32) 1.9×10⁻⁶ 5.5×10⁻⁶ 同上
UMA-s-1p2 (float32) 1.1×10⁻⁶ 7.9×10⁻⁷ 同上
mace-mh-1 (float64) 7.1×10⁻¹⁵ 1.4×10⁻¹⁴ 位级一致
  • OUTCAR 的 Voigt 表头显式写作 xx yy zz yz xz xy(ASE 顺序),并同时给出 eV/ų 与 GPa;压力 -trace(σ)/3 与 Pressure: 行吻合到 10⁻⁶ GPa。用一个只有 xy 剪切的受控形变胞(Na₃PS₄,人为施加,不是平衡态)验证:JSON 与 OUTCAR 均只有 xy 分量非零,yz/xz ≈10⁻⁹ → Voigt 顺序与单位无歧义。
  • 与上游计算器直连对照(同一进程、同一模型文件、同一结构,mliport CLI 的 JSON vs 直接调用上游 ASE 计算器):DPA 能量逐位相同(−42.81268901860108 eV,应力分量差 <1×10⁻⁹);MACE(float64) ΔE = 7.1×10⁻¹⁵ eV、应力差 <1×10⁻¹⁷。→ mliport 没有在 E/F/stress 上引入自己的换算。
  • 非周期输入 fail closed:水分子 + --task bulk → Task 'bulk' requires full 3D periodicity, but the input PBC is [False, False, False]. …mliport will not change it.;同一分子用 --task molecule 正常(exit 0)。2D slab 的拒绝本轮未覆盖:我构造的 slab fixture 实际写成了 3D 周期胞(pbc=[T,T,T],真空在 c 方向),因此它被当作合法 bulk 接受;这一项记为 NOT_RUN,不能声称已验证。
  • 输出目录契约:已存在结果的目录、只读目录、并发写同一目录,全部被拒绝并给出明确原因(见 §9)。

4.2 几何优化(全部真实 GPU)

用例 体系 模型 结果
变胞 OPT Si₈ DPA4C-Nano a = 5.3751 Å,20 步,fmax 4.7×10⁻⁶
变胞 OPT Si₈ mace-mh-1 (omat_pbe) a = 5.4674 Å,15 步(PBE 参考 ≈5.47 Å)
变胞 OPT Cu₃₂ DPA4C-Nano a = 3.6114 Å(实验 3.615)
固定胞 OPT Na₃PS₄(2×1×2) DPA4C-Nano 晶胞逐位不变(Δcell = 0.0),fmax 0.0125
变胞 OPT Na₃PS₄ DPA4C-Nano a 6.924→7.029、c 14.154→14.328,fmax 0.0107
固定胞 OPT Li₆PS₅Cl DPA4C-Nano 晶胞不变,fmax 0.0176(阈值 0.02)
变胞 OPT Si 菱方原胞(60°,非正交) DPA4C-Nano a=3.8003 Å,60° 保持 → 与 5.3751/√2=3.8008 自洽
变胞 OPT 受控剪切 Na₃PS₄(88.28°) DPA4C-Nano 剪切回复到 90.007°,晶胞 (7.028, 7.059, 7.167),与上一行独立结果一致
步数不足 320-atom 级无关;见注 DPA4C-Nano Converged: No 被正确打印;修复前退出码却是 0 → PR #110

注:最初的"步数不足"用例用了理想 fcc Cu₃₂(DPA 在该几何上力已 ≈10⁻⁷),第 0 步即收敛,因此标记 INCONCLUSIVE,另用 0.08 Å 扰动的 Cu₃₂ 重测(cu32_rattled),得到 Converged: No、exit 0(main)/ exit 2(#110)。

4.3 批处理(14 个输入,首次真实验收)

输入集:8 个 VASP、2 个 CIF、2 个 extxyz(含扩展 XYZ)、两组同名结构(a_si/si.vasp 与 g_rep1/si.vasp、h_rep2/si.vasp)、扰动结构、一个损坏 CIF、一个空 POSCAR、一个非周期水分子、一个受压 Cu₃₂。

步骤 结果
Batch SP 14 任务 = 11 completed + 3 failed(损坏 CIF / 空结构 / 非周期分子),单任务失败不中断整批
--resume(无变化) reused = 11,未重算任何成功项;进度文件不追加
修改 1 个输入内容后 --resume reused = 10,只有被改的 a_si/si 重算(shasum 变化被识别)
换模型后 --resume 拒绝:Cannot resume: batch settings differ … model_path / model_sha256: f1309e91… -> 49a341ff…,不混淆不同模型的结果
Batch OPT(max-steps=4, fmax=1e-3) 9 个结构 converged: false;main 上这些记录 status=completed、success: true、not_converged: 0 → 即 PR #110 的缺陷
中断后 resume SIGINT 中断后 --resume 继续,已完成项复用,未留下"假的 completed";batch_summary.json 总数为 14 且与 14 个 batch_status.json 一致

5. 分子动力学:NVE / NVT / NPT / 高温

全部为 DPA4C-Nano 真实 GPU 运行,产物为 raw/md.csv + raw/trajectory.traj + artifacts.json。注意一个物理事实,避免误判为 bug:从弛豫到极小值的构型出发、按目标温度赋 Maxwell-Boltzmann 速度做 NVE,系统会按能量均分自发降温到 ≈T/2(初始全部能量在动能上,平衡后一半进入势能)。Si₆₄ 300 K → 149 K、Na₃PS₄ 600 K → 303 K 都是这个效应,不是恒温器或积分器错误。

| 体系 | 步长 | 物理时长 | |ΔE_total| | 漂移斜率 (eV/atom/ps) | T 初→均值 | 帧数 |
|---|---|---|---|---|---|
| Si₆₄ | 0.25 fs | 3 ps | 1.2×10⁻⁵ eV | −5.6×10⁻⁹ | 300→149.5 K | 1201 |
| Si₆₄ | 0.50 fs | 3 ps | 2.5×10⁻⁴ eV | −1.8×10⁻⁸ | 300→149.5 K | 601 |
| Si₆₄ | 1.00 fs | 3 ps | 1.3×10⁻³ eV | −7.6×10⁻⁹ | 300→149.6 K | 301 |
| Na₃PS₄₁₂₈ | 0.50 fs | 3 ps | 1.1×10⁻³ eV | +1.2×10⁻⁷ | 600→303 K | 601 |
| Na₃PS₄₁₂₈ | 1.00 fs | 3 ps | 4.5×10⁻³ eV | +5.2×10⁻⁷ | 600→304 K | 301 |

NVT(三种恒温器)

  • E_kin + E_pot = E_tot 残差为 0.0(所有 NVE 用例,逐帧核对),能量记账没有错位。
  • 同种子重复:两次 dt=1.0 fs 运行逐帧最大差 2.7×10⁻⁵ eV、温度 5×10⁻⁴ K → 复现合同成立(float32 噪声级)。
  • 保存频率不影响物理:--save-interval 1 与 10 在同一物理时刻的最大能量差 2.7×10⁻⁵ eV。
  • 换种子:最大差 6.6×10⁻⁴ eV、温度 52.6 K → 不同轨迹,符合预期。
  • 独立积分器对照(ASE VelocityVerlet + 同一 DPA 模型,从 mliport 第 0 帧的坐标和动量起步):0.25 fs 漂移 −9.8×10⁻⁹、ΔE=1.0×10⁻⁴ eV;1.0 fs 漂移 +3.7×10⁻⁸、ΔE=1.4×10⁻³ eV。与 mliport 同量级,说明 mliport 的 NVE 就是 ASE 的 Velocity Verlet 语义,漂移由 float32 模型噪声主导,而不是积分器缺陷;这也解释了为什么 Si₆₄ 上漂移对步长不单调(噪声底 ≈ 积分误差)。

NPT(各向同性 MTK)

体系 温度设定 恒温器 平衡+生产 T 均值 ± 块 SEM fmax
Si₆₄ 300 K Langevin (γ=0.01) 1+5 ps 294.5 ± 6.1 K 3.18
Si₆₄ 300 K Bussi/CSVR (τ=100 fs) 1+5 ps 293.7 ± 8.2 K 3.26
Si₆₄ 300 K NHC (4 链, 2 子步) 1+5 ps 294.8 ± 8.1 K 3.60
Na₃PS₄₁₂₈ 600 K Langevin 1+5 ps 587.0 ± 7.1 K 6.10
Na₃PS₄₁₂₈ 600 K NHC 1+5 ps 579.8 ± 17.8 K 5.47
Na₃PS₄₁₂₈ 800 K Bussi 1+5 ps 784.0 ± 18.5 K 6.90

6. NEB / CI-NEB

体系与构建方式完全按 docs/test-systems.md 配方 2:3×3×3 fcc Cu(a=3.63 Å,107 原子)+ 一个空位 + 最近邻跳入空位,跳距 2.5668 Å。端点未预弛豫时 fmax=0.141 eV/Å,mliport 按 --endpoint-fmax 0.02 拒绝启动(fail closed),改用文档的 --endpoint-policy relax 在运行内弛豫端点。

用例 结果
普通 NEB(5 内插,非爬坡) 23 步收敛,最大带力 0.0444 eV/Å(目标 0.05);前向 0.739337 eV / 反向 0.739353 eV(差 1.6×10⁻⁵ eV,对称路径自洽)
CI-NEB(5 内插,爬坡) 收敛;前向 0.750910 / 反向 0.750900 eV;barrier_status = converged_path_estimate,saddle_validation = not_performed(未做 Hessian,因此不称其为已验证过渡态)
端点未弛豫时的行为 Endpoint 0 fmax=0.141 eV/A exceeds endpoint_fmax_eV_A=0.02 → 拒绝启动(fail closed),改用 --endpoint-policy relax 后正常
中断(SIGINT) 在 checkpoints/step_000002 处中断:exit 130、status=cancelled、converged=false,band 7 帧完整保留
用完全相同的选项 resume exit 0、status=completed、收敛,势垒 0.751460 eV(与未中断的 0.750910 eV 差 5.5×10⁻⁴ eV;几何续算、不重放优化器历史,与文档一致)
resume 换模型 拒绝:Resume fingerprint is incompatible with this model, backend, task/head/dtype, cell/PBC, identity, constraints, or band setup,checkpoint 未被破坏
resume 改 --max-steps 未拒绝:exit 0 并跑完;resolved_config.json 里 max_steps 仍是记录值 200(source 标为 CLI)→ 用户显式改的值被静默忽略(见 §10 NEB-RESUME-01)
resume 去掉 --climb 同上:静默沿用记录的 climb=true,不报错
损坏 checkpoint 早期尝试中把 latest 指针写成普通文件 → 明确拒绝:Refusing to replace non-symlink checkpoint pointer,不静默重头开始

梯度预检真实执行:Pre-flight passed: 12 checks on images [0,2,4,6], max |FD − (−F·u)| = 5.1×10⁻⁴ eV/Å。


7. 构型搜索(confsearch)

体系沿用仓库 confsearch_suite.py 的 LGPS 切片 + 甲醇(模式 B 自动放置),--n-starts 4 --cycles 2,--hold-ps 0.5 --cool-ps 0.5,seed 20261117。confsearch run 没有 --device(R10-F05 复现),设备通过 --settings settings.ini 的 DEVICE = cuda:0 传入,这是当前支持的路径。

项 结果
参考运行 status=completed,7 个唯一极小,最优 −123.911922 eV,2/8 周期出现 md_integrity_abort(吸附质键断裂保护)→ 失败计数如实记录
中断 + confsearch resume <dir> 完成,7 个极小,最优 −123.911918 eV
对照组:同参数重跑(不中断) 7 个极小,与参考的最大能量差 4.4×10⁻⁴ eV(中位 7.7×10⁻⁵)
中断+续算 vs 参考 7 个极小,最大能量差 1.2×10⁻³ eV(中位 5.9×10⁻⁵)
换模型 resume 拒绝:model.model_sha256: recorded 'f1309e91…' != current '49a341ff…'
换 task resume 拒绝:model.task: recorded 'bulk' != current 'molecule'
损坏 checkpoint 未 crash:从仍有效的 checkpoint 继续并完成;无半写结果被当成有效数据

结论:续算偏差与"同参数重跑"的偏差同量级(10⁻³ eV),因此续算是忠实的;DPA float32 后端本身在 MD 重启后不保证位级可重复,这一点由对照组量化,而不是假设。


8. 输运与分析(真实轨迹 + 已知答案)

轨迹:α-Na₃PS₄ 2×2×2(128 原子),NVT-NHC,600/700/800 K,2 ps 平衡 + 10 ps 生产,1 fs 步长,每 10 fs 一帧(1001 帧)。

T (K) D_tracer 中位 (m²/s) σ_NE 中位 (mS/cm) σ_collective (mS/cm) Haven 拟合滞后点数
600 3.18×10⁻¹¹(95% CI 3.1×10⁻¹²…7.0×10⁻¹¹) 17.5(CI 1.7…38.6) 61.0 0.287 802
700 9.63×10⁻¹¹ 45.2 未启用该选项 – 802
800 8.23×10⁻¹¹ 33.8 27.8 1.04 802

这些数字不能当作 Na₃PS₄ 的电导率。 700→800 K 的 D 非单调(3 个温度、每个±2 倍不确定度),Arrhenius 拟合得到 Ea = 0.120 ± 0.121 eV、R² = 0.43,即"没有可分辨的 Arrhenius 行为"。按任务书要求,本轮结论是 pipeline_validated / sampling_insufficient,不是材料结论。collective 与 Nernst-Einstein 的差别(Haven 0.29)在 10 ps 尺度不足以判定为真实相关运动。

已知答案校验(与真实轨迹无关,纯算法):

  • Arrhenius:用 D₀=10⁻⁴ m²/s、Ea=0.3 eV 合成三个温度点 → 反演得到 Ea = 0.29999995 ± 2.6×10⁻⁸ eV、D₀ = 9.99999×10⁻⁵、R² = 0.9999999999。
  • MSD/kinisi/GEMDAT 的已知答案路径由 test_science_analysis_known_answers.py 覆盖,并在具备 kinisi/pymatgen 的环境中实跑通过(33 passed)。

其它分析模块在 800 K 轨迹上全部真实执行:validate、thermo、msd(含 mass-weighted COM 漂移修正)、rdf(Na–S、Na–Na)、vacf、spectrum、density、rmsd、transport、transport --collective-conductivity 均 exit 0 并写出 JSON/CSV/图与 provenance。

GEMDAT 在三温度全部 Error: No jumps found(显式 --sites 用理想 2×2×2 的 Na 位点)。这是 10 ps 内位点跳跃事件不足,属统计不足,不是软件错误;仓库自身的 analysis_suite.py 也要求 ≥24 次事件才给定量结论。


9. 可靠性:队列、故障注入、缓存与性能

队列(隔离 MLIPORT_JOBS_DIR,6+3 个 CPU 任务)

  • max_concurrent=1 下实测并发度始终为 1,作业时间区间零重叠(overlapping_pairs: [])。
  • 第二个调度器被拒绝:Error: A scheduler is already running。
  • 暂停/恢复(单个作业与整队)生效,状态机为 pending → paused → pending → running → done。
  • 作业失败(损坏结构)如实记为 failed,不影响后续作业;汇总 done 6 / failed 1。
  • SIGTERM 到前台调度器:0.5 s 内退出(退出码 −15;一个正在运行作业的 job_worker 属于独立会话,继续跑完并自行写回状态)、重启后队列继续排空、结束时无残留调度器进程。注意:这是"调度器被杀但正在运行的 worker 存活"的设计,若 worker 本身被 SIGKILL,作业会停在 running,直到下一次调度器的 stale 恢复逻辑接手(本轮未单独验证这一点)。

输出与磁盘故障注入

用例 结果
只读输出目录 拒绝并提示:Cannot write in output directory … Permission denied,未产生任何结果文件
已存在结果的目录(无 --overwrite) 拒绝,并列出已存在的文件
两个进程同时写同一目录 第二个被运行目录锁拒绝
轨迹被截断 analyze 硬失败(buffer is smaller than requested size,exit 1),不写分析结果
ENOSPC(定向注入到 JsonWriter.write) 抛出 OSError: [Errno 28],只留下 OUTCAR,没有 mliport_results.json、没有 success 声明、没有残留临时文件
单点运行目录交给 analyze 报 [Errno 21] Is a directory(真实原因是"没有轨迹")——信息误导,见 §10 R10-F07

模型切换、缓存与显存

  • DPA A→B→A:三次同一模型同一结构的能量为 −42.8126919 / −42.8126899(B 之后)/ −42.8126919(另一结构之后)eV。差 2×10⁻⁶ eV,与官方 harness 测得的 DPA float32 重复性下限同量级 → 没有证据表明存在跨模型的状态污染;同时说明 float32 后端本身不位级可重复,任何"逐位相同"的断言都不成立。
  • 模型加载失败(不存在路径)后立刻加载正常模型:能量与失败前完全相同 → 失败不污染后续加载。
  • 连续 8 次创建/销毁 DPA 计算器:torch.cuda.memory_allocated() 恒为 17.039 MB → 无泄漏。
  • GRACE 邻居缓存 ON/OFF:ΔE = 7.2×10⁻⁷ eV、Δσ ≤ 2.5×10⁻⁹ eV/ų(TF float32 噪声)。
  • MACE dtype 切换:info() 正确报告 ['float32'] / ['float64'],能量差 6.6×10⁻⁶ eV,不混淆精度身份。

性能(官方 T8 套件,全部 pass)

模型 SP 冷启动 SP 暖 32/128/512 原子 原子/秒 (512) MD 128 原子 MD 512 原子
DPA-3.1-3M 14.04 s(32 原子,含编译) 165.7 / 167.9 / 177.6 ms 2893 5.94 步/s(760 atom-step/s) 5.64 步/s(2889 atom-step/s)
UMA-s-1p2 2.05 s 103.7 / 105.2 / 205.0 ms 2501 9.50 步/s(1216 atom-step/s) 4.88 步/s(2499 atom-step/s)

这些数字只属于"本机 V100 + 该模型/版本",不推广到其他 GPU。


10. 发现的缺陷与修复

R10-F06(新,P1,已修复 → PR #110,Draft)

  • 症状:DPA 后端返回 converged = np.bool_(False);classify_result 用 converged is False 判断,np.False_ is False 为假 → 落到 COMPLETED。
  • 影响:mliport opt 打印 Converged: No 却 exit 0(文档承诺 status 2);mliport batch --calc-type opt 记为 completed/success: true/not_converged: 0;队列作业按退出码派生状态同样错。不会改变已写入的 converged 字段,但会让下游脚本/汇总把未收敛结构当成成功。
  • 最小复现:mliport opt cu32_rattled.vasp --fmax 0.0001 --max-steps 5(DPA4C-Nano, GPU)。
  • 修复:比较前归一化为 Python bool;无法解释为单一布尔值时 fail closed 为 NOT_CONVERGED。
  • 红/绿:新增 2 个失败用例(batch 参数化 + classify_result),修复后 2317 passed。CI 12/12 checks SUCCESS,MERGEABLE/CLEAN。

R10-F01(旧账,P2,已修复待审 → PR #111,Draft,CI 12/12 SUCCESS)

  • 症状:NPT 的 4 个纯配置冲突(thermostat≠LANGEVIN、--no-write-trajectory、temperature≤0、npt_tdamp/pdamp≤0)在 results/、resolved_config.json、模型加载之后才拒绝。
  • 修复:四条校验前移到 resolver._validate_resolved(),错误文本与 MDRunner 完全一致;MDRunner 内校验保留(API 直调仍受保护)。
  • 证据:修复前 4 个用例全部 results_dir=Y resolved=Y model_loaded=Y;修复后全部 N/N/N,退出码仍为 1、消息不变。合法 NVE / NVT-NHC / NPT 的 md.csv 与修复前逐字节相同(md5 一致)。
  • 仍残留:NHC + FixAtoms 的拒绝仍发生在模型加载之后(它依赖结构,不是纯配置);设备不可用时的报错仍是原始 torch 文本(Attempting to deserialize object on a CUDA device…)。这两条记录为遗留项,不在本轮修改。

NEB-RESUME-01(新,P2,未修复)

docs/workflows/neb.md 明确写:"Step budgets are part of that identity in this
beta, so a resume that changes --max-steps is rejected with
Resume fingerprint is incompatible。"

本轮实测:从同一个已取消(cancelled,3 个 checkpoint)的 CI-NEB 复制出三份,
分别用 ① 完全相同的选项、② --max-steps 123(记录值 200)、③ 去掉 --climb
去 resume。三者全部 exit 0 并跑完。检查 resolved_config.json:max_steps
仍是记录值 200,source 标成 CLI——也就是说 checkpoint 层的 run_options
覆盖了用户显式传入的值,指纹因此永远与记录一致,"改了参数就拒绝"这条文档承诺
没有被执行。

  • 科学影响:不会污染数据(band 与协议保持一致,续算忠实),但会静默忽略用户意图(例如用户以为自己在切换 CI-NEB,实际没有),与 AGENTS.md 第 2/3 条"不得静默改变语义、不满足假设时 fail closed"相悖。
  • 为什么本轮不改:修复点在 cmd_neb 的 resume 分支与 resolver 的层优先级(checkpoint_resolved_layer 作为 incar_layer 注入,实际优先级高于 CLI)。改动会波及 INCAR/设置/alise 等所有 resume 入口的既有语义,属于需要重新设计的公共行为;按任务书"若提前校验需要改变公共 API/重新设计,则停止代码修改,记录技术障碍",本轮只记录。
  • 建议(供 owner 决策):在 resume 分支里对显式传入的 CLI 参数与 checkpoint 记录值逐项比较,冲突时抛出文档承诺的错误(并列出冲突项名);或在文档中明确"resume 一律沿用 checkpoint 的记录值,命令行同名参数被忽略"。两条路都必须二选一,现状是第三种(文档说拒绝、行为是静默覆盖)。

其它已确认但未修改的问题

编号 级别 现象 处理
R10-F07 P3 mliport analyze <单点运行目录> thermo 报 [Errno 21] Is a directory,实际原因是"该目录里没有轨迹";错误信息误导 只记录,建议改成明确的"no trajectory in this run"
R10-F08 P3 分析结果 id 不覆盖 raw/mliport_results.json;结果文件被截断时,若同 id 的分析已完成则直接复用(--force 后在轨迹被截断时报 buffer is smaller than requested size) 只记录;分析主输入是轨迹,复用语义本身有文档
R10-F05 P3 mliport confsearch run 没有 --device,只能经 settings/INCAR 传设备;confsearch resume 只接受计算器选项 只验证既有路径可用,建议补文档说明,不新增选项
R10-F04 P3 mliport/utils.py(191 行、8 函数)无任何 importer,仍随 wheel 发布 确认无引用;按任务书不删除,只记录
— 观察 官方 science harness 有 6 条 fail 记录,全部是 float32 旋转/平移不变性在 10⁻⁶ eV/Å 量级的偏差(阈值 = 10×实测重复性下限) 保留 fail 原样,本轮不擅自改判;这是后端精度而非产品缺陷

11. 未完成项与风险

没有做、或只做了一部分的(按任务书要求如实分类)

项 状态 原因
MLIP vs DFT 精度对照(Na₃PS₄ / LPSC / Li₂S) BLOCKED 本机没有元数据完整且已收敛的 CP2K/VASP 标签;旧 Na₃PS₄ 力标签未充分收敛,按规则不得当真值;本轮无 DFT 授权
LLZO(Li₇La₃Zr₂O₁₂)工作流 部分 只用 mp-942733 做了 SP + E/F/stress(5 个模型);96 原子单胞的 OPT/MD/输运未做
LPSC 深度工作流 部分 SP(7 模型)、固定胞 OPT、1500 K 短时 NVT 完成;未做变胞 OPT、NPT、长 MD、输运(无 Na/Li 迁移统计意义)
LGPS 输运 / Arrhenius NOT_RUN 时间与 GPU 预算优先给了 Na₃PS₄ 三温度;LGPS 只做了 SP、NPT(0/2 GPa)与 examples/ 结构一致性核对
Na₃PS₄ 电导率 / 活化能 INCONCLUSIVE 10 ps×3 温度,D 的不确定度约 2 倍,700→800 K 非单调;Arrhenius Ea=0.120±0.121 eV、R²=0.43 → 只能说"算法正确、采样不足"
GEMDAT 位点/跳跃定量 INCONCLUSIVE 三温度均 No jumps found(10 ps 内事件不足);仓库自身要求 ≥24 事件
NEP89 原生 GPUMD 路径 BLOCKED GPUMD/src/gpumd 缺少 libnvJitLink.so.12(其余 CUDA 库可从 uv 缓存补齐);按授权不重新编译 GPUMD
NEP89 ASE/Calorine 路径 UNSUPPORTED 新版 nep.txt 格式 Calorine 无法解析(已知限制)
官方 science harness T6(MD 套件) NOT_RUN(已启动后中止) 固定 profile 是 DPA-3.1-3M,实测约 4.6 步/s(128 原子),单是 NVE 步长扫描就远超本轮剩余预算;改用本报告 §5 的自建 MD 矩阵覆盖,并在 #75 记录中止事实与速率
官方 science harness T5(NEB 套件) NOT_RUN 其 Na₃PS₄ hop 与 Cu 空位端点需要内置 fixture 的完整 8 条工作流,耗时不可控;改用 §6 的直接 CLI NEB/CI-NEB + 中断/续算覆盖
MACE 官方 pinned profile BLOCKED model_manifest.json 钉的是 mace-omat-0-medium(sha256 d4b14be9…),本机没有该文件;mace-mh-1 是不同 artifact,按规则不得替代
TUI NOT_RUN 交互式界面无法在无人值守下做真实点击;其代码路径由 pytest tests/mliport/mliport/test_tui.py 覆盖(本轮全绿)
INCAR / settings / API / CLI 语义一致性 部分 通过 INCAR(run + npt 用例)、settings(confsearch 的 DEVICE)、CLI(全部用例)、Python API(F3 探针)分别跑通;没有做四入口的穷举笛卡尔积
队列的 stale-PID / 被盗 lease 恢复 部分 SIGTERM 调度器→worker 继续→重启排空已验证;SIGKILL 掉 worker 后的 stale 恢复未单独构造
500 原子以上的长时间 MD/NEB NOT_RUN 只做了 512 原子 SP/短 MD 性能冒烟(T8),符合"不强行跑 500 原子"的约束

12. 统计口径

154 条用例:PASS 119 / CHARACTERIZED 31 / INCONCLUSIVE 1 / FAIL 2 / NOT_RUN 1(两个 FAIL 都是 NEB resume 的"文档承诺未执行",不是算错数)。harness 记录 96 个带日志的用例(83 OK / 13 非零退出,后者含预期中的拒绝与负例),GPU 墙钟约 3 h、磁盘 509 MB、无残留进程/调度器/GPU lease。逐条证据见 CASE_MATRIX.md 与 summary.json。

13. 稳定版本建议

建议:可以打"稳定测试版"标签(2.0.0b4 之后的下一个 beta),但不要打正式 2.0.0。

支持这个判断的证据:

  1. 2316 个测试全绿、lint/docs/hygiene 全清;本轮 154 条验收用例中 119 PASS、31 CHARACTERIZED、1 INCONCLUSIVE、2 FAIL、1 NOT_RUN(后两者都是文档承诺未被执行,不是算错数)。
  2. 真实 V100 上四个后端、16 个本地 checkpoint、7 种晶胞、SP/OPT/batch/NEB/CI-NEB/confsearch/NVE/NVT/NPT/输运分析全部跑通并且产物可追溯。
  3. 全部"失败即静默成功"的路径都做了主动攻击:损坏结构、空结构、非周期输入、只读目录、并发写、ENOSPC、截断轨迹、换模型续算、损坏 checkpoint、非法 MD 组合。除 §10 的两条已修复缺陷外,其余全部 fail closed 并留下可诊断的 provenance。
  4. 单位、符号、Voigt 顺序、PBC、task/head/dtype 身份在 CLI 输出、JSON、OUTCAR、artifacts.json 之间一致,并且有独立已知答案(Arrhenius 反演精确到 1e-8;Si 变胞 a=5.3751 Å 与菱方原胞 3.8003 Å 自洽;Cu 空位势垒 0.739/0.751 eV 落在文档区间)。

打正式版之前仍应处理:

  1. 合并或明确拒绝 fix(run-status): numpy converged flag no longer becomes success (R10-F06/P1) #110(numpy bool → 未收敛被记为成功)。这是本轮唯一会悄悄污染批量/队列统计的缺陷;不合并就应在发布说明里明确"未收敛判定不可靠"。
  2. 裁定 NEB resume 的文档与行为冲突(NEB-RESUME-01)。科学数据不会因此损坏,但"改了参数却被静默忽略"在科研软件里不可接受。
  3. 把 §11 的 INCONCLUSIVE 项(Na₃PS₄ 电导率、GEMDAT 跳跃)从任何"材料结论"里排除;它们目前只是软件闭环证据。
  4. 若要在发布说明里声称"支持 NEP89",必须先解决原生 GPUMD 的 CUDA 运行库依赖(本机缺 libnvJitLink.so.12)或明确只支持旧格式。

一句话:软件工程与科学契约层面可以定版;材料精度与统计结论层面不能。 这两件事不要写在同一个句子里。

CASE_MATRIX.md(154 条用例)
Case ID Material Model Task Device Status Evidence
T1-dpa-sse-legacy Si/Cu/Li2S/Na3PS4/LPSC/LGPS SSE-VASP-dpa2 (legacy runtime) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa-sse-legacy.json: 6 materials, load_error=None
T1-dpa31-3m-omat24 Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA-3.1-3M (Omat24) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa31-3m-omat24.json: 6 materials, load_error=None
T1-dpa4-air Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA4-Air SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa4-air.json: 6 materials, load_error=None
T1-dpa4-neo Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA4-Neo SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa4-neo.json: 6 materials, load_error=None
T1-dpa4-plus Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA4-Plus SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa4-plus.json: 6 materials, load_error=None
T1-dpa41c-air Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA-4.1C-Air SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa41c-air.json: 6 materials, load_error=None
T1-dpa4c-mini Si/Cu/Li2S/Na3PS4/LPSC/LGPS DPA4C-Mini SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa4c-mini.json: 6 materials, load_error=None
T1-dpa4c-nano Si/Cu/Li2S/Na3PS4/LPSC/LGPS/LLZO DPA4C-Nano SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-dpa4c-nano.json: 7 materials, load_error=None
T1-grace-2l-omat-fp64 Si/Cu/Li2S/Na3PS4/LPSC/LGPS GRACE-2L-OMAT-medium-base fp64 SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-grace-2l-omat-fp64.json: 6 materials, load_error=None
T1-grace-2l-omat Si/Cu/Li2S/Na3PS4/LPSC/LGPS/LLZO GRACE-2L-SMAX-OMAT-medium SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-grace-2l-omat.json: 7 materials, load_error=None
T1-mace-mh1-omat Si/Cu/Li2S/Na3PS4/LPSC/LGPS/LLZO mace-mh-1 head=omat_pbe SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-mace-mh1-omat.json: 7 materials, load_error=None
T1-uma-ft-r2scan Si/Li2S UMA-s-1p2p1-ft-r2scan (omat) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-uma-ft-r2scan.json: 2 materials, load_error=None
T1-uma-s1p1-omat Si/Cu/Li2S/Na3PS4/LPSC/LGPS UMA-s-1p1 (omat) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-uma-s1p1-omat.json: 6 materials, load_error=None
T1-uma-s1p2-omat Si/Cu/Li2S/Na3PS4/LPSC/LGPS/LLZO UMA-s-1p2 (omat) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-uma-s1p2-omat.json: 7 materials, load_error=None
T1-uma-s1p2p1-omat Si/Cu/Li2S/Na3PS4/LPSC/LGPS UMA-s-1p2p1 (omat) SP T1 (E/F/stress, identity) cuda:0 PASS probe/T1-uma-s1p2p1-omat.json: 6 materials, load_error=None
SCI-dpa-t1_cu_fcc cu DPA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_cu_fcc/t1_real_inference/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t1_cu_fcc cu UMA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_cu_fcc/t1_real_inference/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t1_mgo_rocksalt mgo DPA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_mgo_rocksalt/t1_real_inference/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t1_mgo_rocksalt mgo UMA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_mgo_rocksalt/t1_real_inference/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t1_si_diamond si DPA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_si_diamond/t1_real_inference/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t1_si_diamond si UMA science harness t1_real_inference cuda:0 PASS science_commit/summary.json[t1_si_diamond/t1_real_inference/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_ab_reversibility/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_ab_reversibility/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_global_rotation cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_distorted/t2_invariance_global_rotation/dpa-upstream-model-defined-24dc1695f3] = fail
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_global_rotation cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_global_rotation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_global_translation cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_distorted/t2_invariance_global_translation/dpa-upstream-model-defined-24dc1695f3] = fail
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_global_translation cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_distorted/t2_invariance_global_translation/uma-upstream-model-defined-f16c843796] = fail
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_pbc_wrap/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_pbc_wrap/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_permutation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_cu_distorted/t2_invariance_permutation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_distorted cu DPA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_distorted/t2_invariance_repeatability/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2_cu_distorted cu UMA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_distorted/t2_invariance_repeatability/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_ab_reversibility/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_ab_reversibility/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_global_rotation cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_global_rotation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_global_rotation cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_fcc/t2_invariance_global_rotation/uma-upstream-model-defined-f16c843796] = fail
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_global_translation cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_global_translation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_global_translation cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_global_translation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_pbc_wrap/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_pbc_wrap/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_permutation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_cu_fcc/t2_invariance_permutation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_cu_fcc cu DPA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_fcc/t2_invariance_repeatability/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2_cu_fcc cu UMA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_cu_fcc/t2_invariance_repeatability/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_ab_reversibility/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_ab_reversibility/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_global_rotation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_global_rotation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_global_rotation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_global_rotation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_global_translation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_global_translation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_global_translation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_global_translation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_pbc_wrap/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_pbc_wrap/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_permutation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_permutation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_mgo_rocksalt mgo DPA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_repeatability/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2_mgo_rocksalt mgo UMA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_mgo_rocksalt/t2_invariance_repeatability/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_ab_reversibility/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_ab_reversibility cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_ab_reversibility/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_global_rotation cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_global_rotation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_global_rotation cuda:0 CHARACTERIZED science_commit/summary.json[t2_si_diamond_prim/t2_invariance_global_rotation/uma-upstream-model-defined-f16c843796] = fail
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_global_translation cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_global_translation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_global_translation cuda:0 CHARACTERIZED science_commit/summary.json[t2_si_diamond_prim/t2_invariance_global_translation/uma-upstream-model-defined-f16c843796] = fail
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_pbc_wrap/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_pbc_wrap cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_pbc_wrap/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_permutation/dpa-upstream-model-defined-24dc1695f3] = pass
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_permutation cuda:0 PASS science_commit/summary.json[t2_si_diamond_prim/t2_invariance_permutation/uma-upstream-model-defined-f16c843796] = pass
SCI-dpa-t2_si_diamond_prim si DPA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_si_diamond_prim/t2_invariance_repeatability/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2_si_diamond_prim si UMA science harness t2_invariance_repeatability cuda:0 CHARACTERIZED science_commit/summary.json[t2_si_diamond_prim/t2_invariance_repeatability/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_cu_distorted cu DPA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_distorted/t2_force_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_cu_distorted cu UMA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_distorted/t2_force_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_cu_distorted cu DPA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_distorted/t2_stress_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_cu_distorted cu UMA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_distorted/t2_stress_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_cu_fcc cu DPA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_fcc/t2_force_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_cu_fcc cu UMA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_fcc/t2_force_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_cu_fcc cu DPA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_fcc/t2_stress_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_cu_fcc cu UMA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_cu_fcc/t2_stress_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_mgo_rocksalt mgo DPA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_mgo_rocksalt/t2_force_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_mgo_rocksalt mgo UMA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_mgo_rocksalt/t2_force_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_mgo_rocksalt mgo DPA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_mgo_rocksalt/t2_stress_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_mgo_rocksalt mgo UMA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_mgo_rocksalt/t2_stress_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_si_diamond_prim si DPA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_si_diamond_prim/t2_force_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_si_diamond_prim si UMA science harness t2_force_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_si_diamond_prim/t2_force_fd/uma-upstream-model-defined-f16c843796] = characterized
SCI-dpa-t2fd_si_diamond_prim si DPA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_si_diamond_prim/t2_stress_fd/dpa-upstream-model-defined-24dc1695f3] = characterized
SCI-uma-t2fd_si_diamond_prim si UMA science harness t2_stress_fd cuda:0 CHARACTERIZED science_commit/summary.json[t2fd_si_diamond_prim/t2_stress_fd/uma-upstream-model-defined-f16c843796] = characterized
opt_cu32_cellopt_dpa cu32 DPA4C-Nano OPT cell cuda:0 PASS a2_opt/opt_cu32_cellopt_dpa.json converged=True nsteps=12 fmax=1.0635959459132094e-06
opt_cu32_maxsteps_dpa cu32 DPA4C-Nano OPT fixed cell cuda:0 INCONCLUSIVE a2_opt/opt_cu32_maxsteps_dpa.json converged=True nsteps=0 fmax=5.65804145670739e-07
opt_lpsc_26_fixed_dpa lpsc DPA4C-Nano OPT fixed cell cuda:0 PASS a2_opt/opt_lpsc_26_fixed_dpa.json converged=True nsteps=22 fmax=0.01756567727352376
opt_na3ps4_32_cellopt_dpa na3ps4 DPA4C-Nano OPT cell cuda:0 PASS a2_opt/opt_na3ps4_32_cellopt_dpa.json converged=True nsteps=28 fmax=0.01068125585152302
opt_na3ps4_32_fixed_dpa na3ps4 DPA4C-Nano OPT fixed cell cuda:0 PASS a2_opt/opt_na3ps4_32_fixed_dpa.json converged=True nsteps=11 fmax=0.012493796122726053
opt_na3ps4_sheared_cellopt_dpa na3ps4 DPA4C-Nano OPT cell cuda:0 PASS a2_opt/opt_na3ps4_sheared_cellopt_dpa.json converged=True nsteps=59 fmax=0.019162205639986496
opt_si8_cellopt_dpa si8 DPA4C-Nano OPT cell cuda:0 PASS a2_opt/opt_si8_cellopt_dpa.json converged=True nsteps=20 fmax=4.729372808543126e-06
opt_si8_cellopt_mace si8 mace-mh-1 OPT cell cuda:0 PASS a2_opt/opt_si8_cellopt_mace.json converged=True nsteps=15 fmax=2.429995817060512e-14
opt_si_prim_nonortho_cellopt_dpa si DPA4C-Nano OPT cell cuda:0 PASS a2_opt/opt_si_prim_nonortho_cellopt_dpa.json converged=True nsteps=11 fmax=6.88305590371141e-08
BATCH-sp_baseline 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_baseline] counts={'completed': 11, 'failed': 3} reused=0 status_files=14
BATCH-sp_resume_noop 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_resume_noop] counts={'completed': 11, 'failed': 3} reused=11 status_files=14
BATCH-sp_resume_changed_input 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_resume_changed_input] counts={'completed': 11, 'failed': 3} reused=10 status_files=14
BATCH-sp_resume_changed_model 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_resume_changed_model] counts={'completed': 11, 'failed': 3} reused=10 status_files=14
BATCH-opt_limited_steps 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[opt_limited_steps] counts={'completed': 11, 'failed': 3} reused=0 status_files=14
BATCH-sp_interrupted 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_interrupted] counts={'completed': 11, 'failed': 3} reused=0 status_files=14
BATCH-sp_interrupted_resume 14 mixed DPA4C-Nano batch sp/opt + resume cuda:0 PASS batch_acceptance.json[sp_interrupted_resume] counts={'completed': 11, 'failed': 3} reused=11 status_files=14
cu_neb_plain Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_plain.json barrier_fwd=None barrier_rev=None converged=True exit=0
cu_neb_climb Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_climb.json barrier_fwd=None barrier_rev=None converged=True exit=0
cu_neb_step_limited Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_step_limited.json barrier_fwd=0.753931999206543 barrier_rev=0.7539281845092773 converged=False exit=2
cu_neb_resumed_from_limit Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_resumed_from_limit.json barrier_fwd=0.753931999206543 barrier_rev=0.7539281845092773 converged=False exit=1
cu_neb_interrupted Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_interrupted.json barrier_fwd=0.7509031295776367 barrier_rev=0.7509031295776367 converged=True exit=0
cu_neb_resumed_after_interrupt Cu (vacancy, 107 atoms) DPA4C-Nano NEB/CI-NEB cuda:0 PASS neb/cu_neb_resumed_after_interrupt.json barrier_fwd=0.7509021759033203 barrier_rev=0.7509098052978516 converged=True exit=0
NEB-interrupt-resume Cu (vacancy) DPA4C-Nano CI-NEB interrupted then resumed cuda:0 PASS interrupted status=cancelled exit=130; resumed barrier=0.751460075378418 eV vs uninterrupted None eV
NEB-resume-other-model Cu (vacancy) DPA4C-Nano vs Mini resume with another model cuda:0 PASS refused: ['NEB calculation failed: Resume fingerprint is incompatible with this model, backend, task/head/dtype, cell/PBC, identi
NEB-resume identical options Cu (vacancy) DPA4C-Nano resume identical options cuda:0 PASS exit=0 status=['NEB status: completed'] checkpoint step_000002:7:f1b6b98abf53 -> step_000025:7:ec2229e5a0f9
NEB-resume with --max-steps changed Cu (vacancy) DPA4C-Nano resume with --max-steps changed cuda:0 FAIL exit=0 status=['NEB status: completed'] checkpoint step_000002:7:f1b6b98abf53 -> step_000006:7:4203b56b5f31 — docs promise a rejection; the recorded options silently win
NEB-resume with --climb removed Cu (vacancy) DPA4C-Nano resume with --climb removed cuda:0 FAIL exit=0 status=['NEB status: completed'] checkpoint step_000002:7:f1b6b98abf53 -> step_000025:7:d59ef7bee311 — docs promise a rejection; the recorded options silently win
A1-voigt-dpa Na3PS4 sheared (artificial) DPA SP stress component order cuda PASS voigt=['-8.15e-03', '-1.86e-02', '8.53e-03', '3.44e-09', '-1.55e-08', '7.32e-03'] (xy non-zero, yz/xz ~0)
A1-voigt-mace Na3PS4 sheared (artificial) MACE SP stress component order cuda PASS voigt=['-5.53e-03', '-1.50e-02', '7.86e-03', '-1.23e-17', '1.08e-17', '6.90e-03'] (xy non-zero, yz/xz ~0)
A1-pbc-molecule H2O (non-periodic) DPA4C-Nano SP with task=bulk cuda PASS bulk task refused a non-periodic input
A1-pbc-molecule-ok H2O (non-periodic) mace-mh-1 SP with task=molecule cuda PASS molecule task accepted the same input
A1-pbc-slab2d Cu 2D slab DPA4C-Nano SP with task=bulk cuda NOT_RUN the slab fixture was written as a 3D-periodic cell, so 2D rejection was not exercised
A1-upstream-dpa Si8 DPA mliport CLI vs upstream ASE calculator cuda PASS dE=0.0 eV, upstream=deepmd 3.2.1.dev63+g0edeaef56 (upstream ASE DP calculator)
A1-upstream-mace Si8 MACE mliport CLI vs upstream ASE calculator cuda PASS dE=7.105427357601002e-15 eV, upstream=mace 0.3.16 (upstream MACECalculator)
CONFSEARCH-ref CH3OH on LGPS slab DPA4C-Nano confsearch mode B cuda:0 PASS confsearch_resume.json reference
CONFSEARCH-resume CH3OH on LGPS slab DPA4C-Nano confsearch interrupted + resume cuda:0 PASS confsearch_determinism.json
CONFSEARCH-refusal CH3OH on LGPS slab DPA4C-Nano resume with other model/task; corrupt state cuda:0 PASS confsearch_resume.json resume_other_model/task, corrupt
ARRHENIUS-known-answer - synthetic D0/Ea arrhenius fit cpu PASS Ea=0.29999995 eV recovered from a known 0.3 eV synthetic set
QUEUE-serialisation - DPA4C-Nano (cpu) queue max_concurrent=1 cpu PASS queue_acceptance.json overlaps=[]
QUEUE-race - - two schedulers cpu PASS second scheduler exit=1
QUEUE-pause-resume - - pause/resume job and queue cpu PASS queue_acceptance.json pause/resume_job/resume_queue
QUEUE-sigterm - - SIGTERM scheduler + restart cpu PASS exit=-15 delay=0.5s final={'t0': 'done', 't1': 'done', 't2': 'done', 't3': 'done'}
FAULT-output-writes - DPA4C-Nano readonly dir / existing dir / concurrency cuda:0 PASS f2_output_faults.json
FAULT-enospc - stub calculator ENOSPC at JsonWriter cpu PASS f2b_faults.json enospc_at_json_writer: OSError(28), no results json
FAULT-truncated-evidence - - truncated trajectory/results cpu PASS f2b_faults.json / faults2/f2c_md_faults.json
CACHE-A-B-A - DPA4C-Nano model switch + failed load recovery cuda:0 CHARACTERIZED A_first vs A_second delta=1.91e-06 eV (float32 backend floor); bad load then good load identical=True
CACHE-grace-neighbor Na3PS4 GRACE-2L neighbor cache ON vs OFF cuda:0 PASS |dE|=7.15e-07 eV
DTYPE-mace-32-64 Si mace-mh-1 float32 vs float64 identity cuda:0 PASS reported dtypes ['float32'] / ['float64']
npt_lgps_200_0GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=0.0144±0.0368 GPa, axis spread=6.661338147750939e-16, angle spread=0.0
npt_lgps_200_2GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=1.996±0.0276 GPa, axis spread=6.661338147750939e-16, angle spread=0.0
npt_na3ps4_32_0GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=0.0107±0.0279 GPa, axis spread=4.440892098500626e-16, angle spread=0.0
npt_na3ps4_32_2GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=1.9872±0.0353 GPa, axis spread=4.440892098500626e-16, angle spread=0.0
npt_na3ps4_sheared_0GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=0.0006±0.113 GPa, axis spread=4.440892098500626e-16, angle spread=0.0
npt_si8_0GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=0.011±0.0118 GPa, axis spread=0.0, angle spread=0.0
npt_si8_2GPa Si8/Na3PS4/LGPS DPA4C-Nano MD NPT (isotropic MTK) cuda:0 PASS P=2.0541±0.0226 GPa, axis spread=0.0, angle spread=0.0
nve_na3ps4_dt0.5 Na3PS4 DPA4C-Nano MD NVE cuda:0 PASS drift=1.2312704013890578e-07 eV/atom/ps over 3.0 ps
nve_na3ps4_dt1.0 Na3PS4 DPA4C-Nano MD NVE cuda:0 PASS drift=5.211118188851388e-07 eV/atom/ps over 3.0 ps
nve_si64_dt0.25 Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=-5.638071929660157e-09 eV/atom/ps over 3.0 ps
nve_si64_dt0.5 Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=-1.8014370387028793e-08 eV/atom/ps over 3.0 ps
nve_si64_dt1.0 Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=-7.587542154682241e-09 eV/atom/ps over 3.0 ps
nve_si64_dt1.0_repeat Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=1.5627503020165184e-08 eV/atom/ps over 3.0 ps
nve_si64_dt1.0_save1 Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=-5.100257581699348e-08 eV/atom/ps over 3.0 ps
nve_si64_dt1.0_seedB Si64 DPA4C-Nano MD NVE cuda:0 PASS drift=3.16893338315932e-07 eV/atom/ps over 3.0 ps
nvt_na3ps4_600K_LANGEVIN Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=587.02±7.075 K (target in name)
nvt_na3ps4_600K_NHC Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=579.8±17.838 K (target in name)
nvt_na3ps4_800K_BUSSI Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=784.03±18.488 K (target in name)
nvt_si64_300K_BUSSI Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=293.66±8.217 K (target in name)
nvt_si64_300K_LANGEVIN Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=294.45±6.094 K (target in name)
nvt_si64_300K_NHC Si64/Na3PS4 DPA4C-Nano MD NVT cuda:0 PASS T=294.77±8.135 K (target in name)

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant