-
stab60cdff44f2 · ·
follow stab59, 1. 完成 72/72 个 exact-config MultiGrid 对照单元和 144/144 side case，config/input/fair/残差/参考 warmup 门禁全部通过; 2. 新增 detached 进程组超时清理与回归测试，避免 mpirun/QUDA worker 超时后残留占用 GPU; 3. 新增 final-matrix 聚合器，导出 units/stages/references CSV、trace evidence 和 72 个 fair combined 文档的可重放流程; 4. 最终协议采用 c64 coarse-tol 0.003、c128 coarse-tol 3e-5、1 cold+2 warmup+5 steady，MG 48 组胜 38 组，P100/V100 中位 QUDA/PyQCU 为 1.158/2.115，并如实保留 P100 大格 c64 MG-2 持平边界; 5. 更新最终与快速迭代报告、PDF、AGENTS.md 和验证结果，149 个 CPU/协议测试、Tier 0、Strict CUDA、fused FGMRES、MPI preflight 全部通过; [codex].
-
stab59b2dadb85 · ·
follow stab58, 1. 为 Strict-MG complex64 分布式路径新增非阻塞 Halo 重叠与远端修正融合，并为 complex128 保留安全回退; 2. 修复 BiCGStab reference 的 cold/warmup/steady 合并、QUDA mapped-memory 重试和非 UTF-8 日志处理; 3. 修正矩阵层级/trace 标签与 reference 派生语义，新增 fail-closed 证据校验和 overlap 汇总工具; 4. 补充双 P100 MPI 真残差、最终 PyQCU/QUDA 对照数据与 MultiGrid 报告; [codex].
-
stab58165ff0dd · ·
follow stab57, 1. 修复双 P100 clover-gauge MPI halo 设备绑定与 pack 越界; 2. 新增 rank-sliced HDF5 输入与 trace-off/trace-on runtime cache 复用; 3. 完成 Strict halo 活跃 slot、link cache 与 prolongation 融合优化; 4. 完成 144 side-case QUDA/PyQCU 对照、报告、残差图和 89 项回归; [codex].
-
stab57bd34d64e · ·
follow stab56, 1. 补齐 c128/V100 矩阵：c64 canonical 输入回退、PyQCU MG-2/3 全格点结果、QUDA 小/中格结果及大格容量缺口; 2. 重构 strict Galerkin 显存生命周期，通过临时块释放、方向流式搬运与 blocked-basis offload 完成 V100 大格 c128; 3. 修复矩阵 resume 的 config-hash 冲突并让 runtime 资产绑定对非 CUDA 资产 fail-closed; 4. 同步 c128 分析报告、P100 复验记录与 benchmark skill，清理旧 report-build/render 产物; 5. 84 项快速回归和 history tag-chain 检查通过; [codex].
-
stab5618a1db9c · ·
follow stab55_35, 1. implement distributed Strict-MG runtime: fine face halo by default, coarse full/compact vector+link halos, R/P boundary contracts, global FGMRES reductions and multi-rank fused-coarse fallback; 2. fix dot_many device write-back, vector-halo component stride, compact link-halo packing and backward-correction sign found by the new MPI probes; 3. add distributed Galerkin setup (periodic face-exchange roll + halo-padded fine operator) with process_grid/comm plumbing and fail-closed gates; 4. align finest-level total iteration semantics with QUDA outer GCR/FGMRES total_iter and add strict trace v3 iteration_count records; 5. land collector phase/device/MPI extensions, full matrix orchestrator and JSON->CSV/SVG/PDF/LaTeX report generator with the V100 c64 matrix and documented MG-1/c128/P100 gaps; evidence: c128 coarse operators match the global reference to 1e-16 on 1/2/4 ranks and distributed solves reach <1e-6 true residual. [codex].
-
stab55_35305e455f · ·
follow stab55_34, 1. release duplicate Python setup hierarchy before cold C++ applyInitQcu; 2. close c128 cold C=1 OOM with setup 2709.45 s and passing solve; 3. preserve CUDA/MPI macro and architecture provenance fixes. [opencode].
-
stab55_3441888526 · ·
follow stab55_33, 1. record post-macro-fix full regression PASS 5/5 in 755.5 s. [opencode].
-
stab55_33b8eb6bac · ·
follow stab55_32, 1. fix CUDA/MPI error macro double evaluation; 2. confirm cold c128 C=1 setup fails with real OOM error 0002; 3. add static regression and update evidence. [opencode].
-
stab55_321d27645d · ·
follow stab55_31, 1. test requested CUDA architecture provenance. [opencode].
-
stab55_3100448030 · ·
follow stab55_30, 1. add libqcu path/SHA/cubin architecture provenance; 2. test cuobjdump parser; 3. document architecture checks. [opencode].
-
stab55_30d356ed8c · ·
follow stab55_29, 1. clarify mixed CUDA-architecture scope in the final table. [opencode].
-
stab55_298a3d721f · ·
follow stab55_28, 1. correct CUDA architecture list syntax in report/testing docs. [opencode].
-
stab55_28248a0055 · ·
follow stab55_27, 1. adopt conservative c64 large ratio 2.0052x; 2. retain 1.9898x as reproducibility range; 3. sync report and skills. [opencode].
-
stab55_275c5f9216 · ·
follow stab55_26, 1. adopt conservative c128 large formal ratio 7.8417x; 2. retain first run 8.3507x as reproducibility range; 3. update BiCGStab and skill records. [opencode].
-
stab55_261ab4615c · ·
follow stab55_25, 1. record full dev87 regression on native sm70 cubins; 2. verify PASS 5/5 in 751.2 s. [opencode].
-
stab55_25816808d3 · ·
follow stab55_24, 1. verify and document PyQCU sm70 native rebuild; 2. record side-only solve deltas versus default sm60 JIT; 3. preserve architecture provenance. [opencode].
-
stab55_24710dbab4 · ·
follow stab55_23, 1. conclude parity asset equivalence and cold-setup gain; 2. record multi-GPU hardware limitation in the final conclusions. [opencode].
-
stab55_23b61d64e9 · ·
follow stab55_22, 1. document real multi-GPU capability gap; 2. reject duplicate-UUID pseudo multi-GPU evidence. [opencode].
-
stab55_22f0eda106 · ·
follow stab55_21, 1. confirm PyQCU trace outer iteration count matches formal median; 2. document fixed level mapping and distinguish 20/60 from parity-layer errors. [opencode].
-
stab55_219d6b23f6 · ·
follow stab55_20, 1. reconcile c128 large capability status; 2. distinguish extended C=4 full-double success from formal cold C=1 limitation; 3. update testing skill. [opencode].