feat: report a per-operator speedup range across two runs

The speedup column in a result table compares gems against torch, not one build against another, so an A/B needs the two runs paired by shape. batch_summary gains an --ab mode that indexes both run.logs by operator, dtype and size detail, computes before/after per pair, and prints the per-shape rows plus the range. Only rows that succeeded on both sides with a positive latency count, so an incomplete pairing cannot masquerade as a complete range. When either side carries a BENCHMARK_DIRECT_NO_CUDAGRAPH marker the range is annotated, because a fallback row was timed under a different regime than a captured one. ab_rlc_test.sh drives the pairing for the FlagTree RemoveLayoutConversions switch the way ab_fold_test.sh does for the FlagGems one: per operator, run with the enhancement off while recording autotune configs, then replay those configs with it on, so the two sides differ by that switch alone. Each operator therefore leaves exactly two run.log files, and the summary of each pair is printed as it completes. Checked on an S5000 against gelu_backward over five shapes: both sides captured with no fallback, five rows paired, range -0.43% to +0.09%.
feat: drop stale CUDA-only skip guards on MThreads
2026-09-02 12:19:26 +08:00 · 2026-09-02 12:19:26 +08:00 · 2026-09-02 12:19:07 +08:00 · 2026-09-02 12:19:06 +08:00
26 changed files with 537 additions and 1650 deletions
@@ -20,11 +20,6 @@ OP=softmax OP_FILE=softmax SHAPE_FILE=my_shapes.yaml bash run_pytest.sh
 # 测 FlagGems-vllm 仓库的算子：换 FLAGGEMS_DIR 即可（插件自动适配其包名 flaggems_vllm）
 FLAGGEMS_DIR=/workspace/dev/FlagGems-vllm OP=fused_marlin_moe OP_FILE=fused_marlin_moe bash run_pytest.sh
 # 对齐 FlagGems 原生节点（不传 --shape_file，沿用 --level core 的 core_shapes）
 ZL_BACKEND=cuda \
 TEST_NODE=/workspace/FlagGems/benchmark/test_fused_marlin_moe.py::test_fused_marlin_moe_mxfp4 \
 bash run_pytest.sh
 ```
 上游现为两个仓库，benchmark 体系同构、均受支持：主仓 `/workspace/dev/FlagGems`（Python 包名 `flag_gems`，默认；`fused_marlin_moe_mxfp4` 仅此仓有）与 `/workspace/dev/FlagGems-vllm`（包名 `flaggems_vllm`）。两包需各自 `pip install -e <仓库> --no-deps` 安装（`--no-deps` 避免动 FlagTree 的 triton）。
@@ -40,7 +35,7 @@ softmax:
 上面这条命令是 **record 模式**：现场 autotune、把选中的 config 存档。它用于单次摸底或给 replay 提供基准，**两次 record 的数字不能互相比**——做对比走下面的「A/B 对比测试标准流程」。
-跑之前确认目标卡空闲：共享机器上别的任务会把两侧一起等比拖慢，出一份看似自洽实则作废的数据。用 `ZL_BACKEND` 选运行时（默认 `cuda`），再用该后端的可见性变量选卡（CUDA 为 `CUDA_VISIBLE_DEVICES`，也可设 `ZL_DEVICE`）。**测量始终单进程单设备串行**，这是计时可比的前提；多卡仅用于可选的 sweep 加速（见「加速：多卡并行 sweep」）。非 CUDA 后端默认走上游 `do_bench`，不会尝试 CUDA Graph。
+跑之前先 `nvidia-smi` 确认目标卡空闲：共享机器上别的任务会把两侧一起等比拖慢，出一份看似自洽实则作废的数据。多卡机器用 `CUDA_VISIBLE_DEVICES=<空闲卡号>` 选卡。**测量始终单进程单卡串行**，这是计时可比的前提；多卡仅用于可选的 sweep 加速（见「加速：多卡并行 sweep」）。
 脚本内固定了 `--level core --mode kernel`，其余可调项都走环境变量（见下）；需要改动这两个口径本身时直接编辑 `run_pytest.sh` 中的 pytest 命令行。
@@ -50,14 +45,10 @@ softmax:
 | 变量 | 默认 | 说明 |
 |-----|------|------|
-| `OP` | `fused_marlin_moe_mxfp4` | 测试函数名（`test_` 之后的部分）；设了 `TEST_NODE` 时由节点覆盖 |
+| `OP` | `fused_marlin_moe_mxfp4` | 测试函数名（`test_` 之后的部分） |
-| `OP_FILE` | `fused_marlin_moe` | benchmark 文件名（`test_` 与 `.py` 之间的部分）；设了 `TEST_NODE` 时由节点覆盖 |
+| `OP_FILE` | `fused_marlin_moe` | benchmark 文件名（`test_` 与 `.py` 之间的部分） |
-| `TEST_NODE` | 空 | FlagGems pytest 节点（`…/benchmark/test_<op_file>.py::test_<op>`）。设了且未指定 `SHAPE_FILE` 时默认不传 `--shape_file` |
+| `SHAPE_FILE` | 空（用脚本内置 yaml） | shape yaml 路径 |
-| `SHAPE_FILE` | 空 | shape yaml 路径；空且无 `TEST_NODE` 时用脚本内置 yaml；`-`/`upstream`/`none` 表示沿用上游在 `BENCH_LEVEL` 下的原生 shape 集 |
+| `FLAGGEMS_DIR` | `/workspace/dev/FlagGems` | FlagGems 仓库路径（测 vllm 仓库时指向 `/workspace/dev/FlagGems-vllm`） |
 | `SHAPE_MODE` | 空 | `inline` / `file` / `upstream`；通常不用手设，`TEST_NODE` 会把空 `SHAPE_FILE` 推成 `upstream` |
 | `FLAGGEMS_DIR` | `/workspace/dev/FlagGems` | FlagGems 仓库路径。测试文件缺失时回退 `/workspace/FlagGems`；`TEST_NODE` 也可直接带出仓库根 |
 | `ZL_BACKEND` | `cuda` | 设备后端：`cuda` / `musa` / `rocm` / `ascend`（别名 `nvidia`/`amd`/`npu` 等；未知名直接失败） |
 | `ZL_DEVICE` | 空 | 父进程看到的设备 id，写入 `run_identity.json`；批量层绑卡后也会设 |
 **怎么测**
@@ -65,14 +56,8 @@ softmax:
 |-----|------|------|
 | `REPLAY_FROM` | 空（record 模式） | 指向某次历史 run 目录，replay 其 autotune 选择。做 A/B 时 B 侧必须设（见「A/B 对比测试标准流程」） |
 | `USE_FLAGTUNE` | `0` | `0` 走普通 autotune（跳过搜索、快速验证）；`1` 用 FlagTune 扩展调优空间（首跑全量搜索、慢）。A/B 两侧须取同值 |
 | `BENCH_LEVEL` | `core` | 透传上游 `--level`。`comprehensive` 相对 `core` 做两件事：① 在 `set_shapes()` 里并入子类 `set_more_shapes()`（`mm` 因此从 5 个 shape 变 29 个，含 N/K 到 152064 的大 K 与 `models_shapes/*.yaml` 的模型 shape）；② 在 `get_input_iter()` 里对每个 shape 再产出一次变体（BLAS 是转置侧），用例数翻倍。注意 ① 只在 `SHAPE_MODE=upstream` 下可见——传了 `SHAPE_FILE` 时 `_shape_inject_plugin` 会在上游合并之后再覆盖一次、把 shape 集拉回 yaml 内容（BLAS 那类 arity 不一致的情形除外，插件会退让）；② 不经插件，任何 shape 来源下都生效。sweep 成本同步放大，shape 多时配合 `PARALLEL_WARMUP_GPUS`。A/B 两侧须取同值 |
 | `DTYPES` | 空（上游默认 dtype 扫描） | 空格分隔的 dtype 白名单（如 `"bfloat16 float16"`），逐个转为上游 `--dtypes`。算子不支持时上游报 `can't be supported by this op` |
 | `PARALLEL_WARMUP_GPUS` | 空（串行） | 设为 `N`（≥2）时把 autotune sweep 分片到 N 卡并行，测量仍单卡串行；不设或 `<2` 则完全串行（见「加速：多卡并行 sweep」） |
 | `ZL_ENABLE_CUDAGRAPH` | `auto` | `auto`/`1` 仅在 CUDA 上走 Graph 计时；非 CUDA 即使设 `1` 也拒绝并保持 `timer=direct` |
 | `ZL_ENABLE_AUTOTUNE` | `1` | `0` 关闭 autotune record/replay 插件 |
 | `ZL_ENABLE_IR` | `1` | `0` 关闭 TTGIR / 编译元数据插件 |
 | `ZL_ENABLE_MM_CLUSTER` | `0` | `1` 且后端支持时加载 Hopper mm cluster 规避 |
 | `ZL_EXTRA_PLUGINS` | 空 | 空格分隔、追加到 `-p` 列表的插件名 |
 **输出**
@@ -87,11 +72,9 @@ softmax:
 ```
 runs/softmax_20260716-031752/
 ├── run.log                  # 完整日志（含 SUCCESS 行的 latency/speedup 表；已去 ANSI 色的纯文本）
 ├── run_identity.json        # provider/backend/device/timer/vendor（插件合并写入）
 ├── shapes.yaml              # 本次实际使用的 shape（存档）
 ├── autotune_records/        # Triton autotune 选中的 config（供 replay）
-│   ├── softmax.json
+│   └── softmax.json
 │   └── meta.json            # provider/backend；跨后端 replay 会被拒绝
 └── ttgir/                   # 按 shape 分组的 IR 落盘
    ├── 1024x1024/
    │   └── softmax_kernel_inner/
@@ -169,7 +152,7 @@ shape 多时墙上时间几乎全在 autotune sweep，而不在测量——被
 PARALLEL_WARMUP_GPUS=8 SHAPE_FILE=shape/DeepSeek-V4-Flash-p32768d1024.yaml bash run_pytest.sh
 ```
-加速比取决于 sweep 占比：shape 多、config 空间大的算子收益最明显，可达数倍。不设该变量（或设 <2）时插件静默 no-op。**产物结构、`REPLAY_FROM` 用法、latency 表格式全不变**——合并后的 config 就写进本次 run 的 `autotune_records/<op>.json`，分片临时目录跑完即删。
+加速比取决于 sweep 占比：shape 多、config 空间大的算子收益最明显，可达数倍。不设该变量（或设 <2）时插件静默 no-op，常驻 `PLUGINS` 数组即可。**产物结构、`REPLAY_FROM` 用法、latency 表格式全不变**——合并后的 config 就写进本次 run 的 `autotune_records/<op>.json`，分片临时目录跑完即删。
 **测量本身不并行**：N 个进程压满同机多卡会通过功耗墙/散热耦合，单卡 latency 被邻居拖慢且不可复现。并行只用于"决定哪个 config 胜出"，测量仍是单进程单卡、与不开插件走同一条路径。
@@ -179,17 +162,7 @@ PARALLEL_WARMUP_GPUS=8 SHAPE_FILE=shape/DeepSeek-V4-Flash-p32768d1024.yaml bash
 WARNING 1 config key(s) got different winners across shards — parallel interference reached the sweep
 ```
-**冲突数就是这次加速的可信度指标**：0 可放心用；偏多说明 config 选择已被污染，出正式结论前不设该变量重跑一遍。
+**冲突数就是这次加速的可信度指标**：0 可放心用；偏多说明 config 选择已被污染，出正式结论前不设该变量重跑一遍。其余行为（`REPLAY_FROM` 已设或可见卡不足 2 张时跳过、分片失败回退串行、轮转分片而非按块切、用子进程而非 xdist 的原因）见插件 docstring。分片绑卡遵守继承来的 `CUDA_VISIBLE_DEVICES`——用它选过卡时，分片只会落在你选的那几张上。
 插件会在三种情况下**主动放弃分片、退回串行 sweep**，日志各有独立措辞，都不是静默的：
 - **`BENCH_LEVEL` 不是 `core`**。`comprehensive` 会让上游在 `set_shapes()` 里并入 `set_more_shapes()`，那部分是各子类自己的 Python（硬编码元组加 `models_shapes/*.yaml`），插件只看得到 yaml 里的 shape 列表。按子集分片会让未覆盖的 key 全部现场自调优，而这一轮却自称在 replay——`mm` 上实测是 29 个 shape 只扫到 5 个、90 次 `AUTOTUNE_REPLAY_FALLBACK reason=key_missing`。
 - **任一分片非零退出**。死掉的分片的 shape 从未被扫过，合并出来的 config 集是带洞的，publish 之后每个洞都会静默自调优。因此整轮 sweep 直接作废，worker 日志保留在临时目录里备查。
 - yaml 里查不到该算子的 shape（见下）、`REPLAY_FROM` 已设、或可见卡不足 2 张。
 yaml key 的查找与上游对齐：先按 `op_name`，查不到再沿 Benchmark 类链逐级尝试类名（`mm` 因此命中 `BlasBenchmark`）。类链是静态解析测试模块里 `op_name=` 那次构造调用得到的，不 import FlagGems。日志会打出命中的 key；一个都没命中时会列出尝试过的候选。
 其余行为（轮转分片而非按块切、用子进程而非 xdist 的原因）见插件 docstring。分片绑卡遵守当前后端的可见性变量（CUDA 为 `CUDA_VISIBLE_DEVICES`）——用它选过卡时，分片只会落在你选的那几张上。
 ## 批量多算子测试（run_batch.py）
@@ -197,8 +170,7 @@ yaml key 的查找与上游对齐：先按 `op_name`，查不到再沿 Benchmark
 ```bash
 # 全量筛查（双仓库全部可收集算子，排除带无条件 skip 标记的与 ops/dismiss.txt 里的）
-python run_batch.py --all --devices 0,1,2,3,4,5,6,7 --backend cuda --dtypes bfloat16 --op-timeout 1800
+python run_batch.py --all --gpus 0,1,2,3,4,5,6,7 --dtypes bfloat16 --op-timeout 1800
 # --gpus 仍是 --devices 的别名；两者同时给出会报错
 # 指定子集：默认 main 仓，vllm 仓加前缀；同名测试函数用 @文件名 消歧
 python run_batch.py --ops softmax vllm:fused_marlin_moe nextafter_@nextafter_
@@ -223,7 +195,7 @@ python run_batch.py --ops-file my_ops.txt --replay-root runs/batch_xxx
 **执行语义**：每算子独立子进程（CUDA crash 只废单个算子）；`--op-timeout` 对进程组 SIGTERM→SIGKILL，记 `TIMEOUT` 不重试、批次继续；pytest rc=1 与信号杀最多重试 2 次，连续同 rc 视为确定性失败提前止损；`--dtypes` 有两级降级——遇上游 `can't be supported by this op` 即去掉限制重跑（备注 `dtype_fallback`），其余失败且无成功行时也去掉限制最后救一次（备注 `dtype_rescue`，覆盖 torch baseline 对受限 dtype 编译失败的场景）。runtime skipif 与上游 `xfail` 都判 SKIP（上游声明的无信号状态）；`.complete` 标记只给 PASS/SKIP，`--resume` 据此复用。
-**产物**：`<batch>/<repo>/<op>/` 与单算子 run 目录结构完全一致（run.log、run_identity.json、autotune_records/、ttgir/……），所以任何一个算子都可以事后单独 `REPLAY_FROM` 复测。批量层额外产出 `ops_status.csv`（每算子一行：backend/device/timer、状态/耗时/行数/备注，边跑边原子更新；`gpu` 列与 `device` 同值）与 `summary.csv`（每测量行一行：backend/device/timer、dtype、latency、speedup、Size Detail，来源是 run.log 的结果表，`batch_summary.py` 也可单独对旧批次重跑）。备注列聚合 `no_cudagraph_rows`/`replay_fallback` 计数，解读口径时先看这列。不同后端、不同 timer 的行不要当成同口径。
+**产物**：`<batch>/<repo>/<op>/` 与单算子 run 目录结构完全一致（run.log、autotune_records/、ttgir/……），所以任何一个算子都可以事后单独 `REPLAY_FROM` 复测。批量层额外产出 `ops_status.csv`（每算子一行：状态/耗时/行数/备注，边跑边原子更新）与 `summary.csv`（每测量行一行：dtype、latency、speedup、Size Detail，来源是 run.log 的结果表，`batch_summary.py` 也可单独对旧批次重跑）。备注列聚合 `no_cudagraph_rows`/`replay_fallback` 计数，解读口径时先看这列。
 ## 计时口径：cudagraph 的预热、回退与精度
@@ -236,28 +208,27 @@ python run_batch.py --ops-file my_ops.txt --replay-root runs/batch_xxx
 回退本身不影响 A/B 公平性（两侧同一算子回退行为一致），但**回退口径的测量误差更大**：do_bench 每次迭代都走完整的 Python → launch 路径，kernel 越小，launch 开销和 CPU 侧抖动在数字里占比越高——亚毫秒级 kernel 上两种口径可差 2 倍以上，且行间波动更明显。解读这类算子的结果时：
 - 小 shape 行的绝对值和小幅（<10%）差异不要过度解读，优先看大 shape 行；
- 需要更高置信度时，同一配置多跑几次取中位数，或设 `ZL_ENABLE_CUDAGRAPH=0` 统一用 `timer=direct`（消除同表混两种口径的问题）。
+- 需要更高置信度时，同一配置多跑几次取中位数，或对该算子直接注释掉 `_cudagraph_plugin` 统一用 do_bench 口径（消除同表混两种口径的问题）。
 回退是按测量点发生的，同一份 latency 表里可能混有两种口径的行；若不确定，先 `grep BENCHMARK_DIRECT_NO_CUDAGRAPH run.log` 确认哪些行是回退口径再下结论。
 ## 插件说明
-`-p` 列表由 `plugin_plan.py` 按 `ZL_BACKEND` 生成，不再在 `run_pytest.sh` 里写死。`_cudagraph_plugin` **不**出现在默认列表上：只有 CUDA 且 `_timing_plugin` 选中 Graph 时才会 import 它。
+脚本通过 `-p` 加载以下插件（`run_pytest.sh` 的 `PLUGINS` 数组，可按需注释）：
 | 插件 | 作用 | 何时关闭 |
 |-----|------|---------|
-| `_device_setup_plugin` | 按后端设 `GEMS_VENDOR`（cuda→nvidia），跳过 flag_gems 启动时可能挂死的 `nvidia-smi` 探测，并写入 `run_identity.json`。旧名 `_device_guard_plugin` 仍是兼容 shim | 已设 `GEMS_VENDOR` 时不覆盖 |
+| `_device_guard_plugin` | 在 import flag_gems 之前直接用 torch 探测到 NVIDIA 卡就设 `GEMS_VENDOR=nvidia`，跳过 flag_gems 启动时 `nvidia-smi` 子进程探测（该探测在部分 fork 环境下会挂在 `wait4` 上导致 import 卡死）。副作用：import 期即初始化 CUDA context，故与 pytest-xdist 不兼容（本框架不用 xdist） | 一般无需关：已设 `GEMS_VENDOR`/`FLAGGEMS_VENDOR` 等 env 时自动跳过，非 NVIDIA 卡上自动 no-op |
 | `_parallel_warmup_plugin` | 由 `PARALLEL_WARMUP_GPUS=N` 激活（未设则静默 no-op）：autotune sweep 分片到 N 卡并行，测量仍单卡串行、产物结构不变（见「加速：多卡并行 sweep」） | 不设该变量即关闭 |
 | `_seed_plugin` | 固定 random/numpy/torch 种子，数据相关算子（sort/topk 等）输入逐字节一致 | 不关 |
 | `_shape_inject_plugin` | 让 shape yaml 覆盖子类硬编码的 `set_shapes()` | 不关 |
 | `_shape_iter_inject_plugin` | 覆盖在 `get_input_iter` 里硬编码 shape 的类（conv/pool 等） | 不关 |
 | `_bespoke_shape_plugin` | 覆盖特殊输入构造的算子（upsample/flash_mla/cutlass 等，按类名注册） | 测这些算子之外可关 |
-| `_autotune_record_plugin` | record/replay Triton autotune 选择；`autotune_records/meta.json` 记录 backend，跨后端 replay 拒绝 | `ZL_ENABLE_AUTOTUNE=0` |
+| `_autotune_record_plugin` | record/replay Triton autotune 选择（由 RECORD/REPLAY 环境变量二选一激活） | 不关 |
-| `_timing_plugin` | 按后端选择计时路径：CUDA 可走 Graph（内部调用 `_cudagraph_plugin`），其余后端一开始就是上游 `do_bench`（`timer=direct`） | `ZL_ENABLE_CUDAGRAPH=0` 强制 direct |
+| `_cudagraph_plugin` | `do_bench` → `do_bench_cudagraph`（kernel 纯耗时；内部先做跨流同步，修过一个间歇性 illegal instruction）。**capture 前按 warmup 时间预算显式预热**（先一次丢弃跑吸收 autotune/JIT 编译，再稳态预热），消除首轮抖动。无法 graph capture 的 kernel 自动回退并打 `BENCHMARK_DIRECT_NO_CUDAGRAPH` 标记 | 需要与他人的普通 do_bench 数据对齐时注释掉 |
-| `_cudagraph_plugin` | 仅由 `_timing_plugin` 在 CUDA 上调用。`do_bench` → `do_bench_cudagraph`，capture 前按 warmup 预算显式预热；无法 capture 时回退并打 `BENCHMARK_DIRECT_NO_CUDAGRAPH` | 不要直接 `-p` 加到非 CUDA 运行 |
+| `_pretty_report_plugin` | 结果表整理 + 着色：全表相同的输入折叠成表头下一行图例，每行 Size Detail 只留随行变化的部分（shape 保持 `torch.Size([...])` 原样，MoE 类算子单行从 ~400 字符缩到一屏内）；SUCCESS 绿 / FAILED 红。列名与 `SUCCESS` 字样保持上游原文，`run.log` 的 grep/解析不受影响 | 需要与上游原始表格逐字对齐时注释掉 |
-| `_pretty_report_plugin` | 结果表整理 + 着色：全表相同的输入折叠成表头下一行图例，每行 Size Detail 只留随行变化的部分（shape 保持 `torch.Size([...])` 原样，MoE 类算子单行从 ~400 字符缩到一屏内）；SUCCESS 绿 / FAILED 红。列名与 `SUCCESS` 字样保持上游原文，`run.log` 的 grep/解析不受影响 | 需要与上游原始表格逐字对齐时去掉 |
+| `_ir_meta_plugin` | 编译期记录每个变体的 constexpr/签名/特化；launch 钩子统计每个 (kernel, shape) 的实际使用；退出时按上述结构 dump ttgir | 不关 |
-| `_ir_meta_plugin` | 编译期记录每个变体的 constexpr/签名/特化；launch 钩子统计每个 (kernel, shape) 的实际使用；退出时按上述结构 dump ttgir。Triton knobs 缺失时 fail-soft | `ZL_ENABLE_IR=0` |
+| `_mm_cluster_fix_plugin` | Hopper fp16 mm cluster kernel 越界崩溃的运行时规避 | 默认注释；测 fp16 mm 崩溃时打开 |
 | `_mm_cluster_fix_plugin` | Hopper fp16 mm cluster kernel 越界崩溃的运行时规避 | 默认关；`ZL_ENABLE_MM_CLUSTER=1` 或 `ZL_EXTRA_PLUGINS` |
 ## 常见问题
@@ -286,7 +257,5 @@ shape 多时这部分会主导墙上时间，可设 `PARALLEL_WARMUP_GPUS=N` 并
 ## 依赖假设
 - FlagGems benchmark 体系（`benchmark/base.py` 的 `Benchmark` 类、conftest 的 `--shape_file/--level/--mode` 选项）；主仓与 FlagGems-vllm 仓库同构，`_autotune_record_plugin` 会按实际被 import 的包（`flag_gems` / `flaggems_vllm`）挂 LibTuner 补丁；
- 设备后端由 `ZL_BACKEND` / `--backend` 选择（`cuda` / `musa` / `rocm` / `ascend`）。inventory 仍是 FlagGems 逻辑算子清单，同一 case 可在多个后端上跑，不要按硬件复制 CSV；
+- Triton 需支持 `knobs.compilation.listener`、`kernel_load_end_hook`、`launch_enter_hook`（当前 FlagTree 的 triton 3.6 满足）；
 - Triton 需支持 `knobs.compilation.listener`、`kernel_load_end_hook`、`launch_enter_hook`（当前 FlagTree 的 triton 3.6 满足）。缺 hook 时 IR / autotune 插件 fail-soft，不会把进程打挂；
 - 插件通过 monkeypatch 挂钩上游内部结构，FlagGems/Triton 大版本升级后若行为异常，优先检查各插件 pytest_configure 输出的注册日志是否还正常打印。
 - 无 GPU 的适配器自检：`python test_backends.py`。
@@ -214,57 +214,6 @@ def _replay_run(original):
    return runner
 def _write_record_meta(record_dir: str) -> None:
    """Sidecar next to <op>.json. Replay refuses a different backend.
    The config JSON itself stays `{kernel: {key: cfg}}` so existing REPLAY_FROM
    directories keep working.
    """
    try:
        from backends import get_backend
        backend_name = get_backend().name
    except Exception:
        backend_name = os.environ.get("ZL_BACKEND", "")
    meta = {
        "schema_version": 1,
        "provider": "flaggems",
        "backend": backend_name,
        "op": _op_name(),
    }
    path = Path(record_dir) / "meta.json"
    try:
        path.parent.mkdir(parents=True, exist_ok=True)
        path.write_text(json.dumps(meta, indent=2, ensure_ascii=False) + "\n")
    except Exception as exc:
        from _term_style import tag
        print(f"{tag('[autotune-record-plugin]')} meta.json write failed: {exc}",
              file=sys.stderr, flush=True)
 def _replay_backend_ok(replay_dir: str) -> bool:
    path = Path(replay_dir) / "meta.json"
    if not path.is_file():
        return True
    try:
        meta = json.loads(path.read_text())
    except Exception:
        return True
    if not isinstance(meta, dict):
        return True
    recorded = str(meta.get("backend") or "").strip()
    if not recorded:
        return True
    try:
        from backends import get_backend
        current = get_backend().name
    except Exception:
        current = os.environ.get("ZL_BACKEND", "").strip()
    if recorded != current:
        _emit_marker(f"backend_mismatch recorded={recorded} current={current}")
        return False
    return True
 def _load_replay_dir(replay_dir: str) -> int:
    path = _record_path(replay_dir)
    if not path.is_file():
@@ -307,48 +256,29 @@ def _dump_record(record_dir: str) -> None:
 def pytest_configure(config):
    global _wrap
    from _term_style import tag
    from _run_identity import update_identity
    record_dir = os.environ.get(_RECORD_DIR_ENV, "").strip()
    replay_dir = os.environ.get(_REPLAY_DIR_ENV, "").strip()
    if record_dir and replay_dir:
        # Mutually exclusive: recording while replaying is pointless. Surface loudly.
        print(f"{tag('[autotune-record-plugin]')} error: both RECORD_DIR and "
              "REPLAY_DIR set; ignoring both (no-op)", file=sys.stderr, flush=True)
        update_identity(autotune_mode="uncontrolled")
        return
    if not record_dir and not replay_dir:
        update_identity(autotune_mode="uncontrolled")
        return
    try:
    import triton.runtime.autotuner as _autotuner
    except Exception as exc:
        print(f"{tag('[autotune-record-plugin]')} Triton Autotuner unavailable "
              f"({exc}); autotune_mode=uncontrolled", file=sys.stderr, flush=True)
        update_identity(autotune_mode="uncontrolled")
        return
    if replay_dir and not _replay_backend_ok(replay_dir):
        print(f"{tag('[autotune-record-plugin]')} refusing replay across backends",
              file=sys.stderr, flush=True)
        update_identity(autotune_mode="replay_fallback")
        return
    _wrap = _record_run if record_dir else _replay_run
    _autotuner.Autotuner.run = _wrap(_autotuner.Autotuner.run)
    if record_dir:
        # atexit (not sessionfinish): persist whatever was recorded even if an op
        # crash kills the session; a later replay run falls back for missing keys.
        _write_record_meta(record_dir)
        atexit.register(_dump_record, record_dir)
        update_identity(autotune_mode="record")
        print(f"{tag('[autotune-record-plugin]')} recording autotune configs to "
              f"{_record_path(record_dir)} (Autotuner)",
              file=sys.stderr, flush=True)
    else:
        loaded = _load_replay_dir(replay_dir)
        update_identity(autotune_mode="replay")
        print(f"{tag('[autotune-record-plugin]')} replaying {loaded} recorded entries from "
              f"{_record_path(replay_dir)} (Autotuner)",
              file=sys.stderr, flush=True)
@@ -1,18 +1,25 @@
-"""pytest plugin: swap `triton.testing.do_bench` for cudagraph-based timing.
+"""pytest plugin: swap `triton.testing.do_bench` for graph-capture timing.
-Invoked by `_timing_plugin` when the active backend declares
+Loaded via `-p _cudagraph_plugin` in run_pytest.sh's PLUGINS list (comment that
-`supports_cuda_graph` (CUDA only). Do not put this module on the default `-p`
+line out to time with plain do_bench). Monkey-patches do_bench →
 list for MUSA/ROCm/Ascend: those backends must start on plain do_bench, not
 attempt capture-then-fallback. Monkey-patches do_bench →
 do_bench_cudagraph (kernel-only latency) at session configure, after Triton is
 loaded but before any benchmark calls it. Kernels that can't be graph-captured
 fall back to plain do_bench and print a BENCHMARK_DIRECT_NO_CUDAGRAPH marker
 into the log. Before capturing, fn is explicitly warmed up for the caller's
 warmup time budget (do_bench_cudagraph's own 5-iter warmup is too short to
 settle autotune/JIT for the MoE kernels, which makes the first captured
-measurement unstable run-to-run). The capture itself runs under a scratch CUDA
+measurement unstable run-to-run). The capture itself runs under a scratch RNG
-RNG state so that an aborted capture cannot poison the process-wide generator
+state so that an aborted capture cannot poison the process-wide generator
 (rationale: _capture_with_scratch_rng).
 Everything here goes through Triton's active device interface rather than
 torch.cuda. FlagTree's do_bench is already written that way, but its
 do_bench_cudagraph is not: it hardcodes torch.cuda.stream/Event/CUDAGraph, so
 on a non-CUDA backend the graph path raised and every measurement silently fell
 back to plain do_bench. On MThreads the interface is torch_musa, which supplies
 Stream, Event, synchronize, graph and MUSAGraph, so the same capture works once
 the calls are taken from the interface. CUDA still calls upstream's
 do_bench_cudagraph, so its numbers are unchanged.
 """
 from __future__ import annotations
@@ -26,8 +33,33 @@ import triton.testing as _tt
 _ORIGINAL_DO_BENCH = _tt.do_bench
-def _is_fatal_cuda_error(exc: Exception) -> bool:
+def _device_interface():
-    """True for CUDA context errors (illegal address, OOM, device-side assert):
+    """Triton's active device interface: torch.cuda on NVIDIA, torch_musa on MThreads."""
    from triton.runtime import driver
    return driver.active.get_device_interface()
 def _is_cuda(di) -> bool:
    import torch
    return di is torch.cuda
 def _graph_class(di):
    """The interface's graph type: CUDAGraph, MUSAGraph, or whatever it names."""
    for name in ("CUDAGraph", "MUSAGraph"):
        graph = getattr(di, name, None)
        if graph is not None:
            return graph
    for name in dir(di):
        if name.endswith("Graph"):
            return getattr(di, name)
    raise RuntimeError("device interface exposes no graph type")
 def _is_fatal_device_error(exc: Exception) -> bool:
    """True for device context errors (illegal address, OOM, device-side assert):
    the context is poisoned, so re-raise rather than fall back and record a bogus
    latency as success.
@@ -49,8 +81,12 @@ def _is_fatal_cuda_error(exc: Exception) -> bool:
    return any(m in msg for m in markers)
-def _warmup_before_capture(fn, warmup_ms, grad_to_none):
+# Retained under its historical name; the marker list is backend independent.
-    """Warm fn up before cudagraph capture.
+_is_fatal_cuda_error = _is_fatal_device_error
 def _warmup_before_capture(di, fn, warmup_ms, grad_to_none):
    """Warm fn up before graph capture.
    do_bench_cudagraph only warms up 5 iterations internally — too few for the
    MoE kernels, where the first calls trigger Triton autotune config
@@ -68,8 +104,6 @@ def _warmup_before_capture(fn, warmup_ms, grad_to_none):
    """
    import time
    import torch
    def _run():
        if grad_to_none is not None:
            for x in grad_to_none:
@@ -78,23 +112,23 @@ def _warmup_before_capture(fn, warmup_ms, grad_to_none):
    # First call absorbs autotune/JIT compilation (can take seconds); discard it.
    _run()
-    torch.cuda.synchronize()
+    di.synchronize()
    # Size the warmup loop from one timed steady-state run. dt_ms includes
    # launch overhead, so n errs low for tiny kernels; warmup needn't be exact.
    t0 = time.perf_counter()
    _run()
-    torch.cuda.synchronize()
+    di.synchronize()
    dt_ms = (time.perf_counter() - t0) * 1000.0
    budget = warmup_ms if (warmup_ms and warmup_ms > 0) else 25.0
    n = int(budget / dt_ms) if dt_ms > 0 else 25
    n = max(5, min(n, 200))
    for _ in range(n):
        _run()
-    torch.cuda.synchronize()
+    di.synchronize()
-def _capture_with_scratch_rng(call):
+def _capture_with_scratch_rng(di, call):
-    """Run a graph capture under a throwaway CUDA RNG state.
+    """Run a graph capture under a throwaway device RNG state.
    torch.cuda.graph's capture_begin unconditionally registers the default CUDA
    RNG generator and sets its capturing flag; when a capture aborts midway the
@@ -106,12 +140,21 @@ def _capture_with_scratch_rng(call):
    before capturing and restore the original state object afterwards, success
    or failure. The real generator never takes part in a capture, so poisoning
    can only land on the discarded scratch state, and _seed_plugin determinism
-    is left untouched."""
+    is left untouched.
    An interface that does not expose default_generators has no such registration
    to protect against (torch_musa is one), so the capture simply runs.
    """
    import torch
-    dev = torch.cuda.current_device()
+
-    gen = torch.cuda.default_generators[dev]
+    generators = getattr(di, "default_generators", None)
    if not generators:
        return call()
    gen = generators[di.current_device()]
    if not hasattr(gen, "graphsafe_get_state"):
        return call()
    saved = gen.graphsafe_get_state()
-    scratch = torch.Generator(device="cuda")
+    scratch = torch.Generator(device=torch.device(gen.device))
    scratch.manual_seed(gen.initial_seed())
    gen.graphsafe_set_state(scratch)
    try:
@@ -120,24 +163,85 @@ def _capture_with_scratch_rng(call):
        gen.graphsafe_set_state(saved)
 def _do_bench_graph(di, fn, rep, grad_to_none, quantiles, return_mode, n_retries=10):
    """do_bench_cudagraph, taken from the device interface instead of torch.cuda.
    A step-for-step port of upstream's implementation: estimate one iteration,
    unroll that many calls into a captured graph so host launch overhead drops
    out, then time independent replays. Upstream is used verbatim on CUDA; this
    exists because upstream names torch.cuda directly and so cannot run on any
    other backend.
    The replay timings are handed to _summarize_statistics as a plain list of
    floats, which is what both do_bench and upstream's do_bench_cudagraph pass:
    this fork's helper runs Python's statistics module over them, so wrapping
    them in a tensor makes "mean" raise and makes a quantile request return
    tensors that later fail to serialize.
    """
    with di.stream(di.Stream()):
        fn()
        if grad_to_none is not None:
            for x in grad_to_none:
                x.detach_()
                x.requires_grad_(True)
                x.grad = None
        start_event = di.Event(enable_timing=True)
        end_event = di.Event(enable_timing=True)
        start_event.record()
        for _ in range(5):
            fn()
        end_event.record()
        di.synchronize()
        estimate_ms = start_event.elapsed_time(end_event) / 5
        n_repeat = 1000 if estimate_ms == 0 else max(1, int(rep / estimate_ms))
        graph = _graph_class(di)()
        with di.graph(graph):
            for _ in range(n_repeat):
                if grad_to_none is not None:
                    for x in grad_to_none:
                        x.grad = None
                fn()
        di.synchronize()
        ret = []
        for _ in range(n_retries):
            start_event = di.Event(enable_timing=True)
            end_event = di.Event(enable_timing=True)
            start_event.record()
            graph.replay()
            end_event.record()
            di.synchronize()
            ret.append(start_event.elapsed_time(end_event) / n_repeat)
        del graph
        return _tt._summarize_statistics(ret, quantiles, return_mode)
 def _patched_do_bench(fn, warmup=25, rep=100, grad_to_none=None,
                      quantiles=None, return_mode="mean"):
-    """Drop-in replacement for triton.testing.do_bench using cudagraph capture."""
+    """Drop-in replacement for triton.testing.do_bench using graph capture."""
    op = os.environ.get("FLAGGEMS_PERF_CURRENT_OP", "<unknown>")
-    phase = "sync"  # which step rejected cudagraph, for the fallback marker
+    phase = "sync"  # which step rejected graph capture, for the fallback marker
    try:
-        # do_bench_cudagraph runs fn on a fresh side stream without syncing
+        di = _device_interface()
-        # with the default stream first; input tensors produced just before
+        # The graph path runs fn on a fresh side stream without syncing with the
-        # (quantize/topk in the input iter) may still be in flight, and racing
+        # default stream first; input tensors produced just before (quantize/topk
-        # on them intermittently kills the context (illegal instruction from a
+        # in the input iter) may still be in flight, and racing on them
-        # device trap on garbage indices). Sync before switching streams.
+        # intermittently kills the context (illegal instruction from a device
-        import torch
+        # trap on garbage indices). Sync before switching streams.
-        torch.cuda.synchronize()
+        di.synchronize()
        phase = "warmup"
        # Settle autotune/JIT before capture (rationale: _warmup_before_capture).
-        _warmup_before_capture(fn, warmup, grad_to_none)
+        _warmup_before_capture(di, fn, warmup, grad_to_none)
        phase = "capture"
-        return _capture_with_scratch_rng(lambda: _tt.do_bench_cudagraph(
+        if _is_cuda(di):
            return _capture_with_scratch_rng(di, lambda: _tt.do_bench_cudagraph(
                fn,
                rep=rep,
                grad_to_none=grad_to_none,
                quantiles=quantiles,
                return_mode=return_mode,
            ))
        return _capture_with_scratch_rng(di, lambda: _do_bench_graph(
            di,
            fn,
            rep=rep,
            grad_to_none=grad_to_none,
@@ -147,7 +251,7 @@ def _patched_do_bench(fn, warmup=25, rep=100, grad_to_none=None,
    except Exception as exc:
        # Re-raise genuine device failures; anything else falls back to plain
        # do_bench, with the failing phase (sync/warmup/capture) in the marker.
-        if _is_fatal_cuda_error(exc):
+        if _is_fatal_device_error(exc):
            raise
        reason = type(exc).__name__
        detail = " ".join(str(exc).split())[:160]
@@ -169,17 +273,11 @@ def _patched_do_bench(fn, warmup=25, rep=100, grad_to_none=None,
 def pytest_configure(config):
    from _term_style import tag
    try:
        from backends import get_backend
        if not get_backend().supports_cuda_graph:
            print(f"{tag('[cudagraph-plugin]')} skipped: backend "
                  f"{get_backend().name!r} does not support CUDA Graph",
                  file=sys.stderr, flush=True)
            return
    except Exception as exc:
        print(f"{tag('[cudagraph-plugin]')} backend probe failed ({exc}); "
              "not patching do_bench", file=sys.stderr, flush=True)
        return
    _tt.do_bench = _patched_do_bench
-    print(f"{tag('[cudagraph-plugin]')} triton.testing.do_bench → do_bench_cudagraph",
+    try:
        backend = getattr(_device_interface(), "__name__", "?")
    except Exception:  # pragma: no cover - reported, never fatal at configure
        backend = "?"
    print(f"{tag('[cudagraph-plugin]')} triton.testing.do_bench → graph capture "
          f"via {backend}",
          file=sys.stderr, flush=True)
@@ -1,7 +1,45 @@
-"""Backward-compatible alias of `_device_setup_plugin`.
+"""Guard against FlagGems' device-detection hang (benchmark-side fix).
-Kept so an old `-p _device_guard_plugin` line still sets GEMS_VENDOR. New
+FlagGems' `runtime.backend.device_finder` falls back to
-plugin lists should load `_device_setup_plugin` instead.
+`subprocess.run(nvidia-smi)` **without a timeout**. Under this conda env's
 forked/inconsistent subprocess (`_posixsubprocess` symbol mismatch), that
 child can hang indefinitely, leaving `import flag_gems` stuck in `wait4`
 (seen as run_pytest hanging with no output, GPU at 0%).
 This plugin runs at import time — before any test module imports flag_gems —
 detects the vendor via torch (no subprocess), and sets GEMS_VENDOR so
 device_finder takes its env fast-path (`_get_vendor_from_env`) and never
 reaches the hang-prone subprocess probe. No change to the FlagGems repo.
 Side effect: get_device_properties initializes the CUDA context very early in
 the pytest process. Fine for the current single-process runs; revisit if
 fork-based parallelism (e.g. pytest-xdist) is ever introduced.
 """
-from _device_setup_plugin import *  # noqa: F401,F403
+import os
-from _device_setup_plugin import pytest_configure  # noqa: F401
+import sys
 _VENDOR_ENV_KEYS = ("GEMS_VENDOR", "FLAGGEMS_VENDOR", "GEMS_BACKEND", "FLAGGEMS_BACKEND")
 def _guard_device_vendor():
    # Respect an explicit choice if the user already set one.
    if any(k in os.environ for k in _VENDOR_ENV_KEYS):
        return
    try:
        import torch
        if torch.cuda.is_available():
            name = torch.cuda.get_device_properties(0).name.upper()
            if "NVIDIA" in name:
                os.environ["GEMS_VENDOR"] = "nvidia"
                from _term_style import tag
                print(f"{tag('[device-guard-plugin]')} set GEMS_VENDOR=nvidia "
                      "(skip flag_gems nvidia-smi subprocess probe, avoids import hang)",
                      file=sys.stderr, flush=True)
    except Exception as e:  # pragma: no cover
        print(f"[device-guard-plugin] torch vendor probe failed ({e}); "
              "leaving detection to flag_gems",
              file=sys.stderr, flush=True)
 _guard_device_vendor()
@@ -1,69 +0,0 @@
 """Set FlagGems vendor + record backend identity before flag_gems is imported.
 Replaces the NVIDIA-only `_device_guard_plugin`. Setting GEMS_VENDOR from the
 zl_bench backend name skips FlagGems' hang-prone nvidia-smi subprocess probe
 on CUDA, and maps musa/rocm/ascend onto the matching FlagGems vendor.
 Import-time side effect is intentional: FlagGems reads GEMS_VENDOR while the
 benchmark module is being collected. Device name / arch are filled later in
 pytest_configure so import does not initialize a device context.
 """
 from __future__ import annotations
 import os
 import sys
 from backends import get_backend
 from providers.flaggems import apply_vendor_env
 from _run_identity import update_identity
 def _setup_at_import() -> None:
    try:
        backend = get_backend()
    except Exception as exc:
        print(f"[device-setup-plugin] cannot resolve backend ({exc}); "
              "leaving GEMS_VENDOR to flag_gems",
              file=sys.stderr, flush=True)
        return
    vendor = apply_vendor_env(backend)
    os.environ["ZL_BACKEND"] = backend.name
    device = os.environ.get("ZL_DEVICE", "").strip()
    if not device:
        for key in backend.visibility_env_keys:
            raw = os.environ.get(key, "").strip()
            if raw:
                device = raw.split(",")[0].strip()
                break
    update_identity(
        backend=backend.name,
        device=device,
        flaggems_vendor=vendor or "",
        visibility_env=backend.visibility_env_keys[0] if backend.visibility_env_keys else "",
    )
    if vendor:
        from _term_style import tag
        print(f"{tag('[device-setup-plugin]')} backend={backend.name} "
              f"GEMS_VENDOR={vendor}",
              file=sys.stderr, flush=True)
 def pytest_configure(config):
    try:
        backend = get_backend()
        ident = backend.device_identity()
        update_identity(
            backend=backend.name,
            device_name=ident.get("device_name", ""),
            device_arch=ident.get("device_arch", ""),
            timer_api=ident.get("timer_api", ""),
            flaggems_vendor=ident.get("flaggems_vendor", ""),
        )
    except Exception as exc:
        from _term_style import tag
        print(f"{tag('[device-setup-plugin]')} identity probe failed ({exc})",
              file=sys.stderr, flush=True)
 _setup_at_import()
@@ -402,33 +402,17 @@ def _dump_ttgir(cache_dir: str, dump_dir: str) -> None:
 def pytest_configure(config):
    global _dump_enabled
    from _term_style import tag
    try:
    import triton
    import triton.runtime.autotuner as _autotuner
    except Exception as exc:
        print(f"{tag('[ir-meta-plugin]')} Triton knobs unavailable ({exc}); "
              "IR collection disabled", file=sys.stderr, flush=True)
        return
    if not hasattr(triton, "knobs"):
        print(f"{tag('[ir-meta-plugin]')} triton.knobs missing; "
              "IR collection disabled", file=sys.stderr, flush=True)
        return
-    try:
+    prev = triton.knobs.compilation.listener
        compilation = triton.knobs.compilation
        prev = compilation.listener
    except Exception as exc:
        print(f"{tag('[ir-meta-plugin]')} compilation listener unavailable "
              f"({exc}); IR collection disabled", file=sys.stderr, flush=True)
        return
    if prev is None:
-        compilation.listener = _compile_listener
+        triton.knobs.compilation.listener = _compile_listener
    else:
        def chained(**kwargs):
            prev(**kwargs)
            _compile_listener(**kwargs)
-        compilation.listener = chained
+        triton.knobs.compilation.listener = chained
    dump_dir = os.environ.get(_DUMP_DIR_ENV, "").strip()
    if dump_dir:
@@ -105,18 +105,6 @@ def _warmup_load_backend():
 def pytest_configure(config):
    from _term_style import tag
    try:
        from backends import get_backend
        if not get_backend().supports_mm_cluster_fix:
            print(f"{tag('[mm-cluster-fix-plugin]')} skipped: backend "
                  f"{get_backend().name!r} is not the CUDA Hopper path",
                  file=sys.stderr, flush=True)
            return
    except Exception as exc:
        print(f"{tag('[mm-cluster-fix-plugin]')} backend probe failed ({exc}); "
              "leaving cluster_remote_mm_scenario unpatched",
              file=sys.stderr, flush=True)
        return
    _patch_loaded()
    _warmup_load_backend()
    _patch_loaded()
@@ -0,0 +1,89 @@
 """Narrow compatibility fixes for the frozen MUSA product corpus.
 Several upstream FlagGems tests gate themselves on ``torch.cuda.is_available()``
 even though the code under test runs on any accelerator FlagGems dispatches to.
 On MThreads those guards skip the test instead of exercising the operator, which
 turns a working path into a verdict of "not measured".
 Every override here is pinned to one file and one exact skip reason, applies
 only when FlagGems has dispatched to the MThreads vendor, and is only added
 after the unskipped test was observed to pass on the device.
 Two files were probed on S5000 and deliberately left out, because their guards
 are not stale:
 - ``test_cutlass_scaled_mm.py`` gates on vLLM as well as CUDA, so removing the
  guard would raise on the missing dependency instead of measuring anything.
 - ``test_grid_sample.py`` builds its inputs through a helper whose device
  argument defaults to the literal ``"cuda"``. Its guard is consistent with its
  own body: unskipped, all 90 collected cases fail in ``aten::empty`` on a CUDA
  backend that does not exist here. Making that file portable means editing
  upstream test source, which is outside what a compatibility plugin may do.
 """
 from __future__ import annotations
 import sys
 from pathlib import Path
 # file name -> exact skipif reasons that are stale on a dispatching backend.
 _STALE_CUDA_GUARDS: dict[str, frozenset[str]] = {
    "test_fft.py": frozenset({"CUDA is not available"}),
    "test_top_k_per_row_decode.py": frozenset({"CUDA device required"}),
 }
 # Guards that are only stale for one test in the file stay pinned to it.
 _PINNED_TESTS: dict[str, str] = {
    "test_fft.py": "test_perf_fft",
 }
 def _is_stale_cuda_guard(item, mark) -> bool:
    if mark.name != "skipif":
        return False
    name = Path(str(item.path)).name
    reasons = _STALE_CUDA_GUARDS.get(name)
    if reasons is None or mark.kwargs.get("reason") not in reasons:
        return False
    pinned = _PINNED_TESTS.get(name)
    return pinned is None or item.name == pinned
 # Retained under its historical name: the sealed campaign's tests reference it.
 def _is_stale_fft_cuda_guard(item, mark) -> bool:
    return Path(str(item.path)).name == "test_fft.py" and _is_stale_cuda_guard(
        item, mark
    )
 def pytest_collection_modifyitems(items):
    import flag_gems
    if flag_gems.vendor_name != "mthreads":
        return
    removed: dict[str, int] = {}
    # A guard can sit on the test, its class, or the module's pytestmark, and
    # skipif is resolved by walking that whole chain. Strip every level.
    seen: set[int] = set()
    for item in items:
        for node in item.listchain():
            if id(node) in seen:
                continue
            seen.add(id(node))
            markers = getattr(node, "own_markers", None)
            if not markers:
                continue
            kept = [mark for mark in markers if not _is_stale_cuda_guard(item, mark)]
            dropped = len(markers) - len(kept)
            if dropped:
                markers[:] = kept
                name = Path(str(item.path)).name
                removed[name] = removed.get(name, 0) + dropped
    for name, count in sorted(removed.items()):
        print(
            f"[musa-benchmark-compat] removed {count} stale CUDA-only skip "
            f"guard(s) from {name}",
            file=sys.stderr,
            flush=True,
        )
@@ -33,7 +33,7 @@ import sys
 import tempfile
 import time
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Tuple
 import pytest
@@ -60,8 +60,8 @@ def _requested_gpus() -> int:
    if want < 2:
        return 0
    try:
-        from backends import get_backend
+        import torch
-        have = get_backend().probe_device_count()
+        have = torch.cuda.device_count()
    except Exception:
        return 0
    return max(0, min(want, have))
@@ -70,132 +70,30 @@ def _requested_gpus() -> int:
 def _visible_devices() -> List[str]:
    """The device ids a shard may be pinned to, in parent-visible order.
-    Must respect inherited visibility env for the *active* backend. Writing a
+    Must respect an inherited CUDA_VISIBLE_DEVICES: writing a bare shard index
-    bare shard index into the child would re-interpret it as an absolute id,
+    into the child would re-interpret it as an absolute id, so a caller who
-    so a caller who picked idle cards (CUDA_VISIBLE_DEVICES=5,6) would silently
+    picked idle cards (CUDA_VISIBLE_DEVICES=5,6) would silently get cards 0,1 —
-    get cards 0,1 — exactly the busy-GPU case the caller was avoiding.
+    exactly the busy-GPU case the caller was avoiding.
    """
-    from backends import get_backend
+    raw = os.environ.get("CUDA_VISIBLE_DEVICES", "").strip()
    backend = get_backend()
    for key in backend.visibility_env_keys:
        raw = os.environ.get(key, "").strip()
    if raw:
        return [d.strip() for d in raw.split(",") if d.strip()]
    try:
-        n = backend.probe_device_count()
+        import torch
-        return [str(i) for i in range(n)] if n else []
+        return [str(i) for i in range(torch.cuda.device_count())]
    except Exception:
        return []
-def _attr_name(node: Any) -> str:
+def _read_shape_yaml(path: str, op: str) -> Tuple[List[Any], Dict[str, Any]]:
-    """Last component of a Name/Attribute node: `base.BlasBenchmark` -> `BlasBenchmark`."""
+    """Shapes for this op, plus the sibling keys to carry into each shard."""
    import ast
    if isinstance(node, ast.Name):
        return node.id
    if isinstance(node, ast.Attribute):
        return node.attr
    return ""
 def _class_bases(py_path: Path) -> Dict[str, List[str]]:
    """{class name: [base class names]} parsed statically from one module."""
    import ast
    try:
        tree = ast.parse(py_path.read_text(errors="replace"))
    except Exception:
        return {}
    return {
        node.name: [n for n in (_attr_name(b) for b in node.bases) if n]
        for node in ast.walk(tree)
        if isinstance(node, ast.ClassDef)
    }
 def _class_chain(test_py: Path, op: str) -> List[str]:
    """Benchmark class names to try for `op`, nearest ancestor first.
    FlagGems' `Benchmark.set_shapes` looks up `op_name` and, failing that, walks
    `type(self).__mro__` trying each class name as a yaml key -- which is how
    BLAS ops reach the shared `BlasBenchmark` entry in core_shapes.yaml. This
    plugin runs at pytest_configure, before any Benchmark exists, so the real
    MRO is unavailable; the class is recovered statically instead, from the
    `op_name=` keyword of the constructor call inside the test module, then its
    bases from that module plus benchmark/base.py.
    Breadth-first over the inheritance graph, not a true C3 linearisation: the
    benchmark classes are single-inheritance in practice, and a miss only costs
    the sharding (see the caller's skip path), never correctness.
    """
    import ast
    try:
        tree = ast.parse(test_py.read_text(errors="replace"))
    except Exception:
        return []
    start = ""
    for node in ast.walk(tree):
        if not isinstance(node, ast.Call):
            continue
        for kw in node.keywords:
            if (kw.arg == "op_name" and isinstance(kw.value, ast.Constant)
                    and kw.value.value == op):
                start = _attr_name(node.func)
                break
        if start:
            break
    if not start:
        return []
    graph = _class_bases(test_py)
    for name, bases in _class_bases(test_py.parent / "base.py").items():
        graph.setdefault(name, bases)
    chain: List[str] = []
    seen: set = set()
    queue = [start]
    while queue:
        name = queue.pop(0)
        if not name or name in seen:
            continue
        seen.add(name)
        chain.append(name)
        queue.extend(graph.get(name, []))
    return chain
 def _test_module_path(argv: List[str]) -> Optional[Path]:
    """The benchmark module under test, from pytest's own argv."""
    for a in argv:
        if a.startswith("-"):
            continue
        p = Path(a.split("::", 1)[0])
        if p.suffix == ".py" and p.is_file():
            return p
    return None
 def _read_shape_yaml(path: str, op: str, test_py: Optional[Path] = None
                     ) -> Tuple[List[Any], Dict[str, Any], str]:
    """Shapes for this op, the sibling keys to carry into each shard, and the
    yaml key they came from ("" when nothing matched).
    Tries `op` first, then the benchmark class chain, mirroring FlagGems. Without
    the class fallback every BLAS op (mm/addmm/bmm/baddbmm, whose shapes live
    under `BlasBenchmark`) resolved to zero shapes and sharding silently turned
    itself off.
    """
    import yaml
    with open(path) as f:
        doc = yaml.safe_load(f) or {}
-    keys = [op]
+    entry = doc.get(op) or {}
-    if not isinstance(doc.get(op), dict) and test_py is not None:
+    shapes = entry.get("shapes") or []
        keys += _class_chain(test_py, op)
    for key in keys:
        entry = doc.get(key)
        if isinstance(entry, dict) and entry.get("shapes"):
    extra = {k: v for k, v in entry.items() if k != "shapes"}
-            return list(entry["shapes"]), extra, key
+    return list(shapes), extra
    return [], {}, ""
 def _shard(shapes: List[Any], n: int) -> List[List[Any]]:
@@ -213,8 +111,8 @@ def _spawn(shards: List[List[Any]], op: str, extra: Dict[str, Any],
           ) -> Tuple[List[Path], List[int]]:
    """One subprocess per shard, each pinned to its own GPU.
-    Subprocesses rather than pytest-xdist: `_device_setup_plugin` may initialize
+    Subprocesses rather than pytest-xdist: _device_guard_plugin initializes a
-    a device context at import time, which does not survive fork-based parallelism.
+    CUDA context at import time, which does not survive fork-based parallelism.
    """
    import yaml
    procs: List[Tuple[int, subprocess.Popen, Any]] = []
@@ -245,8 +143,7 @@ def _spawn(shards: List[List[Any]], op: str, extra: Dict[str, Any],
        child_argv += ["--shape_file", str(w / "shapes.yaml")]
        env = dict(os.environ)
-        from backends import get_backend
+        env["CUDA_VISIBLE_DEVICES"] = devices[i]
        env = get_backend().apply_visibility(env, devices[i])
        env[_WORKER_ENV] = "1"                    # stops recursion
        env[_RECORD_DIR_ENV] = str(rec)           # shard records its own picks
        env.pop(_REPLAY_DIR_ENV, None)            # a shard must sweep
@@ -325,23 +222,6 @@ def pytest_configure(config):
        return
    op = _op_name()
    # Only --level core is shardable. At comprehensive, FlagGems' set_shapes()
    # merges set_more_shapes() on top of whatever the yaml holds, and that
    # addition is arbitrary per-subclass Python (hardcoded tuples plus
    # models_shapes/*.yaml) which this plugin cannot see. Sharding the yaml list
    # alone would sweep a strict subset -- measured on mm: 5 of 29 shapes, 90
    # AUTOTUNE_REPLAY_FALLBACK reason=key_missing -- and every uncovered key
    # then self-tunes while the run reports itself as replaying. Refusing costs
    # wall time; sharding a subset costs the same-config premise.
    level = str(getattr(config.option, "level", "") or "").strip().lower()
    if level and level != "core":
        print(f"{tag('[parallel-warmup-plugin]')} --level {level} merges shapes "
              "this plugin cannot enumerate (set_more_shapes); sharding a subset "
              "would break the same-config premise, so sweeping serially instead",
              flush=True)
        return
    shape_file = getattr(config.option, "shape_file", "") or ""
    if not shape_file or not Path(shape_file).is_file():
        print(f"{tag('[parallel-warmup-plugin]')} no --shape_file; "
@@ -352,24 +232,15 @@ def pytest_configure(config):
              "already pinned, nothing to sweep, skipping", flush=True)
        return
    test_py = _test_module_path(list(config.invocation_params.args))
    try:
-        shapes, extra, key = _read_shape_yaml(shape_file, op, test_py)
+        shapes, extra = _read_shape_yaml(shape_file, op)
    except Exception as exc:
        print(f"{tag('[parallel-warmup-plugin]')} cannot read shapes "
              f"({type(exc).__name__}: {exc}); skipping", flush=True)
        return
    if not key:
        # Distinct from "found a short list": nothing in the yaml describes this
        # op at all, so the shard files would be built from an empty set.
        print(f"{tag('[parallel-warmup-plugin]')} no yaml entry for {op!r} "
              f"(tried: {', '.join([op] + _class_chain(test_py, op)) if test_py else op}); "
              "skipping", flush=True)
        return
    if len(shapes) < 2:
-        print(f"{tag('[parallel-warmup-plugin]')} only {len(shapes)} shape(s) "
+        print(f"{tag('[parallel-warmup-plugin]')} only {len(shapes)} shape(s); "
-              f"under yaml key {key!r}; sharding would not pay off, skipping",
+              "sharding would not pay off, skipping", flush=True)
              flush=True)
        return
    devices = _visible_devices()
@@ -380,9 +251,8 @@ def pytest_configure(config):
    shards = _shard(shapes, min(gpus, len(shapes), len(devices)))
    scratch = Path(tempfile.mkdtemp(prefix=f"warmup_{op}_"))
-    print(f"{tag('[parallel-warmup-plugin]')} sweeping {len(shapes)} shapes "
+    print(f"{tag('[parallel-warmup-plugin]')} sweeping {len(shapes)} shapes on "
-          f"(yaml key {key!r}) on {len(shards)} GPUs "
+          f"{len(shards)} GPUs (shards: {[len(s) for s in shards]}); this phase's "
          f"(shards: {[len(s) for s in shards]}); this phase's "
          "latency is discarded, only configs are kept", flush=True)
    keep_scratch = False
@@ -390,22 +260,8 @@ def pytest_configure(config):
    try:
        recs, bad = _spawn(shards, op, extra, scratch,
                           list(config.invocation_params.args), devices)
        elapsed = time.time() - t0
        if bad:
            # A dead shard means its shapes were never swept, so publishing the
            # merge would hand the measurement a config set with holes; every
            # hole then silently self-tunes under a run that claims to replay.
            # Discard the whole sweep rather than publish a partial one.
            keep_scratch = True
            print(f"{tag('[parallel-warmup-plugin]')} shard(s) {bad} exited "
                  f"non-zero after {elapsed:.0f}s; their shapes were never swept, "
                  "so the merged configs would have holes. Discarding this sweep "
                  f"and falling back to normal serial autotune (worker logs kept "
                  f"in {scratch})", flush=True)
            return
        merged, total, conflicts = _merge(recs, op)
        elapsed = time.time() - t0
        if not total:
            keep_scratch = True
@@ -431,6 +287,9 @@ def pytest_configure(config):
        msg = (f"{tag('[parallel-warmup-plugin]')} sweep done in {elapsed:.0f}s: "
               f"{len(merged)} kernels / {total} configs -> {out / f'{op}.json'}; "
               "measurement continues serially on one GPU")
        if bad:
            msg += f" (WARNING shard(s) {bad} exited non-zero; missing keys will "
            msg += "fall back to live autotune)"
        print(msg, flush=True)
        if conflicts:
            print(f"{tag('[parallel-warmup-plugin]')} WARNING {conflicts} config "
@@ -1,91 +0,0 @@
 """Merge-update `$OUT_DIR/run_identity.json` so plugins can record backend facts.
 Shell writes a stub before pytest; plugins add device_name / timer / vendor
 without clobbering earlier keys. Missing OUT_DIR is a silent no-op so unit
 tests and ad-hoc pytest invocations do not fail.
 """
 from __future__ import annotations
 import argparse
 import json
 import os
 from pathlib import Path
 from typing import Any, Dict, Optional
 IDENTITY_NAME = "run_identity.json"
 def identity_path(out_dir: Optional[str] = None) -> Optional[Path]:
    root = out_dir or os.environ.get("OUT_DIR", "").strip()
    if not root:
        return None
    return Path(root) / IDENTITY_NAME
 def load_identity(out_dir: Optional[str] = None) -> Dict[str, Any]:
    path = identity_path(out_dir)
    if path is None or not path.is_file():
        return {}
    try:
        data = json.loads(path.read_text(encoding="utf-8"))
    except Exception:
        return {}
    return data if isinstance(data, dict) else {}
 def update_identity(out_dir: Optional[str] = None, **fields: Any) -> Dict[str, Any]:
    path = identity_path(out_dir)
    if path is None:
        return dict(fields)
    path.parent.mkdir(parents=True, exist_ok=True)
    current = load_identity(out_dir)
    for key, value in fields.items():
        if value is None:
            continue
        current[key] = value
    tmp = path.with_name(path.name + ".tmp")
    tmp.write_text(json.dumps(current, indent=2, ensure_ascii=False) + "\n",
                   encoding="utf-8")
    os.replace(tmp, path)
    return current
 def init_identity(out_dir: str, **fields: Any) -> Dict[str, Any]:
    payload = {"schema_version": 1, "provider": "flaggems"}
    payload.update(fields)
    os.environ["OUT_DIR"] = str(out_dir)
    return update_identity(out_dir, **payload)
 def main(argv=None) -> int:
    p = argparse.ArgumentParser()
    p.add_argument("out_dir")
    p.add_argument("--backend", default=os.environ.get("ZL_BACKEND", "cuda"))
    p.add_argument("--op", default=os.environ.get("OP", ""))
    p.add_argument("--op-file", default=os.environ.get("OP_FILE", ""))
    p.add_argument("--flaggems-dir", default=os.environ.get("FLAGGEMS_DIR", ""))
    p.add_argument("--autotune-mode", default="")
    p.add_argument("--device", default=os.environ.get("ZL_DEVICE", ""))
    p.add_argument("--test-file", default=os.environ.get("TEST_FILE", ""))
    p.add_argument("--shape-mode", default=os.environ.get("SHAPE_MODE", ""))
    p.add_argument("--bench-level", default=os.environ.get("BENCH_LEVEL", "core"))
    args = p.parse_args(argv)
    init_identity(
        args.out_dir,
        backend=args.backend,
        op=args.op,
        op_file=args.op_file,
        flaggems_dir=args.flaggems_dir,
        autotune_mode=args.autotune_mode,
        device=args.device,
        test_file=args.test_file,
        shape_mode=args.shape_mode,
        bench_level=args.bench_level,
    )
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -1,47 +0,0 @@
 """Select the timing path from backend capabilities, before any shape runs.
 CUDA keeps the existing `_cudagraph_plugin` behaviour (graph capture, with
 per-shape fallback to plain do_bench). Other backends do not attempt CUDA
 Graph at all: they leave `triton.testing.do_bench` in place and record
 `timer=direct`. Trying capture-then-fallback on a non-CUDA device would
 produce FATAL_DEVICE noise and mixed-caliber rows.
 """
 from __future__ import annotations
 import os
 import sys
 from backends import get_backend
 from _run_identity import update_identity
 def _graph_requested(backend) -> bool:
    raw = os.environ.get("ZL_ENABLE_CUDAGRAPH", "auto").strip().lower()
    if raw in ("0", "false", "no", "off", "never", "direct"):
        return False
    if raw in ("1", "true", "yes", "on", "always"):
        return bool(backend.supports_cuda_graph)
    return bool(backend.supports_cuda_graph)
 def pytest_configure(config):
    from _term_style import tag
    backend = get_backend()
    raw = os.environ.get("ZL_ENABLE_CUDAGRAPH", "auto").strip().lower()
    force_on = raw in ("1", "true", "yes", "on", "always")
    if force_on and not backend.supports_cuda_graph:
        print(f"{tag('[timing-plugin]')} ZL_ENABLE_CUDAGRAPH=1 ignored on "
              f"backend={backend.name} (no CUDA Graph); using timer=direct",
              file=sys.stderr, flush=True)
    if _graph_requested(backend):
        import _cudagraph_plugin as _cg
        _cg.pytest_configure(config)
        update_identity(timer="cuda_graph", graph_capture="attempted")
        print(f"{tag('[timing-plugin]')} timer=cuda_graph backend={backend.name}",
              file=sys.stderr, flush=True)
        return
    update_identity(timer="direct", graph_capture="disabled")
    print(f"{tag('[timing-plugin]')} timer=direct backend={backend.name} "
          "(upstream do_bench; graph capture not selected)",
          file=sys.stderr, flush=True)
@@ -0,0 +1,68 @@
 #!/bin/bash
 # A/B 对比：FlagTree RemoveLayoutConversions 增强关 vs 开，逐算子两轮。
 #
 # 口径：每个算子先跑关（record autotune config），再用 REPLAY_FROM 回放同一套
 # config 跑开，两侧只差 FLAGTREE_<VENDOR>_RLC_ENHANCE 这一个开关。产物沿用
 # run_pytest.sh 的目录结构，所以每个算子恰好留下两份 run.log：
 #
 #   runs/<op>_<时间戳>/run.log   关（before）
 #   runs/<op>_<时间戳>/run.log   开（after）
 #
 # 跑完用 batch_summary.py --ab <before> <after> 得到该算子的 speedup 区间。
 #
 # 用法：
 #   OPS="softmax gelu_backward" bash ab_rlc_test.sh
 #   OPS="sort:perf_sort topk" bash ab_rlc_test.sh     # 冒号分隔 OP:OP_FILE
 #
 # 注意：RLC 开关名带厂商前缀，用 VENDOR 指定（默认 MUSA）。套用前先确认开关有效：
 # grep 该开关名于 FlagTree 源码应有命中，否则两侧跑的是同一份代码。
 set -uo pipefail
 cd "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 VENDOR="${VENDOR:-MUSA}"
 ENHANCE_VAR="FLAGTREE_${VENDOR}_RLC_ENHANCE"
 # 空格分隔的算子表。元素为 OP 或 OP:OP_FILE（省略时 OP_FILE=OP）。
 OPS="${OPS:-softmax}"
 if [[ -t 1 && -z "${NO_COLOR:-}" ]]; then
  export FLAGGEMS_PERF_COLOR=always
 fi
 LOG="runs/ab_rlc_$(date +%Y%m%d-%H%M%S).log"
 mkdir -p runs
 {
  echo ">>> vendor=$VENDOR  switch=$ENHANCE_VAR  ops=$OPS"
  for entry in $OPS; do
    op="${entry%%:*}"
    op_file="${entry#*:}"
    [[ "$op_file" == "$entry" ]] && op_file="$op"
    echo "########## $op | $ENHANCE_VAR=0 before (record) ##########"
    if ! OP="$op" OP_FILE="$op_file" env "$ENHANCE_VAR=0" bash run_pytest.sh; then
      echo "!!! FAILED before: $op"
      continue
    fi
    BEFORE=$(ls -dt "runs/${op}_"* | head -1)
    echo "########## $op | $ENHANCE_VAR=1 after (replay $BEFORE) ##########"
    if ! OP="$op" OP_FILE="$op_file" REPLAY_FROM="$BEFORE" env "$ENHANCE_VAR=1" bash run_pytest.sh; then
      echo "!!! FAILED after: $op"
      continue
    fi
    AFTER=$(ls -dt "runs/${op}_"* | head -1)
    echo "########## $op | summary ##########"
    python batch_summary.py --ab "$BEFORE/run.log" "$AFTER/run.log" --label "$op" \
      || echo "!!! summary failed: $op"
  done
  echo "########## ALL DONE ##########"
 } 2>&1 | tee "$LOG"
 # 去掉 ANSI 转义以便 grep/diff（终端输出保留颜色）
 sed -i -E $'s/\x1b\\[[0-9;]*[A-Za-z]//g' "$LOG"
 echo ">>> 总 log: $(pwd)/$LOG"
 echo ">>> 每算子两份 run.log 在 runs/<op>_<时间戳>/ 下，按时间先后即 before / after"
@@ -1,76 +0,0 @@
 """Registry of zl_bench device backends.
 Aliases map common vendor spellings onto the canonical backend name. Unknown
 names fail closed rather than defaulting to CUDA: a silent CUDA fallback would
 load the cudagraph plugin on the wrong device.
 """
 from __future__ import annotations
 import os
 from typing import Dict, Optional
 from .ascend import AscendBackend
 from .base import DeviceBackend, parse_device_list
 from .cuda import CudaBackend
 from .musa import MusaBackend
 from .rocm import RocmBackend
 ALIASES: Dict[str, str] = {
    "nvidia": "cuda",
    "nv": "cuda",
    "amd": "rocm",
    "hip": "rocm",
    "npu": "ascend",
    "huawei": "ascend",
 }
 _INSTANCES: Dict[str, DeviceBackend] = {
    "cuda": CudaBackend(),
    "musa": MusaBackend(),
    "rocm": RocmBackend(),
    "ascend": AscendBackend(),
 }
 class UnknownBackendError(ValueError):
    pass
 def known_backends() -> tuple[str, ...]:
    return tuple(sorted(_INSTANCES))
 def normalize_backend_name(name: str) -> str:
    key = (name or "").strip().lower()
    if not key:
        return "cuda"
    return ALIASES.get(key, key)
 def get_backend(name: Optional[str] = None) -> DeviceBackend:
    raw = name if name is not None else os.environ.get("ZL_BACKEND", "cuda")
    key = normalize_backend_name(raw)
    backend = _INSTANCES.get(key)
    if backend is None:
        raise UnknownBackendError(
            f"unknown backend {raw!r}; known: {', '.join(known_backends())} "
            f"(aliases: {', '.join(sorted(ALIASES))})"
        )
    return backend
 def enumerate_devices(backend: Optional[DeviceBackend] = None,
                      explicit: Optional[str] = None) -> list[str]:
    return (backend or get_backend()).enumerate_devices(explicit)
 __all__ = [
    "DeviceBackend",
    "UnknownBackendError",
    "enumerate_devices",
    "get_backend",
    "known_backends",
    "normalize_backend_name",
    "parse_device_list",
 ]
@@ -1,17 +0,0 @@
 from __future__ import annotations
 from .base import DeviceBackend
 class AscendBackend(DeviceBackend):
    name = "ascend"
    visibility_env_keys = (
        "ASCEND_RT_VISIBLE_DEVICES",
        "ASCEND_VISIBLE_DEVICES",
    )
    torch_namespace = "npu"
    flaggems_vendor = "ascend"
    supports_cuda_graph = False
    supports_mm_cluster_fix = False
    supports_triton_hooks = True
    list_command = None
@@ -1,148 +0,0 @@
 """Device backend contract for zl_bench screening runs.
 This is the orchestration-layer adapter: device enumeration, process visibility
 env, and capability flags that decide which pytest plugins are safe to load.
 It does not implement kernel timing receipts or compiler-artifact identity;
 those stay with the caller (FlagGems pytest for screening, a separate harness
 for qualification).
 """
 from __future__ import annotations
 import os
 import subprocess
 from typing import Dict, Iterable, List, Optional, Sequence, Tuple
 # Every vendor visibility variable zl_bench knows about. apply_visibility()
 # clears the whole set before writing the active backend's keys so a parent
 # CUDA_VISIBLE_DEVICES cannot leak onto a MUSA/Ascend child.
 ALL_VISIBILITY_KEYS: Tuple[str, ...] = (
    "CUDA_VISIBLE_DEVICES",
    "MUSA_VISIBLE_DEVICES",
    "HIP_VISIBLE_DEVICES",
    "ROCR_VISIBLE_DEVICES",
    "ASCEND_RT_VISIBLE_DEVICES",
    "ASCEND_VISIBLE_DEVICES",
 )
 def parse_device_list(raw: Optional[str]) -> List[str]:
    if not raw:
        return []
    return [tok.strip() for tok in raw.split(",") if tok.strip()]
 class DeviceBackend:
    """One hardware runtime (CUDA, MUSA, ROCm, Ascend, ...).
    Subclasses only fill class attributes and optional probe hooks. Callers
    should not branch on ``backend.name`` for visibility or plugin policy;
    they should read these attributes / methods.
    """
    name: str = ""
    # First key is the canonical one written into run_identity.json.
    visibility_env_keys: Tuple[str, ...] = ()
    # torch.<namespace> used for optional device_count / identity probes.
    torch_namespace: str = ""
    # FlagGems GEMS_VENDOR value. Lives on the backend object because zl_bench
    # still only drives FlagGems, but the mapping is owned by providers.flaggems
    # and copied here so DeviceBackend.prepare_environment() can stay generic.
    flaggems_vendor: str = ""
    supports_cuda_graph: bool = False
    supports_mm_cluster_fix: bool = False
    supports_triton_hooks: bool = True
    # Optional listing command; (argv, line_prefix). prefix="" means count
    # non-empty lines. Must be a real vendor CLI; do not invent flags.
    list_command: Optional[Tuple[Sequence[str], str]] = None
    def visibility_env(self, device: str) -> Dict[str, str]:
        return {key: device for key in self.visibility_env_keys}
    def apply_visibility(self, env: Dict[str, str], device: str) -> Dict[str, str]:
        out = dict(env)
        for key in ALL_VISIBILITY_KEYS:
            out.pop(key, None)
        out.update(self.visibility_env(device))
        return out
    def enumerate_devices(self, explicit: Optional[str] = None) -> List[str]:
        listed = parse_device_list(explicit)
        if listed:
            return listed
        for key in self.visibility_env_keys:
            listed = parse_device_list(os.environ.get(key, ""))
            if listed:
                return listed
        n = self.probe_device_count()
        if n > 0:
            return [str(i) for i in range(n)]
        return ["0"]
    def probe_device_count(self) -> int:
        n = self._probe_list_command()
        if n > 0:
            return n
        return self._probe_torch_count()
    def _probe_list_command(self) -> int:
        if not self.list_command:
            return 0
        argv, prefix = self.list_command
        try:
            proc = subprocess.run(
                list(argv), text=True, capture_output=True, timeout=30,
            )
        except Exception:
            return 0
        if proc.returncode != 0:
            return 0
        lines = [ln for ln in proc.stdout.splitlines() if ln.strip()]
        if prefix:
            lines = [ln for ln in lines if ln.startswith(prefix)]
        return len(lines)
    def _probe_torch_count(self) -> int:
        if not self.torch_namespace:
            return 0
        try:
            import torch
            api = getattr(torch, self.torch_namespace, None)
            if api is None or not hasattr(api, "device_count"):
                return 0
            return int(api.device_count())
        except Exception:
            return 0
    def device_identity(self) -> Dict[str, str]:
        """Best-effort identity for screening logs. Empty strings are allowed."""
        info = {
            "backend": self.name,
            "device_name": "",
            "device_arch": "",
            "timer_api": f"torch.{self.torch_namespace}" if self.torch_namespace else "",
            "flaggems_vendor": self.flaggems_vendor,
        }
        if not self.torch_namespace:
            return info
        try:
            import torch
            api = getattr(torch, self.torch_namespace, None)
            if api is None:
                return info
            count = int(api.device_count()) if hasattr(api, "device_count") else 0
            if count <= 0:
                return info
            index = int(api.current_device()) if hasattr(api, "current_device") else 0
            if hasattr(api, "get_device_name"):
                info["device_name"] = str(api.get_device_name(index))
            if hasattr(api, "get_device_capability"):
                cap = api.get_device_capability(index)
                info["device_arch"] = f"{cap[0]}.{cap[1]}"
            elif hasattr(api, "get_device_properties"):
                props = api.get_device_properties(index)
                arch = getattr(props, "name", "") or getattr(props, "arch", "")
                info["device_arch"] = str(arch)
        except Exception:
            return info
        return info
@@ -1,14 +0,0 @@
 from __future__ import annotations
 from .base import DeviceBackend
 class CudaBackend(DeviceBackend):
    name = "cuda"
    visibility_env_keys = ("CUDA_VISIBLE_DEVICES",)
    torch_namespace = "cuda"
    flaggems_vendor = "nvidia"
    supports_cuda_graph = True
    supports_mm_cluster_fix = True
    supports_triton_hooks = True
    list_command = (("nvidia-smi", "--list-gpus"), "GPU ")
@@ -1,16 +0,0 @@
 from __future__ import annotations
 from .base import DeviceBackend
 class MusaBackend(DeviceBackend):
    name = "musa"
    visibility_env_keys = ("MUSA_VISIBLE_DEVICES",)
    torch_namespace = "musa"
    flaggems_vendor = "musa"
    supports_cuda_graph = False
    supports_mm_cluster_fix = False
    supports_triton_hooks = True
    # Device count comes from torch.musa or MUSA_VISIBLE_DEVICES. Do not guess
    # a vendor CLI here; confirm the real listing tool before adding it.
    list_command = None
@@ -1,16 +0,0 @@
 from __future__ import annotations
 from .base import DeviceBackend
 class RocmBackend(DeviceBackend):
    name = "rocm"
    # HIP runtimes honor both; write both so a parent CUDA/HIP mix cannot
    # leave the child looking at the wrong cards.
    visibility_env_keys = ("HIP_VISIBLE_DEVICES", "ROCR_VISIBLE_DEVICES")
    torch_namespace = "cuda"  # ROCm PyTorch still exposes torch.cuda
    flaggems_vendor = "amd"
    supports_cuda_graph = False
    supports_mm_cluster_fix = False
    supports_triton_hooks = True
    list_command = None
@@ -39,31 +39,12 @@ DIRECT_NO_CUDAGRAPH_RE = re.compile(r"BENCHMARK_DIRECT_NO_CUDAGRAPH\b")
 REPLAY_FALLBACK_RE = re.compile(r"AUTOTUNE_REPLAY_FALLBACK\b")
 SUMMARY_FIELDS = [
-    "repo", "op", "backend", "device", "timer",
+    "repo", "op", "fg_op_name", "dtype", "case_id", "row_status",
    "fg_op_name", "dtype", "case_id", "row_status",
    "torch_lat", "gems_lat", "speedup", "extra_metrics", "size_detail",
    "op_status", "notes",
 ]
 def _identity_from_status(status_row: Dict[str, str], op_dir: Path) -> Dict[str, str]:
    """Backend/device/timer from ops_status.csv, falling back to run_identity.json."""
    backend = (status_row.get("backend") or "").strip()
    device = (status_row.get("device") or status_row.get("gpu") or "").strip()
    timer = (status_row.get("timer") or "").strip()
    ident_path = op_dir / "run_identity.json"
    if ident_path.is_file() and not (backend and device and timer):
        try:
            from _run_identity import load_identity
            ident = load_identity(str(op_dir))
            backend = backend or str(ident.get("backend") or "")
            device = device or str(ident.get("device") or ident.get("gpu") or "")
            timer = timer or str(ident.get("timer") or "")
        except Exception:
            pass
    return {"backend": backend, "device": device, "timer": timer}
 def parse_run_log(text: str) -> List[Dict[str, str]]:
    """把 run.log 里的全部结果表解析为行列表（跨多个 dtype 表）。"""
    rows: List[Dict[str, str]] = []
@@ -152,11 +133,10 @@ def summarize(batch_dir: Path, out_path: Optional[Path] = None) -> Path:
            text = log.read_text(errors="replace")
            status_row = status_by_dir.get(f"{repo}/{op_dir.name}", {})
            op = status_row.get("op") or op_dir.name
            ident = _identity_from_status(status_row, op_dir)
            table_rows = parse_run_log(text)
            if not table_rows:
                all_rows.append({
-                    "repo": repo, "op": op, **ident, "fg_op_name": "", "dtype": "",
+                    "repo": repo, "op": op, "fg_op_name": "", "dtype": "",
                    "case_id": "", "row_status": "", "torch_lat": "",
                    "gems_lat": "", "speedup": "", "extra_metrics": "",
                    "size_detail": "",
@@ -170,8 +150,7 @@ def summarize(batch_dir: Path, out_path: Optional[Path] = None) -> Path:
                idx = case_counter.get(key, 0)
                case_counter[key] = idx + 1
                all_rows.append({
-                    "repo": repo, "op": op, **ident,
+                    "repo": repo, "op": op, "case_id": f"case_{idx:03d}",
                    "case_id": f"case_{idx:03d}",
                    "op_status": status_row.get("status", ""),
                    "notes": status_row.get("notes", ""),
                    **row,
@@ -186,12 +165,82 @@ def summarize(batch_dir: Path, out_path: Optional[Path] = None) -> Path:
    return out_path
 def ab_rows(before_text: str, after_text: str) -> List[Dict[str, str]]:
    """把两次 run.log 的测量行按 (op, dtype, size_detail) 配对。
    speedup 列是 gems 对 torch，不是前后对比，所以这里用 gems_lat 自己算
    before/after - 1。只有两侧都 SUCCESS 且延迟为正的行才计入；对不上的行
    单独返回，避免用不完整的配对报一个看起来完整的区间。
    """
    def index(text: str) -> Dict[Tuple[str, str, str], float]:
        out: Dict[Tuple[str, str, str], float] = {}
        for row in parse_run_log(text):
            if row.get("row_status") != "SUCCESS":
                continue
            try:
                lat = float(row.get("gems_lat", ""))
            except ValueError:
                continue
            if lat > 0:
                out[(row.get("fg_op_name", ""), row.get("dtype", ""),
                     row.get("size_detail", ""))] = lat
        return out
    before, after = index(before_text), index(after_text)
    paired: List[Dict[str, str]] = []
    for key in sorted(before.keys() & after.keys()):
        op, dtype, size = key
        b, a = before[key], after[key]
        paired.append({
            "fg_op_name": op, "dtype": dtype, "size_detail": size,
            "before_lat": f"{b:.6f}", "after_lat": f"{a:.6f}",
            "speedup_percent": f"{(b / a - 1.0) * 100.0:.2f}",
        })
    return paired
 def report_ab(before: Path, after: Path, label: str = "") -> List[Dict[str, str]]:
    """打印一个算子的前后对比：逐行 speedup 与整体区间。"""
    before_text = before.read_text(encoding="utf-8", errors="replace")
    after_text = after.read_text(encoding="utf-8", errors="replace")
    rows = ab_rows(before_text, after_text)
    name = label or (rows[0]["fg_op_name"] if rows else "?")
    if not rows:
        print(f"[ab] {name}: 没有可配对的测量行"
              f"（before={before} after={after}）")
        return rows
    values = [float(r["speedup_percent"]) for r in rows]
    for r in rows:
        print(f"  {r['dtype']:<16} {float(r['speedup_percent']):+7.2f}%  "
              f"{r['before_lat']} -> {r['after_lat']}  {r['size_detail']}")
    fallback = ""
    if DIRECT_NO_CUDAGRAPH_RE.search(before_text) or \
            DIRECT_NO_CUDAGRAPH_RE.search(after_text):
        # 回退行用的是 do_bench 口径，与 cudagraph 行不同表可比性
        fallback = "  [含 BENCHMARK_DIRECT_NO_CUDAGRAPH 回退行]"
    print(f"[ab] {name}: {len(rows)} 行配对，speedup "
          f"{min(values):+.2f}% ~ {max(values):+.2f}%{fallback}")
    return rows
 def main() -> None:
    p = argparse.ArgumentParser(description=__doc__.splitlines()[0])
-    p.add_argument("batch_dir", type=Path, help="run_batch.py 的产物目录")
+    p.add_argument("batch_dir", type=Path, nargs="?",
                   help="run_batch.py 的产物目录")
    p.add_argument("-o", "--output", type=Path, default=None,
                   help="输出 CSV（默认 <batch_dir>/summary.csv）")
    p.add_argument("--ab", nargs=2, type=Path, metavar=("BEFORE", "AFTER"),
                   help="两份 run.log，按 shape 配对输出该算子的 speedup 区间")
    p.add_argument("--label", default="", help="--ab 报告里的算子名")
    args = p.parse_args()
    if args.ab:
        for path in args.ab:
            if not path.is_file():
                raise SystemExit(f"run.log 不存在: {path}")
        report_ab(args.ab[0], args.ab[1], args.label)
        return
    if args.batch_dir is None:
        raise SystemExit("需要 batch_dir，或用 --ab BEFORE AFTER")
    if not args.batch_dir.is_dir():
        raise SystemExit(f"目录不存在: {args.batch_dir}")
    out = summarize(args.batch_dir, args.output)
@@ -41,8 +41,6 @@ import yaml
 OPS_DIR = Path(__file__).resolve().parent
 # Inventory is FlagGems-logical (repo/op/op_file), not per-hardware.
 # One case is reused across backends; do not duplicate CSVs per vendor.
 DEFAULT_REPOS = [
    ("main", Path("/workspace/dev/FlagGems")),
    ("vllm", Path("/workspace/dev/FlagGems-vllm")),
@@ -1,107 +0,0 @@
 """Decide which pytest plugins a zl_bench run should load.
 CUDA-specific plugins (cudagraph patch, Hopper mm cluster workaround) are not
 placed on the default list; `_timing_plugin` enables graph capture only when
 the backend declares `supports_cuda_graph`. IR / autotune plugins stay on for
 Triton-compatible compilers and can be turned off with env flags.
 """
 from __future__ import annotations
 import argparse
 import os
 from typing import List, Optional, Sequence
 from backends import get_backend
 CORE_PLUGINS: tuple[str, ...] = (
    "_device_setup_plugin",
    "_parallel_warmup_plugin",
    "_seed_plugin",
    "_shape_inject_plugin",
    "_shape_iter_inject_plugin",
    "_bespoke_shape_plugin",
 )
 ALWAYS_AFTER: tuple[str, ...] = (
    "_pretty_report_plugin",
 )
 def _env_flag(name: str, default: str = "1") -> bool:
    raw = os.environ.get(name, default).strip().lower()
    return raw not in ("0", "false", "no", "off", "")
 def plan_plugins(
    backend_name: Optional[str] = None,
    extra: Optional[Sequence[str]] = None,
    enable_autotune: Optional[bool] = None,
    enable_ir: Optional[bool] = None,
    enable_mm_cluster: Optional[bool] = None,
 ) -> List[str]:
    backend = get_backend(backend_name)
    plugins: List[str] = list(CORE_PLUGINS)
    autotune = (
        backend.supports_triton_hooks
        if enable_autotune is None
        else enable_autotune
    )
    if enable_autotune is None:
        autotune = autotune and _env_flag("ZL_ENABLE_AUTOTUNE", "1")
    if autotune:
        plugins.append("_autotune_record_plugin")
    plugins.append("_timing_plugin")
    plugins.extend(ALWAYS_AFTER)
    ir = backend.supports_triton_hooks if enable_ir is None else enable_ir
    if enable_ir is None:
        ir = ir and _env_flag("ZL_ENABLE_IR", "1")
    if ir:
        plugins.append("_ir_meta_plugin")
    mm = enable_mm_cluster
    if mm is None:
        mm = backend.supports_mm_cluster_fix and _env_flag("ZL_ENABLE_MM_CLUSTER", "0")
    elif mm:
        mm = backend.supports_mm_cluster_fix
    if mm:
        plugins.append("_mm_cluster_fix_plugin")
    if extra is None:
        extras = [p for p in os.environ.get("ZL_EXTRA_PLUGINS", "").split() if p]
    else:
        extras = [p for p in extra if p]
    for name in extras:
        if name not in plugins:
            plugins.append(name)
    return plugins
 def pytest_args(plugins: Optional[Sequence[str]] = None, **kwargs) -> List[str]:
    names = list(plugins) if plugins is not None else plan_plugins(**kwargs)
    args: List[str] = []
    for name in names:
        args.extend(("-p", name))
    return args
 def main(argv: Optional[Sequence[str]] = None) -> int:
    p = argparse.ArgumentParser(description="Print zl_bench pytest -p args")
    p.add_argument("--backend", default=None)
    p.add_argument("--pytest-args", action="store_true",
                   help="print `-p name` tokens on one line")
    p.add_argument("--one-per-line", action="store_true")
    args = p.parse_args(argv)
    names = plan_plugins(backend_name=args.backend)
    if args.one_per_line:
        print("\n".join(names))
    else:
        print(" ".join(pytest_args(names)))
    return 0
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -1 +0,0 @@
 """Benchmark providers. Phase 1 only implements FlagGems."""
@@ -1,219 +0,0 @@
 """FlagGems-specific mapping kept out of the generic scheduler.
 GEMS_VENDOR values are FlagGems' vendor names, not zl_bench backend names.
 cuda → nvidia is the important one: FlagGems' device_finder otherwise shells
 out to nvidia-smi without a timeout.
 """
 from __future__ import annotations
 import argparse
 import os
 import shlex
 import sys
 from pathlib import Path
 from typing import Dict, Optional
 _ROOT = Path(__file__).resolve().parent.parent
 if str(_ROOT) not in sys.path:
    sys.path.insert(0, str(_ROOT))
 from backends.base import DeviceBackend
 # Live checkouts used by the NV FlagGems product path. Order is the fallback
 # sequence when the configured root does not contain the requested test file.
 KNOWN_FLAGGEMS_ROOTS = (
    "/workspace/FlagGems",
    "/workspace/dev/FlagGems",
 )
 VENDOR_ENV_KEYS = (
    "GEMS_VENDOR", "FLAGGEMS_VENDOR", "GEMS_BACKEND", "FLAGGEMS_BACKEND",
 )
 VENDOR_BY_BACKEND: Dict[str, str] = {
    "cuda": "nvidia",
    "musa": "musa",
    "rocm": "amd",
    "ascend": "ascend",
 }
 def flaggems_vendor(backend: DeviceBackend) -> str:
    return backend.flaggems_vendor or VENDOR_BY_BACKEND[backend.name]
 def apply_vendor_env(backend: DeviceBackend) -> Optional[str]:
    """Set GEMS_VENDOR unless the user already chose one. Returns the value."""
    for key in VENDOR_ENV_KEYS:
        if key in os.environ and os.environ[key].strip():
            return os.environ[key].strip()
    vendor = flaggems_vendor(backend)
    if vendor:
        os.environ["GEMS_VENDOR"] = vendor
    return vendor or None
 def make_test_node(flaggems_dir: str, op_file: str, op: str) -> str:
    root = flaggems_dir.replace("\\", "/").rstrip("/")
    return f"{root}/benchmark/test_{op_file}.py::test_{op}"
 # Back-compat alias. Do not call this from emit_run_env: the parameter of the
 # same name would shadow it and raise TypeError.
 test_node = make_test_node
 def parse_test_node(node: str) -> Dict[str, Optional[str]]:
    """Parse a FlagGems pytest node into op / op_file / optional repo root.
    Example:
      /workspace/FlagGems/benchmark/test_fused_marlin_moe.py::test_fused_marlin_moe_mxfp4
    """
    raw = (node or "").strip()
    if "::" not in raw:
        raise ValueError(f"pytest node must contain '::': {node!r}")
    file_part, func = raw.rsplit("::", 1)
    if not func.startswith("test_"):
        raise ValueError(f"test function must start with 'test_': {func!r}")
    stem = Path(file_part.replace("\\", "/").rsplit("/", 1)[-1]).stem
    if not stem.startswith("test_"):
        raise ValueError(f"benchmark filename must start with 'test_': {file_part!r}")
    posix_file = file_part.replace("\\", "/")
    parent = posix_file.rsplit("/", 1)[0] if "/" in posix_file else ""
    parent_name = parent.rsplit("/", 1)[-1] if parent else ""
    flaggems_dir = (
        parent.rsplit("/", 1)[0] if parent_name == "benchmark" and "/" in parent else None
    )
    return {
        "test_file": f"{file_part}::{func}",
        "file": file_part,
        "op": func[len("test_"):],
        "op_file": stem[len("test_"):],
        "flaggems_dir": flaggems_dir,
    }
 def _test_py(flaggems_dir: str, op_file: str) -> Path:
    return Path(flaggems_dir) / "benchmark" / f"test_{op_file}.py"
 def resolve_flaggems_dir(configured: str, op_file: str) -> str:
    """Return the first root that actually contains benchmark/test_<op_file>.py."""
    candidates = []
    for item in (configured, *KNOWN_FLAGGEMS_ROOTS):
        if item and item not in candidates:
            candidates.append(item)
    for root in candidates:
        if _test_py(root, op_file).is_file():
            return root
    return configured
 def resolve_existing_test_file(test_file: str, op_file: str) -> str:
    """If the node path is missing, retry the same relative path under known roots."""
    if "::" not in test_file:
        return test_file
    file_part, func = test_file.rsplit("::", 1)
    path = Path(file_part)
    if path.is_file():
        return test_file
    for root in KNOWN_FLAGGEMS_ROOTS:
        candidate = _test_py(root, op_file)
        if candidate.is_file():
            return f"{candidate.as_posix()}::{func}"
    return test_file
 def emit_run_env(
    test_node: str = "",
    op: str = "",
    op_file: str = "",
    flaggems_dir: str = "",
    shape_file: str = "",
    shape_mode: str = "",
 ) -> Dict[str, str]:
    """Resolve OP / OP_FILE / TEST_FILE / SHAPE_MODE for run_pytest.sh."""
    parsed = parse_test_node(test_node) if test_node.strip() else None
    if parsed:
        op = parsed["op"] or op
        op_file = parsed["op_file"] or op_file
        if parsed["flaggems_dir"]:
            flaggems_dir = parsed["flaggems_dir"]
        test_file = parsed["test_file"] or ""
    else:
        flaggems_dir = resolve_flaggems_dir(flaggems_dir, op_file)
        test_file = make_test_node(flaggems_dir, op_file, op)
    flaggems_dir = resolve_flaggems_dir(flaggems_dir, op_file)
    test_file = resolve_existing_test_file(test_file, op_file)
    shape_file = shape_file.strip()
    shape_mode = shape_mode.strip().lower()
    if shape_file in ("-", "upstream", "none"):
        shape_file = ""
        shape_mode = "upstream"
    if not shape_mode:
        shape_mode = "upstream" if (test_node.strip() and not shape_file) else "inline"
    if shape_file:
        shape_mode = "file"
        pass_shape_file = "1"
    elif shape_mode == "upstream":
        pass_shape_file = "0"
    else:
        pass_shape_file = "1"
        shape_mode = "inline"
    return {
        "OP": op,
        "OP_FILE": op_file,
        "FLAGGEMS_DIR": flaggems_dir,
        "TEST_FILE": test_file,
        "SHAPE_FILE": shape_file,
        "SHAPE_MODE": shape_mode,
        "PASS_SHAPE_FILE": pass_shape_file,
    }
 def _print_shell(env: Dict[str, str]) -> None:
    for key, value in env.items():
        print(f"{key}={shlex.quote(str(value))}")
 def main(argv: Optional[list] = None) -> int:
    p = argparse.ArgumentParser(description="FlagGems pytest node helpers")
    p.add_argument("--parse-node", default="")
    p.add_argument("--emit-shell", action="store_true")
    p.add_argument("--op", default="")
    p.add_argument("--op-file", default="")
    p.add_argument("--flaggems-dir", default="")
    p.add_argument("--shape-file", default="")
    p.add_argument("--shape-mode", default="")
    p.add_argument("--shell", action="store_true",
                   help="with --parse-node, print OP/OP_FILE/FLAGGEMS_DIR/TEST_FILE")
    args = p.parse_args(argv)
    if args.emit_shell:
        _print_shell(emit_run_env(
            test_node=args.parse_node,
            op=args.op,
            op_file=args.op_file,
            flaggems_dir=args.flaggems_dir,
            shape_file=args.shape_file,
            shape_mode=args.shape_mode,
        ))
        return 0
    if args.parse_node:
        parsed = parse_test_node(args.parse_node)
        if args.shell:
            _print_shell({k: v or "" for k, v in parsed.items()
                          if k in ("op", "op_file", "flaggems_dir", "test_file")})
            return 0
        for key, value in parsed.items():
            print(f"{key}={value or ''}")
        return 0
    p.error("pass --emit-shell or --parse-node")
    return 2
 if __name__ == "__main__":
    raise SystemExit(main())
@@ -1,16 +1,13 @@
 #!/usr/bin/env python3
 """多算子批量测试驱动：按算子清单逐个调 run_pytest.sh，多卡分片、串行测量。
-每个算子一个独立子进程（设备 crash 只废单个算子）、独立产物目录
+每个算子一个独立子进程（CUDA crash 只废单个算子）、独立产物目录
 <batch>/<repo>/<op>/（与单算子跑 run_pytest.sh 的产物结构完全一致，含
-run.log / shapes.yaml / autotune_records / ttgir / run_identity.json）。
+run.log / shapes.yaml / autotune_records / ttgir）。批量层只做编排：
 批量层只做编排：
- 分片：按算子名稳定哈希到设备（每卡内部串行，测量口径与单算子一致；
+- 分片：按算子名稳定哈希到 GPU（每卡内部串行，测量口径与单算子一致；
  跨卡并行测量存在功耗耦合噪声，本模式定位为筛查口径，可疑算子回
  单算子串行流程确认）；
 - 后端：--backend / ZL_BACKEND 选择 cuda/musa/rocm/ascend，设备可见性
  由 backends/ 写入，不再写死 CUDA_VISIBLE_DEVICES；
 - 超时：--op-timeout 杀死挂死算子（记 TIMEOUT，不重试，继续跑）；
 - 重试：pytest rc=1（用例失败/flaky）与信号杀（rc>=128）最多重试 2 次，
  连续两次同 rc 视为确定性失败提前止损；--dtypes 下两级降级——遇上游
@@ -26,7 +23,7 @@ run.log / shapes.yaml / autotune_records / ttgir / run_identity.json）。
 示例：
  # 全量筛查（两仓库全部可收集算子，排除 marked_skip 与 dismiss）
-  python run_batch.py --all --devices 0,1,2,3,4,5,6,7 --backend cuda --op-timeout 1800
+  python run_batch.py --all --gpus 0,1,2,3,4,5,6,7 --op-timeout 1800
  # 指定算子子集（repo 前缀默认 main；重名用 @文件名 消歧）
  python run_batch.py --ops softmax vllm:fused_marlin_moe nextafter_@nextafter_
@@ -53,8 +50,6 @@ from pathlib import Path
 from typing import Dict, List, Optional, Set, Tuple
 import batch_summary
 from backends import UnknownBackendError, get_backend
 from _run_identity import load_identity
 ZL_BENCH = Path(__file__).resolve().parent
 OPS_DIR = ZL_BENCH / "ops"
@@ -67,9 +62,7 @@ REPO_DIRS = {
 }
 STATUS_FIELDS = [
-    "repo", "op", "op_file", "run_name",
+    "repo", "op", "op_file", "run_name", "gpu", "status", "return_code",
    "backend", "device", "gpu", "timer",
    "status", "return_code",
    "elapsed_ms", "rows_total", "rows_success", "notes", "out_dir",
 ]
@@ -201,44 +194,48 @@ def resolve_selection(args: argparse.Namespace,
    return cases
-# ---------------- 设备分片 ----------------
+# ---------------- GPU 分片 ----------------
-def resolve_device_arg(devices: Optional[str], gpus: Optional[str]) -> Optional[str]:
+def visible_gpus(arg: Optional[str]) -> List[str]:
-    """Prefer --devices; --gpus is an alias. Passing both is an error."""
+    if arg:
-    if devices is not None and gpus is not None:
+        return [g.strip() for g in arg.split(",") if g.strip()]
-        raise ValueError("do not pass both --devices and --gpus")
+    env = os.environ.get("CUDA_VISIBLE_DEVICES", "").strip()
-    return devices if devices is not None else gpus
+    if env:
        return [g.strip() for g in env.split(",") if g.strip()]
    try:
        out = subprocess.run(["nvidia-smi", "--list-gpus"], text=True,
                             capture_output=True, timeout=30).stdout
        n = sum(1 for line in out.splitlines() if line.startswith("GPU "))
        if n:
            return [str(i) for i in range(n)]
    except Exception:
        pass
    return ["0"]
-def visible_devices(backend, explicit: Optional[str] = None) -> List[str]:
+def partition(cases: List[Case], gpus: List[str]) -> Dict[str, List[Case]]:
    return list(backend.enumerate_devices(explicit))
 def partition(cases: List[Case], devices: List[str]) -> Dict[str, List[Case]]:
    """按 (repo, run_name) 稳定哈希分片：与选择顺序无关，resume 时同卡复用。"""
-    parts: Dict[str, List[Case]] = {d: [] for d in devices}
+    parts: Dict[str, List[Case]] = {g: [] for g in gpus}
    for c in cases:
        digest = hashlib.sha256(f"{c.repo}/{c.run_name}".encode()).digest()
-        parts[devices[int.from_bytes(digest[:8], "big") % len(devices)]].append(c)
+        parts[gpus[int.from_bytes(digest[:8], "big") % len(gpus)]].append(c)
    return parts
 # ---------------- 单算子执行 ----------------
-def _spawn(case: Case, out_dir: Path, device: str, args: argparse.Namespace,
+def _spawn(case: Case, out_dir: Path, gpu: str, args: argparse.Namespace,
           dtypes: str) -> subprocess.Popen:
    env = os.environ.copy()
    env.pop("PARALLEL_WARMUP_GPUS", None)   # 批量层已分片，算子内不再并行
    env.pop("REPLAY_FROM", None)
    env.pop("OUT_DIR", None)
    env = args.backend_obj.apply_visibility(env, device)
    env.update({
        "OP": case.op,
        "OP_FILE": case.op_file,
        "FLAGGEMS_DIR": REPO_DIRS[case.repo],
        "OUT_DIR": str(out_dir),
-        "ZL_BACKEND": args.backend,
+        "CUDA_VISIBLE_DEVICES": gpu,
        "ZL_DEVICE": device,
        "FLAGGEMS_PERF_COLOR": "never",
        "NO_COLOR": "1",
        "USE_FLAGTUNE": str(args.use_flagtune),
@@ -274,14 +271,13 @@ def _wait(proc: subprocess.Popen, timeout: int) -> Tuple[int, bool]:
        return 124, True
-def run_case(case: Case, batch_dir: Path, device: str,
+def run_case(case: Case, batch_dir: Path, gpu: str,
             args: argparse.Namespace) -> Dict[str, str]:
    out_dir = batch_dir / case.repo / case.run_name
    marker = out_dir / ".complete"
    if args.resume and marker.is_file() and (out_dir / "run.log").is_file():
-        row = _classify(case, out_dir, device, 0, 0, "", timed_out=False,
+        row = _classify(case, out_dir, gpu, 0, 0, "", timed_out=False)
                        args=args)
        row["notes"] = _join("resumed", row["notes"])
        return row
@@ -299,7 +295,7 @@ def run_case(case: Case, batch_dir: Path, device: str,
            log = out_dir / "run.log"
            if log.is_file():
                log.rename(out_dir / f"run.attempt{attempt - 1}.log")
-        rc, timed_out = _wait(_spawn(case, out_dir, device, args, dtypes),
+        rc, timed_out = _wait(_spawn(case, out_dir, gpu, args, dtypes),
                              args.op_timeout)
        rcs.append(rc)
        if timed_out:
@@ -319,22 +315,20 @@ def run_case(case: Case, batch_dir: Path, device: str,
    # 覆盖 baseline 对受限 dtype 编译失败（torch jiterator 的 special 函数族等）
    # 但报错文案不是上游 "can't be supported by this op" 的场景。
    if dtypes and not timed_out:
-        probe = _classify(case, out_dir, device, rcs[-1], 0, "", False,
+        probe = _classify(case, out_dir, gpu, rcs[-1], 0, "", False)
                          args=args)
        if probe["status"] == "FAIL" and probe["rows_success"] == "0":
            log = out_dir / "run.log"
            if log.is_file():
                log.rename(out_dir / f"run.attempt{len(rcs) - 1}.log")
            notes = _join(notes, "dtype_rescue")
-            rc, timed_out = _wait(_spawn(case, out_dir, device, args, ""),
+            rc, timed_out = _wait(_spawn(case, out_dir, gpu, args, ""),
                                  args.op_timeout)
            rcs.append(rc)
    elapsed_ms = int((time.time() - t0) * 1000)
    if len(rcs) > 1:
        notes = _join(notes, "retry_rc=" + "->".join(map(str, rcs)))
-    row = _classify(case, out_dir, device, rcs[-1], elapsed_ms, notes, timed_out,
+    row = _classify(case, out_dir, gpu, rcs[-1], elapsed_ms, notes, timed_out)
                    args=args)
    if row["status"] in ("PASS", "SKIP"):
        tmp = marker.with_name(".complete.tmp")
        tmp.write_text(row["status"] + "\n")
@@ -351,21 +345,8 @@ def _join(*notes: str) -> str:
    return ";".join(n for n in notes if n)
-def _identity_fields(out_dir: Path, device: str, backend: str) -> Dict[str, str]:
+def _classify(case: Case, out_dir: Path, gpu: str, rc: int, elapsed_ms: int,
-    ident = load_identity(str(out_dir))
+              notes: str, timed_out: bool) -> Dict[str, str]:
    backend_v = str(ident.get("backend") or backend or "")
    device_v = str(ident.get("device") or device or "")
    timer_v = str(ident.get("timer") or "")
    return {
        "backend": backend_v,
        "device": device_v,
        "gpu": device_v,
        "timer": timer_v,
    }
 def _classify(case: Case, out_dir: Path, device: str, rc: int, elapsed_ms: int,
              notes: str, timed_out: bool, args: argparse.Namespace) -> Dict[str, str]:
    text = _read_log(out_dir)
    rows = batch_summary.parse_run_log(text)
    if timed_out:
@@ -375,13 +356,12 @@ def _classify(case: Case, out_dir: Path, device: str, rc: int, elapsed_ms: int,
    n_skipped = batch_summary.pytest_skipped_count(text)
    if status == "PASS" and n_skipped:
        notes = _join(notes, f"partial_skip={n_skipped}")
    ident = _identity_fields(out_dir, device, args.backend)
    return {
        "repo": case.repo,
        "op": case.op,
        "op_file": case.op_file,
        "run_name": case.run_name,
-        **ident,
+        "gpu": gpu,
        "status": status,
        "return_code": str(rc),
        "elapsed_ms": str(elapsed_ms),
@@ -413,12 +393,8 @@ def parse_args() -> argparse.Namespace:
                     help="--all 时不剔除带无条件 skip 标记的算子")
    run = p.add_argument_group("执行")
    run.add_argument("--backend", default=os.environ.get("ZL_BACKEND", "cuda"),
                     help="设备后端：cuda/musa/rocm/ascend（默认 ZL_BACKEND 或 cuda）")
    run.add_argument("--devices", default=None,
                     help="逗号分隔设备列表（默认该后端可见设备；优先于 --gpus）")
    run.add_argument("--gpus", default=None,
-                     help="--devices 的别名（两者同时给出则报错）")
+                     help="逗号分隔 GPU 列表（默认 CUDA_VISIBLE_DEVICES 或全部）")
    run.add_argument("--batch-dir", type=Path, default=None,
                     help="产物目录（默认 runs/batch_<时间戳>；resume 时必须指定）")
    run.add_argument("--op-timeout", type=int, default=1800,
@@ -453,30 +429,19 @@ def parse_args() -> argparse.Namespace:
 def main() -> None:
    args = parse_args()
    try:
        backend = get_backend(args.backend)
    except UnknownBackendError as exc:
        raise SystemExit(f"{LOG} {exc}") from exc
    args.backend_obj = backend
    args.backend = backend.name
    try:
        explicit = resolve_device_arg(args.devices, args.gpus)
    except ValueError as exc:
        raise SystemExit(f"{LOG} {exc}") from exc
    devices = visible_devices(backend, explicit)
    inv = load_inventory()
    cases = resolve_selection(args, inv)
    if not cases:
        raise SystemExit(f"{LOG} 选择结果为空")
    gpus = visible_gpus(args.gpus)
    batch_dir = (args.batch_dir or
                 ZL_BENCH / "runs" / f"batch_{datetime.now():%Y%m%d-%H%M%S}")
    batch_dir = batch_dir.resolve()
    batch_dir.mkdir(parents=True, exist_ok=True)
-    parts = {d: cs for d, cs in partition(cases, devices).items() if cs}
+    parts = {g: cs for g, cs in partition(cases, gpus).items() if cs}
-    print(f"{LOG} backend={args.backend}  {len(cases)} 个算子 -> {len(parts)} 台设备 "
+    print(f"{LOG} {len(cases)} 个算子 -> {len(parts)} 张卡 "
          f"(shards: {[len(cs) for cs in parts.values()]})  out={batch_dir}")
    if args.replay_root:
        print(f"{LOG} replay 模式: REPLAY_FROM={args.replay_root}/<repo>/<op>")
@@ -496,19 +461,17 @@ def main() -> None:
            writer.writerows(rows)
        os.replace(tmp, status_csv)
-    def run_partition(device: str, part: List[Case]) -> None:
+    def run_partition(gpu: str, part: List[Case]) -> None:
        for case in part:
            try:
-                row = run_case(case, batch_dir, device, args)
+                row = run_case(case, batch_dir, gpu, args)
            except Exception as exc:  # 编排层异常不拖垮整卡分片
                out_dir = batch_dir / case.repo / case.run_name
                ident = _identity_fields(out_dir, device, args.backend)
                row = {
                    "repo": case.repo, "op": case.op, "op_file": case.op_file,
-                    "run_name": case.run_name, **ident, "status": "FAIL",
+                    "run_name": case.run_name, "gpu": gpu, "status": "FAIL",
                    "return_code": "EXC", "elapsed_ms": "0", "rows_total": "0",
                    "rows_success": "0", "notes": f"harness_exc={exc}",
-                    "out_dir": str(out_dir),
+                    "out_dir": str(batch_dir / case.repo / case.run_name),
                }
            with lock:
                done[0] += 1
@@ -517,16 +480,16 @@ def main() -> None:
                print(f"  ({done[0]:>4}/{len(cases)})  {row['status']:<7} "
                      f"{row['elapsed_ms']:>8}ms  rows={row['rows_success']}/"
                      f"{row['rows_total']:<4} {case.repo}:{case.run_name:<{wname}} "
-                      f"[device{device}]"
+                      f"[gpu{gpu}]"
                      + (f"  {row['notes']}" if row["notes"] else ""),
                      flush=True)
    if len(parts) == 1:
-        device, part = next(iter(parts.items()))
+        gpu, part = next(iter(parts.items()))
-        run_partition(device, part)
+        run_partition(gpu, part)
    else:
        with ThreadPoolExecutor(max_workers=len(parts)) as pool:
-            futures = [pool.submit(run_partition, d, p) for d, p in parts.items()]
+            futures = [pool.submit(run_partition, g, p) for g, p in parts.items()]
            for f in futures:
                f.result()
@@ -8,21 +8,15 @@ set -euo pipefail
 # 每项均可用同名环境变量覆盖，详见 README「环境变量一览」。
 # 被测算子，对应 $FLAGGEMS_DIR/benchmark/test_<OP_FILE>.py::test_<OP>
 # 也可直接传 FlagGems pytest 节点，例如：
 #   TEST_NODE=/workspace/FlagGems/benchmark/test_fused_marlin_moe.py::test_fused_marlin_moe_mxfp4
 OP="${OP:-fused_marlin_moe_mxfp4}"
 OP_FILE="${OP_FILE:-fused_marlin_moe}"
 # 上游已拆分为两个仓库：主仓 /workspace/dev/FlagGems（包名 flag_gems）与
 # /workspace/dev/FlagGems-vllm（包名 flaggems_vllm）。测 vllm 仓库算子时覆盖本变量即可。
 # NV 产品路径也常见 /workspace/FlagGems；TEST_NODE 或缺失测试文件时会回退到该路径。
 FLAGGEMS_DIR="${FLAGGEMS_DIR:-/workspace/dev/FlagGems}"
 TEST_NODE="${TEST_NODE:-${ZL_TEST_NODE:-}}"
-# shape 来源：非空则用该 yaml；SHAPE_MODE=upstream 则不传 --shape_file，沿用 FlagGems
+# shape 来源：非空则用该 yaml，否则用 INLINE_YAML。顶层 key 必须是 op 名。
 # 在 $BENCH_LEVEL 下的原生 shape 集。设了 TEST_NODE 且未指定 SHAPE_FILE 时默认 upstream。
 SHAPE_FILE="${SHAPE_FILE:-}"
 SHAPE_MODE="${SHAPE_MODE:-}"
 read -r -d '' INLINE_YAML <<'YAML' || true
 fused_marlin_moe_mxfp4:
  shapes:
@@ -59,34 +53,25 @@ REPLAY_FROM="${REPLAY_FROM:-}"
 # always/never 强制开关终端颜色，空则按 tty 判断（run.log 始终去色）
 FLAGGEMS_PERF_COLOR="${FLAGGEMS_PERF_COLOR:-}"
-# 设备后端：cuda / musa / rocm / ascend（别名见 backends/__init__.py）。
+# 各插件作用见 README「插件说明」；注释掉某行即停用该插件
-# 插件列表由 plugin_plan.py 按后端能力生成，不再在此硬编码 -p。
+PLUGINS=(
-# 非 CUDA 默认走上游 do_bench，不会加载 _cudagraph_plugin。
+  -p _device_guard_plugin
-ZL_BACKEND="${ZL_BACKEND:-cuda}"
+  -p _parallel_warmup_plugin
-
+  -p _seed_plugin
-# 上游 --level：core（默认）只用各算子的 core_shapes；comprehensive 额外并入
+  -p _shape_inject_plugin
-# 子类的 set_more_shapes()，并对同一 shape 再产出一次变体（BLAS 是转置侧）。
+  -p _shape_iter_inject_plugin
-# 差距可以很大：mm 在 core 下是 5 个 shape，comprehensive 下并入 4 个大 K
+  -p _bespoke_shape_plugin
-# （N/K 到 152064）与 models_shapes/*.yaml 的模型 shape，共 29 个，且用例翻倍。
+  # -p _mm_cluster_fix_plugin
-# 代价同步放大——sweep 是墙上时间的大头，shape 多时配合 PARALLEL_WARMUP_GPUS。
+  -p _autotune_record_plugin
-BENCH_LEVEL="${BENCH_LEVEL:-core}"
+  -p _cudagraph_plugin
-case "$BENCH_LEVEL" in
+  -p _pretty_report_plugin
-  core|comprehensive) ;;
+  -p _ir_meta_plugin
-  *) echo "!!! BENCH_LEVEL 只能是 core 或 comprehensive（当前：$BENCH_LEVEL）" >&2; exit 2 ;;
+)
 esac
 # ============================== 执行逻辑 ==============================
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-export PYTHONPATH="$SCRIPT_DIR${PYTHONPATH:+:$PYTHONPATH}"
+TEST_FILE="$FLAGGEMS_DIR/benchmark/test_${OP_FILE}.py::test_${OP}"
 eval "$(python "$SCRIPT_DIR/providers/flaggems.py" --emit-shell \
  --parse-node "$TEST_NODE" \
  --op "$OP" \
  --op-file "$OP_FILE" \
  --flaggems-dir "$FLAGGEMS_DIR" \
  --shape-file "$SHAPE_FILE" \
  --shape-mode "$SHAPE_MODE")"
 : "${TEST_FILE:=$FLAGGEMS_DIR/benchmark/test_${OP_FILE}.py::test_${OP}}"
 # 产物目录，可用 OUT_DIR 指定。时间戳只到秒，并发启动会撞名，故用不带 -p 的
 # mkdir 抢占（已存在即失败），撞上就退让到 -2、-3……
@@ -119,12 +104,9 @@ else
  PYTEST_COLOR=()
 fi
 export PYTHONPATH="$SCRIPT_DIR${PYTHONPATH:+:$PYTHONPATH}"
 export FLAGGEMS_PERF_CURRENT_OP="$OP"
 export PYTHONUNBUFFERED=1  # 实时输出不缓冲
 export ZL_BACKEND
 # 按 ZL_BACKEND / ZL_ENABLE_* 生成 -p 列表（见 plugin_plan.py）
 read -r -a PLUGINS <<< "$(python "$SCRIPT_DIR/plugin_plan.py" --backend "$ZL_BACKEND" --pytest-args)"
 # DTYPES 非空时逐个转为上游 --dtypes 选项（action=append，每个 dtype 一次）
 DTYPE_ARGS=()
@@ -144,39 +126,17 @@ else
  MODE_DESC=record
 fi
-# 未指定 shape 文件时：inline 用 INLINE_YAML；upstream 不传 --shape_file。
+# 未指定 shape 文件时，把 INLINE_YAML 落到临时文件供 pytest 读取
-SHAPE_ARGS=()
+if [[ -z "$SHAPE_FILE" ]]; then
 if [[ "${PASS_SHAPE_FILE:-1}" == "0" ]]; then
  if [[ -f "$FLAGGEMS_DIR/benchmark/core_shapes.yaml" ]]; then
    cp -f "$FLAGGEMS_DIR/benchmark/core_shapes.yaml" "$OUT_DIR/shapes.yaml"
  else
    printf '%s\n' "source: upstream" "level: $BENCH_LEVEL" > "$OUT_DIR/shapes.yaml"
  fi
 else
  if [[ -z "$SHAPE_FILE" ]]; then
  SHAPE_FILE="$(mktemp --suffix=.yaml)"
  printf '%s\n' "$INLINE_YAML" > "$SHAPE_FILE"
  trap 'rm -f "$SHAPE_FILE"' EXIT
  fi
  cp -f "$SHAPE_FILE" "$OUT_DIR/shapes.yaml"  # 存档本次实际使用的 shape
  SHAPE_ARGS=(--shape_file "$SHAPE_FILE")
 fi
-
+cp -f "$SHAPE_FILE" "$OUT_DIR/shapes.yaml"  # 存档本次实际使用的 shape
 python "$SCRIPT_DIR/_run_identity.py" "$OUT_DIR" \
  --backend "$ZL_BACKEND" \
  --op "$OP" \
  --op-file "$OP_FILE" \
  --flaggems-dir "$FLAGGEMS_DIR" \
  --autotune-mode "$MODE_DESC" \
  --device "${ZL_DEVICE:-}" \
  --test-file "$TEST_FILE" \
  --shape-mode "$SHAPE_MODE" \
  --bench-level "$BENCH_LEVEL"
 status=0
 {
-  echo "${C_BOLD}>>> op=$OP  backend=$ZL_BACKEND  mode=$MODE_DESC  USE_FLAGTUNE=$USE_FLAGTUNE  shape=$SHAPE_MODE  level=$BENCH_LEVEL${C_RESET}"
+  echo "${C_BOLD}>>> op=$OP  mode=$MODE_DESC  USE_FLAGTUNE=$USE_FLAGTUNE${C_RESET}"
  echo "${C_DIM}>>> node=$TEST_FILE${C_RESET}"
  echo "${C_DIM}>>> out=$OUT_DIR${C_RESET}"
  # 每次用独立的 Triton 缓存目录，跑完即删：保证编译过程可复现，且 ttgir 落盘
@@ -189,8 +149,8 @@ status=0
  env "$AUTOTUNE_ENV" \
  USE_FLAGTUNE=$USE_FLAGTUNE python -u -m pytest -s "$TEST_FILE" \
    "${PLUGINS[@]}" "${PYTEST_COLOR[@]}" \
-    ${SHAPE_ARGS[@]+"${SHAPE_ARGS[@]}"} ${DTYPE_ARGS[@]+"${DTYPE_ARGS[@]}"} \
+    --shape_file "$SHAPE_FILE" ${DTYPE_ARGS[@]+"${DTYPE_ARGS[@]}"} \
-    --level "$BENCH_LEVEL" --mode kernel || status=$?
+    --level core --mode kernel || status=$?
  rm -rf "$CACHE_DIR"
@@ -1,285 +0,0 @@
 #!/usr/bin/env python3
 """No-GPU unit tests for zl_bench backend adapters and plugin policy.
 FlagGems pytest invocations pass an explicit node, so this file is not
 collected during operator runs. Run it with `python test_backends.py`.
 """
 from __future__ import annotations
 import json
 import os
 import sys
 import tempfile
 import unittest
 from pathlib import Path
 ROOT = Path(__file__).resolve().parent
 if str(ROOT) not in sys.path:
    sys.path.insert(0, str(ROOT))
 from backends import (  # noqa: E402
    UnknownBackendError,
    get_backend,
    normalize_backend_name,
 )
 from backends.base import ALL_VISIBILITY_KEYS  # noqa: E402
 from batch_summary import SUMMARY_FIELDS  # noqa: E402
 from plugin_plan import plan_plugins  # noqa: E402
 from providers.flaggems import (  # noqa: E402
    VENDOR_BY_BACKEND,
    emit_run_env,
    flaggems_vendor,
    parse_test_node,
 )
 from run_batch import STATUS_FIELDS, resolve_device_arg  # noqa: E402
 from _run_identity import init_identity, load_identity, update_identity  # noqa: E402
 class BackendResolveTests(unittest.TestCase):
    def test_empty_and_alias_resolve_to_cuda(self):
        self.assertEqual(normalize_backend_name(""), "cuda")
        self.assertEqual(get_backend("nvidia").name, "cuda")
        self.assertEqual(get_backend("nv").name, "cuda")
        self.assertEqual(get_backend("").name, "cuda")
    def test_other_aliases(self):
        self.assertEqual(get_backend("amd").name, "rocm")
        self.assertEqual(get_backend("hip").name, "rocm")
        self.assertEqual(get_backend("npu").name, "ascend")
        self.assertEqual(get_backend("huawei").name, "ascend")
    def test_unknown_backend_fails_closed(self):
        with self.assertRaises(UnknownBackendError):
            get_backend("tpu")
        with self.assertRaises(UnknownBackendError):
            get_backend("not-a-backend")
 class VisibilityTests(unittest.TestCase):
    def test_musa_drops_leftover_cuda(self):
        env = {"CUDA_VISIBLE_DEVICES": "7", "PATH": "/bin"}
        out = get_backend("musa").apply_visibility(env, "1")
        self.assertNotIn("CUDA_VISIBLE_DEVICES", out)
        self.assertEqual(out["MUSA_VISIBLE_DEVICES"], "1")
        self.assertEqual(out["PATH"], "/bin")
        for key in ALL_VISIBILITY_KEYS:
            if key != "MUSA_VISIBLE_DEVICES":
                self.assertNotIn(key, out)
    def test_rocm_writes_hip_and_rocr(self):
        out = get_backend("rocm").apply_visibility(
            {"CUDA_VISIBLE_DEVICES": "0"}, "3")
        self.assertEqual(out["HIP_VISIBLE_DEVICES"], "3")
        self.assertEqual(out["ROCR_VISIBLE_DEVICES"], "3")
        self.assertNotIn("CUDA_VISIBLE_DEVICES", out)
    def test_ascend_writes_both_keys(self):
        out = get_backend("ascend").apply_visibility({}, "2")
        self.assertEqual(out["ASCEND_RT_VISIBLE_DEVICES"], "2")
        self.assertEqual(out["ASCEND_VISIBLE_DEVICES"], "2")
        self.assertNotIn("CUDA_VISIBLE_DEVICES", out)
    def test_cuda_only_sets_cuda(self):
        out = get_backend("cuda").apply_visibility(
            {"MUSA_VISIBLE_DEVICES": "9"}, "0")
        self.assertEqual(out["CUDA_VISIBLE_DEVICES"], "0")
        self.assertNotIn("MUSA_VISIBLE_DEVICES", out)
 class FlagGemsVendorTests(unittest.TestCase):
    def test_vendor_map(self):
        self.assertEqual(VENDOR_BY_BACKEND["cuda"], "nvidia")
        self.assertEqual(VENDOR_BY_BACKEND["rocm"], "amd")
        self.assertEqual(flaggems_vendor(get_backend("cuda")), "nvidia")
        self.assertEqual(flaggems_vendor(get_backend("rocm")), "amd")
        self.assertEqual(flaggems_vendor(get_backend("musa")), "musa")
        self.assertEqual(flaggems_vendor(get_backend("ascend")), "ascend")
 class FlagGemsNodeTests(unittest.TestCase):
    _NODE = (
        "/workspace/FlagGems/benchmark/test_fused_marlin_moe.py"
        "::test_fused_marlin_moe_mxfp4"
    )
    def test_parse_user_command_node(self):
        parsed = parse_test_node(self._NODE)
        self.assertEqual(parsed["op"], "fused_marlin_moe_mxfp4")
        self.assertEqual(parsed["op_file"], "fused_marlin_moe")
        self.assertEqual(parsed["flaggems_dir"], "/workspace/FlagGems")
        self.assertTrue(parsed["test_file"].endswith(
            "test_fused_marlin_moe.py::test_fused_marlin_moe_mxfp4"))
    def test_test_node_defaults_to_upstream_shapes(self):
        env = emit_run_env(
            test_node=self._NODE,
            op="softmax",
            op_file="softmax",
            flaggems_dir="/workspace/dev/FlagGems",
        )
        self.assertEqual(env["OP"], "fused_marlin_moe_mxfp4")
        self.assertEqual(env["OP_FILE"], "fused_marlin_moe")
        self.assertEqual(env["FLAGGEMS_DIR"], "/workspace/FlagGems")
        self.assertEqual(env["SHAPE_MODE"], "upstream")
        self.assertEqual(env["PASS_SHAPE_FILE"], "0")
        self.assertEqual(env["TEST_FILE"], self._NODE)
    def test_plain_run_keeps_inline_shapes(self):
        env = emit_run_env(
            op="fused_marlin_moe_mxfp4",
            op_file="fused_marlin_moe",
            flaggems_dir="/workspace/dev/FlagGems",
        )
        self.assertEqual(env["SHAPE_MODE"], "inline")
        self.assertEqual(env["PASS_SHAPE_FILE"], "1")
        self.assertIn("test_fused_marlin_moe.py::test_fused_marlin_moe_mxfp4",
                      env["TEST_FILE"])
    def test_explicit_shape_file_wins(self):
        env = emit_run_env(
            test_node=self._NODE,
            shape_file="/tmp/my_shapes.yaml",
        )
        self.assertEqual(env["SHAPE_MODE"], "file")
        self.assertEqual(env["PASS_SHAPE_FILE"], "1")
    def test_resolve_uses_existing_local_root(self):
        with tempfile.TemporaryDirectory() as tmp:
            root = Path(tmp) / "FlagGems"
            py = root / "benchmark" / "test_fused_marlin_moe.py"
            py.parent.mkdir(parents=True)
            py.write_text("# stub\n", encoding="utf-8")
            env = emit_run_env(
                op="fused_marlin_moe_mxfp4",
                op_file="fused_marlin_moe",
                flaggems_dir=str(root),
            )
            self.assertEqual(env["FLAGGEMS_DIR"], str(root))
            self.assertTrue(Path(env["TEST_FILE"].split("::", 1)[0]).is_file())
 class PluginPlanTests(unittest.TestCase):
    def _plan(self, backend, **kwargs):
        kwargs.setdefault("extra", [])
        kwargs.setdefault("enable_autotune", True)
        kwargs.setdefault("enable_ir", True)
        kwargs.setdefault("enable_mm_cluster", False)
        return plan_plugins(backend, **kwargs)
    def test_cuda_does_not_put_cudagraph_on_the_list(self):
        names = self._plan("cuda")
        self.assertIn("_timing_plugin", names)
        self.assertIn("_device_setup_plugin", names)
        self.assertNotIn("_cudagraph_plugin", names)
        self.assertNotIn("_mm_cluster_fix_plugin", names)
    def test_musa_same_core_without_graph_or_cluster(self):
        names = self._plan("musa")
        self.assertIn("_timing_plugin", names)
        self.assertNotIn("_cudagraph_plugin", names)
        self.assertNotIn("_mm_cluster_fix_plugin", names)
    def test_mm_cluster_ignored_on_non_cuda(self):
        names = self._plan("musa", enable_mm_cluster=True)
        self.assertNotIn("_mm_cluster_fix_plugin", names)
        names = self._plan("cuda", enable_mm_cluster=True)
        self.assertIn("_mm_cluster_fix_plugin", names)
    def test_extra_plugins_isolated(self):
        names = self._plan("cuda", extra=["_cudagraph_plugin"])
        self.assertIn("_cudagraph_plugin", names)
        names = self._plan("cuda", extra=[])
        self.assertNotIn("_cudagraph_plugin", names)
    def test_extra_none_reads_env(self):
        old = os.environ.get("ZL_EXTRA_PLUGINS")
        os.environ["ZL_EXTRA_PLUGINS"] = "_cudagraph_plugin"
        try:
            names = plan_plugins(
                "cuda", extra=None, enable_autotune=True,
                enable_ir=True, enable_mm_cluster=False)
            self.assertIn("_cudagraph_plugin", names)
        finally:
            if old is None:
                os.environ.pop("ZL_EXTRA_PLUGINS", None)
            else:
                os.environ["ZL_EXTRA_PLUGINS"] = old
 class DeviceArgTests(unittest.TestCase):
    def test_devices_preferred(self):
        self.assertEqual(resolve_device_arg("0,1", None), "0,1")
        self.assertEqual(resolve_device_arg(None, "2"), "2")
        self.assertIsNone(resolve_device_arg(None, None))
    def test_both_set_is_error(self):
        with self.assertRaises(ValueError):
            resolve_device_arg("0", "1")
 class CsvFieldTests(unittest.TestCase):
    def test_status_fields(self):
        for name in ("backend", "device", "gpu", "timer"):
            self.assertIn(name, STATUS_FIELDS)
    def test_summary_fields(self):
        for name in ("backend", "device", "timer"):
            self.assertIn(name, SUMMARY_FIELDS)
        self.assertLess(SUMMARY_FIELDS.index("backend"),
                        SUMMARY_FIELDS.index("fg_op_name"))
 class IdentityMergeTests(unittest.TestCase):
    def test_init_then_update(self):
        old_out = os.environ.get("OUT_DIR")
        try:
            with tempfile.TemporaryDirectory() as tmp:
                init_identity(tmp, backend="cuda", device="0", op="softmax")
                update_identity(tmp, timer="direct")
                ident = load_identity(tmp)
                self.assertEqual(ident["schema_version"], 1)
                self.assertEqual(ident["provider"], "flaggems")
                self.assertEqual(ident["backend"], "cuda")
                self.assertEqual(ident["device"], "0")
                self.assertEqual(ident["timer"], "direct")
                self.assertEqual(ident["op"], "softmax")
        finally:
            if old_out is None:
                os.environ.pop("OUT_DIR", None)
            else:
                os.environ["OUT_DIR"] = old_out
 class AutotuneReplayBackendTests(unittest.TestCase):
    def setUp(self):
        self._old = os.environ.get("ZL_BACKEND")
        os.environ["ZL_BACKEND"] = "cuda"
    def tearDown(self):
        if self._old is None:
            os.environ.pop("ZL_BACKEND", None)
        else:
            os.environ["ZL_BACKEND"] = self._old
    def test_missing_meta_still_replays(self):
        from _autotune_record_plugin import _replay_backend_ok
        with tempfile.TemporaryDirectory() as tmp:
            self.assertTrue(_replay_backend_ok(tmp))
    def test_matching_backend_ok(self):
        from _autotune_record_plugin import _replay_backend_ok
        with tempfile.TemporaryDirectory() as tmp:
            Path(tmp, "meta.json").write_text(
                json.dumps({"backend": "cuda"}), encoding="utf-8")
            self.assertTrue(_replay_backend_ok(tmp))
    def test_mismatch_refuses_replay(self):
        from _autotune_record_plugin import _replay_backend_ok
        with tempfile.TemporaryDirectory() as tmp:
            Path(tmp, "meta.json").write_text(
                json.dumps({"backend": "musa"}), encoding="utf-8")
            self.assertFalse(_replay_backend_ok(tmp))
 if __name__ == "__main__":
    unittest.main()
Author	SHA1	Message	Date
Aharrypotter	abf9475ff3	feat: report a per-operator speedup range across two runs The speedup column in a result table compares gems against torch, not one build against another, so an A/B needs the two runs paired by shape. batch_summary gains an --ab mode that indexes both run.logs by operator, dtype and size detail, computes before/after per pair, and prints the per-shape rows plus the range. Only rows that succeeded on both sides with a positive latency count, so an incomplete pairing cannot masquerade as a complete range. When either side carries a BENCHMARK_DIRECT_NO_CUDAGRAPH marker the range is annotated, because a fallback row was timed under a different regime than a captured one. ab_rlc_test.sh drives the pairing for the FlagTree RemoveLayoutConversions switch the way ab_fold_test.sh does for the FlagGems one: per operator, run with the enhancement off while recording autotune configs, then replay those configs with it on, so the two sides differ by that switch alone. Each operator therefore leaves exactly two run.log files, and the summary of each pair is printed as it completes. Checked on an S5000 against gelu_backward over five shapes: both sides captured with no fallback, five rows paired, range -0.43% to +0.09%.	2026-09-02 12:19:26 +08:00
Aharrypotter	7522143747	feat: drop stale CUDA-only skip guards on MThreads Several upstream FlagGems tests gate themselves on torch.cuda.is_available() even though the code under test runs on whatever accelerator FlagGems dispatches to. On MThreads those guards skip the operator instead of exercising it, so a working path reports as "not measured". A guard is dropped only when it is pinned to one file and one exact skip reason, only when FlagGems has dispatched to the MThreads vendor, and only after the unskipped test was observed to pass on an S5000. The plugin walks the node chain rather than the item's own markers, because a module-level pytestmark does not appear on the item. Two guards were probed on device and deliberately left in place, which is why they are named in the docstring rather than silently absent. test_grid_sample.py builds its inputs through a helper whose device argument defaults to the literal "cuda": unskipped, all 90 collected cases fail in aten::empty on a CUDA backend that is not present. test_cutlass_scaled_mm.py gates on vLLM as well as CUDA, so removing its guard would raise on the missing dependency instead of measuring anything.	2026-09-02 12:19:26 +08:00
Aharrypotter	3c9529dab2	fix: take graph-capture timing from Triton's device interface The cudagraph plugin swapped do_bench for do_bench_cudagraph, but upstream's do_bench_cudagraph names torch.cuda directly: torch.cuda.stream, Event, synchronize, CUDAGraph, graph. FlagTree's do_bench already resolves the device through driver.active.get_device_interface(), its cudagraph twin does not. On a non-CUDA backend every capture therefore raised and every measurement silently fell back to plain do_bench, so the plugin looked enabled while nothing was being captured. The plugin now resolves the interface itself and takes the stream, event, synchronize and graph calls from it. CUDA still calls upstream's do_bench_cudagraph verbatim, so its numbers do not move; other backends get an otherwise step-for-step port. The scratch-RNG guard around capture is skipped when the interface exposes no default_generators, which is the case on MThreads and where there is no generator registration to protect against. The replay timings are handed to _summarize_statistics as a plain list of floats, matching do_bench and upstream's cudagraph path. This fork's helper runs Python's statistics module over them, so wrapping them in a tensor makes "mean" raise and makes a quantile request return tensors that later fail to serialize. Verified on a Moore Threads S5000 through torch_musa, which supplies Stream, Event, synchronize, graph and MUSAGraph. softmax and gelu_backward capture and replay with no fallback marker. RNG operators cannot be captured there yet: dropout aborts with "Cannot call MUSAGeneratorImpl::current_seed during MUSA graph capture" and takes the documented BENCHMARK_DIRECT_NO_CUDAGRAPH fallback, so a corpus mixing RNG and non-RNG operators mixes two timing regimes and the marker is the thing to grep before comparing them.	2026-09-02 12:19:07 +08:00
Aharrypotter	64baa11a23	chore: pin shipped text formats to LF The drivers here are shell scripts executed on Linux accelerator hosts, and consumers record the SHA-256 of the files they ran. With Git's default autocrlf on Windows a checkout rewrites those files with CRLF, which breaks the shell drivers outright and changes every digest a receipt refers to.	2026-09-02 12:19:06 +08:00
		`@@ -1 +0,0 @@`
			`"""Benchmark providers. Phase 1 only implements FlagGems."""`