[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78ZTPJBCDAK2PC81CSF08AS":3,"skill-same:01M78ZTPJBCDAK2PC81CSF08AS":66},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":32,"install":33,"resources":45},"01M78ZTPJBCDAK2PC81CSF08AS","training-evaluation",null,"Guides safe MapTR training and chamfer-evaluation command construction, preflight validation, checkpoint handling, reproducibility, and resource-aware recovery.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"MIT","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: training-evaluation\ndescription: \"Guides safe MapTR training and chamfer-evaluation command\n  construction, preflight validation, checkpoint handling, reproducibility, and\n  resource-aware recovery.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: MIT\n---\n\n# MapTR Training And Evaluation\n\nUse this sub-skill to construct and review commands for MapTR training or\nvector-map evaluation. It is deliberately conservative: prove inputs and\nbackend readiness before starting GPU work, prefer dry runs, and never infer\nthat a documented command succeeded on the current system.\n\n## Scope And Evidence Status\n\nThis guidance covers the MapTR `main` behavior represented by commit\n`a6872d8d9670bde17b4b01560f1221f88b443d55`:\n\n- single-process and one-node distributed training;\n- one-node distributed test\u002Fevaluation with the MapTR `chamfer` metric;\n- work directories, seeds, deterministic mode, resume, FP16 configuration,\n  checkpoints, result directories, and failure recovery.\n\nThe project documents Python 3.8, PyTorch 1.9.1 with CUDA 11.1,\n`mmcv-full==1.4.0`, `mmdet==2.14.0`, `mmsegmentation==0.14.1`, the bundled\nmmdetection3d 0.17.2 code, and a compiled GeometryKernelAttention extension.\nThose are **documented compatibility targets**, not proof that a current host\ncan import the plugin or run its CUDA kernels. Full data, checkpoints,\ntraining, evaluation, and the custom extension were not executed while this\nskill was produced.\n\n## Route First\n\n- For nuScenes\u002FAV2 download, conversion, annotation files, and directory\n  layouts, route to `data-preparation`.\n- For architecture selection, BEV\u002Fquery settings, losses, or config authoring,\n  route to `model-configuration`.\n- For image\u002Fvideo rendering, qualitative output, or log plotting, route to\n  `visualization-benchmarking`.\n- Stay here for command semantics, launch sizing, metric selection,\n  checkpoint\u002Foutput behavior, and run-time troubleshooting.\n\n## Required Run Contract\n\nBefore constructing a launch, obtain all of these values. Keep an unknown as a\nstop condition rather than filling it with a guess.\n\n1. **Intent:** train, resume training, or evaluate.\n2. **Config:** exact MapTR config and any `--cfg-options` overrides.\n3. **Data:** dataset family and proof that every configured annotation\u002Fmap file\n   and camera asset exists.\n4. **Weights:** required backbone initialization for a new run, or the exact\n   resume\u002Fevaluation checkpoint.\n5. **Backend:** compatible legacy OpenMMLab stack, CUDA, `mmcv-full` ops, and\n   GeometryKernelAttention import\u002Fbuild evidence.\n6. **Resources:** visible GPU count, memory per GPU, requested process count,\n   storage, wall-time, and scheduler rules.\n7. **Outputs:** unique `--work-dir`, checkpoint retention plan, and evaluation\n   result location.\n8. **Reproducibility:** seed, deterministic policy, final merged config, source\n   revision, and environment inventory.\n\nRead [workflows.md](references\u002Fworkflows.md) for complete gates, commands, and\nexpected observations. Read [cli-reference.md](references\u002Fcli-reference.md)\nbefore adding flags or overrides.\n\n## Safe Command Procedure\n\n### 1. Stop Before Compute\n\nFrom the MapTR project root, run only parser\u002Fconfig checks first:\n\n```bash\npython tools\u002Ftrain.py --help\npython tools\u002Ftest.py --help\npython tools\u002Fmisc\u002Fprint_config.py projects\u002Fconfigs\u002Fmaptr\u002Fmaptr_tiny_r50_24e.py\n```\n\nThe config print should show `plugin=True`, MapTR dataset\u002Fmodel types,\n`evaluation.metric='chamfer'`, and resolved data\u002Fcheckpoint paths. Printing a\nconfig does not import the plugin, read all data, build CUDA extensions, or\nprove training readiness.\n\nDo not proceed when a path is missing, an override is unclear, requested GPUs\nexceed visible GPUs, the config\u002Fcheckpoint pairing is unverified, or a required\nCUDA\u002Fcustom-op probe has not passed.\n\n### 2. Select The Smallest Launch\n\nFor a supported single-process training attempt:\n\n```bash\npython tools\u002Ftrain.py CONFIG.py \\\n  --work-dir work_dirs\u002FRUN_ID \\\n  --gpus 1 --seed 42 --deterministic --launcher none\n```\n\nFor one-node distributed training, use the bundled validator from this\nsub-skill directory. It prints a command by default and starts nothing:\n\n```bash\npython scripts\u002Flaunch_distributed.py train \\\n  --project-root MAPTR_ROOT --config projects\u002Fconfigs\u002Fmaptr\u002FCONFIG.py \\\n  --gpus N -- --work-dir work_dirs\u002FRUN_ID --seed 42\n```\n\nFor evaluation, use the same validator even with one GPU. The checked-in test\nentry point intentionally rejects `--launcher none`:\n\n```bash\npython scripts\u002Flaunch_distributed.py test \\\n  --project-root MAPTR_ROOT --config projects\u002Fconfigs\u002Fmaptr\u002FCONFIG.py \\\n  --checkpoint work_dirs\u002FRUN_ID\u002Fepoch_N.pth --gpus N\n```\n\nThe test command appends `--eval chamfer`. Never substitute the generic\n3D-detection `bbox` launcher for MapTR. The vector-map datasets accept\n`chamfer` and `iou`; project train\u002Feval instructions and MapTR configs use\n`chamfer`.\n\nReview [launch_distributed.py](scripts\u002Flaunch_distributed.py), then add\n`--execute` only after its dry-run output and warnings pass human review.\nExecution can allocate GPUs and run for hours.\n\n### 3. Preserve Run Evidence\n\nA training work directory should receive the merged config, timestamped text\nlog, logger outputs, and checkpoints such as `epoch_1.pth` according to\n`checkpoint_config.interval`. MapTR 24-epoch configs normally evaluate every\ntwo epochs and save every epoch. Confirm actual config values rather than\nassuming those intervals for another config.\n\nChamfer evaluation should print per-class AP at thresholds 0.5, 1.0, and 1.5,\nplus `NuscMap_chamfer\u002FmAP`-style detail keys, and write formatted predictions\nbelow a timestamped `test\u002F\u003Cconfig-name>\u002F` directory. Treat those as expected\nshapes only, not benchmark claims.\n\nRecord the exact command, `CUDA_VISIBLE_DEVICES`, seed, deterministic flag,\nmerged config, checkpoint identity, environment versions, and complete log.\nDo not compare scores from different configs, class maps, point sampling, data\nversions, or checkpoints as if they were equivalent.\n\n## High-Risk Semantics\n\n- `--resume-from` restores training state; a nonexistent path is silently not\n  applied by the training entry point. Validate it before launch.\n- A backbone `pretrained` path initializes a new model; it is not a training\n  resume checkpoint. `load_from` loads weights without promising optimizer and\n  epoch continuity.\n- Most MapTR configs enable static FP16 with `loss_scale=512`. Confirm the\n  merged config and backend support. Do not add a second AMP mechanism.\n- The distributed training wrapper forces deterministic mode. Determinism may\n  reduce speed and does not guarantee bitwise identity across different GPU,\n  driver, CUDA, library, or process-count combinations.\n- `tools\u002Ftest.py --out` is parsed but reaches an intentional assertion instead\n  of dumping predictions. `--show`\u002F`--show-dir` are not forwarded through its\n  distributed test path. Route artifact visualization elsewhere.\n\n## Recovery\n\nUse [troubleshooting.md](references\u002Ftroubleshooting.md) for symptoms, causes,\nrecovery, and explicit stop conditions. In particular:\n\n- `metric bbox is not supported` → replace `--eval bbox` with\n  `--eval chamfer`; do not reinterpret a detector metric as a map metric.\n- eight requested processes with one visible GPU → stop and use `--gpus 1`, or\n  move to a proven eight-GPU host; never oversubscribe silently.\n- resume\u002Fconfig mismatch → stop, locate the config dumped beside the\n  checkpoint, reconcile its model\u002Fdata\u002Fclass\u002Foptimizer\u002Frunner settings, and\n  resume with that config or start a clearly new run.\n\nDo not run merely to discover whether a prerequisite was satisfied. If\nconfiguration, data, checkpoint provenance, backend compatibility, GPU count,\nor budget remains unresolved, return a blocked preflight with the missing\nevidence and the next safe probe.\n","# MapTR Training And Evaluation\n\nUse this sub-skill to construct and review commands for MapTR training or\nvector-map evaluation. It is deliberately conservative: prove inputs and\nbackend readiness before starting GPU work, prefer dry runs, and never infer\nthat a documented command succeeded on the current system.\n\n## Scope And Evidence Status\n\nThis guidance covers the MapTR `main` behavior represented by commit\n`a6872d8d9670bde17b4b01560f1221f88b443d55`:\n\n- single-process and one-node distributed training;\n- one-node distributed test\u002Fevaluation with the MapTR `chamfer` metric;\n- work directories, seeds, deterministic mode, resume, FP16 configuration,\n  checkpoints, result directories, and failure recovery.\n\nThe project documents Python 3.8, PyTorch 1.9.1 with CUDA 11.1,\n`mmcv-full==1.4.0`, `mmdet==2.14.0`, `mmsegmentation==0.14.1`, the bundled\nmmdetection3d 0.17.2 code, and a compiled GeometryKernelAttention extension.\nThose are **documented compatibility targets**, not proof that a current host\ncan import the plugin or run its CUDA kernels. Full data, checkpoints,\ntraining, evaluation, and the custom extension were not executed while this\nskill was produced.\n\n## Route First\n\n- For nuScenes\u002FAV2 download, conversion, annotation files, and directory\n  layouts, route to `data-preparation`.\n- For architecture selection, BEV\u002Fquery settings, losses, or config authoring,\n  route to `model-configuration`.\n- For image\u002Fvideo rendering, qualitative output, or log plotting, route to\n  `visualization-benchmarking`.\n- Stay here for command semantics, launch sizing, metric selection,\n  checkpoint\u002Foutput behavior, and run-time troubleshooting.\n\n## Required Run Contract\n\nBefore constructing a launch, obtain all of these values. Keep an unknown as a\nstop condition rather than filling it with a guess.\n\n1. **Intent:** train, resume training, or evaluate.\n2. **Config:** exact MapTR config and any `--cfg-options` overrides.\n3. **Data:** dataset family and proof that every configured annotation\u002Fmap file\n   and camera asset exists.\n4. **Weights:** required backbone initialization for a new run, or the exact\n   resume\u002Fevaluation checkpoint.\n5. **Backend:** compatible legacy OpenMMLab stack, CUDA, `mmcv-full` ops, and\n   GeometryKernelAttention import\u002Fbuild evidence.\n6. **Resources:** visible GPU count, memory per GPU, requested process count,\n   storage, wall-time, and scheduler rules.\n7. **Outputs:** unique `--work-dir`, checkpoint retention plan, and evaluation\n   result location.\n8. **Reproducibility:** seed, deterministic policy, final merged config, source\n   revision, and environment inventory.\n\nRead [workflows.md](references\u002Fworkflows.md) for complete gates, commands, and\nexpected observations. Read [cli-reference.md](references\u002Fcli-reference.md)\nbefore adding flags or overrides.\n\n## Safe Command Procedure\n\n### 1. Stop Before Compute\n\nFrom the MapTR project root, run only parser\u002Fconfig checks first:\n\n```bash\npython tools\u002Ftrain.py --help\npython tools\u002Ftest.py --help\npython tools\u002Fmisc\u002Fprint_config.py projects\u002Fconfigs\u002Fmaptr\u002Fmaptr_tiny_r50_24e.py\n```\n\nThe config print should show `plugin=True`, MapTR dataset\u002Fmodel types,\n`evaluation.metric='chamfer'`, and resolved data\u002Fcheckpoint paths. Printing a\nconfig does not import the plugin, read all data, build CUDA extensions, or\nprove training readiness.\n\nDo not proceed when a path is missing, an override is unclear, requested GPUs\nexceed visible GPUs, the config\u002Fcheckpoint pairing is unverified, or a required\nCUDA\u002Fcustom-op probe has not passed.\n\n### 2. Select The Smallest Launch\n\nFor a supported single-process training attempt:\n\n```bash\npython tools\u002Ftrain.py CONFIG.py \\\n  --work-dir work_dirs\u002FRUN_ID \\\n  --gpus 1 --seed 42 --deterministic --launcher none\n```\n\nFor one-node distributed training, use the bundled validator from this\nsub-skill directory. It prints a command by default and starts nothing:\n\n```bash\npython scripts\u002Flaunch_distributed.py train \\\n  --project-root MAPTR_ROOT --config projects\u002Fconfigs\u002Fmaptr\u002FCONFIG.py \\\n  --gpus N -- --work-dir work_dirs\u002FRUN_ID --seed 42\n```\n\nFor evaluation, use the same validator even with one GPU. The checked-in test\nentry point intentionally rejects `--launcher none`:\n\n```bash\npython scripts\u002Flaunch_distributed.py test \\\n  --project-root MAPTR_ROOT --config projects\u002Fconfigs\u002Fmaptr\u002FCONFIG.py \\\n  --checkpoint work_dirs\u002FRUN_ID\u002Fepoch_N.pth --gpus N\n```\n\nThe test command appends `--eval chamfer`. Never substitute the generic\n3D-detection `bbox` launcher for MapTR. The vector-map datasets accept\n`chamfer` and `iou`; project train\u002Feval instructions and MapTR configs use\n`chamfer`.\n\nReview [launch_distributed.py](scripts\u002Flaunch_distributed.py), then add\n`--execute` only after its dry-run output and warnings pass human review.\nExecution can allocate GPUs and run for hours.\n\n### 3. Preserve Run Evidence\n\nA training work directory should receive the merged config, timestamped text\nlog, logger outputs, and checkpoints such as `epoch_1.pth` according to\n`checkpoint_config.interval`. MapTR 24-epoch configs normally evaluate every\ntwo epochs and save every epoch. Confirm actual config values rather than\nassuming those intervals for another config.\n\nChamfer evaluation should print per-class AP at thresholds 0.5, 1.0, and 1.5,\nplus `NuscMap_chamfer\u002FmAP`-style detail keys, and write formatted predictions\nbelow a timestamped `test\u002F\u003Cconfig-name>\u002F` directory. Treat those as expected\nshapes only, not benchmark claims.\n\nRecord the exact command, `CUDA_VISIBLE_DEVICES`, seed, deterministic flag,\nmerged config, checkpoint identity, environment versions, and complete log.\nDo not compare scores from different configs, class maps, point sampling, data\nversions, or checkpoints as if they were equivalent.\n\n## High-Risk Semantics\n\n- `--resume-from` restores training state; a nonexistent path is silently not\n  applied by the training entry point. Validate it before launch.\n- A backbone `pretrained` path initializes a new model; it is not a training\n  resume checkpoint. `load_from` loads weights without promising optimizer and\n  epoch continuity.\n- Most MapTR configs enable static FP16 with `loss_scale=512`. Confirm the\n  merged config and backend support. Do not add a second AMP mechanism.\n- The distributed training wrapper forces deterministic mode. Determinism may\n  reduce speed and does not guarantee bitwise identity across different GPU,\n  driver, CUDA, library, or process-count combinations.\n- `tools\u002Ftest.py --out` is parsed but reaches an intentional assertion instead\n  of dumping predictions. `--show`\u002F`--show-dir` are not forwarded through its\n  distributed test path. Route artifact visualization elsewhere.\n\n## Recovery\n\nUse [troubleshooting.md](references\u002Ftroubleshooting.md) for symptoms, causes,\nrecovery, and explicit stop conditions. In particular:\n\n- `metric bbox is not supported` → replace `--eval bbox` with\n  `--eval chamfer`; do not reinterpret a detector metric as a map metric.\n- eight requested processes with one visible GPU → stop and use `--gpus 1`, or\n  move to a proven eight-GPU host; never oversubscribe silently.\n- resume\u002Fconfig mismatch → stop, locate the config dumped beside the\n  checkpoint, reconcile its model\u002Fdata\u002Fclass\u002Foptimizer\u002Frunner settings, and\n  resume with that config or start a clearly new run.\n\nDo not run merely to discover whether a prerequisite was satisfied. If\nconfiguration, data, checkpoint provenance, backend compatibility, GPU count,\nor budget remains unresolved, return a blocked preflight with the missing\nevidence and the next safe probe.\n",{"name":5,"license":14,"metadata":23,"description":7,"disable-model-invocation":25},{"disco-role":24},"operating",true,"main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","skills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":25},[35,38,41],{"tool":36,"command":37},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill training-evaluation",{"tool":39,"command":40},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill training-evaluation",{"tool":42,"command":43},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill training-evaluation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002FSKILL.md",[46,52,56,60],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fcli-reference.md","REFERENCE","参考文档",8227,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002Freferences\u002Fcli-reference.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Ftroubleshooting.md",10211,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002Freferences\u002Ftroubleshooting.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Fworkflows.md",12236,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002Freferences\u002Fworkflows.md",{"path":61,"resourceType":62,"resourceTypeDesc":63,"sizeBytes":64,"sourceHtmlUrl":65},"scripts\u002Flaunch_distributed.py","SCRIPT","脚本",8923,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Fmaptr\u002Fsub-skills\u002Ftraining-evaluation\u002Fscripts\u002Flaunch_distributed.py",[67,75,82,89,96,103],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":18},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.","Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":18},"01M799NS5APEC0N1FZ7XZVW6BN","passnet-solve","END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":18},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":18},"01M799NHH6NP7Z1RK77KGXAN6S","passnet-pattern-fusion","Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":97,"name":98,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":100,"categories":101,"keywords":102,"publishedAt":18},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":104,"name":105,"nameZh":6,"description":106,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":71,"licenseFileUrl":6,"sourceHtmlUrl":107,"categories":108,"keywords":109,"publishedAt":18},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]