[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7A3Y2W7JW04GV87528S6RJS":3,"skill-same:01M7A3Y2W7JW04GV87528S6RJS":73},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":37,"commitSha":38,"skillDirectory":39,"forksCount":40,"repositoryHtmlUrl":41,"repositoryHomepageUrl":6,"ownerHtmlUrl":42,"ownerBlogUrl":6,"ownerDescription":43,"updatedAt":44,"install":45,"resources":58},"01M7A3Y2W7JW04GV87528S6RJS","cdf-anomalygen-release",null,"Use when building the cdf-anomalygen Docker image (develop \u002F product \u002F air-gapped) from docker\u002FDockerfile — build the chosen target from source and run it. Not for training or synthetic defect-image generation (SDG).","leemiracle\u002Fwork4ai","leemiracle","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F10385913?v=4","01M79YCHARF56CHMGGTECMZN8Z","lee","01M79YCHBY9VH2CJW4P65D3DHR",4,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: cdf-anomalygen-release\ndescription: >-\n  Use when building the cdf-anomalygen Docker image (develop \u002F product \u002F\n  air-gapped) from docker\u002FDockerfile — build the chosen target from source and\n  run it. Not for training or synthetic defect-image generation (SDG).\nlicense: Apache-2.0\ncompatibility: >-\n  Requires Docker 23+ (BuildKit) + nvidia-container-toolkit and a CUDA GPU. The\n  build compiles the heavy CUDA extensions from source and takes >1 h.\nmetadata:\n  owner: NVIDIA\n  service: docker\n  version: 1.2.0\n  author: NVIDIA \u003Cnvidia@nvidia.com>\n  tags:\n      - physical-ai\n      - docker\n      - release\n      - air-gapped\n  languages: [shell]\n  frameworks: [docker]\n---\n\n# Skill: cdf-anomalygen-release\n\n## Purpose\n\nBuild and run the **cdf-anomalygen** container. `docker\u002FDockerfile` reproduces\n`scripts\u002Fenv_setup.sh` as a two-stage image (compile heavy wheels → slim non-root runtime).\nThis is a **release\u002Fbuild** workflow — not training or inference\n(use the `cdf-anomalygen` skill for those). Build targets, GPU-arch lists, and troubleshooting\ndetail live in `docker\u002FREADME.md`. Run every command from the repo root.\n\n> **Check whether a build is needed at all.** Prebuilt images are published on\n> [NGC](https:\u002F\u002Fcatalog.ngc.nvidia.com\u002Forgs\u002Fnvidia\u002Fcontainers\u002Fcdf-anomalygen):\n> `docker pull nvcr.io\u002Fnvidia\u002Fcdf-anomalygen:1.2.0` needs no login and no compile.\n> Build only when the user wants a modified image, a target NGC does not publish, or an\n> air-gapped bundle. **A build from source takes >1 h** — say so before starting one.\n\n## When to Use This Skill\n\nUse this skill when the user wants to build or ship the anomalygen container —\nbuild a **develop** \u002F **product** \u002F **air-gapped** image, push it to a registry, or\nrun it. E.g. \"build the anomalygen product container\", \"build an air-gapped develop\nimage\".\n\nDo not use it to run training or SDG generation — that's the `cdf-anomalygen` skill.\n\n## Prerequisites\n\n- **Docker 23+** (BuildKit on by default) + **nvidia-container-toolkit** + a CUDA GPU.\n- Enough headroom for the source compile: the heavy extensions are memory-hungry and the\n  build needs tens of GB of disk for intermediate layers.\n- The repo **`.venv`** (`scripts\u002Fenv_setup.sh`: py3.13, torch 2.13.0+cu132) is **optional** —\n  it is used only to auto-derive `TORCH_CUDA_ARCH_LIST` below, via\n  `torch.cuda.get_arch_list()`. Without it the documented default is used.\n- **Air-gapped only:** the model checkpoints under `.\u002Fcheckpoints\u002F`. `scripts\u002Fdownload_checkpoints.sh`\n  populates this directory.\n\n## Targets\n\n`docker\u002FDockerfile` exposes four targets (`--target`, default `develop`):\n\n| Target              | What                                                                          |\n| ------------------- | ----------------------------------------------------------------------------- |\n| `develop` (default) | non-root interactive image + dev tooling (`pytest`, `ruff`, `pre-commit`)     |\n| `product`           | non-root, app code **read-only**, `ANOMALYGEN_PRODUCT_MODE=1`, no dev tooling |\n| `airgapped-develop` | `develop` + the checkpoints **baked in** for offline runs                     |\n| `airgapped-product` | `product` + the checkpoints **baked in** for offline runs                     |\n\n## Instructions\n\n**Default flow: Step 1, then stop with the built image.** Push (Step 2) and run\n(Step 3) are opt-in — only do them when the user explicitly asks to publish or run the image.\n\nSet the shared variables once (arm64 host → use the commented values):\n\n```shell\nexport TARGET=develop                       # develop | product | airgapped-develop | airgapped-product\nexport IMAGE=cdf-anomalygen:cuda-13.2.1-${TARGET}-ubuntu24.04-amd64\n# tag = \u003Cbase-cuda>-\u003Ctarget>-\u003Cubuntu>-\u003Carch>; arm64 host → …-arm64. Set TARGET *first*: it feeds both\n# the tag and --target below, so they cannot disagree. Hardcoding \"develop\" here and then building\n# --target product silently ships a product image labelled develop.\nexport PLATFORM=linux\u002Famd64                 # arm64 host: linux\u002Farm64\n# GPU arch (>=8.0), derived from .venv torch when present; forwarded to the build as a --build-arg.\n# The `:-` matters: get_arch_list() is [] with no GPU (and the command fails outright with no .venv),\n# and a bare --build-arg forwards that empty value, overriding the Dockerfile default.\nexport TORCH_CUDA_ARCH_LIST=\"$(.venv\u002Fbin\u002Fpython -c \"import torch; print(' '.join(c for c in sorted({f'{a[3:-1]}.{a[-1]}' for a in torch.cuda.get_arch_list() if a.startswith('sm_')}, key=float) if float(c) >= 8.0))\" 2>\u002Fdev\u002Fnull)\"\nexport TORCH_CUDA_ARCH_LIST=\"${TORCH_CUDA_ARCH_LIST:-8.0 8.6 9.0 10.0 12.0}\"  # arm64: '9.0 10.0 10.3 12.0'\n```\n\n### Step 1 — build the image\n\nStage 1 compiles flash-attn \u002F transformer-engine \u002F apex \u002F flash-attn-3-nv \u002F natten from\nsource. **This takes >1 h** — it is not a hang. `TORCH_CUDA_ARCH_LIST` is *required* on\nthis path: `docker build` never sees a GPU, so nothing can be autodetected.\n\n**Standard** (`develop` \u002F `product`):\n\n```shell\ndocker build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --target \"$TARGET\" -t \"$IMAGE\" .        # TARGET from the shared vars; tag and target stay in sync\n```\n\n**Air-gapped** (`airgapped-develop` \u002F `airgapped-product`) — needs `buildx` + a named\n`ckpts` context (which bypasses `.dockerignore`). The image is self-contained, so save it\nto `.\u002Fresults\u002F` as a gzipped bundle for offline transfer and **report that path to the user**:\n\n```shell\n# --load is required: buildx's default driver keeps the result in its own build cache, so without it\n# the tag never reaches the local daemon and the `docker save` below fails with \"reference does not\n# exist\". (Single-platform builds only — --load cannot export a multi-arch manifest.)\ndocker buildx build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --build-context ckpts=.\u002Fcheckpoints \\\n    --target \"$TARGET\" -t \"$IMAGE\" --load . # set TARGET=airgapped-develop|airgapped-product first\n\nmkdir -p results\nBUNDLE=\"results\u002F$(basename \"$IMAGE\" | tr ':' '_').tar.gz\"\ndocker save \"$IMAGE\" | gzip > \"$BUNDLE\"\necho \"air-gapped bundle: $BUNDLE\"      # tell the user this location\n```\n\n### Step 2 — push (only when the user asks to publish; skip by default)\n\nAsk the user which registry to publish to — there is no default. Tag the local image for\nthat registry, authenticate, then push:\n\n```shell\nexport REGISTRY=\u003Cyour-registry>          # e.g. registry.example.com or an org path\ndocker tag \"$IMAGE\" \"$REGISTRY\u002F$IMAGE\"\ndocker login \"$REGISTRY\"                 # credentials per your registry; --password-stdin in CI\ndocker push \"$REGISTRY\u002F$IMAGE\"\n```\n\n### Step 3 — run (only when the user asks to run the container; skip by default)\n\n`docker run` invocation, the `--user` \u002F `-e USER` pairing that makes `getpass.getuser()` resolve,\n`--shm-size`, `HF_TOKEN`, and loading an air-gapped bundle:\n[`running.md`](references\u002Frunning.md).\n\n## Verification\n\nRun these **after Step 1 has produced an image** — they check that image, and the smoke test needs the\nStep 3 `docker run` invocation, so they are not a standalone entry point.\n\n1. Image exists: `docker images \"$IMAGE\"`.\n2. **product only** — guardrails hold: `ANOMALYGEN_PRODUCT_MODE=1` is set, it runs as\n   the non-root `nvidia` user, and the app code is read-only. If any fails, the\n   image is **not** productized.\n3. Smoke test — in the Step 3 container (repo + `checkpoints\u002F` mounted; baked in for\n   air-gapped), a 20-iteration dry run finishes without error.\n\n   > **The output root must be fresh, or this check is worthless.** A reused root resumes a leftover\n   > checkpoint and finishes instantly, which looks identical to success. The `date +%s` suffix\n   > below is what makes it fresh — keep it, and confirm the run actually logs 20 iterations.\n\n```shell\nIMAGINAIRE_OUTPUT_ROOT=\"$PWD\u002Fresults\u002Fdryrun-$(date +%s)\" torchrun --nproc_per_node=1 \\\n    anomalygen\u002Fscripts\u002Ftexture\u002Ftrain.py --config=cosmos_framework\u002Fconfigs\u002Fbase\u002Fconfig.py \\\n    --recipe=ag_config\u002Fexp_texture_ft_phone_screen.yaml \\\n    -- experiment=anomalygen_texture_ft trainer.max_iter=20 trainer.validation_iter=10 checkpoint.save_iter=10\n```\n\n## Examples\n\n**\"Build the anomalygen product container\"** — shared variables, then Step 1 with the product target;\nstop there (no push unless asked):\n\n```shell\nexport TARGET=product   # re-export IMAGE so the tag follows the target\nexport IMAGE=cdf-anomalygen:cuda-13.2.1-${TARGET}-ubuntu24.04-amd64\ndocker build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --target \"$TARGET\" -t \"$IMAGE\" .\n```\n\n**\"Build an air-gapped develop image\"** — Step 1's `buildx` variant with the `ckpts` named context,\nthen load the saved bundle on the offline host. The checkpoints must already be under\n`.\u002Fcheckpoints\u002F`.\n\n## Troubleshooting\n\n- **\"TORCH_CUDA_ARCH_LIST is empty and could not be derived\"** → something overrode the Dockerfile\n  default with an empty value: a bare `--build-arg` forwards an empty env var, and `get_arch_list()`\n  is `[]` with no GPU. The shared vars guard this with `:-`; otherwise set it by hand (amd64\n  `8.0 8.6 9.0 10.0 12.0`, arm64 `9.0 10.0 10.3 12.0`).\n- **Build seems stuck for an hour** → that is the source compile of flash-attn \u002F TE \u002F apex, which is\n  expected. Don't kill it; watch the layer output instead.\n- **Build OOM** (flash-attn \u002F TE \u002F apex) → lower the job count with\n  `--build-arg MAX_JOBS=…` (default 16).\n- **Missing `--shm-size` (min `16g`)** → \"Bus error\" \u002F silent hangs: the dataloader\n  uses `\u002Fdev\u002Fshm` and Docker's 64 MB default is too small. Always pass it.\n- **`failed to find stage \"…\"`** → bad `--target`; valid: `develop`, `product`,\n  `airgapped-develop`, `airgapped-product`.\n- **Don't build inside a product runtime** (`ANOMALYGEN_PRODUCT_MODE=1`) — build from\n  a clean clone or a develop container.\n","# Skill: cdf-anomalygen-release\n\n## Purpose\n\nBuild and run the **cdf-anomalygen** container. `docker\u002FDockerfile` reproduces\n`scripts\u002Fenv_setup.sh` as a two-stage image (compile heavy wheels → slim non-root runtime).\nThis is a **release\u002Fbuild** workflow — not training or inference\n(use the `cdf-anomalygen` skill for those). Build targets, GPU-arch lists, and troubleshooting\ndetail live in `docker\u002FREADME.md`. Run every command from the repo root.\n\n> **Check whether a build is needed at all.** Prebuilt images are published on\n> [NGC](https:\u002F\u002Fcatalog.ngc.nvidia.com\u002Forgs\u002Fnvidia\u002Fcontainers\u002Fcdf-anomalygen):\n> `docker pull nvcr.io\u002Fnvidia\u002Fcdf-anomalygen:1.2.0` needs no login and no compile.\n> Build only when the user wants a modified image, a target NGC does not publish, or an\n> air-gapped bundle. **A build from source takes >1 h** — say so before starting one.\n\n## When to Use This Skill\n\nUse this skill when the user wants to build or ship the anomalygen container —\nbuild a **develop** \u002F **product** \u002F **air-gapped** image, push it to a registry, or\nrun it. E.g. \"build the anomalygen product container\", \"build an air-gapped develop\nimage\".\n\nDo not use it to run training or SDG generation — that's the `cdf-anomalygen` skill.\n\n## Prerequisites\n\n- **Docker 23+** (BuildKit on by default) + **nvidia-container-toolkit** + a CUDA GPU.\n- Enough headroom for the source compile: the heavy extensions are memory-hungry and the\n  build needs tens of GB of disk for intermediate layers.\n- The repo **`.venv`** (`scripts\u002Fenv_setup.sh`: py3.13, torch 2.13.0+cu132) is **optional** —\n  it is used only to auto-derive `TORCH_CUDA_ARCH_LIST` below, via\n  `torch.cuda.get_arch_list()`. Without it the documented default is used.\n- **Air-gapped only:** the model checkpoints under `.\u002Fcheckpoints\u002F`. `scripts\u002Fdownload_checkpoints.sh`\n  populates this directory.\n\n## Targets\n\n`docker\u002FDockerfile` exposes four targets (`--target`, default `develop`):\n\n| Target              | What                                                                          |\n| ------------------- | ----------------------------------------------------------------------------- |\n| `develop` (default) | non-root interactive image + dev tooling (`pytest`, `ruff`, `pre-commit`)     |\n| `product`           | non-root, app code **read-only**, `ANOMALYGEN_PRODUCT_MODE=1`, no dev tooling |\n| `airgapped-develop` | `develop` + the checkpoints **baked in** for offline runs                     |\n| `airgapped-product` | `product` + the checkpoints **baked in** for offline runs                     |\n\n## Instructions\n\n**Default flow: Step 1, then stop with the built image.** Push (Step 2) and run\n(Step 3) are opt-in — only do them when the user explicitly asks to publish or run the image.\n\nSet the shared variables once (arm64 host → use the commented values):\n\n```shell\nexport TARGET=develop                       # develop | product | airgapped-develop | airgapped-product\nexport IMAGE=cdf-anomalygen:cuda-13.2.1-${TARGET}-ubuntu24.04-amd64\n# tag = \u003Cbase-cuda>-\u003Ctarget>-\u003Cubuntu>-\u003Carch>; arm64 host → …-arm64. Set TARGET *first*: it feeds both\n# the tag and --target below, so they cannot disagree. Hardcoding \"develop\" here and then building\n# --target product silently ships a product image labelled develop.\nexport PLATFORM=linux\u002Famd64                 # arm64 host: linux\u002Farm64\n# GPU arch (>=8.0), derived from .venv torch when present; forwarded to the build as a --build-arg.\n# The `:-` matters: get_arch_list() is [] with no GPU (and the command fails outright with no .venv),\n# and a bare --build-arg forwards that empty value, overriding the Dockerfile default.\nexport TORCH_CUDA_ARCH_LIST=\"$(.venv\u002Fbin\u002Fpython -c \"import torch; print(' '.join(c for c in sorted({f'{a[3:-1]}.{a[-1]}' for a in torch.cuda.get_arch_list() if a.startswith('sm_')}, key=float) if float(c) >= 8.0))\" 2>\u002Fdev\u002Fnull)\"\nexport TORCH_CUDA_ARCH_LIST=\"${TORCH_CUDA_ARCH_LIST:-8.0 8.6 9.0 10.0 12.0}\"  # arm64: '9.0 10.0 10.3 12.0'\n```\n\n### Step 1 — build the image\n\nStage 1 compiles flash-attn \u002F transformer-engine \u002F apex \u002F flash-attn-3-nv \u002F natten from\nsource. **This takes >1 h** — it is not a hang. `TORCH_CUDA_ARCH_LIST` is *required* on\nthis path: `docker build` never sees a GPU, so nothing can be autodetected.\n\n**Standard** (`develop` \u002F `product`):\n\n```shell\ndocker build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --target \"$TARGET\" -t \"$IMAGE\" .        # TARGET from the shared vars; tag and target stay in sync\n```\n\n**Air-gapped** (`airgapped-develop` \u002F `airgapped-product`) — needs `buildx` + a named\n`ckpts` context (which bypasses `.dockerignore`). The image is self-contained, so save it\nto `.\u002Fresults\u002F` as a gzipped bundle for offline transfer and **report that path to the user**:\n\n```shell\n# --load is required: buildx's default driver keeps the result in its own build cache, so without it\n# the tag never reaches the local daemon and the `docker save` below fails with \"reference does not\n# exist\". (Single-platform builds only — --load cannot export a multi-arch manifest.)\ndocker buildx build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --build-context ckpts=.\u002Fcheckpoints \\\n    --target \"$TARGET\" -t \"$IMAGE\" --load . # set TARGET=airgapped-develop|airgapped-product first\n\nmkdir -p results\nBUNDLE=\"results\u002F$(basename \"$IMAGE\" | tr ':' '_').tar.gz\"\ndocker save \"$IMAGE\" | gzip > \"$BUNDLE\"\necho \"air-gapped bundle: $BUNDLE\"      # tell the user this location\n```\n\n### Step 2 — push (only when the user asks to publish; skip by default)\n\nAsk the user which registry to publish to — there is no default. Tag the local image for\nthat registry, authenticate, then push:\n\n```shell\nexport REGISTRY=\u003Cyour-registry>          # e.g. registry.example.com or an org path\ndocker tag \"$IMAGE\" \"$REGISTRY\u002F$IMAGE\"\ndocker login \"$REGISTRY\"                 # credentials per your registry; --password-stdin in CI\ndocker push \"$REGISTRY\u002F$IMAGE\"\n```\n\n### Step 3 — run (only when the user asks to run the container; skip by default)\n\n`docker run` invocation, the `--user` \u002F `-e USER` pairing that makes `getpass.getuser()` resolve,\n`--shm-size`, `HF_TOKEN`, and loading an air-gapped bundle:\n[`running.md`](references\u002Frunning.md).\n\n## Verification\n\nRun these **after Step 1 has produced an image** — they check that image, and the smoke test needs the\nStep 3 `docker run` invocation, so they are not a standalone entry point.\n\n1. Image exists: `docker images \"$IMAGE\"`.\n2. **product only** — guardrails hold: `ANOMALYGEN_PRODUCT_MODE=1` is set, it runs as\n   the non-root `nvidia` user, and the app code is read-only. If any fails, the\n   image is **not** productized.\n3. Smoke test — in the Step 3 container (repo + `checkpoints\u002F` mounted; baked in for\n   air-gapped), a 20-iteration dry run finishes without error.\n\n   > **The output root must be fresh, or this check is worthless.** A reused root resumes a leftover\n   > checkpoint and finishes instantly, which looks identical to success. The `date +%s` suffix\n   > below is what makes it fresh — keep it, and confirm the run actually logs 20 iterations.\n\n```shell\nIMAGINAIRE_OUTPUT_ROOT=\"$PWD\u002Fresults\u002Fdryrun-$(date +%s)\" torchrun --nproc_per_node=1 \\\n    anomalygen\u002Fscripts\u002Ftexture\u002Ftrain.py --config=cosmos_framework\u002Fconfigs\u002Fbase\u002Fconfig.py \\\n    --recipe=ag_config\u002Fexp_texture_ft_phone_screen.yaml \\\n    -- experiment=anomalygen_texture_ft trainer.max_iter=20 trainer.validation_iter=10 checkpoint.save_iter=10\n```\n\n## Examples\n\n**\"Build the anomalygen product container\"** — shared variables, then Step 1 with the product target;\nstop there (no push unless asked):\n\n```shell\nexport TARGET=product   # re-export IMAGE so the tag follows the target\nexport IMAGE=cdf-anomalygen:cuda-13.2.1-${TARGET}-ubuntu24.04-amd64\ndocker build --platform \"$PLATFORM\" -f docker\u002FDockerfile \\\n    --build-arg TORCH_CUDA_ARCH_LIST \\\n    --target \"$TARGET\" -t \"$IMAGE\" .\n```\n\n**\"Build an air-gapped develop image\"** — Step 1's `buildx` variant with the `ckpts` named context,\nthen load the saved bundle on the offline host. The checkpoints must already be under\n`.\u002Fcheckpoints\u002F`.\n\n## Troubleshooting\n\n- **\"TORCH_CUDA_ARCH_LIST is empty and could not be derived\"** → something overrode the Dockerfile\n  default with an empty value: a bare `--build-arg` forwards an empty env var, and `get_arch_list()`\n  is `[]` with no GPU. The shared vars guard this with `:-`; otherwise set it by hand (amd64\n  `8.0 8.6 9.0 10.0 12.0`, arm64 `9.0 10.0 10.3 12.0`).\n- **Build seems stuck for an hour** → that is the source compile of flash-attn \u002F TE \u002F apex, which is\n  expected. Don't kill it; watch the layer output instead.\n- **Build OOM** (flash-attn \u002F TE \u002F apex) → lower the job count with\n  `--build-arg MAX_JOBS=…` (default 16).\n- **Missing `--shm-size` (min `16g`)** → \"Bus error\" \u002F silent hangs: the dataloader\n  uses `\u002Fdev\u002Fshm` and Docker's 64 MB default is too small. Always pass it.\n- **`failed to find stage \"…\"`** → bad `--target`; valid: `develop`, `product`,\n  `airgapped-develop`, `airgapped-product`.\n- **Don't build inside a product runtime** (`ANOMALYGEN_PRODUCT_MODE=1`) — build from\n  a clean clone or a develop container.\n",{"name":5,"license":15,"metadata":24,"description":7,"compatibility":36},{"tags":25,"owner":30,"author":31,"service":27,"version":32,"languages":33,"frameworks":35},[26,27,28,29],"physical-ai","docker","release","air-gapped","NVIDIA","NVIDIA \u003Cnvidia@nvidia.com>","1.2.0",[34],"shell",[27],"Requires Docker 23+ (BuildKit) + nvidia-container-toolkit and a CUDA GPU. The build compiles the heavy CUDA extensions from source and takes >1 h.","main","ffa016d3c9a8d32821f71f6d9d916c30e04fcb67","讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release",0,"https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai","https:\u002F\u002Fgithub.com\u002Fleemiracle","i have some experience that make me better, so i want more. follow and star me would be a good choice.","2026-11-12 17:16:40",{"source":8,"skillName":5,"commands":46,"prompt":56,"zipAvailable":57},[47,50,53],{"tool":48,"command":49},"npx","npx skills add leemiracle\u002Fwork4ai --skill cdf-anomalygen-release",{"tool":51,"command":52},"bunx","bunx skills add leemiracle\u002Fwork4ai --skill cdf-anomalygen-release",{"tool":54,"command":55},"pnpm","pnpm dlx skills add leemiracle\u002Fwork4ai --skill cdf-anomalygen-release","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release\u002FSKILL.md",true,[59,65,69],{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"BENCHMARK.md","REFERENCE","参考文档",7382,"https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release\u002FBENCHMARK.md",{"path":66,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":67,"sourceHtmlUrl":68},"references\u002Frunning.md",1696,"https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release\u002Freferences\u002Frunning.md",{"path":70,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":71,"sourceHtmlUrl":72},"skill-card.md",4189,"https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fpaidf-anomalygen\u002Frepo-docs\u002Fskills\u002Fcdf-anomalygen-release\u002Fskill-card.md",[74,81,88,95,102,109],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7A57J4E78QJZS8MA72P82DV","agentic-ecology-ui","Guidelines and templates for creating a generic web UI to visualize, rank, filter, and annotate vector databases. The UI has a hierarchical structure covering database selection, label selection\u002Fdefinition, ranking\u002Ffiltering configuration, and row visualization\u002Fannotation. Use when building, serving, or adapting an interactive web UI for Hoplite vector databases.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透google-deepmind\u002Fagentic_ecology\u002Frepo-docs\u002Fskills\u002Fagentic-ecology-ui\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M7A57C4RGS1A7VM6CNPNQDBH","agentic-ecology-init","Initializes a local uv-managed project directory for agentic ecology workloads. Sets up Python dependencies using reference pyproject.toml and uv.lock, configures workspace rules, and ensures Agentic Ecology skills are discoverable. Use when initializing a new project workspace or bootstrapping a clean environment.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透google-deepmind\u002Fagentic_ecology\u002Frepo-docs\u002Fskills\u002Fagentic-ecology-init\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M7A576Q1GYF5GAQCWGPPDQWZ","agentic-ecology-camera-traps","Provides capabilities to run SpeciesNet detector and classifier on camera trap images, extract crop-level feature embeddings, and populate a Hoplite vector database for downstream search and agile modeling. Use when ingesting camera trap images, classifying wildlife photos, or indexing camera trap data.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透google-deepmind\u002Fagentic_ecology\u002Frepo-docs\u002Fskills\u002Fagentic-ecology-camera-traps\u002FSKILL.md",[],[],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":19},"01M7A5708M54MMTKXQARDCJQMW","agentic-ecology-bioacoustics","Provides bioacoustic analysis capabilities for ecologists and researchers using the perch-hoplite Python package. A typical use case is to use agile modeling to bootstrap the creation and deployment of a bespoke detector for targeted species on an existing collection of passive acoustic monitoring recordings. Use when processing and annotating audio recordings for bioacoustic applications.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透google-deepmind\u002Fagentic_ecology\u002Frepo-docs\u002Fskills\u002Fagentic-ecology-bioacoustics\u002FSKILL.md",[],[],{"skillNo":103,"name":104,"nameZh":6,"description":105,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":106,"categories":107,"keywords":108,"publishedAt":19},"01M7A53GYCZY3AT9C2FR44TC35","sample-skill","Sample skill used by agent-env integration tests. Exercises SKILL.md parsing plus supporting scripts and references.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透browser-use\u002Fagentenv-framework\u002Frepo-docs\u002Ftst\u002Fdata\u002Fskills\u002Fsample_skill\u002FSKILL.md",[],[],{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":19},"01M7A4KD50DHSNH1SCNKP1J6S1","simready-foundation-validate-foundation-change","Use for auditing SimReady requirement, validator, feature, profile, adapter, test, and skill consistency.","https:\u002F\u002Fgithub.com\u002Fleemiracle\u002Fwork4ai\u002Fblob\u002Fmain\u002F讲透宇宙\u002F05-侦\u002F讲透NVIDIA\u002Fsimready-foundation\u002Frepo-docs\u002F.codex\u002Fskills\u002Fsimready-foundation-validate-foundation-change\u002FSKILL.md",[],[]]