[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M71PXTAVQ3KJ8CSR46X0QAZ9":3,"skill-same:01M71PXTAVQ3KJ8CSR46X0QAZ9":101},{"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":23,"publishedAt":44,"linkOnly":45,"markdownContent":46,"markdownBody":47,"frontMatter":48,"refName":56,"commitSha":57,"skillDirectory":58,"forksCount":14,"repositoryHtmlUrl":59,"repositoryHomepageUrl":60,"ownerHtmlUrl":61,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":62,"install":63,"resources":76},"01M71PXTAVQ3KJ8CSR46X0QAZ9","tao-run-on-local-docker",null,"Local or remote Docker execution for TAO SDK job containers using a Docker daemon with NVIDIA GPU runtime. Use when running TAO jobs on the current machine, a directly attached Docker host, or a remote GPU box exposed through DOCKER_HOST. Trigger phrases include \"run locally\", \"local Docker\", \"remote Docker\", \"use my GPU\", \"run on my machine\", \"host Docker daemon\".","pattanation\u002FNorvik","pattanation","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F235419268?v=4","01M71M5THHYYB15ME8R2E4FCZG","Patta","01M71M5TJM1NDM992M8PDKYR2K",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000011","运维与部署","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43],"tao run on local docker","run","docker","tao","local","NVIDIA","running","extra","my","install","read","Bash","you","when","host","remote","only","daemon","nvidia-tao","branch","2026-11-10 13:02:02",false,"---\nname: tao-run-on-local-docker\ndescription: Local or remote Docker execution for TAO SDK job containers using a Docker daemon with NVIDIA GPU runtime. Use\n  when running TAO jobs on the current machine, a directly attached Docker host, or a remote GPU box exposed through\n  DOCKER_HOST. Trigger phrases include \"run locally\", \"local Docker\", \"remote Docker\", \"use my GPU\", \"run on my\n  machine\", \"host Docker daemon\".\nlicense: Apache-2.0\ncompatibility: Requires NVIDIA driver branch 580, CUDA Toolkit 13.0, Docker, and NVIDIA Container Toolkit 1.19.0. The TAO SDK with the docker extra (pip install 'nvidia-tao-sdk[docker]') is needed only if you want Job handles, S3 I\u002FO wrapping, or run-folder durability via ActionWorkflow.\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\nallowed-tools: Read Bash\ntags:\n- platform\n- local\n- docker\n---\n\n# Local Docker\n\nSingle-node execution platform that runs TAO jobs as named Docker containers on\na Docker daemon. The daemon can be local to the agent host or remote through\n`DOCKER_HOST=ssh:\u002F\u002Fuser@host` \u002F a Docker context. It is useful for development,\ndebugging, small runs, and workflows where a local coding agent submits jobs to\na remote GPU box.\n\nUse local Docker when the data is local to the Docker host or accessible through\nmounted volumes\u002Fcloud credentials. Do not use it for remote cluster scheduling,\nmulti-node training, or jobs that need SLURM queueing.\n\nUse remote Docker when the agent is running on a workstation or laptop but the\nDocker daemon and GPUs are on another single GPU server. In remote Docker mode,\nall local filesystem paths in specs are interpreted on the remote Docker host,\nnot on the agent machine.\n\n## Preflight\n\nThe workflow must verify the host GPU runtime before starting Docker jobs. If\nthe check fails, prompt the user to approve the install, run the printed install\ncommand, and rerun the preflight.\n\n```bash\n# Host GPU runtime: NVIDIA driver 580, CUDA 13.0, NVIDIA Container Toolkit 1.19.0.\nTAO_SKILL_BANK_ROOT=\"${TAO_SKILL_BANK_ROOT:-$PWD}\"\nSETUP_SCRIPT=\"${TAO_SKILL_BANK_ROOT}\u002Fskills\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002Fscripts\u002Fsetup-nvidia-gpu-host.sh\"\n\nbash \"$SETUP_SCRIPT\" --backend docker --check-only || {\n  echo \"MISSING: TAO GPU host runtime is not ready.\"\n  echo \"After user approval, run:\"\n  echo \"  bash \\\"$SETUP_SCRIPT\\\" --backend docker --install --yes\"\n  exit 1\n}\n\n# Mode 1 — direct docker (no Python). All you need is docker + the GPU runtime.\ndocker info >\u002Fdev\u002Fnull 2>&1 || { echo \"MISSING: docker daemon not reachable. Start Docker.\"; exit 1; }\ndocker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi >\u002Fdev\u002Fnull 2>&1 || {\n  echo \"MISSING: NVIDIA Container Toolkit not installed\u002Fconfigured. See:\"\n  echo \"  bash \\\"$SETUP_SCRIPT\\\" --backend docker --install --yes\"\n  exit 1\n}\n\n# Mode 2 — TAO SDK wrapper. Adds Job handles, S3 I\u002FO wrapping, ActionWorkflow.\n# Skip this block if Mode 1 is sufficient for the user's request.\n# When Mode 2 is in scope, read `tao-skill-bank:tao-run-platform` for the DockerSDK\n# kwarg contract, build_entrypoint, and monitoring patterns.\n# nvidia-tao-sdk is on public PyPI; pin lives in versions.yaml (wheels.tao_sdk_docker).\nPIN=$(\"${TAO_SKILL_BANK_PATH:?}\u002Fscripts\u002Fresolve_versions_key.py\" wheels.tao_sdk_docker)\npython -c \"import tao_sdk\" 2>\u002Fdev\u002Fnull || python -m pip install \"$PIN\"\npython -c \"import docker\" 2>\u002Fdev\u002Fnull || python -m pip install \"$PIN\"\npython -c \"import tao_sdk, docker\"\n\n# DockerSDK attaches every job container to ${DOCKER_NETWORK:-tao_default}.\n# Create the network if it is missing; the operation is local and idempotent.\nDOCKER_NETWORK_NAME=\"${DOCKER_NETWORK:-tao_default}\"\ndocker network inspect \"$DOCKER_NETWORK_NAME\" >\u002Fdev\u002Fnull 2>&1 || \\\n  docker network create \"$DOCKER_NETWORK_NAME\" >\u002Fdev\u002Fnull\n```\n\nIf a check fails, the agent prompts the user to authorize the install\u002Ffix via Bash before proceeding. Pip-installable Python requirements and Docker network creation above are exceptions: install\u002Fcreate them automatically, then rerun preflight.\n\n## Credentials\n\nThere are no platform credentials required beyond access to the Docker daemon.\n\nOptional environment:\n\n- **DOCKER_HOST**: Optional Docker daemon URL. If unset, the SDK uses the\n  Docker Python client's normal environment\u002Fdefault socket resolution. Required\n  for the `remote-docker` platform option.\n- **DOCKER_NETWORK**: Docker network for job containers. Default is\n  `tao_default`.\n- **DOCKER_USERNAME**: Registry username. Default is `$oauthtoken` for NGC.\n- **NGC_KEY**: Used when pulling private images from `nvcr.io`.\n- **HOST_SSH_PATH**: Mounted into AutoML brain containers when they need SSH keys\n  to monitor remote SLURM child jobs.\n- **ACCESS_KEY**, **SECRET_KEY**, **S3_ENDPOINT_URL**, **S3_BUCKET_NAME**:\n  Optional S3-compatible storage settings for jobs that still read\u002Fwrite cloud\n  storage from a local container.\n\n## Launch Preflight\n\nBefore generating scripts or starting containers:\n\n1. Verify the Docker daemon is reachable, NVIDIA Container Toolkit is registered\n   as a Docker runtime, GPUs and driver version are reported, and a smoke\n   container can see GPUs before launch. For remote Docker, query GPUs through\n   `docker run ... nvidia-smi` against the remote daemon; do not use local\n   `nvidia-smi` from the agent machine.\n2. Verify every local\u002Ffile dataset annotation and media path exists on the\n   Docker host.\n3. For `s3:\u002F\u002F` datasets\u002Fresults, verify `ACCESS_KEY` and `SECRET_KEY` are set\n   and the exact paths are readable with `aws s3 ls`. If `aws` is missing,\n   report the missing dependency and ask before installing it; rerun preflight\n   after installation.\n4. Verify model-specific credentials such as `HF_TOKEN` before launch.\n5. Check current GPU occupancy with `nvidia-smi` and avoid GPUs already used by\n   other running jobs when the user requested that constraint. Show the selected\n   GPU ids in the launch review.\n6. For model\u002Fcontainer combinations with known architecture limits, compare\n   host GPU compute capability with the container stack before launch. If the\n   selected image cannot JIT or run kernels for the host architecture, block\n   early and ask for a compatible image or platform.\n\nUse the packaged helper for these checks when possible:\n\n```bash\n${TAO_SKILL_BANK_PATH:-~\u002Ftao-skills-external}\u002Fscripts\u002Fcheck_tao_launch_preflight.py \\\n  --platform local-docker \\\n  --container-image \"\u003Cselected-image>\" \\\n  --path train_annotation=\u002Fabs\u002Fpath\u002Fto\u002Fannotations.json \\\n  --path train_media=\u002Fabs\u002Fpath\u002Fto\u002Fmedia\n```\n\nFor a remote Docker daemon, use the `remote-docker` platform and pass or export\n`DOCKER_HOST`. The helper verifies remote GPU\u002Fruntime readiness and checks\nremote-host dataset paths through read-only bind mounts:\n\n```bash\n${TAO_SKILL_BANK_PATH:-~\u002Ftao-skills-external}\u002Fscripts\u002Fcheck_tao_launch_preflight.py \\\n  --platform remote-docker \\\n  --docker-host ssh:\u002F\u002Fuser@gpu-host \\\n  --container-image \"\u003Cselected-image>\" \\\n  --gpu-smoke-image ubuntu:22.04 \\\n  --path train_annotation=\u002Fremote\u002Fdata\u002Ftrain\u002Fannotations.json \\\n  --path train_media=\u002Fremote\u002Fdata\u002Ftrain\n```\n\nThe `--path` values above must exist on the remote Docker host. Do not pass\npaths that exist only on the local laptop or Codex host.\n\n## Multi-GPU and multi-node\n\n**Multi-node is not supported on local Docker.** One job runs on the local Docker daemon's host with no cross-host coordination.\n\nMulti-GPU **on the local host** is supported via the NVIDIA Container Toolkit's `--gpus` flag (`--gpus all` or `--gpus '\"device=0,1,2,3\"'`). `DockerSDK.create_job(gpu_count=N)` plumbs through to `--gpus`. Single-host distributed init uses `localhost`; `torchrun --nproc-per-node=N` or PyTorch DDP work as usual.\n\n## Backend Details\n\nUse the SDK backend value `local-docker`. The local backend schema has no extra\nbackend details, so most routing is controlled by environment and job\nparameters:\n\n```json\n{\n  \"backend_type\": \"local-docker\",\n  \"num_gpu\": 1\n}\n```\n\nFollowing the Brev SDK design, platform\u002Fcontrol-plane values stay in SDK\nstate and Docker labels. The SDK does not inject `BACKEND`, `HOST_PLATFORM`,\n`MONGOSECRET`, `DOCKER_HOST`, or `DOCKER_NETWORK` into the training container.\n\n## Container Execution\n\nThe TAO SDK local Docker handler starts containers through the Docker Python\nclient:\n\n- Backend job name uses the `tao-job-\u003Cjob_id>` form used by SDK handlers.\n- Command is usually `[\"\u002Fbin\u002Fbash\", \"-c\", \"\u003Cjob command>\"]`.\n- Containers run detached. The SDK keeps containers by default so status and\n  logs remain inspectable, unless `DOCKER_AUTO_REMOVE=true`.\n- `\u002Fdev\u002Fshm` is mounted as tmpfs.\n- The configured Docker network is applied by the Docker daemon for the job\n  container; it is not passed through as a process environment variable.\n- Existing containers with the same job id are stopped and removed before a\n  replacement starts.\n\nFor GPU access, the handler auto-detects the host type:\n\n- Tegra or Jetson hosts use `runtime=\"nvidia\"` plus\n  `NVIDIA_VISIBLE_DEVICES` and `NVIDIA_DRIVER_CAPABILITIES=all`.\n- Standard x86 hosts use Docker `device_requests` with GPU capabilities.\n\nIf `num_gpus` is `0`, no GPUs are assigned. If `num_gpus` is `-1`, all visible\nGPUs are requested. Prefer explicit GPU counts for shared development machines.\nWhen explicit device ids are available, prefer them over count-only selection\non shared machines so the launch does not steal GPUs occupied by other tasks.\n\n## Storage\n\nLocal Docker accepts local and `file:\u002F\u002F` paths because the container runs on the\nsame Docker host. Make sure every path in the spec is either:\n\n- mounted into the container by the handler or surrounding service,\n- reachable from inside the container already, or\n- a cloud URI with matching credentials.\n\nFor remote\u002Fshared filesystems, prefer the platform that owns that filesystem.\nFor example, use SLURM plus `lustre:\u002F\u002F\u002F...` for Lustre paths on a cluster.\n\n## Monitoring\n\n- The SDK handler maps Docker container state directly: created -> Pending,\n  running\u002Frestarting -> Running, paused -> Paused, exit code 0 -> Complete,\n  nonzero exit -> Error.\n- Logs come directly from the named container through the Docker Python client\n  (`docker logs tao-job-\u003Cjob_id>`).\n\nIf the container has exited, died, is being removed, or cannot be found, status\nreconciliation treats the backend process as terminated.\n\n## Cancellation\n\nCancellation stops the named container. GPU ownership is managed by Docker \u002F\nthe NVIDIA runtime, not by TAO Core's local GPU manager.\n\n## Optional: via the TAO SDK\n\nIf you want Job handles, S3 I\u002FO wrapping via the SDK's `script_runner`, or\ndurability across sessions:\n\n```python\nfrom tao_sdk.platforms.docker import DockerSDK\n\nsdk = DockerSDK()  # reads DOCKER_HOST, NGC_KEY, S3 creds from env\njob = sdk.create_job(\n    image='nvcr.io\u002Fnvidia\u002Ftao\u002Ftao-toolkit:6.26.3-pyt',\n    command='dino train -e \u002Ftmp\u002Fspec.yaml',\n    gpu_count=1,\n    inputs={'\u002Fdata\u002Ftrain.json': 's3:\u002F\u002Fbucket\u002Fcoco\u002Ftrain.json'},\n    outputs=['\u002Fresults\u002F'],\n)\n\nstatus = sdk.get_job_status(job.id)\nlogs = sdk.get_job_logs(job.id, tail=200)\n```\n\nThis wraps the same `docker run` invocation under a `Job` handle and routes\nthe entrypoint through `script_runner` so `inputs`\u002F`outputs` get downloaded\nfrom \u002F uploaded to S3 automatically. If you don't need those, just use\n`docker run` directly — no SDK install required.\n\n## Failure Modes\n\n**Docker client not initialized**: Verify the Docker Python package is installed,\nset `DOCKER_HOST` if you are not using the default local socket, and confirm the\nprocess can talk to the daemon.\n\n**GPU assignment failed**: Requested GPUs are unavailable, the NVIDIA Container\nToolkit is not configured, or the Docker daemon cannot create GPU device\nrequests. Use fewer GPUs, wait for another job to finish, or verify\n`docker run --gpus ...` works on the host.\n\n**Image pull auth failed**: Set a valid `NGC_KEY` for private `nvcr.io` images\nor run `docker login nvcr.io -u '$oauthtoken'` on the Docker host.\n\n**Container exited unexpectedly**: Check `docker logs tao-job-\u003Cjob_id>`, the\nconfigured `DOCKER_NETWORK`, and the command produced by the SDK action runner.\n\n**Path missing inside container**: A local path on the host is not necessarily\nmounted into the job container. Use a path convention supported by the action\nrunner or configure an explicit volume through the surrounding service.\n","# Local Docker\n\nSingle-node execution platform that runs TAO jobs as named Docker containers on\na Docker daemon. The daemon can be local to the agent host or remote through\n`DOCKER_HOST=ssh:\u002F\u002Fuser@host` \u002F a Docker context. It is useful for development,\ndebugging, small runs, and workflows where a local coding agent submits jobs to\na remote GPU box.\n\nUse local Docker when the data is local to the Docker host or accessible through\nmounted volumes\u002Fcloud credentials. Do not use it for remote cluster scheduling,\nmulti-node training, or jobs that need SLURM queueing.\n\nUse remote Docker when the agent is running on a workstation or laptop but the\nDocker daemon and GPUs are on another single GPU server. In remote Docker mode,\nall local filesystem paths in specs are interpreted on the remote Docker host,\nnot on the agent machine.\n\n## Preflight\n\nThe workflow must verify the host GPU runtime before starting Docker jobs. If\nthe check fails, prompt the user to approve the install, run the printed install\ncommand, and rerun the preflight.\n\n```bash\n# Host GPU runtime: NVIDIA driver 580, CUDA 13.0, NVIDIA Container Toolkit 1.19.0.\nTAO_SKILL_BANK_ROOT=\"${TAO_SKILL_BANK_ROOT:-$PWD}\"\nSETUP_SCRIPT=\"${TAO_SKILL_BANK_ROOT}\u002Fskills\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002Fscripts\u002Fsetup-nvidia-gpu-host.sh\"\n\nbash \"$SETUP_SCRIPT\" --backend docker --check-only || {\n  echo \"MISSING: TAO GPU host runtime is not ready.\"\n  echo \"After user approval, run:\"\n  echo \"  bash \\\"$SETUP_SCRIPT\\\" --backend docker --install --yes\"\n  exit 1\n}\n\n# Mode 1 — direct docker (no Python). All you need is docker + the GPU runtime.\ndocker info >\u002Fdev\u002Fnull 2>&1 || { echo \"MISSING: docker daemon not reachable. Start Docker.\"; exit 1; }\ndocker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi >\u002Fdev\u002Fnull 2>&1 || {\n  echo \"MISSING: NVIDIA Container Toolkit not installed\u002Fconfigured. See:\"\n  echo \"  bash \\\"$SETUP_SCRIPT\\\" --backend docker --install --yes\"\n  exit 1\n}\n\n# Mode 2 — TAO SDK wrapper. Adds Job handles, S3 I\u002FO wrapping, ActionWorkflow.\n# Skip this block if Mode 1 is sufficient for the user's request.\n# When Mode 2 is in scope, read `tao-skill-bank:tao-run-platform` for the DockerSDK\n# kwarg contract, build_entrypoint, and monitoring patterns.\n# nvidia-tao-sdk is on public PyPI; pin lives in versions.yaml (wheels.tao_sdk_docker).\nPIN=$(\"${TAO_SKILL_BANK_PATH:?}\u002Fscripts\u002Fresolve_versions_key.py\" wheels.tao_sdk_docker)\npython -c \"import tao_sdk\" 2>\u002Fdev\u002Fnull || python -m pip install \"$PIN\"\npython -c \"import docker\" 2>\u002Fdev\u002Fnull || python -m pip install \"$PIN\"\npython -c \"import tao_sdk, docker\"\n\n# DockerSDK attaches every job container to ${DOCKER_NETWORK:-tao_default}.\n# Create the network if it is missing; the operation is local and idempotent.\nDOCKER_NETWORK_NAME=\"${DOCKER_NETWORK:-tao_default}\"\ndocker network inspect \"$DOCKER_NETWORK_NAME\" >\u002Fdev\u002Fnull 2>&1 || \\\n  docker network create \"$DOCKER_NETWORK_NAME\" >\u002Fdev\u002Fnull\n```\n\nIf a check fails, the agent prompts the user to authorize the install\u002Ffix via Bash before proceeding. Pip-installable Python requirements and Docker network creation above are exceptions: install\u002Fcreate them automatically, then rerun preflight.\n\n## Credentials\n\nThere are no platform credentials required beyond access to the Docker daemon.\n\nOptional environment:\n\n- **DOCKER_HOST**: Optional Docker daemon URL. If unset, the SDK uses the\n  Docker Python client's normal environment\u002Fdefault socket resolution. Required\n  for the `remote-docker` platform option.\n- **DOCKER_NETWORK**: Docker network for job containers. Default is\n  `tao_default`.\n- **DOCKER_USERNAME**: Registry username. Default is `$oauthtoken` for NGC.\n- **NGC_KEY**: Used when pulling private images from `nvcr.io`.\n- **HOST_SSH_PATH**: Mounted into AutoML brain containers when they need SSH keys\n  to monitor remote SLURM child jobs.\n- **ACCESS_KEY**, **SECRET_KEY**, **S3_ENDPOINT_URL**, **S3_BUCKET_NAME**:\n  Optional S3-compatible storage settings for jobs that still read\u002Fwrite cloud\n  storage from a local container.\n\n## Launch Preflight\n\nBefore generating scripts or starting containers:\n\n1. Verify the Docker daemon is reachable, NVIDIA Container Toolkit is registered\n   as a Docker runtime, GPUs and driver version are reported, and a smoke\n   container can see GPUs before launch. For remote Docker, query GPUs through\n   `docker run ... nvidia-smi` against the remote daemon; do not use local\n   `nvidia-smi` from the agent machine.\n2. Verify every local\u002Ffile dataset annotation and media path exists on the\n   Docker host.\n3. For `s3:\u002F\u002F` datasets\u002Fresults, verify `ACCESS_KEY` and `SECRET_KEY` are set\n   and the exact paths are readable with `aws s3 ls`. If `aws` is missing,\n   report the missing dependency and ask before installing it; rerun preflight\n   after installation.\n4. Verify model-specific credentials such as `HF_TOKEN` before launch.\n5. Check current GPU occupancy with `nvidia-smi` and avoid GPUs already used by\n   other running jobs when the user requested that constraint. Show the selected\n   GPU ids in the launch review.\n6. For model\u002Fcontainer combinations with known architecture limits, compare\n   host GPU compute capability with the container stack before launch. If the\n   selected image cannot JIT or run kernels for the host architecture, block\n   early and ask for a compatible image or platform.\n\nUse the packaged helper for these checks when possible:\n\n```bash\n${TAO_SKILL_BANK_PATH:-~\u002Ftao-skills-external}\u002Fscripts\u002Fcheck_tao_launch_preflight.py \\\n  --platform local-docker \\\n  --container-image \"\u003Cselected-image>\" \\\n  --path train_annotation=\u002Fabs\u002Fpath\u002Fto\u002Fannotations.json \\\n  --path train_media=\u002Fabs\u002Fpath\u002Fto\u002Fmedia\n```\n\nFor a remote Docker daemon, use the `remote-docker` platform and pass or export\n`DOCKER_HOST`. The helper verifies remote GPU\u002Fruntime readiness and checks\nremote-host dataset paths through read-only bind mounts:\n\n```bash\n${TAO_SKILL_BANK_PATH:-~\u002Ftao-skills-external}\u002Fscripts\u002Fcheck_tao_launch_preflight.py \\\n  --platform remote-docker \\\n  --docker-host ssh:\u002F\u002Fuser@gpu-host \\\n  --container-image \"\u003Cselected-image>\" \\\n  --gpu-smoke-image ubuntu:22.04 \\\n  --path train_annotation=\u002Fremote\u002Fdata\u002Ftrain\u002Fannotations.json \\\n  --path train_media=\u002Fremote\u002Fdata\u002Ftrain\n```\n\nThe `--path` values above must exist on the remote Docker host. Do not pass\npaths that exist only on the local laptop or Codex host.\n\n## Multi-GPU and multi-node\n\n**Multi-node is not supported on local Docker.** One job runs on the local Docker daemon's host with no cross-host coordination.\n\nMulti-GPU **on the local host** is supported via the NVIDIA Container Toolkit's `--gpus` flag (`--gpus all` or `--gpus '\"device=0,1,2,3\"'`). `DockerSDK.create_job(gpu_count=N)` plumbs through to `--gpus`. Single-host distributed init uses `localhost`; `torchrun --nproc-per-node=N` or PyTorch DDP work as usual.\n\n## Backend Details\n\nUse the SDK backend value `local-docker`. The local backend schema has no extra\nbackend details, so most routing is controlled by environment and job\nparameters:\n\n```json\n{\n  \"backend_type\": \"local-docker\",\n  \"num_gpu\": 1\n}\n```\n\nFollowing the Brev SDK design, platform\u002Fcontrol-plane values stay in SDK\nstate and Docker labels. The SDK does not inject `BACKEND`, `HOST_PLATFORM`,\n`MONGOSECRET`, `DOCKER_HOST`, or `DOCKER_NETWORK` into the training container.\n\n## Container Execution\n\nThe TAO SDK local Docker handler starts containers through the Docker Python\nclient:\n\n- Backend job name uses the `tao-job-\u003Cjob_id>` form used by SDK handlers.\n- Command is usually `[\"\u002Fbin\u002Fbash\", \"-c\", \"\u003Cjob command>\"]`.\n- Containers run detached. The SDK keeps containers by default so status and\n  logs remain inspectable, unless `DOCKER_AUTO_REMOVE=true`.\n- `\u002Fdev\u002Fshm` is mounted as tmpfs.\n- The configured Docker network is applied by the Docker daemon for the job\n  container; it is not passed through as a process environment variable.\n- Existing containers with the same job id are stopped and removed before a\n  replacement starts.\n\nFor GPU access, the handler auto-detects the host type:\n\n- Tegra or Jetson hosts use `runtime=\"nvidia\"` plus\n  `NVIDIA_VISIBLE_DEVICES` and `NVIDIA_DRIVER_CAPABILITIES=all`.\n- Standard x86 hosts use Docker `device_requests` with GPU capabilities.\n\nIf `num_gpus` is `0`, no GPUs are assigned. If `num_gpus` is `-1`, all visible\nGPUs are requested. Prefer explicit GPU counts for shared development machines.\nWhen explicit device ids are available, prefer them over count-only selection\non shared machines so the launch does not steal GPUs occupied by other tasks.\n\n## Storage\n\nLocal Docker accepts local and `file:\u002F\u002F` paths because the container runs on the\nsame Docker host. Make sure every path in the spec is either:\n\n- mounted into the container by the handler or surrounding service,\n- reachable from inside the container already, or\n- a cloud URI with matching credentials.\n\nFor remote\u002Fshared filesystems, prefer the platform that owns that filesystem.\nFor example, use SLURM plus `lustre:\u002F\u002F\u002F...` for Lustre paths on a cluster.\n\n## Monitoring\n\n- The SDK handler maps Docker container state directly: created -> Pending,\n  running\u002Frestarting -> Running, paused -> Paused, exit code 0 -> Complete,\n  nonzero exit -> Error.\n- Logs come directly from the named container through the Docker Python client\n  (`docker logs tao-job-\u003Cjob_id>`).\n\nIf the container has exited, died, is being removed, or cannot be found, status\nreconciliation treats the backend process as terminated.\n\n## Cancellation\n\nCancellation stops the named container. GPU ownership is managed by Docker \u002F\nthe NVIDIA runtime, not by TAO Core's local GPU manager.\n\n## Optional: via the TAO SDK\n\nIf you want Job handles, S3 I\u002FO wrapping via the SDK's `script_runner`, or\ndurability across sessions:\n\n```python\nfrom tao_sdk.platforms.docker import DockerSDK\n\nsdk = DockerSDK()  # reads DOCKER_HOST, NGC_KEY, S3 creds from env\njob = sdk.create_job(\n    image='nvcr.io\u002Fnvidia\u002Ftao\u002Ftao-toolkit:6.26.3-pyt',\n    command='dino train -e \u002Ftmp\u002Fspec.yaml',\n    gpu_count=1,\n    inputs={'\u002Fdata\u002Ftrain.json': 's3:\u002F\u002Fbucket\u002Fcoco\u002Ftrain.json'},\n    outputs=['\u002Fresults\u002F'],\n)\n\nstatus = sdk.get_job_status(job.id)\nlogs = sdk.get_job_logs(job.id, tail=200)\n```\n\nThis wraps the same `docker run` invocation under a `Job` handle and routes\nthe entrypoint through `script_runner` so `inputs`\u002F`outputs` get downloaded\nfrom \u002F uploaded to S3 automatically. If you don't need those, just use\n`docker run` directly — no SDK install required.\n\n## Failure Modes\n\n**Docker client not initialized**: Verify the Docker Python package is installed,\nset `DOCKER_HOST` if you are not using the default local socket, and confirm the\nprocess can talk to the daemon.\n\n**GPU assignment failed**: Requested GPUs are unavailable, the NVIDIA Container\nToolkit is not configured, or the Docker daemon cannot create GPU device\nrequests. Use fewer GPUs, wait for another job to finish, or verify\n`docker run --gpus ...` works on the host.\n\n**Image pull auth failed**: Set a valid `NGC_KEY` for private `nvcr.io` images\nor run `docker login nvcr.io -u '$oauthtoken'` on the Docker host.\n\n**Container exited unexpectedly**: Check `docker logs tao-job-\u003Cjob_id>`, the\nconfigured `DOCKER_NETWORK`, and the command produced by the SDK action runner.\n\n**Path missing inside container**: A local path on the host is not necessarily\nmounted into the job container. Use a path convention supported by the action\nrunner or configure an explicit volume through the surrounding service.\n",{"name":5,"tags":49,"license":15,"metadata":51,"description":7,"allowed-tools":54,"compatibility":55},[50,28,26],"platform",{"author":52,"version":53},"NVIDIA Corporation","0.1.0","Read Bash","Requires NVIDIA driver branch 580, CUDA Toolkit 13.0, Docker, and NVIDIA Container Toolkit 1.19.0. The TAO SDK with the docker extra (pip install 'nvidia-tao-sdk[docker]') is needed only if you want Job handles, S3 I\u002FO wrapping, or run-folder durability via ActionWorkflow.","claude\u002Finstall-gsd-build-foQdV","d86cd127f527124b5fd0fd0bc679a4d79b7d01d7",".agents\u002Fskills\u002Ftao-run-on-local-docker","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik","https:\u002F\u002Fnorvik-orpin.vercel.app","https:\u002F\u002Fgithub.com\u002Fpattanation","2026-07-17 11:17:04",{"source":8,"skillName":5,"commands":64,"prompt":74,"zipAvailable":75},[65,68,71],{"tool":66,"command":67},"npx","npx skills add pattanation\u002FNorvik --skill tao-run-on-local-docker",{"tool":69,"command":70},"bunx","bunx skills add pattanation\u002FNorvik --skill tao-run-on-local-docker",{"tool":72,"command":73},"pnpm","pnpm dlx skills add pattanation\u002FNorvik --skill tao-run-on-local-docker","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002FSKILL.md",true,[77,83,87,91,95],{"path":78,"resourceType":79,"resourceTypeDesc":80,"sizeBytes":81,"sourceHtmlUrl":82},"BENCHMARK.md","REFERENCE","参考文档",3989,"https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002FBENCHMARK.md",{"path":84,"resourceType":79,"resourceTypeDesc":80,"sizeBytes":85,"sourceHtmlUrl":86},"evals\u002Fevals.json",862,"https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002Fevals\u002Fevals.json",{"path":88,"resourceType":79,"resourceTypeDesc":80,"sizeBytes":89,"sourceHtmlUrl":90},"references\u002Fskill_info.yaml",833,"https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002Freferences\u002Fskill_info.yaml",{"path":92,"resourceType":79,"resourceTypeDesc":80,"sizeBytes":93,"sourceHtmlUrl":94},"skill-card.md",3663,"https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002Fskill-card.md",{"path":96,"resourceType":97,"resourceTypeDesc":98,"sizeBytes":99,"sourceHtmlUrl":100},"skill.oms.sig","OTHER","其他",4837,"https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Ftao-run-on-local-docker\u002Fskill.oms.sig",[102,132,158,182,204,219],{"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":106,"sourceHtmlUrl":107,"categories":108,"keywords":115,"publishedAt":44},"01M71QDV4DM8T19HEHWATEF1K4","webapp-testing","Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fwebapp-testing\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fwebapp-testing\u002FSKILL.md",[109,112],{"categoryNo":110,"categoryName":111,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000009","前端开发",[116,117,118,119,120,28,121,122,123,124,125,126,127,128,129,130,131],"webapp","testing","webapp testing","for","applications","frontend","debugging","toolkit","logs","browser","with","playwright","web","behavior","ui","screenshots",{"skillNo":133,"name":134,"nameZh":6,"description":135,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":136,"sourceHtmlUrl":137,"categories":138,"keywords":140,"publishedAt":44},"01M71QDH2GGMXGA9105C9W1N7B","web-artifacts-builder","Suite of tools for creating elaborate, multi-component claude.ai HTML artifacts using modern frontend web technologies (React, Tailwind CSS, shadcn\u002Fui). Use for complex artifacts requiring state management, routing, or shadcn\u002Fui components - not for simple single-file HTML\u002FJSX artifacts.","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fweb-artifacts-builder\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fweb-artifacts-builder\u002FSKILL.md",[139],{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},[141,142,143,128,144,145,146,119,147,148,121,149,150,151,152,153,154,155,156,157],"builder","artifacts builder","artifacts","web artifacts builder","single-file","css","Tailwind","ui components","single","use","file","components","react","multi","state","routing","state management",{"skillNo":159,"name":160,"nameZh":6,"description":161,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":162,"licenseFileUrl":163,"sourceHtmlUrl":164,"categories":165,"keywords":167,"publishedAt":44},"01M71QD9T2XC3EWMQNN9ACNWQ1","vueuse-functions","Apply VueUse composables where appropriate to build concise, maintainable Vue.js \u002F Nuxt features.","MIT","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvueuse-functions\u002FLICENSE.md","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvueuse-functions\u002FSKILL.md",[166],{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},[168,169,170,171,172,173,174,175,176,177,178,179,180,181],"functions","vueuse","vueuse functions","project","nuxt","apply","github","concise","build","https","features","vue","js","composables",{"skillNo":183,"name":184,"nameZh":6,"description":185,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":162,"licenseFileUrl":186,"sourceHtmlUrl":187,"categories":188,"keywords":191,"publishedAt":44},"01M71QD51HWSGYMCMB7RW7PGK8","vue-testing-best-practices","Use for Vue.js testing. Covers Vitest, Vue Test Utils, component testing, mocking, testing patterns, and Playwright for E2E testing.","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-testing-best-practices\u002FLICENSE.md","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-testing-best-practices\u002FSKILL.md",[189,190],{"categoryNo":110,"categoryName":111,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},[192,117,193,179,194,127,195,196,197,174,198,119,199,200,201,150,202,180,203],"vue testing best practices","testing best practices","best practices","test","testing patterns","component","vitest","mocking","e2e","utils","patterns","e2e testing",{"skillNo":205,"name":206,"nameZh":6,"description":207,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":162,"licenseFileUrl":208,"sourceHtmlUrl":209,"categories":210,"keywords":212,"publishedAt":44},"01M71QD358H2DJ9THJRBP77T3T","vue-router-best-practices","Vue Router 4 patterns, navigation guards, route params, and route-component lifecycle interactions.","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-router-best-practices\u002FLICENSE.md","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-router-best-practices\u002FSKILL.md",[211],{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},[213,179,214,194,215,197,216,174,202,217,218],"router","vue router best practices","route","guards","navigation","lifecycle",{"skillNo":220,"name":221,"nameZh":6,"description":222,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":162,"licenseFileUrl":223,"sourceHtmlUrl":224,"categories":225,"keywords":227,"publishedAt":44},"01M71QCYPNWDE2MQZ1TF3ED5S4","vue-best-practices","MUST be used for Vue.js tasks. Strongly recommends Composition API with `\u003Cscript setup>` and TypeScript as the standard approach. Covers Vue 3, SSR, Volar, vue-tsc. Load for any Vue, .vue files, Vue Router, Pinia, or Vite with Vue work. ALWAYS use Composition API unless the project explicitly requires Options API.","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-best-practices\u002FLICENSE.md","https:\u002F\u002Fgithub.com\u002Fpattanation\u002FNorvik\u002Fblob\u002Fclaude\u002Finstall-gsd-build-foQdV\u002F.agents\u002Fskills\u002Fvue-best-practices\u002FSKILL.md",[226],{"categoryNo":113,"categoryName":114,"taxonomyType":21,"taxonomyTypeDesc":22},[228,179,194,229,171,230,231,232,213,233,234,235,236,237,238,119,239,150,240,241],"vue best practices","Composition","files","tasks","api","pinia","work","options","ssr","typescript","as","vite","load","standard"]