[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-tao-finetune-cosmos-reason":3,"mdc-idbx18-key":34,"related-repo-nvidia-tao-finetune-cosmos-reason":1587,"related-org-nvidia-tao-finetune-cosmos-reason":1691},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"tao-finetune-cosmos-reason","fine-tune Cosmos3-Nano video QA models","Cosmos3-Nano video QA supervised fine-tuning with FSDP parallelism. Use when training or evaluating video question-answering models, fine-tuning Cosmos3-Nano or compatible Cosmos Reason models with SFT\u002FLoRA, or working with Cosmos-RL. Trigger phrases include \"fine-tune Cosmos\", \"Cosmos3 Nano Reasoner\", \"Cosmos-RL SFT\", \"video QA fine-tune\", \"Cosmos3-Nano training\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Deep Learning","deep-learning","tag",{"name":17,"slug":18,"type":15},"Machine Learning","machine-learning",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Video","video",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:27:42.846826","Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Ftao-finetune-cosmos-reason","---\nname: tao-finetune-cosmos-reason\ndescription: Cosmos3-Nano video QA supervised fine-tuning with FSDP parallelism. Use when training or evaluating video\n  question-answering models, fine-tuning Cosmos3-Nano or compatible Cosmos Reason models with SFT\u002FLoRA, or working with\n  Cosmos-RL. Trigger phrases include \"fine-tune Cosmos\", \"Cosmos3 Nano Reasoner\", \"Cosmos-RL SFT\",\n  \"video QA fine-tune\", \"Cosmos3-Nano training\".\nlicense: Apache-2.0\ncompatibility: Requires docker + nvidia-container-toolkit.\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\nallowed-tools: Read Bash\ntags:\n- video\n- qa\n- cosmos\n- sft\n- reasoning\n- vlm\n---\n\n# Cosmos-RL\n\nSupervised fine-tuning (SFT) of Cosmos Reason video QA models. The packaged\ndefault base model is **hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano**. Pretrained weights\nare sourced from HuggingFace, not NGC. Gated HuggingFace models require\n`HF_TOKEN`. Some Cosmos-RL images cannot load the native Cosmos3 Omni checkpoint\nformat directly; for those images, convert Cosmos3-Nano to a Qwen3-VL HF\nsafetensors directory before train\u002Fevaluate and use that converted directory as\nthe PTM path.\n\nUses FSDP-based parallelism with `dp_shard_size` for GPU count and `dp_replicate_size` for node count (not the standard `num_gpus`\u002F`num_nodes`).\n\nRequests for \"Cosmos Reason 3\", \"Cosmos3 Nano Reasoner\", or\n`nvidia\u002FCosmos3-Nano` are handled by this skill. There is no separate Cosmos3\nmodel directory in the skill bank; route those requests here. Override the base\nHuggingFace model only when the user explicitly asks for a different model.\n\nDeep detail lives in references; load the smallest one that matches the task:\n\n- `references\u002Fcosmos-reason-launch.md` — launch intake, preflight, per-action dataset requirements, spec construction, typical overrides.\n- `references\u002Fcosmos-reason-evaluate.md` — evaluate (flat TOML, task types, LoRA eval, selective download, results) and datasets.\n- `references\u002Fcosmos-reason-automl.md` — AutoML\u002FHPO policy and search-space guidance.\n- `references\u002Fcosmos-reason-parameters.md` — important parameters, hardware, error patterns, DEFT\u002Fgap analysis, parent-model inference mappings.\n\n## Dataclass Schemas\n\nGenerated TAO Core schemas are packaged in `schemas\u002F\u003Caction>.schema.json`, with `schemas\u002Fmanifest.json` listing available actions. Each generated schema also emits `references\u002Fspec_template_\u003Caction>.yaml` from the schema top-level `default` field. AutoML enablement is declared at the model layer in `references\u002Fskill_info.yaml` via `automl_enabled`. Runnable AutoML still requires `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` to exist and parse. Use the packaged train schema for `automl_default_parameters`, `automl_disabled_parameters`, defaults, min\u002Fmax bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect `~\u002Ftao-core` at runtime; maintainers regenerate schemas\u002Ftemplates before packaging the skill bank.\n\n## Train Action Policy\n\nThis model is AutoML-enabled at the model layer. Before handling any train-stage request, read `references\u002Fskill_info.yaml` and resolve the run override from either an explicit `automl_policy` value or the user's workflow request. Use `automl_policy: on` by default and only expose `on` \u002F `off` in new launch prompts. Treat phrases like \"turn off AutoML\", \"disable AutoML\", \"no HPO\", or \"plain training\" as `automl_policy: off` for this run only. When `automl_policy: on`, `automl_enabled: true`, and both `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` are packaged, route the train action through `tao-skill-bank:tao-run-automl` by default with this model's `skill_dir`. Preserve workflow\u002Fapplication overrides for datasets, specs, output directories, GPU\u002Fplatform settings, parent checkpoints, and `automl_policy`. Use direct model training only when `automl_policy: off` or the packaged train schema\u002Ftemplate is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.\n\nNon-train actions such as `evaluate`, `inference`, and `quantize` stay in this model skill. The per-run `automl_policy` override does not change model metadata.\n\n## Credentials\n\n- **HF_TOKEN** (required for gated models): HuggingFace access token. For the\n  packaged default, the user must accept the model agreement at\n  \u003Chttps:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FCosmos3-Nano> and provide a token with read\n  access. If the user explicitly overrides the base model, they must accept\n  that target model's agreement too. Passed to the container as a\n  `docker_env_var`.\n\n## Cosmos3 Checkpoint Conversion\n\nWhen a selected image cannot load the native Cosmos3 checkpoint format\n(`model_type=\"cosmos3_omni\"` or `Cosmos3ForConditionalGeneration`), do not patch\nQwenVL, Transformers, or vLLM first. Use the upstream Cosmos Framework VLM\nconversion path to produce a Qwen3-VL HF safetensors directory, then point\nCosmos-RL specs at that converted directory.\n\nThe model skill packages a helper:\n\n```bash\npython skills\u002Fmodels\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py \\\n  --checkpoint-path \u002Fabs\u002Fpath\u002FCosmos3-Nano \\\n  --output-path \u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM \\\n  --secrets-env ~\u002F.tao\u002Fsecrets.env \\\n  --validate-with-image \u003Ccosmos-rl-image>\n```\n\nAfter conversion, use the converted directory consistently as the PTM:\n\n```text\ntrain:    policy.model_name_or_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.model_name=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.base_model_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\n```\n\nFor local Docker, mount the converted directory read-only into the Cosmos-RL\ncontainer and set the spec to the container path. If a converted copy already\nexists and validates, reuse it for PTM baseline evaluation, AutoML\nrecommendations, and final best-checkpoint evaluation rather than converting\nagain.\n\n## Training Requirements\n\n- **Dataset type:** vlm\n- **Formats:** llava, daft\n- **Accepted dataset intents:** training, evaluation, testing\n- **Monitoring metric:** val\u002Favg_loss, val\u002Freward_avg, val\u002Floss\n- **Dataset URI examples:** `s3:\u002F\u002Fbucket\u002Fcosmos\u002Ftrain`, `s3:\u002F\u002Fbucket\u002Fcosmos\u002Feval`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Ftrain`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Feval`\n- **Input modes:** accept either dataset roots or direct spec-key paths. Root mode maps `\u003Croot>\u002Fannotations.json` plus `\u003Croot>` as the media path. Direct spec mode is valid when annotations and media live in different locations, for example `custom.train_dataset.annotation_path=\u002Flustre\u002F...\u002Ftrain.json` and `custom.train_dataset.media_path=\u002Flustre\u002F...\u002Fvideos.tar.gz`.\n- **Media handling:** do not ask the user to choose `videos.tar.gz` vs `images.tar.gz` unless they are using direct spec mode or the model\u002Faction requires a single media archive. In root mode, pass the dataset root as the media path.\n- **Annotation validation:** before launching train\u002FAutoML\u002Fevaluate, verify the\n  annotation JSON is readable and the referenced media path or archive is\n  visible from the selected platform. Do not block, patch, or mutate\n  annotations solely because optional fields are absent.\n- **Per-record video FPS:** the packaged train template uses\n  `custom.vision.nframes`, so per-record `video_fps` is not required by\n  default. If the user switches to `custom.vision.fps`, selects a dataset\n  profile that requires per-record timing, or uses an image\u002Fversion that\n  requires `video_fps`, make it a preflight requirement with\n  `--json-required-field train_annotation=video_fps` and\n  `--json-required-field val_annotation=video_fps` before any download or\n  job launch.\n\n## Spec construction\n\ncosmos-rl is `mode: config`. **Always start from the packaged\n`references\u002Fspec_template_\u003Caction>.yaml` for the requested action** — load it\nas your base spec via `yaml.safe_load(...)` and apply user overrides on top.\nDon't rebuild from scratch.\n\n```python\nimport yaml\nfrom pathlib import Path\n\nskill = Path.home() \u002F \"tao-sdk\u002Ftao-skills-external\u002Fskills\u002Fmodels\u002Ftao-finetune-cosmos-reason\"\naction = \"train\"  # train, evaluate, inference, or quantize\nspecs = yaml.safe_load((skill \u002F f\"references\u002Fspec_template_{action}.yaml\").read_text())\n# Now apply your overrides on top of `specs`.\n```\n\nThe reference TOML (and the spec the model actually consumes) is **nested\ndicts**, not flat dotted keys. Dotted notation in override examples denotes\n*paths into the nested spec* — walk the path and assign at the leaf. See\n`skills\u002Fplatform\u002Ftao-run-platform\u002FSKILL.md`'s \"spec is nested dicts\" callout.\nData source overrides are **mandatory for every action**.\n\nThe packaged template keeps `custom.vision.nframes=8` for bounded 1-GPU memory;\nswitch to `fps` only after checking token budget and GPU memory, and delete\n`custom.vision.nframes` from the spec when you do.\n\nSee `references\u002Fcosmos-reason-launch.md` for launch intake, the full\n`check_tao_launch_preflight.py` slurm\u002Flocal-Docker examples, the\n`video_fps` preflight example, S3 staging, the GPU resource\u002Farchitecture gate,\nthe per-action dataset requirements table, the `\u002Fworkspace` mount caveat,\nthe quantize compatibility shim, and the full typical-overrides list.\n\n## Critical Overrides (Train)\n\nThese are the keys whose template defaults are wrong or where omission flips the run into a different mode:\n\n| Parameter | Template Default | Required Value | Why |\n|---|---|---|---|\n| `policy.model_name_or_path` | `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano` | Direct Docker: `nvidia\u002FCosmos3-Nano`, `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`, or a local HF snapshot path. SDK\u002Fmanaged platform predownload: `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`. | Keep the train and evaluate base model aligned. |\n| `policy.model_max_length` | 40960 | Keep at 40960 or higher | Smaller than ~40k causes `vision_embeds` shape mismatch on video inputs |\n| `train.train_batch_per_replica` | 32 | Any multiple of `train.train_policy.mini_batch` | Mismatch raises an immediate AssertionError |\n| `train.train_policy.type` | `\"sft\"` | Keep as `\"sft\"` for SFT workflows | If dropped during agent regeneration, cosmos-rl flips to RL mode → rollout replica allocated → multi-node attempted → hostname errors when `num_nodes=1` |\n\n## Evaluate\n\nThe evaluator reads a flat TOML config (`dataset`, `model`, `task`,\n`evaluation`, `vision`, `generation`, `metrics`, `results`, `num_gpus`,\n`results_dir`); the `actions.evaluate` block in `references\u002Fskill_info.yaml`\ndeclares inputs and outputs. See `references\u002Fcosmos-reason-evaluate.md` for the\nflat-TOML config detail, task types (`\"\"` General Evaluator vs\n`\"its_directionality\"`), LoRA evaluation via spec_overrides, selective download,\nresults\u002Fmetrics, and the datasets section.\n\n## AutoML \u002F HPO Notes\n\nThe packaged default base model is `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`; apply it\nconsistently to train (`policy.model_name_or_path`) and post-training evaluation\n(`model.base_model_path`) unless the user provides a different model. See\n`references\u002Fcosmos-reason-automl.md` for accuracy-vs-`val\u002Favg_loss` objective\nselection, the `eval_fn` per-recommendation evaluate flow, the knob mapping\n(learning rate, batch size, epochs, weight decay, warmup ratio), example\n`custom_param_ranges`, `train_sample_count` batch-size capping,\n`ordered_int` requirements, and the pre-launch recommendation summary.\n\n## Parameters, Hardware, Errors, DEFT, Inference\n\nFor parallelism, set `policy.parallelism.dp_shard_size` = GPUs per node and\n`policy.parallelism.dp_replicate_size` = node count (1 for single node).\nCosmos-RL handles distributed init internally via FSDP and does not rely on\nplatform-level `MASTER_ADDR`\u002F`WORLD_SIZE`; submit with\n`gpu_count=\u003Cgpus_per_node>` and `num_nodes=\u003CN>` and the spec keys drive\nsharding. Cosmos-RL models are 8B parameters; recommended 8x A100 or H100\n(80GB each).\n\nSee `references\u002Fcosmos-reason-parameters.md` for important parameters (training\nloop, model\u002Fpolicy, parallelism incl. multi-node FSDP, optimization, vision\nencoders, checkpointing incl. the `best` symlink\u002F`epoch_*` resolution,\nvalidation, logging), hardware sizing, the full error-pattern catalog (CUDA OOM,\nLoRA-eval OOM, NaN loss, `vision_embeds` mismatch, quantize token mismatch,\nbatch-size divisibility and per-rank limits, stale cache, scheduler-None,\ngated-repo `HF_TOKEN`, GPU resource\u002Farchitecture gate, status-logging warnings),\nDEFT support and `scripts\u002Fanalyze_gaps.py` gap analysis, and the parent-model\ninference mapping table.\n",{"data":35,"body":47},{"name":4,"description":6,"license":26,"compatibility":36,"metadata":37,"allowed-tools":40,"tags":41},"Requires docker + nvidia-container-toolkit.",{"author":38,"version":39},"NVIDIA Corporation","0.1.0","Read Bash",[22,42,43,44,45,46],"qa","cosmos","sft","reasoning","vlm",{"type":48,"children":49},"root",[50,59,82,119,132,137,186,193,286,292,401,436,442,472,478,499,504,623,628,638,643,649,862,868,903,970,1004,1032,1067,1073,1078,1277,1282,1396,1402,1475,1481,1532,1581],{"type":51,"tag":52,"props":53,"children":55},"element","h1",{"id":54},"cosmos-rl",[56],{"type":57,"value":58},"text","Cosmos-RL",{"type":51,"tag":60,"props":61,"children":62},"p",{},[63,65,71,73,80],{"type":57,"value":64},"Supervised fine-tuning (SFT) of Cosmos Reason video QA models. The packaged\ndefault base model is ",{"type":51,"tag":66,"props":67,"children":68},"strong",{},[69],{"type":57,"value":70},"hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano",{"type":57,"value":72},". Pretrained weights\nare sourced from HuggingFace, not NGC. Gated HuggingFace models require\n",{"type":51,"tag":74,"props":75,"children":77},"code",{"className":76},[],[78],{"type":57,"value":79},"HF_TOKEN",{"type":57,"value":81},". Some Cosmos-RL images cannot load the native Cosmos3 Omni checkpoint\nformat directly; for those images, convert Cosmos3-Nano to a Qwen3-VL HF\nsafetensors directory before train\u002Fevaluate and use that converted directory as\nthe PTM path.",{"type":51,"tag":60,"props":83,"children":84},{},[85,87,93,95,101,103,109,111,117],{"type":57,"value":86},"Uses FSDP-based parallelism with ",{"type":51,"tag":74,"props":88,"children":90},{"className":89},[],[91],{"type":57,"value":92},"dp_shard_size",{"type":57,"value":94}," for GPU count and ",{"type":51,"tag":74,"props":96,"children":98},{"className":97},[],[99],{"type":57,"value":100},"dp_replicate_size",{"type":57,"value":102}," for node count (not the standard ",{"type":51,"tag":74,"props":104,"children":106},{"className":105},[],[107],{"type":57,"value":108},"num_gpus",{"type":57,"value":110},"\u002F",{"type":51,"tag":74,"props":112,"children":114},{"className":113},[],[115],{"type":57,"value":116},"num_nodes",{"type":57,"value":118},").",{"type":51,"tag":60,"props":120,"children":121},{},[122,124,130],{"type":57,"value":123},"Requests for \"Cosmos Reason 3\", \"Cosmos3 Nano Reasoner\", or\n",{"type":51,"tag":74,"props":125,"children":127},{"className":126},[],[128],{"type":57,"value":129},"nvidia\u002FCosmos3-Nano",{"type":57,"value":131}," are handled by this skill. There is no separate Cosmos3\nmodel directory in the skill bank; route those requests here. Override the base\nHuggingFace model only when the user explicitly asks for a different model.",{"type":51,"tag":60,"props":133,"children":134},{},[135],{"type":57,"value":136},"Deep detail lives in references; load the smallest one that matches the task:",{"type":51,"tag":138,"props":139,"children":140},"ul",{},[141,153,164,175],{"type":51,"tag":142,"props":143,"children":144},"li",{},[145,151],{"type":51,"tag":74,"props":146,"children":148},{"className":147},[],[149],{"type":57,"value":150},"references\u002Fcosmos-reason-launch.md",{"type":57,"value":152}," — launch intake, preflight, per-action dataset requirements, spec construction, typical overrides.",{"type":51,"tag":142,"props":154,"children":155},{},[156,162],{"type":51,"tag":74,"props":157,"children":159},{"className":158},[],[160],{"type":57,"value":161},"references\u002Fcosmos-reason-evaluate.md",{"type":57,"value":163}," — evaluate (flat TOML, task types, LoRA eval, selective download, results) and datasets.",{"type":51,"tag":142,"props":165,"children":166},{},[167,173],{"type":51,"tag":74,"props":168,"children":170},{"className":169},[],[171],{"type":57,"value":172},"references\u002Fcosmos-reason-automl.md",{"type":57,"value":174}," — AutoML\u002FHPO policy and search-space guidance.",{"type":51,"tag":142,"props":176,"children":177},{},[178,184],{"type":51,"tag":74,"props":179,"children":181},{"className":180},[],[182],{"type":57,"value":183},"references\u002Fcosmos-reason-parameters.md",{"type":57,"value":185}," — important parameters, hardware, error patterns, DEFT\u002Fgap analysis, parent-model inference mappings.",{"type":51,"tag":187,"props":188,"children":190},"h2",{"id":189},"dataclass-schemas",[191],{"type":57,"value":192},"Dataclass Schemas",{"type":51,"tag":60,"props":194,"children":195},{},[196,198,204,206,212,214,220,222,228,230,236,238,244,246,252,254,260,262,268,270,276,278,284],{"type":57,"value":197},"Generated TAO Core schemas are packaged in ",{"type":51,"tag":74,"props":199,"children":201},{"className":200},[],[202],{"type":57,"value":203},"schemas\u002F\u003Caction>.schema.json",{"type":57,"value":205},", with ",{"type":51,"tag":74,"props":207,"children":209},{"className":208},[],[210],{"type":57,"value":211},"schemas\u002Fmanifest.json",{"type":57,"value":213}," listing available actions. Each generated schema also emits ",{"type":51,"tag":74,"props":215,"children":217},{"className":216},[],[218],{"type":57,"value":219},"references\u002Fspec_template_\u003Caction>.yaml",{"type":57,"value":221}," from the schema top-level ",{"type":51,"tag":74,"props":223,"children":225},{"className":224},[],[226],{"type":57,"value":227},"default",{"type":57,"value":229}," field. AutoML enablement is declared at the model layer in ",{"type":51,"tag":74,"props":231,"children":233},{"className":232},[],[234],{"type":57,"value":235},"references\u002Fskill_info.yaml",{"type":57,"value":237}," via ",{"type":51,"tag":74,"props":239,"children":241},{"className":240},[],[242],{"type":57,"value":243},"automl_enabled",{"type":57,"value":245},". Runnable AutoML still requires ",{"type":51,"tag":74,"props":247,"children":249},{"className":248},[],[250],{"type":57,"value":251},"schemas\u002Ftrain.schema.json",{"type":57,"value":253}," and ",{"type":51,"tag":74,"props":255,"children":257},{"className":256},[],[258],{"type":57,"value":259},"references\u002Fspec_template_train.yaml",{"type":57,"value":261}," to exist and parse. Use the packaged train schema for ",{"type":51,"tag":74,"props":263,"children":265},{"className":264},[],[266],{"type":57,"value":267},"automl_default_parameters",{"type":57,"value":269},", ",{"type":51,"tag":74,"props":271,"children":273},{"className":272},[],[274],{"type":57,"value":275},"automl_disabled_parameters",{"type":57,"value":277},", defaults, min\u002Fmax bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ",{"type":51,"tag":74,"props":279,"children":281},{"className":280},[],[282],{"type":57,"value":283},"~\u002Ftao-core",{"type":57,"value":285}," at runtime; maintainers regenerate schemas\u002Ftemplates before packaging the skill bank.",{"type":51,"tag":187,"props":287,"children":289},{"id":288},"train-action-policy",[290],{"type":57,"value":291},"Train Action Policy",{"type":51,"tag":60,"props":293,"children":294},{},[295,297,302,304,310,312,318,320,326,328,334,336,342,344,349,350,356,358,363,364,369,371,377,379,385,387,392,394,399],{"type":57,"value":296},"This model is AutoML-enabled at the model layer. Before handling any train-stage request, read ",{"type":51,"tag":74,"props":298,"children":300},{"className":299},[],[301],{"type":57,"value":235},{"type":57,"value":303}," and resolve the run override from either an explicit ",{"type":51,"tag":74,"props":305,"children":307},{"className":306},[],[308],{"type":57,"value":309},"automl_policy",{"type":57,"value":311}," value or the user's workflow request. Use ",{"type":51,"tag":74,"props":313,"children":315},{"className":314},[],[316],{"type":57,"value":317},"automl_policy: on",{"type":57,"value":319}," by default and only expose ",{"type":51,"tag":74,"props":321,"children":323},{"className":322},[],[324],{"type":57,"value":325},"on",{"type":57,"value":327}," \u002F ",{"type":51,"tag":74,"props":329,"children":331},{"className":330},[],[332],{"type":57,"value":333},"off",{"type":57,"value":335}," in new launch prompts. Treat phrases like \"turn off AutoML\", \"disable AutoML\", \"no HPO\", or \"plain training\" as ",{"type":51,"tag":74,"props":337,"children":339},{"className":338},[],[340],{"type":57,"value":341},"automl_policy: off",{"type":57,"value":343}," for this run only. When ",{"type":51,"tag":74,"props":345,"children":347},{"className":346},[],[348],{"type":57,"value":317},{"type":57,"value":269},{"type":51,"tag":74,"props":351,"children":353},{"className":352},[],[354],{"type":57,"value":355},"automl_enabled: true",{"type":57,"value":357},", and both ",{"type":51,"tag":74,"props":359,"children":361},{"className":360},[],[362],{"type":57,"value":251},{"type":57,"value":253},{"type":51,"tag":74,"props":365,"children":367},{"className":366},[],[368],{"type":57,"value":259},{"type":57,"value":370}," are packaged, route the train action through ",{"type":51,"tag":74,"props":372,"children":374},{"className":373},[],[375],{"type":57,"value":376},"tao-skill-bank:tao-run-automl",{"type":57,"value":378}," by default with this model's ",{"type":51,"tag":74,"props":380,"children":382},{"className":381},[],[383],{"type":57,"value":384},"skill_dir",{"type":57,"value":386},". Preserve workflow\u002Fapplication overrides for datasets, specs, output directories, GPU\u002Fplatform settings, parent checkpoints, and ",{"type":51,"tag":74,"props":388,"children":390},{"className":389},[],[391],{"type":57,"value":309},{"type":57,"value":393},". Use direct model training only when ",{"type":51,"tag":74,"props":395,"children":397},{"className":396},[],[398],{"type":57,"value":341},{"type":57,"value":400}," or the packaged train schema\u002Ftemplate is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.",{"type":51,"tag":60,"props":402,"children":403},{},[404,406,412,413,419,421,427,429,434],{"type":57,"value":405},"Non-train actions such as ",{"type":51,"tag":74,"props":407,"children":409},{"className":408},[],[410],{"type":57,"value":411},"evaluate",{"type":57,"value":269},{"type":51,"tag":74,"props":414,"children":416},{"className":415},[],[417],{"type":57,"value":418},"inference",{"type":57,"value":420},", and ",{"type":51,"tag":74,"props":422,"children":424},{"className":423},[],[425],{"type":57,"value":426},"quantize",{"type":57,"value":428}," stay in this model skill. The per-run ",{"type":51,"tag":74,"props":430,"children":432},{"className":431},[],[433],{"type":57,"value":309},{"type":57,"value":435}," override does not change model metadata.",{"type":51,"tag":187,"props":437,"children":439},{"id":438},"credentials",[440],{"type":57,"value":441},"Credentials",{"type":51,"tag":138,"props":443,"children":444},{},[445],{"type":51,"tag":142,"props":446,"children":447},{},[448,452,454,462,464,470],{"type":51,"tag":66,"props":449,"children":450},{},[451],{"type":57,"value":79},{"type":57,"value":453}," (required for gated models): HuggingFace access token. For the\npackaged default, the user must accept the model agreement at\n",{"type":51,"tag":455,"props":456,"children":460},"a",{"href":457,"rel":458},"https:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FCosmos3-Nano",[459],"nofollow",[461],{"type":57,"value":457},{"type":57,"value":463}," and provide a token with read\naccess. If the user explicitly overrides the base model, they must accept\nthat target model's agreement too. Passed to the container as a\n",{"type":51,"tag":74,"props":465,"children":467},{"className":466},[],[468],{"type":57,"value":469},"docker_env_var",{"type":57,"value":471},".",{"type":51,"tag":187,"props":473,"children":475},{"id":474},"cosmos3-checkpoint-conversion",[476],{"type":57,"value":477},"Cosmos3 Checkpoint Conversion",{"type":51,"tag":60,"props":479,"children":480},{},[481,483,489,491,497],{"type":57,"value":482},"When a selected image cannot load the native Cosmos3 checkpoint format\n(",{"type":51,"tag":74,"props":484,"children":486},{"className":485},[],[487],{"type":57,"value":488},"model_type=\"cosmos3_omni\"",{"type":57,"value":490}," or ",{"type":51,"tag":74,"props":492,"children":494},{"className":493},[],[495],{"type":57,"value":496},"Cosmos3ForConditionalGeneration",{"type":57,"value":498},"), do not patch\nQwenVL, Transformers, or vLLM first. Use the upstream Cosmos Framework VLM\nconversion path to produce a Qwen3-VL HF safetensors directory, then point\nCosmos-RL specs at that converted directory.",{"type":51,"tag":60,"props":500,"children":501},{},[502],{"type":57,"value":503},"The model skill packages a helper:",{"type":51,"tag":505,"props":506,"children":511},"pre",{"className":507,"code":508,"language":509,"meta":510,"style":510},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","python skills\u002Fmodels\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py \\\n  --checkpoint-path \u002Fabs\u002Fpath\u002FCosmos3-Nano \\\n  --output-path \u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM \\\n  --secrets-env ~\u002F.tao\u002Fsecrets.env \\\n  --validate-with-image \u003Ccosmos-rl-image>\n","bash","",[512],{"type":51,"tag":74,"props":513,"children":514},{"__ignoreMap":510},[515,539,557,575,593],{"type":51,"tag":516,"props":517,"children":520},"span",{"class":518,"line":519},"line",1,[521,527,533],{"type":51,"tag":516,"props":522,"children":524},{"style":523},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[525],{"type":57,"value":526},"python",{"type":51,"tag":516,"props":528,"children":530},{"style":529},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[531],{"type":57,"value":532}," skills\u002Fmodels\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py",{"type":51,"tag":516,"props":534,"children":536},{"style":535},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[537],{"type":57,"value":538}," \\\n",{"type":51,"tag":516,"props":540,"children":542},{"class":518,"line":541},2,[543,548,553],{"type":51,"tag":516,"props":544,"children":545},{"style":529},[546],{"type":57,"value":547},"  --checkpoint-path",{"type":51,"tag":516,"props":549,"children":550},{"style":529},[551],{"type":57,"value":552}," \u002Fabs\u002Fpath\u002FCosmos3-Nano",{"type":51,"tag":516,"props":554,"children":555},{"style":535},[556],{"type":57,"value":538},{"type":51,"tag":516,"props":558,"children":560},{"class":518,"line":559},3,[561,566,571],{"type":51,"tag":516,"props":562,"children":563},{"style":529},[564],{"type":57,"value":565},"  --output-path",{"type":51,"tag":516,"props":567,"children":568},{"style":529},[569],{"type":57,"value":570}," \u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM",{"type":51,"tag":516,"props":572,"children":573},{"style":535},[574],{"type":57,"value":538},{"type":51,"tag":516,"props":576,"children":578},{"class":518,"line":577},4,[579,584,589],{"type":51,"tag":516,"props":580,"children":581},{"style":529},[582],{"type":57,"value":583},"  --secrets-env",{"type":51,"tag":516,"props":585,"children":586},{"style":529},[587],{"type":57,"value":588}," ~\u002F.tao\u002Fsecrets.env",{"type":51,"tag":516,"props":590,"children":591},{"style":535},[592],{"type":57,"value":538},{"type":51,"tag":516,"props":594,"children":596},{"class":518,"line":595},5,[597,602,608,613,618],{"type":51,"tag":516,"props":598,"children":599},{"style":529},[600],{"type":57,"value":601},"  --validate-with-image",{"type":51,"tag":516,"props":603,"children":605},{"style":604},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[606],{"type":57,"value":607}," \u003C",{"type":51,"tag":516,"props":609,"children":610},{"style":529},[611],{"type":57,"value":612},"cosmos-rl-imag",{"type":51,"tag":516,"props":614,"children":615},{"style":535},[616],{"type":57,"value":617},"e",{"type":51,"tag":516,"props":619,"children":620},{"style":604},[621],{"type":57,"value":622},">\n",{"type":51,"tag":60,"props":624,"children":625},{},[626],{"type":57,"value":627},"After conversion, use the converted directory consistently as the PTM:",{"type":51,"tag":505,"props":629,"children":633},{"className":630,"code":632,"language":57,"meta":510},[631],"language-text","train:    policy.model_name_or_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.model_name=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.base_model_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\n",[634],{"type":51,"tag":74,"props":635,"children":636},{"__ignoreMap":510},[637],{"type":57,"value":632},{"type":51,"tag":60,"props":639,"children":640},{},[641],{"type":57,"value":642},"For local Docker, mount the converted directory read-only into the Cosmos-RL\ncontainer and set the spec to the container path. If a converted copy already\nexists and validates, reuse it for PTM baseline evaluation, AutoML\nrecommendations, and final best-checkpoint evaluation rather than converting\nagain.",{"type":51,"tag":187,"props":644,"children":646},{"id":645},"training-requirements",[647],{"type":57,"value":648},"Training Requirements",{"type":51,"tag":138,"props":650,"children":651},{},[652,662,672,682,692,729,769,795,805],{"type":51,"tag":142,"props":653,"children":654},{},[655,660],{"type":51,"tag":66,"props":656,"children":657},{},[658],{"type":57,"value":659},"Dataset type:",{"type":57,"value":661}," vlm",{"type":51,"tag":142,"props":663,"children":664},{},[665,670],{"type":51,"tag":66,"props":666,"children":667},{},[668],{"type":57,"value":669},"Formats:",{"type":57,"value":671}," llava, daft",{"type":51,"tag":142,"props":673,"children":674},{},[675,680],{"type":51,"tag":66,"props":676,"children":677},{},[678],{"type":57,"value":679},"Accepted dataset intents:",{"type":57,"value":681}," training, evaluation, testing",{"type":51,"tag":142,"props":683,"children":684},{},[685,690],{"type":51,"tag":66,"props":686,"children":687},{},[688],{"type":57,"value":689},"Monitoring metric:",{"type":57,"value":691}," val\u002Favg_loss, val\u002Freward_avg, val\u002Floss",{"type":51,"tag":142,"props":693,"children":694},{},[695,700,702,708,709,715,716,722,723],{"type":51,"tag":66,"props":696,"children":697},{},[698],{"type":57,"value":699},"Dataset URI examples:",{"type":57,"value":701}," ",{"type":51,"tag":74,"props":703,"children":705},{"className":704},[],[706],{"type":57,"value":707},"s3:\u002F\u002Fbucket\u002Fcosmos\u002Ftrain",{"type":57,"value":269},{"type":51,"tag":74,"props":710,"children":712},{"className":711},[],[713],{"type":57,"value":714},"s3:\u002F\u002Fbucket\u002Fcosmos\u002Feval",{"type":57,"value":269},{"type":51,"tag":74,"props":717,"children":719},{"className":718},[],[720],{"type":57,"value":721},"\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Ftrain",{"type":57,"value":269},{"type":51,"tag":74,"props":724,"children":726},{"className":725},[],[727],{"type":57,"value":728},"\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Feval",{"type":51,"tag":142,"props":730,"children":731},{},[732,737,739,745,747,753,755,761,762,768],{"type":51,"tag":66,"props":733,"children":734},{},[735],{"type":57,"value":736},"Input modes:",{"type":57,"value":738}," accept either dataset roots or direct spec-key paths. Root mode maps ",{"type":51,"tag":74,"props":740,"children":742},{"className":741},[],[743],{"type":57,"value":744},"\u003Croot>\u002Fannotations.json",{"type":57,"value":746}," plus ",{"type":51,"tag":74,"props":748,"children":750},{"className":749},[],[751],{"type":57,"value":752},"\u003Croot>",{"type":57,"value":754}," as the media path. Direct spec mode is valid when annotations and media live in different locations, for example ",{"type":51,"tag":74,"props":756,"children":758},{"className":757},[],[759],{"type":57,"value":760},"custom.train_dataset.annotation_path=\u002Flustre\u002F...\u002Ftrain.json",{"type":57,"value":253},{"type":51,"tag":74,"props":763,"children":765},{"className":764},[],[766],{"type":57,"value":767},"custom.train_dataset.media_path=\u002Flustre\u002F...\u002Fvideos.tar.gz",{"type":57,"value":471},{"type":51,"tag":142,"props":770,"children":771},{},[772,777,779,785,787,793],{"type":51,"tag":66,"props":773,"children":774},{},[775],{"type":57,"value":776},"Media handling:",{"type":57,"value":778}," do not ask the user to choose ",{"type":51,"tag":74,"props":780,"children":782},{"className":781},[],[783],{"type":57,"value":784},"videos.tar.gz",{"type":57,"value":786}," vs ",{"type":51,"tag":74,"props":788,"children":790},{"className":789},[],[791],{"type":57,"value":792},"images.tar.gz",{"type":57,"value":794}," unless they are using direct spec mode or the model\u002Faction requires a single media archive. In root mode, pass the dataset root as the media path.",{"type":51,"tag":142,"props":796,"children":797},{},[798,803],{"type":51,"tag":66,"props":799,"children":800},{},[801],{"type":57,"value":802},"Annotation validation:",{"type":57,"value":804}," before launching train\u002FAutoML\u002Fevaluate, verify the\nannotation JSON is readable and the referenced media path or archive is\nvisible from the selected platform. Do not block, patch, or mutate\nannotations solely because optional fields are absent.",{"type":51,"tag":142,"props":806,"children":807},{},[808,813,815,821,823,829,831,837,839,844,846,852,854,860],{"type":51,"tag":66,"props":809,"children":810},{},[811],{"type":57,"value":812},"Per-record video FPS:",{"type":57,"value":814}," the packaged train template uses\n",{"type":51,"tag":74,"props":816,"children":818},{"className":817},[],[819],{"type":57,"value":820},"custom.vision.nframes",{"type":57,"value":822},", so per-record ",{"type":51,"tag":74,"props":824,"children":826},{"className":825},[],[827],{"type":57,"value":828},"video_fps",{"type":57,"value":830}," is not required by\ndefault. If the user switches to ",{"type":51,"tag":74,"props":832,"children":834},{"className":833},[],[835],{"type":57,"value":836},"custom.vision.fps",{"type":57,"value":838},", selects a dataset\nprofile that requires per-record timing, or uses an image\u002Fversion that\nrequires ",{"type":51,"tag":74,"props":840,"children":842},{"className":841},[],[843],{"type":57,"value":828},{"type":57,"value":845},", make it a preflight requirement with\n",{"type":51,"tag":74,"props":847,"children":849},{"className":848},[],[850],{"type":57,"value":851},"--json-required-field train_annotation=video_fps",{"type":57,"value":853}," and\n",{"type":51,"tag":74,"props":855,"children":857},{"className":856},[],[858],{"type":57,"value":859},"--json-required-field val_annotation=video_fps",{"type":57,"value":861}," before any download or\njob launch.",{"type":51,"tag":187,"props":863,"children":865},{"id":864},"spec-construction",[866],{"type":57,"value":867},"Spec construction",{"type":51,"tag":60,"props":869,"children":870},{},[871,873,879,881,893,895,901],{"type":57,"value":872},"cosmos-rl is ",{"type":51,"tag":74,"props":874,"children":876},{"className":875},[],[877],{"type":57,"value":878},"mode: config",{"type":57,"value":880},". ",{"type":51,"tag":66,"props":882,"children":883},{},[884,886,891],{"type":57,"value":885},"Always start from the packaged\n",{"type":51,"tag":74,"props":887,"children":889},{"className":888},[],[890],{"type":57,"value":219},{"type":57,"value":892}," for the requested action",{"type":57,"value":894}," — load it\nas your base spec via ",{"type":51,"tag":74,"props":896,"children":898},{"className":897},[],[899],{"type":57,"value":900},"yaml.safe_load(...)",{"type":57,"value":902}," and apply user overrides on top.\nDon't rebuild from scratch.",{"type":51,"tag":505,"props":904,"children":907},{"className":905,"code":906,"language":526,"meta":510,"style":510},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","import yaml\nfrom pathlib import Path\n\nskill = Path.home() \u002F \"tao-sdk\u002Ftao-skills-external\u002Fskills\u002Fmodels\u002Ftao-finetune-cosmos-reason\"\naction = \"train\"  # train, evaluate, inference, or quantize\nspecs = yaml.safe_load((skill \u002F f\"references\u002Fspec_template_{action}.yaml\").read_text())\n# Now apply your overrides on top of `specs`.\n",[908],{"type":51,"tag":74,"props":909,"children":910},{"__ignoreMap":510},[911,919,927,936,944,952,961],{"type":51,"tag":516,"props":912,"children":913},{"class":518,"line":519},[914],{"type":51,"tag":516,"props":915,"children":916},{},[917],{"type":57,"value":918},"import yaml\n",{"type":51,"tag":516,"props":920,"children":921},{"class":518,"line":541},[922],{"type":51,"tag":516,"props":923,"children":924},{},[925],{"type":57,"value":926},"from pathlib import Path\n",{"type":51,"tag":516,"props":928,"children":929},{"class":518,"line":559},[930],{"type":51,"tag":516,"props":931,"children":933},{"emptyLinePlaceholder":932},true,[934],{"type":57,"value":935},"\n",{"type":51,"tag":516,"props":937,"children":938},{"class":518,"line":577},[939],{"type":51,"tag":516,"props":940,"children":941},{},[942],{"type":57,"value":943},"skill = Path.home() \u002F \"tao-sdk\u002Ftao-skills-external\u002Fskills\u002Fmodels\u002Ftao-finetune-cosmos-reason\"\n",{"type":51,"tag":516,"props":945,"children":946},{"class":518,"line":595},[947],{"type":51,"tag":516,"props":948,"children":949},{},[950],{"type":57,"value":951},"action = \"train\"  # train, evaluate, inference, or quantize\n",{"type":51,"tag":516,"props":953,"children":955},{"class":518,"line":954},6,[956],{"type":51,"tag":516,"props":957,"children":958},{},[959],{"type":57,"value":960},"specs = yaml.safe_load((skill \u002F f\"references\u002Fspec_template_{action}.yaml\").read_text())\n",{"type":51,"tag":516,"props":962,"children":964},{"class":518,"line":963},7,[965],{"type":51,"tag":516,"props":966,"children":967},{},[968],{"type":57,"value":969},"# Now apply your overrides on top of `specs`.\n",{"type":51,"tag":60,"props":971,"children":972},{},[973,975,980,982,988,990,996,998,1003],{"type":57,"value":974},"The reference TOML (and the spec the model actually consumes) is ",{"type":51,"tag":66,"props":976,"children":977},{},[978],{"type":57,"value":979},"nested\ndicts",{"type":57,"value":981},", not flat dotted keys. Dotted notation in override examples denotes\n",{"type":51,"tag":983,"props":984,"children":985},"em",{},[986],{"type":57,"value":987},"paths into the nested spec",{"type":57,"value":989}," — walk the path and assign at the leaf. See\n",{"type":51,"tag":74,"props":991,"children":993},{"className":992},[],[994],{"type":57,"value":995},"skills\u002Fplatform\u002Ftao-run-platform\u002FSKILL.md",{"type":57,"value":997},"'s \"spec is nested dicts\" callout.\nData source overrides are ",{"type":51,"tag":66,"props":999,"children":1000},{},[1001],{"type":57,"value":1002},"mandatory for every action",{"type":57,"value":471},{"type":51,"tag":60,"props":1005,"children":1006},{},[1007,1009,1015,1017,1023,1025,1030],{"type":57,"value":1008},"The packaged template keeps ",{"type":51,"tag":74,"props":1010,"children":1012},{"className":1011},[],[1013],{"type":57,"value":1014},"custom.vision.nframes=8",{"type":57,"value":1016}," for bounded 1-GPU memory;\nswitch to ",{"type":51,"tag":74,"props":1018,"children":1020},{"className":1019},[],[1021],{"type":57,"value":1022},"fps",{"type":57,"value":1024}," only after checking token budget and GPU memory, and delete\n",{"type":51,"tag":74,"props":1026,"children":1028},{"className":1027},[],[1029],{"type":57,"value":820},{"type":57,"value":1031}," from the spec when you do.",{"type":51,"tag":60,"props":1033,"children":1034},{},[1035,1037,1042,1044,1050,1052,1057,1059,1065],{"type":57,"value":1036},"See ",{"type":51,"tag":74,"props":1038,"children":1040},{"className":1039},[],[1041],{"type":57,"value":150},{"type":57,"value":1043}," for launch intake, the full\n",{"type":51,"tag":74,"props":1045,"children":1047},{"className":1046},[],[1048],{"type":57,"value":1049},"check_tao_launch_preflight.py",{"type":57,"value":1051}," slurm\u002Flocal-Docker examples, the\n",{"type":51,"tag":74,"props":1053,"children":1055},{"className":1054},[],[1056],{"type":57,"value":828},{"type":57,"value":1058}," preflight example, S3 staging, the GPU resource\u002Farchitecture gate,\nthe per-action dataset requirements table, the ",{"type":51,"tag":74,"props":1060,"children":1062},{"className":1061},[],[1063],{"type":57,"value":1064},"\u002Fworkspace",{"type":57,"value":1066}," mount caveat,\nthe quantize compatibility shim, and the full typical-overrides list.",{"type":51,"tag":187,"props":1068,"children":1070},{"id":1069},"critical-overrides-train",[1071],{"type":57,"value":1072},"Critical Overrides (Train)",{"type":51,"tag":60,"props":1074,"children":1075},{},[1076],{"type":57,"value":1077},"These are the keys whose template defaults are wrong or where omission flips the run into a different mode:",{"type":51,"tag":1079,"props":1080,"children":1081},"table",{},[1082,1111],{"type":51,"tag":1083,"props":1084,"children":1085},"thead",{},[1086],{"type":51,"tag":1087,"props":1088,"children":1089},"tr",{},[1090,1096,1101,1106],{"type":51,"tag":1091,"props":1092,"children":1093},"th",{},[1094],{"type":57,"value":1095},"Parameter",{"type":51,"tag":1091,"props":1097,"children":1098},{},[1099],{"type":57,"value":1100},"Template Default",{"type":51,"tag":1091,"props":1102,"children":1103},{},[1104],{"type":57,"value":1105},"Required Value",{"type":51,"tag":1091,"props":1107,"children":1108},{},[1109],{"type":57,"value":1110},"Why",{"type":51,"tag":1112,"props":1113,"children":1114},"tbody",{},[1115,1165,1200,1233],{"type":51,"tag":1087,"props":1116,"children":1117},{},[1118,1128,1136,1160],{"type":51,"tag":1119,"props":1120,"children":1121},"td",{},[1122],{"type":51,"tag":74,"props":1123,"children":1125},{"className":1124},[],[1126],{"type":57,"value":1127},"policy.model_name_or_path",{"type":51,"tag":1119,"props":1129,"children":1130},{},[1131],{"type":51,"tag":74,"props":1132,"children":1134},{"className":1133},[],[1135],{"type":57,"value":70},{"type":51,"tag":1119,"props":1137,"children":1138},{},[1139,1141,1146,1147,1152,1154,1159],{"type":57,"value":1140},"Direct Docker: ",{"type":51,"tag":74,"props":1142,"children":1144},{"className":1143},[],[1145],{"type":57,"value":129},{"type":57,"value":269},{"type":51,"tag":74,"props":1148,"children":1150},{"className":1149},[],[1151],{"type":57,"value":70},{"type":57,"value":1153},", or a local HF snapshot path. SDK\u002Fmanaged platform predownload: ",{"type":51,"tag":74,"props":1155,"children":1157},{"className":1156},[],[1158],{"type":57,"value":70},{"type":57,"value":471},{"type":51,"tag":1119,"props":1161,"children":1162},{},[1163],{"type":57,"value":1164},"Keep the train and evaluate base model aligned.",{"type":51,"tag":1087,"props":1166,"children":1167},{},[1168,1177,1182,1187],{"type":51,"tag":1119,"props":1169,"children":1170},{},[1171],{"type":51,"tag":74,"props":1172,"children":1174},{"className":1173},[],[1175],{"type":57,"value":1176},"policy.model_max_length",{"type":51,"tag":1119,"props":1178,"children":1179},{},[1180],{"type":57,"value":1181},"40960",{"type":51,"tag":1119,"props":1183,"children":1184},{},[1185],{"type":57,"value":1186},"Keep at 40960 or higher",{"type":51,"tag":1119,"props":1188,"children":1189},{},[1190,1192,1198],{"type":57,"value":1191},"Smaller than ~40k causes ",{"type":51,"tag":74,"props":1193,"children":1195},{"className":1194},[],[1196],{"type":57,"value":1197},"vision_embeds",{"type":57,"value":1199}," shape mismatch on video inputs",{"type":51,"tag":1087,"props":1201,"children":1202},{},[1203,1212,1217,1228],{"type":51,"tag":1119,"props":1204,"children":1205},{},[1206],{"type":51,"tag":74,"props":1207,"children":1209},{"className":1208},[],[1210],{"type":57,"value":1211},"train.train_batch_per_replica",{"type":51,"tag":1119,"props":1213,"children":1214},{},[1215],{"type":57,"value":1216},"32",{"type":51,"tag":1119,"props":1218,"children":1219},{},[1220,1222],{"type":57,"value":1221},"Any multiple of ",{"type":51,"tag":74,"props":1223,"children":1225},{"className":1224},[],[1226],{"type":57,"value":1227},"train.train_policy.mini_batch",{"type":51,"tag":1119,"props":1229,"children":1230},{},[1231],{"type":57,"value":1232},"Mismatch raises an immediate AssertionError",{"type":51,"tag":1087,"props":1234,"children":1235},{},[1236,1245,1254,1266],{"type":51,"tag":1119,"props":1237,"children":1238},{},[1239],{"type":51,"tag":74,"props":1240,"children":1242},{"className":1241},[],[1243],{"type":57,"value":1244},"train.train_policy.type",{"type":51,"tag":1119,"props":1246,"children":1247},{},[1248],{"type":51,"tag":74,"props":1249,"children":1251},{"className":1250},[],[1252],{"type":57,"value":1253},"\"sft\"",{"type":51,"tag":1119,"props":1255,"children":1256},{},[1257,1259,1264],{"type":57,"value":1258},"Keep as ",{"type":51,"tag":74,"props":1260,"children":1262},{"className":1261},[],[1263],{"type":57,"value":1253},{"type":57,"value":1265}," for SFT workflows",{"type":51,"tag":1119,"props":1267,"children":1268},{},[1269,1271],{"type":57,"value":1270},"If dropped during agent regeneration, cosmos-rl flips to RL mode → rollout replica allocated → multi-node attempted → hostname errors when ",{"type":51,"tag":74,"props":1272,"children":1274},{"className":1273},[],[1275],{"type":57,"value":1276},"num_nodes=1",{"type":51,"tag":187,"props":1278,"children":1279},{"id":411},[1280],{"type":57,"value":1281},"Evaluate",{"type":51,"tag":60,"props":1283,"children":1284},{},[1285,1287,1293,1294,1300,1301,1307,1309,1315,1316,1322,1323,1329,1330,1336,1337,1343,1344,1349,1350,1356,1358,1364,1366,1371,1373,1378,1380,1386,1388,1394],{"type":57,"value":1286},"The evaluator reads a flat TOML config (",{"type":51,"tag":74,"props":1288,"children":1290},{"className":1289},[],[1291],{"type":57,"value":1292},"dataset",{"type":57,"value":269},{"type":51,"tag":74,"props":1295,"children":1297},{"className":1296},[],[1298],{"type":57,"value":1299},"model",{"type":57,"value":269},{"type":51,"tag":74,"props":1302,"children":1304},{"className":1303},[],[1305],{"type":57,"value":1306},"task",{"type":57,"value":1308},",\n",{"type":51,"tag":74,"props":1310,"children":1312},{"className":1311},[],[1313],{"type":57,"value":1314},"evaluation",{"type":57,"value":269},{"type":51,"tag":74,"props":1317,"children":1319},{"className":1318},[],[1320],{"type":57,"value":1321},"vision",{"type":57,"value":269},{"type":51,"tag":74,"props":1324,"children":1326},{"className":1325},[],[1327],{"type":57,"value":1328},"generation",{"type":57,"value":269},{"type":51,"tag":74,"props":1331,"children":1333},{"className":1332},[],[1334],{"type":57,"value":1335},"metrics",{"type":57,"value":269},{"type":51,"tag":74,"props":1338,"children":1340},{"className":1339},[],[1341],{"type":57,"value":1342},"results",{"type":57,"value":269},{"type":51,"tag":74,"props":1345,"children":1347},{"className":1346},[],[1348],{"type":57,"value":108},{"type":57,"value":1308},{"type":51,"tag":74,"props":1351,"children":1353},{"className":1352},[],[1354],{"type":57,"value":1355},"results_dir",{"type":57,"value":1357},"); the ",{"type":51,"tag":74,"props":1359,"children":1361},{"className":1360},[],[1362],{"type":57,"value":1363},"actions.evaluate",{"type":57,"value":1365}," block in ",{"type":51,"tag":74,"props":1367,"children":1369},{"className":1368},[],[1370],{"type":57,"value":235},{"type":57,"value":1372},"\ndeclares inputs and outputs. See ",{"type":51,"tag":74,"props":1374,"children":1376},{"className":1375},[],[1377],{"type":57,"value":161},{"type":57,"value":1379}," for the\nflat-TOML config detail, task types (",{"type":51,"tag":74,"props":1381,"children":1383},{"className":1382},[],[1384],{"type":57,"value":1385},"\"\"",{"type":57,"value":1387}," General Evaluator vs\n",{"type":51,"tag":74,"props":1389,"children":1391},{"className":1390},[],[1392],{"type":57,"value":1393},"\"its_directionality\"",{"type":57,"value":1395},"), LoRA evaluation via spec_overrides, selective download,\nresults\u002Fmetrics, and the datasets section.",{"type":51,"tag":187,"props":1397,"children":1399},{"id":1398},"automl-hpo-notes",[1400],{"type":57,"value":1401},"AutoML \u002F HPO Notes",{"type":51,"tag":60,"props":1403,"children":1404},{},[1405,1407,1412,1414,1419,1421,1427,1429,1434,1436,1442,1444,1450,1452,1458,1459,1465,1467,1473],{"type":57,"value":1406},"The packaged default base model is ",{"type":51,"tag":74,"props":1408,"children":1410},{"className":1409},[],[1411],{"type":57,"value":70},{"type":57,"value":1413},"; apply it\nconsistently to train (",{"type":51,"tag":74,"props":1415,"children":1417},{"className":1416},[],[1418],{"type":57,"value":1127},{"type":57,"value":1420},") and post-training evaluation\n(",{"type":51,"tag":74,"props":1422,"children":1424},{"className":1423},[],[1425],{"type":57,"value":1426},"model.base_model_path",{"type":57,"value":1428},") unless the user provides a different model. See\n",{"type":51,"tag":74,"props":1430,"children":1432},{"className":1431},[],[1433],{"type":57,"value":172},{"type":57,"value":1435}," for accuracy-vs-",{"type":51,"tag":74,"props":1437,"children":1439},{"className":1438},[],[1440],{"type":57,"value":1441},"val\u002Favg_loss",{"type":57,"value":1443}," objective\nselection, the ",{"type":51,"tag":74,"props":1445,"children":1447},{"className":1446},[],[1448],{"type":57,"value":1449},"eval_fn",{"type":57,"value":1451}," per-recommendation evaluate flow, the knob mapping\n(learning rate, batch size, epochs, weight decay, warmup ratio), example\n",{"type":51,"tag":74,"props":1453,"children":1455},{"className":1454},[],[1456],{"type":57,"value":1457},"custom_param_ranges",{"type":57,"value":269},{"type":51,"tag":74,"props":1460,"children":1462},{"className":1461},[],[1463],{"type":57,"value":1464},"train_sample_count",{"type":57,"value":1466}," batch-size capping,\n",{"type":51,"tag":74,"props":1468,"children":1470},{"className":1469},[],[1471],{"type":57,"value":1472},"ordered_int",{"type":57,"value":1474}," requirements, and the pre-launch recommendation summary.",{"type":51,"tag":187,"props":1476,"children":1478},{"id":1477},"parameters-hardware-errors-deft-inference",[1479],{"type":57,"value":1480},"Parameters, Hardware, Errors, DEFT, Inference",{"type":51,"tag":60,"props":1482,"children":1483},{},[1484,1486,1492,1494,1500,1502,1508,1509,1515,1517,1523,1524,1530],{"type":57,"value":1485},"For parallelism, set ",{"type":51,"tag":74,"props":1487,"children":1489},{"className":1488},[],[1490],{"type":57,"value":1491},"policy.parallelism.dp_shard_size",{"type":57,"value":1493}," = GPUs per node and\n",{"type":51,"tag":74,"props":1495,"children":1497},{"className":1496},[],[1498],{"type":57,"value":1499},"policy.parallelism.dp_replicate_size",{"type":57,"value":1501}," = node count (1 for single node).\nCosmos-RL handles distributed init internally via FSDP and does not rely on\nplatform-level ",{"type":51,"tag":74,"props":1503,"children":1505},{"className":1504},[],[1506],{"type":57,"value":1507},"MASTER_ADDR",{"type":57,"value":110},{"type":51,"tag":74,"props":1510,"children":1512},{"className":1511},[],[1513],{"type":57,"value":1514},"WORLD_SIZE",{"type":57,"value":1516},"; submit with\n",{"type":51,"tag":74,"props":1518,"children":1520},{"className":1519},[],[1521],{"type":57,"value":1522},"gpu_count=\u003Cgpus_per_node>",{"type":57,"value":253},{"type":51,"tag":74,"props":1525,"children":1527},{"className":1526},[],[1528],{"type":57,"value":1529},"num_nodes=\u003CN>",{"type":57,"value":1531}," and the spec keys drive\nsharding. Cosmos-RL models are 8B parameters; recommended 8x A100 or H100\n(80GB each).",{"type":51,"tag":60,"props":1533,"children":1534},{},[1535,1536,1541,1543,1549,1551,1557,1559,1564,1566,1571,1573,1579],{"type":57,"value":1036},{"type":51,"tag":74,"props":1537,"children":1539},{"className":1538},[],[1540],{"type":57,"value":183},{"type":57,"value":1542}," for important parameters (training\nloop, model\u002Fpolicy, parallelism incl. multi-node FSDP, optimization, vision\nencoders, checkpointing incl. the ",{"type":51,"tag":74,"props":1544,"children":1546},{"className":1545},[],[1547],{"type":57,"value":1548},"best",{"type":57,"value":1550}," symlink\u002F",{"type":51,"tag":74,"props":1552,"children":1554},{"className":1553},[],[1555],{"type":57,"value":1556},"epoch_*",{"type":57,"value":1558}," resolution,\nvalidation, logging), hardware sizing, the full error-pattern catalog (CUDA OOM,\nLoRA-eval OOM, NaN loss, ",{"type":51,"tag":74,"props":1560,"children":1562},{"className":1561},[],[1563],{"type":57,"value":1197},{"type":57,"value":1565}," mismatch, quantize token mismatch,\nbatch-size divisibility and per-rank limits, stale cache, scheduler-None,\ngated-repo ",{"type":51,"tag":74,"props":1567,"children":1569},{"className":1568},[],[1570],{"type":57,"value":79},{"type":57,"value":1572},", GPU resource\u002Farchitecture gate, status-logging warnings),\nDEFT support and ",{"type":51,"tag":74,"props":1574,"children":1576},{"className":1575},[],[1577],{"type":57,"value":1578},"scripts\u002Fanalyze_gaps.py",{"type":57,"value":1580}," gap analysis, and the parent-model\ninference mapping table.",{"type":51,"tag":1582,"props":1583,"children":1584},"style",{},[1585],{"type":57,"value":1586},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":1588,"total":1690},[1589,1606,1620,1634,1646,1661,1676],{"slug":1590,"name":1590,"fn":1591,"description":1592,"org":1593,"tags":1594,"stars":23,"repoUrl":24,"updatedAt":1605},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1595,1598,1601,1602],{"name":1596,"slug":1597,"type":15},"Data Analysis","data-analysis",{"name":1599,"slug":1600,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":1603,"slug":1604,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":1607,"name":1607,"fn":1608,"description":1609,"org":1610,"tags":1611,"stars":23,"repoUrl":24,"updatedAt":1619},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1612,1615,1618],{"name":1613,"slug":1614,"type":15},"Deployment","deployment",{"name":1616,"slug":1617,"type":15},"Infrastructure","infrastructure",{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1621,"name":1621,"fn":1622,"description":1623,"org":1624,"tags":1625,"stars":23,"repoUrl":24,"updatedAt":1633},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1626,1629,1630],{"name":1627,"slug":1628,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1631,"slug":1632,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1635,"name":1635,"fn":1636,"description":1637,"org":1638,"tags":1639,"stars":23,"repoUrl":24,"updatedAt":1645},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1640,1641,1642],{"name":1596,"slug":1597,"type":15},{"name":9,"slug":8,"type":15},{"name":1643,"slug":1644,"type":15},"Testing","testing","2026-07-17T05:29:03.913266",{"slug":1647,"name":1647,"fn":1648,"description":1649,"org":1650,"tags":1651,"stars":23,"repoUrl":24,"updatedAt":1660},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1652,1655,1658,1659],{"name":1653,"slug":1654,"type":15},"Automation","automation",{"name":1656,"slug":1657,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-17T05:28:53.905004",{"slug":1662,"name":1662,"fn":1663,"description":1664,"org":1665,"tags":1666,"stars":23,"repoUrl":24,"updatedAt":1675},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1667,1668,1671,1672],{"name":1613,"slug":1614,"type":15},{"name":1669,"slug":1670,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1673,"slug":1674,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1677,"name":1677,"fn":1678,"description":1679,"org":1680,"tags":1681,"stars":23,"repoUrl":24,"updatedAt":1689},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1682,1683,1686],{"name":9,"slug":8,"type":15},{"name":1684,"slug":1685,"type":15},"Quantum Computing","quantum-computing",{"name":1687,"slug":1688,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305,{"items":1692,"total":1839},[1693,1711,1726,1737,1749,1763,1776,1788,1798,1807,1821,1830],{"slug":1694,"name":1694,"fn":1695,"description":1696,"org":1697,"tags":1698,"stars":1708,"repoUrl":1709,"updatedAt":1710},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1699,1702,1705],{"name":1700,"slug":1701,"type":15},"Documentation","documentation",{"name":1703,"slug":1704,"type":15},"MCP","mcp",{"name":1706,"slug":1707,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1712,"name":1712,"fn":1713,"description":1714,"org":1715,"tags":1716,"stars":1723,"repoUrl":1724,"updatedAt":1725},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1717,1720,1721],{"name":1718,"slug":1719,"type":15},"Containers","containers",{"name":1613,"slug":1614,"type":15},{"name":1722,"slug":526,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1727,"name":1727,"fn":1728,"description":1729,"org":1730,"tags":1731,"stars":1723,"repoUrl":1724,"updatedAt":1736},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1732,1735],{"name":1733,"slug":1734,"type":15},"CI\u002FCD","ci-cd",{"name":1613,"slug":1614,"type":15},"2026-07-14T05:25:59.97109",{"slug":1738,"name":1738,"fn":1739,"description":1740,"org":1741,"tags":1742,"stars":1723,"repoUrl":1724,"updatedAt":1748},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1743,1744,1745],{"name":1733,"slug":1734,"type":15},{"name":1613,"slug":1614,"type":15},{"name":1746,"slug":1747,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1750,"name":1750,"fn":1751,"description":1752,"org":1753,"tags":1754,"stars":1723,"repoUrl":1724,"updatedAt":1762},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1755,1758,1759],{"name":1756,"slug":1757,"type":15},"Debugging","debugging",{"name":1746,"slug":1747,"type":15},{"name":1760,"slug":1761,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1764,"name":1764,"fn":1765,"description":1766,"org":1767,"tags":1768,"stars":1723,"repoUrl":1724,"updatedAt":1775},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1769,1772],{"name":1770,"slug":1771,"type":15},"Best Practices","best-practices",{"name":1773,"slug":1774,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1777,"name":1777,"fn":1778,"description":1779,"org":1780,"tags":1781,"stars":1723,"repoUrl":1724,"updatedAt":1787},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1782,1783,1786],{"name":17,"slug":18,"type":15},{"name":1784,"slug":1785,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1789,"name":1789,"fn":1790,"description":1791,"org":1792,"tags":1793,"stars":1723,"repoUrl":1724,"updatedAt":1797},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1794,1796],{"name":1795,"slug":42,"type":15},"QA",{"name":1643,"slug":1644,"type":15},"2026-07-14T05:25:53.673039",{"slug":1799,"name":1799,"fn":1800,"description":1801,"org":1802,"tags":1803,"stars":1723,"repoUrl":1724,"updatedAt":1806},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1804,1805],{"name":1613,"slug":1614,"type":15},{"name":1616,"slug":1617,"type":15},"2026-07-14T05:25:49.362534",{"slug":1808,"name":1808,"fn":1809,"description":1810,"org":1811,"tags":1812,"stars":1723,"repoUrl":1724,"updatedAt":1820},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1813,1816,1817],{"name":1814,"slug":1815,"type":15},"Code Review","code-review",{"name":1746,"slug":1747,"type":15},{"name":1818,"slug":1819,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1822,"name":1822,"fn":1823,"description":1824,"org":1825,"tags":1826,"stars":1723,"repoUrl":1724,"updatedAt":1829},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1827,1828],{"name":1795,"slug":42,"type":15},{"name":1643,"slug":1644,"type":15},"2026-07-14T05:25:54.928983",{"slug":1831,"name":1831,"fn":1832,"description":1833,"org":1834,"tags":1835,"stars":1723,"repoUrl":1724,"updatedAt":1838},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[1836,1837],{"name":1653,"slug":1654,"type":15},{"name":1733,"slug":1734,"type":15},"2026-07-30T05:29:03.275638",496]