[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-tao-finetune-cosmos-embed":3,"mdc--hack7x-key":34,"related-org-nvidia-tao-finetune-cosmos-embed":3414,"related-repo-nvidia-tao-finetune-cosmos-embed":3572},{"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-embed","fine-tune Cosmos-Embed video models","Cosmos-Embed1 video-text embedding for text-to-video retrieval, video-to-video search, semantic deduplication, and fine-tuning. Use when the user asks to \"fine-tune Cosmos-Embed1\", \"run cosmos-embed inference\", \"export Cosmos-Embed1\", \"embed videos\", or \"search videos with text\".",{"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","2026-07-14T05:29:20.68","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-embed","---\nname: tao-finetune-cosmos-embed\ndescription: >-\n  Cosmos-Embed1 video-text embedding for text-to-video retrieval, video-to-video search, semantic deduplication, and\n  fine-tuning. Use when the user asks to \"fine-tune Cosmos-Embed1\", \"run cosmos-embed inference\", \"export Cosmos-Embed1\",\n  \"embed videos\", or \"search videos with text\".\nlicense: Apache-2.0\ncompatibility: Requires docker + nvidia-container-toolkit, the published Cosmos-Embed TAO container from versions.yaml, and a HuggingFace token when downloading pretrained `nvidia\u002FCosmos-Embed1-*` weights.\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\nallowed-tools: Read Bash\ntags:\n- video\n- vision-language\n- vlm\n- multimodal\n- retrieval\n- embedding\n- cosmos\n- fine-tuning\n---\n\n# Cosmos-Embed\n\nCosmos-Embed1 is a joint video-text embedder for text-to-video retrieval, video-to-video search, zero-shot\u002FkNN classification, and semantic deduplication. The packaged CLI is `cosmos-embed1` and supports `train`, `evaluate`, `inference`, and `export`.\n\nContainer image and per-action commands are in `references\u002Fskill_info.yaml`. Compact starting specs are in `references\u002Fspec_template_*.yaml`.\n\n## Train Action Policy\n\nAutoML is not packaged for this model skill because there are no Cosmos-Embed schemas under `schemas\u002F`. Always use the direct model skill actions for `train`, `evaluate`, `inference`, and `export`, even when a higher-level request includes `automl_policy: on`. Do not route Cosmos-Embed through workflow or AutoML skills until model-specific train schemas and templates are added.\n\nNon-train actions such as `evaluate`, `inference`, `export`, and deploy flows stay in this model skill. The per-run `automl_policy` override does not change model metadata.\n\n## Quick Start\n\nUse the published Cosmos-Embed container declared by `references\u002Fskill_info.yaml`\nand resolved through `versions.yaml`. Do not build from the private\nCosmos-Embed1 source tree for normal skill use; build from source only when\ndeveloping the container itself.\n\n```bash\nTAO_SKILL_BANK_PATH=\"${TAO_SKILL_BANK_PATH:-$PWD}\"\nCOSMOS_EMBED_IMAGE=\"${COSMOS_EMBED_IMAGE:-$(\n  python \"$TAO_SKILL_BANK_PATH\u002Fscripts\u002Fresolve_tao_image.py\" \\\n    --skill-bank \"$TAO_SKILL_BANK_PATH\" \\\n    --model tao-finetune-cosmos-embed \\\n    --action train \\\n    --format json |\n  python -c 'import json,sys; print(json.load(sys.stdin)[\"image\"])'\n)}\"\ndocker pull \"$COSMOS_EMBED_IMAGE\"\n```\n\nExpected local workspace layout:\n\n```text\nworkspace\u002F\n├── data\u002F\n│   ├── msrvtt_test_1k.json\n│   └── video\u002F\n│       ├── video7020.mp4\n│       └── ...\n├── model\u002F\n│   └── Cosmos-Embed1-224p\u002F        # optional if using HF repo id\n├── specs\u002F\n│   ├── train.yaml\n│   ├── evaluate.yaml\n│   ├── inference.yaml\n│   ├── export_onnx.yaml\n│   └── export_hf.yaml\n└── results\u002F\n```\n\nUse these Docker options for all actions unless the local Docker\u002Fplatform skill gives a stricter environment-specific command:\n\n```bash\nTAO_SKILL_BANK_PATH=\"${TAO_SKILL_BANK_PATH:-$PWD}\"\nCOSMOS_EMBED_IMAGE=\"${COSMOS_EMBED_IMAGE:-$(\n  python \"$TAO_SKILL_BANK_PATH\u002Fscripts\u002Fresolve_tao_image.py\" \\\n    --skill-bank \"$TAO_SKILL_BANK_PATH\" \\\n    --model tao-finetune-cosmos-embed \\\n    --action train \\\n    --format json |\n  python -c 'import json,sys; print(json.load(sys.stdin)[\"image\"])'\n)}\"\nRUN_ROOT=\"${RUN_ROOT:-$PWD}\"\nDOCKER_COMMON=(\n  --rm --gpus all --ipc=host --network=host\n  --shm-size=64g\n  --ulimit memlock=-1\n  --ulimit stack=67108864\n  -e HF_TOKEN\n  -e WANDB_DISABLED=true\n  -e WANDB_MODE=disabled\n  -e HUGGINGFACE_HUB_CACHE=\u002Fhf_cache\n  -v \"$RUN_ROOT\u002Fdata:\u002Fdata:ro\"\n  -v \"$RUN_ROOT\u002Fmodel:\u002Fmodel\"\n  -v \"$RUN_ROOT\u002Fspecs:\u002Fspecs:ro\"\n  -v \"$RUN_ROOT\u002Fresults:\u002Fresults\"\n  -v \"$RUN_ROOT\u002Fhf_cache:\u002Fhf_cache\"\n)\n```\n\nFor Cosmos-Embed images that ship `protobuf==7.x`, run a small startup\npreamble before every action:\n\n```bash\npython -m pip install \"protobuf\u003C7\"\n```\n\nThe image contains `wandb==0.21.0` with `protobuf==7.x`; importing W&B fails before training\u002Fevaluation unless protobuf is pinned below 7. Use `WANDB_DISABLED=true` and `WANDB_MODE=disabled` for smoke or offline runs. Cosmos-Embed may still download the public `google-bert\u002Fbert-base-uncased` Q-Former component even when the model checkpoint is disabled, so pass `HF_TOKEN` as an environment variable or mount a persistent HuggingFace cache. Do not write the token into specs, logs, or reports.\n\nTrain:\n\n```bash\ndocker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 train -e \u002Fspecs\u002Ftrain.yaml results_dir=\u002Fresults\"\n```\n\nEvaluate:\n\n```bash\ndocker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 evaluate -e \u002Fspecs\u002Fevaluate.yaml results_dir=\u002Fresults\"\n```\n\nInference:\n\n```bash\ndocker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 inference -e \u002Fspecs\u002Finference.yaml \\\n  'inference.query.input_texts=[\\\"a man is singing on stage\\\"]' \\\n  inference.k=5 \\\n  results_dir=\u002Fresults\"\n```\n\nExport ONNX:\n\n```bash\ndocker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_onnx.yaml \\\n  export.checkpoint=\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_000000001.pt \\\n  export.onnx_file=\u002Fresults\u002Fexport\u002Fcosmos_embed1_combined.onnx \\\n  results_dir=\u002Fresults\"\n```\n\nExport HuggingFace format:\n\n```bash\ndocker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_hf.yaml \\\n  export.checkpoint=\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_000000001.pt \\\n  export.hf_output_dir=\u002Fresults\u002Fexport_hf\u002Fcosmos_embed1_hf \\\n  results_dir=\u002Fresults\"\n```\n\n## Smoke Overrides\n\nFor a small functional check, keep the same specs and override the expensive knobs:\n\n```bash\ntrain.max_iter=1\ntrain.validation_iter=2\ntrain.checkpoint_iter=1\ntrain.optim.optim=adamw\ntrain.optim.warmup_steps=0\ntrain.optim.lr_decay_iters=1\ndataset.train_dataset.batch_size=1\ndataset.val_dataset.batch_size=1\ndataset.train_dataset.workers=0\ndataset.val_dataset.workers=0\n```\n\nWhen shortening the cosine scheduler for smoke runs, keep\n`train.optim.lr_decay_iters` greater than `train.optim.warmup_steps`, or set\n`train.optim.warmup_steps=0` as shown above. The scheduler divides by\n`lr_decay_iters - warmup_steps`, so equal values fail before the checkpoint is\nwritten.\n\nIf no local Cosmos-Embed1 pretrained checkpoint is available, set `model.pretrained_model_path=null` for a plumbing-only smoke train. The model quality is meaningless in that mode, but the train\u002Fevaluate\u002Finference\u002Fexport action paths can still be exercised. In the current container, the Q-Former path can still fetch `google-bert\u002Fbert-base-uncased`; provide `HF_TOKEN` or a mounted HuggingFace cache for fresh ephemeral containers.\n\nFor evaluation and inference smoke tests on a tiny subset:\n\n```bash\nevaluate.callbacks.embedding_visualization=false\nevaluate.callbacks.max_eval_samples=8\ndataset.test_dataset.batch_size=1\ndataset.test_dataset.workers=0\ninference.k=2\ndataset.inference_dataset.batch_size=1\ndataset.inference_dataset.workers=0\n```\n\n## Data Format\n\nThe MSR-VTT path expects a local video glob and a JSON metadata file:\n\n```yaml\ndataset:\n  train_dataset:\n    dataset_type: msrvtt\n    mp4_urls: \u002Fdata\u002Fvideo\u002F*.mp4\n    metadata: \u002Fdata\u002Fmsrvtt_test_1k.json\n```\n\nList-format metadata rows must include at least `video` and `caption`:\n\n```json\n{\"video_id\": \"video7020\", \"video\": \"video7020.mp4\", \"caption\": \"a woman creating a fondant baby and flower\"}\n```\n\nThe dataset loader derives the video id from the local `.mp4` filename and filters to videos present in the metadata. If a run finds zero videos, check that `mp4_urls` points to a container-local glob and that metadata `video` names match the filenames.\n\n## Model Weights\n\n- Local HF directory: mount it under `\u002Fmodel` and set `model.pretrained_model_path=\u002Fmodel\u002FCosmos-Embed1-224p`.\n- HuggingFace repo: set `model.pretrained_model_path=nvidia\u002FCosmos-Embed1-224p` and pass `HF_TOKEN` if access is gated.\n- Fine-tuned checkpoint: set downstream actions to the resolver-selected `\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_#########.pt` file.\n\nTraining writes full checkpoints under `results\u002Ftrain\u002Fcheckpoints\u002Fiter_#########.pt`, updates `results\u002Ftrain\u002Fcheckpoints\u002Flatest_checkpoint.txt`, and creates a `cosmos_embed1_model_latest.pth` symlink. For `evaluate.checkpoint`, `inference.checkpoint`, `export.checkpoint`, and `train.resume_training_checkpoint_path`, resolve and pass the exact `iter_#########.pt` file for the intended iteration. The action spec templates intentionally leave these checkpoint fields null so the model-skill runner or the user must provide the resolver-selected checkpoint. Use the latest symlink only when the user explicitly asks for latest.\n\nFor single-GPU resume\u002Fretrain from a consolidated checkpoint, set `model.fsdp_shard_size: 1`. The container default is 8, which sends resumed training through an FSDP apply path that Cosmos-Embed1 does not implement for this model class.\n\nVariants:\n\n| Variant | Resolution | Frames | Embedding dim |\n|---|---:|---:|---:|\n| `Cosmos-Embed1-224p` | 224 x 224 | 8 | 256 |\n| `Cosmos-Embed1-336p` | 336 x 336 | 8 | 768 |\n| `Cosmos-Embed1-448p` | 448 x 448 | 8 | 768 |\n\nKeep `model.network.embed_dim`, `model.input_hw`, and `model.network.spatial_resolution` aligned with the selected variant.\n\n## Important Parameters\n\n| Parameter | Notes |\n|---|---|\n| `train.num_gpus` | `1` for single GPU, `>1` auto-launches `torchrun`, `-1` auto-detects visible GPUs. |\n| `train.max_iter` | Main training length. Use `1` only for smoke testing. |\n| `train.optim.optim` | `fused_adamw` is faster when available; `adamw` is safer for smoke and portability. |\n| `model.lora.enabled` | Enables LoRA. Set `model.network.visual_encoder.transformer_engine=false` when LoRA is on. |\n| `model.lora.lora_rank` | LoRA rank. Start with `8`; try `4`, `8`, or `16` for manual or AutoML-style sweeps. |\n| `model.lora.lora_alpha` | LoRA scaling factor. Start with `16`; keep near `2 * lora_rank` unless experiments show otherwise. |\n| `model.lora.lora_dropout` | LoRA dropout. Start with `0.1`; sweep `0.0`, `0.05`, and `0.1` for small datasets. |\n| `model.lora.bias` | Bias policy: `none`, `all`, or `lora_only`. Keep `none` unless intentionally training biases. |\n| `model.lora.use_rslora` \u002F `use_dora` | Optional LoRA variants. Enable one at a time and record the setting with the checkpoint. |\n| `model.lora.target_modules` | Optional module-name patterns for LoRA injection. Leave empty for the default ViT + Q-Former attention\u002FMLP targets. |\n| `model.lora.modules_to_save` | Optional modules to keep fully trainable alongside LoRA. Leave empty unless preserving a task-specific head. |\n| `evaluate.load_dataset_pkl` \u002F `save_dataset_pkl` | Cache evaluation embeddings. |\n| `inference.load_dataset_pkl` \u002F `save_dataset_pkl` | Cache the search database for repeated retrieval. |\n| `export.mode` | `video`, `text`, `combined`, or `huggingface`. |\n| `export.on_cpu` | Recommended for export to avoid device mismatch issues. |\n\n### LoRA and AutoML Notes\n\nFor parameter-efficient fine-tuning, set `model.lora.enabled=true` and keep\n`model.network.visual_encoder.transformer_engine=false`; TAO Core's\nCosmos-Embed1 config notes that PEFT cannot inject adapters into Transformer\nEngine layers. Treat the LoRA fields above as the first candidate parameters\nfor manual tuning or AutoML-style search before unfreezing larger model blocks.\nAvoid changing `target_modules` or `modules_to_save` unless the user explicitly\nneeds custom adapter placement.\n\n## S3 Staging\n\nThe Cosmos-Embed1 CLI consumes local paths and Python globs, not raw `s3:\u002F\u002F...\u002F*.mp4` URIs. For S3-backed runs, first stage a subset or full dataset to the execution host\u002Fcontainer filesystem, then use local paths such as `\u002Fdata\u002Fvideo\u002F*.mp4` in the spec.\n\nRecommended S3 layout for staged MSR-VTT data:\n\n```text\ns3:\u002F\u002Fbucket\u002Fpath\u002Fcosmos-embed\u002Fmsrvtt-subset\u002F\n├── msrvtt_test_1k.json\n└── video\u002F\n    ├── video7020.mp4\n    └── ...\n```\n\nAfter downloading\u002Fsyncing that prefix into the mounted `data\u002F` directory, use the same Docker commands above.\n\n## Outputs\n\n```text\nresults\u002F\n├── train\u002F\n│   ├── cosmos_embed1_model_latest.pth\n│   ├── cosmos_embed1_model_\u003Citer>.pth\n│   └── experiment.yaml\n├── evaluate\u002F\n│   ├── metrics.json\n│   └── experiment.yaml\n├── inference\u002F\n│   ├── results.json\n│   └── experiment.yaml\n├── export\u002F\n│   ├── cosmos_embed1_combined.onnx\n│   └── export_config.yaml\n└── export_hf\u002F\n    └── cosmos_embed1_hf\u002F\n```\n\n## Known Pitfalls\n\n| Symptom | Cause | Fix |\n|---|---|---|\n| `MSRVTTDataset: 0 videos found` | `mp4_urls` is not a local glob or metadata filenames do not match videos. | Mount data into the container and set `mp4_urls=\u002Fdata\u002Fvideo\u002F*.mp4`. |\n| HF download\u002Fauth failure | Missing or invalid `HF_TOKEN`, or model agreement not accepted. | Accept the model terms and pass `-e HF_TOKEN`. |\n| `cannot import name 'Imports' from 'wandb.proto.wandb_telemetry_pb2'` | `wandb==0.21.0` in the container is incompatible with `protobuf==7.x`. | Run `python -m pip install \"protobuf\u003C7\"` in the container before invoking `cosmos-embed1`. |\n| Resume fails with `Model does not implement 'apply_fsdp'` | Single-GPU resume loaded a consolidated checkpoint while `model.fsdp_shard_size` stayed at the default 8. | Set `model.fsdp_shard_size=1` for local single-GPU resume\u002Fretrain. |\n| LoRA injection failure | Transformer Engine visual encoder is enabled. | Set `model.network.visual_encoder.transformer_engine=false`. |\n| ONNX\u002FHF export complains about missing components | Export checkpoint is partial or adapter-only. | Use a full checkpoint or configure pretrained visual\u002Ftext sources before export. |\n| CUDA OOM | Batch\u002Fresolution too high for the GPU. | Reduce batch size, use 224p, enable LoRA, or use more GPUs. |\n",{"data":35,"body":49},{"name":4,"description":6,"license":26,"compatibility":36,"metadata":37,"allowed-tools":40,"tags":41},"Requires docker + nvidia-container-toolkit, the published Cosmos-Embed TAO container from versions.yaml, and a HuggingFace token when downloading pretrained `nvidia\u002FCosmos-Embed1-*` weights.",{"author":38,"version":39},"NVIDIA Corporation","0.1.0","Read Bash",[22,42,43,44,45,46,47,48],"vision-language","vlm","multimodal","retrieval","embedding","cosmos","fine-tuning",{"type":50,"children":51},"root",[52,61,107,127,134,180,212,218,238,506,511,521,526,1090,1103,1146,1198,1203,1278,1283,1353,1358,1484,1489,1594,1599,1703,1709,1714,1845,1880,1907,1912,2010,2016,2021,2108,2126,2245,2273,2279,2337,2403,2416,2421,2538,2565,2571,3051,3058,3094,3100,3121,3126,3135,3148,3154,3163,3169,3408],{"type":53,"tag":54,"props":55,"children":57},"element","h1",{"id":56},"cosmos-embed",[58],{"type":59,"value":60},"text","Cosmos-Embed",{"type":53,"tag":62,"props":63,"children":64},"p",{},[65,67,74,76,82,84,90,91,97,99,105],{"type":59,"value":66},"Cosmos-Embed1 is a joint video-text embedder for text-to-video retrieval, video-to-video search, zero-shot\u002FkNN classification, and semantic deduplication. The packaged CLI is ",{"type":53,"tag":68,"props":69,"children":71},"code",{"className":70},[],[72],{"type":59,"value":73},"cosmos-embed1",{"type":59,"value":75}," and supports ",{"type":53,"tag":68,"props":77,"children":79},{"className":78},[],[80],{"type":59,"value":81},"train",{"type":59,"value":83},", ",{"type":53,"tag":68,"props":85,"children":87},{"className":86},[],[88],{"type":59,"value":89},"evaluate",{"type":59,"value":83},{"type":53,"tag":68,"props":92,"children":94},{"className":93},[],[95],{"type":59,"value":96},"inference",{"type":59,"value":98},", and ",{"type":53,"tag":68,"props":100,"children":102},{"className":101},[],[103],{"type":59,"value":104},"export",{"type":59,"value":106},".",{"type":53,"tag":62,"props":108,"children":109},{},[110,112,118,120,126],{"type":59,"value":111},"Container image and per-action commands are in ",{"type":53,"tag":68,"props":113,"children":115},{"className":114},[],[116],{"type":59,"value":117},"references\u002Fskill_info.yaml",{"type":59,"value":119},". Compact starting specs are in ",{"type":53,"tag":68,"props":121,"children":123},{"className":122},[],[124],{"type":59,"value":125},"references\u002Fspec_template_*.yaml",{"type":59,"value":106},{"type":53,"tag":128,"props":129,"children":131},"h2",{"id":130},"train-action-policy",[132],{"type":59,"value":133},"Train Action Policy",{"type":53,"tag":62,"props":135,"children":136},{},[137,139,145,147,152,153,158,159,164,165,170,172,178],{"type":59,"value":138},"AutoML is not packaged for this model skill because there are no Cosmos-Embed schemas under ",{"type":53,"tag":68,"props":140,"children":142},{"className":141},[],[143],{"type":59,"value":144},"schemas\u002F",{"type":59,"value":146},". Always use the direct model skill actions for ",{"type":53,"tag":68,"props":148,"children":150},{"className":149},[],[151],{"type":59,"value":81},{"type":59,"value":83},{"type":53,"tag":68,"props":154,"children":156},{"className":155},[],[157],{"type":59,"value":89},{"type":59,"value":83},{"type":53,"tag":68,"props":160,"children":162},{"className":161},[],[163],{"type":59,"value":96},{"type":59,"value":98},{"type":53,"tag":68,"props":166,"children":168},{"className":167},[],[169],{"type":59,"value":104},{"type":59,"value":171},", even when a higher-level request includes ",{"type":53,"tag":68,"props":173,"children":175},{"className":174},[],[176],{"type":59,"value":177},"automl_policy: on",{"type":59,"value":179},". Do not route Cosmos-Embed through workflow or AutoML skills until model-specific train schemas and templates are added.",{"type":53,"tag":62,"props":181,"children":182},{},[183,185,190,191,196,197,202,204,210],{"type":59,"value":184},"Non-train actions such as ",{"type":53,"tag":68,"props":186,"children":188},{"className":187},[],[189],{"type":59,"value":89},{"type":59,"value":83},{"type":53,"tag":68,"props":192,"children":194},{"className":193},[],[195],{"type":59,"value":96},{"type":59,"value":83},{"type":53,"tag":68,"props":198,"children":200},{"className":199},[],[201],{"type":59,"value":104},{"type":59,"value":203},", and deploy flows stay in this model skill. The per-run ",{"type":53,"tag":68,"props":205,"children":207},{"className":206},[],[208],{"type":59,"value":209},"automl_policy",{"type":59,"value":211}," override does not change model metadata.",{"type":53,"tag":128,"props":213,"children":215},{"id":214},"quick-start",[216],{"type":59,"value":217},"Quick Start",{"type":53,"tag":62,"props":219,"children":220},{},[221,223,228,230,236],{"type":59,"value":222},"Use the published Cosmos-Embed container declared by ",{"type":53,"tag":68,"props":224,"children":226},{"className":225},[],[227],{"type":59,"value":117},{"type":59,"value":229},"\nand resolved through ",{"type":53,"tag":68,"props":231,"children":233},{"className":232},[],[234],{"type":59,"value":235},"versions.yaml",{"type":59,"value":237},". Do not build from the private\nCosmos-Embed1 source tree for normal skill use; build from source only when\ndeveloping the container itself.",{"type":53,"tag":239,"props":240,"children":245},"pre",{"className":241,"code":242,"language":243,"meta":244,"style":244},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","TAO_SKILL_BANK_PATH=\"${TAO_SKILL_BANK_PATH:-$PWD}\"\nCOSMOS_EMBED_IMAGE=\"${COSMOS_EMBED_IMAGE:-$(\n  python \"$TAO_SKILL_BANK_PATH\u002Fscripts\u002Fresolve_tao_image.py\" \\\n    --skill-bank \"$TAO_SKILL_BANK_PATH\" \\\n    --model tao-finetune-cosmos-embed \\\n    --action train \\\n    --format json |\n  python -c 'import json,sys; print(json.load(sys.stdin)[\"image\"])'\n)}\"\ndocker pull \"$COSMOS_EMBED_IMAGE\"\n","bash","",[246],{"type":53,"tag":68,"props":247,"children":248},{"__ignoreMap":244},[249,291,322,356,381,399,417,436,464,477],{"type":53,"tag":250,"props":251,"children":254},"span",{"class":252,"line":253},"line",1,[255,261,267,272,276,281,286],{"type":53,"tag":250,"props":256,"children":258},{"style":257},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[259],{"type":59,"value":260},"TAO_SKILL_BANK_PATH",{"type":53,"tag":250,"props":262,"children":264},{"style":263},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[265],{"type":59,"value":266},"=",{"type":53,"tag":250,"props":268,"children":269},{"style":263},[270],{"type":59,"value":271},"\"${",{"type":53,"tag":250,"props":273,"children":274},{"style":257},[275],{"type":59,"value":260},{"type":53,"tag":250,"props":277,"children":278},{"style":263},[279],{"type":59,"value":280},":-",{"type":53,"tag":250,"props":282,"children":283},{"style":257},[284],{"type":59,"value":285},"$PWD",{"type":53,"tag":250,"props":287,"children":288},{"style":263},[289],{"type":59,"value":290},"}\"\n",{"type":53,"tag":250,"props":292,"children":294},{"class":252,"line":293},2,[295,300,304,308,312,316],{"type":53,"tag":250,"props":296,"children":297},{"style":257},[298],{"type":59,"value":299},"COSMOS_EMBED_IMAGE",{"type":53,"tag":250,"props":301,"children":302},{"style":263},[303],{"type":59,"value":266},{"type":53,"tag":250,"props":305,"children":306},{"style":263},[307],{"type":59,"value":271},{"type":53,"tag":250,"props":309,"children":310},{"style":257},[311],{"type":59,"value":299},{"type":53,"tag":250,"props":313,"children":314},{"style":263},[315],{"type":59,"value":280},{"type":53,"tag":250,"props":317,"children":319},{"style":318},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[320],{"type":59,"value":321},"$(\n",{"type":53,"tag":250,"props":323,"children":325},{"class":252,"line":324},3,[326,331,336,341,346,351],{"type":53,"tag":250,"props":327,"children":328},{"style":257},[329],{"type":59,"value":330},"  python",{"type":53,"tag":250,"props":332,"children":333},{"style":263},[334],{"type":59,"value":335}," \"",{"type":53,"tag":250,"props":337,"children":338},{"style":257},[339],{"type":59,"value":340},"$TAO_SKILL_BANK_PATH",{"type":53,"tag":250,"props":342,"children":343},{"style":318},[344],{"type":59,"value":345},"\u002Fscripts\u002Fresolve_tao_image.py",{"type":53,"tag":250,"props":347,"children":348},{"style":263},[349],{"type":59,"value":350},"\"",{"type":53,"tag":250,"props":352,"children":353},{"style":318},[354],{"type":59,"value":355}," \\\n",{"type":53,"tag":250,"props":357,"children":359},{"class":252,"line":358},4,[360,365,369,373,377],{"type":53,"tag":250,"props":361,"children":362},{"style":257},[363],{"type":59,"value":364},"    --skill-bank",{"type":53,"tag":250,"props":366,"children":367},{"style":263},[368],{"type":59,"value":335},{"type":53,"tag":250,"props":370,"children":371},{"style":257},[372],{"type":59,"value":340},{"type":53,"tag":250,"props":374,"children":375},{"style":263},[376],{"type":59,"value":350},{"type":53,"tag":250,"props":378,"children":379},{"style":318},[380],{"type":59,"value":355},{"type":53,"tag":250,"props":382,"children":384},{"class":252,"line":383},5,[385,390,395],{"type":53,"tag":250,"props":386,"children":387},{"style":257},[388],{"type":59,"value":389},"    --model",{"type":53,"tag":250,"props":391,"children":392},{"style":257},[393],{"type":59,"value":394}," tao-finetune-cosmos-embed",{"type":53,"tag":250,"props":396,"children":397},{"style":318},[398],{"type":59,"value":355},{"type":53,"tag":250,"props":400,"children":402},{"class":252,"line":401},6,[403,408,413],{"type":53,"tag":250,"props":404,"children":405},{"style":257},[406],{"type":59,"value":407},"    --action",{"type":53,"tag":250,"props":409,"children":410},{"style":257},[411],{"type":59,"value":412}," train",{"type":53,"tag":250,"props":414,"children":415},{"style":318},[416],{"type":59,"value":355},{"type":53,"tag":250,"props":418,"children":420},{"class":252,"line":419},7,[421,426,431],{"type":53,"tag":250,"props":422,"children":423},{"style":257},[424],{"type":59,"value":425},"    --format",{"type":53,"tag":250,"props":427,"children":428},{"style":257},[429],{"type":59,"value":430}," json",{"type":53,"tag":250,"props":432,"children":433},{"style":318},[434],{"type":59,"value":435}," |\n",{"type":53,"tag":250,"props":437,"children":439},{"class":252,"line":438},8,[440,444,449,454,459],{"type":53,"tag":250,"props":441,"children":442},{"style":257},[443],{"type":59,"value":330},{"type":53,"tag":250,"props":445,"children":446},{"style":257},[447],{"type":59,"value":448}," -c",{"type":53,"tag":250,"props":450,"children":451},{"style":263},[452],{"type":59,"value":453}," '",{"type":53,"tag":250,"props":455,"children":456},{"style":318},[457],{"type":59,"value":458},"import json,sys; print(json.load(sys.stdin)[\"image\"])",{"type":53,"tag":250,"props":460,"children":461},{"style":263},[462],{"type":59,"value":463},"'\n",{"type":53,"tag":250,"props":465,"children":467},{"class":252,"line":466},9,[468,473],{"type":53,"tag":250,"props":469,"children":470},{"style":318},[471],{"type":59,"value":472},")",{"type":53,"tag":250,"props":474,"children":475},{"style":263},[476],{"type":59,"value":290},{"type":53,"tag":250,"props":478,"children":480},{"class":252,"line":479},10,[481,487,492,496,501],{"type":53,"tag":250,"props":482,"children":484},{"style":483},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[485],{"type":59,"value":486},"docker",{"type":53,"tag":250,"props":488,"children":489},{"style":318},[490],{"type":59,"value":491}," pull",{"type":53,"tag":250,"props":493,"children":494},{"style":263},[495],{"type":59,"value":335},{"type":53,"tag":250,"props":497,"children":498},{"style":257},[499],{"type":59,"value":500},"$COSMOS_EMBED_IMAGE",{"type":53,"tag":250,"props":502,"children":503},{"style":263},[504],{"type":59,"value":505},"\"\n",{"type":53,"tag":62,"props":507,"children":508},{},[509],{"type":59,"value":510},"Expected local workspace layout:",{"type":53,"tag":239,"props":512,"children":516},{"className":513,"code":515,"language":59,"meta":244},[514],"language-text","workspace\u002F\n├── data\u002F\n│   ├── msrvtt_test_1k.json\n│   └── video\u002F\n│       ├── video7020.mp4\n│       └── ...\n├── model\u002F\n│   └── Cosmos-Embed1-224p\u002F        # optional if using HF repo id\n├── specs\u002F\n│   ├── train.yaml\n│   ├── evaluate.yaml\n│   ├── inference.yaml\n│   ├── export_onnx.yaml\n│   └── export_hf.yaml\n└── results\u002F\n",[517],{"type":53,"tag":68,"props":518,"children":519},{"__ignoreMap":244},[520],{"type":59,"value":515},{"type":53,"tag":62,"props":522,"children":523},{},[524],{"type":59,"value":525},"Use these Docker options for all actions unless the local Docker\u002Fplatform skill gives a stricter environment-specific command:",{"type":53,"tag":239,"props":527,"children":529},{"className":241,"code":528,"language":243,"meta":244,"style":244},"TAO_SKILL_BANK_PATH=\"${TAO_SKILL_BANK_PATH:-$PWD}\"\nCOSMOS_EMBED_IMAGE=\"${COSMOS_EMBED_IMAGE:-$(\n  python \"$TAO_SKILL_BANK_PATH\u002Fscripts\u002Fresolve_tao_image.py\" \\\n    --skill-bank \"$TAO_SKILL_BANK_PATH\" \\\n    --model tao-finetune-cosmos-embed \\\n    --action train \\\n    --format json |\n  python -c 'import json,sys; print(json.load(sys.stdin)[\"image\"])'\n)}\"\nRUN_ROOT=\"${RUN_ROOT:-$PWD}\"\nDOCKER_COMMON=(\n  --rm --gpus all --ipc=host --network=host\n  --shm-size=64g\n  --ulimit memlock=-1\n  --ulimit stack=67108864\n  -e HF_TOKEN\n  -e WANDB_DISABLED=true\n  -e WANDB_MODE=disabled\n  -e HUGGINGFACE_HUB_CACHE=\u002Fhf_cache\n  -v \"$RUN_ROOT\u002Fdata:\u002Fdata:ro\"\n  -v \"$RUN_ROOT\u002Fmodel:\u002Fmodel\"\n  -v \"$RUN_ROOT\u002Fspecs:\u002Fspecs:ro\"\n  -v \"$RUN_ROOT\u002Fresults:\u002Fresults\"\n  -v \"$RUN_ROOT\u002Fhf_cache:\u002Fhf_cache\"\n)\n",[530],{"type":53,"tag":68,"props":531,"children":532},{"__ignoreMap":244},[533,564,591,618,641,656,671,686,709,720,752,766,814,832,856,878,892,910,932,954,981,1006,1031,1056,1081],{"type":53,"tag":250,"props":534,"children":535},{"class":252,"line":253},[536,540,544,548,552,556,560],{"type":53,"tag":250,"props":537,"children":538},{"style":257},[539],{"type":59,"value":260},{"type":53,"tag":250,"props":541,"children":542},{"style":263},[543],{"type":59,"value":266},{"type":53,"tag":250,"props":545,"children":546},{"style":263},[547],{"type":59,"value":271},{"type":53,"tag":250,"props":549,"children":550},{"style":257},[551],{"type":59,"value":260},{"type":53,"tag":250,"props":553,"children":554},{"style":263},[555],{"type":59,"value":280},{"type":53,"tag":250,"props":557,"children":558},{"style":257},[559],{"type":59,"value":285},{"type":53,"tag":250,"props":561,"children":562},{"style":263},[563],{"type":59,"value":290},{"type":53,"tag":250,"props":565,"children":566},{"class":252,"line":293},[567,571,575,579,583,587],{"type":53,"tag":250,"props":568,"children":569},{"style":257},[570],{"type":59,"value":299},{"type":53,"tag":250,"props":572,"children":573},{"style":263},[574],{"type":59,"value":266},{"type":53,"tag":250,"props":576,"children":577},{"style":263},[578],{"type":59,"value":271},{"type":53,"tag":250,"props":580,"children":581},{"style":257},[582],{"type":59,"value":299},{"type":53,"tag":250,"props":584,"children":585},{"style":263},[586],{"type":59,"value":280},{"type":53,"tag":250,"props":588,"children":589},{"style":318},[590],{"type":59,"value":321},{"type":53,"tag":250,"props":592,"children":593},{"class":252,"line":324},[594,598,602,606,610,614],{"type":53,"tag":250,"props":595,"children":596},{"style":257},[597],{"type":59,"value":330},{"type":53,"tag":250,"props":599,"children":600},{"style":263},[601],{"type":59,"value":335},{"type":53,"tag":250,"props":603,"children":604},{"style":257},[605],{"type":59,"value":340},{"type":53,"tag":250,"props":607,"children":608},{"style":318},[609],{"type":59,"value":345},{"type":53,"tag":250,"props":611,"children":612},{"style":263},[613],{"type":59,"value":350},{"type":53,"tag":250,"props":615,"children":616},{"style":318},[617],{"type":59,"value":355},{"type":53,"tag":250,"props":619,"children":620},{"class":252,"line":358},[621,625,629,633,637],{"type":53,"tag":250,"props":622,"children":623},{"style":257},[624],{"type":59,"value":364},{"type":53,"tag":250,"props":626,"children":627},{"style":263},[628],{"type":59,"value":335},{"type":53,"tag":250,"props":630,"children":631},{"style":257},[632],{"type":59,"value":340},{"type":53,"tag":250,"props":634,"children":635},{"style":263},[636],{"type":59,"value":350},{"type":53,"tag":250,"props":638,"children":639},{"style":318},[640],{"type":59,"value":355},{"type":53,"tag":250,"props":642,"children":643},{"class":252,"line":383},[644,648,652],{"type":53,"tag":250,"props":645,"children":646},{"style":257},[647],{"type":59,"value":389},{"type":53,"tag":250,"props":649,"children":650},{"style":257},[651],{"type":59,"value":394},{"type":53,"tag":250,"props":653,"children":654},{"style":318},[655],{"type":59,"value":355},{"type":53,"tag":250,"props":657,"children":658},{"class":252,"line":401},[659,663,667],{"type":53,"tag":250,"props":660,"children":661},{"style":257},[662],{"type":59,"value":407},{"type":53,"tag":250,"props":664,"children":665},{"style":257},[666],{"type":59,"value":412},{"type":53,"tag":250,"props":668,"children":669},{"style":318},[670],{"type":59,"value":355},{"type":53,"tag":250,"props":672,"children":673},{"class":252,"line":419},[674,678,682],{"type":53,"tag":250,"props":675,"children":676},{"style":257},[677],{"type":59,"value":425},{"type":53,"tag":250,"props":679,"children":680},{"style":257},[681],{"type":59,"value":430},{"type":53,"tag":250,"props":683,"children":684},{"style":318},[685],{"type":59,"value":435},{"type":53,"tag":250,"props":687,"children":688},{"class":252,"line":438},[689,693,697,701,705],{"type":53,"tag":250,"props":690,"children":691},{"style":257},[692],{"type":59,"value":330},{"type":53,"tag":250,"props":694,"children":695},{"style":257},[696],{"type":59,"value":448},{"type":53,"tag":250,"props":698,"children":699},{"style":263},[700],{"type":59,"value":453},{"type":53,"tag":250,"props":702,"children":703},{"style":318},[704],{"type":59,"value":458},{"type":53,"tag":250,"props":706,"children":707},{"style":263},[708],{"type":59,"value":463},{"type":53,"tag":250,"props":710,"children":711},{"class":252,"line":466},[712,716],{"type":53,"tag":250,"props":713,"children":714},{"style":318},[715],{"type":59,"value":472},{"type":53,"tag":250,"props":717,"children":718},{"style":263},[719],{"type":59,"value":290},{"type":53,"tag":250,"props":721,"children":722},{"class":252,"line":479},[723,728,732,736,740,744,748],{"type":53,"tag":250,"props":724,"children":725},{"style":257},[726],{"type":59,"value":727},"RUN_ROOT",{"type":53,"tag":250,"props":729,"children":730},{"style":263},[731],{"type":59,"value":266},{"type":53,"tag":250,"props":733,"children":734},{"style":263},[735],{"type":59,"value":271},{"type":53,"tag":250,"props":737,"children":738},{"style":257},[739],{"type":59,"value":727},{"type":53,"tag":250,"props":741,"children":742},{"style":263},[743],{"type":59,"value":280},{"type":53,"tag":250,"props":745,"children":746},{"style":257},[747],{"type":59,"value":285},{"type":53,"tag":250,"props":749,"children":750},{"style":263},[751],{"type":59,"value":290},{"type":53,"tag":250,"props":753,"children":755},{"class":252,"line":754},11,[756,761],{"type":53,"tag":250,"props":757,"children":758},{"style":257},[759],{"type":59,"value":760},"DOCKER_COMMON",{"type":53,"tag":250,"props":762,"children":763},{"style":263},[764],{"type":59,"value":765},"=(\n",{"type":53,"tag":250,"props":767,"children":769},{"class":252,"line":768},12,[770,775,780,785,791,795,800,805,809],{"type":53,"tag":250,"props":771,"children":772},{"style":318},[773],{"type":59,"value":774},"  --rm",{"type":53,"tag":250,"props":776,"children":777},{"style":318},[778],{"type":59,"value":779}," --gpus",{"type":53,"tag":250,"props":781,"children":782},{"style":318},[783],{"type":59,"value":784}," all",{"type":53,"tag":250,"props":786,"children":788},{"style":787},"--shiki-light:#9C3EDA;--shiki-default:#C792EA;--shiki-dark:#C792EA",[789],{"type":59,"value":790}," --ipc",{"type":53,"tag":250,"props":792,"children":793},{"style":263},[794],{"type":59,"value":266},{"type":53,"tag":250,"props":796,"children":797},{"style":318},[798],{"type":59,"value":799},"host",{"type":53,"tag":250,"props":801,"children":802},{"style":787},[803],{"type":59,"value":804}," --network",{"type":53,"tag":250,"props":806,"children":807},{"style":263},[808],{"type":59,"value":266},{"type":53,"tag":250,"props":810,"children":811},{"style":318},[812],{"type":59,"value":813},"host\n",{"type":53,"tag":250,"props":815,"children":817},{"class":252,"line":816},13,[818,823,827],{"type":53,"tag":250,"props":819,"children":820},{"style":787},[821],{"type":59,"value":822},"  --shm-size",{"type":53,"tag":250,"props":824,"children":825},{"style":263},[826],{"type":59,"value":266},{"type":53,"tag":250,"props":828,"children":829},{"style":318},[830],{"type":59,"value":831},"64g\n",{"type":53,"tag":250,"props":833,"children":835},{"class":252,"line":834},14,[836,841,846,850],{"type":53,"tag":250,"props":837,"children":838},{"style":318},[839],{"type":59,"value":840},"  --ulimit",{"type":53,"tag":250,"props":842,"children":843},{"style":787},[844],{"type":59,"value":845}," memlock",{"type":53,"tag":250,"props":847,"children":848},{"style":263},[849],{"type":59,"value":266},{"type":53,"tag":250,"props":851,"children":853},{"style":852},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[854],{"type":59,"value":855},"-1\n",{"type":53,"tag":250,"props":857,"children":859},{"class":252,"line":858},15,[860,864,869,873],{"type":53,"tag":250,"props":861,"children":862},{"style":318},[863],{"type":59,"value":840},{"type":53,"tag":250,"props":865,"children":866},{"style":787},[867],{"type":59,"value":868}," stack",{"type":53,"tag":250,"props":870,"children":871},{"style":263},[872],{"type":59,"value":266},{"type":53,"tag":250,"props":874,"children":875},{"style":852},[876],{"type":59,"value":877},"67108864\n",{"type":53,"tag":250,"props":879,"children":881},{"class":252,"line":880},16,[882,887],{"type":53,"tag":250,"props":883,"children":884},{"style":318},[885],{"type":59,"value":886},"  -e",{"type":53,"tag":250,"props":888,"children":889},{"style":318},[890],{"type":59,"value":891}," HF_TOKEN\n",{"type":53,"tag":250,"props":893,"children":895},{"class":252,"line":894},17,[896,900,905],{"type":53,"tag":250,"props":897,"children":898},{"style":318},[899],{"type":59,"value":886},{"type":53,"tag":250,"props":901,"children":902},{"style":787},[903],{"type":59,"value":904}," WANDB_DISABLED",{"type":53,"tag":250,"props":906,"children":907},{"style":263},[908],{"type":59,"value":909},"=true\n",{"type":53,"tag":250,"props":911,"children":913},{"class":252,"line":912},18,[914,918,923,927],{"type":53,"tag":250,"props":915,"children":916},{"style":318},[917],{"type":59,"value":886},{"type":53,"tag":250,"props":919,"children":920},{"style":787},[921],{"type":59,"value":922}," WANDB_MODE",{"type":53,"tag":250,"props":924,"children":925},{"style":263},[926],{"type":59,"value":266},{"type":53,"tag":250,"props":928,"children":929},{"style":318},[930],{"type":59,"value":931},"disabled\n",{"type":53,"tag":250,"props":933,"children":935},{"class":252,"line":934},19,[936,940,945,949],{"type":53,"tag":250,"props":937,"children":938},{"style":318},[939],{"type":59,"value":886},{"type":53,"tag":250,"props":941,"children":942},{"style":787},[943],{"type":59,"value":944}," HUGGINGFACE_HUB_CACHE",{"type":53,"tag":250,"props":946,"children":947},{"style":263},[948],{"type":59,"value":266},{"type":53,"tag":250,"props":950,"children":951},{"style":318},[952],{"type":59,"value":953},"\u002Fhf_cache\n",{"type":53,"tag":250,"props":955,"children":957},{"class":252,"line":956},20,[958,963,967,972,977],{"type":53,"tag":250,"props":959,"children":960},{"style":318},[961],{"type":59,"value":962},"  -v",{"type":53,"tag":250,"props":964,"children":965},{"style":263},[966],{"type":59,"value":335},{"type":53,"tag":250,"props":968,"children":969},{"style":257},[970],{"type":59,"value":971},"$RUN_ROOT",{"type":53,"tag":250,"props":973,"children":974},{"style":318},[975],{"type":59,"value":976},"\u002Fdata:\u002Fdata:ro",{"type":53,"tag":250,"props":978,"children":979},{"style":263},[980],{"type":59,"value":505},{"type":53,"tag":250,"props":982,"children":984},{"class":252,"line":983},21,[985,989,993,997,1002],{"type":53,"tag":250,"props":986,"children":987},{"style":318},[988],{"type":59,"value":962},{"type":53,"tag":250,"props":990,"children":991},{"style":263},[992],{"type":59,"value":335},{"type":53,"tag":250,"props":994,"children":995},{"style":257},[996],{"type":59,"value":971},{"type":53,"tag":250,"props":998,"children":999},{"style":318},[1000],{"type":59,"value":1001},"\u002Fmodel:\u002Fmodel",{"type":53,"tag":250,"props":1003,"children":1004},{"style":263},[1005],{"type":59,"value":505},{"type":53,"tag":250,"props":1007,"children":1009},{"class":252,"line":1008},22,[1010,1014,1018,1022,1027],{"type":53,"tag":250,"props":1011,"children":1012},{"style":318},[1013],{"type":59,"value":962},{"type":53,"tag":250,"props":1015,"children":1016},{"style":263},[1017],{"type":59,"value":335},{"type":53,"tag":250,"props":1019,"children":1020},{"style":257},[1021],{"type":59,"value":971},{"type":53,"tag":250,"props":1023,"children":1024},{"style":318},[1025],{"type":59,"value":1026},"\u002Fspecs:\u002Fspecs:ro",{"type":53,"tag":250,"props":1028,"children":1029},{"style":263},[1030],{"type":59,"value":505},{"type":53,"tag":250,"props":1032,"children":1034},{"class":252,"line":1033},23,[1035,1039,1043,1047,1052],{"type":53,"tag":250,"props":1036,"children":1037},{"style":318},[1038],{"type":59,"value":962},{"type":53,"tag":250,"props":1040,"children":1041},{"style":263},[1042],{"type":59,"value":335},{"type":53,"tag":250,"props":1044,"children":1045},{"style":257},[1046],{"type":59,"value":971},{"type":53,"tag":250,"props":1048,"children":1049},{"style":318},[1050],{"type":59,"value":1051},"\u002Fresults:\u002Fresults",{"type":53,"tag":250,"props":1053,"children":1054},{"style":263},[1055],{"type":59,"value":505},{"type":53,"tag":250,"props":1057,"children":1059},{"class":252,"line":1058},24,[1060,1064,1068,1072,1077],{"type":53,"tag":250,"props":1061,"children":1062},{"style":318},[1063],{"type":59,"value":962},{"type":53,"tag":250,"props":1065,"children":1066},{"style":263},[1067],{"type":59,"value":335},{"type":53,"tag":250,"props":1069,"children":1070},{"style":257},[1071],{"type":59,"value":971},{"type":53,"tag":250,"props":1073,"children":1074},{"style":318},[1075],{"type":59,"value":1076},"\u002Fhf_cache:\u002Fhf_cache",{"type":53,"tag":250,"props":1078,"children":1079},{"style":263},[1080],{"type":59,"value":505},{"type":53,"tag":250,"props":1082,"children":1084},{"class":252,"line":1083},25,[1085],{"type":53,"tag":250,"props":1086,"children":1087},{"style":263},[1088],{"type":59,"value":1089},")\n",{"type":53,"tag":62,"props":1091,"children":1092},{},[1093,1095,1101],{"type":59,"value":1094},"For Cosmos-Embed images that ship ",{"type":53,"tag":68,"props":1096,"children":1098},{"className":1097},[],[1099],{"type":59,"value":1100},"protobuf==7.x",{"type":59,"value":1102},", run a small startup\npreamble before every action:",{"type":53,"tag":239,"props":1104,"children":1106},{"className":241,"code":1105,"language":243,"meta":244,"style":244},"python -m pip install \"protobuf\u003C7\"\n",[1107],{"type":53,"tag":68,"props":1108,"children":1109},{"__ignoreMap":244},[1110],{"type":53,"tag":250,"props":1111,"children":1112},{"class":252,"line":253},[1113,1118,1123,1128,1133,1137,1142],{"type":53,"tag":250,"props":1114,"children":1115},{"style":483},[1116],{"type":59,"value":1117},"python",{"type":53,"tag":250,"props":1119,"children":1120},{"style":318},[1121],{"type":59,"value":1122}," -m",{"type":53,"tag":250,"props":1124,"children":1125},{"style":318},[1126],{"type":59,"value":1127}," pip",{"type":53,"tag":250,"props":1129,"children":1130},{"style":318},[1131],{"type":59,"value":1132}," install",{"type":53,"tag":250,"props":1134,"children":1135},{"style":263},[1136],{"type":59,"value":335},{"type":53,"tag":250,"props":1138,"children":1139},{"style":318},[1140],{"type":59,"value":1141},"protobuf\u003C7",{"type":53,"tag":250,"props":1143,"children":1144},{"style":263},[1145],{"type":59,"value":505},{"type":53,"tag":62,"props":1147,"children":1148},{},[1149,1151,1157,1159,1164,1166,1172,1174,1180,1182,1188,1190,1196],{"type":59,"value":1150},"The image contains ",{"type":53,"tag":68,"props":1152,"children":1154},{"className":1153},[],[1155],{"type":59,"value":1156},"wandb==0.21.0",{"type":59,"value":1158}," with ",{"type":53,"tag":68,"props":1160,"children":1162},{"className":1161},[],[1163],{"type":59,"value":1100},{"type":59,"value":1165},"; importing W&B fails before training\u002Fevaluation unless protobuf is pinned below 7. Use ",{"type":53,"tag":68,"props":1167,"children":1169},{"className":1168},[],[1170],{"type":59,"value":1171},"WANDB_DISABLED=true",{"type":59,"value":1173}," and ",{"type":53,"tag":68,"props":1175,"children":1177},{"className":1176},[],[1178],{"type":59,"value":1179},"WANDB_MODE=disabled",{"type":59,"value":1181}," for smoke or offline runs. Cosmos-Embed may still download the public ",{"type":53,"tag":68,"props":1183,"children":1185},{"className":1184},[],[1186],{"type":59,"value":1187},"google-bert\u002Fbert-base-uncased",{"type":59,"value":1189}," Q-Former component even when the model checkpoint is disabled, so pass ",{"type":53,"tag":68,"props":1191,"children":1193},{"className":1192},[],[1194],{"type":59,"value":1195},"HF_TOKEN",{"type":59,"value":1197}," as an environment variable or mount a persistent HuggingFace cache. Do not write the token into specs, logs, or reports.",{"type":53,"tag":62,"props":1199,"children":1200},{},[1201],{"type":59,"value":1202},"Train:",{"type":53,"tag":239,"props":1204,"children":1206},{"className":241,"code":1205,"language":243,"meta":244,"style":244},"docker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 train -e \u002Fspecs\u002Ftrain.yaml results_dir=\u002Fresults\"\n",[1207],{"type":53,"tag":68,"props":1208,"children":1209},{"__ignoreMap":244},[1210,1252],{"type":53,"tag":250,"props":1211,"children":1212},{"class":252,"line":253},[1213,1217,1222,1227,1231,1236,1240,1244,1248],{"type":53,"tag":250,"props":1214,"children":1215},{"style":483},[1216],{"type":59,"value":486},{"type":53,"tag":250,"props":1218,"children":1219},{"style":318},[1220],{"type":59,"value":1221}," run",{"type":53,"tag":250,"props":1223,"children":1224},{"style":263},[1225],{"type":59,"value":1226}," \"${",{"type":53,"tag":250,"props":1228,"children":1229},{"style":257},[1230],{"type":59,"value":760},{"type":53,"tag":250,"props":1232,"children":1233},{"style":263},[1234],{"type":59,"value":1235},"[@]}\"",{"type":53,"tag":250,"props":1237,"children":1238},{"style":263},[1239],{"type":59,"value":335},{"type":53,"tag":250,"props":1241,"children":1242},{"style":257},[1243],{"type":59,"value":500},{"type":53,"tag":250,"props":1245,"children":1246},{"style":263},[1247],{"type":59,"value":350},{"type":53,"tag":250,"props":1249,"children":1250},{"style":257},[1251],{"type":59,"value":355},{"type":53,"tag":250,"props":1253,"children":1254},{"class":252,"line":293},[1255,1260,1265,1269,1274],{"type":53,"tag":250,"props":1256,"children":1257},{"style":318},[1258],{"type":59,"value":1259},"  bash",{"type":53,"tag":250,"props":1261,"children":1262},{"style":318},[1263],{"type":59,"value":1264}," -lc",{"type":53,"tag":250,"props":1266,"children":1267},{"style":263},[1268],{"type":59,"value":335},{"type":53,"tag":250,"props":1270,"children":1271},{"style":318},[1272],{"type":59,"value":1273},"python -m pip install 'protobuf\u003C7' && cosmos-embed1 train -e \u002Fspecs\u002Ftrain.yaml results_dir=\u002Fresults",{"type":53,"tag":250,"props":1275,"children":1276},{"style":263},[1277],{"type":59,"value":505},{"type":53,"tag":62,"props":1279,"children":1280},{},[1281],{"type":59,"value":1282},"Evaluate:",{"type":53,"tag":239,"props":1284,"children":1286},{"className":241,"code":1285,"language":243,"meta":244,"style":244},"docker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 evaluate -e \u002Fspecs\u002Fevaluate.yaml results_dir=\u002Fresults\"\n",[1287],{"type":53,"tag":68,"props":1288,"children":1289},{"__ignoreMap":244},[1290,1329],{"type":53,"tag":250,"props":1291,"children":1292},{"class":252,"line":253},[1293,1297,1301,1305,1309,1313,1317,1321,1325],{"type":53,"tag":250,"props":1294,"children":1295},{"style":483},[1296],{"type":59,"value":486},{"type":53,"tag":250,"props":1298,"children":1299},{"style":318},[1300],{"type":59,"value":1221},{"type":53,"tag":250,"props":1302,"children":1303},{"style":263},[1304],{"type":59,"value":1226},{"type":53,"tag":250,"props":1306,"children":1307},{"style":257},[1308],{"type":59,"value":760},{"type":53,"tag":250,"props":1310,"children":1311},{"style":263},[1312],{"type":59,"value":1235},{"type":53,"tag":250,"props":1314,"children":1315},{"style":263},[1316],{"type":59,"value":335},{"type":53,"tag":250,"props":1318,"children":1319},{"style":257},[1320],{"type":59,"value":500},{"type":53,"tag":250,"props":1322,"children":1323},{"style":263},[1324],{"type":59,"value":350},{"type":53,"tag":250,"props":1326,"children":1327},{"style":257},[1328],{"type":59,"value":355},{"type":53,"tag":250,"props":1330,"children":1331},{"class":252,"line":293},[1332,1336,1340,1344,1349],{"type":53,"tag":250,"props":1333,"children":1334},{"style":318},[1335],{"type":59,"value":1259},{"type":53,"tag":250,"props":1337,"children":1338},{"style":318},[1339],{"type":59,"value":1264},{"type":53,"tag":250,"props":1341,"children":1342},{"style":263},[1343],{"type":59,"value":335},{"type":53,"tag":250,"props":1345,"children":1346},{"style":318},[1347],{"type":59,"value":1348},"python -m pip install 'protobuf\u003C7' && cosmos-embed1 evaluate -e \u002Fspecs\u002Fevaluate.yaml results_dir=\u002Fresults",{"type":53,"tag":250,"props":1350,"children":1351},{"style":263},[1352],{"type":59,"value":505},{"type":53,"tag":62,"props":1354,"children":1355},{},[1356],{"type":59,"value":1357},"Inference:",{"type":53,"tag":239,"props":1359,"children":1361},{"className":241,"code":1360,"language":243,"meta":244,"style":244},"docker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 inference -e \u002Fspecs\u002Finference.yaml \\\n  'inference.query.input_texts=[\\\"a man is singing on stage\\\"]' \\\n  inference.k=5 \\\n  results_dir=\u002Fresults\"\n",[1362],{"type":53,"tag":68,"props":1363,"children":1364},{"__ignoreMap":244},[1365,1404,1429,1460,1472],{"type":53,"tag":250,"props":1366,"children":1367},{"class":252,"line":253},[1368,1372,1376,1380,1384,1388,1392,1396,1400],{"type":53,"tag":250,"props":1369,"children":1370},{"style":483},[1371],{"type":59,"value":486},{"type":53,"tag":250,"props":1373,"children":1374},{"style":318},[1375],{"type":59,"value":1221},{"type":53,"tag":250,"props":1377,"children":1378},{"style":263},[1379],{"type":59,"value":1226},{"type":53,"tag":250,"props":1381,"children":1382},{"style":257},[1383],{"type":59,"value":760},{"type":53,"tag":250,"props":1385,"children":1386},{"style":263},[1387],{"type":59,"value":1235},{"type":53,"tag":250,"props":1389,"children":1390},{"style":263},[1391],{"type":59,"value":335},{"type":53,"tag":250,"props":1393,"children":1394},{"style":257},[1395],{"type":59,"value":500},{"type":53,"tag":250,"props":1397,"children":1398},{"style":263},[1399],{"type":59,"value":350},{"type":53,"tag":250,"props":1401,"children":1402},{"style":257},[1403],{"type":59,"value":355},{"type":53,"tag":250,"props":1405,"children":1406},{"class":252,"line":293},[1407,1411,1415,1419,1424],{"type":53,"tag":250,"props":1408,"children":1409},{"style":318},[1410],{"type":59,"value":1259},{"type":53,"tag":250,"props":1412,"children":1413},{"style":318},[1414],{"type":59,"value":1264},{"type":53,"tag":250,"props":1416,"children":1417},{"style":263},[1418],{"type":59,"value":335},{"type":53,"tag":250,"props":1420,"children":1421},{"style":318},[1422],{"type":59,"value":1423},"python -m pip install 'protobuf\u003C7' && cosmos-embed1 inference -e \u002Fspecs\u002Finference.yaml ",{"type":53,"tag":250,"props":1425,"children":1426},{"style":257},[1427],{"type":59,"value":1428},"\\\n",{"type":53,"tag":250,"props":1430,"children":1431},{"class":252,"line":324},[1432,1437,1442,1447,1451,1456],{"type":53,"tag":250,"props":1433,"children":1434},{"style":318},[1435],{"type":59,"value":1436},"  'inference.query.input_texts=[",{"type":53,"tag":250,"props":1438,"children":1439},{"style":257},[1440],{"type":59,"value":1441},"\\\"",{"type":53,"tag":250,"props":1443,"children":1444},{"style":318},[1445],{"type":59,"value":1446},"a man is singing on stage",{"type":53,"tag":250,"props":1448,"children":1449},{"style":257},[1450],{"type":59,"value":1441},{"type":53,"tag":250,"props":1452,"children":1453},{"style":318},[1454],{"type":59,"value":1455},"]' ",{"type":53,"tag":250,"props":1457,"children":1458},{"style":257},[1459],{"type":59,"value":1428},{"type":53,"tag":250,"props":1461,"children":1462},{"class":252,"line":358},[1463,1468],{"type":53,"tag":250,"props":1464,"children":1465},{"style":318},[1466],{"type":59,"value":1467},"  inference.k=5 ",{"type":53,"tag":250,"props":1469,"children":1470},{"style":257},[1471],{"type":59,"value":1428},{"type":53,"tag":250,"props":1473,"children":1474},{"class":252,"line":383},[1475,1480],{"type":53,"tag":250,"props":1476,"children":1477},{"style":318},[1478],{"type":59,"value":1479},"  results_dir=\u002Fresults",{"type":53,"tag":250,"props":1481,"children":1482},{"style":263},[1483],{"type":59,"value":505},{"type":53,"tag":62,"props":1485,"children":1486},{},[1487],{"type":59,"value":1488},"Export ONNX:",{"type":53,"tag":239,"props":1490,"children":1492},{"className":241,"code":1491,"language":243,"meta":244,"style":244},"docker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_onnx.yaml \\\n  export.checkpoint=\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_000000001.pt \\\n  export.onnx_file=\u002Fresults\u002Fexport\u002Fcosmos_embed1_combined.onnx \\\n  results_dir=\u002Fresults\"\n",[1493],{"type":53,"tag":68,"props":1494,"children":1495},{"__ignoreMap":244},[1496,1535,1559,1571,1583],{"type":53,"tag":250,"props":1497,"children":1498},{"class":252,"line":253},[1499,1503,1507,1511,1515,1519,1523,1527,1531],{"type":53,"tag":250,"props":1500,"children":1501},{"style":483},[1502],{"type":59,"value":486},{"type":53,"tag":250,"props":1504,"children":1505},{"style":318},[1506],{"type":59,"value":1221},{"type":53,"tag":250,"props":1508,"children":1509},{"style":263},[1510],{"type":59,"value":1226},{"type":53,"tag":250,"props":1512,"children":1513},{"style":257},[1514],{"type":59,"value":760},{"type":53,"tag":250,"props":1516,"children":1517},{"style":263},[1518],{"type":59,"value":1235},{"type":53,"tag":250,"props":1520,"children":1521},{"style":263},[1522],{"type":59,"value":335},{"type":53,"tag":250,"props":1524,"children":1525},{"style":257},[1526],{"type":59,"value":500},{"type":53,"tag":250,"props":1528,"children":1529},{"style":263},[1530],{"type":59,"value":350},{"type":53,"tag":250,"props":1532,"children":1533},{"style":257},[1534],{"type":59,"value":355},{"type":53,"tag":250,"props":1536,"children":1537},{"class":252,"line":293},[1538,1542,1546,1550,1555],{"type":53,"tag":250,"props":1539,"children":1540},{"style":318},[1541],{"type":59,"value":1259},{"type":53,"tag":250,"props":1543,"children":1544},{"style":318},[1545],{"type":59,"value":1264},{"type":53,"tag":250,"props":1547,"children":1548},{"style":263},[1549],{"type":59,"value":335},{"type":53,"tag":250,"props":1551,"children":1552},{"style":318},[1553],{"type":59,"value":1554},"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_onnx.yaml ",{"type":53,"tag":250,"props":1556,"children":1557},{"style":257},[1558],{"type":59,"value":1428},{"type":53,"tag":250,"props":1560,"children":1561},{"class":252,"line":324},[1562,1567],{"type":53,"tag":250,"props":1563,"children":1564},{"style":318},[1565],{"type":59,"value":1566},"  export.checkpoint=\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_000000001.pt ",{"type":53,"tag":250,"props":1568,"children":1569},{"style":257},[1570],{"type":59,"value":1428},{"type":53,"tag":250,"props":1572,"children":1573},{"class":252,"line":358},[1574,1579],{"type":53,"tag":250,"props":1575,"children":1576},{"style":318},[1577],{"type":59,"value":1578},"  export.onnx_file=\u002Fresults\u002Fexport\u002Fcosmos_embed1_combined.onnx ",{"type":53,"tag":250,"props":1580,"children":1581},{"style":257},[1582],{"type":59,"value":1428},{"type":53,"tag":250,"props":1584,"children":1585},{"class":252,"line":383},[1586,1590],{"type":53,"tag":250,"props":1587,"children":1588},{"style":318},[1589],{"type":59,"value":1479},{"type":53,"tag":250,"props":1591,"children":1592},{"style":263},[1593],{"type":59,"value":505},{"type":53,"tag":62,"props":1595,"children":1596},{},[1597],{"type":59,"value":1598},"Export HuggingFace format:",{"type":53,"tag":239,"props":1600,"children":1602},{"className":241,"code":1601,"language":243,"meta":244,"style":244},"docker run \"${DOCKER_COMMON[@]}\" \"$COSMOS_EMBED_IMAGE\" \\\n  bash -lc \"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_hf.yaml \\\n  export.checkpoint=\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_000000001.pt \\\n  export.hf_output_dir=\u002Fresults\u002Fexport_hf\u002Fcosmos_embed1_hf \\\n  results_dir=\u002Fresults\"\n",[1603],{"type":53,"tag":68,"props":1604,"children":1605},{"__ignoreMap":244},[1606,1645,1669,1680,1692],{"type":53,"tag":250,"props":1607,"children":1608},{"class":252,"line":253},[1609,1613,1617,1621,1625,1629,1633,1637,1641],{"type":53,"tag":250,"props":1610,"children":1611},{"style":483},[1612],{"type":59,"value":486},{"type":53,"tag":250,"props":1614,"children":1615},{"style":318},[1616],{"type":59,"value":1221},{"type":53,"tag":250,"props":1618,"children":1619},{"style":263},[1620],{"type":59,"value":1226},{"type":53,"tag":250,"props":1622,"children":1623},{"style":257},[1624],{"type":59,"value":760},{"type":53,"tag":250,"props":1626,"children":1627},{"style":263},[1628],{"type":59,"value":1235},{"type":53,"tag":250,"props":1630,"children":1631},{"style":263},[1632],{"type":59,"value":335},{"type":53,"tag":250,"props":1634,"children":1635},{"style":257},[1636],{"type":59,"value":500},{"type":53,"tag":250,"props":1638,"children":1639},{"style":263},[1640],{"type":59,"value":350},{"type":53,"tag":250,"props":1642,"children":1643},{"style":257},[1644],{"type":59,"value":355},{"type":53,"tag":250,"props":1646,"children":1647},{"class":252,"line":293},[1648,1652,1656,1660,1665],{"type":53,"tag":250,"props":1649,"children":1650},{"style":318},[1651],{"type":59,"value":1259},{"type":53,"tag":250,"props":1653,"children":1654},{"style":318},[1655],{"type":59,"value":1264},{"type":53,"tag":250,"props":1657,"children":1658},{"style":263},[1659],{"type":59,"value":335},{"type":53,"tag":250,"props":1661,"children":1662},{"style":318},[1663],{"type":59,"value":1664},"python -m pip install 'protobuf\u003C7' && cosmos-embed1 export -e \u002Fspecs\u002Fexport_hf.yaml ",{"type":53,"tag":250,"props":1666,"children":1667},{"style":257},[1668],{"type":59,"value":1428},{"type":53,"tag":250,"props":1670,"children":1671},{"class":252,"line":324},[1672,1676],{"type":53,"tag":250,"props":1673,"children":1674},{"style":318},[1675],{"type":59,"value":1566},{"type":53,"tag":250,"props":1677,"children":1678},{"style":257},[1679],{"type":59,"value":1428},{"type":53,"tag":250,"props":1681,"children":1682},{"class":252,"line":358},[1683,1688],{"type":53,"tag":250,"props":1684,"children":1685},{"style":318},[1686],{"type":59,"value":1687},"  export.hf_output_dir=\u002Fresults\u002Fexport_hf\u002Fcosmos_embed1_hf ",{"type":53,"tag":250,"props":1689,"children":1690},{"style":257},[1691],{"type":59,"value":1428},{"type":53,"tag":250,"props":1693,"children":1694},{"class":252,"line":383},[1695,1699],{"type":53,"tag":250,"props":1696,"children":1697},{"style":318},[1698],{"type":59,"value":1479},{"type":53,"tag":250,"props":1700,"children":1701},{"style":263},[1702],{"type":59,"value":505},{"type":53,"tag":128,"props":1704,"children":1706},{"id":1705},"smoke-overrides",[1707],{"type":59,"value":1708},"Smoke Overrides",{"type":53,"tag":62,"props":1710,"children":1711},{},[1712],{"type":59,"value":1713},"For a small functional check, keep the same specs and override the expensive knobs:",{"type":53,"tag":239,"props":1715,"children":1717},{"className":241,"code":1716,"language":243,"meta":244,"style":244},"train.max_iter=1\ntrain.validation_iter=2\ntrain.checkpoint_iter=1\ntrain.optim.optim=adamw\ntrain.optim.warmup_steps=0\ntrain.optim.lr_decay_iters=1\ndataset.train_dataset.batch_size=1\ndataset.val_dataset.batch_size=1\ndataset.train_dataset.workers=0\ndataset.val_dataset.workers=0\n",[1718],{"type":53,"tag":68,"props":1719,"children":1720},{"__ignoreMap":244},[1721,1734,1747,1759,1772,1785,1797,1809,1821,1833],{"type":53,"tag":250,"props":1722,"children":1723},{"class":252,"line":253},[1724,1729],{"type":53,"tag":250,"props":1725,"children":1726},{"style":483},[1727],{"type":59,"value":1728},"train.max_iter",{"type":53,"tag":250,"props":1730,"children":1731},{"style":318},[1732],{"type":59,"value":1733},"=1\n",{"type":53,"tag":250,"props":1735,"children":1736},{"class":252,"line":293},[1737,1742],{"type":53,"tag":250,"props":1738,"children":1739},{"style":483},[1740],{"type":59,"value":1741},"train.validation_iter",{"type":53,"tag":250,"props":1743,"children":1744},{"style":318},[1745],{"type":59,"value":1746},"=2\n",{"type":53,"tag":250,"props":1748,"children":1749},{"class":252,"line":324},[1750,1755],{"type":53,"tag":250,"props":1751,"children":1752},{"style":483},[1753],{"type":59,"value":1754},"train.checkpoint_iter",{"type":53,"tag":250,"props":1756,"children":1757},{"style":318},[1758],{"type":59,"value":1733},{"type":53,"tag":250,"props":1760,"children":1761},{"class":252,"line":358},[1762,1767],{"type":53,"tag":250,"props":1763,"children":1764},{"style":483},[1765],{"type":59,"value":1766},"train.optim.optim",{"type":53,"tag":250,"props":1768,"children":1769},{"style":318},[1770],{"type":59,"value":1771},"=adamw\n",{"type":53,"tag":250,"props":1773,"children":1774},{"class":252,"line":383},[1775,1780],{"type":53,"tag":250,"props":1776,"children":1777},{"style":483},[1778],{"type":59,"value":1779},"train.optim.warmup_steps",{"type":53,"tag":250,"props":1781,"children":1782},{"style":318},[1783],{"type":59,"value":1784},"=0\n",{"type":53,"tag":250,"props":1786,"children":1787},{"class":252,"line":401},[1788,1793],{"type":53,"tag":250,"props":1789,"children":1790},{"style":483},[1791],{"type":59,"value":1792},"train.optim.lr_decay_iters",{"type":53,"tag":250,"props":1794,"children":1795},{"style":318},[1796],{"type":59,"value":1733},{"type":53,"tag":250,"props":1798,"children":1799},{"class":252,"line":419},[1800,1805],{"type":53,"tag":250,"props":1801,"children":1802},{"style":483},[1803],{"type":59,"value":1804},"dataset.train_dataset.batch_size",{"type":53,"tag":250,"props":1806,"children":1807},{"style":318},[1808],{"type":59,"value":1733},{"type":53,"tag":250,"props":1810,"children":1811},{"class":252,"line":438},[1812,1817],{"type":53,"tag":250,"props":1813,"children":1814},{"style":483},[1815],{"type":59,"value":1816},"dataset.val_dataset.batch_size",{"type":53,"tag":250,"props":1818,"children":1819},{"style":318},[1820],{"type":59,"value":1733},{"type":53,"tag":250,"props":1822,"children":1823},{"class":252,"line":466},[1824,1829],{"type":53,"tag":250,"props":1825,"children":1826},{"style":483},[1827],{"type":59,"value":1828},"dataset.train_dataset.workers",{"type":53,"tag":250,"props":1830,"children":1831},{"style":318},[1832],{"type":59,"value":1784},{"type":53,"tag":250,"props":1834,"children":1835},{"class":252,"line":479},[1836,1841],{"type":53,"tag":250,"props":1837,"children":1838},{"style":483},[1839],{"type":59,"value":1840},"dataset.val_dataset.workers",{"type":53,"tag":250,"props":1842,"children":1843},{"style":318},[1844],{"type":59,"value":1784},{"type":53,"tag":62,"props":1846,"children":1847},{},[1848,1850,1855,1857,1862,1864,1870,1872,1878],{"type":59,"value":1849},"When shortening the cosine scheduler for smoke runs, keep\n",{"type":53,"tag":68,"props":1851,"children":1853},{"className":1852},[],[1854],{"type":59,"value":1792},{"type":59,"value":1856}," greater than ",{"type":53,"tag":68,"props":1858,"children":1860},{"className":1859},[],[1861],{"type":59,"value":1779},{"type":59,"value":1863},", or set\n",{"type":53,"tag":68,"props":1865,"children":1867},{"className":1866},[],[1868],{"type":59,"value":1869},"train.optim.warmup_steps=0",{"type":59,"value":1871}," as shown above. The scheduler divides by\n",{"type":53,"tag":68,"props":1873,"children":1875},{"className":1874},[],[1876],{"type":59,"value":1877},"lr_decay_iters - warmup_steps",{"type":59,"value":1879},", so equal values fail before the checkpoint is\nwritten.",{"type":53,"tag":62,"props":1881,"children":1882},{},[1883,1885,1891,1893,1898,1900,1905],{"type":59,"value":1884},"If no local Cosmos-Embed1 pretrained checkpoint is available, set ",{"type":53,"tag":68,"props":1886,"children":1888},{"className":1887},[],[1889],{"type":59,"value":1890},"model.pretrained_model_path=null",{"type":59,"value":1892}," for a plumbing-only smoke train. The model quality is meaningless in that mode, but the train\u002Fevaluate\u002Finference\u002Fexport action paths can still be exercised. In the current container, the Q-Former path can still fetch ",{"type":53,"tag":68,"props":1894,"children":1896},{"className":1895},[],[1897],{"type":59,"value":1187},{"type":59,"value":1899},"; provide ",{"type":53,"tag":68,"props":1901,"children":1903},{"className":1902},[],[1904],{"type":59,"value":1195},{"type":59,"value":1906}," or a mounted HuggingFace cache for fresh ephemeral containers.",{"type":53,"tag":62,"props":1908,"children":1909},{},[1910],{"type":59,"value":1911},"For evaluation and inference smoke tests on a tiny subset:",{"type":53,"tag":239,"props":1913,"children":1915},{"className":241,"code":1914,"language":243,"meta":244,"style":244},"evaluate.callbacks.embedding_visualization=false\nevaluate.callbacks.max_eval_samples=8\ndataset.test_dataset.batch_size=1\ndataset.test_dataset.workers=0\ninference.k=2\ndataset.inference_dataset.batch_size=1\ndataset.inference_dataset.workers=0\n",[1916],{"type":53,"tag":68,"props":1917,"children":1918},{"__ignoreMap":244},[1919,1937,1950,1962,1974,1986,1998],{"type":53,"tag":250,"props":1920,"children":1921},{"class":252,"line":253},[1922,1927,1931],{"type":53,"tag":250,"props":1923,"children":1924},{"style":483},[1925],{"type":59,"value":1926},"evaluate.callbacks.embedding_visualization",{"type":53,"tag":250,"props":1928,"children":1929},{"style":318},[1930],{"type":59,"value":266},{"type":53,"tag":250,"props":1932,"children":1934},{"style":1933},"--shiki-light:#6182B8;--shiki-default:#82AAFF;--shiki-dark:#82AAFF",[1935],{"type":59,"value":1936},"false\n",{"type":53,"tag":250,"props":1938,"children":1939},{"class":252,"line":293},[1940,1945],{"type":53,"tag":250,"props":1941,"children":1942},{"style":483},[1943],{"type":59,"value":1944},"evaluate.callbacks.max_eval_samples",{"type":53,"tag":250,"props":1946,"children":1947},{"style":318},[1948],{"type":59,"value":1949},"=8\n",{"type":53,"tag":250,"props":1951,"children":1952},{"class":252,"line":324},[1953,1958],{"type":53,"tag":250,"props":1954,"children":1955},{"style":483},[1956],{"type":59,"value":1957},"dataset.test_dataset.batch_size",{"type":53,"tag":250,"props":1959,"children":1960},{"style":318},[1961],{"type":59,"value":1733},{"type":53,"tag":250,"props":1963,"children":1964},{"class":252,"line":358},[1965,1970],{"type":53,"tag":250,"props":1966,"children":1967},{"style":483},[1968],{"type":59,"value":1969},"dataset.test_dataset.workers",{"type":53,"tag":250,"props":1971,"children":1972},{"style":318},[1973],{"type":59,"value":1784},{"type":53,"tag":250,"props":1975,"children":1976},{"class":252,"line":383},[1977,1982],{"type":53,"tag":250,"props":1978,"children":1979},{"style":483},[1980],{"type":59,"value":1981},"inference.k",{"type":53,"tag":250,"props":1983,"children":1984},{"style":318},[1985],{"type":59,"value":1746},{"type":53,"tag":250,"props":1987,"children":1988},{"class":252,"line":401},[1989,1994],{"type":53,"tag":250,"props":1990,"children":1991},{"style":483},[1992],{"type":59,"value":1993},"dataset.inference_dataset.batch_size",{"type":53,"tag":250,"props":1995,"children":1996},{"style":318},[1997],{"type":59,"value":1733},{"type":53,"tag":250,"props":1999,"children":2000},{"class":252,"line":419},[2001,2006],{"type":53,"tag":250,"props":2002,"children":2003},{"style":483},[2004],{"type":59,"value":2005},"dataset.inference_dataset.workers",{"type":53,"tag":250,"props":2007,"children":2008},{"style":318},[2009],{"type":59,"value":1784},{"type":53,"tag":128,"props":2011,"children":2013},{"id":2012},"data-format",[2014],{"type":59,"value":2015},"Data Format",{"type":53,"tag":62,"props":2017,"children":2018},{},[2019],{"type":59,"value":2020},"The MSR-VTT path expects a local video glob and a JSON metadata file:",{"type":53,"tag":239,"props":2022,"children":2026},{"className":2023,"code":2024,"language":2025,"meta":244,"style":244},"language-yaml shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","dataset:\n  train_dataset:\n    dataset_type: msrvtt\n    mp4_urls: \u002Fdata\u002Fvideo\u002F*.mp4\n    metadata: \u002Fdata\u002Fmsrvtt_test_1k.json\n","yaml",[2027],{"type":53,"tag":68,"props":2028,"children":2029},{"__ignoreMap":244},[2030,2044,2056,2074,2091],{"type":53,"tag":250,"props":2031,"children":2032},{"class":252,"line":253},[2033,2039],{"type":53,"tag":250,"props":2034,"children":2036},{"style":2035},"--shiki-light:#E53935;--shiki-default:#F07178;--shiki-dark:#F07178",[2037],{"type":59,"value":2038},"dataset",{"type":53,"tag":250,"props":2040,"children":2041},{"style":263},[2042],{"type":59,"value":2043},":\n",{"type":53,"tag":250,"props":2045,"children":2046},{"class":252,"line":293},[2047,2052],{"type":53,"tag":250,"props":2048,"children":2049},{"style":2035},[2050],{"type":59,"value":2051},"  train_dataset",{"type":53,"tag":250,"props":2053,"children":2054},{"style":263},[2055],{"type":59,"value":2043},{"type":53,"tag":250,"props":2057,"children":2058},{"class":252,"line":324},[2059,2064,2069],{"type":53,"tag":250,"props":2060,"children":2061},{"style":2035},[2062],{"type":59,"value":2063},"    dataset_type",{"type":53,"tag":250,"props":2065,"children":2066},{"style":263},[2067],{"type":59,"value":2068},":",{"type":53,"tag":250,"props":2070,"children":2071},{"style":318},[2072],{"type":59,"value":2073}," msrvtt\n",{"type":53,"tag":250,"props":2075,"children":2076},{"class":252,"line":358},[2077,2082,2086],{"type":53,"tag":250,"props":2078,"children":2079},{"style":2035},[2080],{"type":59,"value":2081},"    mp4_urls",{"type":53,"tag":250,"props":2083,"children":2084},{"style":263},[2085],{"type":59,"value":2068},{"type":53,"tag":250,"props":2087,"children":2088},{"style":318},[2089],{"type":59,"value":2090}," \u002Fdata\u002Fvideo\u002F*.mp4\n",{"type":53,"tag":250,"props":2092,"children":2093},{"class":252,"line":383},[2094,2099,2103],{"type":53,"tag":250,"props":2095,"children":2096},{"style":2035},[2097],{"type":59,"value":2098},"    metadata",{"type":53,"tag":250,"props":2100,"children":2101},{"style":263},[2102],{"type":59,"value":2068},{"type":53,"tag":250,"props":2104,"children":2105},{"style":318},[2106],{"type":59,"value":2107}," \u002Fdata\u002Fmsrvtt_test_1k.json\n",{"type":53,"tag":62,"props":2109,"children":2110},{},[2111,2113,2118,2119,2125],{"type":59,"value":2112},"List-format metadata rows must include at least ",{"type":53,"tag":68,"props":2114,"children":2116},{"className":2115},[],[2117],{"type":59,"value":22},{"type":59,"value":1173},{"type":53,"tag":68,"props":2120,"children":2122},{"className":2121},[],[2123],{"type":59,"value":2124},"caption",{"type":59,"value":2068},{"type":53,"tag":239,"props":2127,"children":2131},{"className":2128,"code":2129,"language":2130,"meta":244,"style":244},"language-json shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","{\"video_id\": \"video7020\", \"video\": \"video7020.mp4\", \"caption\": \"a woman creating a fondant baby and flower\"}\n","json",[2132],{"type":53,"tag":68,"props":2133,"children":2134},{"__ignoreMap":244},[2135],{"type":53,"tag":250,"props":2136,"children":2137},{"class":252,"line":253},[2138,2143,2147,2152,2156,2160,2164,2169,2173,2178,2182,2186,2190,2194,2198,2203,2207,2211,2215,2219,2223,2227,2231,2236,2240],{"type":53,"tag":250,"props":2139,"children":2140},{"style":263},[2141],{"type":59,"value":2142},"{",{"type":53,"tag":250,"props":2144,"children":2145},{"style":263},[2146],{"type":59,"value":350},{"type":53,"tag":250,"props":2148,"children":2149},{"style":787},[2150],{"type":59,"value":2151},"video_id",{"type":53,"tag":250,"props":2153,"children":2154},{"style":263},[2155],{"type":59,"value":350},{"type":53,"tag":250,"props":2157,"children":2158},{"style":263},[2159],{"type":59,"value":2068},{"type":53,"tag":250,"props":2161,"children":2162},{"style":263},[2163],{"type":59,"value":335},{"type":53,"tag":250,"props":2165,"children":2166},{"style":318},[2167],{"type":59,"value":2168},"video7020",{"type":53,"tag":250,"props":2170,"children":2171},{"style":263},[2172],{"type":59,"value":350},{"type":53,"tag":250,"props":2174,"children":2175},{"style":263},[2176],{"type":59,"value":2177},",",{"type":53,"tag":250,"props":2179,"children":2180},{"style":263},[2181],{"type":59,"value":335},{"type":53,"tag":250,"props":2183,"children":2184},{"style":787},[2185],{"type":59,"value":22},{"type":53,"tag":250,"props":2187,"children":2188},{"style":263},[2189],{"type":59,"value":350},{"type":53,"tag":250,"props":2191,"children":2192},{"style":263},[2193],{"type":59,"value":2068},{"type":53,"tag":250,"props":2195,"children":2196},{"style":263},[2197],{"type":59,"value":335},{"type":53,"tag":250,"props":2199,"children":2200},{"style":318},[2201],{"type":59,"value":2202},"video7020.mp4",{"type":53,"tag":250,"props":2204,"children":2205},{"style":263},[2206],{"type":59,"value":350},{"type":53,"tag":250,"props":2208,"children":2209},{"style":263},[2210],{"type":59,"value":2177},{"type":53,"tag":250,"props":2212,"children":2213},{"style":263},[2214],{"type":59,"value":335},{"type":53,"tag":250,"props":2216,"children":2217},{"style":787},[2218],{"type":59,"value":2124},{"type":53,"tag":250,"props":2220,"children":2221},{"style":263},[2222],{"type":59,"value":350},{"type":53,"tag":250,"props":2224,"children":2225},{"style":263},[2226],{"type":59,"value":2068},{"type":53,"tag":250,"props":2228,"children":2229},{"style":263},[2230],{"type":59,"value":335},{"type":53,"tag":250,"props":2232,"children":2233},{"style":318},[2234],{"type":59,"value":2235},"a woman creating a fondant baby and flower",{"type":53,"tag":250,"props":2237,"children":2238},{"style":263},[2239],{"type":59,"value":350},{"type":53,"tag":250,"props":2241,"children":2242},{"style":263},[2243],{"type":59,"value":2244},"}\n",{"type":53,"tag":62,"props":2246,"children":2247},{},[2248,2250,2256,2258,2264,2266,2271],{"type":59,"value":2249},"The dataset loader derives the video id from the local ",{"type":53,"tag":68,"props":2251,"children":2253},{"className":2252},[],[2254],{"type":59,"value":2255},".mp4",{"type":59,"value":2257}," filename and filters to videos present in the metadata. If a run finds zero videos, check that ",{"type":53,"tag":68,"props":2259,"children":2261},{"className":2260},[],[2262],{"type":59,"value":2263},"mp4_urls",{"type":59,"value":2265}," points to a container-local glob and that metadata ",{"type":53,"tag":68,"props":2267,"children":2269},{"className":2268},[],[2270],{"type":59,"value":22},{"type":59,"value":2272}," names match the filenames.",{"type":53,"tag":128,"props":2274,"children":2276},{"id":2275},"model-weights",[2277],{"type":59,"value":2278},"Model Weights",{"type":53,"tag":2280,"props":2281,"children":2282},"ul",{},[2283,2304,2324],{"type":53,"tag":2284,"props":2285,"children":2286},"li",{},[2287,2289,2295,2297,2303],{"type":59,"value":2288},"Local HF directory: mount it under ",{"type":53,"tag":68,"props":2290,"children":2292},{"className":2291},[],[2293],{"type":59,"value":2294},"\u002Fmodel",{"type":59,"value":2296}," and set ",{"type":53,"tag":68,"props":2298,"children":2300},{"className":2299},[],[2301],{"type":59,"value":2302},"model.pretrained_model_path=\u002Fmodel\u002FCosmos-Embed1-224p",{"type":59,"value":106},{"type":53,"tag":2284,"props":2305,"children":2306},{},[2307,2309,2315,2317,2322],{"type":59,"value":2308},"HuggingFace repo: set ",{"type":53,"tag":68,"props":2310,"children":2312},{"className":2311},[],[2313],{"type":59,"value":2314},"model.pretrained_model_path=nvidia\u002FCosmos-Embed1-224p",{"type":59,"value":2316}," and pass ",{"type":53,"tag":68,"props":2318,"children":2320},{"className":2319},[],[2321],{"type":59,"value":1195},{"type":59,"value":2323}," if access is gated.",{"type":53,"tag":2284,"props":2325,"children":2326},{},[2327,2329,2335],{"type":59,"value":2328},"Fine-tuned checkpoint: set downstream actions to the resolver-selected ",{"type":53,"tag":68,"props":2330,"children":2332},{"className":2331},[],[2333],{"type":59,"value":2334},"\u002Fresults\u002Ftrain\u002Fcheckpoints\u002Fiter_#########.pt",{"type":59,"value":2336}," file.",{"type":53,"tag":62,"props":2338,"children":2339},{},[2340,2342,2348,2350,2356,2358,2364,2366,2372,2373,2379,2380,2386,2387,2393,2395,2401],{"type":59,"value":2341},"Training writes full checkpoints under ",{"type":53,"tag":68,"props":2343,"children":2345},{"className":2344},[],[2346],{"type":59,"value":2347},"results\u002Ftrain\u002Fcheckpoints\u002Fiter_#########.pt",{"type":59,"value":2349},", updates ",{"type":53,"tag":68,"props":2351,"children":2353},{"className":2352},[],[2354],{"type":59,"value":2355},"results\u002Ftrain\u002Fcheckpoints\u002Flatest_checkpoint.txt",{"type":59,"value":2357},", and creates a ",{"type":53,"tag":68,"props":2359,"children":2361},{"className":2360},[],[2362],{"type":59,"value":2363},"cosmos_embed1_model_latest.pth",{"type":59,"value":2365}," symlink. For ",{"type":53,"tag":68,"props":2367,"children":2369},{"className":2368},[],[2370],{"type":59,"value":2371},"evaluate.checkpoint",{"type":59,"value":83},{"type":53,"tag":68,"props":2374,"children":2376},{"className":2375},[],[2377],{"type":59,"value":2378},"inference.checkpoint",{"type":59,"value":83},{"type":53,"tag":68,"props":2381,"children":2383},{"className":2382},[],[2384],{"type":59,"value":2385},"export.checkpoint",{"type":59,"value":98},{"type":53,"tag":68,"props":2388,"children":2390},{"className":2389},[],[2391],{"type":59,"value":2392},"train.resume_training_checkpoint_path",{"type":59,"value":2394},", resolve and pass the exact ",{"type":53,"tag":68,"props":2396,"children":2398},{"className":2397},[],[2399],{"type":59,"value":2400},"iter_#########.pt",{"type":59,"value":2402}," file for the intended iteration. The action spec templates intentionally leave these checkpoint fields null so the model-skill runner or the user must provide the resolver-selected checkpoint. Use the latest symlink only when the user explicitly asks for latest.",{"type":53,"tag":62,"props":2404,"children":2405},{},[2406,2408,2414],{"type":59,"value":2407},"For single-GPU resume\u002Fretrain from a consolidated checkpoint, set ",{"type":53,"tag":68,"props":2409,"children":2411},{"className":2410},[],[2412],{"type":59,"value":2413},"model.fsdp_shard_size: 1",{"type":59,"value":2415},". The container default is 8, which sends resumed training through an FSDP apply path that Cosmos-Embed1 does not implement for this model class.",{"type":53,"tag":62,"props":2417,"children":2418},{},[2419],{"type":59,"value":2420},"Variants:",{"type":53,"tag":2422,"props":2423,"children":2424},"table",{},[2425,2455],{"type":53,"tag":2426,"props":2427,"children":2428},"thead",{},[2429],{"type":53,"tag":2430,"props":2431,"children":2432},"tr",{},[2433,2439,2445,2450],{"type":53,"tag":2434,"props":2435,"children":2436},"th",{},[2437],{"type":59,"value":2438},"Variant",{"type":53,"tag":2434,"props":2440,"children":2442},{"align":2441},"right",[2443],{"type":59,"value":2444},"Resolution",{"type":53,"tag":2434,"props":2446,"children":2447},{"align":2441},[2448],{"type":59,"value":2449},"Frames",{"type":53,"tag":2434,"props":2451,"children":2452},{"align":2441},[2453],{"type":59,"value":2454},"Embedding dim",{"type":53,"tag":2456,"props":2457,"children":2458},"tbody",{},[2459,2487,2513],{"type":53,"tag":2430,"props":2460,"children":2461},{},[2462,2472,2477,2482],{"type":53,"tag":2463,"props":2464,"children":2465},"td",{},[2466],{"type":53,"tag":68,"props":2467,"children":2469},{"className":2468},[],[2470],{"type":59,"value":2471},"Cosmos-Embed1-224p",{"type":53,"tag":2463,"props":2473,"children":2474},{"align":2441},[2475],{"type":59,"value":2476},"224 x 224",{"type":53,"tag":2463,"props":2478,"children":2479},{"align":2441},[2480],{"type":59,"value":2481},"8",{"type":53,"tag":2463,"props":2483,"children":2484},{"align":2441},[2485],{"type":59,"value":2486},"256",{"type":53,"tag":2430,"props":2488,"children":2489},{},[2490,2499,2504,2508],{"type":53,"tag":2463,"props":2491,"children":2492},{},[2493],{"type":53,"tag":68,"props":2494,"children":2496},{"className":2495},[],[2497],{"type":59,"value":2498},"Cosmos-Embed1-336p",{"type":53,"tag":2463,"props":2500,"children":2501},{"align":2441},[2502],{"type":59,"value":2503},"336 x 336",{"type":53,"tag":2463,"props":2505,"children":2506},{"align":2441},[2507],{"type":59,"value":2481},{"type":53,"tag":2463,"props":2509,"children":2510},{"align":2441},[2511],{"type":59,"value":2512},"768",{"type":53,"tag":2430,"props":2514,"children":2515},{},[2516,2525,2530,2534],{"type":53,"tag":2463,"props":2517,"children":2518},{},[2519],{"type":53,"tag":68,"props":2520,"children":2522},{"className":2521},[],[2523],{"type":59,"value":2524},"Cosmos-Embed1-448p",{"type":53,"tag":2463,"props":2526,"children":2527},{"align":2441},[2528],{"type":59,"value":2529},"448 x 448",{"type":53,"tag":2463,"props":2531,"children":2532},{"align":2441},[2533],{"type":59,"value":2481},{"type":53,"tag":2463,"props":2535,"children":2536},{"align":2441},[2537],{"type":59,"value":2512},{"type":53,"tag":62,"props":2539,"children":2540},{},[2541,2543,2549,2550,2556,2557,2563],{"type":59,"value":2542},"Keep ",{"type":53,"tag":68,"props":2544,"children":2546},{"className":2545},[],[2547],{"type":59,"value":2548},"model.network.embed_dim",{"type":59,"value":83},{"type":53,"tag":68,"props":2551,"children":2553},{"className":2552},[],[2554],{"type":59,"value":2555},"model.input_hw",{"type":59,"value":98},{"type":53,"tag":68,"props":2558,"children":2560},{"className":2559},[],[2561],{"type":59,"value":2562},"model.network.spatial_resolution",{"type":59,"value":2564}," aligned with the selected variant.",{"type":53,"tag":128,"props":2566,"children":2568},{"id":2567},"important-parameters",[2569],{"type":59,"value":2570},"Important Parameters",{"type":53,"tag":2422,"props":2572,"children":2573},{},[2574,2590],{"type":53,"tag":2426,"props":2575,"children":2576},{},[2577],{"type":53,"tag":2430,"props":2578,"children":2579},{},[2580,2585],{"type":53,"tag":2434,"props":2581,"children":2582},{},[2583],{"type":59,"value":2584},"Parameter",{"type":53,"tag":2434,"props":2586,"children":2587},{},[2588],{"type":59,"value":2589},"Notes",{"type":53,"tag":2456,"props":2591,"children":2592},{},[2593,2639,2662,2692,2717,2763,2795,2841,2887,2912,2929,2946,2970,2993,3034],{"type":53,"tag":2430,"props":2594,"children":2595},{},[2596,2605],{"type":53,"tag":2463,"props":2597,"children":2598},{},[2599],{"type":53,"tag":68,"props":2600,"children":2602},{"className":2601},[],[2603],{"type":59,"value":2604},"train.num_gpus",{"type":53,"tag":2463,"props":2606,"children":2607},{},[2608,2614,2616,2622,2624,2630,2631,2637],{"type":53,"tag":68,"props":2609,"children":2611},{"className":2610},[],[2612],{"type":59,"value":2613},"1",{"type":59,"value":2615}," for single GPU, ",{"type":53,"tag":68,"props":2617,"children":2619},{"className":2618},[],[2620],{"type":59,"value":2621},">1",{"type":59,"value":2623}," auto-launches ",{"type":53,"tag":68,"props":2625,"children":2627},{"className":2626},[],[2628],{"type":59,"value":2629},"torchrun",{"type":59,"value":83},{"type":53,"tag":68,"props":2632,"children":2634},{"className":2633},[],[2635],{"type":59,"value":2636},"-1",{"type":59,"value":2638}," auto-detects visible GPUs.",{"type":53,"tag":2430,"props":2640,"children":2641},{},[2642,2650],{"type":53,"tag":2463,"props":2643,"children":2644},{},[2645],{"type":53,"tag":68,"props":2646,"children":2648},{"className":2647},[],[2649],{"type":59,"value":1728},{"type":53,"tag":2463,"props":2651,"children":2652},{},[2653,2655,2660],{"type":59,"value":2654},"Main training length. Use ",{"type":53,"tag":68,"props":2656,"children":2658},{"className":2657},[],[2659],{"type":59,"value":2613},{"type":59,"value":2661}," only for smoke testing.",{"type":53,"tag":2430,"props":2663,"children":2664},{},[2665,2673],{"type":53,"tag":2463,"props":2666,"children":2667},{},[2668],{"type":53,"tag":68,"props":2669,"children":2671},{"className":2670},[],[2672],{"type":59,"value":1766},{"type":53,"tag":2463,"props":2674,"children":2675},{},[2676,2682,2684,2690],{"type":53,"tag":68,"props":2677,"children":2679},{"className":2678},[],[2680],{"type":59,"value":2681},"fused_adamw",{"type":59,"value":2683}," is faster when available; ",{"type":53,"tag":68,"props":2685,"children":2687},{"className":2686},[],[2688],{"type":59,"value":2689},"adamw",{"type":59,"value":2691}," is safer for smoke and portability.",{"type":53,"tag":2430,"props":2693,"children":2694},{},[2695,2704],{"type":53,"tag":2463,"props":2696,"children":2697},{},[2698],{"type":53,"tag":68,"props":2699,"children":2701},{"className":2700},[],[2702],{"type":59,"value":2703},"model.lora.enabled",{"type":53,"tag":2463,"props":2705,"children":2706},{},[2707,2709,2715],{"type":59,"value":2708},"Enables LoRA. Set ",{"type":53,"tag":68,"props":2710,"children":2712},{"className":2711},[],[2713],{"type":59,"value":2714},"model.network.visual_encoder.transformer_engine=false",{"type":59,"value":2716}," when LoRA is on.",{"type":53,"tag":2430,"props":2718,"children":2719},{},[2720,2729],{"type":53,"tag":2463,"props":2721,"children":2722},{},[2723],{"type":53,"tag":68,"props":2724,"children":2726},{"className":2725},[],[2727],{"type":59,"value":2728},"model.lora.lora_rank",{"type":53,"tag":2463,"props":2730,"children":2731},{},[2732,2734,2739,2741,2747,2748,2753,2755,2761],{"type":59,"value":2733},"LoRA rank. Start with ",{"type":53,"tag":68,"props":2735,"children":2737},{"className":2736},[],[2738],{"type":59,"value":2481},{"type":59,"value":2740},"; try ",{"type":53,"tag":68,"props":2742,"children":2744},{"className":2743},[],[2745],{"type":59,"value":2746},"4",{"type":59,"value":83},{"type":53,"tag":68,"props":2749,"children":2751},{"className":2750},[],[2752],{"type":59,"value":2481},{"type":59,"value":2754},", or ",{"type":53,"tag":68,"props":2756,"children":2758},{"className":2757},[],[2759],{"type":59,"value":2760},"16",{"type":59,"value":2762}," for manual or AutoML-style sweeps.",{"type":53,"tag":2430,"props":2764,"children":2765},{},[2766,2775],{"type":53,"tag":2463,"props":2767,"children":2768},{},[2769],{"type":53,"tag":68,"props":2770,"children":2772},{"className":2771},[],[2773],{"type":59,"value":2774},"model.lora.lora_alpha",{"type":53,"tag":2463,"props":2776,"children":2777},{},[2778,2780,2785,2787,2793],{"type":59,"value":2779},"LoRA scaling factor. Start with ",{"type":53,"tag":68,"props":2781,"children":2783},{"className":2782},[],[2784],{"type":59,"value":2760},{"type":59,"value":2786},"; keep near ",{"type":53,"tag":68,"props":2788,"children":2790},{"className":2789},[],[2791],{"type":59,"value":2792},"2 * lora_rank",{"type":59,"value":2794}," unless experiments show otherwise.",{"type":53,"tag":2430,"props":2796,"children":2797},{},[2798,2807],{"type":53,"tag":2463,"props":2799,"children":2800},{},[2801],{"type":53,"tag":68,"props":2802,"children":2804},{"className":2803},[],[2805],{"type":59,"value":2806},"model.lora.lora_dropout",{"type":53,"tag":2463,"props":2808,"children":2809},{},[2810,2812,2818,2820,2826,2827,2833,2834,2839],{"type":59,"value":2811},"LoRA dropout. Start with ",{"type":53,"tag":68,"props":2813,"children":2815},{"className":2814},[],[2816],{"type":59,"value":2817},"0.1",{"type":59,"value":2819},"; sweep ",{"type":53,"tag":68,"props":2821,"children":2823},{"className":2822},[],[2824],{"type":59,"value":2825},"0.0",{"type":59,"value":83},{"type":53,"tag":68,"props":2828,"children":2830},{"className":2829},[],[2831],{"type":59,"value":2832},"0.05",{"type":59,"value":98},{"type":53,"tag":68,"props":2835,"children":2837},{"className":2836},[],[2838],{"type":59,"value":2817},{"type":59,"value":2840}," for small datasets.",{"type":53,"tag":2430,"props":2842,"children":2843},{},[2844,2853],{"type":53,"tag":2463,"props":2845,"children":2846},{},[2847],{"type":53,"tag":68,"props":2848,"children":2850},{"className":2849},[],[2851],{"type":59,"value":2852},"model.lora.bias",{"type":53,"tag":2463,"props":2854,"children":2855},{},[2856,2858,2864,2865,2871,2872,2878,2880,2885],{"type":59,"value":2857},"Bias policy: ",{"type":53,"tag":68,"props":2859,"children":2861},{"className":2860},[],[2862],{"type":59,"value":2863},"none",{"type":59,"value":83},{"type":53,"tag":68,"props":2866,"children":2868},{"className":2867},[],[2869],{"type":59,"value":2870},"all",{"type":59,"value":2754},{"type":53,"tag":68,"props":2873,"children":2875},{"className":2874},[],[2876],{"type":59,"value":2877},"lora_only",{"type":59,"value":2879},". Keep ",{"type":53,"tag":68,"props":2881,"children":2883},{"className":2882},[],[2884],{"type":59,"value":2863},{"type":59,"value":2886}," unless intentionally training biases.",{"type":53,"tag":2430,"props":2888,"children":2889},{},[2890,2907],{"type":53,"tag":2463,"props":2891,"children":2892},{},[2893,2899,2901],{"type":53,"tag":68,"props":2894,"children":2896},{"className":2895},[],[2897],{"type":59,"value":2898},"model.lora.use_rslora",{"type":59,"value":2900}," \u002F ",{"type":53,"tag":68,"props":2902,"children":2904},{"className":2903},[],[2905],{"type":59,"value":2906},"use_dora",{"type":53,"tag":2463,"props":2908,"children":2909},{},[2910],{"type":59,"value":2911},"Optional LoRA variants. Enable one at a time and record the setting with the checkpoint.",{"type":53,"tag":2430,"props":2913,"children":2914},{},[2915,2924],{"type":53,"tag":2463,"props":2916,"children":2917},{},[2918],{"type":53,"tag":68,"props":2919,"children":2921},{"className":2920},[],[2922],{"type":59,"value":2923},"model.lora.target_modules",{"type":53,"tag":2463,"props":2925,"children":2926},{},[2927],{"type":59,"value":2928},"Optional module-name patterns for LoRA injection. Leave empty for the default ViT + Q-Former attention\u002FMLP targets.",{"type":53,"tag":2430,"props":2930,"children":2931},{},[2932,2941],{"type":53,"tag":2463,"props":2933,"children":2934},{},[2935],{"type":53,"tag":68,"props":2936,"children":2938},{"className":2937},[],[2939],{"type":59,"value":2940},"model.lora.modules_to_save",{"type":53,"tag":2463,"props":2942,"children":2943},{},[2944],{"type":59,"value":2945},"Optional modules to keep fully trainable alongside LoRA. Leave empty unless preserving a task-specific head.",{"type":53,"tag":2430,"props":2947,"children":2948},{},[2949,2965],{"type":53,"tag":2463,"props":2950,"children":2951},{},[2952,2958,2959],{"type":53,"tag":68,"props":2953,"children":2955},{"className":2954},[],[2956],{"type":59,"value":2957},"evaluate.load_dataset_pkl",{"type":59,"value":2900},{"type":53,"tag":68,"props":2960,"children":2962},{"className":2961},[],[2963],{"type":59,"value":2964},"save_dataset_pkl",{"type":53,"tag":2463,"props":2966,"children":2967},{},[2968],{"type":59,"value":2969},"Cache evaluation embeddings.",{"type":53,"tag":2430,"props":2971,"children":2972},{},[2973,2988],{"type":53,"tag":2463,"props":2974,"children":2975},{},[2976,2982,2983],{"type":53,"tag":68,"props":2977,"children":2979},{"className":2978},[],[2980],{"type":59,"value":2981},"inference.load_dataset_pkl",{"type":59,"value":2900},{"type":53,"tag":68,"props":2984,"children":2986},{"className":2985},[],[2987],{"type":59,"value":2964},{"type":53,"tag":2463,"props":2989,"children":2990},{},[2991],{"type":59,"value":2992},"Cache the search database for repeated retrieval.",{"type":53,"tag":2430,"props":2994,"children":2995},{},[2996,3005],{"type":53,"tag":2463,"props":2997,"children":2998},{},[2999],{"type":53,"tag":68,"props":3000,"children":3002},{"className":3001},[],[3003],{"type":59,"value":3004},"export.mode",{"type":53,"tag":2463,"props":3006,"children":3007},{},[3008,3013,3014,3019,3020,3026,3027,3033],{"type":53,"tag":68,"props":3009,"children":3011},{"className":3010},[],[3012],{"type":59,"value":22},{"type":59,"value":83},{"type":53,"tag":68,"props":3015,"children":3017},{"className":3016},[],[3018],{"type":59,"value":59},{"type":59,"value":83},{"type":53,"tag":68,"props":3021,"children":3023},{"className":3022},[],[3024],{"type":59,"value":3025},"combined",{"type":59,"value":2754},{"type":53,"tag":68,"props":3028,"children":3030},{"className":3029},[],[3031],{"type":59,"value":3032},"huggingface",{"type":59,"value":106},{"type":53,"tag":2430,"props":3035,"children":3036},{},[3037,3046],{"type":53,"tag":2463,"props":3038,"children":3039},{},[3040],{"type":53,"tag":68,"props":3041,"children":3043},{"className":3042},[],[3044],{"type":59,"value":3045},"export.on_cpu",{"type":53,"tag":2463,"props":3047,"children":3048},{},[3049],{"type":59,"value":3050},"Recommended for export to avoid device mismatch issues.",{"type":53,"tag":3052,"props":3053,"children":3055},"h3",{"id":3054},"lora-and-automl-notes",[3056],{"type":59,"value":3057},"LoRA and AutoML Notes",{"type":53,"tag":62,"props":3059,"children":3060},{},[3061,3063,3069,3071,3076,3078,3084,3086,3092],{"type":59,"value":3062},"For parameter-efficient fine-tuning, set ",{"type":53,"tag":68,"props":3064,"children":3066},{"className":3065},[],[3067],{"type":59,"value":3068},"model.lora.enabled=true",{"type":59,"value":3070}," and keep\n",{"type":53,"tag":68,"props":3072,"children":3074},{"className":3073},[],[3075],{"type":59,"value":2714},{"type":59,"value":3077},"; TAO Core's\nCosmos-Embed1 config notes that PEFT cannot inject adapters into Transformer\nEngine layers. Treat the LoRA fields above as the first candidate parameters\nfor manual tuning or AutoML-style search before unfreezing larger model blocks.\nAvoid changing ",{"type":53,"tag":68,"props":3079,"children":3081},{"className":3080},[],[3082],{"type":59,"value":3083},"target_modules",{"type":59,"value":3085}," or ",{"type":53,"tag":68,"props":3087,"children":3089},{"className":3088},[],[3090],{"type":59,"value":3091},"modules_to_save",{"type":59,"value":3093}," unless the user explicitly\nneeds custom adapter placement.",{"type":53,"tag":128,"props":3095,"children":3097},{"id":3096},"s3-staging",[3098],{"type":59,"value":3099},"S3 Staging",{"type":53,"tag":62,"props":3101,"children":3102},{},[3103,3105,3111,3113,3119],{"type":59,"value":3104},"The Cosmos-Embed1 CLI consumes local paths and Python globs, not raw ",{"type":53,"tag":68,"props":3106,"children":3108},{"className":3107},[],[3109],{"type":59,"value":3110},"s3:\u002F\u002F...\u002F*.mp4",{"type":59,"value":3112}," URIs. For S3-backed runs, first stage a subset or full dataset to the execution host\u002Fcontainer filesystem, then use local paths such as ",{"type":53,"tag":68,"props":3114,"children":3116},{"className":3115},[],[3117],{"type":59,"value":3118},"\u002Fdata\u002Fvideo\u002F*.mp4",{"type":59,"value":3120}," in the spec.",{"type":53,"tag":62,"props":3122,"children":3123},{},[3124],{"type":59,"value":3125},"Recommended S3 layout for staged MSR-VTT data:",{"type":53,"tag":239,"props":3127,"children":3130},{"className":3128,"code":3129,"language":59,"meta":244},[514],"s3:\u002F\u002Fbucket\u002Fpath\u002Fcosmos-embed\u002Fmsrvtt-subset\u002F\n├── msrvtt_test_1k.json\n└── video\u002F\n    ├── video7020.mp4\n    └── ...\n",[3131],{"type":53,"tag":68,"props":3132,"children":3133},{"__ignoreMap":244},[3134],{"type":59,"value":3129},{"type":53,"tag":62,"props":3136,"children":3137},{},[3138,3140,3146],{"type":59,"value":3139},"After downloading\u002Fsyncing that prefix into the mounted ",{"type":53,"tag":68,"props":3141,"children":3143},{"className":3142},[],[3144],{"type":59,"value":3145},"data\u002F",{"type":59,"value":3147}," directory, use the same Docker commands above.",{"type":53,"tag":128,"props":3149,"children":3151},{"id":3150},"outputs",[3152],{"type":59,"value":3153},"Outputs",{"type":53,"tag":239,"props":3155,"children":3158},{"className":3156,"code":3157,"language":59,"meta":244},[514],"results\u002F\n├── train\u002F\n│   ├── cosmos_embed1_model_latest.pth\n│   ├── cosmos_embed1_model_\u003Citer>.pth\n│   └── experiment.yaml\n├── evaluate\u002F\n│   ├── metrics.json\n│   └── experiment.yaml\n├── inference\u002F\n│   ├── results.json\n│   └── experiment.yaml\n├── export\u002F\n│   ├── cosmos_embed1_combined.onnx\n│   └── export_config.yaml\n└── export_hf\u002F\n    └── cosmos_embed1_hf\u002F\n",[3159],{"type":53,"tag":68,"props":3160,"children":3161},{"__ignoreMap":244},[3162],{"type":59,"value":3157},{"type":53,"tag":128,"props":3164,"children":3166},{"id":3165},"known-pitfalls",[3167],{"type":59,"value":3168},"Known Pitfalls",{"type":53,"tag":2422,"props":3170,"children":3171},{},[3172,3193],{"type":53,"tag":2426,"props":3173,"children":3174},{},[3175],{"type":53,"tag":2430,"props":3176,"children":3177},{},[3178,3183,3188],{"type":53,"tag":2434,"props":3179,"children":3180},{},[3181],{"type":59,"value":3182},"Symptom",{"type":53,"tag":2434,"props":3184,"children":3185},{},[3186],{"type":59,"value":3187},"Cause",{"type":53,"tag":2434,"props":3189,"children":3190},{},[3191],{"type":59,"value":3192},"Fix",{"type":53,"tag":2456,"props":3194,"children":3195},{},[3196,3230,3262,3309,3349,3372,3390],{"type":53,"tag":2430,"props":3197,"children":3198},{},[3199,3208,3218],{"type":53,"tag":2463,"props":3200,"children":3201},{},[3202],{"type":53,"tag":68,"props":3203,"children":3205},{"className":3204},[],[3206],{"type":59,"value":3207},"MSRVTTDataset: 0 videos found",{"type":53,"tag":2463,"props":3209,"children":3210},{},[3211,3216],{"type":53,"tag":68,"props":3212,"children":3214},{"className":3213},[],[3215],{"type":59,"value":2263},{"type":59,"value":3217}," is not a local glob or metadata filenames do not match videos.",{"type":53,"tag":2463,"props":3219,"children":3220},{},[3221,3223,3229],{"type":59,"value":3222},"Mount data into the container and set ",{"type":53,"tag":68,"props":3224,"children":3226},{"className":3225},[],[3227],{"type":59,"value":3228},"mp4_urls=\u002Fdata\u002Fvideo\u002F*.mp4",{"type":59,"value":106},{"type":53,"tag":2430,"props":3231,"children":3232},{},[3233,3238,3250],{"type":53,"tag":2463,"props":3234,"children":3235},{},[3236],{"type":59,"value":3237},"HF download\u002Fauth failure",{"type":53,"tag":2463,"props":3239,"children":3240},{},[3241,3243,3248],{"type":59,"value":3242},"Missing or invalid ",{"type":53,"tag":68,"props":3244,"children":3246},{"className":3245},[],[3247],{"type":59,"value":1195},{"type":59,"value":3249},", or model agreement not accepted.",{"type":53,"tag":2463,"props":3251,"children":3252},{},[3253,3255,3261],{"type":59,"value":3254},"Accept the model terms and pass ",{"type":53,"tag":68,"props":3256,"children":3258},{"className":3257},[],[3259],{"type":59,"value":3260},"-e HF_TOKEN",{"type":59,"value":106},{"type":53,"tag":2430,"props":3263,"children":3264},{},[3265,3274,3290],{"type":53,"tag":2463,"props":3266,"children":3267},{},[3268],{"type":53,"tag":68,"props":3269,"children":3271},{"className":3270},[],[3272],{"type":59,"value":3273},"cannot import name 'Imports' from 'wandb.proto.wandb_telemetry_pb2'",{"type":53,"tag":2463,"props":3275,"children":3276},{},[3277,3282,3284,3289],{"type":53,"tag":68,"props":3278,"children":3280},{"className":3279},[],[3281],{"type":59,"value":1156},{"type":59,"value":3283}," in the container is incompatible with ",{"type":53,"tag":68,"props":3285,"children":3287},{"className":3286},[],[3288],{"type":59,"value":1100},{"type":59,"value":106},{"type":53,"tag":2463,"props":3291,"children":3292},{},[3293,3295,3301,3303,3308],{"type":59,"value":3294},"Run ",{"type":53,"tag":68,"props":3296,"children":3298},{"className":3297},[],[3299],{"type":59,"value":3300},"python -m pip install \"protobuf\u003C7\"",{"type":59,"value":3302}," in the container before invoking ",{"type":53,"tag":68,"props":3304,"children":3306},{"className":3305},[],[3307],{"type":59,"value":73},{"type":59,"value":106},{"type":53,"tag":2430,"props":3310,"children":3311},{},[3312,3323,3336],{"type":53,"tag":2463,"props":3313,"children":3314},{},[3315,3317],{"type":59,"value":3316},"Resume fails with ",{"type":53,"tag":68,"props":3318,"children":3320},{"className":3319},[],[3321],{"type":59,"value":3322},"Model does not implement 'apply_fsdp'",{"type":53,"tag":2463,"props":3324,"children":3325},{},[3326,3328,3334],{"type":59,"value":3327},"Single-GPU resume loaded a consolidated checkpoint while ",{"type":53,"tag":68,"props":3329,"children":3331},{"className":3330},[],[3332],{"type":59,"value":3333},"model.fsdp_shard_size",{"type":59,"value":3335}," stayed at the default 8.",{"type":53,"tag":2463,"props":3337,"children":3338},{},[3339,3341,3347],{"type":59,"value":3340},"Set ",{"type":53,"tag":68,"props":3342,"children":3344},{"className":3343},[],[3345],{"type":59,"value":3346},"model.fsdp_shard_size=1",{"type":59,"value":3348}," for local single-GPU resume\u002Fretrain.",{"type":53,"tag":2430,"props":3350,"children":3351},{},[3352,3357,3362],{"type":53,"tag":2463,"props":3353,"children":3354},{},[3355],{"type":59,"value":3356},"LoRA injection failure",{"type":53,"tag":2463,"props":3358,"children":3359},{},[3360],{"type":59,"value":3361},"Transformer Engine visual encoder is enabled.",{"type":53,"tag":2463,"props":3363,"children":3364},{},[3365,3366,3371],{"type":59,"value":3340},{"type":53,"tag":68,"props":3367,"children":3369},{"className":3368},[],[3370],{"type":59,"value":2714},{"type":59,"value":106},{"type":53,"tag":2430,"props":3373,"children":3374},{},[3375,3380,3385],{"type":53,"tag":2463,"props":3376,"children":3377},{},[3378],{"type":59,"value":3379},"ONNX\u002FHF export complains about missing components",{"type":53,"tag":2463,"props":3381,"children":3382},{},[3383],{"type":59,"value":3384},"Export checkpoint is partial or adapter-only.",{"type":53,"tag":2463,"props":3386,"children":3387},{},[3388],{"type":59,"value":3389},"Use a full checkpoint or configure pretrained visual\u002Ftext sources before export.",{"type":53,"tag":2430,"props":3391,"children":3392},{},[3393,3398,3403],{"type":53,"tag":2463,"props":3394,"children":3395},{},[3396],{"type":59,"value":3397},"CUDA OOM",{"type":53,"tag":2463,"props":3399,"children":3400},{},[3401],{"type":59,"value":3402},"Batch\u002Fresolution too high for the GPU.",{"type":53,"tag":2463,"props":3404,"children":3405},{},[3406],{"type":59,"value":3407},"Reduce batch size, use 224p, enable LoRA, or use more GPUs.",{"type":53,"tag":3409,"props":3410,"children":3411},"style",{},[3412],{"type":59,"value":3413},"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":3415,"total":3571},[3416,3434,3451,3462,3474,3488,3501,3513,3526,3537,3551,3560],{"slug":3417,"name":3417,"fn":3418,"description":3419,"org":3420,"tags":3421,"stars":3431,"repoUrl":3432,"updatedAt":3433},"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},[3422,3425,3428],{"name":3423,"slug":3424,"type":15},"Documentation","documentation",{"name":3426,"slug":3427,"type":15},"MCP","mcp",{"name":3429,"slug":3430,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":3435,"name":3435,"fn":3436,"description":3437,"org":3438,"tags":3439,"stars":3448,"repoUrl":3449,"updatedAt":3450},"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},[3440,3443,3446],{"name":3441,"slug":3442,"type":15},"Containers","containers",{"name":3444,"slug":3445,"type":15},"Deployment","deployment",{"name":3447,"slug":1117,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":3452,"name":3452,"fn":3453,"description":3454,"org":3455,"tags":3456,"stars":3448,"repoUrl":3449,"updatedAt":3461},"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},[3457,3460],{"name":3458,"slug":3459,"type":15},"CI\u002FCD","ci-cd",{"name":3444,"slug":3445,"type":15},"2026-07-14T05:25:59.97109",{"slug":3463,"name":3463,"fn":3464,"description":3465,"org":3466,"tags":3467,"stars":3448,"repoUrl":3449,"updatedAt":3473},"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},[3468,3469,3470],{"name":3458,"slug":3459,"type":15},{"name":3444,"slug":3445,"type":15},{"name":3471,"slug":3472,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":3475,"name":3475,"fn":3476,"description":3477,"org":3478,"tags":3479,"stars":3448,"repoUrl":3449,"updatedAt":3487},"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},[3480,3483,3484],{"name":3481,"slug":3482,"type":15},"Debugging","debugging",{"name":3471,"slug":3472,"type":15},{"name":3485,"slug":3486,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":3489,"name":3489,"fn":3490,"description":3491,"org":3492,"tags":3493,"stars":3448,"repoUrl":3449,"updatedAt":3500},"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},[3494,3497],{"name":3495,"slug":3496,"type":15},"Best Practices","best-practices",{"name":3498,"slug":3499,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":3502,"name":3502,"fn":3503,"description":3504,"org":3505,"tags":3506,"stars":3448,"repoUrl":3449,"updatedAt":3512},"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},[3507,3508,3511],{"name":17,"slug":18,"type":15},{"name":3509,"slug":3510,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":3514,"name":3514,"fn":3515,"description":3516,"org":3517,"tags":3518,"stars":3448,"repoUrl":3449,"updatedAt":3525},"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},[3519,3522],{"name":3520,"slug":3521,"type":15},"QA","qa",{"name":3523,"slug":3524,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":3527,"name":3527,"fn":3528,"description":3529,"org":3530,"tags":3531,"stars":3448,"repoUrl":3449,"updatedAt":3536},"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},[3532,3533],{"name":3444,"slug":3445,"type":15},{"name":3534,"slug":3535,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":3538,"name":3538,"fn":3539,"description":3540,"org":3541,"tags":3542,"stars":3448,"repoUrl":3449,"updatedAt":3550},"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},[3543,3546,3547],{"name":3544,"slug":3545,"type":15},"Code Review","code-review",{"name":3471,"slug":3472,"type":15},{"name":3548,"slug":3549,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":3552,"name":3552,"fn":3553,"description":3554,"org":3555,"tags":3556,"stars":3448,"repoUrl":3449,"updatedAt":3559},"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},[3557,3558],{"name":3520,"slug":3521,"type":15},{"name":3523,"slug":3524,"type":15},"2026-07-14T05:25:54.928983",{"slug":3561,"name":3561,"fn":3562,"description":3563,"org":3564,"tags":3565,"stars":3448,"repoUrl":3449,"updatedAt":3570},"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},[3566,3569],{"name":3567,"slug":3568,"type":15},"Automation","automation",{"name":3458,"slug":3459,"type":15},"2026-07-30T05:29:03.275638",496,{"items":3573,"total":3666},[3574,3591,3601,3615,3625,3638,3652],{"slug":3575,"name":3575,"fn":3576,"description":3577,"org":3578,"tags":3579,"stars":23,"repoUrl":24,"updatedAt":3590},"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},[3580,3583,3586,3587],{"name":3581,"slug":3582,"type":15},"Data Analysis","data-analysis",{"name":3584,"slug":3585,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":3588,"slug":3589,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":3592,"name":3592,"fn":3593,"description":3594,"org":3595,"tags":3596,"stars":23,"repoUrl":24,"updatedAt":3600},"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},[3597,3598,3599],{"name":3444,"slug":3445,"type":15},{"name":3534,"slug":3535,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":3602,"name":3602,"fn":3603,"description":3604,"org":3605,"tags":3606,"stars":23,"repoUrl":24,"updatedAt":3614},"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},[3607,3610,3611],{"name":3608,"slug":3609,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":3612,"slug":3613,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":3616,"name":3616,"fn":3617,"description":3618,"org":3619,"tags":3620,"stars":23,"repoUrl":24,"updatedAt":3624},"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},[3621,3622,3623],{"name":3581,"slug":3582,"type":15},{"name":9,"slug":8,"type":15},{"name":3523,"slug":3524,"type":15},"2026-07-17T05:29:03.913266",{"slug":3626,"name":3626,"fn":3627,"description":3628,"org":3629,"tags":3630,"stars":23,"repoUrl":24,"updatedAt":3637},"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},[3631,3632,3635,3636],{"name":3567,"slug":3568,"type":15},{"name":3633,"slug":3634,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"2026-07-17T05:28:53.905004",{"slug":3639,"name":3639,"fn":3640,"description":3641,"org":3642,"tags":3643,"stars":23,"repoUrl":24,"updatedAt":3651},"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},[3644,3645,3647,3648],{"name":3444,"slug":3445,"type":15},{"name":3646,"slug":486,"type":15},"Docker",{"name":9,"slug":8,"type":15},{"name":3649,"slug":3650,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":3653,"name":3653,"fn":3654,"description":3655,"org":3656,"tags":3657,"stars":23,"repoUrl":24,"updatedAt":3665},"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},[3658,3659,3662],{"name":9,"slug":8,"type":15},{"name":3660,"slug":3661,"type":15},"Quantum Computing","quantum-computing",{"name":3663,"slug":3664,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]