Tao Finetune Cosmos Reason

Cosmos-RL

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

Supervised fine-tuning (SFT) of Cosmos Reason video QA models. The packaged default base model is hf_model://nvidia/Cosmos3-Nano. Pretrained weights are sourced from HuggingFace, not NGC. Gated HuggingFace models require HF_TOKEN. Some Cosmos-RL images cannot load the native Cosmos3 Omni checkpoint format directly; for those images, convert Cosmos3-Nano to a Qwen3-VL HF safetensors directory before train/evaluate and use that converted directory as the PTM path.

Uses FSDP-based parallelism with dp_shard_size for GPU count and dp_replicate_size for node count (not the standard num_gpus/num_nodes).

Requests for "Cosmos Reason 3", "Cosmos3 Nano Reasoner", or nvidia/Cosmos3-Nano are handled by this skill. There is no separate Cosmos3 model directory in the skill bank; route those requests here. Override the base HuggingFace model only when the user explicitly asks for a different model.

Execution Environment — Hard Gate

references/skill_info.yaml:container_image is the default runtime for every Cosmos-RL action, including AutoML child jobs. Resolve and preflight that image before creating training files, installing training dependencies, downloading the checkpoint, or launching a smoke test. A user-supplied image=<override> replaces the resolved default for that run.

Do not create or use a host venv, uv environment, generic NGC PyTorch image, or custom Transformers training project merely because the base model is on Hugging Face. Venv-based training is allowed only when the user explicitly selects venv execution. A host Python environment used to run TAO SDK or the AutoML controller is control-plane-only and does not change the child-job runtime: training and evaluation still run in the resolved Cosmos-RL image.

If this skill and tao-finetune-huggingface-model both appear to match, this dedicated model skill wins.

Deep detail lives in references; load the smallest one that matches the task:

• references/cosmos-reason-launch.md — launch intake, preflight, per-action dataset requirements, spec construction, typical overrides.
• references/cosmos-reason-evaluate.md — evaluate (flat TOML, task types, LoRA eval, selective download, results) and datasets.
• references/cosmos-reason-automl.md — AutoML/HPO policy and search-space guidance.
• references/cosmos-reason-parameters.md — important parameters, hardware, error patterns, DEFT/gap analysis, parent-model inference mappings.
• references/cosmos-reason-wts-gb300.md — deterministic WTS dataset/runtime helpers and single-GB300 launch guards.

Quick Start (docker run)

Docker-native launch in the resolved Cosmos-RL image — no TAO SDK and no host Python. This satisfies the Execution Environment hard gate above: the container is the runtime. Use the local Docker/platform skill instead when it gives a stricter environment-specific command.

COSMOS_RL_IMAGE_DEFAULT=nvcr.io/nvidia/tao/tao-toolkit:7.1.0-cosmos-rl  # versions-key: images.tao_toolkit.cosmos_rl
COSMOS_RL_IMAGE="${COSMOS_RL_IMAGE:-$COSMOS_RL_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --ipc=host
  --shm-size=64g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -e HF_TOKEN
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)

Train:

docker run "${DOCKER_COMMON[@]}" "$COSMOS_RL_IMAGE" \
  cosmos-rl --config /specs/train.toml /opt/cosmos_rl/tao_sft_example.py

Evaluate:

docker run "${DOCKER_COMMON[@]}" "$COSMOS_RL_IMAGE" \
  cosmos-rl-evaluate --config /specs/evaluate.toml

Specs are TOML for this skill, not YAML. inference and quantize are driven by spec-reading wrappers rather than a single flag-for-flag command; run them through the same image and see references/cosmos-reason-launch.md for the exact form. Pass HF_TOKEN as an environment variable — never in a spec or log.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML still requires schemas/train.schema.json and references/spec_template_train.yaml to exist and parse. Use the packaged train schema for automl_default_parameters, automl_disabled_parameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Non-train actions such as evaluate, inference, and quantize stay in this model skill. The per-run automl_policy override does not change model metadata.

Credentials

• HF_TOKEN (required for gated models): HuggingFace access token. For the packaged default, the user must accept the model agreement at https://huggingface.co/nvidia/Cosmos3-Nano and provide a token with read access. If the user explicitly overrides the base model, they must accept that target model's agreement too. Passed to the container as a docker_env_var.

Cosmos3 Checkpoint Conversion

When a selected image cannot load the native Cosmos3 checkpoint format (model_type="cosmos3_omni" or Cosmos3ForConditionalGeneration), do not patch QwenVL, Transformers, or vLLM first. Use the upstream Cosmos Framework VLM conversion path to produce a Qwen3-VL HF safetensors directory, then point Cosmos-RL specs at that converted directory.

The model skill packages a helper:

python skills/models/tao-finetune-cosmos-reason/scripts/prepare_cosmos3_vlm_checkpoint.py \
  --checkpoint-path /abs/path/Cosmos3-Nano \
  --output-path /abs/path/Cosmos3-Nano-VLM \
  --secrets-env ~/.tao/secrets.env \
  --validate-with-image <cosmos-rl-image>

After conversion, use the converted directory consistently as the PTM:

train:    policy.model_name_or_path=/abs/path/Cosmos3-Nano-VLM
evaluate: model.model_name=/abs/path/Cosmos3-Nano-VLM
evaluate: model.base_model_path=/abs/path/Cosmos3-Nano-VLM

For local Docker, mount the converted directory read-only into the Cosmos-RL container and set the spec to the container path. If a converted copy already exists and validates, reuse it for PTM baseline evaluation, AutoML recommendations, and final best-checkpoint evaluation rather than converting again.

Training Requirements

The model-level GPU host requirements in references/skill_info.yaml override the TAO-wide platform defaults for this workflow. Apply that profile through the shared host setup check before launch, as shown in references/cosmos-reason-launch.md. Treat the versions as lower bounds, not exact pins, and retain the selected-image native GPU smoke test.

• Dataset type: vlm
• Formats: llava, daft
• Accepted dataset intents: training, evaluation, testing
• Monitoring metric: val/avg_loss, val/reward_avg, val/loss
• Dataset URI examples: s3://bucket/cosmos/train, s3://bucket/cosmos/eval, /lustre/fsw/tao_datasets/cosmos_rl/train, /lustre/fsw/tao_datasets/cosmos_rl/eval
• Input modes: accept either dataset roots or direct spec-key paths. Root mode maps <root>/annotations.json plus <root> as the media path. Direct spec mode is valid when annotations and media live in different locations, for example custom.train_dataset.annotation_path=/lustre/.../train.json and custom.train_dataset.media_path=/lustre/.../videos.tar.gz.
• Media handling: do not ask the user to choose videos.tar.gz vs images.tar.gz unless they are using direct spec mode or the model/action requires a single media archive. In root mode, pass the dataset root as the media path.
• Annotation validation: before launching train/AutoML/evaluate, verify the annotation JSON is readable and the referenced media path or archive is visible from the selected platform. Do not block, patch, or mutate annotations solely because optional fields are absent.
• Per-record video FPS: the packaged train template uses custom.vision.nframes, so per-record video_fps is not required by default. If the user switches to custom.vision.fps, selects a dataset profile that requires per-record timing, or uses an image/version that requires video_fps, make it a preflight requirement with --json-required-field train_annotation=video_fps and --json-required-field val_annotation=video_fps before any download or job launch.

Spec construction

cosmos-rl is mode: config. Always start from the packaged references/spec_template_<action>.yaml for the requested action — load it as your base spec via yaml.safe_load(...) and apply user overrides on top. Don't rebuild from scratch.

import yaml
from pathlib import Path

skill = Path.home() / "tao-sdk/tao-skills-external/skills/models/tao-finetune-cosmos-reason"
action = "train"  # train, evaluate, inference, or quantize
specs = yaml.safe_load((skill / f"references/spec_template_{action}.yaml").read_text())
# Now apply your overrides on top of `specs`.

The reference TOML (and the spec the model actually consumes) is nested dicts, not flat dotted keys. Dotted notation in override examples denotes paths into the nested spec — walk the path and assign at the leaf. See skills/platform/tao-run-platform/SKILL.md's "spec is nested dicts" callout. Data source overrides are mandatory for every action.

The packaged template keeps custom.vision.nframes=8 for bounded 1-GPU memory; switch to fps only after checking token budget and GPU memory, and delete custom.vision.nframes from the spec when you do.

Epoch-based cadence policy

Training duration, checkpointing, and validation are epoch-based by default:

• Set training duration with train.epoch.
• Keep train.ckpt.save_freq_in_epoch=1 and do not disable it or replace it with train.ckpt.save_freq unless the user explicitly requests step-based checkpointing.
• Keep validation.freq_in_epoch=1 and do not replace it with validation.freq unless the user explicitly requests step-based validation.

Dataset, GPU, topology, memory, and runtime-image guidance must not change these cadence defaults. A hardware- or dataset-specific reference may explain how to implement an explicitly requested step cadence, but must never infer or select step cadence on the user's behalf.

See references/cosmos-reason-launch.md for launch intake, the full check_tao_launch_preflight.py slurm/local-Docker examples, the video_fps preflight example, S3 staging, the GPU resource/architecture gate, the per-action dataset requirements table, the /workspace mount caveat, the quantize compatibility shim, and the full typical-overrides list.

Critical Overrides (Train)

These are the keys whose template defaults are wrong or where omission flips the run into a different mode:

Evaluate

The evaluator reads a flat TOML config (dataset, model, task, evaluation, vision, generation, metrics, results, num_gpus, results_dir); the actions.evaluate block in references/skill_info.yaml declares inputs and outputs. See references/cosmos-reason-evaluate.md for the flat-TOML config detail, task types ("" General Evaluator vs "its_directionality"), LoRA evaluation via spec_overrides, selective download, results/metrics, and the datasets section.

AutoML / HPO Notes

The packaged default base model is hf_model://nvidia/Cosmos3-Nano; apply it consistently to train (policy.model_name_or_path) and post-training evaluation (model.base_model_path) unless the user provides a different model. See references/cosmos-reason-automl.md for accuracy-vs-val/avg_loss objective selection, the eval_fn per-recommendation evaluate flow, the knob mapping (learning rate, batch size, epochs, weight decay, warmup ratio), example custom_param_ranges, train_sample_count batch-size capping, ordered_int requirements, and the pre-launch recommendation summary.

Parameters, Hardware, Errors, DEFT, Inference

For parallelism, set policy.parallelism.dp_shard_size = GPUs per node and policy.parallelism.dp_replicate_size = node count (1 for single node). Cosmos-RL handles distributed init internally via FSDP and does not rely on platform-level MASTER_ADDR/WORLD_SIZE; submit with gpu_count=<gpus_per_node> and num_nodes=<N> and the spec keys drive sharding. SFT requires at least 256 GB of cumulative visible GPU memory rather than a fixed device count or per-device capacity. Set dp_shard_size and the platform GPU request to the actual visible GPU count, and use dp_replicate_size=1 for a single node. The selected image must support every visible GPU architecture and pass its runtime CUDA-stack smoke test.

See references/cosmos-reason-parameters.md for important parameters (training loop, model/policy, parallelism incl. multi-node FSDP, optimization, vision encoders, checkpointing incl. the best symlink/epoch_* resolution, validation, logging), hardware sizing, the full error-pattern catalog (CUDA OOM, LoRA-eval OOM, NaN loss, vision_embeds mismatch, quantize token mismatch, batch-size divisibility and per-rank limits, stale cache, scheduler-None, gated-repo HF_TOKEN, GPU resource/architecture gate, status-logging warnings), DEFT support and scripts/analyze_gaps.py gap analysis, and the parent-model inference mapping table.