Chain runs into a pipeline

tp submit my-sft --profile h200:8
tp submit my-dpo --profile h200:8 --init-from run/run-4f2a91/checkpoint

Stage two starts from stage one’s checkpoint. The server resolves the reference, checks the artifact exists and that you may read it, and injects the real path into the job. Post-training is rarely one step: SFT, then DPO or GRPO, then export and evaluation. Two flags cover it — --init-from for the hand-off, --then for what follows automatically.

Start from a previous run’s checkpoint

# the latest checkpoint of that run
tp submit my-dpo --profile h200:8 --init-from run/run-4f2a91/checkpoint

# a specific step
tp submit my-grpo --profile h200:8 --init-from "run/run-8b1c02/checkpoint@step=120"

Every run’s checkpoint is registered when it succeeds, so the next stage only needs the run id. The new run records a Derivation back to the old one, and the Artifacts tab on either job page shows the whole chain. A Derivation crosses projects and owners; Pipeline membership does not.

Run export and evaluation automatically

tp submit my-grpo --profile h200:8 --then export --then eval

Each --then becomes a Step of the same Pipeline, submitted once training succeeds. The job list shows one row for the Pipeline, with a state rolled up from its Steps. When training succeeds and a later Step does not happen, that state is partial — see job states.

Export to HuggingFace format by hand

tp export my-grpo \
  --checkpoint checkpoints/step_500 \
  --checkpoint-format nemo-dcp

Formats: nemo-dcp, nemo-megatron, verl-fsdp, verl-megatron, huggingface. The checkpoint path is the one the artifact registry recorded, which the job page lists under Artifacts. To push the result to the Hub:

tp export my-grpo --checkpoint checkpoints/step_500 \
  --checkpoint-format nemo-dcp \
  --push-repo myorg/my-model

The server injects your own linked Hugging Face token. Link one first — see the Hugging Face integration.

Run the framework’s own evaluation

# NeMo-RL, against an exported model
tp eval my-grpo --model /shared/hf_export/my-model

# NeMo-RL, with an explicit evaluation config
tp eval my-grpo --model Qwen/Qwen3.5-9B --eval-config configs/eval/gsm8k.yaml

# verl SFT, reporting validation samples back onto the training run
tp eval my-verl-sft --run-id run-4f2a91 --step 200 --data data/val.parquet

tp eval reuses the training framework’s own inference path. For scores you can compare across models and teams, use tp bench instead — those land in the benchmark dashboard. See benchmarks.

A pipeline end to end

# 1. SFT as the base
tp new sft-qwen --method nemo-rl/sft
tp submit sft-qwen --profile h200:8

# 2. GRPO from the SFT checkpoint, exporting when it finishes
tp new grpo-qwen --method nemo-rl/grpo
tp submit grpo-qwen --profile h200:8 \
  --init-from run/run-4f2a91/checkpoint --then export

# 3. Score it against standard benchmarks
tp bench run bench-qwen -m run:run-8b1c02 --suites gsm8k,mmlu

# 4. Try it
tp serve start run:run-8b1c02 --gpus 1

Check it worked

The job page’s Artifacts tab lists what this run registered and the run it derived from. A Pipeline that reads succeeded ran every declared Step; partial means one of them did not happen.

Model registry to turn a checkpoint into a version · Benchmarks · Playground

Get started

Core concepts

Train a model

Evaluate and serve

Agent environments

Extend the platform

Collaborate

Integrations

Reference

Help

Chain runs into a pipeline

Start from a previous run’s checkpoint

Run export and evaluation automatically

Export to HuggingFace format by hand

Run the framework’s own evaluation

A pipeline end to end

Check it worked

Next

​Start from a previous run’s checkpoint

​Run export and evaluation automatically

​Export to HuggingFace format by hand

​Run the framework’s own evaluation

​A pipeline end to end

​Check it worked

​Next

Start from a previous run’s checkpoint

Run export and evaluation automatically

Export to HuggingFace format by hand

Run the framework’s own evaluation

A pipeline end to end

Check it worked

Next