coreweave/inference-cookbook v0.1.1

Initial release

Repository: coreweave/inference-cookbook

Tag: v0.1.1

Published: 2026-10-13T23:01:43Z

Prerelease: no

Release notes: The first release of the CoreWeave Inference Cookbook and of cwtrain, its Python library and CLI.

Use them to train with reinforcement learning against a live CoreWeave Dedicated Inference deployment. After each training step, the trainer publishes its new weights by checkpoint hot-load (RL Rollouts). Every replica swaps them in place, and the next rollouts come from the deployment's OpenAI-compatible endpoint. The trainer runs no inference engines of its own.

Install

pip install cwtrain
pip install "cwtrain[nemo-rl]" # adds torch and safetensors for the NeMo-RL integration

cwtrain needs Python 3.12 or newer and is tested on 3.12, 3.13 and 3.14.

cwtrain

Deployments. The cwtrain deployment commands (create, get, delete, parameters) and cwtrain.infra handle the deployment that a run samples from:

• create it from a CreateDeploymentRequest body;
• wait until it is READY and get its rollout_base_url;
• delete it at the end.

Every create sends an ID chosen by the client, so a retry never creates a second resource. A request file that still has a `` is refused.

Checkpoint hot-load. The cwtrain hotload commands (publish, trigger, wait, status, list, inspect, check) and HotLoader publish the trainer's weights and wait until every replica serves them.

• Snapshot formats. A snapshot is either a full HuggingFace safetensors checkpoint or a delta.

A delta holds only the changed tensors, as a zstd-compressed XOR with an Adler-32 checksum per tensor. The engine checks the checksums before it serves the weights.

• Immutable identities. Uploads record the SHA-256 of every object. Retrying an interrupted

upload is safe. Different bytes under an existing identity are refused with SnapshotConflictError.

• Safe retries. Only one hot-load runs per deployment at a time. A retried create request

returns the hot-load that already exists.

• Rollback. HotLoader.rollback(identity) hot-loads an earlier completed snapshot again.
• Permission check. Before training, cwtrain hotload check and

python -m cwtrain.integrations.preflight check that the trainer can list, read and write the snapshot prefix. They leave nothing behind.

Framework integrations.

• miles and slime: the post-write hook cwtrain.integrations.slime.hook publishes every weight

version. In delta mode it keeps a local weight mirror and publishes a full snapshot every 20 versions (CWTRAIN_HOTLOAD_FULL_EVERY).

• NeMo-RL: the coreweave generation backend, HotLoadWeightSynchronizer and

CoreWeaveDTensorPolicyWorker.

• Rollout sampler (cwtrain.integrations.rollout), which every recipe uses:
• it sends one /v1/completions request per GRPO group and retries failed requests;
• where the endpoint supports them, it uses token-ID prompts and exact response token IDs;
• it is configured with the CWTRAIN_ROLLOUT_* variables.

Recipes

| Recipe | Framework | What it runs | Tested with | | --- | --- | --- | --- | | miles | miles | GRPO on Qwen3-4B, synchronous or one step ahead | radixark/miles:v0.1.1 | | slime | slime | GRPO on Qwen3-4B | slime 8ee9c1e | | NeMo-RL | NeMo-RL | GRPO on Nemotron 3.5 Lightning BF16, optional NeMo Gym rollouts | NeMo-RL v0.7.0 | | NeMo Gym | NeMo Gym | Scored evaluation against a temporary deployment | NeMo Gym v0.6.0 |

The training recipes run on Kubernetes and train on DAPO-Math-17k. Each one ships:

• Kubernetes manifests;
• an example deployment request;
• an example bucket policy;
• an example organization access policy for the trainer.

The trainer runs as a Kubernetes service account of its own, through Workload Identity Federation. It can write only under the hot-load prefix, and it has no rights to delete objects or to change the policy. Every training recipe writes snapshots under the same prefix, hotload. The slime and NeMo-RL launchers apply small patches to their framework; miles needs no changes.

Requirements

• RL Rollouts enabled for your organization. RL Rollouts is in preview; CoreWeave support turns

it on. Until then, creating a deployment with a hotLoad section fails with hot-load is not enabled for this organization.

• A hot-load deployment. It must use engine dynamo-vllm and have a hotLoad section, and all

its gateways must be in one zone. You can enable hot-load only when you create a deployment.

• An API token with the Inference Admin role.
• An AI Object Storage bucket whose policy lets the Dedicated Inference service account read

it. The storage region has no default: set AWS_REGION to the bucket's Availability Zone.

Known limitations

• cwtrain is alpha, so its API may change before 1.0.
• A replica that starts with no peer to copy from replays the chain from object storage, and it

handles at most 63 deltas. Publish a full snapshot at least every 64 versions. The integrations publish one every 20 versions by default.

• The API has no rollback. rollback needs the earlier snapshot's objects to still be in the

bucket.

• On large models, a hot-load takes minutes. By default, the integrations pause training until it

completes.

• The integrations keep one or two extra copies of the checkpoint on the trainer's local disk.

Documentation

• Repository README
• cwtrain README:

CLI, API layers, snapshot layout and configuration

• RL Rollouts product documentation
• Report security issues as described in

SECURITY.md, not in a public issue.

Licensed under Apache-2.0.