START WITH YOUR WORKSPACE

Connect to see real runs.

Use the API token configured for this Weavely server. Start from an editable definition. The YAML contains the image, inputs, resources, and execution settings.

No sample workloads are created. Training submission and model deployment use your configured backend.

EXPERIMENT · INSPECT · REPRODUCE

Training runs

Find what changed. Keep the exact binary. Re-run with confidence.

Select for comparison	Run / image	Status	Latest scalar	Allocation cost	Created

Your next experiment starts here.

Submit a training definition, then inspect status, logs, W&B results, and allocation costs in one place.

weavely run submit training.yaml

Training curves stay in W&B. Scalars shown here are reported values from driver logs; costs are estimates from configured allocation rates.

VIBE TESTS · DEMOS · RESEARCH

Research inference

Deploy a model for an experiment. Inspect its endpoint and runtime.

One model replica for the MVP. Your existing Kubernetes networking carries requests directly to the model runtime. Weavely manages deployment and observation.

Endpoint / model	Status	Requests in flight	Allocation cost	Created

Give your model a place to run.

Choose a pinned runtime image, model artifact, and GPU resources. Review the complete definition before deploying.

weavely serve deploy serving.yaml

Runtime metrics appear only when the backend reports them. Missing observations are not zero traffic. Production routing and autoscaling are outside this MVP.

TRAINING / RL · DEPLOYED INFERENCE

Evaluation

Evaluate training checkpoints in their RL runtime, or test an existing inference deployment. Compare both with a shared scoring protocol.

Prototype · exact-match evaluation. The harness code, dataset, prompt format, generation settings, and scorer identify the protocol. Runtime images and topology are tracked separately; matching protocols do not imply identical outputs.

Select for comparison	Evaluation / target	Status	Exact match	Allocation cost	Created

A shared test for every model.

Point the same evaluation harness at a checkpoint or endpoint, then inspect comparable scores and result artifacts.

weavely eval submit evaluation.yaml

Evaluation scores come from the completed harness result. No score is displayed until a result is available.

Name optionalImage · immutable digestThe exact executable image is part of this definition.Model pathDataset paths one per line

WorkersGPUs per worker

Execution settings All fields also remain editable in YAML

CPU per workerMemory per worker

Model parallelism

Tensor · TPExpert · EPPipeline · PP

Context · CPExpert tensor · ETP

Training

Model name

RolloutsRollout batch sizeSamples per promptResponse token limitLearning rateSave checkpoint

Rollout TPRollout EPGPUs per rollout engine

Serving limits

Maximum running requestsContext length

Generation & scoring

TemperatureTop PMaximum tokensSeedExample limit optionalNormalize whitespace

Prototype scoring uses exact_match. Generation and scoring settings are included in the protocol identity.

Scheduling

PriorityNot before optional

Image, data, GPU layout, parallelism, and training settings are stored inline. Use Edit YAML for every additional option.

training.yaml

One file describes the work.

Preview the complete definition, then inspect the Kubernetes objects it generates. Keep the YAML with your experiment.

THE CLI WAYweavely run submit training.yaml

Connect to see real runs.

Training runs

Your next experiment starts here.

Research inference

Give your model a place to run.

Evaluation

A shared test for every model.

Connect your workspace

An explicit definition.

One file describes the work.

Inspect what will run.

Compare training runs

Choose an exported checkpoint.

Connect to see real runs.

Training runs

Runs 0

Your next experiment starts here.

Research inference

Endpoints 0

Give your model a place to run.

Evaluation

Evaluations 0

A shared test for every model.