Deploy a checkpoint

Serve a full model checkpoint or merged fine-tune as a private model on Boardwalk.

Repository requirements

Export the model to a Hugging Face repository with:

• config.json describing the model architecture.
• Safetensors weights: a single file, or weight shards with model.safetensors.index.json.
• The tokenizer files and intended chat template for chat inference.

config.json
model.safetensors
tokenizer.json
tokenizer_config.json

Use the complete export from your training tools. For a merged LoRA, upload the merged model weights and configuration. A repository containing adapter_config.json is treated as an adapter; keep adapter and checkpoint exports separate.

Compatibility

Boardwalk checks the architecture, weight format, tokenizer, context length, and serving engine. Safetensors are required; pickle weight files are not supported. Qualification confirms whether the configuration can be served.

Boardwalk selects capacity and verifies the serving representation before activation. Your callable model retains its weight identity. See how serving works.

Deploy the checkpoint

1. As an organization owner or admin, open Models → Deploy a model in the console.
2. Enter the repository ID and optional branch, tag, or commit. Connect Hugging Face for private or gated repositories, and accept any required license.
3. Inspect the repository. Choose the model name and review the context and serving configuration.
4. Select Deploy. Boardwalk pins the source commit, prepares the model, runs checks, and benchmarks token rates before activation.

Large checkpoints take longer to download and prepare. Follow the phase and check results in the console. Qualification is free; inference is billed at the published token rates.

Call your model

Once active, copy its Call as value into the request’s model field. Use a key with inference:write and sufficient credits. This example uses support-agent.

curl https://api.boardwalk.cloud/v1/chat/completions \
  -H "Authorization: Bearer $BOARDWALK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"support-agent","messages":[{"role":"user","content":"Hello"}],"max_completion_tokens":256}'

Your deployment is private to your organization. Use revision pins when you need to select a specific served weight commit.

Troubleshooting

• Missing weights: confirm all shards referenced by the weight index are present.
• Unsupported configuration: read the failed screening or engine check. Changing a file extension does not convert a weight format.
• Incorrect chat output: verify that the exported tokenizer and chat template match training.
• Deployment still running: check the current phase before retrying; provisioning and large downloads can take several minutes.

Redeploy to publish a new checkpoint. The existing revision keeps serving until the new one passes qualification. If a check fails, correct the repository and try again; include the deployment ID when contacting support.