Checkpoint and adapter requirements
Use a complete export from your training tools. Boardwalk serves the pinned source in its published precision. To use lower-precision weights, deploy a quantized checkpoint.
Full checkpoints
A Hugging Face repository needs the model configuration, safetensors weights, and tokenizer. Include the intended chat template for chat inference.
config.json
model.safetensors
tokenizer.json
tokenizer_config.jsonSharded weights require model.safetensors.index.json and every referenced shard. Pickle weight files are unsupported. Use a standard tokenizer.json; custom tokenizer code is accepted only when it can be converted without changing token IDs or text.
LoRA adapters
Keep adapter and full-checkpoint exports in separate repositories. An adapter repository needs these files and an accessible base model identified in its configuration:
adapter_config.json
adapter_model.safetensorsBoardwalk resolves the base model and serves the adapter under its own private model name. Connect Hugging Face if the adapter or base is private or gated.
Compatibility
The selected engine must support the architecture and weight format, and the weights plus a full-context request must fit on the selected replica. A rejected deployment explains what to change. Repository code does not run implicitly on a GPU host.
Common fixes
| Problem | What to check |
|---|---|
| Cannot read the source | Repository ID, pinned ref, Hugging Face connection, and any required license acceptance. |
| Missing weights or tokenizer | Upload the complete export, including all shards and the index file. |
| Model does not fit | Review GPU size, context length, KV precision, or a quantized checkpoint. |
| Unsupported architecture or settings | Read the run log and choose a compatible engine and configuration. |
Ready to run it? Follow Deploy a model.