Skip to main content
rlcli sample lets you run quick raw generation against a running Tinker server. It does not apply a chat template. Pass either a base model name or a tinker:// checkpoint path, and the prompt text is encoded with the model tokenizer before being sent.

​
Usage

​
Flags

​
string
required
Prompt text to send to the model.
​
string
Base HuggingFace model name or path. Required unless --checkpoint is provided.
​
string
Tinker checkpoint path, e.g. tinker://.... Required unless --model is provided.
​
string
default:"http://localhost:8000"
Tinker API server URL.
​
int
default:"256"
Maximum number of tokens to generate.
​
float
default:"1.0"
Sampling temperature.

​
Tokenizer resolution

  • When --model is provided, rlcli uses that model name to load the tokenizer.
  • When only --checkpoint is provided, rlcli falls back to the RLCLI_TOKENIZER environment variable for the tokenizer name or path. If neither is set, the command fails.

​
Examples

Sample from a base model:
Sample from a trained checkpoint: