What's Changed
- Add python API for programmatic access
- Unified client for CLI and API: Refactored to use a common client for consistent behavior and reduced code duplication.
- Dynamic Slurm script generation: Replaced static slurm scripts with job-specific dynamically generated slurm scripts, the generated scripts will live in the log directory and can be reused without needing to provide any additional argument.
- vLLM engine arguments dynamic parsing: Replaced the selected few vLLM engine arguments with a generic
vllm_args
parameter so that users can pass in any vLLM engine argument they need, updated model config to accommodate this change.
- Decoupled code base from Vector cluster environment: Stripped out all Vector cluster environment related variables to a public facing python file so that VI package can be easily adapted to other slurm environments.
- Migrate documentation build to mkdocs
- Disable singularity automatic home directory mounting
- Add support for multi-version docs hosted via github pages
New Contributors
Other Contributors
@XkunW @amrit110 @kohankhaki @jwilles