Principle:SqueezeAILab ETS Dependency Installation
| Knowledge Sources | |
|---|---|
| Domains | Infrastructure, Setup |
| Last Updated | 2026-02-14 02:00 GMT |
Overview
A setup procedure that installs all required software dependencies before running inference-time tree search experiments.
Description
Before any ETS experiment can execute, the runtime environment must be provisioned with the correct versions of all software dependencies. The primary dependency is sglang-ets, a custom fork of the SGLang serving framework that adds support for process reward model (PRM) scoring, KV cache state forking for tree search, and batched multi-threaded inference. Additionally, the outlines library (pinned to version 0.0.44) is required for structured generation support.
This principle ensures reproducibility by specifying exact repository URLs and version pins rather than relying on generic package indices.
Usage
Apply this principle at the very beginning of any ETS experiment setup, before launching model servers or running tree search. It is a prerequisite for all subsequent steps in the ETS Experiment Pipeline.
Theoretical Basis
Dependency management for research code follows the principle of environment reproducibility: pinning exact versions of forked libraries ensures that experimental results can be replicated across different machines and time points. The use of a custom SGLang fork (rather than the upstream release) is necessary because the ETS algorithm requires modifications to the inference server that are not available in the standard SGLang distribution, specifically:
- State forking: The ability to fork KV cache states during tree expansion
- Score backend: Integration of PRM reward scoring within the generation pipeline
- Batched execution: Multi-threaded batch processing via
run_batch