Libra-Internal

SGLang v0.4.10 patched to record per-token expert routing for Libra (ICLR 2026). It generates the Token-to-Expert Mapping Tables (TEMTs), builds Lina's expert-selection-path tables, and simulates imbalance ratios offline.

This repository is a submodule of SNU-ARC/Libra. See the README there for installation, and REPRODUCE.md for reproducing the paper.

Libra is built on SGLang v0.4.10 and released under the Apache License 2.0 (see LICENSE).

Name		Name	Last commit message	Last commit date
.devcontainer		.devcontainer
3rdparty/amd		3rdparty/amd
assets		assets
benchmark		benchmark
docker		docker
docs		docs
examples		examples
lina		lina
prefetch_rebalance		prefetch_rebalance
python		python
scripts		scripts
sgl-kernel		sgl-kernel
sgl-router		sgl-router
test		test
.clang-format-ignore		.clang-format-ignore
.editorconfig		.editorconfig
.gitignore		.gitignore
.gitmodules		.gitmodules
.isort.cfg		.isort.cfg
.pre-commit-config.yaml		.pre-commit-config.yaml
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
bench_one_batch_token2expert_expt.sh		bench_one_batch_token2expert_expt.sh
generate_temt_glm45.sh		generate_temt_glm45.sh
generate_temt_qwen3.sh		generate_temt_qwen3.sh
get_imbalance_ratio_fluctuation.py		get_imbalance_ratio_fluctuation.py

Navigation Menu

Repository navigation

About

Repository files navigation

Libra-Internal

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Libra-Internal

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages