FlagScale v2.1.0

FlagScale v2.1.0

Part of FlagOS 2.2.

Changes since v2.0.0 (FlagOS 2.1).

33 pull requests and 2 additional commits.

Source: v2.1.0, commit 4371924fbfb4.

New Features

• Add perf monitor integration (#1216) by @lzy-edu
• [Model] Support to train glm5 (#1227) by @heavyrain-lzy
• [ Train] [Features] add new features of Orca model in qwen_gr00t (#1228) by @1016qqz
• Add low-overhead GPU progress heartbeat monitoring (#1243) by @lzy-edu
• [Tsingmicro] feat(backend): add txda (tsingmicro) platform (#1254) by @tsingmicro-public-e
• Add support for Qwen36 LLM Backbone and CKPT Conversion (#1273) by @lxd-cumt
• [Model] Add KERV training and inference integration (#1278) by @zhengzihaoPKU
• add profiler kernel reports (#1281) by @zhaoyinglia
• [Model] Add KERV optimized runtime operators (#1291) by @zhengzihaoPKU
• [Model] Update qwen3vl for current Megatron-LM-FL (#1235) (54957dfb479f; commit)

Hardware Support

• [Ascend] Native Integration of MegatronAdaptor (FlagScale Module) on FlagOS (#1226) by @ltllt1

Improvements

• refactor: rename VLA serve entrypoint (#1225) by @yilongma0110

Documentation

• docs(qwen35): identify main commit for training guide (#1290) by @zhaoyinglia

CI/Infrastructure

• [CICD] Add build and test CI workflows for Ascend and Metax (#1250) by @AlexMa616
• [CICD] Add S5000 image build and end-to-end validation pipeline (#1260) by @AlexMa616
• [CICD] Add Hygon BW1000 CI support (#1272) by @AlexMa616
• [CICD] Add Enflame image build and test support (#1275) by @hooray03
• [CICD] Add Kunlunxin P800 training support (#1282) by @AlexMa616
• [CICD] Remove inference and serve pipelines, focus on training… (#1283) by @AlexMa616
• [CICD]: update golden value of functional_test hetero_train and train. 2026/7/21 (#1244) (5c694d2f6dba; commit)

Bug Fixes

• fix gr00t checkpoint portability (#1219) by @yilongma0110
• fix pi05 pretrained loading memory peak (#1221) by @yilongma0110
• [Fix] Follow up on the latest Megatron-LM-FL version in the MUSA test environment (#1224) by @Darryl233
• fix qwen35 ckpt convert (#1231) by @wenqingqian
• Fix ckpt convert for qwen35 moe model (#1233) by @wenqingqian
• Fix chat template return type compatibility (#1262) by @lihongyang1990
• fix engram transformer integration (#1279) by @zhaoyinglia
• [cherry-pick][Fix] fix flagcx non hetero init (#1303) (#1304) by @zhaoyinglia

Other

• Straggler detection rebased (#1215) by @lzy-edu
• [train]: Add DeepSeek-V4 TFLOPs. (#1230) by @LiJunscs
• Dsv4 fix (#1232) by @LiJunscs
• chore: add Apache 2.0 copyright header to all source files (#1252) by @tengqm
• Configure FL backends before TP overlap initialization (#1263) by @lihongyang1990
• Upgrade/megatron v0.18.2 (#1284) by @zhaoyinglia
• Cp from main (#1301) by @zhaoyinglia

Contributors

Thanks to the contributors to the pull requests included in this release.

• @zhaoyinglia (6 PRs)
• @AlexMa616 (5 PRs)
• @lzy-edu (3 PRs)
• @yilongma0110 (3 PRs)
• @LiJunscs (2 PRs)
• @lihongyang1990 (2 PRs)
• @wenqingqian (2 PRs)
• @zhengzihaoPKU (2 PRs)
• @1016qqz
• @Darryl233
• @heavyrain-lzy
• @hooray03
• @ltllt1
• @lxd-cumt
• @tengqm
• @tsingmicro-public-e