sglang-plugin-FL v0.2.0

Part of FlagOS 2.2.

Changes since v0.1.0 (FlagOS 2.1).

41 pull requests and 0 additional commits.

Source: v0.2.0, commit 1c84db70fc33.

New Features

feat: add multi-node inference verification examples for Qwen3.6 models (#16) by @yixiaodapeng
feat: add pipeline parallelism support to multi-node examples (#23) by @yixiaodapeng
feat: complete FlagCX full communication replacement for PP support (#26) by @yixiaodapeng
feat: add thead (Thead/PPU) CUDA-compatible vendor routing (#27) by @yixiaodapeng
feat: multi-platform attention backend support (#28) by @huangxu0914
feat: enable cuda graph for ascend and musa platforms (#31) by @huangxu0914
feat: enable FlagCX comm on Ascend / MUSA and add platform-aware multinode examples (#32) by @huangxu0914
feat: add TXDA vendor backend with op implementations and platform support (#33) by @tsingmicro-public-e
feat: add Iluvatar backend for SGLang-FL op dispatch (#34) by @DannyP0
feat: add hcu vendor (#35) by @wangxshuai
feat: add end-to-end throughput benchmark script for sglang server (#40) by @huangxu0914
feat: add Kunlunxin (KLX) out-of-tree vendor backend (#41) by @liquanfeng
feat: add enflame gcu (#42) by @EF-Knight
feat: add empty device support for national platform deployment (#43) by @yixiaodapeng
[NPU] Add ascend vendor patch for PP comm and qwen3-vl (#49) by @huangxu0914
[e2e] Feat/musa sglang0516 (#54) by @lentil1016
[e2e] Feat/ascend npu e2e (#57) by @lentil1016
feat: add MTP (Multi-Token Prediction) support for Qwen3.6-27B (#58) by @yixiaodapeng
feat(disagg): add FlagCX KV transfer backend for PD disaggregation (#59) by @leoda1
[PPU] Add PPU default attention_backend into _ATTN_BACKEND_MAP (#63) by @huangxu0914
[e2e] feat(thead): add T-Head PPU (PPU-ZW810E) CI platform (#64) by @lentil1016

Hardware Support

[NPU] Maintaining the FlagGems blacklist using YAML configuration (#55) by @huangxu0914
[MUSA] Maintaining the FlagGems blacklist using YAML configuration (#56) by @huangxu0914
[MUSA]: bypass torch_musa isin op and fix vl accuracy (#79) by @huangxu0914

Improvements

refactor: DeviceInfo service class + vendor early-patches mechanism (#73) by @huangxu0914

CI/Infrastructure

ci: add CUDA CI pipeline with lint, build, and test stages (#37) by @lentil1016
[CICD] Add cuda CICD test (#45) by @xinxin12345
ci(feishu): align Feishu notification with vllm-plugin-FL (#60) by @lentil1016
ci: add /rerun-failed-ci slash command for PR authors (#87) by @lentil1016

Bug Fixes

fix: add dispatch cache for strict-mode call() to eliminate L2 overhead (#21) by @yixiaodapeng
fix: compat with FlagGems 5.3.0-rc2 DeviceDetector path change (#29) (#30) by @yixiaodapeng
[MUSA]Fix/pp multi req hang (#36) by @superwjfeng
fix: align SGLANG_FL_STRICT semantics (#69) by @xinxin12345
fix: rename hcu backend module path to hygon (#78) by @limylily
fix(examples): validate full prime sequence in multinode tests (#80) by @xinxin12345
fix: restore cuda cicd container_volumes path (#83) by @huangxu0914
fix: update pyproject version to 0.2.0 to match release tag (#119) by @huangxu0914