All pull requests

Pull requests

• feat(metax): add BGE-M3 initial model adaptation

  buildcoredocsexamplestests
  #636·cshujie131-bot opened Oct 25, 2026·3 tasks done·

  ·
  1

• feat(metax): add Qwen3.8-Flash-Next BF16 baseline

  #635·cshujie131-bot opened Oct 25, 2026·3 tasks done·

  ·
  4

• feat(thead): support MiniMax-M3.1 W8A8 inference

  coretests
  #634·ZenithJoyH opened Oct 23, 2026Contributor·2 of 3 tasks·

  ·
  1

• feat(ascend): add NpuProfilerWrapper for torch_npu profiling

  core
  #629·yequ6154 opened Oct 22, 2026·

  ·
  1

• [CICD]: report pytest results to FlagCICD

  cidocstests
  #627·hooray03 opened Oct 22, 2026Contributor·3 tasks·

  ·

• feat(ascend): support BF16 MoE expert parallelism

  coretests
  #625·xiemingda-1002 opened Oct 21, 2026Contributor·1 of 2 tasks·

  ·
  4

• [debug, never merge] probe: musa e2e concurrency without e2e_serial

  citests
  #624·lentil1016 opened Oct 21, 2026Contributor·

  ·
  2

• ci: block newly added Chinese characters in .py diffs

  ci
  #623·lentil1016 opened Oct 21, 2026Contributor·

  ·

• feat(hy4): add INT8 sparse indexer fallback on pre-FP8 GPUs

  #622·CherryLemon opened Oct 21, 2026Collaborator·

  ·

• fix(ascend): backport GDN patch imports for vllm 0.24.0 module split

  core
  #621·yequ6154 opened Oct 21, 2026·

  ·
  3

• feat(nvidia): adapt native CUDA vLLM 0.32 runtime and CI

  buildcicoredocstests
  #615·physics31415926 opened Oct 20, 2026Collaborator·

  ·
  6

• feat(hy4/metax): graph-mode investigation for Hy4-preview on MetaX C550 (superset of #611; graph does not converge)

  coretests
  #614·tonyh168 opened Oct 19, 2026·5 of 6 tasks·

  ·
  1

• feat(hy4): day-0 support for Hy4-preview on MetaX C550

  coretests
  #611·tonyh168 opened Oct 19, 2026·3 tasks·

  ·
  1

• fix(thead): remove unused FlashAttention capability imports

  core
  #608·ZenithJoyH opened Oct 19, 2026Contributor·3 tasks·

  ·

• perf(metax): optimize Kimi K3 TP8/PP4 inference

  buildcoredocsexamplestests
  #607·cshujie131-bot opened Oct 18, 2026·3 tasks done·

  ·
  1

• Day0 qwen3.8 flash next 1015 metax

  buildcoredocstests
  #605·tonyh168 opened Oct 16, 2026·3 tasks·

  ·
  1

• [CICD]: update FlagCICD metrics endpoint

  ci
  #604·hooray03 opened Oct 16, 2026Contributor·

  ·
  5

• ci: move H20 NCCL NVLS workaround into per-device config (device_env)

  ci
  #601·lentil1016 opened Oct 16, 2026Contributor·

  ·

• [CICD] Add platform-aware routing and controls

  cidocs
  #599·hooray03 opened Oct 15, 2026Contributor·3 tasks·

  ·

• Use existing FlagGems-vllm operator imports in active platform CI

  cicoredocstests
  #597·cyber-pioneer opened Oct 15, 2026Collaborator·3 tasks done·

  ·

• perf(metax): accelerate Qwen3-VL top-k/top-p sampling

  core
  #596·liyu030 opened Oct 15, 2026·3 tasks·

  ·

• perf(metax): add Qwen3-VL TP2 MoE tuning config

  core
  #595·liyu030 opened Oct 15, 2026·3 tasks·

  ·

• ci: port the workflow rework (stage model, routing, slash commands) onto 0.4.0-dev

  buildcitests
  #592·lentil1016 opened Oct 14, 2026Contributor·

  ·
  1

• fix(metax): route reshape_and_cache_flash through FL dispatch (port to 0.4.0-dev)

  core
  #591·tonyh168 opened Oct 14, 2026·3 tasks·

  ·
  1

• discover multi-node Prefill TP addresses without a forward

  coredocsexamplestests
  #590·ymmyys opened Oct 14, 2026·

  ·