Reporting period: September 30, 2026 (Wednesday) to October 7, 2026 (Wednesday), 8 days total — National Day holiday special summary period Sources: The last pre-holiday issue (2026-09-30) and the first post-holiday issue (2026-10-08, for window-tail continuity verification) of this publication’s daily FlagOS updates; daily reports were suspended from October 1 to 7, so holiday activity is based on verification of commits, PRs, and tags across all 54 GitHub repositories of the organization (verification extended to the morning of 10-08), while news coverage is based on public searches using multiple keyword sets over the holiday window (about 9 days) Editorial note: Development within the organization did not stop during the holiday window: a total of 801 commits (712 for FlagQuantum), 412 PRs merged (312 for FlagQuantum), and 447 new PRs opened (331 for FlagQuantum); 16 repositories had pushes, and 12 repositories recorded commits to the default branch within the window. On the external news side, component-level searches remained quiet, while the industry focus was DeepSeek’s September 30 open-sourcing of infrastructure components for Ascend (see Chapter 5). The post-holiday first day (10-08) saw work resume on schedule: as of verification that morning, 14 PRs had already been recorded as merged.

I. Weekly Highlights

  • Development intensity undiminished over the holiday: 801 commits and 412 PRs merged in an 8-day window (09-30 ~ 10-07): FlagQuantum alone contributed 712 commits (88.9%) and 312 merges, making it the undisputed mainstay of the organization’s holiday development; the remaining repos totaled 89 commits, concentrated in multi-vendor CI migration and documentation batches. Commit cadence rose steadily from 10-01 onward (FlagQuantum alone: 60 commits on 10-01, 101 on 10-05, 194 on 10-06, 122 on 10-07), with intensity peaking in the middle-to-late part of the holiday.
  • FlagQuantum distributed certification campaign concludes: two-node fully enters the certification release track (10-01 ~ 10-08): On 10-01, #312 promoted two-node state vector, MPS, and tensor network training to production_supported; #318 / #321 re-recorded the two-node claim with empirical evidence and back-derived blockers; on 10-06, #535 certified two-node state vector and measured MPS / TN capacity ceilings; on the morning of 10-08, #587 certified two-node MPS training and recorded the TN capacity ceiling — two_node_mps_training was promoted to release_certified. The certification tolerance gate (#292) and evidence pinning (#348) were completed in tandem.
  • FlagGems: major KernelGen supply-side progress + multi-vendor CI migration (10-01 ~ 10-06): The kernelgen-dev branch saw 34 commits — 226 operator definitions and tests (#6853), 70 original operator definitions (#6860), test suites for batches 1–28 and the final batch (#6799 ~ #6839), and missing definition backfill (#6862, thnn_conv2d); the main branch saw 14 commits — backend tests for hygon / metax / mthreads / iluvatar migrated into runtime images (#6844), CPU runner preinstall flattening (#6847), and flagtree 0.7.0 backend parsing and PPU support (#6863 / #6852). A total of 47 merges in the window.
  • build-infra finalizes 2.2, opens 2.3 (10-01): At 17:30 in the evening, annotated tag v2.2.0 was cut (labeled “Release 2.2.0”; this repo manages its version line via tags rather than GitHub Release objects); afterward, “Bump base version to 2.3.0” (#1316) and root directory cleanup (#1317) were merged, and a workflow preparation batch (#1318 ~ #1321) was completed in the evening — the infrastructure repo’s development baseline officially switched to the 2.3 line.
  • FlagCX CRL transport hardening streak of six (09-30 ~ 10-04): #633 kernel proxy integrated into shared RMA transport, #635 collective proxy integrated with ordered multi-channel / multi-QP transport, #636 symmetric window and MR lifecycle hardening, #638 GET data flushed before publish completion, #639 unified remote visibility flush semantics (plus #637 cleaning up the now-defunct image build workflow) — transport-layer correctness hardening ran through the entire holiday.
  • FlagFFT multi-platform optimization branches advance over the holiday (10-02 ~ 10-05): MetaX maca-3d-opt and Ascend npu-2d / npu-3d-opt branches saw multiple pushes on 10-02 ~ 10-03; on the Hygon side, codex/hcu-fourstep-c2r, codex/hcu-3d-route-fix, and -dev sub-branches were opened in succession on 10-03 ~ 10-04; Moore Threads musa-3d-opt continued advancing after launching before the holiday. No new merges on the default branch; results await review in branch form; the repo’s last push was 10-05 16:10.
  • Member unit holiday batches continue to flow (09-30 ~ 10-08): MetaX W8A8 INT8 matrix multiplication (#6572) merged, Hygon bmm / baddbmm new tickets (#6865 / #6866), Kunlunxin XPU3 SDNN 46-marker fix batch (#6867), DAMO XuanTie ADDMM series performance batch (#6859), Ascend fix streak of three (#6848 ~ #6850) and Qwen3.6 inference path optimization (sglang-plugin-FL #137), Enflame specialized operator streak of four (FlagGems-vllm #889 ~ #892), Cambricon rotary embedding backend (FlagAttention #84) — see Chapter 4 for details.
  • Industry focus: DeepSeek open-sources Ascend-oriented infrastructure components (09-30): Covering TileLang high-level language compilation tools, compute libraries, and distributed communication libraries, with all components corresponding one-to-one with its NVIDIA-oriented open-source components (DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect); concurrently disclosed a joint effort with Huawei on a 128-card supernode solution based on Ascend 950 (UBL128 networking and ASC-COMM communication library support). Not a FlagOS organizational development; included as industry background (see Chapter 5).

II. Versions and Release Activity

Cycle Transition: 2.2 Finalized, 2.3 Baseline Opened

  • build-infra v2.2.0 (10-01 17:30): Annotated tag “Release 2.2.0” was pushed, pointing to the finalization of its 2.2-cycle build infrastructure; this repo manages version nodes via tags (v2.1.0 / v2.1.1 / v2.2.0), with no GitHub Release objects published.
  • 2.3 Baseline Opened (10-01 evening): #1316 “Bump base version to 2.3.0” merged at 17:37; #1317 root directory cleanup; evening #1318 ~ #1321 workflow housekeeping (unified display names, filenames standardized to .yml, release workbook triggered by filename, base image build logic inlined) — the infrastructure repo entered 2.3 development mode.
  • Pre-holiday Handoff (09-30 early morning): 5 early-morning merges completed 2.2 wrap-up — build-infra #1312 (EXTRA_PYPI fallback switched to Tencent mirror), #1313 (megatron application image tag 2.2.0-0.2.3 recorded), Torch-FL #483 / #484 (wheel pre-release fix), docs #530.

Component Tag and Release Inventory (within window)

  • Full-repo tag scan: Checked tags / releases atomic sources across all 54 repos in the organization — the only new tag in the window was build-infra v2.2.0 (10-01); no other components released new versions (the 2.2 official tag batch completed on 09-24, and the holiday was the first frozen quiet period thereafter).
  • FlagQuantum: Latest tag remains the v0.1.0 series (June); the development line is currently in the release_certified certification stage (see Chapter III), with no new tags during the holiday.
  • Development line status: FlagGems 5.5 development line (KernelGen test batch), FlagGems-Experimental (Nvidia KernelGen operator batch continuously merging), FlagGems-vllm 0.3.0-dev1, FlagTree (PPU / mthreads backend advancing) — 2.3-cycle components remained in development state during the holiday.

Images and Artifacts

  • Application image line: build-infra holiday-side work focused on record wrap-up — megatron_rl mthreads-musa5.2.0 image record (#1313), mthreads sglang documentation fix (#1314); sync-to-remote workflow removed the authorize job (09-30).
  • runtime image matrix: FlagGems migrated hygon / metax / mthreads / iluvatar backend tests to runtime images in one batch (#6844, 10-01), with accompanying single-repo CI adjustments (#6840 ~ #6843); flagtree 0.7.0 version alignment and PPU support are covered in Chapter III.

III. Components and Code Co-building

FlagQuantum (712 commits / 312 merged)

  • Distributed & certification: #312 (two-node state-vector / MPS / tensor-network training promoted to production_supported, 10-01), #318 / #321 (empirical-leg evidencing and blocker derivation, 10-01), #292 (certification tolerance gate excludes non-finite numbers, 10-01), #348 (claims pinned to published history, 10-02), #535 (state-vector certification + MPS / TN capacity ceilings, 10-06), #563 / #573 (certification lane reads and anchor instrumentation, 10-06 ~ 10-07), #587 (MPS training certification + TN capacity caps, 10-08).
  • Kernel catalog routing & rollout: on the first day of the window, merged the full catalog routing for MPS contraction, projected contraction, single-qubit matrices, fused CX / RY-RZ / RX-RZ sequences, and layout BMM (#259 ~ #281); multiple unblocking-style rollouts (#388 / #391 / #394 / #398 / #415, 10-02 ~ 10-03); the SV-002 ~ SV-010 bounded dispatch evidence series (#474 ~ #581) and NUM / MEAS / GR series evidence revisions (#595 declared sixteen undeclared revisions in one pass, 10-07).
  • QEC (error-correction line aligned with CUDA-Q): detector error model construction (#382), PyMatching cross-validation (#397), CSS matrix blocks (#404), decoder registry (#411), alignment contract with CUDA-Q (#347, 10-02), belief propagation decoding (#533), sliding-window decoding (#590, 10-08), stim text reading (#543), alignment-row convergence (#592) — the full chain from model construction and noise injection to decoding.
  • Gradients & training: fq.gradient (#384), SPSA (#395), zero-noise extrapolation (#390, all 10-02), jacobian / jvp / vjp (#562, 10-07), second-order derivatives (#575), Fubini-Study metric tensor (#585, 10-08).
  • Compiler: SABRE layout and lookahead routing (#307 / #310 / #317, 10-01), pass-manager contract (#301), arbitrary two-qubit unitary decomposition (#387, 10-02), optimization-level parameter surface (#526), routing quality quantification (#531).
  • Vocabulary migration wrap-up: WQ-3 ~ WQ-8 and WD-1 / WD-2 landed on main in one pass via #506 on 10-05; on 10-08 docs(api-changes) recorded the final slice — the “wire → qubit” terminology migration fully complete.
  • Runtime & structure: density_matrix output kind (#541), two-qubit dense gates (#544), three-qubit bounded reversible kernels (#561, all 10-06), sampling collapse (#572, 10-07), Circuit.control / Circuit.power construction-time composition (#552 / #542).
  • Cadence: merges per day 24 / 60 / 61 / 46 / 39 / 16 / 41 / 25 (09-30 ~ 10-07); 10-06’s single-day 194 commits was the peak of the entire period.

FlagGems (14 commits / 47 merged)

  • 14 commits on main: multi-vendor runtime image migration (#6844), CPU runner preinstall flattening (#6847), flagtree 0.7.0 / PPU support (#6863 / #6852), MetaX W8A8 (#6572), Kunlunxin median fix and robust testing (#6685 / #6801), Ascend global leak fix (#6770), op id normalization (#6792 / #6794), reference baseline declaration (#6800), command.yaml migration (#6798), explicit operator invocation tests (#6736), KernelGen Nvidia batch (#3951 linalg_eigh / #6610 dim).
  • 34 commits on the kernelgen-dev branch: 29 test commits (#6799 ~ #6839) + 3 definition commits (#6853 / #6860 / #6862).

FlagGems-Experimental (9 commits + tail-batch merges)

  • 9 [KernelGen][Nvidia] operators merged on 09-30 (long-tail batch of internal tensor and view classes, including _indices, _make_dual, col_indices_copy, etc., numbered #645 ~ #705); 6 more tail-batch merges on the first day after the holiday (#647 / #681 / #683 / #687 / #688 / #695).

docs (31 commits)

  • Full 2.2 documentation batch (#521 “Cheng’s 2.2 flagos drafts”, 09-30), sglang plugin docs (#527), two ARM CPU docs (#533 vllm-plugin-fl / #534 FlagTree), FlagTree naming and torch-fl rename (#535 / #539 / #540), two model-list updates (#531 / #532, further update #541 on 10-03), what’s news (#536), link and path fixes (#537 / #538).

build-infra (19 commits)

  • See Chapter II: 2.2 wrap-up (#1312 ~ #1314) → v2.2.0 tag → 2.3.0 baseline switch and workflow preparation (#1315 ~ #1321, of which #1315 is a dependency upgrade).

Other repositories

  • FlagCX (6 commits): six consecutive CRL transport hardening commits (#633 ~ #639, see Top News).
  • FlagTree (6 merged + 4 issues filed): CI fix batch (#1303 / #1315 pid path checks, #1318 / #1319 perf error display fixes including Ascend), mthreads local artifacts bumped to v0.7.0 (#1311), tle.dsa.ascend.raw interface docs (#1301); issue side: TLE gpu.alloc alias support extended to mthreads / hcu (#1316) and PPU (#1320), [TLE][Ascend] INT8 GM-to-L1 ND2NZ primitive (#1323), PPU baseline workflow (#1324, 10-08).
  • FlagAttention (3 commits): three attention frontends merged — parallax (#57), Hy3 (#56), moba (#71); issue filed for [Cambricon] rotary embedding backend (#84).
  • FlagTrain (1 commit): blocked_flash and vendor operators registered by name (#4, KMCompiler, merged 09-30); issue #5 filed ([Ascend][Metax][Thead] blocked_flash).
  • FlagGems-vllm (1 commit + 6 issues filed): dequantize_and_gather_k_cache (#873, [Ascend][MThreads] KMCompiler) merged; issues filed for four Enflame specialized operators (#889 ~ #892), [MThreads] fp8_einsum PH1 optimization (#893), [Ascend] block-scaled INT8 einsum (#894).
  • vllm-plugin-FL (2 merges on the first day after the holiday): #559 (CI self-healing + vendor plugin prebaking), #561 (Kunlunxin P800 bumped to vLLM 0.24.0), both merged on the morning of 10-08; #568 / #569 (XiWang PTPU static graph / torch 2.10 support) under review with updates.
  • Torch-FL (1 commit): #483 / #484 wheel release order fix (merged in the early hours of 09-30).
  • FlagOS-Compressor (1 commit): #8 initial version of the documentation site (09-30).
  • FlagFFT / FlagSparse: no new commits on the default branch — FlagFFT had branch-level activity (see Top News), FlagSparse only branch cleanup.
  • Repositories silent over the holiday: community, FlagScale, FlagBLAS, FlagDNN, FlagTensor, FlagAudio, KernelGen, KernelGenBench, TransformerEngine-FL, Megatron-LM-FL, FlagPrism, FlagOS-Robo and a batch of other repositories had no commits within the window (last pushes all on 09-27 ~ 09-29 or earlier); FlagGems-sglang had no commits on main, with changes within the window concentrated in the competition PR queue (see Chapter V).

IV. Member Units and Vendor Adaptation

  • BAAI (lead): The holiday period focused on organizing code and certification efforts, with external channels remaining silent (no new component-level announcements in the news feed across the entire window).
  • Hygon: runtime image migration (#6844, merged 10-01); bmm optimization (#6865, 10-06) and baddbmm optimization (#6866, 10-07) tickets filed in succession, addmm (#6696) continues to be updated post-holiday; FlagTree tle.gpu.alloc alias support (#1316); FlagFFT hcu-3d-route-fix / -fourstep-c2r branches advancing.
  • MetaX: W8A8 INT8 matrix multiplication merged (#6572); addmm / baddbmm / router GEMM specialization (#6807) and C550 MM / BMM optimization (#6811) tickets filed; FlagFFT maca-3d-opt branch saw intensive pushes from 10-02 to 10-03.
  • Moore Threads: FlagTree mthreads local artifacts bumped to v0.7.0 (#1311 merged); fp8_einsum (block-scaled fp8 bmm) PH1 optimization (#893) ticket filed; FlagFFT musa-3d-opt branch advancing.
  • Kunlunxin: median empty complex fix and vendor-robust tests (#6685), CI runtime-ization (#6801) merged; XPU3 SDNN 46-marker fix batch (#6867, early morning 10-08) ticket filed; vllm-plugin-FL P800 upgraded to vLLM 0.24.0 (#561, merged on the first day post-holiday).
  • Ascend: global leak fix merged (#6770); fix triple filed (#6848 empty batch early return / #6849 softmax intra-row chunking / #6850 pow exponent bit-width, 10-02); Qwen3.6 prefill / decode path optimization (sglang-plugin-FL #137); FlagTree CI batch (#1315 / #1318 / #1319) and INT8 primitive (#1323); FlagGems-vllm block-scaled INT8 einsum (#894); AscendNPU-IR L1 preload fix (#6, 10-03).
  • Enflame: four Enflame-specialized operator tickets filed in succession — swiglu / add_rms_norm / causal_conv1d_fn / causal_conv1d_update (FlagGems-vllm #889 ~ #892, 09-30); CI runtime-ization (#6843); vllm-plugin-FL vendor plugin pre-baking (included in #559, 10-08).
  • Iluvatar CoreX: CI runtime-ization (#6843, merged as part of #6844).
  • DAMO XuanTie: ADDMM / ADDMM_ / router GEMM / BMM / BADDBMM performance batch filed (#6859 perf(thead), 10-03); FlagTrain blocked_flash (#5) tagged with [Thead] target.
  • Cambricon: rotary embedding backend ticket filed (FlagAttention #84, 09-30).
  • Qingwei / Horizon: no new activity lines in this window (maintaining late-September status).

V. Ecosystem and Community

  • Industry side (background inclusion): DeepSeek open-sources Ascend-version infrastructure components (09-30): The open-source components correspond one-to-one with their NVIDIA platform versions — TileLang high-level language compilation tool (encapsulating Ascend C low-level instructions, carrying most operators for V4-series model training), DeepGEMM (matrix operations), DeepEP (cross-device communication), TileKernels (vector computation and memory access), FlashMLA (sparse attention), DeepSelect (data filtering); jointly advancing the Ascend 950 128-card super-node solution with Huawei (UBL128 full interconnect + ASC-COMM communication library), with multiple test groups showing compute and communication performance approaching hardware limits. This aligns with FlagOS’s multi-chip unified software stack narrative and is the most closely watched industry development of this cycle (not a dynamic of this stack).
  • Competition: FlagOS Competition Track1 (SGLang cross-chip operator optimization) submission queue continues to flow through the FlagGems-sglang main repo — gelu_tanh_and_mul (#121), deepep_permute (#123), dsv3_fused_a_gemm (#125), group_norm_silu (#127), concat_mla_k (#128), l2norm (#133), add3 (#134) and 15 others in queue; the holiday window saw a new build_trtllm_mha_page_table submission (#135, 09-30), and on the first day post-holiday add3 (#134) resumed review progress.
  • Conference: 2026 AI Open Computing Conference & Zhongzhi FlagOS Technology Conference (October 17–18, Beijing) — entering the final preparation period post-holiday (fewer than ten days countdown); the conference’s quantum-intelligence fusion session aligns with the holiday FlagQuantum certification / compiler line progress.
  • External channels: Official WeChat account and community channels silent during the holiday; public queries on components and release lines continue a calm sequence (the 09-30 daily report was the 23rd consecutive calm window, with no component-level new announcements across the entire holiday window); the BAAI community during the holiday only reposted third-party technical content (unrelated to this stack, not included in this report).

VI. Trend Observations

  • Holiday concentrated development refreshes understanding again: FlagQuantum became the absolute engine of the window with roughly 90% of commits (including a single-day peak of 194 on 10-06), while the rest of the organization ran in low-power mode (CI migration, documentation, ticket filing) — the post-holiday review queue will be digested in one burst (14 merges on the first day already signal this).
  • Release engineering completes the 2.3 handover: build-infra accomplished “2.2 final tag → 2.3 baseline switch → workflow preparation” in one smooth sequence; the 2.3 development line (FlagGems 5.5, FlagGems-vllm 0.3.0, vLLM 0.28 adaptation) is expected to accelerate convergence post-holiday.
  • KernelGen moves from “generating operators” to “definition and test supply” phase: 226 + 70 definitions and 29 test suites merged in a concentrated batch — the operator supply side is laying industrial groundwork for the 2.3 coverage sprint.
  • Vendor collaboration’s holiday low-amplitude normalization: Four vendors’ backend runtime image migrations and multiple optimization / fix tickets continued to flow during the holiday, showing that collaboration cadence no longer depends on the workday rhythm; large tasks (Hygon bmm series, Ascend fix triple) chose the rhythm of filing before the holiday and reviewing after.
  • The quantum line’s “evidence-first” methodology output: From certification campaigns to evidence revision declarations (#595 declaring sixteen revisions in one statement), the acquisition of release_certified status is grounded in reproducible measurements and auditable declarations — this engineering discipline has methodological reference value for other component lines in 2.3.
  • Next observation points: The 10-17 conference’s external releases and quantum-intelligence fusion session; 2.3 development line convergence and first tags; review and landing of Hygon bmm series, Ascend fix triple, and FlagFFT branch optimizations; the convergence direction of KernelGen batches; community manifest snapshot refresh for the 2.2 wrap-up.

Appendix: Sources and Verification