Research window: 2026-10-10 09:40 ~ 2026-10-12 06:00 (Beijing time, approx. 44 hours; Monday morning report, window starts from the cutoff of the previous daily report (10-10 makeup report for the adjusted holiday), covering Saturday morning through early Monday morning, including all of Sunday).

Sources: GitHub (org: flagos-ai, full verification of pushed_at across 55 repos, full verification of 102 commit searches, verification of merged and open PR searches (69 / 165), atomic verification of Releases across 26 repos and tags across 7 repos, cross-supplementary checks of atom feeds and branch-level data across 5 repos), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia, Tavily supplementary search, official CSDN column and BAAI Community verification (see appendix source list for details).

Issue Index

  • Today’s Highlights: Version window opens — three FlagQuantum 0.3.0 RCs in a row, Tsingmicro / Kunlunxin dual-track onboarding to Torch-FL (10-10 ~ 10-12)
  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 Organizational Cadence: 19 repos pushed, 102 commits, 69 merges (10-10 ~ 10-12)
    • 1.2 FlagQuantum: Three Consecutive 0.3.0 RCs — stable ABI, cross-platform wheels for three platforms, and Jiuding multi-device integration (10-10 ~ 10-11)
    • 1.3 FlagGems: 19 merges — Enflame S60 joins test backends, Kunlunxin optimization batch and KernelGen pipeline (10-10)
    • 1.4 FlagGems Under Review: Competition Track1 continuation, new KernelGen faces, and XPU3 serial (10-10 ~ 10-12)
    • 1.5 FlagTree: TLE-Raw backported to 3.5.x, Ascend tle.pipe and mthreads guard fixes (10-10 ~ 10-11)
    • 1.6 Torch-FL: Tsingmicro / Kunlunxin onboarding lands and runtime consistency suite opens (10-10 ~ 10-12)
    • 1.7 FlagCX and build-infra: CRL artifact pool / IPC wrap-up, wheel directory automation (10-10 ~ 10-11)
    • 1.8 Other Tracks: FlagOS-Compressor / FlagGems-vllm / vllm-plugin-FL / training side and docs (10-10 ~ 10-12)
    • 1.9 Governance and Repo Landscape: FlagOS 2.3 cycle opens and new repo FlagQAI added (10-10)
  • II. News Coverage and Ecosystem
    • 2.1 Ecosystem Roundup: MetaX university open-source tour Shanghai stop visits three campuses next week (10-10)
    • 2.2 Supplementary Coverage: Biren Technology’s HK$4 billion placement closes — 70% directed to supply chain and capacity (10-08 / 10-11)
    • 2.3 Search and Official Channel Verification: Component lines remain quiet (10-10 ~ 10-12)
  • III. Member Deep Dives
    • 3.1 MetaX: Runtime packaging protection, GDN fused kernels, and university open-source tour (10-10 ~ 10-12)
    • 3.2 Kunlunxin: Torch-FL onboarding, vLLM FlagTree support, and XPU3 batch progress (10-10 ~ 10-12)
    • 3.3 Tsingmicro: Torch-FL “FlagGems-only” operator route lands (10-10)
    • 3.4 DAMO XuanTie / Ascend: PPU CI dual-repo landing, Ascend fix batch (10-10 ~ 10-11)
    • 3.5 Moore Threads / Enflame / Iluvatar: MUSA 0.28 and S60 backend, conv and attention patches (10-10 ~ 10-12)
    • 3.6 Other Members: Hygon / Cambricon / Biren / Horizon (10-10 ~ 10-11)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Full Source List

Today’s Highlights: Version Window Opens — FlagQuantum 0.3.0 Ships Three Consecutive RCs, Tsingmicro / Kunlunxin Dual-Track Integration into Torch-FL (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Sources: FlagQuantum v0.3.0rc3 · FlagQuantum #613 · Torch-FL #495 · Torch-FL #497 · build-infra #1331 (full list in appendix)

The weekend window (including all of Sunday) sustained the post-holiday high: 19 of 55 repos received pushes during the window, commit search returned 102 hits, 69 PRs were merged (all after the previous cutoff), and 165 open PRs were updated within the window; 3 new Releases appeared (three FlagQuantum release candidates), no other new tags, and 1 new repo was added (FlagQAI). Three main threads:

  • Version window opens: The prior observation point — “when will the first new Release since 09-28 appear” — materialized in this window. FlagQuantum shipped v0.3.0rc1 / rc2 / rc3 within 18 hours (10-10 14:42 / 20:38 / 10-11 08:52), completing the PyTorch Stable ABI migration for native CPU extensions and wheel validation across Linux / macOS / Windows; the same window saw Jiuding multi-device job integration and MetaX runtime packaging protection (see 1.2).
  • Vendor integration lands on two tracks (Torch-FL): Tsingmicro’s “FlagGems-only” operator path (two PRs, 10-10) and Kunlunxin Kunlun P800 XPU “CUDA-boxing” backend (10-11 22:09) were merged in succession, with DAMO XuanTie PPU’s FlagTree 0.7.0 pinning landing in the same window; a hardware-agnostic runtime consistency suite followed (#497, 35 files +1947, see 1.6).
  • Release infrastructure surges: build-infra completed directory registration and release automation for all 20 FlagCX v0.14.0 wheels across 8 merges (#1329 / #1331 “wheel release ingestion” hooks), and added a flagos-compressor wheel release pipeline (#1328 / #1330) — component releases shifting from “manual record-keeping” to “pipeline bookkeeping” (see 1.7).

Assessment: This window merged 69 PRs across 14 repos, matching the per-unit-time intensity of the post-holiday surge, and exhibits a distinct “release preparation season” character — ABI, wheel, directory, and version-number engineering actions all clustered in the same window, echoing the external release cadence on the eve of the October 17-18 technology conference; on the competition Track1 side, new submissions continued to flow in beyond the backlog clearing, with no sign of community supply tapering off.

I. Open-Source Project Progress (GitHub Activity)

1.1 Organization Cadence: 19 Repos Pushed, 102 Commits, 69 Merges (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: org: flagos-ai repository list · org repos API (verification baseline)

  • Repository scope: 55 repositories (+1 vs. previous period, with FlagQAI newly added); 19 repos had pushes within the window — docs (10-12 05:36), FlagQuantum (10-12 02:59), vllm-plugin-FL (10-12 00:22), FlagTree (10-11 22:49), Torch-FL (10-11 22:24), FlagCX (10-11 09:56), FlagSparse (10-11 00:02), build-infra (10-10 22:59), community (10-10 20:50), FlagGems (10-10 19:06), FlagAttention (10-10 18:25), FlagGems-Experimental (10-10 16:02), FlagOS-Compressor (10-10 15:50), FlagGems-vllm (10-10 15:19), FlagQAI (10-10 14:17), TransformerEngine-FL (10-10 14:01), Megatron-LM-FL (10-10 13:15), FlagDNN (10-10 11:48), FlagScale (10-10 11:27).
  • Commits and merges: 102 commits found in search (FlagQuantum 47 / FlagGems 19 / build-infra 12 / Torch-FL 6 / FlagCX 3 / FlagTree 3 / FlagGems-Experimental 2 / FlagOS-Compressor 2 / FlagGems-vllm 2 / FlagScale 2 / FlagAttention 1 / FlagDNN 1 / FlagQAI 1 / vllm-plugin-FL 1); all 69 merges postdate the previous period cutoff (14 repos had merges; details in 1.2 ~ 1.8).
  • Under review: 165 PRs were updated within the window (first 100 verified individually: primarily FlagGems 65 / vllm-plugin-FL 7 / FlagGems-Experimental 7 / FlagGems-vllm 5 / docs 3 / FlagTree 3 / FlagGems-sglang 3 / FlagCX 2).
  • Supplementary verification: window pushes for 5 repos were not found in the commit search index and were filled in via atom feeds and branch-level verification — FlagSparse (zyq/update branch csc_coo_opt, 10-11 00:02), Megatron-LM-FL (#184 / #187), TransformerEngine-FL (#129 / #132), community (#126 branch), docs (update/model-list-aihuanxin branch).
  • Trend over the 3 days outside the window: FlagGems-sglang (10-09 13:24) and FlagTrain (10-09 11:03) had pushes the day before the window; this window was dominated by PR review activity.
  • Releases: 3 new Releases within the window (FlagQuantum rc1 / rc2 / rc3); no other new Releases and no other new tags (the FlagCX v0.14.0 tag predates this, from 09-28).

1.2 FlagQuantum: 0.3.0 Triple RC — Stable ABI, Cross-Platform Wheels, and Jiuding Multi-Device Integration (10-10 ~ 10-11)

Date: 2026-10-10 ~ 2026-10-11　Source: v0.3.0rc1 · v0.3.0rc2 · v0.3.0rc3 · #613 · #618 · #620 (full list in appendix)

  • rc1 (10-10 14:42): First candidate of the 0.3 series, packaging user-visible work since 0.2.0 — circuit composition and output API extensions, gradient and vector derivative APIs, OpenQASM import, compiler optimization levels, distributed workflows, digital twin validation, and updated Jiuding / Quafu provider guides; known issue is a qsteed adapter 0.1.1 dependency conflict (not yet uploaded to PyPI, retained as an immutable record).
  • rc2 (10-10 20:38): Corrective release — introduces cp310-abi3 wheels (covering Python 3.10 ~ 3.12), native CPU extensions aligned to the PyTorch 2.10 Stable ABI baseline (runtime verified on 2.13 / 2.14), bundles flagquantum-compiler-qsteed 0.2.0, and incorporates Jiuding and Quafu workflow improvements.
  • rc3 (10-11 08:52): Cross-platform RC — a single cp310-abi3 wheel per supported platform (Linux x86_64 / macOS arm64 / Windows amd64), completing MSVC portability fixes (Stable ABI flag injection, OpenMP SIMD, bit lookup).
  • Engineering track (15 merges): recoverable release validation (#610), rc1 wheel compatibility fix (#611), native extension Stable ABI migration (#613), macOS / Windows wheel build fixes (#615), retention of the Jiuding CPU job queue model (#617), Jiuding multi-device jobs (#618), Jiuding submission quickstart docs (#619), vendor PyTorch runtime packaging guard (#620, merged 10-12 02:59).
  • Positioning: FlagQuantum is the quantum track component of FlagOS; this triple RC is its largest version action since 0.2.0 and the organization’s first new version batch since 09-28; the “Jiuding (cluster management platform) integration” brings quantum workflows into the multi-device job system.

1.3 FlagGems: 19 Merges — Enflame S60 Joins Test Backends, Kunlunxin Optimization Batch, and KernelGen Pipeline (10-10)

Date: 2026-10-10　Source: #6917 · #6937 · #6951 · #5680 and 19 others (see appendix)

  • New backend onboarding (Enflame): #6917 adds the Enflame S60 backend to ops tests (merged 16:17) — following the previous period’s CI migration to flagcicd, Enflame test coverage moves from CI infrastructure into the operator test matrix.
  • Kunlunxin, 6 merges: kunlunxin CI fix (#6937, 11:34); fake_quantize_per_channel_affine family optimized kernels (#6951); hardsigmoid_ and native_channel_shuffle optimizations (#6975); skip FP8 benchmarks on devices without FP8 capability (#6974); cumsum_out / linalg_det / searchsorted_scalar_out optimizations (#6981).
  • KernelGen (Nvidia), 4 merges: matrix_exp (#5680), _linalg_solve_ex (#6664), _pack_padded_sequence (#5256), _scaled_mm_v2 (#6532) — the Triton kernel production line continues to fill in against the PyTorch operator checklist.
  • CI infrastructure, 5 merges: the robust-checkout composite action series (#6952 added, #6965 abort on speed, #6963 precheck wiring, #6970 10-second probe cap, #6977 precheck rounds raised from 30 to 60) — ongoing hardening against checkout hangs in proxy environments.
  • Ascend and misc: KMCompiler index_fill launch grid cap fix (#6946); ops tests switched to Shanghai timezone (#6958); fourth batch of explicit operator invocation tests (#6947); fmod_ scalar broadcast reference mutation fix (#6790).

1.4 FlagGems Under Review: Competition Track1 Continues, New KernelGen Faces, and XPU3 Serial (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: FlagGems-sglang #144 · #6985 · #7003 · #7013 and others (see appendix)

  • Competition Track1 continues (FlagGems-sglang): 3 new competition kernel PRs within the window — unpad_draft_extend_output (#144), hash_topk (#145), pack_topk_ids (#146), authored by competition-channel contributors; new submissions keep flowing in beyond the cleanup.
  • KernelGen (Nvidia) batch of new opens: chx7514 opened 10 convolution and adjoint operators in a single window (slow_conv series, glu_jvp and glu_backward_jvp, mT and mH, etc., #6987 ~ #6996); ShawnsYing filed 8 in a row (geqrf, gradient, grid_sampler, split_copy, select_copy, view_as_real_copy, view_as_complex_copy, _reshape_alias_copy, etc., #6998 ~ #7011); plus _sparse_broadcast_to (#6979), _linalg_det (#5301, existing).
  • Kunlunxin XPU3 serial (llaboon): SDNN v12 — conv1d/conv2d carriers, cudnn bf16, flash_mla and attention backward reference (#7003); conv2d / conv_transpose2d / conv_depthwise2d routed through vendor launch tables (#6914 / #6918 / #6934); addmm device baseline (#6910), cholesky_solve Triton FP32 fast path (#6911), outplace_fused_experts device baseline (#6913); plus a dozen or more existing batches — SDPA routing, flash-attention backward, sparse MLA, scaled_grouped_mm — continuing to update (still pushing at 04:00 on 10-12).
  • Ascend (KMCompiler) fix batch: cholesky_solve bishengir-compile segfault fix (#6948), mul large contiguous input launch grid overflow fix (#6949), index_copy grid cap (#6978); addmm / bmm local auto-blockify performance restoration (#7013).
  • Others: Moore Threads moe_load_balance_loss (#6985) and SDPA forward optimization (#6984); Cambricon philox layout fix (#7009); SiliconFlow argsort timeout and scatter_reduce launch cap (#6940 / #6941); slice complex view (#6953).

1.5 FlagTree: TLE-Raw Backport to 3.5.x, Ascend tle.pipe, and mthreads Guard Fix (10-10 ~ 10-11)

Date: 2026-10-10 ~ 2026-10-11　Source: #1321 · #1275 · #1343 and others (see appendix)

  • 5 merges: TLE-Raw base backport to the 3.5.x maintenance line (#1321, 10-10 19:32); Ascend tle.pipe support (#1275, 19:38); TLE test cases for normalize and fused add-normalize on H100 (#1332); Iluvatar AABS — shrinking dot tile raised back to the 16×16×16 TCU floor (#1339); CI auto-rerun watchdog (set up job failure scenario, #1343).
  • 3 under review: SpacemiT build fix (#1337, modifies setup.py); mthreads Proton header guard in tool registration (#1347, newly opened 10-11); XPU backend sync with internal triton (#1330, updated 10-12 00:02).
  • Observation: The 3.5.x maintenance line runs in parallel with mainline (TLE-Raw backport vs. PPU 3.6 baseline advancement), indicating FlagTree has entered a phase of parallel multi-version-line maintenance.

1.6 Torch-FL: Qingwei / Kunlunxin Onboarding Lands and Runtime Consistency Suite Opens (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: #492 · #495 · #497 and others (see appendix)

  • Qingwei track (2 merges): #492 establishes the “FlagGems-only” operator path — the Qingwei backend is onboarded with the full FlagGems operator supply (no CUDA stack dependency); #493 exempts Qingwei from CUDA-only sparse and SDP linking guards.
  • Kunlunxin onboarding (#495, 10-11 22:09): Kunlun P800 XPU backend onboarded in “CUDA-boxing” form — remapping torch.cuda calls to the XPU device family, the second vendor track after Qingwei.
  • DAMO XuanTie (PPU): #494 pins PPU’s FlagTree to its channel’s current 0.7.0 build (10-11 22:11).
  • Runtime consistency suite (#497 opened): hardware-agnostic runtime contract tests — device / registration / allocator / transfer / streams and events / declared dispatch, wired into all CI configurations; cross-platform testing exposed and fixed 4 real gaps (Enflame GCU stream and event query semantics, non-CUDA current_stream() placeholder, Qingwei stream construction, Ascend matmul’s FlagGems passthrough path), and pinned the CI queue to released FlagTree / FlagCX wheels (35 files, +1947/-91).
  • Others: CUDA compatibility design doc (#487), backend configuration restoration between unit tests (#496).
  • Positioning: In a single window, Torch-FL absorbed the onboarding of two new vendors (Qingwei / Kunlunxin) plus the XuanTie pin, and is becoming the unified entry layer for vendors joining FlagOS.

1.7 FlagCX and build-infra: CRL Artifact Pool / IPC Wrap-Up, Wheel Directory Automation (10-10 ~ 10-11)

Date: 2026-10-10 ~ 2026-10-11　Source: #647 · #649 · build-infra #1331 and others (see appendix)

  • FlagCX, 3 merges: async signal completion validation and fallback for unsupported PUT (#646); CRL adds a pthread worker pool and shared P2P artifacts (#647); stream-based IPC Get with network fallback (#649) — robustness and performance tracks of the communication runtime advance in parallel.
  • FlagCX, 2 under review: Hygon Device API and IR CI tests (#643, updated 10-11); PXN relay transport with 8-GPU CI coverage (#648, newly opened).
  • build-infra, 8 merges (a concentrated day for release infrastructure): #1325 / #1326 complete the release registration and backfill of all 20 FlagCX v0.14.0 wheels; #1327 adds python / os / arch dimensions to the wheel info directory; #1328 / #1330 add and fix the flagos-compressor wheel release pipeline (including a build isolation fix); #1329 / #1331 wire “wheel release” into an automatic ledger hook (release = bookkeeping); #1322 dependency bump.
  • Observation: The wheel directory moves from a “manual table” to “pipeline auto-bookkeeping” — infrastructure catch-up as the 2.2 release cycle wraps up, and groundwork for the next version line (2.3) and releases of new components such as the compressor.

1.8 Remaining Tracks: FlagOS-Compressor / FlagGems-vllm / vllm-plugin-FL / Training Side and docs (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: FlagOS-Compressor #12 · FlagGems-vllm #907 · vllm-plugin-FL #546 · Megatron-LM-FL #200 (see appendix)

  • FlagOS-Compressor: #12 fixes the PyPI dependency lower bound (merged 15:50, synchronized with the wheel pipeline, see 1.7).
  • FlagGems-vllm: #857 MetaX block-scaled INT8 einsum merged (10:36); ops tests Shanghai timezone (#905); under review, new KMCompiler optimizations (#906), MetaX GDN fused kernel (#907), etc.
  • vllm-plugin-FL: #546 Kunlunxin supports FlagTree (merged 10-11 16:22); active under review — MetaX vLLM 0.28 (#575, updated 10-12 02:53), continuous pushes on the MUSA CI branch (codex/musa-ci-20261011), Iluvatar attention cache fix (#579), Xiwang PTPU graph mode (#569), H100 adaptations for Qwen3.8 / HY4 / GLM5.3 (#455 / #453 / #454), etc.
  • Training side: Megatron-LM-FL #184 (PPU CI) and #187 (unit test group reusing dependency wheels) merged, #200 Ascend GDN coverage newly opened; TransformerEngine-FL #129 (PPU unit and integration tests) and #132 (backend test runner standardization) merged; FlagScale #1268 (checkpoint-period heartbeat false-positive fix) and #1271 (diagnostic stop lifecycle) merged; FlagTrain #8 (KMCompiler CI/CD Dockerfile) newly opened.
  • Docs and misc: docs #542 (model list update, updated 10-12 05:38), #547 / #548 (ModelScope doc bot updates); FlagAttention #91 operator API and test standardization refactor merged; FlagDNN documentation commit.

1.9 Governance and Repository Scope: FlagOS 2.3 Cycle Opens and New Repository FlagQAI (10-10)

Date: 2026-10-10 ~ 2026-10-12　Source: community #126 · FlagQAI (see appendix)

  • FlagOS 2.3 cycle opens (community #126, under review): adds release/2.3/schedule.md (bilingual) — FEP merge deadline 11-13 (hard gate), feature freeze + rc0 11-30, rc1 12-07, rc2 12-14, test end / code freeze 12-25, GA 12-30; candidates cut on Mondays.
  • Four process improvements from the 2.2 retrospective: ① the FEP merge deadline is upgraded to a hard gate (it was advisory in 2.2, and 16 FEPs merged only within the rc window); ② a tag gate is set at feature freeze (tag = manifest = package metadata; the RM smoke-verifies the rc0 manifest the same week); ③ Go/No-Go is made explicit (written report 48 hours before GA, TSC vote; release-scope changes only at this gate); ④ the graduation process is bound to tracking issues (executable test plan before rc1, item-by-item acceptance conclusions recorded at GA).
  • New repository FlagQAI (the 55th): created 10-10, containing only an initial LICENSE commit (no description, no README) — placeholder / incubation state; naming and purpose to be observed.

II. News Coverage and Ecosystem

2.1 Ecosystem Inclusion: MetaX University Open Source Tour Shanghai Stop to Visit Three Universities Next Week (10-10)

Date: 2026-10-10　Source: Guandian: MetaX University Open Source Tour Shanghai Stop to Visit Three Universities in Mid-October (10-10)

According to MetaX, the MetaX University Open Source Tour Shanghai Stop will visit three universities in mid-October: East China Normal University (10-13), Donghua University (10-14), and Shanghai University (10-16). The event theme is “Make Open Source Your First Resume,” covering the MetaX open source ecosystem, the Youth Open Source Special Fund Seed Program, and model intelligent routing practices based on MetaX GPUs; the MetaX developer ecosystem has already invested over 100,000 GPU card-hours serving developers, and plans to support more than 80 open source projects. This event aligns with the talent-side operations of the FlagOS ecosystem—MetaX is a FlagOS member, and its open source ecosystem investment (funds, compute, project support) provides developer-side supply for the multi-chip software stack.

2.2 Supplementary Inclusion: Biren Technology’s HK$4 Billion Placement Completed—70% Directed to Supply Chain and Production Capacity (10-08 / 10-11)

Date: 2026-10-08 (announcement) / 2026-10-11 (financial media re-coverage)　Source: Securities Times: Biren Technology Plans to Raise Over HK$4 Billion via Share Placement (10-08) · Chaoqi Network: Placing Shares While Falling, “Cash Incinerator” Biren Technology Prepares to Raise Another HK$4 Billion (10-11)

  • Matter: On 10-08, Biren Technology announced the placement of 130 million new H shares (at HK$31.08 per share, a discount of approximately 9.76%), raising a total of HK$4.04 billion; approximately 70% is for supply chain procurement, production preparation, and commercialization of next-generation products, approximately 20% for R&D capabilities and the software ecosystem (developer community, toolchain, compatibility with mainstream AI models/frameworks), and approximately 10% for working capital. The next-generation flagship product has entered post-tape-out verification, with customer sampling and mass production expected within the coming quarters.
  • Connection to FlagOS: Biren is the newest manufacturer among the domestic GPU “Four Little Dragons” to join the FlagOS adaptation line—Megatron-LM-FL #199 “Biren SUPA Platform Support” was under review in the previous window, and its software ecosystem investment direction (framework compatibility, developer community) directly echoes FlagOS’s integration roadmap; this matter occurred on 10-08 and was not covered in the previous issue, so it is handled as a supplementary inclusion in this issue.

2.3 Search and Official Channel Verification: Component Line Remains Quiet (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: Google News RSS (24 sets of Chinese and English query terms, via proxy) · HN Algolia · FlagOS CSDN Column · BAAI Community

  • Google News RSS: 24 query sets yielded 13 in-window hits, with 1 inclusion (MetaX University Open Source Tour) and 1 supplementary inclusion (Biren placement re-coverage); the remaining 11 were noise (robotics company reports, storage vendors, stock market updates, and BAAI Community miscellany)—component name queries (FlagOS / FlagGems / FlagScale / FlagTree / FlagCX / KernelGen, etc.) continued a quiet window with zero hits.
  • HN Algolia: 7 query sets with no direct hits.
  • Official Channels: No new articles in-window on the CSDN column (the latest homepage entries remain previous competition and livestream content); no FlagOS-related entries on the BAAI Community homepage. The “2026 AI Open Computing Conference & Zhongzhi FlagOS Technology Conference” on October 17-18 was previously covered consecutively (09-18 registration, 09-28 agenda update), with no new agenda developments in this window.

III. Member Deep Dives

3.1 MetaX: Runtime Packaging Protection, GDN Fused Kernel, and University Open Source Tour (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: FlagQuantum #620 · FlagGems-vllm #907 · vllm-plugin-FL #554

  • Runtime packaging protection (FlagQuantum #620, merged 10-12 02:59): Fixed the issue of preserving compatible vendor PyTorch runtimes during packaging (the fix/metax-torch-runtime branch was closed out via #620) — MetaX runtime enters the quantum line’s cross-platform packaging compatibility matrix.
  • Operator line: FlagGems-vllm #857 (block-scaled INT8 einsum) merged (10:36); #907 adds a backend-specific GDN (Gated DeltaNet) fused_recurrent fused kernel (newly opened 10-10 18:57).
  • Inference adaptation: vllm-plugin-FL #554 (MetaX C550 adaptation for vLLM 0.28) updated 10-10, still under review.
  • Ecosystem operations: University Open Source Tour Shanghai stop (three universities on 10-13 / 10-14 / 10-16, see 2.1).

3.2 Kunlunxin: Torch-FL Integration, vLLM FlagTree Support, and XPU3 Batch Progress (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: Torch-FL #495 · vllm-plugin-FL #546 · FlagGems #7003 et al. (see appendix)

  • Framework-layer integration: Torch-FL #495 (Kunlun P800 XPU CUDA-boxing backend, merged 10-11 22:09); vllm-plugin-FL #546 (Kunlunxin support for FlagTree, merged 10-11 16:22) — Kunlunxin lands moves at both the framework and inference plugin layers in the same period.
  • FlagGems merged batch of 6: CI fixes (#6937), fake_quantize family (#6951), hardsigmoid_ and channel_shuffle (#6975), FP8 benchmark skip (#6974), cumsum_out and two other operator optimizations (#6981) — a mid-October “fix + optimize” dual-track closeout.
  • XPU3 batch under review (SDNN line): #7003 SDNN v12 (conv1d/conv2d carriers, cudnn bf16, flash_mla / attention backward reference) newly opened; the conv-family vendor startup table (#6914 / #6918 / #6934), addmm / cholesky_solve / fused_experts device baselines (#6910 / #6911 / #6913), and a dozen-plus other PRs continued updating into the early hours of 10-12.
  • Quantum line: FlagQuantum #617 preserves the Jiuding CPU job queue model (Kunlun capability fix branch).

3.3 Tsingmicro: Torch-FL “FlagGems-only” Operator Path Lands (10-10)

Date: 2026-10-10　Source: Torch-FL #492 · Torch-FL #493

  • Path established (#492, merged 17:57): Established a “FlagGems-only” operator path for the Tsingmicro backend — breaking free from dependence on CUDA-ecosystem operators, with FlagGems supplying the full operator set.
  • Guard exemption (#493, merged 23:10): Excluded Tsingmicro from the CUDA-only sparse and SDP (scaled dot-product) link guards, ensuring its sparse and attention paths are not blocked by CUDA dependency chains.
  • Background linkage: In the previous window, the Tsingmicro TX8110 backend entered the FlagGems full operator test CI (self-hosted runner); in this window, Torch-FL integration landed — Tsingmicro’s integration has expanded from “test regression” to “framework entry point.” The path (FlagGems-only) differs from the prior “CUDA-compatible” strategy, and its inference stack performance warrants further observation.

3.4 DAMO XuanTie / Ascend: PPU CI Lands in Two Repos, Ascend Fix Batch (10-10 ~ 10-11)

Date: 2026-10-10 ~ 2026-10-11　Source: Megatron-LM-FL #184 · TransformerEngine-FL #129 · Torch-FL #494 · FlagGems #6946

  • DAMO XuanTie (PPU): Megatron-LM-FL #184 (PPU CI support) and TransformerEngine-FL #129 (PPU unit and integration tests) merged in the same window — PPU enters the CI matrix of two training-stack repos; Torch-FL #494 pins PPU’s FlagTree to the 0.7.0 build.
  • Ascend: FlagGems KMCompiler fix batch — index_fill grid capping merged (#6946); cholesky_solve segfault (#6948), mul grid overflow (#6949), index_copy (#6978), addmm / bmm auto-blockify (#7013) under review; Megatron-LM-FL #200 adds Ascend GDN coverage (including numerical tests); FlagTree #1275 Ascend tle.pipe merged; FlagTrain #8 KMCompiler CI/CD Dockerfile newly opened.

3.5 Moore Threads / Enflame / Iluvatar: MUSA 0.28 and S60 Backends, conv and Attention Fixes (10-10 ~ 10-12)

Date: 2026-10-10 ~ 2026-10-12　Source: FlagGems #6985 · FlagGems #6917 · FlagTree #1339 et al. (see appendix)

  • Moore Threads: FlagGems newly opens the moe_load_balance_loss operator (#6985) and SDPA forward optimization (V scheduling and accumulator fusion, #6984); FlagTree #1347 Proton header guard fix (newly opened 10-11); vllm-plugin-FL #575 (MUSA adaptation for vLLM 0.28) updated 10-12 02:53, with the MUSA CI branch pushed continuously — the most active participant in the 0.28 adaptation wave.
  • Enflame: FlagGems #6917 Enflame S60 enters the ops test backend (merged 16:17) — a test-coverage move following the previous period’s CI migration to flagcicd; Torch-FL #497’s consistency suite fixed GCU stream / event query semantic gaps.
  • Iluvatar: FlagTree #1339 AABS shrink dot tile fix merged (16×16×16 TCU lower bound); vllm-plugin-FL #579 fixes multi-attention cache binding (newly opened 10-10); FlagGems #6298 conv2d optimization continues under review (updated 10-10 18:44).

3.6 Other Members: Hygon / Cambricon / Biren / Horizon (10-10 ~ 10-11)

Date: 2026-10-10 ~ 2026-10-11　Source: FlagCX #643 · FlagGems #7009

  • Hygon: FlagCX #643 (Hygon Device API and IR CI tests) updated 10-11 09:30, under review.
  • Cambricon: FlagGems #7009 philox layout fix (randn compilation restored, newly opened 10-11).
  • Biren: HK$4 billion placement completed (see 2.2); Megatron-LM-FL #199 SUPA support was an item under review in the previous period, with no new updates in this window.
  • Horizon: No new public activity within the window.

IV. Summary and Trend Observations

  1. Weekend high-level continuity, with both intensity and breadth: 69 merges in a 44-hour window, covering 14 repos, with 19 repos having pushes (including all of Sunday and early Monday); compared with the previous period (about 28 hours, 45 merges), the per-unit-time intensity held steady, and merges continued from late Sunday night into early Monday (FlagQuantum #620 at 02:59, Torch-FL #497 updated at 04:54).
  2. Distinct “release preparation season” characteristics: FlagQuantum’s three consecutive RCs (ABI migration, cross-three-platform wheels), FlagCX’s full registration of 20 wheels, the build-infra wheel auto-accounting hook, and the flagos-compressor pipeline — four release-engineering actions appearing in the same window, marking this cycle’s first “version infrastructure concentration day,” and echoing the external release cadence ahead of the October 17-18 technical conference.
  3. Torch-FL becomes the primary entry point for vendor integration: A single window lands two integrations — Tsingmicro (FlagGems-only path) and Kunlunxin (CUDA-boxing) — plus XuanTie PPU pinning and a hardware-agnostic consistency suite. The pattern of “vendor integration = one new backend + one CI contract + one set of packaging protections” is taking shape.
  4. No decline in competition and community supply: Track1 added 3 new competition kernels (#144 ~ #146) beyond clearing the backlog; new faces in KernelGen (chx7514, ShawnsYing) continue supplying at a pace of 8 ~ 10 per week — community-side long-tail contributions toward “full operator coverage” are still accelerating.
  5. Governance enters the 2.3 cycle: The 2.3 schedule is announced (FEP hard gate 11-13, GA 12-30) and four process improvements land, completing the institutionalization actions from the 2.2 retrospective — from here the organization runs dual-track: “2.2 release wrap-up + 2.3 cycle kickoff.”
  6. Next observation points: FlagQuantum 0.3.0 official release (the closeout cadence after three consecutive RCs); the first merge landing point of the vLLM 0.28 multi-vendor adaptation batch (MetaX / Kunlunxin / Hygon / Ascend / XuanTie five lines still under review); FEP and milestone progress in the first week of the 2.3 cycle; and 2.2 external release materials and workshop preparation ahead of the conference (10-17 / 18).

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full verification of pushed_at across 55 repos 19 repos pushed within window; 1 new repo added (FlagQAI)
GitHub Commit search (committer-date basis) Full verification of 102 commits (incl. pagination) FlagQuantum 47 / FlagGems 19 / build-infra 12 / Torch-FL 6 / FlagCX 3 / FlagTree 3 / others 12
GitHub Merge search (is:merged) Item-by-item verification of 69 merges FlagGems 19 / FlagQuantum 15 / build-infra 8 / Torch-FL 6 / FlagTree 5 / FlagCX 3 / others 13; all later than previous cutoff
GitHub Open search (is:open) 165 updates; item-by-item verification of top 100 Mainly FlagGems 65 / vllm-plugin-FL 7 / FlagGems-Experimental 7 / FlagGems-vllm 5 / docs 3 / FlagTree 3 / FlagGems-sglang 3
GitHub Branch / atom feed supplementary verification 5 repos (FlagSparse / Megatron-LM-FL / TransformerEngine-FL / community / docs) Filled in window commits and branch pushes not covered by search index
Releases Tag / Release atomic verification 26 repos’ releases + 7 repos’ tags 3 new Releases (FlagQuantum v0.3.0rc1 ~ rc3); no other new tags
News Google News RSS (24 sets of Chinese/English query terms, via proxy) 44-hour window filter + item-by-item review 13 hits: 1 included (MetaX university open-source tour), 1 supplemented (Biren placement), remaining 11 excluded
News Tavily supplementary search Re-verification of ecosystem and vendor keywords No new in-window developments for conference agenda / Xiwang / Qingwei etc.; Biren placement details verified via original site (Securities Times)
Community HN Algolia Re-verification of 7 query sets No direct hits
Official channels CSDN column / BAAI community Page-scrape re-verification No new in-window articles on CSDN; no FlagOS entries on BAAI community homepage
Historical comparison 10-10 daily report (previous issue) Deduplication verification Previous issue covered through 10-10 09:40; all items in this issue are later than that cutoff (Biren placement is a 10-08 event, explicitly supplemented)

Complete Source List