Reporting period: October 19, 2026 (Monday) to October 23 (Friday), 5 working days in total.

Sources: 5 issues of this publication’s daily FlagOS update reports (10-19 / 10-20 / 10-21 / 10-22 / 10-23, with no missing issues). Among them, the 10-19 issue is the Monday morning report, whose research window starts from the previous issue’s cutoff at 10-16 06:00 (approximately 72 hours, covering the entire “2026 AI Open Computing Conference & Zhongzhi FlagOS Technology Conference” held from 10-17 to 10-18 and the Liangzhiyun launch); the four issues from 10-20 to 10-23 each cover a 24-hour window, running back-to-back, with the final issue’s last shift covering 10-22 06:00 to 10-23 06:00. Therefore, the factual range of this report is 10-16 to 10-22, with the 10-16 to 10-18 portion labeled according to the original dates.

Compilation notes: Covers five dimensions: versions and releases, component and code co-building, member units and vendor adaptation, ecosystem and community, and trend observations. The five daily reports, accumulated by their respective windows, total 654 commits and 661 merges; if only the four complete 24-hour windows are counted (the merged days from 10-19 to 10-22), the totals are 507 commits and 522 merges. The number of repositories with pushes in each issue is 16, 16, 19, 22, and 16 respectively, out of a total of 61 organization repositories. Versions and tags were independently verified against each component repository’s atomic source (releases.atom), consistent with the daily reports’ methodology.

I. Weekly Highlights

  • The conference wrapped up over two days, and the scale of 2.2 was for the first time broken into four verifiable layers (10-17 ~ 10-18): In the main forum, Lin Yonghua, Vice President and Chief Engineer of the Beijing Academy of Artificial Intelligence (BAAI), presented under the title “Zhongzhi FlagOS 2.2: From the Token Economy to Physical AI” — 21 chip vendors and 35 chips, 9 major operator libraries totaling 1,454 operators (an increase of 890 over the previous version), cross-chip operator performance up 27.6% (covering 10 chips), and 264 code contributors (up 55% quarter-over-quarter). On the community side, there are also more than 730 active contributors, more than 100 member organizations, coverage of more than 30 countries, and more than 55,800 developers; merged PRs from 2.1 to 2.2 rose from 1,890 to 4,890, new issues from 420 to 1,423, and commits from 5,100 to 9,730. On the same day, the cross-chip debugging and performance analysis tool FlagPrism and the unified quantization tool FlagOS-Compressor were released. Reports state that 2.2 was released on September 30, whereas the repository manifest was actually committed between 09-28 and 09-29; the two-day discrepancy is handled in the body text according to the repository timeline.

  • Quantum-AI Cloud was publicly launched, extending the FlagOS execution stack to hybrid quantum-classical scheduling (10-18): The Quantum-AI Fusion sub-forum included a “Quantum-AI Cloud Platform Launch” session. Quantum-AI Cloud was jointly launched by BAAI and the Beijing Academy of Quantum Information Sciences, supporting FlagQuantum through the FlagOS Execution Stack and interfacing with the “Jiuding” AI computing system and the “QUAFU” quantum computer to enable hybrid quantum-classical job scheduling. The code side corroborates this: the release notes for FlagQuantum v0.3.0 already state that Jiuding and Quafu tasks are resumable and that Quafu multi-circuit results preserve order, indicating that at release the task submission, receipt, and result-ordering pipeline was already inside the package rather than merely a framework interface. The boundaries are equally clear — it does not run on quantum hardware, does not include an MCP server or agent runtime, and hardware execution is handled by the Jiuding and QUAFU sides.

  • The 2.3 cycle saw its first same-day loop from “requirement to evidence” (10-21 ~ 10-22): On 10-21, two FEPs were merged — FEP-0140 defines three version lines for vllm-plugin-FL (0.24 completing five vendors, 0.28 as the November main release supporting ten vendors, and 0.32 introducing five vendors), while FEP-0141 locks in the ten-vendor matrix for SGLang 0.5.18 and the DeepSeek-V4 W8A8 target; on 10-22, Hygon completed alignment with vLLM 0.28.0 and SGLang 0.5.18 on the same day in both the vllm-plugin-FL and sglang-plugin-FL repositories (84 files in total, a net increase of about 5,700 lines). From “writing the matrix into the document” to “the first measured pass” took only one day — the fastest requirement-to-evidence loop this quarter.

  • The delivery side closed out six backends to seven within the same week and turned the communication library from a special case into the general rule (10-16 ~ 10-22): build-infra first landed one self-contained 22.04 builder each for FlagCX’s six backends — nvidia, hygon, cambricon, metax, enflame, and iluvatar — lowering the glibc floor from 2.38 to 2.35; on 10-19 it migrated the four Ascend wheels from the 24.04 runtime into the same form and added Tsingmicro and Xiwang; on 10-21 it onboarded Tsingmicro and Moore Threads as FlagTree’s fifth and sixth backends; on 10-22 Kunlunxin became the seventh backend. On the same day, flagcx==0.14.0 was extended from being pinned only on the two nvidia lines to all twenty backends, and a one-click install script was added to condense multi-repository dependencies into a single command.

  • Four infrastructure items were completed on the same day, shifting organizational capacity toward “how to test, how to install, how to share” (10-20): optest received its first substantive commit since the repository was created (PR #1, 38 files), containing 12 backend configurations and hardware probes for 13 vendors; FlagOps elevated FlagGems’ five CI actions into organization-level shared resources; build-infra added a one-click install script (1 file, +308); Torch-FL and TransformerEngine-FL introduced the unified FlagCICD entry point on the same day with support for L0 to L2 tiers, and FlagScale added selective platform CI. Testing, actions, installation, and the CI entry point — all four lines each completed a structural merge within the same window.

  • The quantum stack advanced from design contract to runnable path (10-21 ~ 10-22): On 10-21, FlagQuantum saw 84 commits and 34 merges, downgrading dynamic circuits to “successor qubit plus Boolean SSA results” and landing this in the Program IR, while making exact downgrade and restoration a mandatory prerequisite for deployment; on 10-22, a further 53 commits and 20 merges turned the M3 neutral atom from two design documents into an executable path — a native Rydberg frontend on the Core side (PR #700), plan 1.7 dynamics on the Simulation side (#707), and moving-atom pulse training on the Examples side (#714, with unconditional RR rising from 0.0431 to 0.1228 in a 40-step, 128-interval measured run).

  • Merge volume hit a new high during the week, but the bulk of the increment was test contracts rather than operators (10-22): The day’s 163 merges were the highest in the past week, of which FlagGems alone accounted for 91, or 56% of the whole organization; at the same time, commit volume fell 18% and pushing repositories decreased 27%, indicating digestion of a backlog rather than expansion. The testing mainline was carried by lukatao alone with 61 commits (67% of that repository), opening 10 batch PRs within seven minutes from 23:40 to 23:47, changing test cases from wrapper functions to directly calling operators or injecting candidates. For the full week, KernelGen’s new operators were concentrated in three batches of merges plus the tail end of the sparse family, with genuinely new capability accounting for less than 20%.

  • Three information increments from outside the repositories appeared on the ecosystem side (10-19 ~ 10-22): China.com published the first full-stack technical breakdown after the conference (adding seven quantization cases and grouped evidence); The Beijing News and the Beijing Daily app carried out a second wave of industry-version dissemination, adding a citable figure of “a new model completing adaptation for 12 chips within one day”; Pandaily systematically relayed 2.2 in English, the first complete article aimed at overseas readers. In addition, the Beijing Municipal Science & Technology Commission and the Zhongguancun Administrative Committee published a pre-approval notice for the 2026 Zhongzhi FlagOS Acceleration Program, in which Kuangwei Zhixin’s “On-device Deployment and Application Demonstration of the AlohaMini Embodied Intelligent Robot Based on FlagOS-Robo” was selected — the first external project with deliverable constraints to appear in the embodied intelligence direction.

II. Versions and Release Activity

This Week’s Tags and Release List (Beijing Time)

Time Component Tag / Version Notes
10-18 10:36 FlagCX v0.15.0.dev0 Development-line tag; same window merged Hygon UIL and per-adaptor bitcode
10-19 13:01 FlagTree 0.7.1a1+ascend3.5 Ascend 3.5 build line, first alpha in the 0.7.1 series
10-20 20:11 FlagTree 0.7.1.dev1 Base line moved from a1 to dev1, nightly build artifact
10-20 20:12 FlagTree 0.7.1.dev1+ppu3.6 PPU 3.6 build line, one minute after the previous entry
  • Four tags, three components, none a formal release: all tags in this week’s window belong to development lines or hardware build lines. Apart from the three components above, FlagGems, FlagBLAS, FlagDNN, Torch-FL, vllm-plugin-FL, sglang-plugin-FL, FlagOps, FlagScale, FlagAttention, FlagSparse and other repositories had no new Releases in the window. This is the typical gap between 2.2’s release on September 30 and 2.3’s GA on December 30—public versions are silent while the artifact line stays active.
  • Carried over from the previous issue: FlagQuantum v0.3.0 (10-16 13:34) and v0.3.1 (10-16 19:20) were already listed in the previous weekly’s tag table; this issue treats them as the baseline versions of the quantum stack and does not count them again.

FlagTree: Release Cadence Sliced by Hardware Version

  • Four build lines coexist: the 0.7.1 line now has four hardware build branches—ascend3.5, iluvatar3.6, xpu3.6, ppu3.6—with the base line moving from a1 to dev1. The release notes only say “add version number to pypi,” indicating automated nightly builds rather than feature releases; whether a release has occurred must be judged from the build line, not from version-number semantics.
  • The backend matrix expanded to seven rows this week: SpacemiT TX8 (10-21, fifth), Moore Threads MUSA (10-21, sixth), and Kunlunxin XPU (10-22, seventh) were successively integrated into the FlagTree build line and daily build rows. The Kunlunxin row has five unique constraints: ten precompiled inputs; caches matched by directory name and LLVM 19 packages sharing two tarball roots; frontend selected by environment variable; pybind11 pinned exactly to 3.0.1 (3.0.4 causes the dispatcher to SIGSEGV); and no Release build flags (the vendor wheel itself is rel-with-asserts).
  • Compiler-side feature increments are few and concentrated: AMD RDNA advanced on 10-21 from single-cooperative large-core decode (M=1) to prefill (M>1, PR #1365, 66 files); the RDNA prefill WMMA GEMM microkernel is 10 to 30 times faster than the GEMV loop when M>1, and on the decode side split-K cut lm_head from 17.0 to 14.9 ms/token and per-layer wide projections from 14.9 to 10.5 ms/token; MUSA added per-copy TME cache persistence (#1385); the SpaceMiT RISC-V backend added the smt_rvisa bare-metal eDSL and retired the private TLE dialect (#1381, 10-16, within the Monday window).

FlagGems: 5.4 Is the Formal Line, No New Additions on the 5.5 Development Line

  • Formal-line status unchanged: 5.4 remains the current formal line (released 09-28); 5.5.0.dev0 opened on 10-12 and had no new tags this week, with package version metadata still showing 0.1.0—the Hygon vLLM 0.28 acceptance record gathered evidence by “comparing all 1,944 installed source files one by one against commit e12a220, metadata version untrustworthy,” and this practice of substituting content for VCS metadata appeared again this week.
  • Operator coverage enters the wrap-up phase: this week KernelGen merged 14 Nvidia operators in three batches, all boundary operators for aliasing, views, and sparse semantics. The 10:43 batch covered copy, sparse construction, and nested-tensor table construction (ger, sparse_bsc_tensor); the 15:23 batch covered value and shape views, plus nested-tensor storage offsets and csr-to-coo sparse index conversion (_values_copy, _shape_as_tensor); the 22:23 batch covered dimension reordering and complex views (swapdims, moveaxis, view_as_complex_copy). On the review side, roughly 20 sparse-family PRs (_sparse_mm series, copy_sparse_to_sparse_, coo and csr index conversion, etc.) were opened in one go on the night of 10-19.
  • The leading signal for a version release is in build-infra: the vllm-plugin-native wheel added nvidia, metax, kunlunxin, and hygon backends on 10-21 and registered a release; on 10-22 it uniformly pinned flagcx==0.14.0 for all backends—all artifact preparation ahead of the 2.3 formal release.

III. Components and Collaborative Code Development

Organizational Cadence: Five Periods Accumulate 654 Commits and 661 Merges

Period Window (Beijing Time) Repos Pushed Commits Merges In Review (Updated Within Window)
10-19 10-16 06:00 ~ 10-19 06:00 (72 hours) 16 147 139 239
10-20 10-19 06:00 ~ 10-20 06:00 16 93 88 180
10-21 10-20 06:00 ~ 10-21 06:00 19 115 120 195
10-22 10-21 06:00 ~ 10-22 06:00 22 164 151 154
10-23 10-22 06:00 ~ 10-23 06:00 16 135 163 183
Total — — 654 661 —
  • Merges by Repo (Cumulative, per Daily Report Disclosure): FlagGems 290 (43.9%), build-infra 96 (14.5%), FlagQuantum 82 (12.4%), Torch-FL 53 (8.0%), FlagTree 25, FlagGems-vllm 21, vllm-plugin-FL 18, FlagGems-sglang 15, FlagBLAS 6, FlagCX 5, optest 5, FlagTrain 5, FlagGems-Experimental 3, sglang-plugin-FL 3, FlagOps 2, FlagAttention 2, community 2, FlagOS-Compressor 2, FlagSparse and TransformerEngine-FL 1 each; plus FlagCV, Trident, .github, and flir 1 each. Some periods only counted the first 100 merged PRs by repo, yielding an attributable total of 641, with the 20-commit difference being unexpanded tail repos.
  • Dual-Engine Structure Unchanged: FlagGems and FlagQuantum together account for 56.3% of merges; adding build-infra and Torch-FL brings the top four repos to 79%. The weekly pattern is “infrastructure and quantum stack rise Monday through Wednesday, FlagGems clears its backlog Thursday through Friday.”
  • Individual Output Highly Concentrated: lukatao totaled 119 commits in FlagGems by merge date (23 on 10-20, 35 on 10-21, 61 on 10-22), all test and reference-implementation fixes; tengqm handled 35 commits in build-infra (7 on 10-20, 15 on 10-21, 13 on 10-22); lvyufeng covered Torch-FL’s MPS and platform side; all FlagQuantum merges were completed by the same account.

FlagGems (290 Merges): Test Contracts Become the Primary Theme

  • Three Rounds of Revert and Rebuild: On 10-20, PR #7361 first reverted the earlier shared test-infrastructure changes (17 files +599/−1040), pushing compatibility handling down into each operator’s own tests; 22 subsequent fixes all followed this direction—input snapshots in CPU reference mode, tensor view geometry, complex and half-precision value comparison, sparse construction and conversion, generated-case executability, assertion boundaries, pytest marker normalization. This was an engineering correction that “first acknowledged the shared approach was unworkable, then filled in operator by operator.”
  • lukatao’s Three-Part Test Case Overhaul: On 10-21, 35 commits converted generated cases to run cross-backend (split and squeeze, adjoint and mH and mT, as_strided, moveaxis and reshape_as and swapaxes, permutation views, integer and half-precision conversion), with #7425 through #7440 being sixteen consecutive commits; on 10-22, another 61 commits converted foreach overrides to “direct routing” batches 01 through 09 and moved the last nine operator suites to the common candidate API—making “which implementation is tested” explicit rather than implicit.
  • Backend Skip Governance Begins Documenting Reasons: Ascend and Moore Threads skips were individually scoped by input domain (linalg_vander complex input, logdet complex, row_stack_out memory format, trapz complex32, orgqr MUSA zero columns, the qint32 series, Hygon cuDNN validation), and four sparse suite ports were merged consecutively in the early hours of 10-23 (sparse log-softmax cross-backend portability, sparse validator firmware and direct benchmark candidate invocation, compressed sparse validator, common sparse block constructor).
  • Two Large Vendor-Side Commits: Enflame’s bulk update #7077 (26 files +3520/−511, titled only “update from enflame 20261013,” with no itemized details); Kunlunxin’s fused MoE #7107 (13 files), adding three fused_experts variants and dispatch_fused_moe_kernel, with aggregate score 1.324 to 1.329, and proactively declaring the baseline as an alternative implementation and dispatch_fused_moe_kernel’s 0.626 as substandard.
  • Kunlunxin Optimization Line: 9 commits on 10-20, 17 on 10-21, 13 on 10-22, shifting from adding implementations to gating by precision and shape—w8a8_block_fp8_matmul gains FP8 capability gating, fp64 and grouped-MM switch to capability-based skips with documented reasons, single-layer GRU cells switch to a flat 1D kernel raising the aggregate score from 0.4217 to 0.8253, and sgn’s large-numel path switches to a cheap main body plus configurable 1D tiling.

FlagQuantum (82 Merges): From Contract to Code, Then From Code to Rejection

  • Dynamic Circuits Enter the IR (10-21): PR #681 defines a typed dynamic Program IR contract; #683 lowers mid-circuit measurement to “successor qubit plus Boolean SSA result,” reset to an explicit successor qubit, and conjunction/disjunction to canonical DNF over active Boolean values, restoring only when the lowered result matches the full function graph verbatim; #694 makes exact lowering and restoration a mandatory prerequisite for dynamic deployment, halting unverified programs before calling any provider.
  • M3 Neutral Atoms From Contract to Runnable (10-22): On the Core side, #700 exposes a frontend for native Rydberg control directly on ProgramRecord (RydbergProgram provides three actions—move, pulse, to_ir—and explicitly states “a controlled position does not represent physical presence”); on the Simulation side, #707 routes native Rydberg programs through the existing Lindblad plan, with plan 1.7 sealed to exactly seven outer fields; on the Examples side, #714 trains two pulse amplitudes and moves atom pairs, with 40 steps over 128 intervals measuring unconditional RR rising from 0.0431 to 0.1228, survival-conditioned RR 0.1500, and no-flag selection RR 0.1439.
  • Half of Merges Spent on Rejection and Deletion: On 10-22, #701 makes native physics programs uniformly throw CapabilityError instead of AttributeError at packaging time; #705, per Proposal 082, directly deletes NeutralAtomProvider and two Remote exports that never had a corresponding external service, leaving no aliases and no empty shells; #716 proposes Proposal 083 defining the Runtime’s sample_terminal primitive and prints “proposed, not implemented” at the end of the example. Such actions add no capability but determine whether M3’s boundaries are credible.
  • Plotting and Citability Chain Completed: On 10-21, #668 adds a lazy LaTeX backend and an “encoder—circuit—measurement head” hybrid model view (43 files +14311), with Notebooks preferring SVG display; #684 adds CITATION.cff (version 0.3.1, release date 2026-10-16, Apache-2.0); #687 inventories reusable neutral-atom capabilities and missing contracts, providing independent numerical references of physical survival 0.8187, unconditional RR 0.4094, and survival-conditioned RR 0.5.
  • Compute Expansion: On 10-20, #653 and #661 migrate the CPU suite to eight A800 runners with automatic GPU test triggering, and long tasks switch to parallel—the advancement method from here on is large-scale consistency validation rather than feature rollout.

build-infra (96 Merges): Compressing “N Backends Times M Artifact Types” Into One Line

  • Builder Matrix Takes Shape: The base/ directory expands this week from six to nine vendors and seventeen baseline images—nvidia (cuda12.8 / 13.3), ascend (cann8.5.0 / 9.0.0, including 910c), cambricon (neuware4.4.3 / 4.7.2), enflame (tops1.9.10 / 1.10.6), hygon (dtk26.04), iluvatar (corex4.4.0 / 4.5.0), metax (maca3.7.2.1 / 3.8.1.3), sunrise (tangrt1.2.0), tsingmicro (tsm260610).
  • FlagCX’s 22.04 Consolidation: From 10-16 to 10-18, each of six backends gets a self-contained builder, each using ubuntu:22.04 as the glibc floor, pre-installing vendor SDKs into the image, and leaving a trace in wheel-info; #1407 adds a dispatch-level --no-bitcode switch to avoid direct errors on non-nvidia adaptors when the v0.14.0 tag predates bindings/ir.
  • FlagTree Build Line: The builder merges from “two workflows doing one thing” into a single matrix, with one line in backends.yaml carrying the Containerfile, image repo, runner label, disk budget, and acceptance assertions. Tsingmicro, Moore Threads, and Kunlunxin rows land on the same template, differing only in prebuilt tree and toolchain resolution paths.
  • Installation Path Convergence: On 10-20, #1425 adds a one-click install script (1 file +308), converging for the first time the installation of “multiple repos, multiple dependency sets” into a single command; #1435 restores the vllm native extension wheel for the CUDA distribution backend and points out the pitfall that PEP 440 ignores build tags and short pins install the wrong backend.
  • Build Docs Become a Site: On 10-22, twelve commits (#1462 through #1470) merge design records and site pages for eight components—vllm-plugin-FL, FlagTree, FlagCX, FlagGems, FlagLibs, Megatron, SGLang, verl—consecutively within five hours, the first time this season build design has been written into searchable content.

Torch-FL (53 Merges): Platform Expansion Followed by Item-by-Item Completion

  • MPS Line Completes Its Main Body: Eight MPS-related merges across 10-19 to 10-21—native mm and bmm into the MPSGraph gemm class, softmax and elementwise neg and pow nativized, compiled MPSGraph plan caching, queued-work keepalive and native masked_fill, native MPSGraph kernels for Qwen3.5, advanced indexing collection on Metal; on 10-20, #566 lands with 69 files +22929/−236, CI uses GitHub-hosted Apple Silicon arm64 runners, and explicitly states “x86_64 macOS runners have no Metal device, so environment check failure is expected.”
  • Tsingmicro Triple: torch.compile lands on TX8 (#541, 14 files, author explicitly notes device property queries can only return constants rather than measured hardware values), host-side random number kernels (#604), device timeline into torch.profiler (#630, DeviceTracer based on rocprofiler-sdk, measured on TX8110).
  • Two Hard Fixes for Multi-Card Correctness: Kunlunxin’s CrossDeviceMemcpy previously called the same-card-semantics xpu_memcpy, passing pointers from two cards returned −801 with the destination unwritten, silently invalidating any inter-card copy; on 10-22 it switches to xpu_memcpy_peer (#634); liblaunch_shared.so’s undefined symbol: cudaMemcpy can only be resolved after the Triton launcher dlopen, so Kunlunxin’s CUDA compatibility runtime is preloaded (#632). These two commits produce no performance numbers but require recalculating conclusions from all prior multi-card experiments.
  • CI and Platform Line: BI-V150 shared CI line (#569, 24 files, allowing initial runs to retain known failures to complete bring-up), FlagCICD unified entry with L0 to L2 tiering (#599), AGENTS.md designated as the sole agent guide (#601), Kunlunxin P800’s 2.9 stable branch line (#273), platform pytest results reported to FlagCICD (#542 and #577).

FlagTree (25 Merges): Filling Matrix Cells, Language Layer Awaits Stabilization

  • This Week’s FlagTree Merges Are Almost Entirely on the “Backend Times Version Times Workflow” Three-Dimensional Matrix: Enflame enflame3.6-gcu300’s FlagGems baseline and test workflows, Enflame vLLM benchmarks into workflows, AMD amd3.6 baseline and tests, Ascend 3.5 py312 delivery and fixes, triton_v3.5.x branch codeowner updates, XPU3.6’s vllm benchmark workflow.
  • The Only Structural Change Is AMD Prefill: On 10-21, #1364 adds two targets to FlagMega—amd-gfx1100 (W7900) and amd-gfx1201 (R9700)—enabling bf16 Qwen3-1.7B to complete decoding on RDNA with a single cooperative large core; to avoid fence deadlock from 128 workgroups unable to co-reside, the per-block compute warp cap is pressed to 4; on 10-22, #1365 adds M>1 prefill and “prefill outputs only the last token’s logits.”
  • Interface With RISC-V: smt_rvisa lets authors explicitly specify SEW, VL, tiling, and instruction selection, with kernels lowered through the C++ builder API, plus an LLVM direct-emit path that skips spine-opt entirely; the backend no longer keeps a private TLE dialect copy, instead consuming the common third_party/tle. Such changes appear in no external messaging yet determine whether the RVV backend can go from “runs” to “squeezes out performance.”

Inference Stack: vllm-plugin-FL (18), FlagGems-vllm (21), FlagGems-sglang (15), sglang-plugin-FL (3)

  • Version Alignment Is This Week’s Main Theme: On the vllm-plugin-FL side, 10-19 bumps the DAMO XuanTie backend to vLLM 0.28.0, and 10-22 lands Hygon 0.28.0 (distribution package 0.28.0+empty, running both eager and FULL_AND_PIECEWISE graph modes, TP=2, max model length 4096, memory utilization 0.85); on the sglang-plugin-FL side, 10-21 has Ascend 910C independently adapting 0.5.18 (67 files, acceptance includes four-card full-machine image gate), and 10-22 has Hygon BW1000 upgrading to 0.5.18 (65 files, including single-machine TP2 and physical two-node TP4/PP1, TP2/PP2). Same-version entries for MetaX, SpacemiT, Enflame, and Kunlunxin remain in review or unopened.
  • Graph Mode Spreads to Two Domestic Backends: Graph mode for Sunrise and Enflame GCU300 (10-19, inductor plus cudagraph) merges the same day, so vLLM’s CUDA Graph path is no longer validated only on Ascend and Kunlunxin; on 10-21, Ascend completes decode graph execution and adapts Qwen3.6-27B and 35B-A3B text and image inference (26 files +2513).
  • Model Side Precedes Public Release: FlagGems-vllm’s GLM-5.3 Flash public operators (#900) and Qwen4’s HC / PLE / sparse QSA three operator types (#896) are architecture terms that only appeared in October, with operators landing before the models are public; #977 then adds TP=2 configuration for MetaX fused MoE, averaging 1.065x speedup across 14 M values on C550.
  • Competition Becomes Main-Repo Supply for the First Time: FlagGems-sglang’s ten competition merges on 10-20 were all merged by the same maintainer within 40 minutes (Track 1 tiny-k-gemm, pack-topk-ids, moe-topk-sum, indexed-scale-shift, etc.), accompanied by #163’s single +28940 benchmark snapshot; on 10-22, #204 fused_pack_qkv and #201 fused_sigmoid_mul show competition output has entered the mainline rather than stopping at submissions.

Other Components and Integration Infrastructure

  • FlagBLAS (6 Commits): Hygon gfx936 six-precision GEMM (#151, sgemm through zgemm, +7060/−127, dgemm uses fp64 MMAC dot product and depends on FlagTree’s fp64 MMAC support) and L1 support (#149); on 10-22, Iluvatar L1 optimization (#153) and Moore Threads GEMM (#152). BLAS precision completion exposes compiler-side capability gaps, an empirical instance of cross-repo collaboration this week.
  • FlagCX (5 Commits): Enabling Hygon Device API and unified IR HIP backend (#643), device bitcode generated per adaptor rather than hardcoded nvidia (#668), v0.15.0.dev0 tag, multi-segment VMM windows (#658, 42 files +7667, adding multi-segment callbacks for CUDA, Hygon DU, MetaX, and PPU, exchanging segment file descriptors via bounded descriptors, publishing the window only after collective success), platform routing and CI unification (#665).
  • FlagTrain (5 Commits): Moore Threads backend and two operators (#20), KMCompiler support for Megatron and TransformerEngine (#22), blocked_flash (Iluvatar) and mixed_moe_gemm (#23, #24), multi_tensor_lion (four backends) and quick_gelu (#16, #26). This repo only got its first documentation on 10-15, making it the youngest operator repo currently.
  • Sparse, Attention, and Quantization: FlagSparse merges MUSA optimization (#113, from external contributor NCIC-AlphaSparse branch); FlagAttention replaces roughly 40 host-side small operators and .item() synchronizations in kv-outer’s MiniMax M3 sparse attention prefill with two Triton kernels, building 1.5 to 2.5 times faster and achieving synchronization-free scheduling and full-path graph capture (#107); FlagOS-Compressor supports linear-only INT8 export for DeepSeek V4.1 and MiMo V2.5 (#10) and MXFP8, NVFP4 source checkpoint conversion (#11, 143 tests passing).
  • optest (5 Commits) and FlagOps (2 Commits): optest goes from a script collection to an installable wheel (#4), switches to src layout (#5), exposes console commands (#6), and shows CI job names as repo plus backend (#2); FlagOps takes over five FlagGems actions (#15), after which FlagGems completes migration from local actions to shared actions (#7346, #7416).
  • New and Experimental Repos: FlagCV (created 10-14) lands its first commit late on 10-18 and merges PR #1 skeleton in the early hours of 10-19—interfaces aligned with OpenCV (core and imgproc header groups), multi-backend structure following FlagDNN, kernels written in Triton, with backends/ laying out six backends at once—nvidia, thead, ascend, hygon, iluvatar, mthreads—but no runnable operators yet; Trident is an experimental bridging layer created 10-12 (Triton or Torch-style Python functions compiled to MLIR and LLVM then executed via tvm_ffi), with #3 this week adding several ATen operator tests; flir’s in-review #83 adds the TensorView dialect and tile.load and tile.store views, a substantive move in the CommonIR direction.

IV. Member Units and Vendor Adaptation

  • Kunlunxin: The backend with the highest engineering density this week, and it is not on the public member unit list. Approximately 40 attributable merges (nine on the 10-20 merge date, eighteen on 10-21, thirteen on 10-22; the 10-19 window was not counted separately), spanning three layers: operators, build, and runtime. On the operator side, work shifted from filling in implementations to gating by precision and shape (FP8 capability gating, skipping fp64 and grouped-MM by capability, GRU flattened 1D kernel aggregation score from 0.4217 to 0.8253, fused MoE three-operator from 1.324 to 1.329); on the build side, it became the seventh FlagTree backend; on the runtime side, two hard defects were fixed — silent failure of cross-card copies and preload ordering. The risk is on the operator side — most changes are another tile or fusion of the same operator, there is no unified performance gate, and aggregation scores are not comparable across vendors because of baseline substitution.
  • Hygon: The only vendor this week to complete upstream version alignment in two inference plugin repos simultaneously. vLLM 0.28.0 and SGLang 0.5.18 were merged on the same day (84 files total, net increase of about 5,700 lines), validated on a single-card 64 GiB BW1000 cluster; the vLLM side ships release package 0.28.0+empty, while the SGLang side fixes pipeline-parallel coordinated exit for both physical two-node TP4/PP1 and TP2/PP2. Supporting work: FlagBLAS six-precision GEMM and L1, FlagGems FP8 sparse sum backward firmware, and one hygon fix in FlagDNN. Its reusable value lies in writing the empty-backend upgrade process as a template with a version table.
  • Qingwei Intelligent: Everything added this week addresses infrastructure gaps rather than operator coverage — torch.compile landed on TX8 (#541, device properties can only return constants), host-side random number kernels (#604), device timeline into torch.profiler (#630, measured on TX8110, also pinning the vendor build switch), and FlagTree build lines integrated and hooked to the daily line (#1436, #1440). The order is worth noting: first buildable in CI, then reproducible randomness, then observable performance, and only then optimization.
  • Ascend: The main line this week is versions and graph execution in the two plugin repos — vLLM 0.24 full decode graph execution plus Qwen3.6-27B and 35B-A3B text and image inference (#606), SGLang 0.5.18 standalone adaptation for 910C (#131), W8A16 FP8 TopK (#6030), fp32 capability gating for matmul by SoC (#616), and four wheels migrated to the 22.04 builder (build-infra #1414, glibc floor lowered from 2.38 to 2.35). On the test side, a dozen-plus changes cover FP8 input construction, tile block errors, and reference docs, still mostly fixes.
  • Moore Threads: MUSA became the sixth FlagTree backend and entered the daily line (#1449, #1456); local binaries that change with the ref must be specified explicitly; FlagTree added per-copy TME cache persistence (#1385); FlagBLAS added a GEMM implementation (#152); on the FlagGems side, complex cholesky_solve inputs are now constructed on CPU, and the MUSA specialization of moe_load_balance_loss was merged. Five test skips are concentrated on this backend’s input-domain limits, indicating that failures are still mainly “input unsupported” rather than implementation errors.
  • Enflame: This week’s work covers five stages — baseline, benchmark, packaging, assertions, and graph mode: the enflame3.6-gcu300 baseline and test workflow, vLLM benchmarks into the workflow, builder validation changed to data-driven, toolkit deb preinstalled into the image, and GCU300 graph mode merged; the FlagGems vendor bulk update #7077 (26 files +3520/−511) went into main in one shot, but the title only gives a date with no itemized details. On the FlagTree side, the 02-sparse-mla case was disabled due to an intermittent bus error and the vllm benchmark parameters were updated; stability issues remain at the test-case layer.
  • Iluvatar: Both CI and build sides are in place — the BI-V150 shared CI line merged (#569, CoreX 4.5.0), the iluvatar standalone runtime lets wheels link against the CPU build of torch (#593), and build-infra refreshed the pin to 22.04 and re-released versions; on the code side, direct pushes to master still dominate (four plus three FlagDNN commits, no PR process), and the equivalent performance baseline (equivalent_baseline.cpp et al., +8157) is the repo’s most substantive output this week. Process standardization lags behind Torch-FL and FlagQuantum.
  • MetaX: This week’s effort is concentrated on performance tiers and paired deployment plans under review — the two-node C550, TP16 MiniMax M3 BF16 optimization tier (#540, opened 09-18, still being updated on 10-18) and the four-node C550, TP8 and PP4 Kimi K3 MXFP4 and BF16 implementation (#607, newly opened 10-19, running on the vLLM 0.28 compatibility branch); already merged are the FlagGems-vllm fused MoE TP=2 configuration (#977, averaging 1.065× across 14 M values on C550, also fixing a shared-memory configuration out-of-bounds at M≥4096) and the build-infra change labeling metax native wheels truthfully by platform.
  • DAMO XuanTie (including Zhenwu PPU): FlagBLAS added four-precision GEMM for Zhenwu PPU with NN/TN/NT/TT layout variants (#148, complex GEMM redirected to level3); FlagGems fixed a backend GRU barrier deadlock (#7203); FlagGems-vllm added W8A8 block INT8 matrix multiplication (#954, 9 files +1493, including tuning configs and autotuning integration) and fixed W8A8 MoE integer accumulation; vllm-plugin-FL was upgraded to vLLM 0.28.0 (#574). The four lines correspond to the base library, concurrency correctness, quantized operators, and version tracking.
  • Sunrise and SpacemiT: Sunrise landed a 22.04 self-contained builder on 10-19 and recorded wheels re-released as 22.04-clean; graph mode merged on 10-19 (#569), and the probe entered the optest unified facility. SpacemiT’s smt_rvisa (FlagTree #1381) is the first entry of the RISC-V AI direction into the compiler language-layer extension, and together with the same-backend item under review in FlagGems it constitutes all of this week’s RISC-V line activity.
  • Horizon Robotics and Cambricon: Neither company had any dedicated repository pushes or merges within the organization this week, with both the 10-21 and 10-22 windows silent. The daily report has flagged them as gaps for two consecutive days; if there is still no activity next week, they should be listed separately as a risk item.

V. Ecosystem and Community

  • Conference and post-event coverage formed a three-stage diffusion (10-17 to 10-22): On 10-18, the main forum was first reported on-site by DOIT; on 10-19, China.com published the first post-event full-stack technical breakdown, organized into five sections—”operators and compilation, assisted R&D, plugins and quantization, Token services, and edge-side and platform”—filling in previously uncovered quantization facts: KernelGen now supports nine chips, generated 765 new operators this quarter, of which 416 passed accuracy validation on at least five domestic chips and reached over 80% of native operator performance; all 40 FlagGems-SGLang operators passed on NVIDIA; after W8A8 quantization of Qwen3.5-397B-A17B, the model file shrank from 751.41GB to 382.45GB (a 49.1% reduction) while GPQA-Diamond remained at 84.85%; on 10-20, The Beijing News and Beijing Daily client carried out a second wave of industry-version diffusion, adding a citable formulation: “a new model completes adaptation to 12 chips within one day”; on 10-22, Pandaily provided a systematic English retelling and for the first time disclosed that China Telecom’s Xing4.0-29B-A4B long-input first-token latency dropped 19.87% versus native operators.
  • The boundaries of Liangzhiyun are stated more clearly than in most promotional material: FlagOS handles the execution stack and scheduling, while the Quantum Academy handles hardware and cloud; v0.3.0 and 0.3.1 explicitly state that it does not run on quantum hardware and does not include an MCP server or agent runtime (the MCP adapter is a separate package). The external watchpoints are when Liangzhiyun will open externally, and whether FlagQuantum’s Hygon and Moore Threads backends will be included in cloud scheduling.
  • On the governance side, the acceptance targets of 2.3 were fixed for the first time: FEP-0140 (vLLM three-version line) and FEP-0141 (SGLang 0.5.18, ten vendors) were merged on 10-21; testing reuses the 2.2 workflow, installing the target plugin version from source as needed within vendor or FlagOS images, and acceptance covers installation, dense and MoE text and image inference, performance, concurrency, and 24-hour stability. During the same period, .github added Slack invitation entry points to both the Chinese and English versions of the organization homepage, resolving the issue of new developers being unable to find the existing invitation.
  • The first external project with deliverables appeared in the embodied intelligence direction (10-22): The Beijing Municipal Science & Technology Commission and Zhongguancun Administrative Committee published a pre-approval notice for proposed projects under the 2026 “Central Guidance for Local Development” and “New Generation Information and Communication Technology Innovation” special programs, in the Zhongzhi FlagOS acceleration plan direction; Kuowei Zhixin’s “Edge-side Deployment and Application Demonstration of AlohaMini Embodied Intelligent Robot Based on FlagOS-Robo” was selected. Acceptance covers embodied model training adaptation and inference deployment, domestic chip edge-side applications, multi-chip inference accuracy consistency testing, and coordination among the embodied model, motion-control cerebellum, and body control. This is the third external project with deliverable constraints, following the FEP matrix and competitions.
  • Community influence event (10-19, retrospective published 10-21): Triton Developer Conference 2026 was held in San Jose with over 500 developers in attendance; the FlagOS team’s KernelGen was the only project on site to systematically demonstrate a “multi-chip operator auto-generation Agent,” and it is built on the FlagTree compilation infrastructure; under the 2.2 formulation, the number of auto-generated operators on 8 chips ranged from 77 to 551, with performance comparison ranges from 2.58x to 38.33x (the original text notes these are the test conditions of each speaker and should not be extrapolated across projects).
  • Competitions and developer operations: Track 1 of the SGLang cross-chip operator optimization competition posted on the CSDN FlagOS site has begun producing mergeable results for the main repository, but the pace remains batch-based rather than distributed; on 10-22, a university roundtable livestream was added (teaching implementation under the Global AI Open Computing Faculty Development Program), which is a community activity rather than a technical release.
  • Channel verification conclusion: Across 24 Chinese and English query terms in five windows, direct searches for component names (FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagAttention, FlagCX, KernelGen) consistently returned zero hits; all valid entries came from member organization posts, industry media, and the community’s own channels. The latest original article on the CSDN FlagOS column homepage is still the 08-28 Day0 adaptation post; the BAAI Community search page is client-rendered and yields only a framework page; eight HN Algolia queries produced no direct title matches. Verifiable progress in ecosystem building still rests on the repositories.

VI. Trend Observations

  • Unified delivery surface precedes feature expansion: The heaviest batch of work this week was not new operators, but the glibc baseline, builder form factor, communication dependencies, and installation commands—the 22.04 self-contained builder consolidated from six backends to seven, flagcx==0.14.0 expanded from two nvidia lines to twenty backends, the one-click install script reduced multi-repo dependencies to a single command, and optest turned test infrastructure into an installable wheel. The common effect of these moves is “no more editing install scripts when switching backends,” which more directly determines 2.3’s delivery speed than any single-point performance gain.
  • The loop from requirement to evidence was compressed to one day this week: The two FEPs merged on 10-21 wrote the version-by-vendor matrix, and on 10-22 Hygon lit up the first cell; likewise, after optest and FlagCICD landed on 10-20, the FlagGems test-case rework on 10-22 immediately pivoted to “direct-connect candidates.” The remaining nine vendors and two version lines for 2.3 are the main workload for the next six weeks; the thing to watch is whether the same 0.28.0 reuses Hygon’s empty backend template on SpacemiT, MetaX, Enflame, and Kunlunxin.
  • The quality-over-quantity orientation has stabilized: Of the 661 merges for the week, fewer than 20% actually added operator capability; the bulk was test contracts (at least 119 from a single person), automated output (three KernelGen batches of 14 operators plus the sparse family under review), and build infrastructure. This is consistent with the post-2.2 “measure existing capability accurately first” orientation; the cost is slower operator-count growth and slower updates to external performance numbers.
  • The quantum stack has formed a replicable mode of advance: write the contract as a rejectable implementation first, then open it up block by block: expand the subpackage while locking the core (159 new public names into the subpackage, Stable Core’s 41 exports unchanged), make deployment pre-checks mandatory gates, remove vendors with no external service, and freeze unapproved sampling interfaces. Half of this week’s merges went into rejection and deletion, which determines whether M3’s subsequent boundaries are credible; the risk is that all 20 to 34 merges were done by a single account, and verification of physical semantics depends on the same people writing tests and implementations, leaving no independent external review.
  • Observability and multi-card correctness have become the new weak spots for multi-backend: SpacemiT got a device timeline for the first time, and Kunlunxin’s cross-card copy had been wrong all along (bytes were never written; synchronous calls would not mask it)—both surfaced only after “it runs.” The past two weeks’ test porting and test-case contract rework, and this week’s optest centralization, are essentially different expressions of the same judgment: after “it runs,” the first hurdle for multi-backend is reproducibility, not speed.
  • Three risk items need confirmation in the next issue: First, “whole-batch update” commits lack verifiable detail (Enflame’s 26 files, +3520/−511, cannot be mapped to specific operators and gains), and Kunlunxin’s fused MoE was measured against a substituted baseline—both will distort subsequent cross-vendor comparisons. Second, Horizon and Cambricon have been silent for two consecutive days; if this continues next week, it should be called out separately in the weekly report. Third, on the training side (Megatron-LM-FL, FlagScale), pushes were sparse this week, and the three NPU PRs for TransformerEngine-FL have not yet merged; the training stack is lagging the inference stack in the 2.3 cycle.
  • Next things to watch: whether the FlagTree 0.7.1 development build goes GA before 2.3; whether the second vendor for vLLM 0.28 (SpacemiT or MetaX) reuses the Hygon template; when FlagCV lands its first runnable operator; whether per-session notes from the conference sub-forums and workshops are published; and when Liangzhi Cloud opens to the public.

Appendix: Sources and Verification

  • Sources: Five FlagOS daily reports dated 2026-10-19, 10-20, 10-21, 10-22, and 10-23, covering windows of 10-16 06:00 to 10-19 06:00 (72 hours), 10-19 06:00 to 10-20 06:00, 10-20 06:00 to 10-21 06:00, 10-21 06:00 to 10-22 06:00, and 10-22 06:00 to 10-23 06:00 respectively; no issues were missing during the period.
  • GitHub verification method: Push status was checked daily for all 61 repositories in the flagos-ai organization; merges were verified as ground truth via is:merged plus merged:>=window start, and in-review items via is:open plus updated:>=window start; new commits on the main branch were cross-checked with commits?sha=<branch>&since=<UTC>, and pushes to non-default branches were disambiguated using the org event stream and commits.atom as a fallback. Per-repo merge counts are the sum of the figures disclosed in each daily report; since some issues counted only the first 100 merged PRs, the attributable total is 641, with the 20-item difference being tail repositories not broken out.
  • Version and tag verification: Each was re-checked against the components’ repositories via the releases.atom feed. Four tags were detected within this week’s window (FlagCX v0.15.0.dev0, FlagTree 0.7.1a1+ascend3.5, 0.7.1.dev1, 0.7.1.dev1+ppu3.6); FlagGems, FlagScale, FlagBLAS, FlagDNN, Torch-FL, vllm-plugin-FL, sglang-plugin-FL, build-infra, FlagAttention, FlagSparse, FlagOps and other repositories had no new releases, consistent with the daily reports.
  • News-side verification: 24 sets of Chinese and English Google News RSS queries (14-day window for component names, 7-day for member and industry terms), eight HN Algolia queries, the CSDN FlagOS column, and original posts from BAAI communities and member organizations were cross-verified item by item; valid items were cited on an “original site first” basis, while late-indexed items and republished old news (e.g., market pieces of a Huorong nature, old news about an industry expo award) were excluded per the rules and not counted.
  • Scope notes: All dates and times are Beijing time; the Monday issue covers 72 hours, and its 10-16 to 10-18 content is labeled with original dates, overlapping with the previous weekly report (covering 10-12 to 10-16) on boundary days; tags and FEPs in the overlapping portion were already listed in the previous issue and are not double-counted here; “merged” refers to PRs whose GitHub merge time falls within the window, which is offset by about one day from the daily reports’ day-by-day window counts; this report uniformly adopts the merge-date basis.
  • Limitations: Branch-level activity (e.g., FlagPrism’s Enflame debug branch, FlagGems’ bump-flagtree-tsm branch) cannot be fully enumerated under the default-branch basis; on the training side, Megatron-LM-FL and FlagScale saw sparse activity within the window, so the confidence of related judgments is lower than that of the inference stack; news-side hits for component names remain at zero, and ecosystem content comes mainly from member organizations and industry media; portions that are not official FlagOS announcements are labeled in the body by source nature.

References