Research window: 2026-09-22 10:00 ~ 2026-09-23 10:00 (approx. 24 hours; continues the 09-22 report window with no gaps) Sources: GitHub (org: flagos-ai, full review of pushed_at across 54 repos, 24 repos had pushes within the window; commit search returned 143 hits across 14 repos; Release Info portal and community list checked page by page), Google News RSS (24 sets of Chinese and English query terms, via proxy), AICC2026 conference official site and coverage (Guangming Online / The Beijing News / CBISMB), 2026 Yunqi Conference coverage (Guandian / National Business Daily / East Money / Alibaba Cloud Developer Community), Loongson Technology announcements (ITHome / MyDrivers), etc. (see appendix source list for details)

In This Issue

  • Today’s Highlights: 5 days to 2.2 GA (09-28) — release checklist committed to repo, Release Info portal rebuilt, Megatron application images 2.2.0-0.2.3 cover 14 backend rows (09-22/09-23)
  • I. Open Source Project Progress (GitHub Activity)
    • 1.1 Release Engineering Addendum: FlagCX wheel channel advances to Kunlunxin and all rows, deb/rpm covers every packaged component (09-22/09-23)
    • 1.2 FlagQuantum: 38 commits in a single window — cloud QPU scaffolding, three-framework interoperability, simulation fusion acceleration (09-22)
    • 1.3 FlagGems: KernelGen adds 15 more Nvidia operators; dense merge of fixes and optimizations across six platforms (09-22/09-23)
    • 1.4 FlagGems-vllm: Ascend XCS new operators merged; Moore Threads / Hygon / MetaX three in-review-to-merged (09-22/09-23)
    • 1.5 Torch-FL: 11 commits converged — DCU and Ascend fixes, build matrix unified, retired code cleanup (09-22/09-23)
    • 1.6 Other Activity: FlagCX transport hardening, FlagSparse three-backend progress, TransformerEngine-FL NPU backend, FlagTree PPU CI (09-22/09-23)
  • II. News Coverage and Ecosystem
    • 2.1 AICC2026 held in Beijing: Two FlagOS talks and launch of “Reliable AI” evaluation system (09-21/09-22)
    • 2.2 Yunqi Conference: DAMO XuanTie unveils Zhenwu V900 — FlagOS-supported platform PPU welcomes next-generation hardware (09-22)
    • 2.3 Domestic compute software stack updates: Loongson accelerated computing platform first release; Global Digital Trade Expo opens today (09-22/09-23)
    • 2.4 Component-level retrieval: eighteenth consecutive quiet window (09-22/09-23)
  • III. Member Company Deep Dive
    • 3.1 Hygon: DCU index boundary fix and INT8 block-wise BMM merged (09-22/09-23)
    • 3.2 Moore Threads: W8A8 variable-length attention quantization support merged; AICC2026 talk and “Reliable AI” participation (09-22)
    • 3.3 MetaX: Fused inverse RoPE quantization operator merged; topk W8A16 optimization and MACA FFT default strategy (09-22)
    • 3.4 Ascend: Operator batch and test image updates; TransformerEngine-FL NPU backend implementation (09-22/09-23)
    • 3.5 Iluvatar CoreX: Iluvatar wheel 22.04 rebuild line and container row registration (09-22/09-23)
    • 3.6 Enflame: Multi-row image registration and sglang application row (09-22/09-23)
    • 3.7 Tsingmicro: AICC2026 agenda and runtime row (09-22)
    • 3.8 DAMO XuanTie: Zhenwu V900 and PPU CI wrap-up (09-22/09-23)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Complete Source List

Today’s Highlights: 5 Days to 2.2 GA (09-28) — Release Manifest Landed, Release Info Portal Rebuilt, Megatron Application Images Cover 14 Backend Rows

Date: 2026-09-22 ~ 2026-09-23　Sources: community #111, release-2.2.yaml, Release Info Portal, build-infra #1019 to #1034

Per the official schedule in the community 2.2 release tracking (#47), the rc2 stabilization window closes on 09-24 (hotfixes only thereafter), with the GA Go/No-Go set for 09-28. This window falls squarely in the final stretch, and release engineering advanced on three fronts in a single day:

First, the official release manifest has begun to land. At 00:11 on 09-23, community #111 was merged: it adds release/2.2/release-2.2.yaml — the official release source manifest for FlagOS 2.2 (vcstool format, enabling one-shot retrieval of all component sources via vcs import). The manifest’s first entry is sglang-plugin-fl v0.2.0 (the official tag is placed on the commit where the module passed final rc2 validation, branch 0.2.0-rc2, HEAD 1c84db7). The file header specifies the versioning convention: official tags are always bare version numbers (no rc/post suffix), with FlagTree as the exception, retaining the upstream local version format (e.g., 0.7.0+triton3.6); structurally it is organized by layers such as the L2 “framework adaptation layer,” with remaining components to be added in batches.

Second, application image registration has advanced to 14 backend rows. Among the 33 commits to build-infra in this window, the Megatron-LM-FL application image tag 2.2.0-0.2.3 was registered across 14 platform rows (nvidia-cuda12.8 / cuda13.3, ascend-cann8.5.0 / cann9.0.0, 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-maca3.8.1.3, mthreads-musa4.3.6 / 5.2.0, corresponding to #1019 through #1033); accompanied by two fixes: build steps now read the locked wheel version directly (#1018), and installation versions are precisely pinned with two previously unbuilt rows added (#1017). Also registered in the same period was the vLLM Kunlunxin application image 2.2.0-0.3.0rc2.post2 (#1012 / #1013, with authorized rebuild #1010).

Third, the Release Info portal was rebuilt. From 06:38 to 06:44 on 09-23, the release-info repository’s gh-pages branch received 13 consecutive pushes, rebuilding the Hugo portal (flagos-ai.github.io/release-info). The portal is organized into four tiers — “vendor / base image (base) / runtime image (runtime) / application image (application)” — covering base and runtime images for 20 platform rows, plus application images for vLLM (two matrices: 0.20.2 / 0.24.0), Megatron (training / rl), and sglang 0.5.18; on the Megatron nvidia-cuda12.8 row page, for example, the complete pull instructions for harbor.baai.ac.cn/flagos-app/megatron_training0.17.1-nvidia-cuda12.8:2.2.0-0.2.3 and flagos-runtime-nvidia-cuda12.8:2.2.0 are already visible. Image registration, portal publication, and manifest landing aligned on the same day for the first time in this window, and the 2.2 release artifacts are beginning to become externally visible.

Version branches advanced in parallel: FlagTree 0.7.0-rc2-triton3.6 (pushed 09-23 09:36), vllm-plugin-FL 0.3.0-rc2 (09-22 23:20, #549), FlagScale 2.1.0-rc2 (09-23 08:43, #1301 cherry-picked from main), sglang-plugin-FL 0.2.0-rc2 (v0.2.0 already tagged).

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 54 repos in the org, 24 had pushes during the window; commit search returned 143 hits (across 14 repos), with additional branch-side pushes from release-info (gh-pages), FlagScale, vllm-plugin-FL, FlagTree, docs, sglang-plugin-FL, etc., not counted in the search. Breakdown: FlagQuantum 38, build-infra 33, FlagGems 28, Torch-FL 11, FlagSparse 10, FlagGems-vllm 6, FlagFFT 4, FlagCX 3, FlagTree 2, FlagAttention 2, FlagBLAS 2, whisper-triton 2, community 1, FlagGems-Experimental 1.

This window’s shape = “2.2 release wrap-up” + “quantum stack explosion” + “dense multi-platform fixes”: release engineering is detailed in “Today’s Highlights”; FlagQuantum was the single most active repo with 38 commits; FlagGems / FlagGems-vllm / Torch-FL all entered a “fix + tune” cadence simultaneously.

1.1 Release Engineering Supplement: FlagCX wheel channel advances to Kunlunxin and all rows, deb/rpm covers every packaged component (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: build-infra #1000, #1002, #1007, #1008, #1014

  • FlagCX wheel distribution line: Kunlunxin wheel landed (#1000, following up yesterday’s report’s “under review” item); the wheel channel opened across all remaining eight rows (#1002); per-row pins auto-derived from ref (#1007), every dispatch must carry a wheel ref (#1006), per-row records of published pins and documentation of the “wheel-to-torch binding relationship” (#1005 / #1016); release containers inherit node proxies (#1015). The per-backend precompiled wheel distribution model expanded from “pilot rows” to “all rows + pipeline hardening.”
  • Native package coverage: every packaged component gains a path into the repo (#1001), deb and rpm coverage completed (#1008) — consistent with the acceptance line of FEP-0019 (unified packaging integration, Wave 1 already includes the four-repo merge of FlagCX / FlagTree / FlagSparse / FlagAttention).
  • FlagTree Iluvatar wheel rebuilt on Ubuntu 22.04 and builder published (#1014, Tianshu line).
  • Two additional doc cleanups: retracted the claim of cross-compiler flag_gems ConfigCache pollution (#1034), retired the metax flagtree downgrade rationale (#1020).

1.2 FlagQuantum: 38 commits in a single window — cloud QPU scaffolding, three-framework interop, simulation fusion acceleration (09-22)

Date: 2026-09-22　Source: FlagQuantum #135, #154, #138, #168, #167

FlagQuantum (the PyTorch-first quantum AI framework in the FlagOS ecosystem) saw 38 commits in a single window, with four parallel lines:

  • Cloud QPU execution: added Azure Quantum QPU execution scaffolding (#168) — another cloud backend line following last month’s Braket / CUDA-Q.
  • Cross-framework interop: Amazon Braket circuit adapter implementation (#154, with companion contract #147), Cirq circuit conversion (#135) and symbolic parameter contracts (#166 / #170), CUDA-Q kernel export (#138, fixing finite-angle IR validation #145), cross-framework capability gap tracking (#164), cross-framework circuit semantic conformance tests (#160).
  • Simulation performance: CPU state-vector gate acceleration (#142), tail two-qubit gate acceleration (#143), cross-wire diagonal region fusion (#146), two-wire diagonal matching fusion (#153), single-wire gate layer fusion (#157 / #159 / #165), mixed-width region fusion (#169), plus an optional-switch vectorized trajectory row sampler (#132) and path-decomposition CPU lane + differential kernel network benchmark (#128).
  • Validation line (regional twin): circuit suite validation (#123), holdout set validation (#141), history and evaluation persistence (#137 / #144 / #152 / #163), drift alignment (#155), holdout-evidence gating (#162), publishing validated candidates (#167).
  • CI: runner reuse across optional SDK versions (#158), braket coverage lane (#161), CPU test parallelization (#149), suite deduplication (#150 / #156), queue bottleneck cleanup (#148).

1.3 FlagGems: KernelGen adds 15 more Nvidia operators; dense six-platform fixes and optimizations merged (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: FlagGems #5245, #5845, #6531, #6339, #5063

  • KernelGen (Nvidia) 15 operators: orgqr (#5845), trapz (#5927), matrix_power (#5735), matmul_backward (#5639), _pad_packed_sequence (#5245), _scaled_dot_product_attention_math_for_mps (#5259), _nested_tensor_softmax_with_shape (#5693), _nested_tensor_from_mask (#5691), _sample_dirichlet (#5946), _ctc_loss (#5489), _batch_norm_with_update (#5828), _cudnn_rnn (#5823), _transform_bias_rescale_qkv (#6141), _spdiags (#6102), vander (#5945) — the auto-generation pipeline continues to batch-fill long-tail operators.
  • NVIDIA quality line: restored native FP8 WGMMA for scaled MM (#6531).
  • Hygon: fixed index_reduce_ issues on the Ascend / Tianshu / MetaX platforms (#5063, contributed by SiliconFlow).
  • MetaX: merged and optimized the topk_w8a16_fp8 operator (#6339).
  • Ascend: cat/concat dispatch and implementation improvements (#6548), rrelu_with_noise implementation (#6534, KMCompiler line), nonzero_static benchmark switched to operator mode (#6571), 910B weekly test image update (#6590).
  • Quantization and general: dynamic W4A8 matmul and portable 4-bit packing validation (#6494), block diag merged (#6555), vector_norm negative-order finite-value mask fix (#5883).

1.4 FlagGems-vllm: Ascend XCS new operator merged; three Moore Threads / Hygon / MetaX items move from review to merge (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: FlagGems-vllm #817, #823, #803, #826

  • Ascend XCS: KV pool state compression (kpool) Triton operator merged (#817, 09-23 09:22), continuing the XCS operator batch cadence.
  • Three items move from review to merge: Moore Threads variable-length attention W8A8 FP8 (#823), Hygon INT8 block-wise BMM and INT8 einsum contraction (#803), MetaX fused inverse RoPE FP8 quantization (#826) — all three were “under review pool” items in previous daily reports, landing together in this window.
  • Engineering fixes: TLE paths gated by required capability (#833), test dependencies no longer upgrade existing numpy (#832).

1.5 Torch-FL: 11 commits converge — DCU and Ascend fixes, build matrix unification, retired code cleanup (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: Torch-FL #395, #393, #392, #387, #372

The PyTorch adaptation layer (Torch-FL / PyTorch-Plugin-FL) saw 11 commits in this window, with engineering efforts converging:

  • Device fixes: DCU out-of-bounds index values rejected outright (#395), Ascend kernels changed to execute on the device where operands reside (#392), PPU restores CUDA metadata for CPU torch wheels (#370), SDPA attention preserves the autograd graph in dispatch mode (#390).
  • Build and configuration: platform / build matrix converged into a single table (#387), unrecognized or undetected GEMS_VENDOR now errors outright (#388), added ruff config and removed black/isort (#386), codegen --check gate and registration consistency tests (#382).
  • Cleanup and performance: removed global Tensor.__getitem__ patch (#372), retired code such as the MetaX native tree (#393), qwen-image all-valid prompt mask omission (#371).

1.6 Other Activity: FlagCX transport hardening, FlagSparse three-backend progress, TransformerEngine-FL NPU backend, FlagTree PPU CI (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: FlagCX #594, TransformerEngine-FL #113, FlagTree #1259, FlagSparse #80

  • FlagCX: heterogeneous transport hardening and alignment with accelerator CI (#594), p2p engine rpc serve extension and barex transport upgrade (#611), packaging builds only on packaging changes (#616).
  • FlagSparse: NCIC-AlphaSparse collaboration branch merged two batches in a row (#79 / #80) — DCU SpMM COO optimization, MACA SpMM results, Ascend CI and outputs, MTX tests entering CI.
  • TransformerEngine-FL: added NPU backend LayerNorm implementation (#114) and NPU backend implementation for multi-tensor Adam (#113).
  • Megatron-LM-FL: stabilized vendor backend compatibility and CI integration (#174).
  • FlagTree: main branch PPU CI adds pid.txt validation (#1259, merged 09-23 10:05), GEMS reusable workflow incorporates CORE change detection (#1265), new builtin function documentation update (#1120).
  • FlagAttention: test and benchmark updates (#69), pytest adds --record / --output (#70).
  • FlagBLAS: Ascend L2 performance optimization merged (#123 — SPR / SPR2 / HPR / HPR2), external contributor packaging line (#1, version aligned to the 0.3.0 release line).
  • FlagFFT: MACA “1d single” measurement strategy switched to default, HCU / Ascend CANN9 environment documentation completed.
  • FlagGems-Experimental: KernelGen adds dim operator (#702).
  • whisper-triton: synced to the iFlytek Codeup patch baseline (#6).

II. News Coverage and Ecosystem

2.1 AICC2026 Held in Beijing: Two FlagOS Talks and Launch of the “Reliable AI” Evaluation System (09-21/09-22)

Date: 2026-09-21 (conference dates) / 2026-09-22 (reports published)　Sources: Guangming Online, The Beijing News, CBISMB, AICC Official Agenda

The 7th Artificial Intelligence Computing Conference (AICC2026) was held on 09-21 at the Zhongguancun International Innovation Center in Beijing, with reports appearing from 09-22. Two items directly related to FlagOS: In the 09:50 slot of the main forum, Lin Yonghua, Vice President and Chief Engineer of the Beijing Academy of Artificial Intelligence (BAAI), delivered the talk “From Architectural Innovation to Usable Compute: FlagOS-Driven Software-Hardware Co-Design as a New Paradigm for AI Computing,” noting that “AI workloads will not converge, nor will computing architectures be reduced to a single type; translating hardware innovation into usable compute requires the support of a shared software ecosystem—Zhongzhi FlagOS lowers the repeated adaptation costs of diverse chips through unified compilers, operators, communication, and training/inference frameworks”; according to the conference agenda, BAAI researcher Zhao Shuai also gave a talk in a special forum titled “Zhongzhi FlagOS: Unlocking the Value of Diverse AI Compute with an Open Software Stack.” The conference also held the launch ceremony of the “Reliable AI” Evaluation Working Group (China Electronics Standardization Institute’s Fifth Electronics Research Institute, 10:10 slot), and in the industrial achievements session formally released the “Reliable AI” evaluation system for the Agent era—jointly released by the Fifth Electronics Research Institute of the Ministry of Industry and Information Technology together with BAAI, Inspur Information, Moore Threads, MetaX, the Beijing Institute of Open Source Chip, Baidu AI Cloud, Alibaba Cloud, and more than 20 other industry chain organizations, covering the full chain of chips, complete systems and interconnect, frameworks and systems, models, data, and applications, and conducting full-link evaluation across reliability, security, and stability dimensions. At the same time, IDC and Inspur Information jointly released the “2026 China Artificial Intelligence Computing Power Development Assessment Report”: China’s intelligent computing power is expected to reach 2,576.5 EFLOPS in 2026, up 87.9% year-on-year. The density of member organizations on the agenda (Moore Threads, Qingwei Intelligent, two BAAI FlagOS talks) shows that this stack has become a standing topic in China’s open compute ecosystem.

2.2 Yunqi Conference: DAMO XuanTie Releases Zhenwu V900—FlagOS-Supported PPU Platform Welcomes a New Hardware Generation (09-22)

Date: 2026-09-22　Sources: Guandian, National Business Daily, East Money, Alibaba Cloud Developer Community

At the 2026 Yunqi Conference (09-22, Hangzhou), Alibaba’s DAMO XuanTie (T-Head) released its new-generation training-inference integrated AI chip Zhenwu V900: 216GB memory, 1,200GB/s inter-die interconnect bandwidth, native support for FP8 / FP4 low-precision computing, and 3x the single-chip performance of the previous generation Zhenwu M890, with mass production scheduled for Q1 2027; the conference also previewed the next-generation Zhenwu J900 (March 2028) and two generations of Yitian 720 / 730 server CPUs (2027), and disclosed that the Zhenwu V900 supernode features unified memory addressing and can form thousand-card full-bandwidth interconnect. The Zhenwu series is Alibaba’s in-house PPU (Parallel Processing Unit), and PPU is already one of the supported platforms of the FlagOS unified plugin and compiler (FlagGems-vllm PPU operators, FlagTree PPU CI, FlagDNN PPU architecture)—the release of a new hardware generation means this stack’s PPU backend now has a clear generational upgrade target.

2.3 Domestic Compute Software Stack Updates: Loongson Releases First Version of Its Accelerated Computing Platform; Global Digital Trade Expo Opens Today (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Sources: ITHome, MyDrivers, Global Digital Trade Expo AI Zone Preview

  • Loongson Technology released the first software version of its “Loongson Accelerated Computing Platform” on 09-22: based on the LG200 series general-purpose GPU integrated into Loongson 2K3000 / 9A1000, it covers drivers, compilers, runtime, operator libraries (BLAS / DNN), inference engines, and debugging and performance analysis tools, supports both OpenCL 3.0 and CUDA programming models, enables low-cost migration of CUDA applications, and is planned to achieve generational software compatibility as it evolves with 9A2000 / 9A3000. Another example of a domestic compute “full-stack software platform,” serving as a reference point for this stack’s approach of “a unified software stack lowering adaptation costs.”
  • The 5th Global Digital Trade Expo opened on 09-23 in Hangzhou (through 09-27): it features for the first time a “Digital Trade · Chip Scenarios” integrated circuit application ecosystem section, with companies including MetaX participating in a coordinated “chip + materials + compute” exhibition; 117 compute companies are exhibiting. The industry visibility of member organizations is carried forward by the conference season.

2.4 Component-Level Search: Eighteenth Consecutive Quiet Window (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Sources: Google News RSS (24 sets of Chinese and English query terms, via proxy)

Searches in Chinese and English using component names such as FlagOS / FlagGems / FlagScale / FlagTree / FlagPerf / FlagCX / KernelGen (when:7d and when:14d) returned zero component-level hits in the 24-hour window, constituting the eighteenth consecutive quiet window (the previous window was the seventeenth). This contrasts with the conference side: in this window, external information was carried by member/ecosystem developments at AICC2026 and the Yunqi Conference (see 2.1–2.3). HN Algolia returned no relevant hits for terms such as FlagOS / FlagGems; searches of the BAAI community found no new FlagOS-related articles.

III. Member Unit Deep Dive

3.1 Hygon: DCU Index Boundary Fix and INT8 Block-wise BMM Merged (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: Torch-FL #395, FlagGems-vllm #803

  • Torch-FL #395: Out-of-bounds index values on DCU now directly rejected (preventing silent erroneous results).
  • FlagGems-vllm #803 (in review, transitioning to merge): INT8 block-wise BMM kernel and INT8 einsum contraction (bhr,hdr→bhd) — low-precision matrix multiplication for the inference side, complementing the Hygon implementation.
  • Image rows: megatron_training-hygon-dtk26.04 registered as 2.2.0-0.2.3; runtime hygon-dtk26.04 already published on the portal.

3.2 Moore Threads: W8A8 Variable-Length Attention Quantization Support Merged; AICC2026 Talk and “Reliable AI” Participation (09-22)

Date: 2026-09-22　Source: FlagGems-vllm #823, AICC Official Agenda, CBISMB

  • FlagGems-vllm #823 (in review, transitioning to merge): W8A8 FP8 support for variable-length attention (varlen) — the quantization path on the MUSA platform is now open.
  • AICC2026: Ren Yanping, Deputy General Manager of Moore Threads’ Solutions Department, spoke at a specialized forum; also participated as a co-release unit in the “Reliable AI” evaluation system.
  • Image rows: megatron_training-mthreads-musa4.3.6 / musa5.2.0, two rows registered as 2.2.0-0.2.3.

3.3 MetaX: Fused Inverse RoPE Quantization Operator Merged; topk W8A16 Optimization and MACA FFT Default Strategy (09-22)

Date: 2026-09-22　Source: FlagGems-vllm #826, FlagGems #6339, FlagFFT commits

  • FlagGems-vllm #826 (in review, transitioning to merge): Fused inverse RoPE and FP8 quantization operator added and optimized for MetaX implementation.
  • FlagGems #6339: topk_w8a16_fp8 operator added and optimized (metaX).
  • FlagFFT: MACA “1d single” measurement strategy promoted to default (measurement-driven tuning moves from experimental to official).
  • Torch-FL #393: Retired code including the MetaX native tree removed — the adaptation layer converges toward plug-in architecture.
  • Image rows: megatron_training-metax-maca3.8.1.3 registered as 2.2.0-0.2.3.

3.4 Ascend: Operator Batch and Test Image Updates; TransformerEngine-FL NPU Backend Implementation (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: FlagGems #6548, #6590, FlagGems-vllm #817, TransformerEngine-FL #113

  • FlagGems: cat/concat dispatch and implementation improvements (#6548), rrelu_with_noise (#6534), nonzero_static benchmark fix (#6571), 910B weekly test image update (#6590).
  • FlagGems-vllm #817: XCS line adds KV pool state compression (kpool) operator (merged on the morning of 09-23).
  • TransformerEngine-FL #113 / #114: NPU backend implementations for multi-tensor Adam and LayerNorm — the Ascend backend for the training-side framework is now complete.
  • Torch-FL #392: Ascend kernels now execute on the device where operands reside, fixing cross-device dispatch.
  • Image rows: ascend-cann8.5.0 / cann9.0.0 and 910c variants, multiple rows registered and published on the portal.

3.5 Iluvatar: Iluvatar wheel 22.04 Rebuild Line and Container Row Registration (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: build-infra #1014, Release Info Portal

  • build-infra #1014: Iluvatar wheel rebuilt on Ubuntu 22.04, and its builder image published — the wheel distribution line is now complete for the Iluvatar row.
  • Image rows: iluvatar-corex4.4.0 / corex4.5.0, two rows registered as megatron image 2.2.0-0.2.3; base / runtime rows published on the portal.
  • FlagGems #5063 fixes the index_reduce_ implementation on the Iluvatar platform (contributed by SiliconFlow).

3.6 Enflame: Multi-Row Image Registration and sglang Application Rows (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: Release Info Portal

  • Image registration: megatron_training-enflame-tops1.9.10 / tops1.10.6, two rows registered as 2.2.0-0.2.3; sglang application images include two enflame-tops rows; base / runtime rows updated with the portal rebuild.

3.7 Tsingmicro: AICC2026 Agenda and Runtime Rows (09-22)

Date: 2026-09-22　Source: AICC Official Agenda, Release Info Portal

  • AICC2026 Agenda: Li Bin, Vice President of Software at Tsingmicro, presented “3D Chips, Open Source New Chapter: Open3D-PIMC Reconstructing the Computing Software Foundation” — a public move by a reconfigurable computing vendor on the open software stack side.
  • Image rows: tsingmicro-tsm260610 base / runtime rows updated with the portal rebuild; vllm-plugin-FL 0.3.0-rc2 branch includes tsingmicro txda 0.24.0 adaptation (previously merged in the background, now finalized with the version branch).

3.8 DAMO XuanTie: Zhenwu V900 and PPU CI Finalization (09-22/09-23)

Date: 2026-09-22 ~ 2026-09-23　Source: FlagTree #1259, Torch-FL #370, Guandian

  • Hardware side: Zhenwu V900 (next-generation PPU) released, mass production in 2027 Q1 (see 2.2).
  • Software side: FlagTree PPU CI adds pid.txt validation (#1259, merged 09-23 10:05) — CI integrity finalization for the PPU compilation pipeline; Torch-FL restores CUDA metadata for CPU torch wheel under the PPU path (#370).

IV. Summary and Trend Observations

  • 2.2 enters its final week before GA, and deliverables are becoming “visible”: The first entry in the official release manifest is now in the repository (sglang-plugin-fl v0.2.0), the Release Info portal was rebuilt the same day, and Megatron application images 2.2.0-0.2.3 cover 14 rows. The rc2 stabilization period ends 09-24, with GA set for 09-28 — the observation point for the coming week is the batch-by-batch completion of the manifest and the finalization order of rc2/post tags across components.
  • Distribution formats continue to broaden: FlagCX wheel expanded from pilot rows to all rows + Kunlunxin landing, deb/rpm coverage extended to every packaged component, FlagTree Iluvatar wheel rebuilt — aligning with the FEP-0019 (unified packaging integration) Wave 1 acceptance line, the “beyond images” native package channel is becoming a standard delivery surface for 2.2.
  • The quantum stack is the day’s highest single-repo throughput: FlagQuantum advances 38 commits simultaneously across cloud backends (Azure Quantum / Braket / CUDA-Q / Cirq four-way interoperability), simulation fusion acceleration, and candidate validation gates — moving from a “following component” toward a framework-level development cadence.
  • Three repos simultaneously enter fix/tuning cadence: FlagGems (15 KernelGen + six-platform fixes), FlagGems-vllm (three in-review transitioning to merge), Torch-FL (11 engineering finalizations, including removal of global patches and retired code) — the “acceptance, fix, tune” window shape after feature freeze is consistent with the schedule.
  • External information surface is carried by conference season: The eighteenth quiet window for component-level search vs. two FlagOS talks at AICC2026 and the “Reliable AI” launch, the Yunqi Zhenwu V900 release, and Loongson’s full-stack GPU software platform — industry attention on “unified software stack / diversified compute” is being concentratedly released during conference season, mutually corroborating with the code-side finalization.
  • To watch: Subsequent entries in the 2.2 manifest (FlagTree 0.7.0+triton3.6, vllm-plugin-FL 0.3.0, FlagScale 2.1.0 and other version branches already in progress); software adaptation follow-up for new PPU hardware (V900 mass production in 2027 Q1); FlagSparse / FlagAttention packaging line (FEP-0019 Wave 1) landing before GA.

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full verification of pushed_at across 54 repos 24 repos active within window
GitHub Commit search + per-repo cross-check Item-by-item verification within window 143 items (14 repos) + branch-side pushes (release-info / FlagScale / vllm-plugin-FL / FlagTree / docs)
GitHub Per-repo releases scan releases API across 13 repos No new releases within window (2.2 via manifest / image registry / portal publication)
Release engineering community manifest + Release Info portal + build-infra Full-text retrieval, page-by-page cross-check release-2.2.yaml first entry; portal rebuilt with 20 platform rows; Megatron images 2.2.0-0.2.3 cover 14 rows
News Google News RSS (24 sets of Chinese/English query terms, via proxy) 24-hour window filtering + item-by-item exclusion Zero hits on component terms (18th quiet window); ecosystem-side entries retained for AICC2026 / Yunqi / Loongson, etc.
Media Guangming Online / The Beijing News / CBISMB / Guandian / National Business Daily / East Money / ITHome Full-text retrieval and review AICC2026, Zhenwu V900, Loongson platform
Community HN Algolia / BAAI Community Search and review No new FlagOS-related posts
Project docs community // schedule.md, flagos-ai.github.io/release-info Full-text retrieval 2.2 schedule (rc2 through 09-24, GA 09-28) cross-checked against image matrix

Complete Source List