Research window: 2026-09-18 10:18 ~ 2026-09-20 10:18 (approx. 48 hours, including the weekend; continues from the 09-18 report window with no gap) Sources: GitHub (org: flagos-ai, full check of pushed_at across 54 repos; item-by-item verification of 206 commits across 20 repos within the window), Google News RSS (24 sets of Chinese and English query terms, via proxy), BAAI Community, Zhihu institutional accounts, CSDN FlagOS column, etc. (see the appendix source list for details)

Issue Index

  • Today’s Highlights: Full batch authorized rebuild of vLLM images for all 20 build-infra backends; 2.2 release artifacts enter rc2.post2 assembly (09-19/09-20)
  • I. Open Source Project Progress (GitHub Activity)
    • 1.1 build-infra: Authorized rebuild of vLLM images for 20 backends; plugin wheels switch to “build once, distribute to all vendors” (09-19/09-20)
    • 1.2 build-infra: FlagTree pinning policy tightened to 0.6.x; FlagTune cost model disabled per pinned version (09-20)
    • 1.3 FlagGems: KernelGen continues to land; roughly ten batch fixes for Kunlunxin backend in a single day (09-20)
    • 1.4 FlagGems: DAMO XuanTie PPU backend enables TLE; Hygon DCU Flash Attention fixes (09-20)
    • 1.5 FlagSparse: Intensive adaptation push across MUSA / MACA / DCU backends (09-18~09-20)
    • 1.6 FlagQuantum: Major expansion of algorithm primitive package — QFT, phase estimation, Grover, QUBO, and more than ten others (09-18~09-20)
    • 1.7 Torch-FL: Qwen-Image 2.1 / 2512 multi-chip performance optimization pipeline established (09-18~09-20)
    • 1.8 Other Updates: FlagTree, FlagGems-sglang, FlagGems-vllm, FlagFFT, FlagPrism, etc. (09-18~09-20)
  • II. News Coverage and Ecosystem
    • 2.1 Component-level search: fifteenth consecutive quiet window — zero hits over 48 hours (09-18~09-20)
    • 2.2 2026 AI Open Computing Conference & Zhongzhi FlagOS Technology Conference opens registration: October 17-18, Beijing (09-18)
    • 2.3 FlagOS × MiniCPM Model Inference Throughput Optimization Challenge: submissions open September 21 (recently announced)
  • III. Member Company Deep Dive
    • 3.1 Kunlunxin: FlagGems backend fixes land in batch (09-20)
    • 3.2 Moore Threads: MUSA line operators and images advance in parallel (09-20)
    • 3.3 Hygon: DCU Flash Attention fixes and Qwen-Image performance line (09-20)
    • 3.4 DAMO XuanTie: PPU backend enables TLE; KernelGen operators enter mainline (09-19/09-20)
    • 3.5 MetaX: MACA line FlagSparse adaptation and image records (09-19/09-20)
    • 3.6 Iluvatar CoreX: FlagDNN test surface convergence (09-19/09-20)
    • 3.7 Ascend: Image rebuild, dedicated linear operators, and kda_gate operator (09-20)
    • 3.8 Moore Threads / Enflame / MetaX images and tooling surface (09-20)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Appendix: Complete Source List

Today’s Highlight: build-infra batch-authorizes rebuilds for all 20 backend vLLM images; 2.2 release artifacts enter rc2.post2 assembly

Date: 2026-09-19, 2026-09-20　Source: build-infra #938, build-infra #956, build-infra #934

In this window, build-infra became the densest line of work in the org with 43 commits, and its actions were highly concentrated: authorizing the 2.2.0 / plugin post2 rebuild for all 20 backend vLLM images, and writing the rebuild results back one by one. The 09-20 03:07 app changelogs: authorize the 2.2.0 / plugin-post2 rebuild for all 20 images (#938) was the starting gun; then, over the hour-plus from 09-20 03:07 to 10:28, the bot account recorded the new image tags backend by backend; at 09-20 09:43, another commit, authorize the 2.2.0 / plugin-post2 rebuild for all 20 vllm 0.24.0 images, made the scope and baseline of this rebuild round (vLLM 0.24.0) explicit in the title.

The recorded image tags cover the org’s entire backend matrix: the 2.2.0-0.3.0rc2.post2 tier lands on Ascend (cann8.5.0, cann9.0.0, and two 910c variants), Cambricon (neuware4.4.3 / 4.7.2), Moore Threads (musa4.3.6 / 5.2.0), Enflame (tops1.9.10 / 1.10.6), MetaX (maca3.7.2.1 / 3.8.1.3), Hygon (dtk26.04), Iluvatar (corex4.5.0), Kunlunxin (xre5.37.1), Infinigence (tangrt1.2.0), and NVIDIA (cuda12.8 / 13.3); an earlier tier, 0.2.2rc2.post2 (plugin version), had already landed for a batch first. All 20 backend images were refreshed within a single window, making this the largest single action in the org during this window.

Complementing this is a change in how the plugin wheel is distributed: the 09-19 23:34 vllm-plugin wheel: build once, publish to every vendor index (#934) changed “build once per vendor” to “build once, publish to all vendor indexes,” followed by two commits that changed the version number to be derived from the git ref (#932) and fixed a runs-on parameter unpacking issue (#935). This chain compresses plugin distribution from “N builds” to “1 build + N publishes,” a structural optimization in release engineering.

Another parallel wrap-up line turns “what exactly is installed in the image” into a verifiable fact: vllm app image: read the installed plugin version past the entrypoint (#937) lets the image report its own plugin version; docs: print TBD for an app image that was never built (#948) explicitly marks never-built images as TBD in the docs; changelogs: name the FlagTree each 2.2.0 vllm image actually builds on (#953) annotates each image with the actual compilation base. All three reduce the room for “documentation and artifacts being inconsistent.”

Looking at the actions within the window side by side: 2.2 release artifact assembly has reached the stage of “batch rebuild of all backend images + verifiable versions + merged distribution path.” With 8 days left until GA (2026-09-28) on the release schedule, this is a normal pace of progress within the testing and stabilization period.

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 54 repos in the org, 20 had commits in this window, totaling 206 (spanning 48 hours, including the weekend), distributed as: build-infra 43, FlagSparse 39, FlagGems 38, FlagQuantum 23, FlagGems-sglang 10, FlagGems-Experimental 10, Torch-FL 9, FlagDNN 6, FlagFFT 5, docs 4, FlagTree 4, FlagGems-vllm 4, vllm-plugin-FL 2, FlagPrism 2, FlagBLAS 2, FlagCX 1, FlagAttention 1, FlagAudio 1, FlagScale-Agent 1, TransformerEngine-FL 1.

This window’s shape = two threads: “batch assembly of release artifacts” + “backend fix wave”: build-infra drove full-backend image rebuilds and consolidation of the distribution pipeline; on the application side (FlagGems / FlagSparse / Torch-FL / FlagGems-vllm), dense fixes and performance tuning landed across multiple chip backends, with Kunlunxin and Hygon DCU seeing the most concentrated fixes; on the governance side (FlagQuantum algorithm primitives, FlagTree TLE), work continued along the 2.2 roadmap. On 09-19 (Saturday), the repos totaled roughly 12 commits, with development activity continuing uninterrupted across the weekend.

1.1 build-infra: 20 backend vLLM images rebuilt with authorization, plugin wheels switched to “build once, distribute to all vendors” (09-19/09-20)

Date: 2026-09-19, 2026-09-20　Source: build-infra #938, build-infra #934, build-infra #939

See “Today’s Highlights.” Two additional points: first, the cann9.0.0 backend added the cann-shmem package at 04:00 on 09-20 (#939), filling in the dependency for the shared-memory component of the Ascend 9.0 line; second, Runtime: fetch vendor deps from the vendor index (#930) at 15:47 on 09-18 began sourcing runtime dependencies from the “vendor index,” and together with the removal of old container files at 13:50 on 09-18 (#929), the source and source-of-truth for container builds began converging onto a unified index.

1.2 build-infra: FlagTree pinning tightened to 0.6.x, FlagTune cost model disabled where pinned (09-20)

Date: 2026-09-20　Source: build-infra #940, build-infra #943, build-infra #944, build-infra #947

Three consecutive “pinning” commits landed on the morning of 09-20: nvidia-cuda13.3: hold flagtree at 0.6.1 (#940), metax: hold flagtree at 0.6.1 for both MACA backends (#943), enflame-tops1.9.10: hold flagtree at 0.6.0 (#944) — pinning the compiler foundation of three backends back to the 0.6.x line (a reversal of the earlier version-bump actions); subsequently, runtime: disable the FlagTune cost model where flagtree 0.6.x is pinned (#947) indicated that the FlagTune cost model is incompatible with the 0.6.x pin and must be disabled according to pin status. The three pins plus one compatibility handling form a complete set of “reversal + adaptation.”

There was also one status-matrix correction: status matrix: drop the merged entries from megatron's prs (#933) and status matrix: prs = open PRs, md = recen… (#932, earlier part) hard-coded the status matrix’s statistical definitions (PR column = open PRs, md column = recent commits).

1.3 FlagGems: KernelGen continues to land, roughly ten Kunlunxin backend fixes in a single day (09-20)

Date: 2026-09-20　Source: FlagGems #5917, FlagGems #5878, FlagGems #6317, FlagGems #6476

FlagGems’ 38 commits this period remained underpinned by KernelGen output: on the Nvidia side, two fused optimizer operators were added, _fused_adagrad_ (#5917) and _fused_sgd_ (#5878); the Experimental repo (see 1.8) saw ten more Nvidia operators land that day. The most concentrated thread this period was Kunlunxin backend fixes: binliu landed roughly ten consecutive commits, covering persistent kernel scheduling and shape-based auto-tuning for grouped_mm large-M / small-K shapes, SDPA attention fixes (preserving value-1 parameter non-specialization, converging auto-tuning), mse_loss stage-one switch to full-block masking (#6476), scatter / log_sigmoid_backward / amp_foreach, fill / add_relu / special_log1p / dequantize, pow / pdist / Legendre polynomials, segment_reduce / arcsinh / cosh / acosh / log2 / log10, bessel_y1 / bernoulli, t_copy / selu / reflection_pad2d, log_softmax / logical_not, and a fix for an accidental deletion during lgamma migration — essentially “clearing out” the backend’s recent operator regressions.

Two tooling items: [KMCompiler] operator linalg_lstsq bug fix (#6499), [KMCompiler] fix fp64 error (#6491), and delete uint8 dtype (#6477); on the MThreads side, fixes to feature_dropout_ restoration and splitting (#6490, #6482). Benchmark and CI side: Fix(benchmark): align pytest markers and op_name with operator registry ids (#6489), ci: run pre-check jobs on basic-runner-cpu (#6431).

1.4 FlagGems: DAMO XuanTie PPU backend enables TLE, Hygon DCU Flash Attention fixed (09-20)

Date: 2026-09-20　Source: FlagGems #6423, FlagGems #5822, FlagGems #6456

Three heavyweight backend commits: [T-Head] Enable TLE for PPU backend (#6423) turns on the Triton Language Extension (TLE) on the DAMO XuanTie PPU backend, corresponding to that direction’s place in the 2.2 multi-backend plan; [Hygon] Fix Flash Attention forward/backward kernels on DCU (#5822) fixes the forward/backward Flash Attention kernels on Hygon DCU; [Ascend] Add specialized linear operator (#6456) adds a specialized linear operator for Ascend. Also, [MetaX] Fix masked_fill broadcast OOB and illegal num_warps=16 (#6480) corrects an out-of-bounds and an illegal configuration on MetaX.

1.5 FlagSparse: Dense progress on MUSA / MACA / DCU three-backend adaptation (09-18~09-20)

Date: 2026-09-18 ~ 2026-09-20　Source: FlagSparse #73, FlagSparse #71

FlagSparse’s 39 commits this period came almost entirely from the partner’s dual-branch collaboration flow (two NCIC-AlphaSparse accounts merging alternately), with the theme being three-backend adaptation: the MUSA (Moore Threads) line landed musa updates; the MACA (MetaX) line landed maca spsm debug and maca test commands; the DCU (Hygon) line landed dcu updates, dcu sddmm tests, ci dcu, and dcu spgemm test fp32 fp64 sperated (separating single- and double-precision spgemm tests). There were also cwrapper add (a new C wrapper layer) and one revert of “New update” (#63 rolled back, #64 withdrawn), plus documentation cleanup. Overall, this is the process of making the sparse operator library real across three domestic backends.

1.6 FlagQuantum: Major expansion of the algorithm primitives package — QFT, phase estimation, Grover, QUBO, and more than a dozen others (09-18~09-20)

Date: 2026-09-18 ~ 2026-09-20　Source: FlagQuantum #85, FlagQuantum #91, FlagQuantum #104

FlagQuantum’s 23 commits this period were advanced continuously by a single contributor (Wei LIU), with the theme being the systematic laying down of the algorithm primitives package: starting from the algorithm primitives package and the quantum Fourier transform (#85), then successively landing the operator protocol and phase estimation (#87), state preparation (#88), multi-controlled X and reversible bit-string comparator (#89), truth-table oracle synthesis (#90), Grover search (#91), and amplitude estimation (#92); then moving into the application layer — QUBO-to-Ising mapping (#86), quantum kernel estimation and classical kernel ridge classifier (#100), phase estimation based on quantum principal component analysis (#98), Grover quantized k-medians (#99), QUBO feature selection (#101), amplitude estimation for frequent-item proportions (#102), and phase estimation for estimating singular values (#104). Two engineering-quality items: coverage and type-checking gates continued the recent “tightening” trajectory; on the functional side, one item was the detector error model for the quantum error correction (QEC) record layer (#84, corresponding to stage 2a of proposal 048). Beyond the 01-09 layer, there was also one operational item from the build governance side: limiting the repo’s share of org-level runners (same topic as recent CI resource governance).

1.7 Torch-FL: Qwen-Image 2.1 / 2512 multi-chip performance optimization becomes a thread (09-18~09-20)

Date: 2026-09-18 ~ 2026-09-20　Source: Torch-FL #360, Torch-FL #356, Torch-FL #353

Torch-FL’s 9 commits this period were threaded by a single contributor (nate.river) around Qwen-Image series performance across multiple chips: first establishing a manual test flow with test: add the Qwen-Image-2.1 manual test flow for chip bring-up (#342), then optimizing continuously — fixing DCU’s SDPA backend selection and pointwise routing (#345), reducing DCU-side FlagGems scheduling / layout / auto-tuning overhead (#352), bringing GCU’s Qwen-Image-2512 single step down to 7.1 s/it (#354, SDPA routing + GEMM operand handling), wiring Qwen-Image-2.1’s key-valid mask into the FlagGems SDPA path (#353), further reducing DCU’s FlagGems scheduling overhead (#356), and finally producing Qwen-Image-2.1’s throughput curve and tail-latency / arithmetic-intensity measurements (#360). This is a typical chip bring-up performance thread going “from runnable to running well.”

1.8 Other Activity: FlagTree, FlagGems-sglang, FlagGems-vllm, FlagFFT, FlagPrism, etc. (09-18~09-20)

Date: 2026-09-18 ~ 2026-09-20　Source: FlagTree #1084, FlagGems-sglang #93, FlagGems-Experimental #659, FlagPrism #13

  • FlagTree (4): [KMCompiler][TLERaw] Optimize the sort operator based on tle_raw (#1084) continues the TLE Raw trajectory; [TLE][CommonIR] Added CommonIR conversion integration on Triton 3.6 (#1160) integrates CommonIR conversion into Triton 3.6; [QC][TLE] Support cross-dtype Hopper WGMMA accumulator reuse (#1001); [XPU] Guard do_bench against estimate_ms==0 (#1224) fixes a divide-by-zero.
  • FlagGems-sglang (10): the moe_fused_mul_sum competition operator landed and was registered (#87); a batch of batch3 reference, tests, benchmarks, and docs (#93); do_bench changed to a single vendor-dispatched helper function (#94); Ascend fused router K-loop pipelining limited to 2 stages (#96).
  • FlagGems-Experimental (10): the KernelGen Nvidia line landed ten more operators that day — sparse_dim, _has_same_storage_numel, _dimV, _values, col_indices, adjoint, atleast_2d, detach_copy, ccol_indices_copy, crow_indices_copy (#640~#659), mostly small sparse and view-type operators, a precursor step in ferrying them from Experimental to the mainline.
  • FlagGems-vllm (4): MUSA side optimized deepseek_v4’s fused_q_kv_rmsnorm (#825); the MetaX line optimized varlen flash attention (#808); the Ascend line added the kda_gate_cumsum operator (#809); the KMCompiler line added a vendor-native precision gate to the fused_moe precision benchmark (#824).
  • FlagFFT (5): continuous refactoring of the acceptance framework — captured data flows through a pipeline with the result tree flattened, operator logs conclude with platform-readable verdicts, input writers no longer wait on readers, and captured inputs are regenerated on demand (no longer resident on disk).
  • FlagPrism (2): unified operator acceptance tests and retired the old batch suite (#13); added Enflame debugger and TOPSPTI profiling support (#12).
  • docs (4): online lab help content updates (two rounds, #509, #510).
  • FlagCX (1): added Iluvatar support to the device API default path (#608).
  • vllm-plugin-FL (2): capture common attention metadata in the accelerator graph (#442); Iluvatar BI-V150 adapted to the vLLM 0.24 line (#526).
  • FlagDNN (6): test fixes across the ppu / mthreads / ascend / iluvatar platforms and run_test.py fixes; Ascend platform architecture improvements.
  • FlagBLAS (2): fixed thead’s pyproject.toml.
  • FlagAttention (1): one fix commit (#63).
  • TransformerEngine-FL (1): cherry-pick from the main line (#124).
  • FlagScale-Agent (1): hardened the agent harness — REPL prompt watchdog, cross-session proposal registry, etc.
  • FlagAudio (1): added a Nexus upload path to CI (#8).

II. News Coverage and Ecosystem

2.1 Fifteenth Consecutive Quiet Window for Component-Level Search: Zero Hits Over 48 Hours (09-18~09-20)

Date: 2026-09-18 ~ 2026-09-20　Source: Google News RSS (24 sets of Chinese and English query terms, via proxy)

Searches combining component names — FlagOS / FlagGems / FlagScale / FlagTree / FlagPerf / FlagCX / KernelGen — with terms such as BAAI returned zero hits within the 48-hour window (the gnews script measured 0 items), marking the fifteenth consecutive quiet window. Silence on the component-level news front has become the norm; this report’s external activity facts are based on the GitHub org and community channels (BAAI Community / Zhihu institutional account).

2.2 2026 AI Open Computing Conference and Zhongzhi FlagOS Technical Conference Opens Registration: October 17-18, Beijing (09-18)

Date: 2026-09-18 (published)　Source: Zhihu institutional account post

The Zhongzhi FlagOS community, together with the Beijing Academy of Artificial Intelligence (BAAI), the Beijing Academy of Quantum Information Sciences, and the Zhongguancun AI Open Source Alliance (in preparation), will hold the “2026 AI Open Computing Conference and Zhongzhi FlagOS Technical Conference” on October 17-18, 2026 at the Exhibition Center of Zhongguancun National Independent Innovation Demonstration Zone in Beijing, under the theme “One Stack End-to-End, Boundless Intelligent Computing.” The conference agenda covers heterogeneous computing foundations, open-source AI system software stacks, quantum-intelligence integration, in-vehicle intelligence, embodied intelligence, and AI agent infrastructure; hands-on workshops span the full chain, including the KernelGen high-performance operator auto-generation engineering, FlagOS-empowered on-device LLM inference, Triton-TLE novel Attention kernel optimization, and FlagScale-Agent architecture design and practice. Registration is now open (limited-time early-bird discount). This is the most important external event milestone for the FlagOS ecosystem in the near term, and its workshop topics correspond closely to this issue’s engineering-side observations (TLE, KernelGen, on-device, Agent).

2.3 FlagOS × MiniCPM Model Inference Throughput Performance Optimization Challenge: Submissions Open September 21 (recently announced)

Date: Recently announced (submission phase begins 2026-09-21)　Source: Zhihu institutional account post

FlagOS and MiniCPM are jointly hosting a model inference throughput performance optimization challenge, with the development and submission phase running from September 21 to November 20, 2026 (UTC+8) and judging in early December. The challenge targets inference throughput optimization on the FlagOS stack, continuing the community competition series after the SGLang cross-chip operator optimization contest (the “operator island” battle).

III. Member Unit Deep Dive

3.1 Kunlunxin: FlagGems backend fixes land in batch (09-20)

Date: 2026-09-20　Source: FlagGems #6317, FlagGems #6476

The Kunlunxin line landed roughly ten fixes in this window (see 1.3), covering everything from basic elementwise operators to core operators such as grouped_mm / SDPA. The most substantive is a persistent kernel specialized for the “large M / small K” shape of grouped_mm, along with shape-based autotuning—specialized for the typical shapes of MoE scenarios. Combined with the moe_fused_mul_sum competition operator landing on the SGLang side, the MoE inference path for Kunlunxin continues to be solidified.

3.2 Moore Threads: MUSA line advances on both operators and images (09-20)

Date: 2026-09-20　Source: FlagGems-vllm #825, FlagSparse #73

Two items on the MUSA line: in FlagGems-vllm, fused_q_kv_rmsnorm for deepseek_v4 is optimized for Moore Threads (#825); and the MUSA adaptation update in FlagSparse. On the image side, the 2.2.0 images for the musa4.3.6 / musa5.2.0 backends completed rc2.post2 rebuild records. In addition, the MThreads KernelGen operator batch in FlagGems-eXperimental (#6458) was ferried from Experimental to mainline on 09-19.

3.3 Hygon: DCU Flash Attention fix and Qwen-Image performance line (09-20)

Date: 2026-09-20　Source: FlagGems #5822, Torch-FL #356

Two hard tasks for Hygon this period: FlagGems fixes the Flash Attention forward/backward kernel on DCU (#5822, matured over a long cycle through PR 5822); Torch-FL reduces FlagGems scheduling overhead for Qwen-Image-2.1 on DCU (#356). On the image side, hygon-dtk26.04 completed the 2.2.0 rc2.post2 rebuild.

3.4 DAMO XuanTie: PPU backend enables TLE, KernelGen operators enter mainline (09-19/09-20)

Date: 2026-09-19, 2026-09-20　Source: FlagGems #6423, FlagGems #6460

Two items for XuanTie: [T-Head] Enable TLE for PPU backend (#6423) turns on the Triton language extension in the PPU backend—TLE is the unified optimization language for cross-hardware families in the 2.2 multi-backend plan, and T-Head is now on that path; the KernelGen cherry-pick on 09-19 ferried T-Head operators from Experimental back to mainline (#6460). This forms a continuous follow-up to the “T-Head enters the FlagTree 2.2 multi-backend plan” observed in the 09-18 issue.

3.5 MetaX: MACA line FlagSparse adaptation and image records (09-19/09-20)

Date: 2026-09-19, 2026-09-20　Source: FlagSparse #71, build-infra #943

MACA line: FlagSparse’s maca spsm debug and test commands advance; build-infra unifies the FlagTree pinned versions of the two MACA backends at 0.6.1 (#943); the maca3.7.2.1 / 3.8.1.3 images completed 2.2.0 rebuild records. MetaX’s varlen flash attention optimization on the FlagGems-vllm side belongs to the same line.

3.6 Iluvatar: FlagDNN test surface converges (09-19/09-20)

Date: 2026-09-19, 2026-09-20　Source: FlagDNN commits

FlagDNN’s six commits this period focus mainly on converging the test surface: test fixes for the four platforms iluvatar / ascend / mthreads / ppu and a run_test.py fix, plus one Ascend platform architecture improvement. The corex4.5.0 image completed a 2.2.0 rebuild record. On the FlagCX side, Iluvatar’s Device API default path support (#608) and vllm-plugin-FL’s BI-V150 vLLM 0.24 adaptation (#526) landed in the same window.

3.7 Ascend: image rebuilds, dedicated linear operator, and kda_gate operator (09-20)

Date: 2026-09-20　Source: FlagGems #6456, FlagGems-vllm #809, build-infra #939

Ascend line: FlagGems adds a dedicated linear operator (#6456); FlagGems-vllm adds the kda_gate_cumsum Triton operator (#809); FlagGems-sglang fixes the fused router K loop pipeline (#96); on the image side, four Ascend variants (cann8.5.0 / cann9.0.0 × 910c and non-910c) completed rc2.post2 rebuilds in batch, with the cann-shmem package added (#939).

3.8 Enflame / Zhonghao Xinying / NVIDIA images and tooling (09-20)

Date: 2026-09-20　Source: FlagPrism #12, build-infra commits

Enflame line: FlagPrism adds Enflame debugger and TOPSPTI profiling support (#12); the tops1.9.10 / 1.10.6 images completed rebuilds, with tops1.9.10’s FlagTree pinned version rolled back to 0.6.0 (#944). Zhonghao Xinying tangrt1.2.0 and NVIDIA cuda12.8 / 13.3 images completed rebuild records in sync; NVIDIA cuda13.3’s FlagTree pinned version rolled back to 0.6.1 (#940).

IV. Summary and Trend Observations

  • Releases enter the all-backend batch rebuild phase: all 20 backend vLLM images were refreshed to rc2.post2 within a single window, and three things landed simultaneously—images self-reporting versions, docs marked TBD, and each annotated with its compilation base—making “verifiable releases” the focus of this round of engineering, 8 days from GA (09-28), a normal pace.
  • The shift in plugin distribution paradigm is worth noting: “build once, publish to all vendor indexes” (#934) compresses N builds into 1 build + N publishes; this kind of infrastructure-level simplification will directly reduce the release cost of every subsequent version.
  • A dense period of multi-chip backend fixes: Kunlunxin (about ten), Hygon (Flash Attention + Qwen-Image), MetaX (SDPA fix), and Ascend (dedicated linear operator) all concentrated fixes in the same window, reflecting that the 2.2 test period (09-01 to 09-24) has entered the “clearing accounts by backend” stage; Kunlunxin’s handling of grouped_mm large M / small K specialization is the most direct evidence of MoE-scenario targeting.
  • FlagQuantum’s expansion rate is one of the fastest directions in the whole org: a single contributor made 23 commits in 48 hours, laying out a chain from primitives (QFT, phase estimation, Grover, amplitude estimation) to applications (QUBO feature selection, quantum PCA, k-medians), while simultaneously building the QEC record layer—a concrete manifestation of “quantum-intelligence fusion” at the code level.
  • T-Head enters on two lines: enabling TLE in the PPU backend + ferrying KernelGen operators to mainline, consistent with T-Head’s position in the 2.2 multi-backend plan, as XuanTie’s participation in the FlagOS stack continues to deepen.
  • Ecosystem activities enter the conference cycle: registration opens for the FlagOS Technical Conference on October 17-18, and the workshop topics (KernelGen, edge devices, TLE Attention, FlagScale-Agent) correspond one-to-one with prior engineering observations; the MiniCPM Challenge opens on 9/21—ecosystem operations density will rise significantly from late September onward.

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full verification of pushed_at for 54 repos 20 repos active within the window
GitHub commits API for each repo (since=09-18T10:18Z) Pulled per repo, verified item by item 206 commits
News Google News RSS (24 query term sets) gnews_topic.py, 48h window 0 hits (the 15th quiet window)
Community Zhihu org account / BAAI Community / CSDN column Manual search and review Conference registration post (09-18), challenge post (recent)

Appendix: Complete Source List