Research window: 2026-08-24 10:18 ~ 2026-08-25 10:18 Beijing Time Sources: GitHub (org: flagos-ai 52 repos + commit search + PR/Release API), Google News RSS (9 query sets in Chinese and English, proxy chain operational), HN Algolia, Tavily cross-validation (see appendix)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 vllm-plugin-FL v0.3.0-rc0 released — first new version-line candidate after FlagOS 2.1 (window headline) (08-24)
    • 1.2 build-infra: vllm 0.20.2 validation matrix expanded — Moore Threads F/T dual paths fully passing, MetaX app image landed (08-24/08-25)
    • 1.3 build-infra: validation-driven hardening — “pass” semantics upgraded to true E2E (08-24/08-25)
    • 1.4 Torch-FL: torch.compile three backends landed (MetaX/MUSA/Enflame GCU) (08-24)
    • 1.5 FlagGems: KernelGen Nvidia operator expansion 13+, Ascend/Moore Threads optimization (08-24/08-25)
    • 1.6 FlagTensor/FlagDNN: Iluvatar backend adaptation initiated (08-24/08-25)
    • 1.7 Other repository activity (sglang-plugin-FL Enflame backend, Qwen3.5 compatibility, FlagTree, FlagSparse, FlagOS-Compressor)
  • II. News Coverage and Ecosystem
    • 2.1 Overall news-side assessment (sixth consecutive day of zero component hits)
    • 2.2 Trend over the past 3 days (corroborated by GitHub org pushed)
    • 2.3 Exclusions
  • III. Member Company Deep Dive
  • IV. Summary
  • Appendix: Complete Source List

1. Open-Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 13 had pushes during the window; commit search returned 59 unique in-window commits (FlagGems 20, build-infra 19, Torch-FL 6, FlagSparse 6, FlagTensor 3, vllm-plugin-FL 2, FlagTree 1, FlagOS-Compressor 1, sglang-plugin-FL 1). The window’s headline is the vllm-plugin-FL v0.3.0-rc0 release — the first new version-line candidate since FlagOS 2.1 (6/24), targeting vLLM 0.24.0; this window’s main thread is “inference-side validation matrix expansion (Moore Threads and MetaX join the ranks) + torch.compile multi-backend rollout (MetaX/MUSA/GCU) + new Iluvatar backend adaptation.”

1.1 vllm-plugin-FL v0.3.0-rc0 Released — First New Version-Line Candidate Since FlagOS 2.1 (Window Headline)

Source: Release v0.3.0-rc0 (released 8/24 11:45 Beijing time, target main, no prerelease flag)

  • First v0.3.0-line candidate release: Since vllm-plugin-FL v0.2.0 shipped alongside FlagOS 2.1 on 6/24, this is the component’s first new version-line candidate. The 0.3 line targets vLLM 0.24.0 (echoing the 8/23-24 Kunlunxin 0.24.0 validation sequence on the same line); the release action landed in this window (tag created 8/21, officially released 8/24 11:45).
  • Supporting activity: #403 version doc (8/24 14:26) adds documentation for the release; #383 Qwen3.5 text-only runtime compatibility (merged to main 8/24 15:23) — adds text-only runtime support for Qwen3.5 dense and MoE on the v0.3.0-dev branch (targeting vLLM 0.24.0): upstream vLLM already contains the causal model class but does not fully register the canonical text config, hybrid cache helpers, and VL-prefix checkpoint mapping; the compatibility layer is installed by the generic plugin without modifying the vLLM package or the compiled _moe_C extension. Because it merged after the rc0 cut point (11:45), it is expected to land in a subsequent rc.
  • Continued watch item: the Kunlunxin decode scale fix PR #400 (base release/0.2 line) remains open — when patch_decode_attention routes decode to prefill_attention, it passes the raw decode scale (1/sqrt(head_size)) rather than the adjusted scale it expects (scale*sqrt(head_size)); the roughly 11x-too-small alpha flattens the softmax, causing garbled Qwen3-4B decode output on Kunlunxin P800 XPU; the fix multiplies by sqrt(head_size) (scale=1.0 when head_size=128).

Interpretation: The 0.3.0-rc0 release cadence dovetails directly with the 8/23-24 validation wave — Kunlunxin 0.24.0, NVIDIA cuda13.3, and other backend validations are all based on the 0.24 line, and rc0 promotes that line to an official candidate. Qwen3.5 compatibility (#383) merged to main shortly after, showing that “Day0 compatibility for new models” remains vllm-plugin-FL’s core value proposition (echoing the eight-chip Day0 adaptation for DeepSeek-V4 in April).

1.2 build-infra: vllm 0.20.2 Validation Matrix Expands — Moore Threads Passes Both F/T Paths, MetaX App Image Lands

Source: #506, #508, #510, #512, #514, #516, #521, #522 (8/24 14:58-21:55), #523 (8/24 22:37), #524 (8/25 08:43)

  • Moore Threads (MThreads) passes both F/T paths on both backends: #506 enables vllm0.20.2 app image builds for mthreads-musa4.3.6 and mthreads-musa5.2.0 (the deps_app.vllm0.20.2 key is the matrix gate for the app-image workflow); #508 records that both backends pass all FlagTree (F) and Triton (T) dual-path serve E2E tests on app image 2.1.2-0.2.1 (Qwen3-4B), and corrects the earlier “triton absent” false report (triton 3.6.0 is side-installed at /opt/triton and only becomes visible to importlib.metadata after compiler()).
  • MetaX app image lands: #510 enables 0.20.2 builds for the metax backend; #512/#514 record app image tag 2.1.2-0.2.1_g825c1cd for metax-maca3.7.2.1 and metax-maca3.8.1.3; #516 sets both deps_app and image_tag in the record step; #521/#522 bring vllm-plugin-fl into the 0.20.2 full-install path and derive plugin versions per backend from image_tag.
  • Wrap-up and supporting work: #523 closes the experiment record for the Kunlunxin triton attention bug (experiment closure, corresponding to upstream PR #268); #524 bakes torchvision/torchaudio into runtime deps for the Ascend megatron line.

Interpretation: The vllm 0.20.2 validation matrix has expanded from three vendors (NVIDIA, Kunlunxin, Ascend) to five — Moore Threads becomes the fourth backend to complete full dual-compiler validation (with both musa4.3.6/5.2.0 generations turning green on both backends simultaneously), and MetaX’s app images for two maca versions have been built and tagged into the matrix (F/T pass status not yet explicitly recorded, so the process is still in progress). The “inference-side validation matrix turning green collectively” thread continues to spread from yesterday’s Ascend to Moore Threads and MetaX; the templatization of the validation paradigm means a new backend goes from build enablement to recorded validation in hours.

1.3 build-infra: Validation-Driven Hardening — “Pass” Semantics Upgraded to Real E2E

Source: #517 (8/24 19:20), #518 (8/24 20:22), #519 (8/24 20:34), #520 (8/24 20:47), #525 (8/25 09:14)

  • #517: Validation marker semantics upgraded. Previously the “pass” marker in the matrix was determined solely by the verify script’s exit code, which in effect only proved “install succeeded + import didn’t blow up”; this change wires each backend’s validated E2E recipe into the script, so “pass” = the workload actually ran through with exit code 0 (the serve test in verify-vllm-backend.sh switches to --app-image mode and fails if MODEL_PATH is missing).
  • #518: The plan task previously collected only the default compiler column for each cell to be validated (F if FlagTree was available, otherwise T); now both F/T columns are collected explicitly (both verify scripts support --compiler flagtree|triton).
  • #519/#520: The results branch is rebuilt from a main snapshot (no more rebase); a single failing cell fails the entire verify job — errors are no longer silently swallowed.
  • #525: The serve readiness check changes from a fixed 60s sleep to polling readiness — validation is faster and more reliable.

Interpretation: This is the companion closure to yesterday’s status-matrix system (YAML-ization + automatic write-back on release) — hardening the decision logic of the “validation state machine” itself: marker semantics move from “install-level” to “workload-level,” both compiler columns are forced to be explicitly validated, and failures cannot be masked. As multi-chip validation scales up, the credibility of validation results becomes build-infra’s next engineering priority.

1.4 Torch-FL: torch.compile Lands on Three Backends (MetaX/MUSA/Enflame GCU)

Source: #158 (8/24 17:28), #159 (8/24 20:32), #160 (8/24 16:59), #162 (8/24 20:27), #164 (8/24 21:03)

  • #158: MetaX torch.compile validation lands (PR filed by Claude Code, reviewed by human reviewer @zhaoyinglia): the compile integration suite passes on a real FlagTree MetaX build, fixing two Inductor autotuning failures that only surfaced on the official CPU torch wheel — the MetaX stream shim breaking Event.record(), and the maca Triton backend name leaking into torch.device(); it also exposes the in-image vendor Triton to the CI venv and adds regression tests.
  • #159: MUSA torch.compile binds directly to torch_fl: registers torch.compile(backend="flagos") for ACCELERATOR=musa, implemented via the Moore Threads FlagTree runtime. There was a tradeoff along the way: an initial version used a fabricated compatibility module to satisfy the driver’s import torch_musa, which the maintainer rejected, requiring FlagTree to bind directly to torch-fl before merging.
  • #160: Enflame GCU S60 AMP validation: platform constraints were first measured on S60 hardware before making changes; tests/integration/test_amp.py went from all-skipped to 25/25 passing, with the shared AutocastPrivateUse1 strategy taking effect.
  • #162/#164: A dedicated wheel CI pipeline for Enflame; CUDA CI also incorporates torch.compile tests.

Interpretation: The torch.compile path is moving from “single-backend experiment” to “standard multi-backend capability” — MetaX, MUSA, and Enflame GCU all landed/validated on the same day, all based on real hardware testing. The collaboration model of AI agents (Claude Code) filing PRs with human reviewers gating them has become the norm in Torch-FL (#158/#159/#160 all follow this pattern), a new characteristic of FlagOS’s open-source collaboration workflow.

1.5 FlagGems: KernelGen Nvidia Operator Expansion 13+, Ascend/Moore Threads Optimizations

Source: KernelGen Nvidia operator series (#3408/#4008/#5355/#5491/#5653/#5291/#5367/#5519/#5114/#5116/#5117/#5118/#5211, 8/24 14:07-8/25 07:01), #5383, #5675, #5684

  • Batch expansion of KernelGen Nvidia operator coverage (13+ Triton operators): special function family (special_bessel_y1, special_logit, special_log_ndtr), RNN/LSTM family (mkldnn_rnn_layer single-layer LSTM, _cudnn_rnn_backward single-layer unidirectional LSTM backward), activation backward family (hardsigmoid/hardswish/hardtanh backward), quantization family (quantized_lstm, alpha_dropout_), tensor operations (fliplr, _upsample_nearest_exact1d_backward, sym_constrain_range). This is KernelGen’s (automatic operator generation) continued expansion of torch operator coverage.
  • Ascend: #5383 optimizes AddMM’s layout and bias epilogue; a fix series for torch.linalg.matrix_norm calls under the KMcompiler path (#5707/#5708/#5712, 8/24 15:38-17:29).
  • Moore Threads: MTHREADS backend constant_pad_nd override optimization (#5675) and FillCopy path optimization (#5684).

1.6 FlagTensor/FlagDNN: Iluvatar Backend Adaptation Begins

Source: FlagTensor iluvatar commit (8/24 23:33), FlagTensor #15 (merged 8/25 10:03), FlagDNN Iluvatar WIP (8/24 12:09-14:18)

  • FlagTensor (Triton tensor primitive library: unary 28 / binary 4 / contraction 6, benchmarked against cuTensor, built on FlagTree): fixes Triton 3.6 compatibility (flagtree 0.6.1 + iluvatar3.6) and adds tune_configs for Iluvatar; #15 merges the iluvatar branch into main (8/25 10:03).
  • FlagDNN: WIP Iluvatar backend adaptation (8/24 12:09) + sync merge (#9, 8/24 12:20), plus an ascend fix (8/24 11:29).

Interpretation: Iluvatar becomes the window’s newly active backend — FlagTensor and FlagDNN both began adaptation for it simultaneously, and together with the push cadence of repos such as FlagPrism on 8/21, domestic chip onboarding continues along the “light up the component matrix one by one” path.

1.7 Other Repository Activity

  • sglang-plugin-FL: #42 feat: add enflame gcu (8/24 15:01) — the sglang plugin adds the Enflame GCU backend, tested with qwen3.6-27b and qwen3.6-35b-a3b.
  • FlagTree: #1034 HINT backend selection fix (8/24 11:08); TLE C++ module generates python stubs and adds type annotations (8/24 11:18); #1040 PPU AABS min size fix (8/24 23:43).
  • FlagSparse: DCU test robustness improvements (test robust up) + #47 dcu bsr spmv/spmm per-matrix result validation (8/24 14:46-21:02).
  • FlagOS-Compressor: #6 calibrated GPTQ/AWQ + native AutoRound (8/24 12:55) — pure PyTorch AutoRound supports symmetric W4A16/W8A16 (SignSGD, min/max tuning, optimal parameter recovery, cascaded quantization input), decouples quantization algorithms from checkpoint packaging, and generalizes device backends to CUDA/NPU/MLU/MUSA plus registered custom devices.
  • docs / release-info / FlagCX: pushed_at falls within the window but no in-window merges to the main branch (non-default branch syncs/routine updates), consistent with yesterday’s pattern.

II. News Coverage and Ecosystem

2.1 Overall News Assessment

The Google News RSS proxy pipeline is functioning normally. Across 9 sets of Chinese and English queries (FlagOS/FlagGems/FlagScale+FlagTree+FlagPerf/BAAI open source/BAAI when:3d/BAAI open source, etc.), there were zero direct hits on component keywords within the window — the sixth consecutive day of quiet at the component level (8/20–8/25), consistent with historical patterns. HN Algolia returned only irrelevant false matches (Xiaomi Xring SoC, ID scanner false positives, etc.), which were excluded. Hits within the BAAI-related window were all generic AI news aggregated and reposted by the BAAI community (Anthropic employment impact, GPT-5.6 price cuts, Google’s TPU father departing, DP Technology’s research platform, etc.), with no direct connection to FlagOS. The “Lingjing Zhiyuan WRC 2026 Embodied Brain” series was verified to refer to Shanghai Lingjing Zhiyuan Technology Co., Ltd. (founded by Sun Bo, established 2025-07) — not BAAI — and was excluded (see 2.3). There are no new items eligible for inclusion on the news side; the main body of this report is GitHub activity.

Date Active Repos Main Content
8/22~8/23 (previous window) build-infra(6), Megatron-LM-FL (integration branch), FlagSparse(3) Cambricon dual-backend E2E (shim removal); Kunlunxin 0.20.2 dual-compiler E2E; status-matrix system launched; Ascend vllm 0.20.2 dual-compiler F/T all passing
8/23~8/24 (previous window) build-infra(18), Torch-FL(1) Kunlunxin vllm 0.24.0 dual-compiler verified; NVIDIA cuda13.3 0.20.2 verified; status-matrix YAML-ified; vllm-plugin-FL pure-Python build unified
8/24~8/25 (this window) build-infra(19), FlagGems(20), Torch-FL(6), FlagSparse(6), FlagTensor(3), vllm-plugin-FL(2), FlagTree(1), FlagOS-Compressor(1), sglang-plugin-FL(1) vllm-plugin-FL v0.3.0-rc0 released; Moore Threads vllm 0.20.2 F/T dual-path all passing (dual musa backend); MetaX app image landed; torch.compile landed across three backends (MetaX/MUSA/GCU); FlagGems KernelGen 13+ new operators; Iluvatar adaptation initiated (FlagTensor/FlagDNN)

Trend assessment: The main thread has evolved from “validation matrix collectively turning green” to “validation matrix expansion + validation semantics hardening + new inference/compilation capabilities advancing in parallel.” On the inference side (vllm 0.20.2), the matrix now covers NVIDIA, Kunlunxin, Ascend, and Moore Threads — all four with dual-compiler full passes — plus MetaX app image landing. torch.compile, as a cross-backend compilation capability, landed on the same day across three backends (MetaX/MUSA/GCU). Combined with FlagGems KernelGen operator auto-generation coverage expansion and Qwen3.5 compatibility, all three capability lines of FlagOS — “compilation + operators + inference plugins” — are advancing deeper into multi-chip territory. Iluvatar has become a new backend integration point (FlagTensor/FlagDNN dual components). v0.3.0-rc0 is the first new version-line candidate since 2.1; the release cadence is linked to the Kunlunxin 0.24.0 line and the merge status of Qwen3.5 compatibility (#383).

2.3 Exclusions

  • “WRC 2026 Wrap-Up Lingjing Zhiyuan Defines Industry Ceiling with Embodied Brain Foundation” (Ebrun, 8/24), “WRC 2026 Closing: Embodied Brain Becomes Key Foundation for ‘Landing Battle’” (Sohu, 8/25) — “Lingjing Zhiyuan” refers to Shanghai Lingjing Zhiyuan Technology Co., Ltd. (embodied brain T300/T81, CEO Sun Bo), not BAAI; the name contains “Zhiyuan,” which triggered gnews matching, and was excluded after verification.
  • BAAI community aggregated reposts within the window (Anthropic employment data, GPT-5.6 Sol price cuts, Google’s TPU father departing for Anthropic, DP Technology research platform, UBTech WRC, etc., multiple items on 8/24) — the BAAI community is an AI news aggregation platform; these items are generic AI industry news with no direct connection to FlagOS.
  • “Frontier AI Risk and Emergency Response Research Report (Draft for Comments)” (BAAI community, 8/24) — a BAAI safety research output, not a FlagOS system software stack development.
  • Gambling SEO pollution (“BAAI releases and open-sources new Kaiyuan chess and card games,” “sports betting app download,” etc., womenofchina.com, 8/21–8/24) — excluded in bulk by title keywords.
  • Qwen3.8-2.4T nine-chip adaptation coverage (8/14, BAAI community), AI Compiler event recap (8/20) — old news outside the window.
  • HN hits (Xiaomi Xring SoC, ID scanner false positives, Talk Like Claude Day) — irrelevant false matches with no connection to FlagOS.

III. Deep Dive on Member Organizations

Multi-vendor activity map within the window:

Vendor Chip/Backend Activity within window Evidence
BAAI — build-infra leads vllm matrix expansion (Moore Threads/MetaX) and validation-driven hardening (37 commits); vllm-plugin-FL v0.3.0-rc0 released; sixth consecutive day of zero component hits on the news side build-infra PR series, vllm-plugin-FL release
Moore Threads MThreads MUSA vllm 0.20.2 app image passes both F/T paths (musa4.3.6 + musa5.2.0, image 2.1.2-0.2.1, #506/#508); torch.compile directly binds torch_fl via FlagTree (#159); FlagGems constant_pad_nd optimization (#5675/#5684) build-infra #506/#508, Torch-FL #159, FlagGems #5675/#5684
MetaX MetaX MACA vllm 0.20.2 app image landed (maca3.7.2.1 + maca3.8.1.3, tag 2.1.2-0.2.1_g825c1cd, #510/#512/#514); MetaX FlagTree torch.compile validation + CI coverage (#158) build-infra #510/#512/#514, Torch-FL #158
Enflame Enflame GCU sglang-plugin-FL adds GCU backend (#42, qwen3.6-27b/35b-a3b tested); GCU S60 AMP validation passes all 25 items (#160); standalone wheel CI pipeline (#162) sglang-plugin-FL #42, Torch-FL #160/#162
Iluvatar Iluvatar New backend adaptation initiated: FlagTensor Triton 3.6 compatibility + iluvatar tune_configs + branch merge (#15); FlagDNN WIP Iluvatar backend (#9) FlagTensor commits, FlagDNN commits
KunlunXin KunlunXin XPU triton attention bug experiment wraps up (build-infra #523, corresponding to upstream PR #268); vllm-plugin-FL #400 decode scale fix still open (release/0.2 line) build-infra #523, vllm-plugin-FL #400
Ascend (Huawei) Ascend megatron runtime deps add torchvision/torchaudio (#524); FlagGems AddMM layout optimization (#5383) + KMcompiler matrix_norm fix (#5707/#5708/#5712) build-infra #524, FlagGems commits
NVIDIA (ecosystem reference) CUDA FlagGems KernelGen 13+ new Triton operators (RNN/LSTM, activation backward, special functions, quantization family) FlagGems PR series
Cambricon / Hygon / Tsingmicro / Horizon MLU / DCU / TXDA / BPU No new merges on the open-source side within the window (Cambricon reported on 8/22; Hygon’s FlagSparse DCU side is test robustness improvement #47, not a new feature) —

Trend assessment: The keyword for the chip landscape in this window is “new backends joining the roster + compilation capabilities landing across backends.” With vllm 0.20.2 passing both F/T paths on dual musa backends, Moore Threads becomes the fourth backend in the matrix to achieve “dual-compiler full pass,” with MetaX close behind in landing its app image; torch.compile was validated on the same day across three backends — MetaX, MUSA, and Enflame GCU — extending “multi-chip coverage of the compilation stack” from training frameworks to the PyTorch compilation path; Iluvatar becomes a new onboarding point (FlagTensor/FlagDNN dual components initiate Iluvatar adaptation). AI agent-assisted development (Claude Code submitting PRs + human reviewers gatekeeping) has become the norm in Torch-FL, and the open-source collaboration model itself is changing.

IV. Summary

  1. vllm-plugin-FL v0.3.0-rc0 released (window headline): The first release candidate of a new version line since FlagOS 2.1 (6/24), published 8/24 11:45, targeting vLLM 0.24.0; Qwen3.5 text-only compatibility (#383) was merged into mainline after the rc0 cut point and is expected to land in a subsequent rc; the Kunlunxin decode scale fix (#400) remains open.
  2. vllm 0.20.2 verification matrix expands: Moore Threads’ dual backends (musa4.3.6/5.2.0) pass all F/T dual paths, making it the fourth backend to pass fully under both compilers (#506/#508); MetaX’s dual maca version app images land (#510/#512/#514). “New backend from build enablement to recorded verification” has been templated into an hour-scale process.
  3. Verification-driven hardening: The semantics of “pass” are upgraded from install-level to “real E2E workload runs through with exit code 0” (#517); F/T dual-compiler columns are mandatorily collected explicitly (#518); a failing cell fails the verify job (#520); serve readiness polling replaces fixed sleep (#525) — multi-chip verification status enters a “trustworthy and auditable” phase.
  4. torch.compile lands on three backends: MetaX (#158), MUSA (#159, FlagTree directly binding torch_fl), Enflame GCU AMP (#160, all 25 tests pass), plus Enflame’s standalone wheel CI (#162); the collaboration model of Claude Code submitting PRs with human reviewer gatekeeping has become the norm.
  5. Iluvatar (Tianshu Zhixin) new backend adaptation begins: FlagTensor (Triton 3.6 compatibility + tune_configs, branch merged #15) and FlagDNN (WIP backend) advance in parallel.
  6. FlagGems KernelGen operator coverage expands: 13+ new operators on the Nvidia side (RNN/LSTM forward and backward, activation backward family, special functions, quantization family), Ascend AddMM/KMcompiler fixes, Moore Threads pad optimization; sglang-plugin-FL adds a new Enflame GCU backend (#42); FlagOS-Compressor adds native AutoRound quantization (#6).
  7. News side: sixth consecutive day with zero component hits: no 2.x official release signal (build-infra tag still on the v2.1.x line); no ecosystem-side items eligible for inclusion. Next observation points: v0.3.0 official release (including Qwen3.5 compatibility), vllm-plugin-FL #400 merge, MetaX dual-compiler verification status, and completion of FlagDNN/FlagTensor Iluvatar adaptation.

Limitations note: Commit counts and merge times come from commit search and the repos/commits API (the search index may lag slightly; for a few repositories, pushed_at differs from main-branch commits due to branch synchronization, which was verified using the commits API); build-infra E2E environment details are based on PR descriptions; MetaX’s “app image landing” does not imply any inference about its F/T verification completion status; the v0.3.0-rc0 tag creation time (8/21) precedes the release action (8/24) and has been distinguished as such; gnews redirect links cannot be fetched for full text, so news-side judgments rely on titles and cross-corroboration from multiple sources, and low-weight Chinese sources may have been missed.

Appendix: Complete Source List

No. Event Source Link
1 vllm-plugin-FL v0.3.0-rc0 released https://github.com/flagos-ai/vllm-plugin-FL/releases/tag/v0.3.0-rc0
2 vllm-plugin-FL Qwen3.5 text-only compatibility (#383) https://github.com/flagos-ai/vllm-plugin-FL/pull/383
3 vllm-plugin-FL Kunlunxin decode scale fix (#400, still open) https://github.com/flagos-ai/vllm-plugin-FL/pull/400
4 build-infra Moore Threads vllm 0.20.2 F/T dual-path all passing (#506/#508) https://github.com/flagos-ai/build-infra/pull/508
5 build-infra MetaX vllm 0.20.2 app image (#510/#512/#514) https://github.com/flagos-ai/build-infra/pull/512
6 build-infra verification driver hardening (#517/#518/#519/#520/#525) https://github.com/flagos-ai/build-infra/pull/517
7 build-infra Kunlunxin triton attention experiment wrap-up (#523) https://github.com/flagos-ai/build-infra/pull/523
8 build-infra Ascend megatron runtime deps (#524) https://github.com/flagos-ai/build-infra/pull/524
9 Torch-FL MetaX torch.compile verification (#158) https://github.com/flagos-ai/Torch-FL/pull/158
10 Torch-FL MUSA torch.compile direct binding (#159) https://github.com/flagos-ai/Torch-FL/pull/159
11 Torch-FL Enflame GCU AMP verification (#160) https://github.com/flagos-ai/Torch-FL/pull/160
12 Torch-FL Enflame wheel CI (#162) https://github.com/flagos-ai/Torch-FL/pull/162
13 FlagGems KernelGen Nvidia operator sequence (#5519 and 13+ others) https://github.com/flagos-ai/FlagGems/pull/5519
14 FlagGems Ascend AddMM optimization (#5383) https://github.com/flagos-ai/FlagGems/pull/5383
15 FlagGems MTHREADS pad optimization (#5675/#5684) https://github.com/flagos-ai/FlagGems/pull/5675
16 FlagTensor Iluvatar adaptation (#15) https://github.com/flagos-ai/FlagTensor/pull/15
17 FlagDNN Iluvatar WIP (#9) https://github.com/flagos-ai/FlagDNN/pull/9
18 sglang-plugin-FL Enflame GCU backend (#42) https://github.com/flagos-ai/sglang-plugin-FL/pull/42
19 FlagTree HINT/#1040/TLE stubs https://github.com/flagos-ai/FlagTree/pull/1040
20 FlagSparse DCU bsr result validation (#47) https://github.com/flagos-ai/FlagSparse/pull/47
21 FlagOS-Compressor AutoRound (#6) https://github.com/flagos-ai/FlagOS-Compressor/pull/6
22 org repos overview (52 repos, 13 active within window) https://api.github.com/orgs/flagos-ai/repos?per_page=100&sort=updated
23 commit search (59 unique commits within window) https://api.github.com/search/commits?q=org:flagos-ai+committer-date:%3E2026-08-24T02:18:00Z
24 gnews 9 query sets in Chinese and English (zero component-level hits) https://news.google.com/rss/search?q=FlagOS+when%3A14d&hl=zh-CN&gl=CN&ceid=CN%3Azh-Hans
25 HN Algolia (irrelevant false matches) https://hn.algolia.com/api/v1/search_by_date?query=FlagOS
26 Tavily cross-validation (Lingjing Zhiyuan identity check, no additional FlagOS coverage) https://www.tavily.com