Research window: 2026-09-02 10:18 ~ 2026-09-03 10:18 Beijing Time Sources: GitHub (org: flagos-ai 52 repos + full commit search of 48 entries + tags/releases date verification), Google News RSS (19 query sets in Chinese and English, proxy chain normal), HN Algolia, BAAI Community (see appendix)

Index

  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 build-infra: vLLM 0.24.0 application image expansion — Cambricon dual-backend/Iluvatar/Enflame registered on the same day (09-02)
    • 1.2 build-infra: sglang 0.5.18 application image landed (Phase 5), import-face shim wheel released (09-02)
    • 1.3 FlagGems: Hygon HCU operator optimization, Iluvatar fixes, KernelGen/KMCompiler operator expansion (09-02/09-03)
    • 1.4 Inference plugin layer: sglang-plugin-FL DeviceInfo serviceization, vllm-plugin-FL GDN precision fix, FlagGems-vllm fix (09-02)
    • 1.5 Training stack and remaining repos: Megatron-LM-FL/TransformerEngine-FL CICD alignment, FlagTensor merged into muxi branch, FlagDNN Hygon CPU reference implementation (09-02)
  • II. News Coverage and Ecosystem
    • 2.1 News side: zero direct hits in 24h, review of the main thread over the past 3 days: 2.2 rc0 verification tag intensive period wrapping up (09-01~09-02)
    • 2.2 BAAI institutional updates: SCO framework AI training program concluded, content community remains active (09-02)
  • III. Member Company Deep Dive
    • 3.1 Enflame: final allotment rate of 0.02455315% announced, lower than Moore Threads and MetaX (09-02/09-03)
    • 3.2 MetaX: Yipingji released its first coverage report, Xiangshan Co.’s acquisition of its distributor draws attention (09-02)
    • 3.3 Cambricon: aggressive recruitment of software talent, a signal of software stack investment (09-02)
    • 3.4 Iluvatar: in-depth interim report review and target price adjustment, CMBI maintains Buy (09-02)
    • 3.5 Horizon and industry perspective: from selling solutions to selling foundations; domestic GPUs enter the large-scale delivery test (09-02/09-03)
  • IV. Summary
  • Appendix: Complete Source List

I. Open Source Project Progress (GitHub Activity)

Window Overview: Of the 52 repos in the org, 16 had pushes during the window. Commit search returned 48 in-window commits (build-infra 24, FlagGems 11, FlagFFT 3, sglang-plugin-FL 2, etc.). Tags/releases API verification shows no new tags and no new releases in the window (the previous round of 2.2 rc0.post1 verification tags landed in bulk from 08-26 to 09-01, already reported). Main threads this window: 1) build-infra expanded the vLLM 0.24.0 application image line from yesterday’s “Ascend / MetaX / Moore Threads” to Cambricon (dual backend), Iluvatar, and Enflame, with application image tags for the major domestic AI chip players landing in bulk the same day; 2) the sglang verification matrix advanced from Phase 0 to Phase 5: 0.5.18 application image + workflow + import-face shim wheel landed; 3) FlagGems multi-vendor backend fixes entered a high-frequency phase (Hygon HCU operator performance optimization, Iluvatar Triton 3.6 fixes, KernelGen/KMCompiler Nvidia operator expansion in parallel); 4) the inference plugin layer continued its “testable, fixable, separably deployable” engineering cadence.

1.1 build-infra: vLLM 0.24.0 application image expansion — Cambricon dual backend / Iluvatar / Enflame registered same day (09-02)

Sources: build-infra #678, #681, #687, #688, #698, #680, #686, #696

  • 09-02 11:54 #678 merged: enables vLLM 0.24.0 application images for three backends — cambricon (Cambricon), enflame (Enflame), and iluvatar-4.4.0 (Iluvatar) — continuing the same image management line from yesterday’s Ascend/MetaX/Moore Threads registration, with noticeably broader member coverage.
  • Four application image tags landed in bulk that day (all on the 2.1.2-0.2.1 application image line, carrying source commit fingerprints and dates): 15:56 Cambricon cambricon-neuware4.7.2 (#681), 20:05 Cambricon cambricon-neuware4.4.3 (#687, dual version), 21:37 Iluvatar iluvatar-corex4.4.0 (#688), 22:39 Enflame enflame-tops1.9.10 (#698).
  • Supporting engineering: 15:45 #680 registers the symm-mem guard upstream PR #434 for iluvatar 4.4.0; 18:19 #686 fixes the image tag determination logic (compares against origin rather than the local image, avoiding false positives); 22:35 #696 adds a record_only mode to the vllm application image workflow (register only, no build).

Interpretation: Following yesterday’s first entry of Ascend (ascend-cann8.5.0) into unified vLLM application image management, this window saw concentrated registration of Cambricon (with dual backends neuware4.7.2/4.4.3), Iluvatar, and Enflame — the multi-vendor inference image matrix for the 2.2 cycle now covers more than 6 members, matching up one-for-one with the “10 chips” landscape in yesterday’s Day0 article; image tags carrying source fingerprints (g{sha}) indicate that registration and builds are now traceable to specific commits.

1.2 build-infra: sglang 0.5.18 application image lands (Phase 5), import-face shim wheel released (09-02)

Sources: build-infra #684, #690, #697, #685, #682

  • 09-02 18:13 #684 merged: adds the sglang 0.5.18 application image, build workflow, and the –app-image verify validation parameter (Phase 5) — the sglang verification matrix initiated yesterday via #670 moved from “status recording” to a “buildable, verifiable” stage, completing 5 phase advances that day.
  • 21:38 #690: releases the sgl_kernel import-face shim wheel — provides an import-face compatibility shim for scenarios lacking the native sgl_kernel package, lowering the installation barrier for sglang plugins on heterogeneous backends.
  • 22:40 #697: merges runtime_common entirely into the wheel’s Requires-Dist; 18:45 #685 fixes a metax (MetaX) guard regression in the repack wheel.
  • 17:21 #682 registers upstream plugin PR status into the sglang status-matrix (paired with yesterday’s real-time rendering mechanism for the PR status column).

Interpretation: Within two days the sglang line advanced from Phase 0 (initiation) to Phase 5 (application image + verification + wheel release), the fastest of build-infra’s three verification matrices; the shim wheel and Requires-Dist merge indicate its release artifacts are beginning to be engineered for compatibility with “non-FlagOS standard environments,” lowering the cost of third-party integration.

1.3 FlagGems: Hygon HCU operator optimization, Iluvatar fixes, KernelGen/KMCompiler operator expansion (09-02/09-03)

Sources: FlagGems #4663, #5783, #5393, #5419, #5426, #4443, #5870, #5886, #5902, #5908, #5898 (CI rule-check phase1/2)

  • Hygon HCU backend: 14:26 #4663 batch-optimizes HCU backend operator performance — following the adaptation density of Hygon DCU covering three models in yesterday’s Day0 article, the operator layer continues to be refined.
  • Iluvatar backend: 15:54 #5783 fixes the restoration of rank-1 vectorization for pointwise operators under Triton 3.6 (a regression fix triggered by a compiler version upgrade).
  • KernelGen/KMCompiler (Nvidia-side operator expansion): 15:32 #5393 slice operator (Triton kernel), 15:33 #5419 ctc_loss operator YAML registration, 15:53 #5426 dsplit operator (view implementation), 09-03 09:23 #4443 aten._unsafe_index operator, 09-03 09:31 #5870 linalg_norm operator — operator coverage expands in parallel across both the KernelGen and KMCompiler compilation paths (after the “8 new operators” from the Day0 article two days ago landed, operator registration has entered a routine expansion cadence).
  • Quality and benchmarks: 10:52 #5886 adds INT8 tests for (un)pack_seq_triton (echoing the FlagOS-Compressor INT8 quantization line); 12:16 #5902 isolates MM operator measurement from the benchmark (eliminating cross-talk); 15:02 #5908 enforces _FULL_CONFIG ordering; 11:02 CI rule-check phase2 continues to advance.

Interpretation: The distribution of the 11 commits in the FlagGems window is highly representative — member backends (Hygon HCU, Iluvatar) doing “fix and optimize,” compilation paths (KernelGen/KMCompiler) doing “new operator coverage expansion,” and testing/benchmark engineering providing “stability,” all three lines advancing together; the Iluvatar Triton 3.6 regression fix signals that unified compiler upgrades will continue to impose adaptation work on the operator library — precisely the value of having a common operator library.

1.4 Inference plugin layer: sglang-plugin-FL DeviceInfo service-ization, vllm-plugin-FL GDN precision fix, FlagGems-vllm fix (09-02)

Sources: sglang-plugin-FL #73, #87, vllm-plugin-FL #385, FlagGems-vllm #724

  • sglang-plugin-FL: 14:12 #73 abstracts DeviceInfo into a service class and introduces a vendor early-patches mechanism — vendor early patches can be injected per vendor, moving the plugin’s differentiated adaptation to chip vendors from “scattered code” toward “mechanization”; 16:53 #87 adds a /rerun-failed-ci slash command (PR authors can self-service rerun failed CI, a collaboration efficiency tool).
  • vllm-plugin-FL: 11:28 #385 fixes the GDN (Gated Decay Network, a new structure introduced by Qwen3.8-Flash-Next) packed decode path to keep the beta parameter in FP32 — same origin and theme as the previously reported vLLM upstream PR 53877 (GDN Packed Decode FP32 precision defect fix), with the plugin layer and upstream converging on the precision issue in sync.
  • FlagGems-vllm: 16:28 #724 fixes the [KMCompiler] persistent_topk stale ring slot issue and hardens vLLM tests; 17:29 fixes a crash caused by torch auto-backend auto-loading in CI.

Interpretation: Same-day commits across the three inference plugin repos all point to “adaptation engineering”: vendor patch mechanization (sglang #73), precision issues fixed in sync with upstream (vllm-plugin-FL #385), CI stability hardening (FlagGems-vllm) — the plugin layer supporting yesterday’s “multi-model Day0” is consolidating one-off adaptations into reusable mechanisms.

1.5 Training stack and remaining repos: Megatron-LM-FL/TransformerEngine-FL CICD alignment, FlagTensor merges muxi branch, FlagDNN Hygon CPU reference implementation (09-02)

Sources: Megatron-LM-FL #137, TransformerEngine-FL #111, FlagTensor #17, FlagDNN commits, FlagPrism #7, Torch-FL #244, FlagFFT commits

  • Training stack CICD alignment: 15:43 Megatron-LM-FL #137 updates CUDA runner labels; 15:49 TransformerEngine-FL #111 unifies NVIDIA runner labels to flagcicd-a100 — the CI infrastructure of the two training stack forks wraps up migration to FlagOS’s own runner (flagcicd), shedding dependence on generic CUDA runners.
  • FlagTensor: 16:15 #17 merges the muxi branch into main (the branch name corresponds to MetaX’s pinyin) — following the 08-26 #16 merge of the iluvatar (Iluvatar) branch, another member-vendor adaptation line enters the main branch of FlagTensor, and the repo is becoming a convergence point for the multi-vendor tensor layer.
  • FlagDNN: 16:01 adds a CPU reference implementation for the hygon (Hygon) backend — providing a baseline for Hygon operator correctness verification.
  • FlagPrism: 16:35 #7 fixes ixKN profile suffix compatibility and adds comments (Iluvatar-related toolchain).
  • Torch-FL: 10:44 #244 documents a Hugging Face connection fallback (paired with the “HF official test runner” being built starting from yesterday’s #236, solving the problem of running official tests under restricted networks).
  • FlagFFT: 17:48-23:15 three commits progressively isolate the NumPy FFT verification tool from test inlining into a standalone harness (out-of-tree verification engineering).

II. News Coverage and Ecosystem

2.1 News Side: Zero Direct Hits in 24h; Review of the Past 3 Days’ Main Thread: 2.2 rc0 Validation Tag Peak Period Winds Down (09-01~09-02)

Source: BAAI Community Day0 article (09-01, covered in detail yesterday)

  • In this window (09-02 10:18 ~ 09-03 10:18), component-level and topic-level news queries returned zero direct hits: 24h-7d queries in Chinese and English for FlagOS/FlagGems/FlagScale/FlagTree/FlagPerf/FlagSparse/FlagAttention/FlagCX were all empty or hit only reprints of the 09-01 BAAI official Day0 article across various channels (not new content); HN Algolia returned zero relevant results (all “flag” hits were feature flags noise).
  • Trend over the past 3 days (corroborated by GitHub): 8/31-9/1 was the intensive landing period for 2.2 RC0 validation tags (FlagSparse, vllm-plugin-FL, sglang-plugin-FL, FlagGems-vllm, Torch-FL, FlagOS-Compressor and other rc0.post1 series, all previously reported); this window has entered a release plateau — tags/releases API verification shows no new tags, no new releases; the next action on the 2.2 checklist is expected to be an rc0.postN iteration or the official rc branch cut (signaled by activity in the community checklist repo, which saw no commits in this window).
  • Ecosystem-side narrative is still digesting the “three Day0s in four days”: the article still appeared in multiple Chinese query result sets on 09-02, indicating that the FlagOS multi-chip adaptation capability narrative is spreading through the community, but no new official release picked up the thread within the window.

2.2 BAAI Institutional News: SCO Framework AI Training Project Concludes; Content Community Remains Active (09-02)

Source: BAAI Community “SCO Framework AI Training Project Successfully Concludes” (BAAI Community, 09-02)

  • BAAI Community published “SCO Framework AI Training Project Successfully Concludes”: BAAI’s international AI training project under the SCO framework has completed — a routine move by BAAI as an open-source and AI public platform on the international cooperation side, not directly related to the FlagOS ecosystem but an institutional development.
  • BAAI Community content remained highly active on 09-02/09-03 (spread of embodied intelligence talent migration observations, reprints of the Li Feifei team’s multimodal world model release, the latter not a BAAI achievement); BAAI’s “model R&D + content community + FlagOS” triangular content ecosystem is proceeding at a normal pace.

III. Deep Dive on Member Organizations

3.1 Enflame: Final Allotment Rate of 0.02455315% Announced, Lower Than Moore Threads and MetaX (09-02/09-03)

Sources: Guandian (allotment rate announcement), Sina Finance (announcement overview), Phoenix Finance (allotment rate lower than Moore Threads and MetaX), Phoenix Finance (industry competition under pressure)

  • On the evening of 09-02, Enflame Technology (688801.SH) announced its online issuance subscription status and allotment rate: final allotment rate of 0.02455315% (approximately 2.45 per ten thousand), with 7,032,045 valid online subscription accounts and 42.066 billion valid subscription shares; due to the high preliminary online subscription multiple, the clawback mechanism was triggered, resulting in 10.3285 million shares allocated online after clawback.
  • Media cross-comparison: the allotment rate is lower than Moore Threads (0.0363%) and MetaX, representing a low allotment rate befitting the “third-highest-priced new stock of the year” (issue price of 142.18 yuan; earlier estimates suggested winning one lot could yield over 280,000 yuan in profit); the strategic investment lineup including Tencent and Xiaomi continues to draw interpretation.
  • Risk-perspective coverage appeared simultaneously: Phoenix Finance and others flagged valuation and competition risks with headlines like “industry competition under pressure on all sides” and “the stark reality behind the grand narrative” (1.7% market share), while a Sohu opinion piece noted that “domestic GPUs — what’s the next battleground” has entered a period of discussion.

Interpretation: Enflame’s IPO subscription has entered the “allotment rate realization phase,” with 7.03 million subscription accounts confirming market enthusiasm for domestic AI chips; meanwhile, media are shifting their lens from “IPO wealth creation” to “delivery and competition” — this belongs to the same narrative thread as yesterday’s discussion of MetaX/Iluvatar CoreX’s “profitability quality” (who is actually making money in their core business, who has sustained delivery capability). For FlagOS, Enflame’s post-IPO software ecosystem investment cadence (the enflame backend has registered tops1.9.10 application images in this window) is worth continued tracking.

3.2 MetaX: Yi Rating Initiates Coverage, Xiangshan Co.’s Acquisition of Its Distributor Draws Attention (09-02)

Sources: East Money (Yi Rating initiation), Yicai (initiation report), Sohu (Xiangshan Co. acquisition)

  • 09-02 institutional rating update: Yi Rating released its initiation report on MetaX (688802.SH) (published simultaneously by East Money and Yicai), marking the beginning of sell-side coverage buildout following MetaX’s IPO.
  • A-share mirror development: Xiangshan Co. announced a cross-sector acquisition of a target related to “MetaX’s general distributor” (previously, Ruisheng Intelligent sold its stake in the distributor less than a year after investing, and Joyson Electronics’ private placement has already gained 40%) — MetaX’s distribution system has become an A-share M&A theme, reflecting the commercial heat of its ecosystem.

Interpretation: MetaX remains active in the FlagOS MUSA backend (FlagGems metax guard fix #685, yesterday’s report on MUSA dual-version registration); the dual signals of “initiation coverage + M&A theme” on the A-share side indicate its capital market narrative has entered a normalized phase; progress in embodied scenarios following the joint venture with UBTech (Xixuan Chuangzhi, reported yesterday) remains a future point of interest.

3.3 Cambricon: Aggressive Recruitment of Software Talent, Signaling Software Stack Investment (09-02)

Source: Tiantian Fund (relaying Cambricon’s aggressive software talent recruitment) (09-02)

  • 09-02 media observation: Cambricon (688256.SH) is aggressively recruiting software talent, interpreted as a signal of its shift from a chip hardware narrative toward investment in “software stack/ecosystem” (Cambricon is one of the Day0 ten chips on the FlagOS side, and build-infra has registered neuware4.7.2/4.4.3 dual-backend vLLM application images for it in this window).
  • Note: This is a media observation relay; the specific recruitment scale and direction have not been officially verified by Cambricon, and it is included as a side signal of software ecosystem investment.

Interpretation: Following Goldman Sachs’ target price upgrade (reported yesterday), Cambricon has shifted to coverage of software talent reserves, echoing its active operator/image registration in the FlagOS ecosystem — the competitive focus of domestic AI chips is shifting from “single-chip metrics” to “software stack maturity,” which is mutually causal with FlagOS’s positioning as a unified software stack.

3.4 Iluvatar CoreX: In-Depth Interim Report Reading and Target Price Adjustment, CMBI Maintains Buy (09-02)

Sources: EE Times China (in-depth interim report reading), Guandian (CMBI)

  • EE Times China published an article on 09-02 titled “What Really Matters in Iluvatar CoreX’s Interim Report Is Not the 17.2%”: it argues that the market’s focus on its gross margin (17.2%) may miss the point, advocating for interpreting the interim report from dimensions such as revenue structure and customer quality — continuing the “profitability quality” discussion but offering a different perspective.
  • CMBI slightly adjusted Iluvatar CoreX’s target price to HK$872 on 09-02 while maintaining a buy rating (Hong Kong stocks); geekpark published a long-form piece the same day titled “Domestic GPUs Enter the Scale Delivery Test,” analyzing it alongside peers with “the winning hand written into the three financial statements.”

Interpretation: Iluvatar CoreX has substantive activity on the FlagOS side in this window (FlagGems #5783 Triton 3.6 fix, the iluvatar line already merged into FlagTensor yesterday, build-infra corex4.4.0 application image registration); Hong Kong sell-side maintains buy but slightly lowered the target price, and the financial narrative has entered a phase of “squeezing out excess and focusing on the core business.”

3.5 Horizon Robotics and Industry Perspective: From Selling Solutions to Selling the Base; Domestic GPUs Enter the Scale Delivery Test (09-02/09-03)

Sources: Sina (from selling solutions to selling the base), Sohu (2.218 million chip units shipped), Geekpark (scale delivery test), Phoenix Finance (what’s the next battleground)

  • Horizon Robotics (09660.HK) interim report continued reading (two pieces on the morning of 09-03): Sina Finance’s “From Selling Solutions to Selling the Base” dissects the identity transformation behind its revenue structure (licensing and service revenue accounting for over half) — from an intelligent driving solution provider to a “base/platform” company; Sohu continues to fuel the interim report discussion with a headline highlighting the tension of “2.218 million chip units sold (+12.1%), yet over half of revenue is not from selling chips.” Both are interpretive extensions of yesterday’s interim report, with no new data.
  • Industry commentary (evening of 09-02): Geekpark’s “Domestic GPUs Enter the Scale Delivery Test” and Phoenix’s “Domestic GPUs — What’s the Next Battleground?” approach from the three financial statements and the competitive landscape respectively, pointing to “delivery capability” as the next-stage deciding factor — in sync with the discussions in 3.1/3.4.

Interpretation: Horizon Robotics’ “platformization” transformation narrative as a FlagOS member (D-Robotics BPU) aligns with FlagOS’s positioning as an open-source software stack platform; the collective shift of industry commentary toward “scale delivery” means that the value proposition of unified software stacks like FlagOS (one adaptation, multi-chip delivery) is in a period of rising industry demand.

IV. Summary

In this window (09-02 10:18 ~ 09-03 10:18), the FlagOS ecosystem main threads:

  1. High activity on GitHub, stable on releases: 16 repos pushed, 48 commits within the window; tags/releases API verification shows no new tags/releases within the window — the dense 2.2 rc0.post1 verification tag period before yesterday has concluded, entering a version stabilization phase (next signal to watch is the community manifest repo activity).
  2. build-infra inference image matrix expands: the vLLM 0.24.0 application image line added Cambricon (neuware4.7.2/4.4.3 dual backend), Iluvatar CoreX (corex4.4.0), and Enflame (tops1.9.10) on the same day, stacking on yesterday’s Ascend/MetaX/Moore Threads, bringing multi-vendor inference image registration to over 6 vendors; the sglang verification matrix advanced from Phase 0 to Phase 5 in two days (0.5.18 application image + import-face shim wheel).
  3. High-frequency multi-vendor backend fixes in FlagGems: Hygon HCU operator performance optimization, Iluvatar CoreX’s Triton 3.6 regression fix, five new operator registrations for KernelGen/KMCompiler; member chip backend “fix and optimize” became the main theme of the operator library within the window.
  4. Engineering advances in inference/training layers in parallel: sglang-plugin-FL introduced the vendor early-patches mechanism and DeviceInfo serviceization; vllm-plugin-FL fixed GDN packed decode FP32 precision (same theme as vLLM upstream); Megatron-LM-FL/TransformerEngine-FL CI migrated to the flagcicd-a100 self-owned runner; FlagTensor merged the muxi (MetaX) branch.
  5. Member financial developments: Enflame’s final allotment rate of 0.02455315% (7.03 million subscriptions, lower than Moore Threads/MetaX); MetaX received Yi Rating initiation coverage; Cambricon aggressively recruiting software talent; Iluvatar CoreX’s CMBI target price slightly lowered while maintaining buy; industry commentary shifting toward “scale delivery test.”
  6. News side was a quiet window: zero direct component-level hits in 24h, the official Day0 article continues to spread within the ecosystem but with no new releases to follow up; on the BAAI institutional side, there are routine developments such as the conclusion of the SCO framework AI training program.

Appendix: Complete Source List

Source Verification Result
GitHub org repos API (52 repos) 16 repos pushed within the window, no new repos created (latest FlagPrism still 08-04); build-infra/FlagGems etc. pushed 09-03 early morning 10:0x verified via commits as non-main-branch actions (no corresponding main commits), not counted in the main text
GitHub commit search All 48 commits within the window retrieved (build-infra 24, FlagGems 11, FlagFFT 3, sglang-plugin-FL 2, FlagGems-vllm 2, FlagTensor 2, vllm-plugin-FL/FlagPrism/Torch-FL/FlagDNN 1 each); Megatron-LM-FL #137, TransformerEngine-FL #111 supplemented via direct commits API
GitHub tags API + fixed-date verification No new tags within the window: FlagTree v0.4.0 (2026-01-09), FlagSparse v0.3.0-rc0.post1 (08-27), vllm-plugin-FL v0.3.0-rc0.post1 (08-31), FlagGems v5.4.0-rc0.post1 (09-01), build-infra v2.1.1 (08-03), Megatron-LM-FL v0.15.0rc7+fl (2026-01) etc. were all created before the window
GitHub releases API No new releases (build-infra has no release; FlagGems latest v5.3.0 is 06-24; FlagSparse v0.3.0-rc0.post1 release published 09-01 02:02 CST, before yesterday’s window)
build-infra #676-#698 series Verified item by item: 0.24.0 application image three-backend enablement (#678), Cambricon neuware4.7.2/4.4.3 (#681/#687), Iluvatar corex4.4.0 (#688), Enflame tops1.9.10 (#698), symm-mem guard registration (#680), sglang 0.5.18 Phase5 (#684), shim wheel (#690), Requires-Dist merge (#697), metax guard fix (#685), record_only mode (#696)
FlagGems #4663/#5783/#5393/#5419/#5426/#4443/#5870/#5886/#5902/#5908 Verified item by item: HCU operator optimization, Iluvatar Triton 3.6 fix, KernelGen/KMCompiler operators, INT8 testing, benchmark isolation, _FULL_CONFIG sorting
sglang-plugin-FL #73/#87, vllm-plugin-FL #385, FlagGems-vllm #724 Verified: DeviceInfo servitization + vendor early-patches, /rerun-failed-ci, GDN packed decode FP32, persistent_topk fix
Megatron-LM-FL #137, TransformerEngine-FL #111, FlagTensor #17, FlagDNN, FlagPrism #7, Torch-FL #244, FlagFFT Verified: flagcicd-a100 runner migration, muxi branch merge, hygon CPU reference, ixKN profile fix, HF connection fallback documentation, FFT verification tool isolation
Google News RSS Chinese group FlagOS/component-level zero hits in 24h (only 09-01 Day0 article reprints resurfacing + gambling SEO filtered out); Enflame (allotment rate 0.02455315%, 7.032 million households, lower than Moore/MetaX, industry pressure interpretation), MetaX (Yiping first coverage, XiangShan acquisition general agent), Cambricon (recruiting software talent), Tianshu Zhixin (EET in-depth, CMBI HK$872), Horizon (selling base/2.218 million units), multiple GeekPark/Phoenix industry commentary items verified
Google News RSS English group All empty or noise (ECOticias “tree branches” colliding with FlagTree), no valid hits
HN Algolia Zero relevant hits for FlagOS/FlagGems (all 476/213 items are noise like feature flags, Six Flags, etc.)
BAAI Community (gnews hits) SCO framework AI training program completion (09-02); Day0 article continues to spread
Viewpoint Network/Sina Finance/East Money/Yicai/Sohu/GeekPark/Phoenix/EET China/Tiantian Fund Allotment rate announcement data, Yiping first coverage, recruiting software talent, EET interim report in-depth, target price adjustments etc. verified item by item via title + summary