Research window: 2026-09-24 10:00 ~ 2026-09-25 10:00 (approx. 24 hours; continues from the 09-24 report window with no gap) Sources: GitHub (org: flagos-ai, full pushed_at audit of 54 repos, 24 repos active in window; commit search returned 175 hits across 18 repos; tags and releases verified per repo; community release manifest cross-checked line by line against build-infra PR text), Google News RSS (25 query sets in Chinese and English, via proxy), ITHome / Yicai / ifeng / Sina Finance and other reports (see appendix source list for details)

Index

  • Today’s Highlights: 3 days to 2.2 GA (09-28) — delivery scope audit begins: 18 FEPs retained, 6 deferred to 2.3 (09-24)
  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 Release Engineering: L1 manifest pending merge; development mainlines collectively bumped (09-24)
    • 1.2 build-infra: 910C dual config passes megatron_rl E2E; sglang app pages ship seven in a row (09-24)
    • 1.3 build-infra (II): vendor-neutral naming at the app layer; rpm / deb packaging pipelines split (09-24)
    • 1.4 FlagTree: 0.7.0 platformization tags continue to expand — iluvatar3.6 / metax3.6 / tileir3.6 land (09-24)
    • 1.5 FlagGems: Kunlunxin XPU major fix batch nears 20 commits; KernelGen adds 14 more Nvidia operators (09-24)
    • 1.6 FlagGems (II) and plugin line: Biren SUPA backend merged; Ascend grouped_matmul (09-24)
    • 1.7 FlagSparse: mainline bumped to 0.4.0 after v0.3.0; NCIC-AlphaSparse collaboration merged (09-24)
    • 1.8 FlagQuantum: 13 commits in a single window — PennyLane execution bridge and CPU fusion optimization batch (09-24)
    • 1.9 Other Activity: FlagTrain’s first Triton training operators land (09-24)
  • II. News Coverage and Ecosystem
    • 2.1 Component-level search: twentieth consecutive quiet window (09-24/09-25)
    • 2.2 Open3D-PIMC coverage matrix expands again: Sina Finance publishes article (09-24)
  • III. Member Company Deep Dive
    • 3.1 Moore Threads: Protenix-v2 full-pipeline inference support — AI4S three-scenario loop closed (09-24)
    • 3.2 Ascend: 910C NPU visibility fix and RL training chain ready (09-24)
    • 3.3 Kunlunxin: XPU major fix batch and P800 test machine migration (09-24)
    • 3.4 Biren / Xiwang: SUPA and Sunrise dual-line integration (09-24)
    • 3.5 MetaX / Hygon / Tsingmicro: metax3.6 tags, complex scalar fix, tsingmicro variant (09-24)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Full Source List

Today’s Highlight: 3 Days to 2.2 GA (09-28) — Delivery Scope Audit Begins: 18 FEPs Retained, 6 Deferred to 2.3

Date: 2026-09-24　Source: community #115, community #114

The rc2 stabilization period ended on 09-24 per the official schedule, with GA set for 09-28. Within this window, the main release-engineering activity shifted from “release artifact freeze” to “scope and claims freeze” — at 14:29 on 09-24 (Beijing time), the community repo opened PR #115 “docs: define delivered FlagOS 2.2 FEP scope and defer remaining work to 2.3” (branch docs/fep-2.2-rc2-audit-20260924, pushed twice that day through 16:12):

  • Scope bifurcation: the 2.2 index contains 18 Implemented FEPs (retained); FEP 0019, 0069, 0081, 0084, 0093, and 0100 have their target version changed to 2.3 and are marked Deferred.
  • Tightened retention criteria: each retained FEP is limited to “implemented interfaces + verified configurations,” with all other goals moved into an explicit 2.3 section and the associated deferred-work index.
  • RC2-aligned update sources: source revisions, APIs, packaging details, test commands, and associated PRs updated item by item, while preserving verification provenance, platform failures, and known limitations.
  • Terminology clarification: explicitly distinguishes Tree primitive QA from NVSHMEM fusion / FlagCX-backed fusion, and Compressor artifact testing from full-model inference acceptance — avoiding any framing of “component-level verification” as “end-to-end acceptance.”
  • Style constraints: concise English required; template text and “unsupported completion / performance claims” removed.

In parallel, the formal tag list for the ten L1 operator library components (community #114: flaggems v5.4.0, flagfft v0.2.0, flagsparse v0.3.0, flagdnn v0.3.0, flagblas v0.3.0, flagtensor v0.3.0, flagaudio v0.3.0, flagattention v0.4.0, flaggems-vllm v0.2.0, flaggems-sglang v0.1.0) remains under review (waiting merge).

Assessment: Yesterday’s observation was that “the manifest, tags, and scope cross-validate one another”; today #115 advances that validation to a third layer — the public claims themselves. With 3 days to GA, the focus is not new features but separating “implemented / verified / deferred” into distinct statements, each backed by evidence. This is a release-discipline action: the public introduction of 2.2 will use this version of the FEP document as its factual basis, and the subsequent 2.3 development narrative (including new directions such as 3D compute and quantum) is now formally cut off from the 2.2 delivery boundary.

I. Open-Source Project Progress (GitHub Activity)

1.1 Release Engineering: L1 Manifest Pending Merge; Development Mainlines Bump in Unison (09-24)

Date: 2026-09-24　Source: community #114, FlagSparse #90, FlagAudio #20, FlagAttention #77

  • #114 status: The mapping manifest of the ten L1 components (“official tags ← rc2 baseline”) is still under review. Line-by-line remote readback verification of the ten official tags is complete; once the manifest is merged, it becomes part of the 2.2 official release artifacts.
  • Development mainline bumps: The day after the official tags landed, two mainlines switched to the next cycle ahead of schedule — FlagSparse main bumped to 0.4.0 (09-24 19:10, PR #90 “Bump main version to 0.4.0”); the FlagAudio version line was corrected to 0.4.0 (19:20, PR #20, with one revert and re-apply along the way, finally settling at 0.4.0); FlagAttention likewise carried out an “update version” commit (18:12, PR #77).
  • Implication: The 2.2 release artifacts (rc2 baseline + official tags) and the development mainlines (0.4.0 series) have formally diverged in version numbering — the 2.3 development cycle has in effect begun, proceeding in parallel with the manifest closeout ahead of 2.2 GA.

1.2 build-infra: 910C Dual Configurations Pass megatron_rl E2E; Seven sglang App Pages Land (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: build-infra #1105, #1107, #1095, #1085, #1089, #1090, #1091, #1092, #1093, #1094

  • Ascend 910C dual-configuration RL E2E passes (#1105, closed 09-25 07:58): Two 910C machines (hw115: CANN 9.0.0 + torch 2.10 + torch-npu 2.10; hw114: CANN 8.5.0 + torch 2.9 + torch-npu 2.9) with single-step installation of megatron-core[rl]==0.17.1+fl.0.2.3; train_rl.py GRPO (rollout→training step, 2 rounds) exits end-to-end with code 0. The wheel already bundles three RL fixes (packed_seq gating, KV-append platformize + NPU paged attention, capability-dispatched paged dispatch), so no further patching is needed on the container side.
  • 910C NPU visibility fix (#1107): The verify container switched to privileged startup — testing showed torch.npu.device_count()=0 (drvRetCode 87 / aclInit 507899) was caused by the Ascend-Docker-Runtime (7.3.0) hook failing to map davinci dies for the new container; after the fix, verification passed on four Ascend nodes (910C RL image, 910B training image).
  • megatron_training 0.17.1 on 910C (#1095) + app image tag 2.2.0-0.2.3 dual-configuration records (#1103 / #1104).
  • Seven sglang app pages in a row (09-24 13:40 ~ 18:29): sglang0.5.12-kunlunxin-xre5.37.1 (#1093), sglang0.5.11-ascend-cann8.5.0 (#1092), sglang0.5.11-enflame-tops1.10.6 (#1094), sglang0.5.11-hygon-dtk26.04 (#1089, standard + empty-mode dual pages), sglang0.5.11-mthreads-musa4.3.5 (#1090), sglang0.5.11-iluvatar-corex4.5.0 (#1091), sglang0.5.11-generic-c13.0 (#1085); additionally sglang0.5.12-thead-ppu2.1.0 was updated to the official component tag (#1082). Seven platform app pages landed in a single night, rapidly densifying the 2.2 “platform coverage table” in its final stage.

1.3 build-infra (Part 2): Vendor-Neutral Naming at the App Layer; rpm / deb Packaging Pipelines Split (09-24)

Date: 2026-09-24　Source: build-infra #1086, #1097, #1084, #1083, #1102, #1088, #1099, #1100, #1101, release-info portal

  • Vendor-neutral naming at the app layer (#1086): Application-layer release artifacts switched to vendor-neutral naming (names formerly containing vendor strings were consolidated), standardizing the unified application-layer narrative across diverse chips.
  • Packaging pipeline split: #1097 (FlagDNN / FlagTensor / FlagFFT moved to release-path builds, with rpm packaged from a release branch), #1084 (noarch-rpm: independent build, verification, and release per distribution), #1083 (native-rpm: download each suite package by name at release time), #1102 (flagcx-rpm and flaggems-deb stamped with release versions), #1101 (fixes to derived base images for openEuler findutils and FlagFFT rpm).
  • Other: #1088 (doc generator refuses to overwrite pages it did not generate), #1099 (empty-mode sglang image naming fix), #1100 (proxy relay for megatron app image builds).

1.4 FlagTree: 0.7.0 Platform Tags Continue to Expand — iluvatar3.6 / metax3.6 / tileir3.6 Land (09-24)

Date: 2026-09-24　Source: FlagTree tags page, #1280, #1283, #1274, #1289, #1284, #1279, #1281, #1116, #1216

  • Tag expansion continues: The 0.7.0 series continued to expand and update within the window — four variants appeared: iluvatar3.6 (Iluvatar CoreX, 10:17), metax3.6 (MetaX, 17:13), tileir3.6 (next day 01:00), and tsingmicro3.3 (Tsingmicro, 21:27); the main tag 0.7.0 along with triton3.6, xpu3.6, etc. were updated within the window. The 0.7.0 series now covers three triton lines plus ppu / mthreads / xpu / tsingmicro / ascend / metax / iluvatar / tileir platform variants.
  • CI coverage expansion: Two new FlagGems baseline + test workflows added for mthreads3.6 (Moore Threads) and hcu3.6 (Hygon) (#1283, #1274); MetaX vLLM benchmarks added to workflows (#1280); routine workflows update (#1289).
  • Packaging fixes: Triton module library name unified to _C/libtriton.so (fixing kernel compilation failure after deb installation, #1284), rpm installs clang and ships it with the package (#1279), deb pre-release versions given PEP 440-compliant numbers (#1281).
  • Features and docs: PPU / MTHREADS updated for fp8/fp4 low-precision support via the “backend capability contract” (#1116); TLE supports large-node transfers and limits the number of live tensors (#1216); TLE-Distributed adds 2 triton-distributed examples (#747); allgather in-graph field corrects context_id (#1287); SPEC cpp lands the metax backend (#1107).

1.5 FlagGems: Nearly 20 Kunlunxin XPU Fixes; KernelGen Adds 14 More Nvidia Operators (09-24)

Date: 2026-09-24　Source: FlagGems PR list (#6417, #6445, #6457, #6464, #6473, #6478, #6483, #6520, #6544, #6617, #6622, #6636, #6639, #6646, #6653, #6678, #5912, #5915, #5916, #5794, #6144, #6168, #5322, #6509, #6510, #6511, #6514, #6517, #6527, #6533)

FlagGems saw 46 commits within the window, of which nearly 20 in the Kunlunxin (XPU) direction formed the largest batch of the window:

  • Performance fast paths: aten::_flash_attention_forward cached fast path (#6520), scaled_mm host fast path (captured launch replay, #6473), conv_depthwise2d zero-bias reuse (#6445), small/medium-shape true_divide raw-kernel fast path (#6417), fill_tensor_out (#6678), _functional_sym_constrain_range_for_size (#6636), scalar comparison operator batch (#6639), native_dropout_backward two-level kernel (#6653).
  • Correctness fixes: conv_transpose1d/2d binding and conv2d routing (#6457), erfc/erfinv/bucketize/threshold (#6544), bf16_paged_mqa_logits scalar max abort (#6464), fused-kernel batch (#6266), heaviside_/hypot_ grid parameters (#6483), elementwise/special-function batch (#6265), float64 case skipping (#6478), XPU fix and optimization batch (#6646, #6622).
  • CI migration: test-kunlunxin migrated to the klx-p800 machine type and fixed local action references (#6617).
  • KernelGen / Nvidia adds 14 Triton operators: _unique2, _pad_enum, _philox_normal_, rrelu_, relu6_, softshrink_backward, hardshrink_backward, l1_loss, linalg_solve_ex, _histogramdd dual variants (bin_tensors / bin_cts), _logcumsumexp, infinitely_differentiable_gelu_backward, _scaled_grouped_mm_v2.
  • Engineering side: Fixed operator registration paths pointing to invalid ATen overloads (#6655), benchmarks add reference-only baseline validation (#6681), CI builds only on packaging changes (#6542), test log prefixes made dynamic (#6680).

1.6 FlagGems (Part 2) and Plugin Lines: Biren SUPA Backend Merged; Ascend grouped_matmul (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: FlagGems #6367, #6654, #6579, #6580, #6679, vllm-plugin-FL #555, #530, #422, FlagGems-vllm #847, #837

  • Biren SUPA backend dual-layer integration: FlagGems “[BACKEND] support Biren SUPA backend” merged (#6367, 18:07); vllm-plugin-FL “feat(biren): add Biren SUPA backend” merged in sync (#530, 14:30) — a new chip partner onboarded on the same day at both the operator-library and inference-plugin layers.
  • Sunrise vLLM 0.24 line port: vllm-plugin-FL #555 ports the Sunrise/PTPU 0.2.1 patch set to the vLLM 0.24 mainline, enabling Qwen3.6 27B and 35B-A3B to run correctly in enforce-eager mode on PTPU hardware; 14 new import-time/deferred patch modules added (FLA/GDN rebinding, FlagCX communication and collective ops, cudagraph, etc.), with per-patch guards degrading API mismatches to skips rather than import crashes (merged next day 00:06).
  • vllm-plugin-FL also merged an operator profiling tool (#422, 23:54).
  • Ascend: FlagGems writes grouped_matmul into operators.yaml (#6654), log_ launch grid capped by vector core count (#6579); FlagGems-vllm supplements the XCS operator manifest (#847), suppresses TLE output for flash_mla on certain platforms (#837).
  • Hygon: Fixed handling of complex scalar multiplication when no complex kernel parameters are available (#6679). SiliconFlow: Optimized the generic and Ascend kernels of adaptive_avg_pool2d (#6580). Also flagtree metax parameter updates (#6637).

1.7 FlagSparse: Mainline Bumps to 0.4.0 After v0.3.0; NCIC-AlphaSparse Collaboration Merges (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: FlagSparse PR list (#92, #91, #90, #88, #87, #75, #36)

  • Version: After the v0.3.0 official tag (09-24 01:50), main bumped to 0.4.0 to enter the next development cycle (#90); version alias fix (#88).
  • Collaboration merge: Six merges from the NCIC-AlphaSparse collaboration line — Merge PR #92 (next day 00:20), biv algorithm added (#87 batch), MUSA delivery defect fixes and matrices directory cleanup, three rounds of CI wrap-up (ci / ci2 / ci3).
  • Packaging: rpm dist-info path no longer bound to spec Version (#91), packaging trigger conditions narrowed (#75), pip prefix compatibility fix (#36); dependabot batch (#13/#20/#34/#35).

1.8 FlagQuantum: 13 PRs in a Single Window — PennyLane Execution Bridge and CPU Fusion Optimization Batch (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: FlagQuantum PR list (#207 ~ #219)

  • Execution bridge expands again: Added a PennyLane Lightning execution bridge (#207, 11:16) — the third third-party framework execution channel after Cirq and Qiskit Aer; also added an evidence-constrained simulator selection advisor (#208 / #210, giving backend recommendations based on measured evidence).
  • Cloud integration: Prioritizes the Quafu task API and aligns with backend qubit capacity (#209).
  • CPU optimization batch: SWAP cross-product-state component remapping (#212), noisy-MPS count fallback (#213), CPU CZ graph fusion into a single state-vector pass (#214), shallow Clifford circuit routing by measured cost (#215), CPU controlled-phase graph fusion (#216), product-state Clifford gate specialization (#218), CPU dense-fusion width tuning (#219).
  • Benchmark construction: Cross-framework simulator workload corpus (#211) and native corpus evidence refresh (#217).

1.9 Other Activity: FlagTrain’s First Batch of Triton Training Operators Lands (09-24)

Date: 2026-09-24　Source: FlagTrain #1, #2, FlagBLAS #131, FlagFFT commits, FlagAudio #20, FlagAttention #76, Torch-FL #415, FlagCX #623, FlagScale-Agent #35

  • FlagTrain (new repo, created 09-16) first operators: The Triton lamb training operator submitted by KMCompiler (#1) and test method updates (#2) were merged — this repo is positioned to “implement reusable training operators and compute components for mainstream training frameworks (Megatron / DeepSpeed / Transformer Engine) in Triton”; previously it had only docs and a directory skeleton, and this window marks the first implementations landing.
  • FlagBLAS: Ascend L2 operator optimization batch — banded packed and triangular L2 (#131 batch, 15:57), SYMV launch path simplification, merged into mainline (09-24 21:09).
  • FlagFFT: JIT artifact cross-process isolation and locking + multi-process regression tests (12:10); acceptance matrix restructured by dtype and batch placement; mixed-radix CT size 210 tests.
  • FlagAudio: Version line corrected to 0.4.0 (#20, 19:20); rpm pip fallback installable on openEuler.
  • FlagAttention: Version update (#77), benchmark fixes (#75, #76).
  • FlagDNN: mthreads reuses compiled artifacts instead of going through the legacy conversion API (#21 batch); packaging merge (#12).
  • Torch-FL: bool neg escapes via (op, dtype) when routed through FlagGems (#415); profile selection on PPU and full draining of the CUPTI tracer (#413); HF test subprocess utils binding fix (#414).
  • FlagCX: rpm Requires excludes unversioned CANN sonames (#623); CRL heterogeneous gather aggregation scheduling moved after transport (#621); PAL supports SHCA verbs ABI and 17-bit LID (#618).
  • FlagScale-Agent: skills split by framework for inference upgrades and hardware adaptation (#35, 11:38).
  • libtriton_jit (#70) and FlagTensor (#24): Packaging trigger conditions narrowed in sync.

II. News Coverage and Ecosystem

2.1 Component-Level Searches: Twentieth Consecutive Quiet Window (09-24/09-25)

Date: 2026-09-24 ~ 2026-09-25　Source: Google News RSS (25 groups of Chinese and English query terms, via proxy), HN Algolia

  • Component-name queries (FlagOS / FlagGems / FlagScale / FlagTree / FlagPerf / FlagCX / KernelGen) had zero hits within the 24-hour window — the twentieth consecutive quiet window on the news side.
  • Of the 20 ecosystem-side hits within the window: Moore Threads Protenix-v2 (included in 3.1), Open3D-PIMC follow-up coverage (see 2.2); the rest were stock market pieces (Enflame, semiconductor sector) and think-tank content, excluded by topic.
  • HN Algolia (FlagOS / FlagGems / FlagScale / BAAI) had zero hits.

2.2 Open3D-PIMC Coverage Matrix Expands Again: Sina Finance Publishes Article (09-24)

Date: 2026-09-24　Source: Google News (Sina Finance, 09-24 10:15), Guangming Online (09-23), CNR (09-18)

  • Sina Finance published “Open3D PIMC Open-Sourced, Completing the Software Foundation of the 3D Computing Ecosystem” — the coverage matrix for the 3D computing chip programming framework jointly developed by Tsingmicro and BAAI continues to expand: Xinhua (09-14) → CNR (09-18) → Guangming Online (09-23) → Sina Finance (09-24).
  • The core facts of the coverage are consistent with previous reports (unified programming model, plug-in compatibility with multiple vendors, inclusion in the FlagOS system, targeting next-generation 3D computing chips), with no new technical information; this is a continuation of media momentum and is not expanded upon further in this report.

III. Member Unit Deep Dive

3.1 Moore Threads: Full-Pipeline Inference Support for Protenix-v2 — Closing the Loop on Three AI4S Scenarios (09-24)

Date: 2026-09-24　Source: ITHome (via ifeng), Yicai

Moore Threads announced that, based on its in-house MUSA software stack, it has completed full-pipeline inference support for the open-source biomolecular structure prediction model Protenix-v2 on its training-inference integrated AI compute card MTT S5000. Protenix-v2 was developed by ByteDance’s Seed team (released 2026-04) and is one of the few fully open-source foundation models worldwide capable of high-precision end-to-end prediction of proteins, DNA, RNA, and small-molecule ligand complexes. Measured results: S5000 delivers roughly 26% average speedup in end-to-end inference performance over mainstream international GPUs, with 3D conformation prediction accuracy highly aligned. With this adaptation, Moore Threads has completed end-to-end engineering validation across three core AI4S scenarios: macromolecular structure prediction, whole-genome sequence modeling, and clinical medical imaging. This dovetails dynamically with yesterday’s report on RLinf × S5000 on the training side — the training side (embodied RL) and the inference side (AI4S) each landed a move in the same week, and the scenario narrative of “domestic compute power taking on scientific research workflows” continues to broaden.

3.2 Ascend: 910C NPU Visibility Fix and RL Training Chain Ready (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: build-infra #1105, #1107, #1103, FlagGems #6654

Two 910C configurations (CANN 9.0.0 / CANN 8.5.0) completed end-to-end GRPO validation of megatron_rl 0.17.1 within the same window (see 1.2), and resolved the NPU invisibility issue via privileged containers (an engineering failure where the Ascend-Docker-Runtime hook did not map the davinci die — a classic case of “environment problem ≠ code problem”). Supporting items: records for two megatron application image tags 2.2.0-0.2.3 for cann9.0.0 and cann8.5.0, and the sglang ascend-cann8.5.0 application page registered; on the operator side, grouped_matmul added to the FlagGems list, and FlagGems-vllm XCS supplementarily recorded.

3.3 Kunlunxin: Large XPU Fix Batch and P800 Test Machine Migration (09-24)

Date: 2026-09-24　Source: FlagGems #6617 et al. (see 1.5), build-infra #1093

Nearly 20 commits in a single window, covering fixes and performance fast paths across flash attention, GEMM (scaled_mm), convolution families, element-wise and special functions, and dropout backward — the most intensive round of preparation for this platform ahead of the 2.2 release; test infrastructure simultaneously migrated to the klx-p800 machine model (#6617), with the sglang kunlunxin-xre5.37.1 application page registered (#1093). An intensive fix batch and a machine model migration appearing on the same day typically correspond to a platform acceptance sprint ahead of release.

3.4 Biren / Sunrise: Dual-Track Integration of SUPA and Sunrise (09-24)

Date: 2026-09-24 ~ 2026-09-25　Source: FlagGems #6367, vllm-plugin-FL #530, #555

  • Biren: The SUPA backend entered the line on the same day at two layers — “FlagGems operator library + vllm-plugin-FL inference plugin” (#6367 / #530) — isomorphic to the integration paths of previous chip vendors, marking the first appearance of a new chip partner on the FlagOS code side.
  • Sunrise: vllm-plugin-FL ported the 0.2.1 Sunrise/PTPU patch set to the vLLM 0.24 mainline (#555, merged in the early hours of the next day), fixing enforce-eager execution of Qwen3.6 27B / 35B-A3B on PTPU hardware, and using 14 per-patch guards to prevent import crashes caused by vLLM 0.24 API mismatches — a “version-line follow-up” for an existing partner.

3.5 MetaX / Hygon / Tsingmicro: metax3.6 Tag, complex Scalar Fix, tsingmicro Variants (09-24)

Date: 2026-09-24　Source: FlagTree tags page, FlagGems #6679, FlagTree #1274

  • MetaX: FlagTree 0.7.0+metax3.6 tag landed + SPEC cpp backend included (FlagTree #1107); FlagGems updated flagtree metax parameters in sync (#6637); MetaX vLLM benchmarks added to CI (#1280); sglang mthreads / hygon application pages registered on the same day.
  • Hygon: FlagGems fixed kernel parameter handling for complex scalar multiplication (#6679); FlagTree added hcu3.6 FlagGems baseline and test workflow (#1274); sglang hygon-dtk26.04 application page (standard + empty mode, #1089).
  • Tsingmicro: FlagTree tsingmicro3.3 / tsingmicro3.6 variants complete; Open3D-PIMC media follow-up report (see 2.2).

IV. Summary and Trend Observations

  1. Release engineering enters “scope and claims freeze”: Upgraded from release artifact freeze (tags / lists) to scope audit (community #115) — 18 FEPs retained, 6 deferred to 2.3, claims purified into “evidence-backed completeness.” Actions in the 3 days before GA are dominated by closing out and honest statements, with the 2.3 development narrative and the 2.2 delivery boundary formally cut apart.
  2. Development mainlines collectively step up: The day after ten components’ official tags landed, FlagSparse and FlagAudio mainlines immediately bumped to 0.4.0, and FlagAttention received a version update — the 2.3 development cycle has in effect already started, running in parallel with 2.2 GA.
  3. New chip integration and version follow-up continue on dual tracks: Biren entered the line at two layers on the same day (FlagGems + vllm-plugin-FL), and Sunrise completed its vLLM 0.24 line port; the three-layer isomorphic integration pattern of “operator library + inference plugin + compiler” recurs, becoming the standard path for external chips to integrate with FlagOS.
  4. Platform matrix density rises rapidly before release: build-infra shipped seven sglang application pages in a single night (one per platform across seven platforms), and FlagTree 0.7.0 variants expanded to over ten — 2.2’s “platform coverage table” is still densifying, and the verifiability of coverage is the core asset of this release round.
  5. Ascend 910C chain ready: Dual-configuration RL E2E passing + privileged container fix provide a reproducible evidence chain for 2.2’s Ascend claims in training / RL scenarios (directly corresponding to the “retain verification provenance” requirement emphasized in #115).

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full check of pushed_at across 54 repos 24 repos active within window
GitHub Commit search + per-repo verification Item-by-item verification within window 175 items (across 18 repos)
Tags FlagTree tags / releases atom Item-by-item verification 0.7.0 series includes new variants such as iluvatar3.6 / metax3.6 / tileir3.6 / tsingmicro3.3
Release engineering community #114 / #115 original text Item-by-item comparison L1 checklist still under review; 2.2 FEP scope audit opened (18 retained / 6 deferred)
Release engineering build-infra PRs and release-info portal Original text retrieval and verification 910C RL E2E; seven consecutive sglang application page releases; packaging pipeline split
News Google News RSS (25 query sets in Chinese and English, via proxy) 24-hour window filtering + item-by-item exclusion Zero hits on component terms (20th quiet window); 2 items retained on ecosystem side
Media IT之家 (reposted by ifeng) / Yicai Original text retrieval and review Protenix-v2 full-pipeline inference (3.1)
Media Sina Finance / GMW.cn / CNR.cn Link comparison Open3D-PIMC coverage matrix continues to expand, no new technical information (2.2)
Community HN Algolia Search review Zero hits
Historical comparison FlagOS daily reports from the previous three days Deduplication verification Items such as Protenix and Open3D included based on incremental additions after deduplication against existing coverage; stock market pieces excluded

Complete Source List