Research window: 2026-09-28 10:00 ~ 2026-09-29 10:00 (24 hours; Tuesday window; continues from the 09-28 daily report window with no gap) Sources: GitHub (org: flagos-ai, full review of pushed_at across 54 repositories, 26 repos active within the window; commit search of 285 entries across 15 repositories, plus item-by-item verification of 25 tags/releases across 21 repositories, full status of community #114 ~ #123, the build-infra six-wheel release line and image tag batch, and 20 deployments of the release-info portal), Google News RSS (24 sets of Chinese and English query terms, via proxy), reports from NetEase / Sina Finance / Phoenix News and others (see the appendix source list for details)

Issue Index

  • Today’s Highlights: 2.2 Full-Stack Release Landed — 25 tags/releases across 21 repos opened within the window; six operator libraries re-released after fixes (09-28 ~ 09-29)
  • I. Open Source Project Progress (GitHub Activity)
    • 1.1 Release Engineering: 2.2 Tags and Releases Fully Landed — Three Waves of Rollout (09-28 17:00 ~ 09-29 01:00)
    • 1.2 community: Eight Consecutive Release-Manifest Commits (#116 ~ #123) — From Module Manifest and Six-Tag Reissue to Six-Operator Source Fixes
    • 1.3 build-infra: Six Operator Library Wheels Published to Internal PyPI; vLLM 0.24.0 Application Images Bulk Re-recorded (09-28)
    • 1.4 build-infra (Part 2): flagtree 0.7.0 Runtime Switch and Image Record Wrap-Up (09-28 ~ 09-29)
    • 1.5 FlagGems: cross_attention PPU Specialization and Hygon conv2d; CI Migrated to flagcicd Platform (09-28)
    • 1.6 FlagGems-vllm: Iluvatar TLE Kernels and Eight-Vendor Quantization Batch (09-28 ~ 09-29)
    • 1.7 Operator Library Matrix: Six Libraries Re-released After Fixes and Version Normalization Wrap-Up (FlagFFT / FlagSparse / FlagDNN / FlagBLAS / FlagTensor / FlagAttention)
    • 1.8 Torch-FL: v2.10.0 Version Line in Place — Wheels Bound to PyTorch 2.10 and Vendor Naming (09-28 ~ 09-29)
    • 1.9 FlagTree: MTHREADS / SUNRISE Fixes and metax3.6 CI Baseline (09-28)
    • 1.10 FlagQuantum: Native CPU Adjoint Multithreading and Fused Kernel Batch (09-28)
    • 1.11 Other Activity: FlagCX / FlagScale / FlagTrain / NCC Collaboration Line / docs (09-28 ~ 09-29)
  • II. News Coverage and Ecosystem
    • 2.2 Component-Level Search: Twenty-Second Consecutive Quiet Window (09-28 ~ 09-29)
    • 2.2 Zhongzhi FlagOS Technology Conference Agenda Details Updated: Main Forum Adds “FlagOS 2.2 Release” Session (09-28 Redistributed)
    • 2.3 Hygon × SinoSoft: EngineerAI Computing Appliance Launched — DCU Lands in Industrial Scenarios (09-28)
  • III. Member Unit Deep Dive
    • 3.1 Image Matrix: Member Vendors’ 0.3.0 Application Images Bulk Deployed (09-28 Late Night)
    • 3.2 Hygon: DCU Fix Line, INT8 Attention and LAMB Kernels (09-28)
    • 3.3 Ascend / MetaX / Moore Threads: KMCompiler, metax3.6 CI and MUSA Kernel Fixes (09-28)
    • 3.4 Iluvatar / Enflame: iluvatar TLE Kernel Batch and enflame Runtime Switch (09-28 ~ 09-29)
    • 3.5 Tsingmicro / Cambricon / Kunlunxin: Backend Extensions and Image Deployment (09-28)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Complete Source List

Today’s Highlight: 2.2 Full-Stack Release Landed — 25 Tags/Releases Across 21 Repos Opened Within the Window (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: community #116, #121, #123, FlagGems v5.4.0, FlagScale v2.1.0, release-info portal

09-28 was GA day for 2.2 (the previous daily report window closed at 10:00 that day). This window fully covers the external landing process for the release artifacts — tags, Releases, wheels, images, and the portal, all five artifact categories, were opened within the same window:

  • Release scale: 25 tags/Releases landed across 21 repos within the window (including 4 FlagTree variant tags, the dual version lines of vllm-plugin-FL, and FlagAudio v0.3.0 released in the same window though outside the manifest). Of the 2.2 official manifest (community release-2.2.yaml, 24 items across L0 / L1 / L2 tiers), 21 completed release within the window — Torch-FL is the sole pending item (version number already set at v2.10.0), while the FlagTree triton3.5 / 3.3 variants follow their established tags.
  • Three waves of rollout: First wave (around 17:00 on 09-28) — FlagGems v5.4.0 (17:23), Megatron-LM-FL v0.3.0 (17:23), flagcx v0.14.0 (17:53), FlagScale v2.1.0 and TransformerEngine-FL v0.3.0 (17:53), FlagTree enflame / sunrise variants (17:05 / 17:15); second wave (20:16 ~ 22:15) — SGLang-Plugin-FL v0.2.0 (20:16), with flagtree-cpu v0.1.0, FlagGems-vllm v0.2.0, FlagGems-sglang v0.1.0, vllm-plugin-FL v0.3.0 + v0.2.2, KernelGen v2.2.0, KernelGenBench v0.2.0, FlagOS-Compressor v0.1.0 and others landing in a concentrated burst from 20:27 to 20:28; third wave (01:12 ~ 01:13 on 09-29) — six L1 operator libraries (FlagFFT / FlagSparse / FlagDNN / FlagBLAS / FlagTensor / FlagAttention) completed final release after the source correction was merged (#123).
  • Correction loop closed: The six operator libraries’ sources were switched to use “the last commit on the mainline of the main branch before 2026-09-28 09:00” (FlagDNN / FlagBLAS use the master branch) as the cutoff snapshot — the correction merge (01:12) and the six libraries’ re-release (01:12 ~ 01:13) connected within the same minute-level window, locking the release artifacts to a non-drifting cutoff.
  • Artifact sync: Six pure-Python operator library wheels (FlagAttention 0.4.0, FlagSparse / FlagBLAS / FlagAudio 0.3.0, FlagFFT / FlagGems-vllm 0.2.0) were published to the internal PyPI (build-infra #1250, CI run 36425218655, all pullable via pip download); the vLLM 0.24.0 application image landed in bulk under the 2.2.0-0.3.0 tag across roughly 20 records spanning 11 platform families (see 1.3 / 3.1); the release-info portal saw 20 automated deployments within the window.

Assessment: The “ready to release at any time” status of the previous window was redeemed in this window as “released” — release discipline (better slow than wrong, re-release after correction, cutoff-locked versioning) held throughout; Torch-FL is the only gap, with the formal external narrative milestone anchored to the 10-17 conference (see 2.2).

1. Open-Source Project Progress (GitHub Activity)

1.1 Release Engineering: 2.2 Tags and Releases Fully Landed — Three Waves (09-28 17:00 ~ 09-29 01:00)

Date: 2026-09-28 ~ 2026-09-29　Source: FlagGems v5.4.0, FlagScale v2.1.0, FlagTree Releases, vllm-plugin-FL Releases, Megatron-LM-FL v0.3.0, FlagCX v0.14.0, SGLang-Plugin-FL Releases

  • Release inventory (21 repos / 25 tags, all atomically verified one by one against releases and tags): FlagGems v5.4.0 (1247 PRs since v5.3.0 — new operators including KernelGen-line argsort / isneginf and SiliconFlow-line _scaled_mm / segment_reduce / nanmedian, plus the Op Diff tool); FlagScale v2.1.0 (32 PRs — GLM5 training support, Tsingmicro txda platform backend, KERV training/inference integration, GPU progress heartbeat monitoring); FlagTree 0.7.0+triton3.6 (199 PRs — TLE-line extract_tile / insert_tile (HCU), optional TileIR backend integration, Sunrise / MetaX TLE support); vllm-plugin-FL v0.3.0 (vLLM 0.24.0 compatibility line, 64 PRs — W8A8 quantization adaptation, Ascend E2E port, BreakableCUDAGraphWrapper support for third-party backends) + v0.2.2 (0.20.2 maintenance line); Megatron-LM-FL v0.3.0 (GLM5 / 5.1 / 5.2 DSA structural models, sm90 fused DSA kernels, native Ascend MegatronAdaptor integration); FlagGems-vllm v0.2.0 (125 PRs — TLE split-kv, NSA line, chunk KDA / GLA, DeepSeek-V4 mega MoE input); FlagGems-sglang v0.1.0 (first release — KernelGen-line gemma_rms_norm, fused MoE, fused gated delta rule decoding, etc.); FlagOS-Compressor v0.1.0 (first release — safetensors weight conversion, INT4 / INT8 quantization and W8A8 MoE export); flagtree-cpu v0.1.0 (first release — AArch64 SVE2/NEON I8MM fallback, Qwen3.8-27B workflow on Apple Silicon); KernelGen v2.2.0, KernelGenBench v0.2.0 (multi-chip cuBLAS / CANN backends, S-tier anti-cheat sandbox, MmShapeBench), SGLang-Plugin-FL v0.2.0, TransformerEngine-FL v0.3.0, and six L1 operator libraries (see 1.7).
  • Timeline: The first wave (17:05 ~ 17:53) was the “advance batch”; at 20:27 ~ 20:28 a batch landing of 14 repos occurred (Release objects generated in a burst); at 22:15 and 00:11 the FlagTree mthreads variant and the Sunrise application tag were added respectively; at 01:12 ~ 01:13 the six operator libraries’ corrected re-release was completed.
  • Manifest status: Of the 24 items in the official release-2.2.yaml manifest (final state), 23 are ready (tag-status=existing); only Torch-FL is pending (the v2.10.0 tag does not yet exist, and the target commit remains a proposal); the six operator library entries are marked release-source=main-branch cutoff; flagtree-triton3.5 / 3.3 reuse the existing 09-24 tags.
  • Implication: From “manifest definition” (09-28 morning) to “full-stack landing” (09-28 afternoon through 09-29 early morning) took under 12 hours — 2.2’s release execution has reached a reproducibly installable completion state.

1.2 community: Eight Consecutive Release-Manifest Commits (#116 ~ #123) — From Module Manifest and Six-Tag Re-Issuance to Six-Operator Source Correction

Date: 2026-09-28 ~ 2026-09-29　Source: community #116, #117, #118, #119, #120, #121, #122, #123

  • #116 (12:10) defines the official module manifest: The release-2.2.yaml manifest lands — listing versions and target commits in three tiers: L0 (compilers and communication libraries) / L1 (core and domain operator libraries) / L2 (framework adaptation: PyTorch plugins, training/inference frameworks, and quantization).
  • #117 (14:19) operator snapshot: L1 operator libraries select target commits based on the mainline snapshot before 09-28 09:00.
  • #118 (14:47) flagtree-cpu: Records the independent release sequence v0.1.0.
  • #119 (15:43) vLLM tag alignment: Clarifies the reuse status of official vLLM tags in the 2.2 release.
  • #120 (17:22) release policy: Specifies that this batch “reuses existing tags and does not create or move tags”; FlagAudio is removed from the 2.2 release manifest; FlagTree variants and CPU entries are handled separately.
  • #121 (17:54) six-tag re-issuance: Six previously missing tags (flagcx v0.14.0, TransformerEngine-FL v0.3.0, FlagOS-Compressor v0.1.0, FlagScale v2.1.0, KernelGen v2.2.0, KernelGenBench v0.2.0) are re-issued at the recorded commits; Torch-FL is left for the next batch.
  • #122 (18:01) Torch-FL version finalized: The Torch-FL version in the manifest is set to v2.10.0.
  • #123 (01:12) six-operator source correction: The sources of FlagFFT / FlagSparse / FlagDNN / FlagBLAS / FlagTensor / FlagAttention are corrected from “RC2 branch HEAD” to “main branch 09:00 cutoff”; the three FlagGems-family libraries retain their existing tags.
  • Status notes: The earlier manifest proposal #114 (adding L1 operator libraries to the manifest with RC2 branch HEAD as target) was closed unmerged (18:10), its scope settled by the route restructuring in #116 / #117 / #123; #115 (defining the delivered FEP scope for 2.2, with the remainder deferred to 2.3) remains open with no action in the window.

1.3 build-infra: Six Operator Library Wheels Published to Internal PyPI; vLLM 0.24.0 Application Images Batch Re-Recorded (09-28)

Date: 2026-09-28　Source: build-infra #1246, #1247, #1250, #1249, #1252, tags record batch #1253 ~ #1272

  • Six wheel releases: #1246 establishes the packaging/flaglibs/ release line (build.sh + build-config + decision record + CI workflow) — FlagAttention / FlagSparse / FlagBLAS / FlagAudio / FlagFFT / FlagGems-vllm are built as py3-none-any pure-Python wheels from their respective release tags; #1247 fixes a heredoc escaping issue in build-config parsing; #1250 marks them “published”: the six wheels are uploaded to flagos-pypi-hosted (full CI pipeline passes, verified both via index and pip download). Key decisions for the release line: pure-Python packages are not split into per-vendor indexes (JIT compilation provides backend adaptation at runtime), and the wheel version is a pure function of the build ref.
  • vLLM 0.24.0 image re-recording: #1252 authorizes rebuilding the 0.24.0 application images against plugin v0.3.0; from late night into the next morning, the 2.2.0-0.3.0 tag lands across 20 records spanning nvidia (cuda12.8 / 13.3), Ascend (cann8.5.0 / 9.0.0 / 910c dual variants), Hygon, MetaX (3.7.2.1 / 3.8.1.3), Moore Threads (4.3.6 / 5.2.0), Iluvatar (4.4.0 / 4.5.0), Enflame (1.9.10 / 1.10.6), Cambricon (4.4.3 / 4.7.2), Kunlunxin, Tsingmicro, and Sunrise (19 records batched at 22:18 ~ 22:44, #1253 ~ #1272; Sunrise added at 00:11, #1277).
  • Daytime advance batch (14:38 ~ 16:46): 12 platform records for 2.2.0-0.2.2 (vLLM 0.20.2 line) (#1229 ~ #1240, covering Ascend / Moore Threads / Hygon / nvidia / Cambricon / Enflame / Tsingmicro / Sunrise); Iluvatar added in the evening (#1245, 20:11), and Iluvatar’s 0.24.0 application was authorized in advance for rebuild against the plugin rc2.post3 line (#1249, 21:17).
  • sglang / megatron lines: #1251 (Iluvatar) / #1280 (Sunrise) record sglang application tags; #1243 records the Iluvatar megatron application tag; #1277 ~ #1279 add Sunrise’s vllm (0.3.0 / 0.2.2) and megatron application tags.
  • Implication: The application image matrix completed its overall switch to the vLLM 0.24.0 compatibility line on GA day — release artifacts advanced from “tags” to “pullable images.”

1.4 build-infra (Part 2): flagtree 0.7.0 Runtime Switch and Image Record Wrap-Up (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: build-infra #1242, #1281, #1282, #1283, #1273 ~ #1276

  • Runtime switch: #1242 switches the Enflame dual-line and Sunrise application builds to the flagtree 0.7.0 runtime (the Iluvatar 4.4.0 runtime falls back to the 0.6.1 baseline; #1271 authorizes rebuilding the Sunrise application on the rebuilt runtime); #1281 / #1282 / #1283 complete, on the morning of 09-29, the 0.18.2 relocated wheel, the Enflame 0.7.0 rebuild authorization, and changelog date backfill for superseded 0.20.2 tags.
  • Iluvatar changelog chain: #1273 ~ #1276, four consecutive commits — 4.4.0 rc2.post3 record re-authorization, reason restoration, date backfill, and 4.5.0 rc2.post2 date backfill.
  • Docs: #1241 adds the vllm0.24.0-arm64-triton3.7.2 application page — a standalone entry for the ARM64 platform in the application matrix.
  • Observation: Application image builds are beginning to bind to the flagtree 0.7.0 runtime — compiler and application version pairing advances in sync during release week.

1.5 FlagGems: cross_attention PPU Specialization and Hygon conv2d; CI Migrated to the flagcicd Platform (09-28)

Date: 2026-09-28　Source: FlagGems #6668, #6700, #6740, #6737, #6660, #6717, #6665, #6648

  • Operators: #6668 adds a dense cross attention operator (with a PPU-specialized implementation); #6700 Hygon conv2d optimization; #6715 Kunlunxin logging completion.
  • Ascend KMCompiler: #6665 rewrites block size selection and the sparse / linear paths; #6648 trims Ascend performance and functional test cases.
  • CI infrastructure: #6737 migrates workflows to the flagcicd self-hosted runner platform; #6660 onboards MetaX testing to flagcicd; #6740 upgrades the operator test image to flagtree v0.7.0; #6717 bumps the MetaX version in backends.yaml to 0.7.0.
  • Implication: On the day after the v5.4.0 release (within this window), the FlagGems main branch shows no freeze pause — operators and test infrastructure continue to advance on both fronts.

1.6 FlagGems-vllm: Iluvatar TLE Kernels and an Eight-Vendor Quantization Batch (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: FlagGems-vllm #866, #820, #836, #858, #842, #805, #822, #779, #861, #853

  • Iluvatar: #866 (09-29 09:41) adds a TLE-optimized topk_softplus_sqrt kernel; #861 adds gemma_rms_norm.
  • Hygon: #820 INT8 varlen flash attention; #842 dual decode / prefill backends for top_k_per_row.
  • Ascend: #836 fused inverse-RoPE int8 quantization; #853 mhc_pre specialization (with fine-grained control and truncation).
  • Moore Threads: #858 W8A8 Block FP8 matrix multiplication optimization; #822 ports and optimizes fused inv-RoPE fp8 quantization.
  • DAMO XuanTie (PPU): #805 per-token grouped fp8 quantization (thead backend); #779 flash_attn_varlen w8a8_int8 for PPU.
  • Version line: #865 bumps to 0.3.0-dev1 (the dev line starts on the day v0.2.0 is released); #867 / #869 build and import normalization.
  • Implication: Quantization (INT8 / FP8) and linear-attention-family kernels cover eight vendors in a batch — the inference-side kernel library maintains the densest contribution cadence during release week.

1.7 Operator Library Matrix: Six-Library Corrected Re-Release and Version Unification Wrap-Up (FlagFFT / FlagSparse / FlagDNN / FlagBLAS / FlagTensor / FlagAttention)

Date: 2026-09-28 ~ 2026-09-29　Source: FlagFFT v0.2.0, FlagSparse v0.3.0, FlagDNN v0.3.0, FlagBLAS v0.3.0, FlagTensor v0.3.0, FlagAttention v0.4.0

  • Six-library contents: FlagFFT v0.2.0 (PPU backend and libtriton_jit code generation, Ascend NPU 1D/2D/3D FFT, 2D C2C/Z2Z decomposition and Rader prime-length algorithm); FlagSparse v0.3.0 (spmv_bsr optimization and spsv_sell; CUDA / DCU merge for the NCIC-AlphaSparse collaboration line); FlagDNN v0.3.0 (conv_dgrad / conv_wgrad, sdpa_fp8_backward draft, Ascend op and runtime, Hygon backend); FlagBLAS v0.3.0 (static dispatch, dgemm/cgemm/zgemm, NVIDIA group gemm, remaining Moore Threads FP64 L2 kernels, fp64 skip items treated as passing); FlagTensor v0.3.0 (36-operator C++ wrapper (TritonJIT), PPU / Ascend / MetaX C550 backends, contraction and trinary operators); FlagAttention v0.4.0 (MSA / GLA / SageAttention / GDN2 TLE operators with Enflame / Hygon / MetaX / Moore Threads multi-backend support).
  • Version unification wrap-up: The six libraries (and FlagAudio) complete, within the window, the “dynamic version (read from installed metadata) + drop scikit-build-core + changelog” overhaul — a single source of truth for version numbers, clearing the way for wheel-based packaging; the heredoc fix in #1247 was hands-on troubleshooting for this line.
  • Corrected re-release: The six libraries’ tags/Releases complete their final refresh at 01:12 ~ 01:13 after #123 merges (01:12) (source is the 09-28 09:00 mainline cutoff).
  • Related lines: FlagSparse has 20 commits in the window (biv sddmm optimization and tuning, DCU spmv multi-card test case fixes, CI refinement, PRs #102 ~ #107); FlagBLAS has 15 (#136 ~ #142 range, including Moore Threads FP64 L2 and build fixes); FlagGems-sglang batch4 merged (#132) and lazy-import test overhaul (#129 / #130).

1.8 Torch-FL: v2.10.0 Version Line in Place — Wheels Bound to PyTorch 2.10 and Vendor Naming (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: Torch-FL #468, #471, #465, #467, #462, #461, #460, #457, #455

  • Version and packaging: #468 names the wheel version after the bound PyTorch line (2.10.0); #471 (09-29 09:56, the last key commit in the window) names the wheel version by vendor and explicitly declares FlagTree / FlagGems / FlagCX dependencies — the first time release artifact dependencies are made explicit.
  • DCU line: #465 packages the DCU native communication bridge into the wheel; #402 standalone nccl extension build; #458 pins native_batch_norm to CUDA boxing.
  • MUSA line: #467 standalone stream class with explicit vendor dispatch; #462 migrates fp64 addmm/baddbmm to mudnn and rejects non-runnable items.
  • Architecture: #460 defines the backend autoload and import order contract; #457 separates bootstrap from optional framework hooks; #455 keeps installed PyTorch immutable; #456 replaces existence assertions with sys.modules diff testing; #470 cp310 CI compatibility (tomllib-free).
  • Status: Torch-FL is pending in the community manifest — the version line (v2.10.0) and packaging conventions are in place, awaiting only the tag landing; this is the last piece of the 2.2 release puzzle.

1.9 FlagTree: MTHREADS / SUNRISE Fixes and metax3.6 CI Baseline (09-28)

Date: 2026-09-28　Source: FlagTree #1302, #1300, #1299, #1297, #1298, #1296, #1295, #1290

  • Backend fixes: #1302 removes the generatePtr offset clamp in MTHREADS when block_shape exceeds shape; #1290 trims the Enflame enable_i64 gate per toolchain support (cherry-pick #1254) and makes parse_options idempotent.
  • CI baseline: #1297 adds the metax3.6 FlagGems baseline and test workflow; #1300 fixes the fuser process for the metax benchmark; #1298 fixes the PPU fuser cleanup script; #1296 / #1295 stability fixes (raw tests and VLLM_PLUGINS=fl environment).
  • Tags: 0.7.0+enflame3.6 (17:05) / +sunrise3.6 (17:15) / +mthreads3.6 (22:15) are released within the window — the Sunrise and Moore Threads variants are new release lines.

1.10 FlagQuantum: Native CPU Adjoint Multithreading and a Fused-Kernel Batch (09-28)

Date: 2026-09-28　Source: FlagQuantum #232, #231, #230, #229

  • Native CPU adjoint multithreading (#232), direct Triton compiler identity recording (#231), adjoint execution template reuse (#230), quantum kernel package structure established (#229); the fusion line covers rotation / adjoint segment fusion (#224 ~ #228), Z/ZZ adjoint gradients (#223), and a differentiable simulator corpus (#222).
  • Observation: Component-level engineering for the quantum-AI fusion direction (10-17 conference special track, see 2.2) continues to refine performance on the CPU side.

1.11 Other Activity: FlagCX / FlagScale / FlagTrain / NCC Collaboration Line / docs (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: FlagCX #630, FlagTrain #3, FlagScale v2.1.0, docs commits

  • FlagCX: #630 adds the PAL shared transport ordering foundation layer (a cross-backend transport consistency substrate).
  • FlagScale: Mainly the v2.1.0 release (see 1.1); the Tsingmicro txda backend enters the training framework platform manifest.
  • FlagTrain: #3 fixes LAMB kernel-mode speedup measurement on the Hygon platform.
  • docs: Model manifest update branch advances (modelscope / aihuanxin line, 09-29 morning).
  • NCC collaboration line (NCIC-AlphaSparse): 20 FlagSparse commits, including biv sddmm optimization and DCU multi-card test cases (see 1.7).

II. News Coverage and Ecosystem

2.1 Component-level search: 22nd consecutive quiet window (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia

  • Component and release-line queries (FlagOS / FlagGems / FlagScale / FlagTree / FlagCX / KernelGen / FlagPerf / 2.2) produced no new external release announcements within the 24-hour window — the 22nd consecutive quiet window; all release activity occurred on the code and artifact side (see Chapters I and III).
  • 10 hits on the ecosystem side within the window: the only one directly related to FlagOS was further dissemination of conference registration (see 2.2); one on the Hygon industry side (see 2.3); the rest were stock market pieces (related to Enflame Technology’s IPO), gambling SEO, commentary, and unrelated reposts, all excluded by topic.
  • HN Algolia returned no relevant hits for FlagOS / FlagGems / FlagScale / BAAI (fuzzy-match results were all unrelated topics).

2.2 Zhongzhi FlagOS Technical Conference agenda details updated: main forum to include a “FlagOS 2.2 release” segment (09-28 re-dissemination)

Date: 2026-09-28 (first published 09-23, re-disseminated within the window via channels including DAMO Developer Matrix)　Source: NetEase (AI科技大本营)

  • Main forum agenda: The 2026 Open AI Computing Conference and Zhongzhi FlagOS Technical Conference (10-17 / 10-18, Beijing Zhongguancun National Independent Innovation Demonstration Zone Exhibition Center) will have a main forum themed “From Token Economy to Physical AI,” with a content track including “From the Era’s Imperative of Open Intelligent Computing to the FlagOS 2.2 Release” — the external release segment for 2.2 is anchored to the conference’s main forum; Huang Tiejun (Chairman of BAAI) and Lin Yonghua (Vice President and Chief Engineer) are confirmed to attend.
  • Sessions and workshops: Six tracks (AI Systems / Embodied Intelligence / In-Vehicle Intelligent Computing / Quantum-Intelligence Fusion / International Open Cooperation in AI Computing / AI Agent); the Quantum-Intelligence Fusion track will launch the Quantum-Intelligence Cloud Platform plan on site; five hands-on workshops all day on 10-18 (KernelGen operator engineering, on-device inference, Triton-TLE, quantum-intelligence practical operations, FlagScale-Agent).
  • Note: Conference registration information was covered in the 09-20 issue; this item is an update on agenda and speaker details (early-bird tickets from 59 yuan).

2.3 Hygon × SinoSoft: EngineerAI computing appliance launched — DCU lands in industrial scenarios (09-28)

Date: 2026-09-28　Source: Sina Finance (ZOL), Phoenix Net

  • At the “Second Symposium on Domestic Computing Power Applications for Key Nuclear Power Information Technology,” SinoSoft and Hygon Information jointly launched the EngineerAI computing appliance: a C86-architecture CPU + DCU dual-chip collaborative computing base, available in four versions tiered by crew / station / department / group load levels, targeting role-based AI for industrial vertical scenarios (a next-generation command center intelligence product for nuclear power).
  • Relationship to this stack: Hygon DCU is a FlagOS member platform — in this window the hygon application image 2.2.0-0.3.0 was committed to the repository and the DCU fix line remains active (see 3.2); industrial / energy scenarios are a depth play for DCU beyond intelligent computing centers, mirroring FlagOS’s “cloud-to-edge, all-scenario” direction.

III. Deep Dive into Member Organizations

3.1 Image Matrix: Member Vendors’ 0.3.0 Application Images Land in Batch (Late Night 09-28)

Date: 2026-09-28 ~ 2026-09-29　Source: build-infra #1253 ~ #1272, release-info portal

  • Late-night batch (22:18 ~ 22:44, 19 entries; 20 total after Sunrise was added at 00:11 the next day): vLLM 0.24.0 compatibility line (plugin v0.3.0) application images 2.2.0-0.3.0 landed across 11 platform families — Hygon dtk26.04, MetaX maca3.7.2.1 / 3.8.1.3, Moore Threads musa4.3.6 / 5.2.0, Iluvatar corex4.4.0 / 4.5.0, Enflame tops1.9.10 / 1.10.6, Cambricon neuware4.4.3 / 4.7.2, Kunlunxin xre5.37.1, Tsingmicro tsm260610, Ascend cann8.5.0 / 9.0.0 (including dual 910c variants), Sunrise tangrt1.2.0, and dual nvidia lines.
  • Significance: This is the most direct evidence of 2.2’s “deployability” — seven member vendors plus platforms such as Ascend all landed on the vLLM 0.24.0 compatibility line the day after GA; the portal (release-info) tracked 20 deployment updates in sync.

3.2 Hygon: DCU Fix Line, INT8 Attention, and LAMB Kernel (09-28)

Date: 2026-09-28　Source: FlagGems #6700, FlagGems-vllm #820, #842, Torch-FL #465, #458, FlagTrain #3, FlagSparse commits

  • Operator layer: FlagGems Hygon conv2d optimization (#6700); FlagGems-vllm INT8 varlen flash attention (#820) and top_k_per_row decode / prefill backends (#842).
  • Framework layer: Three Torch-FL DCU entries (native communication bridge into wheel #465, native_batch_norm pinned to CUDA boxing #458, standalone nccl extension #402).
  • Training kernels: FlagTrain LAMB kernel-mode speedup measurement fix (#3).
  • Sparse line: FlagSparse DCU spmv csr multi-card test case fix.
  • Industry side: SinoSoft EngineerAI compute appliance (see 2.3).

3.3 Ascend / MetaX / Moore Threads: KMCompiler, metax3.6 CI, and MUSA Kernel Fixes (09-28)

Date: 2026-09-28　Source: FlagGems #6665, #6648, #6660, #6717, FlagGems-vllm #836, #853, FlagTree #1300, #1297, #1302, Torch-FL #467, #462

  • Ascend: Two FlagGems KMCompiler entries (block size and sparse/linear rewrite #6665; test case trimming #6648); FlagGems-vllm fused inverse-RoPE int8 quantization (#836) and mhc_pre specialization (#853); three image lines (cann8.5.0 / 9.0.0 / 910c) landed.
  • MetaX: FlagTree metax benchmark fuser fix (#1300) and metax3.6 baseline and test workflow (#1297); FlagGems tests migrated to flagcicd (#6660) and backends.yaml bumped to 0.7.0 (#6717); images 3.7.2.1 / 3.8.1.3 landed.
  • Moore Threads: FlagTree generatePtr clamping fix (#1302); Torch-FL standalone stream class and explicit dispatch (#467), fp64 addmm/baddbmm migrated to mudnn (#462); images 4.3.6 / 5.2.0 landed.

3.4 Iluvatar / Enflame: iluvatar TLE Kernel Batch and enflame Runtime Switch (09-28 ~ 09-29)

Date: 2026-09-28 ~ 2026-09-29　Source: FlagGems-vllm #866, #861, build-infra #1242, #1282, FlagTree #1290

  • Iluvatar: FlagGems-vllm TLE-optimized topk_softplus_sqrt (#866) and gemma_rms_norm (#861); on the build-infra side, iluvatar rebuild authorization and changelog chain (#1249 / #1273 ~ #1276); images corex4.4.0 / 4.5.0 landed.
  • Enflame: flagtree 0.7.0 runtime switch covering both Enflame lines (#1242) and rebuild authorization (#1282); FlagTree enable_i64 gating trimmed per toolchain (#1290); FlagGems-sglang adds Enflame backend support; images tops1.9.10 / 1.10.6 landed.

3.5 Tsingmicro / Cambricon / Kunlunxin: Backend Expansion and Image Landing (09-28)

Date: 2026-09-28　Source: FlagScale v2.1.0, FlagGems #6715, build-infra tags record batch

  • Tsingmicro: FlagScale v2.1.0 adds the txda (Tsingmicro) platform backend — Tsingmicro enters the training framework backend matrix; image tsm260610 landed.
  • Cambricon: Images neuware4.4.3 / 4.7.2 dual-line 2.2.0-0.3.0 (and 0.2.2) records landed.
  • Kunlunxin: Image xre5.37.1 landed; FlagGems Kunlunxin logging completed (#6715).

IV. Summary and Trend Observations

  1. Release landing: 2.2 moves from “checklist” into “repositories” — 21 repositories and 25 tags/Releases landed within 12 hours; six wheels uploaded to the internal PyPI, application images re-recorded in batch, and 20 deployments on the portal — with “tag + Release + wheel + image + portal” all five in place, 2.2 enters a reproducibly installable state.
  2. The correction loop demonstrates release discipline — the six-operator-library source fix (#123) and the re-release were connected within the same minute; the checklist wording changed from “RC2 HEAD” to “mainline cutoff snapshot” — release artifacts are anchored to non-drifting commits, “better to adjust the wording than chase a moving source.”
  3. Torch-FL is the only blank — the v2.10.0 version line and the vendor naming packaging spec (explicit wheel dependency declarations) are in place, awaiting only the tag landing — the other side of release discipline: no go-ahead if standards aren’t met.
  4. Development line and release line are clearly on separate tracks — dev version numbers rise collectively for FlagGems-vllm 0.3.0-dev1, FlagSparse / FlagBLAS / FlagAudio 0.4.0-dev1, FlagFFT 0.3.0-dev1, and others; the NCIC collaboration line (FlagSparse) remains active — the 2.3 cycle and the 2.2 wrap-up advance in parallel.
  5. External narrative anchored to the 10-17 conference — the main forum will feature a “FlagOS 2.2 release” segment; the news side marks a second consecutive quiet window, contrasting with off-field industry moves (Hygon × SinoSoft appliance) — the norm is driven by the code side, while release and conference milestones carry the external narrative.
  6. Multi-platform coverage reaches its widest this year — the application image matrix covers 11 platform families, and the quantization kernel batch spans eight vendors — the delivery surface of “develop once, run on multiple chips” peaks during release week; the next observation points are the Torch-FL landing and the convergence of the 2.3 development line.

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full verification of pushed_at across 54 repos 26 repos active within window (15 of which have commits within window)
GitHub Commit search + per-repo cross-check Line-by-line verification of 285 commits from commit search (full total_count) 285 commits across 15 repos (164 from build-infra); no new repos within org
Releases releases / tags atom Line-by-line comparison of 25 tags across 21 repos All recorded; FlagTree ×4, vllm-plugin-FL ×2 are multi-tag repos
Releases releases API re-check published_at sampling across five tiers FlagGems 17:23, Megatron 17:23, FlagScale 17:53, TE-FL 17:53, FlagAttention 01:11 after correction
Release Engineering community #114 ~ #123 original text Line-by-line comparison of status and diffs #116 ~ #123 all merged; #114 closed without merge; #115 still open
Release Engineering 2.2 manifest (release-2.2.yaml) Original text retrieval and cross-check 21 of 24 items released within window; Torch-FL pending; triton3.5/3.3 reuse existing tags
Release Engineering build-infra PR and release-info Original text retrieval and cross-check Six wheels published (CI verification passed) and ~20 image tag records; portal deployed 20 times
News Google News RSS (24 sets of Chinese/English query terms, via proxy) 24-hour window filtering + line-by-line elimination Zero additions on component line (22nd quiet window); 2 items retained on ecosystem side
Media NetEase / Sina Finance / ifeng Original text retrieval and re-check Conference agenda update (re-dissemination); Hygon × SinoSoft all-in-one machine
Community HN Algolia Search re-check No relevant hits
Historical Comparison Previous two days of FlagOS daily reports Deduplication verification Conference registration 09-20 already reported (this time only 2.2 session details); Enflame IPO 09-10 / 09-14 already reported (this time market article eliminated)

Complete Source List