FlagOS Daily Intelligence Report (2026-09-18)
Research window: Past 24 hours (2026-09-17 10:18 ~ 2026-09-18 10:18, Beijing Time) Sources: GitHub (org: flagos-ai, full verification of pushed_at across 54 repositories + full retrieval of 147 commit search results verified by committer-date + commit re-verification on default branches of key repositories + tags/releases metadata + direct raw retrieval of the FEP directory and 2.2 release checklist/schedule from community + key PR patches and file lists from build-infra + full ref verification of the rc2 branch via git ls-remote), Google News RSS (33 sets of query terms in Chinese and English, via proxy), HN Algolia, BAAI Community, Jiwei.com, EE Times China, etc. (see appendix source list for details)
Index
- Today’s Highlights: community merges 20 FEPs in one shot, FlagOS 2.2 feature list and release boundary both finalized (09-17)
- I. Open Source Project Progress (GitHub Activity)
- 1.1 community: 20 FEPs merged in one batch, 2.2 feature list finalized (09-17)
- 1.2 RC2 manifest and integration branches: 24 module entries, ten repos cut rc2 branches in sync (09-17)
- 1.3 build-infra bumps versions twice in a row: FlagGems to rc2.post3, Moore Threads FlagTree to 0.7.0rc2 (09-18)
- 1.4 build-infra: Ascend FlagTree wheel upload pipeline goes through for the first time (09-17)
- 1.5 FlagGems: KernelGen lands another dozen-plus Nvidia operators in a single day, AMD W7900D enters weekly tests (09-17/09-18)
- 1.6 FlagTree: DEB packaging and Nexus publishing folded into build-infra shared workflow (09-17/09-18)
- 1.7 Torch-FL: environment variable surface consolidated under a single FLAGOS prefix (09-17/09-18)
- 1.8 FlagQuantum: coverage floor raised two levels, quantum error correction record layer lands (09-17/09-18)
- 1.9 FlagFFT: MUSA and MACA portable 3D transpose validation records, dependencies pinned uniformly (09-17)
- 1.10 FlagCX: lane mask ABI extended to 64-bit, Fedora 43 enters RPM matrix (09-18)
- 1.11 FlagAudio: first three PRs for the audio operator library merged (09-17)
- 1.12 Other activity: FlagDNN, FlagGems-vllm, FlagGems-sglang, flir, FlagScale (09-17/09-18)
- II. News Coverage and Ecosystem
- 2.1 Component-level search: fourteenth consecutive quiet window, zero hits in 24 hours (09-17~09-18)
- 2.2 ZDTaichu5.0-9B open-sourced via BAAI community relay: no interface with the FlagOS stack, not covered in this issue (09-16~09-17)
- 2.3 Member unit capital markets: MetaX stock price anomaly and institutional reduction, Enflame IPO draft (09-17)
- III. Member Unit Deep Dives
- 3.1 Hygon: fused rrelu backend kernel merged, decoupled DCU wheel fixes Qwen-Image (09-17/09-18)
- 3.2 Moore Threads: MUSA-line FlagTree jumps to 0.7.0rc2, vLLM 0.20.2 image gets plugin installed for the first time (09-18)
- 3.3 Kunlunxin: 0.6.1+xpu3.6 pinning and tle.raw fast-path fallback (09-17)
- 3.4 MetaX: MACA portable 3D transpose validation and unified dependency pinning (09-17)
- 3.5 Iluvatar CoreX: FlagDNN platform architecture reorganization (09-17)
- 3.6 Ascend: CommonIR proof-of-concept marked as implemented, grouped_topk tuned via tle.dsa.raw (09-17)
- 3.7 DAMO XuanTie: T-Head enters FlagTree 2.2 multi-backend planning, PPU-line CI rollout (09-17 verification)
- 3.8 Tsingmicro: distributed primitives enabled on backend, converging with 2.2’s TLE distributed roadmap (09-17)
- 3.9 BAAI (lead party): FEP governance implementation and RC2 manifest maintenance (09-17)
- IV. Summary and Trend Observations
- Appendix: Source Verification Table
- Appendix: Complete Source List
Today’s Highlight: community merges 20 FEPs in one batch, settling both the FlagOS 2.2 feature list and release boundaries
Date: 2026-09-17 Source: community FEP directory, FEP-0099 operator library, FlagOS 2.2 release schedule
Between 23:55 and 23:56 on 09-17, the community repo took 20 consecutive commits in a single minute, all prefixed FEP(sig-...), numbered 0085 through 0100, spanning seven areas: operators, compiler, framework, training, networking, edge, and AI for science. This batch of documents turns “what 2.2 will deliver” from a meeting consensus into files that can be checked off one by one inside the repo.
The most noteworthy aspect is that every FEP includes a “Release Boundary and Evidence” section — the baseline version for that area in 2.1, the version in the 2.2 candidate (RC2 list), and the development window — and closes by stating which parts have corresponding code and which remain roadmap only. Take the operator library: FEP-0099 lays out four compute targets for 2.2: push the total operator count to 635 (301 of them produced by the KernelGen workflow), bring attention-class operators in line with NVIDIA’s advanced level and run them on 5+ domestic chips, key operators based on the Triton Language Extension (TLE), and general optimization techniques such as low bit-width and fused large kernels; it lists six 2.1 baselines including FlagGems v5.3.0, 2.2 candidates including FlagGems v5.4.0-rc2.post1, and specifies the development window as 06-01 to 08-31.
Status differences within the same batch are also clearly stated: on the compiler side, FEP-0095, which introduces a Common IR, is marked implemented, but the body limits its scope to “a proof of concept for Ascend 910B/910C on the Triton 3.5 line,” links to FlagTree #974 (opened 08-13, merged 08-27), and declares that it does not amount to a production-grade unified path across hardware families; meanwhile FEP-0100 on KernelGen capabilities deliberately remains a draft, stating outright that “no corresponding implementation or implementation PR was found in public release tags,” refusing to treat the roadmap as a deliverable. This approach — having both implemented and explicitly unimplemented items in the same batch of documents — is the most practical aspect of this set of FEPs.
On the timeline, this batch lands in the middle of the testing and stabilization period: the release schedule sets feature freeze at 08-31, the testing period from 09-01 to 09-24, and GA at 2026-09-28, with the freeze rule stating that “both doors — FEPs and feature code — close simultaneously; miss it and it slips to the next version.” So these documents serve to complete the registration of the already-frozen scope, not to add anything new to 2.2 — with 10 days left until GA.
1. Open-Source Project Progress (GitHub Activity)
Window Overview: Of the 54 repos in the org, 28 had pushes during the window (up from 19 in the previous period — a clear surge in volume); a single commit search returned 147 commits within the window (both pages taken in full, sorted by committer-date descending), distributed across 14 repos: FlagGems 40, FlagFFT 21, community 19, FlagQuantum 17, Torch-FL 15, build-infra 7, FlagTree 7, FlagGems-sglang 7, FlagDNN 3, FlagCX 3, FlagAudio 3, FlagGems-vllm 3, FlagScale 1, flir 1. Another 14 repos had pushes during the window but no new commits on the default branch (FlagBLAS, FlagTensor, KernelGen, KernelGenBench, Megatron-LM-FL, TransformerEngine-FL, FlagRelease, sglang-plugin-FL, FlagOS-Compressor, FlagAttention, FlagSparse, etc.), clustered within the same minute range from 09-17 23:29 to 23:33; verified via git ls-remote as creation pushes of rc2 integration branches — see 1.2 for details.
The shape of this window = two tracks: “release registration” + “packaging and release artifacts”: On the governance side, the full set of FEP documents was filled in at once and the 2.2 release boundary was locked down; on the engineering side, attention clearly shifted from feature development to the release artifacts themselves — DEB packaging, wheel upload pathways, Nexus artifact sync, coverage gates, environment variable consolidation — all work at the level of “can it install, can it ship, is it correct after install.”
1.1 community: 20 FEPs merged in one batch, 2.2 feature list finalized (09-17)
Date: 2026-09-17 Source: community FEP directory, FEP documentation and version tracking
See “Today’s Highlights.” Two additional facts from the directory side: First, FEPs are stored in subdirectories by SIG, and this batch of new documents landed in sig-operator (0099), sig-compiler (0095 to 0098), sig-framework (0086 to 0091), sig-training (0092), sig-kernelgen (0093, 0100), sig-network (0085), sig-edge (0082 to 0084), sig-ai4s (0081), filling all active SIGs in one pass. Second, the edge-side FEPs are not just documents — they come with executable script directories (two sets under fep/sig-edge/scripts/: Triton 3.7 and arm64 inference), consistent with the edge-side trajectory observed in this report.
1.2 RC2 manifest and integration branches: 24 module entries, ten repos cut rc2 branches in sync (09-17)
Date: 2026-09-17 Source: release-2.2-rc2.yaml, release-branch-tag workflow
From 09-17 23:29 to 23:33, ten module repos showed pushes within the same minute range but no new commits on the default branch. Pulling all refs via git ls-remote confirmed that these repos all now have integration branches in the form 0.3.0-rc2 / 0.2.0-rc2 / 0.1.0-rc2 / 2.2.0-rc2, with corresponding v*-rc2.post1 tags already cut; FlagGems and FlagCX are further ahead, with tags already at rc2.post3 and rc2.post2. The release-branch-tag.yml in community is the executor of this action chain: it calls manage-release.py according to the release manifest, creating and pushing branches and tags in bulk under the release bot identity.
The manifest itself was also upgraded: the RC2 version lists 24 module entries, and a line-by-line comparison with RC1 shows the version lines shifted forward by one notch overall (e.g., FlagGems from rc1.post2 to rc2.post2, FlagAttention from 0.4.0-rc1.post1 to 0.4.0-rc2.post1), with a header note added explaining the convention of “splitting one repo into multiple entries, each tracking a different upstream line.” Two details worth noting: FlagTree, because it splits entries by Triton version, does not follow the global v-prefix tag convention; and FlagScale’s version number in the 2.2 manifest is v2.1.0-rc2.post1, already off the 2.0 line.
1.3 build-infra bumps versions twice in a row: FlagGems to rc2.post3, Moore Threads FlagTree to 0.7.0rc2 (09-18)
Date: 2026-09-18 Source: build-infra #925, build-infra #926, configs.yaml
Following the previous day’s action of bumping the base version to 2.2.0, build-infra landed two more version-bump commits on 09-18 at 07:32 and 08:35, both touching only configs.yaml. The first (#925) changed the flaggems field from 5.4.0-rc2.post2 to 5.4.0-rc2.post3; the second (#926) changed the Moore Threads backend’s flagtree pin from 0.6.1+mthreads3.6 to 0.7.0rc2+mthreads3.6. The same file’s header describes the release action as three steps — update the field, tag v<version>, done — so these two line changes mean that on the 2.2 image line, both the operator library and the Moore Threads compilation base moved to a newer round of candidates.
Viewed side by side: both the operator library and the compiler base are moving toward rc2 iteration positions (post3 / 0.7.0rc2), while the manifest’s default version field remains at 2.2.0 — the release artifacts are still being assembled and have not yet been formally tagged.
1.4 build-infra: Ascend FlagTree wheel upload pathway works end-to-end for the first time (09-17)
Date: 2026-09-17 Source: build-infra #923, build-infra #924
#923 fixed Ascend FlagTree packaging: previously neither Ascend target could produce a wheel (Ascend 3.5 failed on FlagPrism’s Metric.h due to an undeclared uint64_t), because after the packaging script first landed, all fixes were committed to a different branch while the merge went through another branch; this PR also renamed both target files by CANN version and added a cann8.5.0 target. #924 then replaced the open question of “upload path not yet verified” with a tested answer: on 09-17 both wheels reached the flagos-pypi-ascend artifact repository — flagtree-0.7.0rc2+ascend3.5 and flagtree-0.6.0+ascend3.2; the documentation also records the build times on Ascend nodes (26 minutes and 14 minutes, mostly spent compiling AscendNPU-IR and triton under MAX_JOBS=32) and one environment limitation: the pip version on CANN nodes is older than what --break-system-packages requires.
1.5 FlagGems: KernelGen adds another dozen-plus Nvidia operators in a single day, AMD W7900D enters weekly tests (09-17/09-18)
Date: 2026-09-17, 2026-09-18 Source: FlagGems #5846 quantile operators, FlagGems #6312 AMD weekly tests, FlagGems #5731
FlagGems’ 40 commits this period are the densest line in the org, with the majority being Nvidia operators produced by the KernelGen workflow landing in the repo, a dozen-plus added in a single day: quantile and mean types (nanquantile, nanmean), log-determinant (logdet), polar decomposition (linalg_polar), recurrent network backward kernels (GRU / LSTM cells and their backward passes), upsampling backward (Lanczos antialiasing), sparse matrix multiplication, feature dropout, packed linear layers, etc., plus two element-wise composite operators (and_tensor, and_scalar) and all tensor overloads of set_.
Three items on the backend and tooling side: [CI] Add AMD W7900D backend to weekly test workflow adds AMD’s W7900D to the weekly test matrix, a relatively new non-domestic backend line; the SiliconFlow backend adds a zero-copy real operator and switches to registration by dispatch key; a subsequent commit fills in the __init__.py for seven backend subpackages, fixing a structural defect of the “new backend package not picked up by the import chain” kind. There are also several gate fixes (operator export checks, KernelGen test markers scanning only PR-added lines).
1.6 FlagTree: DEB packaging and Nexus publishing merged into build-infra shared workflows (09-17/09-18)
Date: 2026-09-17, 2026-09-18 Source: FlagTree #1063, FlagTree #1066, FlagTree #1082, FlagTree #1170
Three of FlagTree’s seven commits this period are release-artifact engineering: switching release packages to push to the FlagOS Nexus artifact repository via build-infra’s shared workflow (#1063), adding a row to the DEB packaging matrix for ubuntu22.04 / python3.12 to cover the ABI combination (#1066), and stripping debug info from libtriton and libproton in the DEB package while skipping dh_dwz (#1076). The next day, a Bump packaging baseline to 0.6.1 and plumb the wheel version commit connected the packaging baseline to the wheel version number (#1082), so that artifact versions are no longer maintained by hand.
One compiler-side item is directly related to the 2.2 TLE trajectory: the tle.signal / tle.signal_wait frontend exposes synchronization scope and memory ordering as explicit parameters (#1170), corresponding to the semantic interface needed for “distributed primitives and communication-computation fusion” in FEP-0096.
1.7 Torch-FL: environment variable surface consolidated into a single FLAGOS prefix (09-17/09-18)
Date: 2026-09-17, 2026-09-18 Source: Torch-FL #341, Torch-FL #337, Torch-FL #320
Torch-FL’s 15 commits this period form a coherent refactoring chain: keeping only the in-tree build switch and removing the boxed and native mode variables (#336), unifying SDK discovery logic and giving PPU its own accelerator identifier (#337), accepting only vendor-native SDK environment variables (#338), aligning the runtime and routing environment variable tables with the code (#339), and finally unifying into a single variable surface with the FLAGOS_ prefix (#341); a CI commit in the same period retires the old FlagGems switches entirely (#320). Unlike the previous period’s “post-routing-switch fallout fixes,” this line is an interface surface consolidation — merging multiple historical switches into a single contract, which is the kind of work done before release to freeze the externally visible surface.
There are also two substantive platform-side fixes: making torch.cuda available on the decoupled DCU wheel (#332), and several Ascend semantic fixes (boolean SDPA masks, boolean mask indexing, nearest-neighbor upsampling staying on the device side).
1.8 FlagQuantum: coverage floor raised two levels, quantum error correction recording layer lands (09-17/09-18)
Date: 2026-09-17, 2026-09-18 Source: FlagQuantum #64, FlagQuantum #55, FlagQuantum #52
FlagQuantum’s 17 commits this period are almost entirely about establishing engineering quality standards: raising the global coverage floor from 60% to 75% with per-package floors (#55), bringing the entire package under strict type checking (#52), extending lint gates to modern type annotations and defect-category rules, assigning every test file to a tier marker, and wiring previously skipped JAX tests into CI with coverage-gain binding. One functional item is the first phase of the quantum error correction (QEC) recording layer (#64), implemented decoupled from any specific encoding scheme; on the refactoring side, nine twin writers were merged into a single implementation, and eleven nested functions with late-binding loop variable defects were fixed.
Against FEP-0081’s positioning of FlagQuantum in the AI-for-Science direction, this period’s focus is “establish the engineering baseline first,” not yet algorithmic scale expansion.
1.9 FlagFFT: MUSA and MACA portable 3D transpose verification records, dependencies pinned to a unified version (09-17)
Date: 2026-09-17 Source: FlagFFT commit history, FlagFFT repository
FlagFFT’s 21 commits this period are mainly merges of hardware adaptation branches and verification records: merging the hardware profile branch with the MUSA (Moore Threads), MACA (MetaX), and CUDA feature branches respectively, leaving one verification record document for each, with the MACA side having two consecutive commits enabling the portable register-tiled 3D transpose (platform verification first, then formal enablement). On the performance side, one commit changes large complex leaf nodes to be staged through a single shared buffer, and incidentally fixes two issues in native precision collection: the memory upper bound and file cache release. A final commit pins the Triton JIT runtime dependency to a unified backend revision, in the same direction as the version consolidation on the FlagTree side.
1.10 FlagCX: lane mask ABI widened to 64 bits, Fedora 43 enters the RPM matrix (09-18)
Date: 2026-09-18 Source: FlagCX #607, FlagCX #561, FlagCX #562
Three commits on the communication library side all landed on the morning of 09-18: widening the lane mask binary interface from 32 bits to 64 bits (#607, an ABI expansion made necessary as chip scale grows), replacing the standalone Nexus upload process with build-infra’s shared workflow (#561, in step with FlagTree), and adding Fedora 43 to the Nvidia RPM build matrix (#562). Taken together, what the communication library did this period is “interface capacity + artifact pipeline,” not new communication algorithms.
1.11 FlagAudio: first three PRs for the audio operator library merged (09-17)
Date: 2026-09-17 Source: FlagAudio #5 spectrogram operators, FlagAudio #4, FlagAudio #2 packaging
FlagAudio merged three PRs in the same morning, a rare functional action for this repo: one implementing spectrogram operators in Triton with accompanying tests, one fixing the gain identity check and changing the torchaudio import to lazy loading, and one wrapping up packaging (dependency and license declarations). All three came from external contributors’ branches, indicating that developers from outside the repo are starting to get involved; combined with FlagAudio’s inclusion among the 24 module entries in the operator layer in the FEP manifest, it is transitioning from placeholder status to a module with actual deliverables.
1.12 Other activity: FlagDNN, FlagGems-vllm, FlagGems-sglang, flir, FlagScale (09-17/09-18)
Date: 2026-09-17, 2026-09-18 Source: FlagDNN, FlagGems-vllm #776, FlagGems-sglang #84, flir #73
FlagDNN’s three commits restructure the implementation architecture for the Iluvatar and Hygon platforms respectively, and add a text summation operator (psum_text), continuing the repo’s practice of laying out implementations by platform. FlagGems-vllm’s three commits are all inference operators on domestic chips: optimizing grouped Top-K on Ascend using TLE’s DSA raw interface, adding a chunked gated Delta rule forward operator, and FP8 variable-length attention. FlagGems-sglang merged three PRs, importing a fused multimodal rotary position embedding operator and switching to referencing operators by package namespace, while also fixing the pre-commit checks on the batch-two branch. One flir commit fixes a memory effect marker, a continuation of the general intermediate representation trajectory. One FlagScale commit changes fork-repo PR dependency artifacts to be uploaded to avoid cache isolation — a CI availability fix.
II. News Coverage and Ecosystem
2.1 Fourteenth Consecutive Quiet Window for Component-Level Search: Zero Hits Within 24 Hours (09-17~09-18)
Date: 2026-09-17 ~ 2026-09-18 Source: Google News RSS (33 Chinese and English query terms, via proxy)
This issue ran 33 searches, split evenly between Chinese and English, using keywords including FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen, as well as “BAAI open source,” “BAAI chips,” “domestic compute open-source software stack,” and “BAAI open source AI system stack.” There were zero hits on component names within the 24-hour window, and the most recent report directly related to component names remains the 09-15 English article on the open-sourcing of Open3D-PIMC. This is now the fourteenth consecutive quiet window. Of the 15 hits, about 13 were stock market coverage and institutional rating pieces (MetaX share price swings, institutional selling, and Enflame IPO-related items), which were removed as a batch based on headline semantics and not counted in the main text.
2.2 Zidong Taichu ZDTaichu5.0-9B Open-Sourced, Relayed via BAAI Community: No Interface with the FlagOS Stack, Not Included in This Issue (09-16~09-17)
Date: 2026-09-16, 2026-09-17 Source: BAAI Community repost, QbitAI report
On 09-17, the BAAI Community reposted a report on the open-source general multimodal model ZDTaichu5.0-9B from Zidong Taichu; the model itself was released on 09-15, with weights and the data production plan opened simultaneously. This item matched the “BAAI large model open-source stack” query term in the search, but upon verification it was confirmed that the model was released by the Zidong Taichu team, with a language backbone from the Qwen 3.5 series, and that it has no interface with any layer of FlagOS—compiler, operator library, or plugins. It is third-party content on the BAAI Community platform rather than a FlagOS development. In accordance with the guideline to “exclude news that only mentions BAAI models in general terms,” it is not included; this verification is recorded here only to avoid repeated judgment later.
2.3 Capital-Side Developments for Member Organizations: MetaX Share Price Swings and Institutional Reduction, Enflame IPO Article (09-17)
Date: 2026-09-17 Source: Cailian Press, Jiwei
Capital-side news for member organizations was concentrated on 09-17: MetaX shares rose 14.44% that day, leading the compute chip sector, while multiple articles reported that five institutions collectively sold more than RMB 2.3 billion, and another report said its 10,000-card cluster had entered production. For Enflame, there was an industry article on the “market-oriented GP plus local state-owned enterprise LP” structure behind its IPO. These are capital market developments with no direct connection to the FlagOS codebase, and are recorded only as background on member organizations, not counted as technical items.
III. Member Unit Deep Dive
3.1 Hygon: Fused rrelu Backend Kernel Merged, Decoupled DCU Wheel Fixes Qwen-Image (09-17/09-18)
Date: 2026-09-18 Source: FlagGems #6379, Torch-FL #332, FlagDNN
Hygon had two substantive actions this period. On the operator side, FlagGems added a fused backend kernel for noisy random ReLU (#6379), going through the KernelGen compilation path to complete the backend implementation of this activation function. On the framework side, Torch-FL fixed the issue where torch.cuda was unavailable on the decoupled DCU wheel (#332), in the context of running the Qwen-Image image generation model — a decoupled wheel refers to a build that does not bundle the vendor’s full runtime, which previously would fail outright on models requiring the cuda namespace. FlagDNN also reorganized the Hygon platform implementation (see 1.12).
3.2 Moore Threads: MUSA Line FlagTree Jumps to 0.7.0rc2, vLLM 0.20.2 Image Gets Plugin for the First Time (09-18)
Date: 2026-09-18 Source: build-infra #926, build-infra #922
Moore Threads had the most complete set of changes among member units this period. On the compilation foundation, build-infra bumped the pinned FlagTree version for the MUSA backend from 0.6.1+mthreads3.6 to 0.7.0rc2+mthreads3.6 (#926), jumping directly to the 2.2 candidate line. On the inference plugin side, #922 records that two 0.20.2 application images were rebuilt on top of the plugin repository’s v0.2.2-rc2.post1 tag: the previous 2.1.2 tag images had no plugin installed at all, making this the first 0.20.2 image with a vLLM plugin on the Moore Threads release line, shipping with MUSA-related attention backend registries and device interface adaptations. On the FlagFFT side, MUSA 3D transpose platform validation records exist for the same period (1.9).
3.3 Kunlunxin: 0.6.1+xpu3.6 Pinning and tle.raw Fast Path Revert (09-17)
Date: 2026-09-17 Source: FlagGems #6393, FlagGems #6409, FlagGems #6328
Kunlunxin focused on fixes this period: first upgrading FlagTree in FlagGems dependencies to 0.6.1+xpu3.6 (#6393); then reverting the previous tle.raw scalar comparison fast path (#6409) — a deliberate negation, indicating that the fast path failed to deliver stable gains on Kunlunxin; the same batch also included fixes for masked fill, hyperbolic sine, activation, and index fill operators (#6328). On the base image side, build-infra fixed the Kunlunxin containerfile the day before the window, with no new actions this period.
3.4 MetaX: MACA Portable 3D Transpose Validation and Unified Dependency Pinning (09-17)
Date: 2026-09-17 Source: FlagFFT commit log, FlagFFT hardware profile branch log
The MetaX line this period centered on FlagFFT: after merging the hardware profile branch and the MACA feature branch, two validation records were left and the portable register-tiled 3D transpose was formally enabled — platform validation first, then enablement, consistent with the two-step approach throughout this codebase. To wrap up, the Triton JIT runtime dependency was pinned to a unified backend revision to prevent dependency drift between the MACA side and the mainline. Additionally, the MetaX-related plugin entries in the RC2 manifest (one each for the vLLM plugin 0.2 line and 0.3 line) have both been branched to rc2, part of the automation actions in 1.2.
3.5 Iluvatar CoreX: FlagDNN Platform Architecture Reorganization (09-17)
Date: 2026-09-17 Source: FlagDNN commit log
Iluvatar CoreX had one substantive action this period: the FlagDNN platform implementation architecture reorganization (see 1.12). Looking back at the previous period, this platform gained /dev/itrctl device interface support on the FlagGems side and had iluvatar3.6 baseline workflows on the FlagTree side; this period continues with structural cleanup at the operator library layer rather than adding new operators — consistent with the “general optimization techniques” wording among the four operator-layer goals in FEP-0099.
3.6 Ascend: CommonIR Proof-of-Concept Marked as Implemented, grouped_topk Tuned via tle.dsa.raw (09-17)
Date: 2026-09-17 Source: FEP-0095 Common IR, FlagGems-vllm #776, FlagGems #6354
Ascend is a line with clear progress on both the documentation and code sides this period. On the documentation side, FEP-0095 is one of the few among this batch of 20 documents marked as implemented, with its scope explicitly limited to a proof of concept for Ascend 910B/910C on the Triton 3.5 line, linked to FlagTree #974 (merged 08-27), with the path being TLE’s DSA interface lowered through a common intermediate representation to linalg and then into npubin. On the code side, FlagGems-vllm uses TLE’s DSA raw interface on Ascend to optimize grouped Top-K performance (#776); FlagGems added a Top-K operator (#6354). On the artifact side, the Ascend FlagTree wheel upload path was successfully exercised for the first time this period (1.4).
3.7 DAMO XuanTie: T-Head Enters FlagTree 2.2 Multi-Backend Planning, PPU Line CI Rolled Out (09-17 review)
Date: 2026-09-17 Source: FEP-0098 FlagTree Multi-Backend Architecture, vllm-plugin-FL #518, Torch-FL #337
A newly added document this period writes DAMO XuanTie’s T-Head backend into the FlagTree 2.2 multi-backend architecture plan (FEP-0098), alongside NVIDIA TileIR, Enflame, and Moore Threads, with a target Triton line of 3.6. On the engineering side, PPU line integration continues to advance: the inference plugin added PPU CI support the day before the window (#518), Torch-FL allocated a separate accelerator identifier for PPU in this period’s refactoring (#337), and explicitly probes for PPU in the operator test gate (#309). This line is characterized by infrastructure before operators: CI, identifiers, and gates are laid down first, with operators following.
3.8 Tsingmicro: Distributed Primitives Enabled on Backend, Converging with the 2.2 TLE Distributed Route (09-17)
Date: 2026-09-17 Source: FlagTree #1141, FEP-0096 TLE Megakernel and Distributed, flir commit log
A key commit from Tsingmicro this period enabled two TLE primitives — remote access and distributed barrier — on its backend (FlagTree #1141). Comparing this with FEP-0096 reveals the intent: that FEP plans to lower TLE’s distributed interface onto FlagCX in 2.2, achieving fusion of communication and computation within a single Triton kernel, and Tsingmicro is one of the earliest vendors to land this route on the backend side. On the upstream side, Tsingmicro’s patch synchronization to the common intermediate representation repository flir (LLVM22 support) has been ongoing for several weeks prior.
3.9 BAAI (Lead Party): FEP Governance Landed and RC2 Manifest Maintained (09-17)
Date: 2026-09-17 Source: FEP Description and Version Tracking, release directory
The lead party’s actions this period were entirely at the governance layer: a single merge of 20 FEP documents with synchronized version tracking updates (see “Today’s Highlights”); maintenance of the RC2 release manifest and git-level branch/tag automation (1.2); the release process documentation now describes the FEP lifecycle in two stages — “counts upon entry, graduation separate” — and stipulates that FEPs missing the freeze are deferred rather than discarded. At the organizational level, one formulation worth recording: edge and RISC-V have been listed as independent directions (the edge SIG is active, with RISC-V and OS packaging both listed as planned), indicating that the branching intent beyond 2.2 has already been put in writing in the governance documents.
IV. Summary and Trend Observations
In one sentence: FlagOS 2.2 has entered the final stretch of its testing period, where “only bug fixes are accepted” — the real protagonists this period are release artifacts and documentation: 20 FEPs complete the feature list, all 24 modules have been branched to rc2 and tagged, and the operator libraries and compilation foundation on the image line continue to move to newer candidate positions.
Three trackable threads:
First, release engineering has become the primary work surface. This period saw pushes across 28 repositories in the org with 147 commits, but categorized by content, the proportion leans distinctly toward packaging, artifacts, gates, and environment variables: FlagTree’s DEB and Nexus releases migrated to shared workflows, FlagCX replaced its standalone upload process, the Ascend wheel upload path was exercised, Torch-FL consolidated the environment variable surface into a single prefix scheme, and FlagQuantum raised its coverage floor to 75%. This aligns with the release schedule’s stipulation that only bug fixes are accepted from 09-01 to 09-24, and also means the remaining risk before the 09-28 GA falls mainly on “installs and runs correctly” rather than feature gaps.
Second, the frontier of multi-backend adaptation continues to expand, but forms are diverging. Within the same week, Moore Threads got the FlagTree 0.7.0rc2 jump and the first plugin-equipped 0.20.2 image, MetaX and Moore Threads each had 3D transpose validation on FlagFFT, Ascend exercised wheel upload for the first time, Tsingmicro enabled distributed primitives, and Kunlunxin proactively reverted a fast path that yielded no gains. There is both “moving up” (versions and artifacts) and “moving back” (negating ineffective optimizations) — the latter’s appearance is typically a signal that platform adaptation has entered a stable phase.
Third, the outline of the next version is already appearing in the documents. RISC-V experimental support, OS packaging and distribution integration, edge SDK, and the two working groups for embodied intelligence and scientific intelligence all now have positions in the FEP description’s “planned/incubating” table; FEP-0098 has also listed the new backends of T-Head, Enflame, and Moore Threads in the architecture plan. After the 09-28 GA, these directions will likely take over the attention that 2.2 relinquishes from the mainline.
To watch: (1) When build-infra’s version field moves from 2.2.0 to the tagging step, and whether the versions of the remaining backends in configs.yaml will see further bumps; (2) the ratio of FEP statuses transitioning from draft to implemented around the 2.2 GA, especially the KernelGen capability items that deliberately remain in draft; (3) whether previously quieter modules such as FlagAudio, FlagTensor, and FlagSparse have acceptance actions within the rc2 line.
Appendix: Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org (54 repos) | 28 repos had pushes within the window; 14 repos had rc2 branches created with no new commits on the default branch |
| GitHub commit search | 147 commits within the window matched (both pages retrieved in full), spanning 14 repos |
| git ls-remote (all refs) | Verified that the rc2 integration branches and v*-rc2.post1/post2/post3 tags for ten modules do exist |
| community FEP directory | 20 new documents added in this batch, numbered 0085 to 0100, covering seven SIGs |
| community 2.2 release manifest | RC2 manifest contains 24 module entries; item-by-item version comparison between RC1 and RC2 completed |
| community release schedule | Feature freeze 08-31, testing period 09-01 to 09-24, GA 09-28 |
| build-infra PR patches | Item-by-item verification of file lists and version field changes for #922 to #926 |
| Google News RSS (33 query sets) | Zero matches for component names in the 24-hour window; of 15 matches, about 13 were stock market articles and were excluded |
| HN Algolia (FlagOS / FlagGems) | No relevant entries within the window |
| BAAI Hub (hub.baai.ac.cn) | One repost about the Zidong Taichu model; verified to have no interface with the FlagOS stack, not included |
Complete Source List
- [1] community FEP directory — https://github.com/flagos-ai/community/tree/main/fep
- [2] FEP-0099 Operator Library for FlagOS 2.2 — https://github.com/flagos-ai/community/blob/main/fep/sig-operator/0099-operator-library-flagos-2.2.md
- [3] FEP-0100 KernelGen Capabilities for FlagOS 2.2 — https://github.com/flagos-ai/community/blob/main/fep/sig-kernelgen/0100-kernelgen-capability-flagos-2.2.md
- [4] FEP-0098 FlagTree Multi-Backend Architecture — https://github.com/flagos-ai/community/blob/main/fep/sig-compiler/0098-flagtree-multi-backend-architecture.md
- [5] FEP-0095 Common IR Introduced into FlagTree — https://github.com/flagos-ai/community/blob/main/fep/sig-compiler/0095-flagtree-common-ir.md
- [6] FEP-0096 TLE Megakernel and Distributed Primitives — https://github.com/flagos-ai/community/blob/main/fep/sig-compiler/0096-flagtree-tle-megakernel-and-distributed.md
- [7] FEP-0097 FlagTree Compiler Optimization — https://github.com/flagos-ai/community/blob/main/fep/sig-compiler/0097-flagtree-compiler-optimization.md
- [8] FEP-0092 FlagScale for FlagOS 2.2 — https://github.com/flagos-ai/community/blob/main/fep/sig-training/0092-flagscale-flagos-2.2-features.md
- [9] FEP-0085 FlagCX for FlagOS 2.2 — https://github.com/flagos-ai/community/blob/main/fep/sig-network/0085-flagcx-flagos-2.2-new-features.md
- [10] FEP-0081 FlagQuantum Development Direction — https://github.com/flagos-ai/community/blob/main/fep/sig-ai4s/0081-flagquantum.md
- [11] FEP Description and Version Tracking — https://github.com/flagos-ai/community/blob/main/fep/README_CN.md
- [12] FlagOS 2.2 Release Schedule (Chinese) — https://github.com/flagos-ai/community/blob/main/release/2.2/schedule_CN.md
- [13] FlagOS 2.2 RC2 Source Manifest — https://github.com/flagos-ai/community/blob/main/release/2.2/release-2.2-rc2.yaml
- [14] FlagOS 2.2 RC1 Source Manifest — https://github.com/flagos-ai/community/blob/main/release/2.2/release-2.2-rc1.yaml
- [15] Release Branch and Tag Workflow — https://github.com/flagos-ai/community/blob/main/.github/workflows/release-branch-tag.yml
- [16] build-infra #925 FlagGems Version Bump — https://github.com/flagos-ai/build-infra/pull/925
- [17] build-infra #926 Moore Threads FlagTree Upgrade — https://github.com/flagos-ai/build-infra/pull/926
- [18] build-infra #924 Ascend Upload Path Record — https://github.com/flagos-ai/build-infra/pull/924
- [19] build-infra #923 Ascend wheel Build Fix — https://github.com/flagos-ai/build-infra/pull/923
- [20] build-infra #922 Moore Threads Image Change Record — https://github.com/flagos-ai/build-infra/pull/922
- [21] build-infra configs.yaml — https://github.com/flagos-ai/build-infra/blob/main/configs.yaml
- [22] FlagGems #5846 Quantile Operators — https://github.com/flagos-ai/FlagGems/pull/5846
- [23] FlagGems #6312 AMD W7900D Weekly Test — https://github.com/flagos-ai/FlagGems/pull/6312
- [24] FlagGems #5731 SiliconFlow Zero-Copy Registration — https://github.com/flagos-ai/FlagGems/pull/5731
- [25] FlagGems #6379 Hygon Fused rrelu — https://github.com/flagos-ai/FlagGems/pull/6379
- [26] FlagGems #6393 Kunlunxin FlagTree Version Pinning — https://github.com/flagos-ai/FlagGems/pull/6393
- [27] FlagGems #6409 Kunlunxin Fast Path Fallback — https://github.com/flagos-ai/FlagGems/pull/6409
- [28] FlagGems #6328 Kunlunxin Operator Fix — https://github.com/flagos-ai/FlagGems/pull/6328
- [29] FlagGems #6354 Ascend Top-K Operator — https://github.com/flagos-ai/FlagGems/pull/6354
- [30] FlagGems #5972 PPU Quantized Matrix Multiplication — https://github.com/flagos-ai/FlagGems/pull/5972
- [31] FlagGems-vllm #776 Ascend Grouped Top-K Tuning — https://github.com/flagos-ai/FlagGems-vllm/pull/776
- [32] FlagGems-vllm #788 Ascend Chunked Gated Delta Rule — https://github.com/flagos-ai/FlagGems-vllm/pull/788
- [33] FlagGems-vllm #749 FP8 Variable-Length Attention — https://github.com/flagos-ai/FlagGems-vllm/pull/749
- [34] FlagGems-sglang #84 Fused Operator Import — https://github.com/flagos-ai/FlagGems-sglang/pull/84
- [35] FlagTree #1063 Nexus Release Workflow — https://github.com/flagos-ai/FlagTree/pull/1063
- [36] FlagTree #1066 DEB Matrix Expansion — https://github.com/flagos-ai/FlagTree/pull/1066
- [37] FlagTree #1082 Packaging Baseline and Version Alignment — https://github.com/flagos-ai/FlagTree/pull/1082
- [38] FlagTree #1170 TLE Synchronization Scope and Memory Ordering — https://github.com/flagos-ai/FlagTree/pull/1170
- [39] FlagTree #1141 Qingwei Backend Distributed Primitives — https://github.com/flagos-ai/FlagTree/pull/1141
- [40] Torch-FL #341 Environment Variable Surface Unification — https://github.com/flagos-ai/Torch-FL/pull/341
- [41] Torch-FL #337 SDK Discovery and PPU Identification — https://github.com/flagos-ai/Torch-FL/pull/337
- [42] Torch-FL #332 Decouple DCU wheel — https://github.com/flagos-ai/Torch-FL/pull/332
- [43] Torch-FL #320 Legacy FlagGems Switch Retirement — https://github.com/flagos-ai/Torch-FL/pull/320
- [44] FlagQuantum #64 Quantum Error Correction Record Layer — https://github.com/flagos-ai/FlagQuantum/pull/64
- [45] FlagQuantum #55 Coverage Floor 75% — https://github.com/flagos-ai/FlagQuantum/pull/55
- [46] FlagQuantum #52 Strict Type Checking — https://github.com/flagos-ai/FlagQuantum/pull/52
- [47] FlagCX #607 lane mask Expanded to 64-bit — https://github.com/flagos-ai/FlagCX/pull/607
- [48] FlagCX #561 Shared Nexus Upload Workflow — https://github.com/flagos-ai/FlagCX/pull/561
- [49] FlagCX #562 Fedora 43 RPM Matrix — https://github.com/flagos-ai/FlagCX/pull/562
- [50] FlagAudio #5 Spectrogram Operators — https://github.com/flagos-ai/FlagAudio/pull/5
- [51] FlagAudio #2 Packaging Finalization — https://github.com/flagos-ai/FlagAudio/pull/2
- [52] FlagFFT Commit History — https://github.com/flagos-ai/FlagFFT/commits/main
- [53] FlagDNN Commit History — https://github.com/flagos-ai/FlagDNN/commits/master
- [54] flir #73 Memory Effect Annotation Fix — https://github.com/flagos-ai/flir/pull/73
- [55] FlagScale #1298 Fork PR Dependency Artifacts — https://github.com/flagos-ai/FlagScale/pull/1298
- [56] vllm-plugin-FL #518 PPU CI Support — https://github.com/flagos-ai/vllm-plugin-FL/pull/518
- [57] BAAI Hub Zidong Taichu Model Repost (for verification, not included) — https://hub.baai.ac.cn/view/58035
- [58] QbitAI Zidong Taichu Model Report (for verification, not included) — https://www.qbitai.com/2026/09/490839.html
- [59] Cailian Press Computing Power Chip Sector Capital Flows — https://news.google.com/rss/articles/CBMiSEFVX3lxTE1nZUZBSTdjeElYcmlCLVk4YVYydVllaW9BWWt1TDhkUVBPaVVFTkFiVzhoSmtaOXpVS05wR3pWSmNMT2s3Rm0zNQ
- [60] Jiwei Enflame IPO Structure Report — https://news.google.com/rss/articles/CBMiQ0FVX3lxTFBrRmhHZ0M0cW1KQmdjUjEwWUJBNmN0ZHYyQ1BJRU0wVkYtTkZKZ2JUWkNiMlNNejd5WFg0eGxnWGN0dU0