Research window: 2026-11-06 06:00 ~ 2026-11-09 06:00 (Beijing time, 72 hours; continuing from the cutoff of the previous daily report, covering Saturday and Sunday).

Sources: GitHub (full audit of 62 repos under org: flagos-ai, 261 commits within the window, 253 merged PRs, 269 updated in-review PRs within the window, pushes to 21 repos, atomic audit of Releases and Tags across 23 repos, article-by-article verification of build-infra vendor documentation), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia, CSDN FlagOS column, BAAI Community (see the appendix source list for details).


Issue Index

  • Today’s Highlights: T-Head ZW810E PPU joins FlagOS with thead-ppu2.1.0, and 2.3.0 delivers the first runtime for this backend; FlagGems merges 123 PRs in a single day, including 87 Ascend KernelGen operators; TransformerEngine-FL moves to a pure-Python distribution-layer wheel, with a single build serving all non-NVIDIA vendors (11-07 / 11-08)
  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 Organizational cadence: 21 repos pushed, 261 commits, 253 merges, 269 under review (11-09)
    • 1.2 Platform side: T-Head ZW810E PPU onboarded, first runtime delivered in 2.3.0 (11-07)
    • 1.3 FlagGems: 123 merges in a single day, 87 Ascend operators landed at once (11-06)
    • 1.4 FlagTree: Hygon hcu3.6 released, deb and rpm now split into per-backend packages (11-06)
    • 1.5 FlagQuantum: Stim and PyMatching integrated into the error-correction workflow (11-07)
    • 1.6 FlagCX: Iluvatar CoreX unit-test pipeline, Ascend PCI topology gains DCMI fallback (11-07)
    • 1.7 Torch-FL: Cambricon torch.compile tiers and eight operators nativized (11-08)
    • 1.8 FlagPrism: Hygon HIP debugger and DTK profiler enter mainline (11-08)
    • 1.9 Distribution and infrastructure: TE-FL pure-Python wheel serves multiple vendors from one line, vLLM 0.32 adaptation wave (11-08)
    • 1.10 New repos and supporting work: FlagMRI-3D repo created, optest adds NV baseline and speedup provenance (11-06)
  • II. News Coverage and Ecosystem
    • 2.1 Channel check: zero hits at component level and in Chinese media; the latest public in-depth piece remains the 10-20 article in 2.2 (11-09)
    • 2.2 Competition channel: Track 1 SGLang operator submissions keep landing, 17 merged and 29 under review (11-06)
  • III. Member Units and Backend Vendors in Depth
    • 3.1 T-Head: SDK distributed only via PAI-DSW images, FlagTune cost model now open (11-06)
    • 3.2 Ascend: 87 operators landed in bulk, Ascend packaging and communication lines both filled in (11-06)
    • 3.3 Hygon: three HCU lines advanced the same day, twenty dedicated operators queued for review (11-08)
    • 3.4 Cambricon and Iluvatar: MLU operators nativized, CoreX CI coverage expanded (11-08)
    • 3.5 Moore Threads, MetaX, Kunlunxin, Tsingmicro: MUSA bumped to 0.32, MetaX 2.3.0 withdrawn and rebuilt (11-07)
    • 3.6 Moore Threads: Kimi K3.1 Pro Day-0 runs on the native MUSA stack (11-06)
  • IV. Summary and Trend Observations

I. Open-Source Project Progress (GitHub Activity)

1.1 Organizational Cadence: 21 Repos Pushed, 261 Commits, 253 Merges, 269 PRs Under Review (11-09)

Date: 2026-11-09 Sources: flagos-ai organization repositories · Commit search (within window) · Merged PR search (within window) · Open PR search (within window)

  • Totals: 21 of 62 repositories had pushes within the window; 261 new commits on main branches, 253 merged PRs, and 269 open PRs updated within the window.
  • Commit distribution: FlagGems 123 (47.1%), build-infra 34 (13.0%), FlagQuantum and FlagGems-sglang 16 each, FlagGems-vllm 14, Torch-FL and FlagTree 12 each; followed by OpenCourse 11, optest 6, FlagTrain 4, and FlagBLAS, FlagAttention, FlagCX 3 each.
  • Merge distribution: FlagGems 123, build-infra 35, FlagGems-sglang 17, FlagGems-vllm 14, FlagQuantum 13, Torch-FL 12, FlagTree 12 led the pack; the remaining 16 repos totaled 27.
  • Per-person distribution: LoserCheems 87 (all Ascend KernelGen operators), tengqm 27 (build-infra), FlagQuantum bot 13, lvyufeng 12 (Torch-FL), Yukun-Cui 12, c2flowDS 12, xuanzhengdu-eng 10.
  • Cadence: 11-06 (Friday) was the peak day with 224 commits; 11-07 and 11-08 were the weekend, totaling 37 commits across two days, with build-infra, Torch-FL, FlagPrism, FlagQuantum, and FlagCX still advancing.
  • Release surface: Across the Release and Tag atom sources of 23 repositories, only one addition occurred within the window: FlagTree’s 0.7.1a1+hcu3.6; the 2.3.0 record on the image side is covered separately in 1.2.
  • Versus previous period: Commits rose from 123 to 261, merges from 94 to 253, open PRs from 255 to 269, with pushed repos flat at 21. The previous period was a wrap-up day; this period is a batch-landing day.

1.2 Platform Surface: T-Head ZW810E PPU Onboarded, First 2.3.0 Runtime Delivered (11-07)

Date: 2026-11-07 Sources: build-infra #1675 (base image) · build-infra #1679 (runtime changelog) · build-infra #1676 (status matrix key rename) · build-infra #1682 (applicable platforms) · vendor docs thread

  • Onboarding: #1675 adds the backend key thead-ppu2.1.0, corresponding to T-Head ZW810E’s PPU 2.1.0 SDK line; the base layer directly references the vendor’s PAI-DSW training image (~25 GB, SDK at /usr/local/PPU_SDK), strips the platform entrypoint, points pip back to the stack mirror, and adds cmake and libpython3-dev.
  • First runtime: #1679 adds the runtime changelog for this backend under tag 2.3.0—the first runtime delivery for this backend line. Dependencies are torch 2.10.0+v0.1.0.ppu2.1.1 and numpy 1.26.4, default compiler FlagTree 0.7.0+ppu3.6, /opt/triton fallback version 3.5.0+v0.2.0.ppu2.1.1, operator library FlagGems 5.4.0, and no cpp wheel is produced.
  • Auditability: #1676 replaces the old key thead-ppu2.0.0 scattered across the vllm, sglang, and megatron status matrices and two hardcoded backend lists with thead-ppu2.1.0, and rewrites the sglang exclusion note—from “no base image” to “base exists, app layer not yet integrated.”
  • Applicable platforms: #1682 adds a generic platform field to configs.yaml, rendered as the “Applicable Platforms” line in image descriptions; the T-Head entry is filled with Alibaba Cloud PAI-DSW; #1681, #1684, and #1685 in the same batch record the base and runtime changelog dates for 2.3.0.
  • Status: The app layer is not yet integrated (deps_app and launch_docs both false); on the FlagGems side, #6497 was merged in the same window, enabling the FlagTune cost model for MM and MV.

1.3 FlagGems: 123 Merges in a Single Day, 87 Ascend Operators Landed at Once (11-06)

Date: 2026-11-06 Sources: FlagGems #8212 (special_erfc) · FlagGems #8206 (renorm) · FlagGems #8479 (log_softmax_backward_data) · FlagGems #6497 (FlagTune cost model) · FlagGems #6923 (enflame migration to flagcicd)

  • Scale: All 123 merges in the window landed on 11-06, densely between 10:34 and 11:24—within roughly one hour—all authored by LoserCheems, with no duplicate operator names.
  • Ascend dominates: 87 carry the [KernelGen][Ascend] prefix, covering six xlogy variants, three mish variants, two renorm variants, three deg2rad variants, three arctanh variants, plus amin, clamp_, clip_, permute_copy, narrow_copy, smm, euclidean_dist, cumulative_trapezoid, and others.
  • Other backends: Iluvatar 10 (three float_power variants on BI-V150, avg_pool1d, expand_copy, greater_, etc.), Moore Threads 9, NVIDIA 7; also MetaX 2 (addmm, baddbmm with router GEMM specialization, cross_attention) and Kunlunxin 3 (overrides for six operators and a foreach scalar kernel fix).
  • Supporting changes: #6497 enables the FlagTune cost model for T-Head PPU; #6923 migrates Enflame CI to flagcicd and bumps FlagTree to 0.7.0; #8509 updates CODEOWNERS.
  • Under review: FlagGems has another 88 PRs open, including 20 Hygon-specific operators (see 3.3).

1.4 FlagTree: Hygon hcu3.6 Released, deb and rpm Split Per Backend (11-06)

Date: 2026-11-06 Sources: FlagTree 0.7.1a1+hcu3.6 · FlagTree #1417 (one deb per backend) · FlagTree #1461 (rpm uses published wheel) · FlagTree #1583 (MThreads TLE sub-slicing) · build-infra #1660 (hcu into wheel release line)

  • Release: 0.7.1a1+hcu3.6 was published to PyPI on 11-06, the third pre-release of the Hygon backend on the 0.7.1 series (the first two being mthreads3.6 on 10-26 and tsingmicro3.6 on 10-29).
  • Packaging refactor: #1417 switches to producing one deb package per backend based on the backend table; #1461 builds the Ascend rpm directly from the published wheel instead of rebuilding in-repo; #1382 makes runtime depend on Python development headers; #1369 again strips debug sections from libtriton.so; #1368 drops the bundled CUPTI static library.
  • Compiler side: #1583 adds tensor sub-slicing capability to Moore Threads TLE broadcast buffers; #1531 preserves shared arena geometry in pointer and buffer views; #1401 lets mixed-precision SQMMA reuse the FP32 accumulator.
  • Release line: build-infra #1660 adds the hcu target to FlagTree’s wheel release line, following the same approach as the TE-FL release line in 1.9.

1.5 FlagQuantum: Stim and PyMatching Integrated into Error-Correction Workflow (11-07)

Date: 2026-11-07 Sources: FlagQuantum #1087 (Stim and PyMatching workflow) · FlagQuantum #1106 (dynamic record conversion) · FlagQuantum #1102 (Clifford sampler) · FlagQuantum #1104 (Qmod corpus compatibility) · FlagQuantum #1105 (hybrid adaptive state vector)

  • Error-correction chain: #1087 integrates the Stim and PyMatching experimental workflow into the QEC module; #1106 supports explicit dynamic record conversion with embedded Stim; #1102 lets the prepared Clifford sampler run directly on the simulation side.
  • Simulation and corpus: #1105 verifies output consistency of the hybrid adaptive state vector; #1104 hooks up a validated Qmod corpus compatibility layer.
  • Cadence: 13 merges and 16 commits spanning 11-06 and 11-07—the most-merged project in the window after FlagGems.
  • Positioning: FlagQuantum is FlagOS’s quantum-classical hybrid framework, organized under the in-formation wg-ai4s working group in community governance.

1.6 FlagCX: Iluvatar CoreX Unit-Test Pipeline, Ascend PCI Topology Gains DCMI Fallback (11-07)

Date: 2026-11-07 Sources: FlagCX #670 (Iluvatar CoreX unit tests) · FlagCX #692 (PCI topology and DCMI fallback) · FlagCX #696 (Barrier fence) · FlagCX #700 (Ascend IPC and GPU MR) · FlagCX #699 (MetaX AllToAll validation)

  • CI surface: #670 adds a unit-test pipeline and CI image for Iluvatar CoreX, part of the same batch of CoreX CI work as FlagScale, Megatron-LM-FL, and TransformerEngine-FL in 3.4.
  • Topology: #692 enables PCI topology at the PAL layer with AscendCL and DCMI as fallback paths; #696 completes UIL Barrier fence handling and resource setup.
  • Under review: #700 adds IPC and GPU MR support for Ascend and covers the P2P engine; #699 fixes data validation for MetaX AllToAll.
  • Image side: build-infra #1686 fixes the flagcx wheel build, requiring checkout by tag rather than resolved sha.

1.7 Torch-FL: Cambricon torch.compile Tier and Eight Operators Nativized (11-08)

Date: 2026-11-08 Sources: Torch-FL #772 (torch.compile tier) · Torch-FL #770 (eight operators nativized) · Torch-FL #771 (silu via CNNL activation kernel) · Torch-FL #769 (MLU-side strided gather) · Torch-FL #767 (CNNL code generator)

  • Path enablement: #772 adds a torch.compile tier for Cambricon, making this compilation path “at least exist” on that backend; #767 moves training-step operators into the CNNL code generator.
  • Performance: #770 natively serves eight more operators and switches matrix multiplication to batched mode; #771 routes silu directly through CNNL’s activation kernel; #769 and #768 keep strided gather, softmax, and negate on the MLU device instead of going through the host.
  • CI: #776, #765, #764, and #723 focus on cache and yq download robustness, with #723 making every CI lane run shared test cases.
  • Volume: 12 merges and 12 commits—one of the most active projects over the weekend.

1.8 FlagPrism: Hygon HIP Debugger and DTK Profiler Land on Main (11-08)

Date: 2026-11-08 Sources: FlagPrism #19 (HCU debugging and profiling) · FlagPrism #20 (vendor adaptation requirements doc) · FlagPrism repo · FEP-0170

  • Capabilities: #19 adds HIP debugger and DTK profiler support for the Hygon HCU backend, covering both compile-time and runtime observation surfaces.
  • Documentation: #20 specifies chip vendor adaptation requirements, turning “what can be integrated and what must be provided” into checkable items.
  • Positioning: FlagPrism is a multi-backend debugging and profiling tool for Triton programs, serving the FlagTree ecosystem; it corresponds to FEP-0170 submitted on 11-06, targeting five backends in v0.2.
  • Caveat: FEP-0170 remains a Provisional Draft, and the author states GPU and model acceptance testing has not yet been run; this entry therefore records only the merge fact, not capability completion.

1.9 Distribution and Infrastructure: TE-FL Pure-Python Wheel One-Line Multi-Vendor, vLLM 0.32 Adaptation Wave (11-08)

Date: 2026-11-08 Sources: build-infra #1687 (TE-FL distribution-layer release line) · build-infra #1690 (TE catalog entry) · build-infra #1683 (vLLM 0.32 minimum-version assessment) · vllm-plugin-FL #706 (0.32 on T-Head PPU) · sglang-plugin-FL #155 (Kunlunxin 0.5.18)

  • Distribution layer: #1687 adds a pure-Python wheel release line for TransformerEngine-FL, building a py3-none-any package with TE_FL_SKIP_CUDA=1 and pushing a single build to the PyPI and daily index of all non-NVIDIA vendors; versions look like 2.17.0+fl.0.3.0, with the upstream segment used for version gating and the FL version carried as a local label. NVIDIA is excluded because its native path is upstream TE.
  • Supporting: #1690 adds the transformer_engine catalog entry and description; #1689 records the te-fl-release wheel publication; #1688 hardens cloning against unstable GitHub HTTPS on runners.
  • Framework adaptation: A wave of adaptation for vLLM 0.32.0 and SGLang 0.5.18—T-Head PPU (#706), Iluvatar (#674), MUSA (#715), Kunlunxin P800 on 0.28 (#585), and Ascend with FlagOS-Compressor dynamic W8A8 quantization (#643) are all under review; the two SGLang 0.5.18 items for Kunlunxin and PPU are also under review.
  • Minimum versions: #1683 records the assessment conclusions on vLLM 0.32.0’s torch and Python minimums, paving the way for the subsequent matrix.

1.10 New Repos and Supporting Work: FlagMRI-3D Created, optest Adds NV Baseline and Speedup Source (11-06)

Date: 2026-11-06 Sources: FlagMRI-3D · optest #44 (speedup source and NV baseline) · optest #39 (NV baseline workflow) · FlagScale #1322 (Iluvatar training CI) · FlagScale-Agent #50 (Day 0 workflow)

  • New repo: FlagMRI-3D was created on 11-06, bringing the organization’s repository count from 61 to 62, with 1 commit that day; the README has not yet been uploaded and its positioning remains to be seen.
  • Test infrastructure: optest merged 6 PRs; #44 supports speedup-source and automatically updates NV baseline data, and #39 adds an NV baseline workflow—both aimed squarely at reproducibility of cross-vendor performance comparisons.
  • Training and agents: FlagScale #1322 integrates Iluvatar BI training through the shared CI framework; FlagScale-Agent #50 adds the Day 0 workflow and hardware acceptance criteria (under review).
  • Courses and docs: OpenCourse merged 4 PRs (including #11 updating hook documentation); on the docs side, the ModelScope model-list bot opened tickets for three consecutive days (#576, #577, #578 and #542, #545 are all under review).

II. News Coverage and Ecosystem

2.1 Channel Verification: Zero Hits at Component Level and in Chinese Media; the Only Public In-Depth Piece Remains the 2.2 Article from 10-20 (11-09)

Date: 2026-11-09 Sources: FlagOS 2.2 in-depth article (BAAI Hub, 10-20) · InfoQ report on the same topic · CSDN FlagOS column

  • Zero hits: 24 Chinese and English query terms covering FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen, and flagos-ai all returned nothing within the 72-hour window; the Chinese query “智源研究院 开源” returned 9 results, all about large models and industry commentary, unrelated to FlagOS.
  • In-depth article still in circulation: The 10-20 BAAI Hub and InfoQ articles on FlagOS 2.2 report nine operator libraries totaling 1,454 operators, an increase of 890 over the previous version, roughly 27% cross-chip performance improvement across over a thousand operators and ten chips, and GLM-5.3-Flash achieving 2.04x logic-service throughput and 19.4% to 27.6% lower cost on existing Ascend 910B hardware. These are existing 2.2 figures and do not belong to this window.
  • Pending verification: On 11-07, a media outlet reposted an Enflame article with the headline “listed on the STAR Market today, shares up over 200% at peak”; this was verified to be retrospective content from the 9-11 listing day, with a date and facts that do not match, so it was not included (the previous daily brief already verified the same headline).
  • Official site and community: The flagos.io homepage returns a compressed stream that cannot be read directly; the CSDN column in this window contains only an entry point for a replay of a sub-forum session on 11-06, with no new articles; the BAAI Hub search page is client-rendered, so this issue relies on direct reading of FlagOS-related pages.

2.2 Competition Channel: Track 1 SGLang Operator Submissions Keep Landing, 17 Merged and 29 Under Review (11-06)

Date: 2026-11-06 Sources: FlagGems-sglang repository · FlagGems-sglang #273 (grouped_gemm) · FlagGems-sglang #262 (gemma_rmsnorm) · Track 1 description (08-21)

  • Landing volume: In this window, FlagGems-sglang had 17 merges and 29 submissions under review, all titled with the [FlagOS Competition-Track1] prefix; contributors include external contestants such as apex-mochen and ggbondbest.
  • Task types: Merged submissions include grouped_gemm, four rmsnorm variants in the gemma series, kda_gate_chunk_cumsum, kda_prologue, chunk_kda_scaled_dot_kkt, swiglu_oai_split, normalize_and_patchify, causal_conv3d_cat_pad, fused_append_shared_experts, and others; on the review side there are also reduce_head_scores, awq_gemm, the dequantize_k_cache series, and act_and_mul.
  • Competition background: Track 1 focuses on multi-chip performance optimization of SGLang framework operators, organized by Zhongzhi FlagOS together with IEEE and co-organized by SGLang, with MetaX, the Supercomputing Internet, and Kunlunxin participating; development uses Triton and Triton-TLE uniformly; the schedule states that award acceptance and judging will be completed within one month starting 11-13 after the competition.
  • Scope: The repository only reflects code landing and does not mean performance benchmarks have been passed; this issue records “submissions landed” without drawing conclusions on results.

III. Deep Dive into Member Units and Backend Vendors

3.1 T-Head: SDK Distributed Only via PAI-DSW Images, FlagTune Cost Model Enabled (11-06)

Date: 2026-11-06 Sources: Vendor doc thead · FlagGems #6497 · FlagAttention #127

  • Naming clarification: T-Head (Alibaba’s chip company) makes PPU (AI accelerator cards) and Yitian (Arm CPUs); thead is T-Head’s brand name and upstream branch name, and FlagOS’s backend keys thead-ppu*, the vendor doc docs/vendors/thead.md, and PPU_SDK in PAI-DSW all refer to this company. XuanTie was spun off from T-Head in 2024 and moved under DAMO Academy; it only makes RISC-V processors (C950, C908, etc.) and is unrelated to the PPU backend in this case—do not conflate the two.
  • Distribution constraint: T-Head provides no PPU SDK installer and keeps no source in the filestore; the SDK exists only inside its PAI-DSW training image, so the base layer can only reference that image directly (~25 GB), and changing versions means changing a FROM line.
  • Build location: The self-hosted runner for this backend is registered at the org level—it is the Alibaba Cloud PAI-DSW node itself, the only machine that can pull the domestically reachable vendor image; the node has no GitHub egress, and images are built directly on that node. Persistent storage is CPFS 685 GB, plus a writable OSS mount.
  • Compiler side: #6497 enables the FlagTune cost model for MM and MV; FlagAttention #127’s PPU-ZW810E Wall Attention was merged in the previous window, and this period marks its first corresponding base and runtime in the 2.3.0 delivery.
  • Assessment: From backend key versioning on 11-05 to the image and runtime on 11-07, T-Head’s FlagOS integration completed the three steps of “key—image—delivery” in two days; the app layer (vLLM and SGLang) is still under review.

3.2 Ascend: 87 Operators Landed in Bulk, Ascend Packaging and Communication Both Filled In (11-06)

Date: 2026-11-06 Sources: FlagGems #8201 to #8479 (KernelGen Ascend batch) · FlagTree #1461 (rpm uses published wheel) · FlagCX #692 (DCMI fallback) · flir #86 (structured lowering fix)

  • Operators: 87 unique operators merged at once, all Triton implementations, concentrated in math, normalization, pooling, and copy categories—bulk output from KernelGen auto-generation; a single author completed them within one hour, showing the generate—test—merge pipeline can now run at batch scale.
  • Packaging: FlagTree #1461 makes the Ascend rpm build directly from published wheels, avoiding duplicate compilation in the compiler repo; the deb was meanwhile changed to one package per backend (#1417).
  • Communication: FlagCX #692 falls back to AscendCL and DCMI when PCI topology is unavailable; #700 separately covers Ascend IPC and P2P (under review).
  • Compiler defects: flir #86 fixes Ascend structured lowering, integer division/remainder, and dot transfer (under review).

3.3 Hygon: Three HCU Lines Advance on the Same Day, Twenty Dedicated Operators Queued for Review (11-08)

Date: 2026-11-08 Sources: FlagTree 0.7.1a1+hcu3.6 · FlagPrism #19 · FlagGems #8563 (Hygon dedicated operators) · FlagTree #1441 (HYGON TLE communication primitives)

  • Release: 0.7.1a1+hcu3.6 is the latest pre-release of the Hygon compiler backend in the 0.7.1 series, and it enters the build-infra wheel release line in sync (#1660).
  • Debug profiling: FlagPrism #19 hooks up the HIP debugger and DTK profiler; the vendor doc states that in code this backend is called DCU, the SDK is called DTK, and the compiler backend is called hcu, with target gfx936 and warp_size 64.
  • Operators: 20 Hygon dedicated operators under review (author bwbwzzz), including ldexp_, histc, igamma, lstm, convolution_overrideable, grid_sampler_3d, nuclear_norm, and special-series Chebyshev polynomials, none yet merged.
  • Communication: FlagTree #1441 adapts TLE communication primitives for HYGON (under review); TransformerEngine-FL #153, the Hygon BF16 expert GEMM merged in the previous window, has entered mainline.

3.4 Cambricon and Tianshu: MLU Operators Nativized, CoreX CI Surface Expands (11-08)

Date: 2026-11-08 Sources: Torch-FL #770 · FlagCX #670 · FlagScale #1322 · Megatron-LM-FL #213 · TransformerEngine-FL #151

  • Cambricon: Five Torch-FL merges form a complete chain—torch.compile tiers (#772), CNNL code generator (#767), eight operators nativized plus batched matmul (#770), silu routed to the activation kernel (#771), and strided gather kept on device (#769). The fused attention completion from the previous window is the step before this chain.
  • Tianshu: The CoreX CI surface appears simultaneously in four repos—the FlagCX unit-test pipeline (#670) is merged, while FlagScale’s BI training CI (#1322), Megatron-LM-FL’s CI and dev image (#213), and TransformerEngine-FL’s CI and image (#151) are all under review.
  • Operators: On the FlagGems side, 10 Tianshu KernelGen merges landed, concentrated in the three float_power variants for BI-V150 plus pooling and copy categories.
  • Assessment: Tianshu’s moves lean toward “building the pipelines,” Cambricon’s toward “moving operators off the device”; both are adding evidence to the 2.3 acceptance matrix.

3.5 Moore Threads, MetaX, Kunlunxin, Tsingmicro: MUSA Bumped to 0.32, MetaX 2.3.0 Withdrawn and Rebuilt (11-07)

Date: 2026-11-07 Sources: vllm-plugin-FL #715 (MUSA 0.32) · FlagGems #6807 (MetaX specialization) · build-infra #1672 (metax dependency removal) · build-infra #1674 (metax changelog restore) · sglang-plugin-FL #169 (Kunlunxin 0.5.18)

  • Moore Threads: FlagGems merged 9 KernelGen operators and 1 baddbmm backend optimization; vllm-plugin-FL #715 pushes MUSA adaptation to vLLM 0.32.0 (under review).
  • MetaX: FlagGems merged specializations for addmm, baddbmm, and router GEMM plus two cross_attention entries; on the build-infra side, metax 2.3.0 went through a withdraw-and-rebuild—first adding mcoplib and deep_gemm dependencies for the backend (#1670), then on the same day deciding to remove them (#1672), registering the withdraw-and-rebuild (#1673), and restoring the changelog (#1674).
  • Kunlunxin: Three FlagGems merges fix the foreach scalar kernel and override six operators; sglang-plugin-FL #155 and #169 bump the backend to SGLang 0.5.18, and vllm-plugin-FL #585 adds P800 support on vLLM 0.28, all under review.
  • Tsingmicro: No merges or submissions in the window; the most recent remains the 10-29 FlagTree 0.7.1a1+tsingmicro3.6 pre-release.

3.6 Moore Threads: Kimi K3.1 Pro Day-0 Runs on the Native MUSA Stack (11-06)

Date: 2026-11-06 Sources: Moore Threads announcement repost (Jiemian) · Technical details report

  • Fact: On 11-06 Moore Threads announced Day-0 adaptation of Kimi K3.1 Pro based on MTT S5000 accelerator cards and the MUSA software stack, covering model loading, service startup, and basic text and image inference paths.
  • Technical path: SGLang-MUSA carries the upstream implementation, with key compute paths migrated via the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler; adaptations target sparse attention index selection, KV cache layout, and the Prefill and Decode paths, while MXFP4 expert weights are handled with online conversion plus cross-layer reuse of an FP8 workspace.
  • Framing: Neither the announcement nor the reports mention FlagOS or FlagGems; this adaptation runs on the native MUSA stack and is an independent advance by a member unit, not counted as a FlagOS stack result; it is included here as an ecosystem reference.
  • Reference value: The same model has not yet shown a multi-chip Day-0 record on the FlagOS path; the next issue can watch whether corresponding operator coverage appears in FlagGems.

IV. Summary and Trend Observations

  • The platform surface has moved from “having keys” to “being deliverable”: T-Head’s ZW810E PPU aligned backend key, base image, runtime changelog, and status matrix in four places within two days, and 2.3.0 is the first runtime delivery for this backend. The device is obtainable only inside the Alibaba Cloud PAI-DSW image and the runner has no GitHub egress—these two constraints determine a delivery shape different from other backends, and the app layer remains empty in the near term.
  • Operator production has entered a batch rhythm: 123 FlagGems merges in a single day, of which 87 Ascend operators were landed by one person within an hour, with another 20 Hygon dedicated operators queued for review. KernelGen has turned “writing operators” into “reviewing generated results,” shifting the bottleneck from implementation to review and acceptance—which explains why the number of PRs under review (269) stays persistently above the number merged.
  • The distribution layer is converging: TransformerEngine-FL supplies all non-NVIDIA vendors with a single pure-Python wheel, while FlagTree’s deb moves to one package per backend and the rpm builds directly from published wheels; both moves point at the same goal—reducing per-chip duplicate builds and keeping differences in configuration.
  • Framework adaptation comes in waves: Adaptation to vLLM 0.32.0 and SGLang 0.5.18 clustered within the window, covering five vendors—T-Head, Tianshu, MUSA, Kunlunxin, and Ascend—but the vast majority remain under review. Whether they can merge before the 11-30 rc0 is the key variable for whether 2.3 can be delivered per the matrix.
  • 2.3 enters the closing countdown: The community 2.3 schedule sets the FEP merge deadline at 11-13, feature freeze and rc0 at 11-30, rc1 at 12-07, rc2 at 12-14, code lock at 12-25, and GA at 12-30. No new FEPs were submitted this period; the six from 11-06 (0164 to 0168, 0170) remain Provisional Draft and self-describe as not having run GPU acceptance, so closing pressure concentrates in the coming week.
  • Risks and watch points: T-Head’s app layer not yet connected, all 20 Hygon operators under review, three of Tianshu’s four CI lines under review, and competition submissions only ingested without results—all four fall under “action taken, acceptance pending”; next week the focus should be on merge rate rather than new volume.

Appendix: Source Verification Table and Full Source List

Source Verification Table

Source Verification Result
GitHub org flagos-ai (62 repos) 21 repos pushed within the window, 261 commits, 253 merged PRs, 269 open PR updates; 1 new repo (FlagMRI-3D)
GitHub Release and Tag atomic sources (23 repos) 1 addition: FlagTree 0.7.1a1+hcu3.6 (11-06); the 2.3.0 record on the mirror side is counted separately
Google News RSS (24 Chinese and English query terms) All component-level queries returned zero hits; the 9 “BAAI open source” results were unrelated to FlagOS and have been excluded
HN Algolia All hits were false matches on the word “flags”; no FlagOS-related discussion
CSDN FlagOS column Only one sub-forum replay entry on 11-06 within the window; no new articles
BAAI Community Homepage and FlagOS-related pages read directly; the 10-20 in-depth 2.2 article is pre-existing and was not used as news for this issue
Official site flagos.io Homepage returns a compressed stream that cannot be read directly; GitHub and community pages were used instead
Enflame IPO-related repost (11-07) Verified to be a retrospective piece on the 9-11 listing; date and facts do not match, so it was not included

Full Source List


Corrections in This Edition

The PPU backend covered in this issue (thead-ppu2.1.0, ZW810E, PAI-DSW images, PPU_SDK) belongs to T-Head, not XuanTie. Previous editions labeled this backend line as “DAMO XuanTie,” conflating T-Head with XuanTie (the RISC-V processor unit spun off from T-Head in 2024 and transferred to DAMO); thead is simply the open brand name T-Head has continued to use and does not refer to XuanTie. This edition has been corrected throughout to T-Head, with a naming clarification added.