FlagOS Daily Intelligence Report (2026-11-11)
Research window: 2026-11-10 06:00 ~ 2026-11-11 06:00 (Beijing time, 24 hours; continuing from the cutoff of the previous daily report).
Sources: GitHub (full audit of 63 repos under org: flagos-ai, 256 commits within the window, 245 merged PRs, 343 updated in-review PRs within the window, pushes to 21 repos), 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: 17 SGLang operators merged in one batch before the competition closes; Torch-FL sees 14 commits in a single day after renaming, and 17 operators in Cambricon’s Qwen3 training step switch to CNNL; FlagOS runtime 2.3.0’s fifth backend line is finalized as AMD ROCm 7.2.4 (11-10)
- I. Open Source Project Progress (GitHub Activity)
- 1.1 Organizational Cadence: 21 repos pushed, 256 commits, 245 merges, 343 under review (11-10)
- 1.2 Torch-FL: 14 commits in a single day after renaming, Cambricon Qwen3 training step lands (11-10)
- 1.3 FlagGems: 149 merges; of KernelGen’s 130, MetaX 89 and Hygon 20 (11-10)
- 1.4 FlagGems-sglang: 17 SGLang operators from the competition track merged in one batch (11-10)
- 1.5 FlagTree: NVIDIA, Iluvatar, and Moore Threads backend lines advance on the same day (11-10)
- 1.6 build-infra: Ascend CANN 9.2.1 daily build line enabled; runtime 2.3.0 recorded under the ROCm line (11-10)
- 1.7 FlagGems-vllm: T-Head PPU native vLLM wheels added to install options; Ascend gains sparse FlashAttention (11-10)
- 1.8 FlagQuantum: QEC record format and fermion workflow, 10 commits in a single day (11-10)
- 1.9 FlagTrain, FlagCV, optest: training and vision operator libraries close the evaluation loop (11-10)
- 1.10 FlagOS-Compressor, FlagSparse, FlagOmni: quantization and sparsity work wraps up (11-10)
- 1.11 docs: FlagPrism vendor adaptation spec rewritten as an interface contract (11-10)
- II. News Coverage and Ecosystem
- 2.1 Channel Verification: 15 hits across 24 query groups; component names still at zero hits (11-11)
- 2.2 Operator Talent Coverage: BAAI’s Liu Guang’s views included in draft; competition framing must be distinguished (11-10)
- 2.3 Shanghai Open Source System: ModelScope aggregates MetaX and Iluvatar CoreX compute (11-10)
- III. Member Units and Backend Vendors In Depth
- 3.1 Cambricon: 17 operators in Qwen3 training step switch to CNNL, dual-queue out-of-order issue fixed (11-10)
- 3.2 Kunlunxin: all four breakpoints closed on P800, FlagGems routing left empty pending survey (11-10)
- 3.3 MetaX: 89 KernelGen commits in a single day, foreach family rolled out across the board (11-10)
- 3.4 Hygon: 20 specialized operators merged, BW1100 added to CI test tags (11-10)
- 3.5 Iluvatar, Moore Threads, T-Head, Tsingmicro: incremental progress on backend versions and wheel installation (11-10)
- IV. Summary and Trend Observations
- Appendix: Source Verification Table and Complete Source List
I. Open-Source Project Progress (GitHub Activity)
1.1 Organization cadence: 21 repos pushed, 256 commits, 245 merges, 343 under review (11-10)
Date: 2026-11-10 Source: flagos-ai organization repositories · Commits within window · Merged PRs within window
- Totals: 21 of 63 repos had pushes within the window; 256 new commits on main branches, 245 merged PRs, and 343 open PRs updated within the window.
- Merge distribution: FlagGems 149 (58.2%), FlagGems-sglang 20, build-infra 17, Torch-FL 14, FlagQuantum and FlagGems-vllm 10 each, FlagTree 9, FlagTrain 6, FlagCV, FlagOmni, and OpenCourse 4 each, FlagOS-Compressor and FlagSparse 2 each, docs 2, and community, FlagCX, and optest 1 each.
- Versus previous period: Merges rose from 217 to 245, commits from 219 to 256, and pushed repos fell from 22 to 21; open PRs rose from 252 to 343, so the review queue backlog is heavier than last period.
- Version status: No new releases within the window; latest component versions remain FlagTree 0.7.0+triton3.6, FlagCX v0.14.0, FlagGems v5.4.0, FlagQuantum v0.3.1, and Torch-FL v2.10.0.
- Two repos pushed without main-branch commits: Megatron-LM-FL and flagtree-cpu had push timestamps within the window but no new commits on main — the former was PR and issue activity, the latter a push to a separate Triton 3.7 branch; neither counts toward this period’s commit total.
1.2 Torch-FL: 14 merges in a single day after rename; Cambricon Qwen3 training step lands (11-10)
Date: 2026-11-10 Source: Torch-FL repo · Torch-FL #809 (Cambricon Qwen3 training step) · Torch-FL #789 (Kunlunxin FlagGems routing)
- Positioning: Torch-FL is the repo formerly named PyTorch-Plugin-FL; the README now reads torch-fl, exposing a single flagos device that routes each operator among four implementation classes — reusable local kernels, portable compiler kernels, vendor-native implementations, and explicit CPU fallback. The same batch of docs commits also renamed the related documentation to torch-fl.
- Single-day scale: 14 merges, making it the fourth-largest repo by commits this period and the fastest-growing major repo, with another 102 open issues.
- Cambricon: #809 supplies the 17 Cambricon operators needed by the Qwen3 training step directly from the code generator locally, including the in-place arithmetic family, three fused CNNL entry points, four view operators, and zeros, sqrt, and div, while also fixing a lifetime defect in fill_.
- Build defect found along the way: torch_bindings had never been compiled with USE_CAMBRICON, so the binding that tells Triton which cnrt queue to launch into did not exist at all; Triton was effectively running on an empty queue with no ordering guarantee relative to CNNL kernels.
- Performance status: The training step takes 1.86× the time of torch_mlu and has not yet caught up; the PR itself states it only addresses operator coverage and queue ordering, not performance parity.
- Multi-backend in the same batch: #801 adds an experimental TileOPs vendor route, #802 adds argmin and max.dim/min.dim on MPS, #797 moves Tianshu backend type conversion back to the device side, and #790 relaxes the Tianshu fused-attention switch to the model’s own call pattern.
- Under review: A PR routing 316 operators measured on P800 wholesale to FlagGems is still under review — the next major move on the Kunlunxin side.
1.3 FlagGems: 149 merges; of 130 KernelGen commits, 89 MetaX and 20 Hygon (11-10)
Date: 2026-11-10 Source: FlagGems repo · FlagGems #8723 (Qingwei backend upgraded to FlagTree 0.7.0) · FlagGems #8643 (Kunlunxin var and vander)
- Scale: 149 commits and 149 merges, all in this repo, accounting for 60% of organization-wide merges; 130 carry the KernelGen label, split by backend into MetaX 89, Hygon 20, NVIDIA 9, Moore Threads 9, and Kunlunxin 3.
- MetaX foreach family: The 89 are the full foreach-prefixed operator set, covering tensor, scalar, list, and in-place variants of the add, mul, and div families, plus maximum, minimum, log, log2, log10, log1p, abs, acos, neg, addcdiv, and more — a whole-family completion rather than scattered patches.
- Hygon: The 20 dedicated operators are mainly special functions, pooling, deconvolution, embedding_bag backward, hash_tensor, and fused matrix multiply, plus quality-check items adding layer_norm and native_layer_norm.
- Correctness fixes: cdist_backward now dispatches gradient formulas by p-norm; Kunlunxin gains a batch_norm backward override and fixes to the var signature and vander override; the benchmark JSON record for W8A8 INT8 matrix multiply is restored.
- Engineering hygiene: A pre-commit hook for full-width punctuation was added, along with a fix for an unused-variable warning in Moore Threads linalg_tensorinv; trailing commas and import order in the metax and mthreads operator files were also cleaned up.
1.4 FlagGems-sglang: 17 competition-track SGLang operators merged at once (11-10)
Date: 2026-11-10 Source: FlagGems-sglang repo · FlagGems-sglang #315 (mxfp8 block-scaled matrix multiply) · FlagGems-sglang #336 (SGLang baseline workload)
- Scale: 20 commits, 19 merges, of which 17 carry the FlagOS Competition-Track1 label in their titles — the last organized delivery batch of the competition development phase.
- Operator content: Concentrated on low-precision quantization and matrix multiply — fp8 and mxfp8 block-scaled matrix multiply, quantization at tensor, per-token, and per-token-group granularity, fp8 indexer quantization, mxfp8 quantized write-back for KV cache, AWQ dequantization, fused routing of silu and mul, and QK normalization for zimage.
- Evaluation side: The same batch merges the native SGLang baseline workload for optest reuse, giving competition operators a comparable reference score.
- Trajectory: Last period saw 72 competition operator definitions land in one go; this period 17 implementations were merged, one day after the definitions. The competition development phase runs from August 17 to November 13, leaving only enough time for wrap-up and re-testing.
1.5 FlagTree: NVIDIA, Tianshu, and Moore Threads backend lines advance on the same day (11-10)
Date: 2026-11-10 Source: FlagTree repo · FlagTree #1532 (TLE local pointer proof) · FlagTree #1592 (Moore Threads PH1 and SQMMA)
- Scale: 9 merges, with each of the three backend lines taking a share and no single point of concentration.
- NVIDIA: #1532 preserves TLE’s local pointer proof through read/write lowering — a correctness-oriented compiler fix.
- Tianshu: #1409 adds set_layout support for its backend, and #1618 upgrades LLVM to 0.7.0 to fix the attention path; the two land as a pair.
- Moore Threads: #1592 fixes PH1 warp specialization and SQMMA synchronization issues — a fix at the level of hardware synchronization semantics.
- General pipeline: #1517 adds four optional switches to the compilation pipeline — prefetch promotion, wgmma carry, index prefetch, and load merging — turning performance tuning into configuration.
- CI side: Adds an hcu3.6 test label for Hygon BW1100, plus two commits routing scheduled-job and manual-trigger failure notifications to the correct owners.
1.6 build-infra: Ascend CANN 9.2.1 daily build line enabled; runtime 2.3.0 recorded on ROCm line (11-10)
Date: 2026-11-10 Source: build-infra repo · build-infra #1721 (Ascend 9.2.1 daily line) · build-infra #1716 (2.3.0 recorded on ROCm 7.2.4)
- Scale: 17 merges, third-highest commit count this period, all focused on the three tasks of building, packaging, and release registration.
- New Ascend line: #1721 opens the CANN 9.2.1 daily build line, with the same batch adding a cp312 builder, target options, wheel version-tag validation switched to interpreter-derived, and libzstd preinstalled for the 9.2.1 LLVM tree.
- Release defects: #1724 fixes the Ascend builder’s ref parameter being read as a pip version number; #1720 makes official releases produce clean Ascend wheel version numbers; #1722 and #1723 register the wheel and add a 9.2.1 column to the builder docs, respectively.
- runtime 2.3.0: #1716 records the 2.3.0 change date on the AMD ROCm 7.2.4 line — the fifth backend line to land in 2.3.0 within the past week, after Ascend CANN 9.2.1, NVIDIA CUDA 13.3, MetaX, and T-Head PPU 2.1.0.
- Assessment: The 2.3.0 changelog being settled backend by backend, with images built per vendor line, indicates the release is advancing as a packaging matrix rather than a one-shot freeze.
1.7 FlagGems-vllm: T-Head PPU native vLLM wheel added to install options; Ascend gains sparse FlashAttention (11-10)
Date: 2026-11-10 Source: FlagGems-vllm repo · FlagGems-vllm #1148 (T-Head PPU wheel) · FlagGems-vllm #967 (Ascend sparse attention)
- Scale: 10 merges, concentrated on three lines: compiler pipeline, vendor portability, and test coverage.
- T-Head: #1148 adds a thead install option that directly installs T-Head’s vendor vLLM wheel; #874 opens FP8 fused-operator test coverage for T-Head. PPU here refers to T-Head’s AI accelerator card line, unrelated to DAMO XuanTie’s RISC-V processors.
- Ascend: #967 adds a sparse FlashAttention operator; #849 rewrites the indexer K-quantization cache gather operator in a form portable to both Ascend and MetaX.
- Pipeline: #1063 selects FlagTree prefetch depth and load-merge strategy by shape; #993 aliases the quantized self-attention preparation gate output to the packed QKV buffer.
- Testing: #1154 moves reference implementations out of production code; #1135 adds a check that blocks operators exported but not registered in the operator manifest.
1.8 FlagQuantum: QEC record formats and fermionic workflows, 10 merges in a day (11-10)
Date: 2026-11-10 Source: FlagQuantum repo · FlagQuantum #1120 (measurements to detector records) · FlagQuantum #1116 (six native record formats)
- Scale: 10 merges, 13 open issues, latest version still v0.3.1 from October 16.
- Quantum error correction: #1120 converts measurement files into detector records and logical records; #1116 supports streaming detector samples out in six native file formats; #1117 proposes a design for a bounded repetition-code adapter.
- Fermionic: #1121 adds a controlled low-rank evolution workflow, #1119 adds Pauli stabilizer contraction, and #1118 adds a native radix-2 Fourier circuit.
- Dependency removal: The chemistry side now reads MolecularData HDF5 files directly instead of depending on OpenFermion, forming two paths alongside its in-review optional OpenFermion conversion interface.
1.9 FlagTrain, FlagCV, optest: training and vision operator libraries close the evaluation loop (11-10)
Date: 2026-11-10 Source: FlagTrain repo · FlagCV repo · optest repo
- FlagTrain: 6 merges. Adds a Triton kernel for DeepSpeed FP8 matrix multiply, a fused dual-activation operator for NVIDIA, and per-case placeholder candidate validation; also fixes an error in the mixed-precision matrix multiply benchmark and removes a redundant backward implementation of quick_gelu.
- FlagCV: 4 merges. Evaluation candidates are now dispatched per operator, and selected benchmark cases must hit a placeholder candidate, preventing empty runs from being counted as passes.
- optest: 1 merge adding FlagTrain support, bringing the training operator library into the unified evaluation infrastructure.
- Status: All three repos are evaluation and domain libraries created in October; this period’s work connects the new libraries into the existing evaluation loop, and none has an external release yet.
1.10 FlagOS-Compressor, FlagSparse, FlagOmni: quantization and sparsity wrap-up (11-10)
Date: 2026-11-10 Source: FlagOS-Compressor repo · FlagSparse repo · FlagOmni repo
- FlagOS-Compressor: 2 merges. Fixes calibration input retention for the MiMo model and preserves native FP32 tensors; unifies GPTQ inverse-Hessian decomposition across backends, where NPU and GPU previously used two different decompositions.
- FlagSparse: 2 merges, including a PR from the external AlphaSparse project and a sparse optimization targeting Moore Threads MUSA.
- FlagOmni: 4 merges, all on documentation and structure — the README gains design goals, directory structure, and development guidance, and the package layout is regrouped by framework. Still no operator implementations on the second day since the repo was created.
1.11 docs: FlagPrism vendor adaptation spec rewritten as an interface contract (11-10)
Date: 2026-11-10 Source: docs #580 (FlagPrism adaptation spec) · FlagPrism repo
- Change size: About 390 lines added and 30 removed in each of Chinese and English, rewriting what was a one-page adaptation requirement into a full specification.
- Structure: Organized into four sections — goals, Profiler adaptation interface, Debugger integration, and zero impact when disabled — and clarifies three information classes: basic instrumentation information, numerical summaries, and chip performance information, of which only the third depends on the vendor Profiler.
- Interface-ization: Specifies 6 pure virtual interfaces for vendor adapters (backend name, device type, supported hardware metrics, collection plan generation, etc.) and fixes the runtime collection flow at six steps: generate plan, configure collector, start/stop operators, flush, stop, and import.
- Significance: Onboarding a new chip to FlagPrism shifts from code changes to implementing a set of interfaces, consistent with the backend adaptation approach of other FlagOS components.
II. News Coverage and Ecosystem
2.1 Channel Verification: 15 Hits Across 24 Queries, Component Names Still at Zero (11-11)
Date: 2026-11-11 Source: Google News RSS (Chinese) · Google News RSS (English) · HN Algolia
- Component names: All seven groups of Chinese and English queries — FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen — returned zero hits within the 14-day window, marking the thirteenth consecutive quiet window.
- Hit composition: 24 queries yielded 15 hits in total, all from Chinese-language terms, with zero hits on the English side. Enflame returned 8 hits (IPO retrospectives, share price and shareholder count pieces), BAAI chip 3 hits, BAAI open source 1 hit, MetaX open source 1 hit, domestic compute 1 hit, and Tsingmicro 1 hit.
- Exclusions: All 8 Enflame hits were capital-market pieces; the single Tsingmicro hit was about Chipmore Semiconductor’s Hong Kong listing and unrelated; the Heima Morning Report and ETF pieces were general industry content. All were excluded.
- Two usable items: A report on operator talent and a piece on Shanghai’s open source system, covered in the next two sections.
- HN: Queries are still split into the two terms flag and os, returning book reviews and unrelated discussions, with no FlagOS-related content.
2.2 Operator Talent Report: BAAI’s Liu Guang Quoted; Competition Scope Must Be Distinguished (11-10)
Date: 2026-11-10 Source: Sohu (From Competing on Compute to Deepening Operator Work)
- Content: An industry commentary centered on the operator talent gap, citing Liu Guang, head of the Systems Research Group at BAAI, from his remarks at the 2026 BAAI Conference — that operator development is the hardest and most foundational work in AI infrastructure, and that talent scarcity has reached a critical point. The same piece quotes iFlytek’s Liu Cong on the need to tackle underlying bottlenecks one by one in domestic-compute training.
- Data points: The competition cited is the “China Telecom Xingchen Cup · University AI Operator Development Challenge,” co-hosted by China Telecom’s AI company and Huawei, with 177 universities, 539 teams registered, and 28 advancing to the finals; the final problem was a sparse FlashAttention backward operator.
- Scope must be distinguished: This competition is not the same as FlagOS’s SGLang operator optimization challenge, and the report does not mention FlagOS components at all. This issue cites only its industry judgment on operator talent scarcity and does not conflate the competition results with the FlagOS contest.
- Relevance: This corroborates the 17 competition operators merged on the GitHub side in this issue — the bottleneck in operator supply has shifted from “writing” to “review and acceptance,” while the talent gap is a longer-term layer.
2.3 Shanghai Open Source System: Model Ark Aggregates MetaX and Iluvatar CoreX Compute (11-10)
Date: 2026-11-10 Source: Sina Shanghai · PEdaily
- Policy side: In December 2025, Shanghai issued an implementation plan to strengthen its open source ecosystem, targeting 1 to 2 internationally influential open source communities by 2027, 100 open source commercialization companies, more than 200 high-quality open source projects incubated, and over 3 million open source developers gathered; basic development tools receive up to RMB 20 million in subsidies, and high-quality open source projects up to RMB 5 million in awards.
- Platform side: On October 30, the Open Source China headquarters settled in Zhangjiang, Pudong, and the Model Ark base opened the same day. Its compute side already supports domestic heterogeneous compute including Ascend, MetaX, and Iluvatar CoreX, while its model side aggregates invocation interfaces for hundreds of mainstream open source models.
- Relationship to FlagOS: MetaX and Iluvatar CoreX are both FlagOS member organizations. The substance of this item is that member organizations’ chips are aggregated as heterogeneous compute by a third-party platform — an indirect ecosystem-level positive, not an action by FlagOS itself.
- Competition side: The same piece mentions that the Shanghai Open Source Software Application Innovation Competition received 766 registered projects and 269 projects responding to the challenge, unrelated to FlagOS.
III. Deep Dive into Member Units and Backend Vendors
3.1 Cambricon: 17 Operators in the Qwen3 Training Step Switched to CNNL, Dual-Queue Out-of-Order Issue Fixed (11-10)
Date: 2026-11-10 Source: Torch-FL #809
- Coverage: 17 operators now supplied directly from the code generator to Cambricon native implementations; the training step no longer routes these operators through FlagGems.
- Categories: In-place arithmetic family, three fused CNNL entry points (add-divide, add-multiply, linear interpolation), four view operator types (slice, alias, transpose, unsafe view), plus zeros, sqrt, and div.
- Defects: Fixed a lifecycle issue in fill_; more critically, torch_bindings was not compiled with USE_CAMBRICON, so Triton launched on an empty cnrt queue with no ordering constraint against CNNL kernels — this class of silent defect throws no error, only intermittent wrong numbers.
- Performance: This training step still takes 1.86x as long as torch_mlu; the PR explicitly states it only addresses coverage and correctness, not performance parity.
- Effort level: A single change covering 17 operators plus a build-system defect — the heaviest member-unit item this period.
3.2 Kunlunxin: All Four Blockers Closed on P800, FlagGems Routing Left Empty Pending Profiling (11-10)
Date: 2026-11-10 Source: Torch-FL #789 · FlagGems #8643 · FlagGems #8641
- Problem: The Kunlunxin wheel was in fact already compiled into FlagGems, but the backend config routed zero operators to it, so every FlagGems-covered operator fell back to CPU on Kunlunxin.
- Four blockers: The config generator could not route any operator to Kunlunxin (the escape hatch documented was a silent no-op), no registration file existed for the FlagGems slot, the runtime glue needed to run FlagGems under CPU-only torch was missing, and the native-vs-FlagGems priority config did not name the platform.
- Result: All four blockers closed and end-to-end validation passed on P800; however, the allowlist remains empty and the config file is byte-for-byte identical to before the change — per the author, the first route must be backed by P800 profiling data.
- Next: A PR routing all 316 P800-measured operators to FlagGems is under review — the next step for this line.
- Operator side: FlagGems also added a batch_norm backward override for Kunlunxin the same day, and fixed the var signature and vander override.
3.3 MetaX: 89 KernelGen Commits in a Single Day, foreach Family Rolled Out Wholesale (11-10)
Date: 2026-11-10 Source: FlagGems (MetaX KernelGen batch) · FlagGems-vllm #849
- Scale: 89 KernelGen-specific operators merged, 68% of all org-wide KernelGen merges — the largest backend queue this period.
- Family breakdown: Concentrated in foreach-prefixed operators — add, multiply, divide, min/max, log, abs, acos, neg, add-divide, and others, with tensor, scalar, list, and in-place variants nearly all aligned one-to-one: a complete family fill-in.
- Assessment: This pattern indicates MetaX has finished mapping the interface semantics of the foreach family and handed it en masse to the auto-generation pipeline, with humans only reviewing — a different stage of work from last period’s 63 Enflame commits “grinding out defects operator by operator.”
- Portability: The indexer K-quantized cache collection operator now uses a single portable implementation shared by Ascend and MetaX, reducing the fork where each wrote its own.
3.4 Hygon: 20 Specialized Operators Merged, BW1100 Added to CI Test Labels (11-10)
Date: 2026-11-10 Source: FlagGems (Hygon KernelGen batch) · FlagTree #1597 (hcu3.6-bw1100 test label) · FlagTree (FP64 matmul-add support)
- Operators: 20 specialized operators merged, covering 3D and 2D adaptive pooling backward, special Chebyshev polynomial, special i1e, index reduction, Lanczos upsampling backward, divide and dequantize, soft-threshold backward, remainder, in-place multiply, overwritable convolution, depthwise separable convolution, convolution mode, embedding_bag dense backward, masked scaling, MPS-specific attention math variants, convolution double backward, tensor hash, and matmul with bias activation.
- Compiler: FlagTree added Hygon FP64 matmul-add support in the same batch, indicating its FP64 path no longer relies solely on emulation.
- CI: Added Qwen test labels for the hcu3.6 + BW1100 combination — the new-generation chip has entered routine regression.
- Numerics library side: FlagGems quality checks added layer_norm and native layer_norm — correctness backstops for a foundational normalization operator.
3.5 Iluvatar, Moore Threads, T-Head, Tsingmicro: Small Steps on Backend Versions and Wheel Installation (11-10)
Date: 2026-11-10 Source: FlagTree #1409 · FlagGems-vllm #1148 · FlagGems #8723
- Iluvatar: FlagTree added set_layout support and bumped LLVM to 0.7.0 to fix attention; Torch-FL moved type conversion back to the device side and relaxed the enablement conditions for fused attention — both reduce host-side round trips.
- Moore Threads: FlagTree fixed PH1 warp specialization and SQMMA synchronization; on the FlagGems side, 9 KernelGen merges, including a specialized implementation of linalg_householder_product.
- T-Head: FlagGems-vllm added a thead install entry that directly installs the PPU vendor build of the vLLM wheel, and enabled FP8 fused-operator test coverage. This is T-Head’s PPU accelerator card line, not the XuanTie (DAMO) RISC-V processor.
- Tsingmicro: FlagGems bumped the tsm260610 backend to the FlagTree 0.7.0 + tsingmicro3.6 corresponding version — a pinning that follows the compiler mainline.
IV. Summary and Trend Observations
- Competition sprint pushes the operator library up an order of magnitude: 17 SGLang competition operators merged at once, plus MetaX’s 89-commit full foreach family fill-in, brought single-day merges to a recent high of 245; but the review queue also rose to 343, making review bandwidth the main bottleneck.
- Torch-FL, post-rename, has become the second front: 14 merges in one day, with the Cambricon and Kunlunxin lines advancing simultaneously — and both are filling in the “routing” layer rather than operator implementations themselves, indicating FlagOS’s competitive dimension is shifting from operator count to framework-integration completeness.
- A compiler and runtime version matrix is taking shape: Over the past week, runtime 2.3.0 has successively locked in five backend lines — Ascend CANN 9.2.1, NVIDIA CUDA 13.3, MetaX, T-Head PPU 2.1.0, and AMD ROCm 7.2.4 — with Ascend 9.2.1 additionally opening a daily-build line; the release cadence is now split by vendor.
- Adaptation work is becoming contractualized: FlagPrism codifies vendor adaptation into a 6-interface spec, FlagTree turns performance tuning into four compiler switches, and FlagGems-vllm blocks unregistered exported operators at CI — all three moves turn verbal agreements into verifiable contracts.
- Domain libraries are laying skeletons: FlagOmni is still documentation and package structure on its second day since repo creation; FlagCV and FlagTrain have first wired the evaluation loop into optest, with no public release yet; the value of these libraries can only be assessed once the first batch of operators lands.
- To watch: Whether the PR routing 316 P800-measured operators merges and opens Kunlunxin’s FlagGems path; whether results are announced after the competition ends on November 13; whether 2.3.0 ships as a unified release once the remaining backend lines are filled in.
Appendix: Source Verification Table and Full Source List
Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org flagos-ai (63 repos) | 21 repos pushed within the window, 256 commits, 245 merged PRs, 343 updates under review; repo count unchanged |
| GitHub Release and Tag | No new Releases within the window; latest are FlagTree 0.7.0+triton3.6, FlagCX v0.14.0, FlagGems v5.4.0, FlagQuantum v0.3.1, Torch-FL v2.10.0 |
| Google News RSS (24 query term sets) | Seven component-name sets returned zero hits within 14 days; 15 hits on the Chinese side, of which 2 were included after excluding stock-market and general-industry pieces |
| HN Algolia | Query was split into flag and os; hits were unrelated discussions such as book reviews, no FlagOS content |
| CSDN FlagOS column | No new articles within the window |
| BAAI community / flagos.io | Homepage returns a shell page; body requires front-end rendering and cannot be read directly; repo and page direct reads used instead |
| Competition side | Track 1 development phase runs August 17 to November 13; 17 competition operators merged this period; judging begins after the end |
| Member/partner official sites | No full sweep this period; member activity based mainly on GitHub commits |
Full Source List
- [1] FlagOS organization repository overview — https://github.com/orgs/flagos-ai/repositories
- [2] Commit search within the window — https://github.com/search?q=org%3Aflagos-ai+committer-date%3A%3E%3D2026-11-09T22%3A00%3A00Z&type=commits
- [3] Merged PR search within the window — https://github.com/search?q=org%3Aflagos-ai+is%3Apr+is%3Amerged+merged%3A%3E%3D2026-11-09T22%3A00%3A00Z&type=pullrequests
- [4] Open PR search within the window — https://github.com/search?q=org%3Aflagos-ai+is%3Apr+is%3Aopen+updated%3A%3E%3D2026-11-09T22%3A00%3A00Z&type=pullrequests
- [5] Torch-FL repository — https://github.com/flagos-ai/Torch-FL
- [6] Torch-FL #809 (Cambricon Qwen3 training step) — https://github.com/flagos-ai/Torch-FL/pull/809
- [7] Torch-FL #789 (Kunlunxin FlagGems routing) — https://github.com/flagos-ai/Torch-FL/pull/789
- [8] Torch-FL #801 (TileOPs routing experiment) — https://github.com/flagos-ai/Torch-FL/pull/801
- [9] Torch-FL #802 (MPS reduction operators) — https://github.com/flagos-ai/Torch-FL/pull/802
- [10] Torch-FL #797 (Iluvatar device-side type conversion) — https://github.com/flagos-ai/Torch-FL/pull/797
- [11] Torch-FL #790 (Iluvatar fused attention switch) — https://github.com/flagos-ai/Torch-FL/pull/790
- [12] FlagGems repository — https://github.com/flagos-ai/FlagGems
- [13] FlagGems #8723 (Tsingmicro backend upgraded to FlagTree 0.7.0) — https://github.com/flagos-ai/FlagGems/pull/8723
- [14] FlagGems #8643 (Kunlunxin var and vander) — https://github.com/flagos-ai/FlagGems/pull/8643
- [15] FlagGems #8641 (Kunlunxin batch_norm backward) — https://github.com/flagos-ai/FlagGems/pull/8641
- [16] FlagGems #6032 (cdist_backward dispatch by p-norm) — https://github.com/flagos-ai/FlagGems/pull/6032
- [17] FlagGems #8624 (embedding_spmdm_grad) — https://github.com/flagos-ai/FlagGems/pull/8624
- [18] FlagGems-sglang repository — https://github.com/flagos-ai/FlagGems-sglang
- [19] FlagGems-sglang #315 (mxfp8 block-scaled matrix multiply) — https://github.com/flagos-ai/FlagGems-sglang/pull/315
- [20] FlagGems-sglang #336 (SGLang baseline workload) — https://github.com/flagos-ai/FlagGems-sglang/pull/336
- [21] FlagGems-sglang #311 (per-tensor absmax fp8) — https://github.com/flagos-ai/FlagGems-sglang/pull/311
- [22] FlagTree repository — https://github.com/flagos-ai/FlagTree
- [23] FlagTree #1532 (TLE local pointer proof) — https://github.com/flagos-ai/FlagTree/pull/1532
- [24] FlagTree #1409 (Iluvatar set_layout) — https://github.com/flagos-ai/FlagTree/pull/1409
- [25] FlagTree #1618 (Iluvatar LLVM 0.7.0) — https://github.com/flagos-ai/FlagTree/pull/1618
- [26] FlagTree #1592 (Moore Threads PH1 and SQMMA) — https://github.com/flagos-ai/FlagTree/pull/1592
- [27] FlagTree #1517 (four pipeline tuning switches) — https://github.com/flagos-ai/FlagTree/pull/1517
- [28] FlagTree #1597 (Hygon BW1100 test labels) — https://github.com/flagos-ai/FlagTree/pull/1597
- [29] build-infra repository — https://github.com/flagos-ai/build-infra
- [30] build-infra #1721 (Ascend 9.2.1 daily build line) — https://github.com/flagos-ai/build-infra/pull/1721
- [31] build-infra #1716 (2.3.0 records ROCm 7.2.4) — https://github.com/flagos-ai/build-infra/pull/1716
- [32] build-infra #1724 (Ascend ref parameter misread) — https://github.com/flagos-ai/build-infra/pull/1724
- [33] build-infra #1720 (Ascend clean wheel version number) — https://github.com/flagos-ai/build-infra/pull/1720
- [34] build-infra #1713 (2.3.0 pending entries) — https://github.com/flagos-ai/build-infra/pull/1713
- [35] FlagGems-vllm repository — https://github.com/flagos-ai/FlagGems-vllm
- [36] FlagGems-vllm #1148 (T-Head PPU wheel install items) — https://github.com/flagos-ai/FlagGems-vllm/pull/1148
- [37] FlagGems-vllm #967 (Ascend sparse FlashAttention) — https://github.com/flagos-ai/FlagGems-vllm/pull/967
- [38] FlagGems-vllm #1063 (prefetch depth selected by shape) — https://github.com/flagos-ai/FlagGems-vllm/pull/1063
- [39] FlagGems-vllm #1135 (intercept unregistered exported operators) — https://github.com/flagos-ai/FlagGems-vllm/pull/1135
- [40] FlagQuantum repository — https://github.com/flagos-ai/FlagQuantum
- [41] FlagQuantum #1120 (measurement to detector and logic records) — https://github.com/flagos-ai/FlagQuantum/pull/1120
- [42] FlagQuantum #1116 (six native record formats) — https://github.com/flagos-ai/FlagQuantum/pull/1116
- [43] FlagQuantum #1119 (Pauli stabilizer contraction) — https://github.com/flagos-ai/FlagQuantum/pull/1119
- [44] FlagTrain repository — https://github.com/flagos-ai/FlagTrain
- [45] FlagTrain #18 (DeepSpeed FP8 matrix multiply kernel) — https://github.com/flagos-ai/FlagTrain/pull/18
- [46] FlagCV repository — https://github.com/flagos-ai/FlagCV
- [47] optest #45 (FlagTrain support) — https://github.com/flagos-ai/optest/pull/45
- [48] FlagOS-Compressor repository — https://github.com/flagos-ai/FlagOS-Compressor
- [49] FlagSparse #118 (AlphaSparse merge) — https://github.com/flagos-ai/FlagSparse/pull/118
- [50] FlagOmni repository — https://github.com/flagos-ai/FlagOmni
- [51] docs #580 (FlagPrism vendor adaptation spec) — https://github.com/flagos-ai/docs/pull/580
- [52] FlagPrism repository — https://github.com/flagos-ai/FlagPrism
- [53] community #172 (FEP runtime and acceptance requirements) — https://github.com/flagos-ai/community/pull/172
- [54] Sohu (from competing on compute to deepening operators) — https://www.sohu.com/a/1099439543_610727
- [55] Sina Shanghai (Shanghai uses open-source platforms as a lever) — https://sh.sina.cn/news/2026-11-10/detail-iniyfuyy2664047.d.html
- [56] PEdaily (report on Shanghai open-source platform) — https://m.pedaily.cn/99discoveries/142897
- [57] Track 1 description (FlagOS X SGLang Operator Optimization Challenge) — https://segmentfault.com/a/1190000048185827
- [58] CSDN FlagOS column — https://flagos.csdn.net