FlagOS Daily Intelligence Report (2026-11-16)
Research window: 2026-11-13 06:00 ~ 2026-11-16 06:00 (Beijing time, 72 hours; continuing from the cutoff of the previous daily report, covering Saturday and Sunday).
Sources: GitHub (full verification of 64 repos under org: flagos-ai, 166 in-window commits, 223 merged PRs, 269 in-window updated open PRs, pushes to 19 repos, atomic verification of Releases in 31 repos and Tags in 8 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).
This Issue’s Index
- Today’s Highlights: Torch-FL compiles FlagGems’ Python submission path into a native host, cutting Qwen3-0.6B warm-start overhead to 1.17x that of vendor PyTorch (11-14~11-16)
- I. Open-Source Project Progress (GitHub Activity)
- 1.1 Organizational Cadence: 19 repos pushed, 166 commits, 223 merges, with the first day accounting for 80% of merges (11-13~11-15)
- 1.2 Release Milestones: The FEP merge deadline for 2.3 falls on the first day of the window; no new Releases or Tags within the window (11-13)
- 1.3 build-infra: 2.3.0 image lines for four backends land in the same week; PPU builder and MUSA 5.3.0 enter (11-13~11-15)
- 1.4 FlagGems: 112 merges; KernelGen’s 25 and 62 test-alignment commits proceed in parallel (11-13)
- 1.5 FlagGems-vllm: Hygon W8A8 MoE and NVIDIA Hopper pre-attention land in a single persistent kernel (11-13)
- 1.6 Torch-FL: eager host compiler pushes eight commits in the window; submission overhead drops from 211 ms to 28 ms (11-14~11-16)
- 1.7 FlagTree: TLE explicit layout, Thrive cumsum, Moore Threads KV physical layout (11-13~11-14)
- 1.8 Quantum Line: FlagQuantum 16 merges, Hubbard graph and single-site density advance; FlagQAI 10 commits (11-13~11-15)
- 1.9 Framework and Communication: FlagCX Hygon uniRunner, vllm-plugin-FL Iluvatar vLLM 0.32 lands (11-13~11-15)
- 1.10 Competitions and New Projects: 17 Track1 operators under review, FlagOmni’s first batch of LightX2V operators (11-13~11-15)
- II. News Coverage and Ecosystem
- 2.1 Channel Verification: 24 query groups yield 31 candidates; zero hits for FlagOS and component-name queries (11-13~11-15)
- 2.2 Ecosystem Background: Horizon Sunrise chips integrate with open-source embodied tools such as RLinf; Enflame co-founder and Chief Ecosystem Officer departs (11-13)
- III. Member Units and Backend Vendor Deep Dive
- 3.1 Hygon: MegaMoE folds routing and two-level W8A8 into a single persistent kernel (11-13)
- 3.2 Iluvatar CoreX: vLLM 0.32 adaptation merged, corex4.4.0 exits the application layer (11-13~11-15)
- 3.3 Ascend: CANN 9.2.1-910C image line established, A5 splits KDA and causal convolution into separate directories (11-13)
- 3.4 T-Head: PPU 2.1.0 runtime and FlagTree PPU builder paired on the same day (11-13)
- 3.5 Moore Threads: MUSA 5.3.0 line established and a batch of 14 dedicated operators (11-13~11-15)
- 3.6 Kunlunxin, Tsingmicro, MetaX, Enflame, Horizon: gap-filling, migration, and sunrise rows (11-13~11-15)
- IV. Summary and Trend Observations
- Appendix: Source Verification Table and Complete Source List
Today’s Focus: Torch-FL compiles FlagGems’ Python submission path into a native host, cutting Qwen3-0.6B warm-start overhead to 1.17x that of vendor PyTorch (11-14~11-16)
Date: 2026-11-14 ~ 2026-11-16 Sources: Torch-FL #824 · Torch-FL #839 · Torch-FL #822 (requirements issue) · Project 14
- Results: #824 compiles FlagGems’ Python-side host logic—metadata, memory allocation, heuristics, and Triton launch preparation—into a guarded native host program, while retaining the existing device kernels. Under full Qwen3-0.6B inference, warm-start CPU submission time for prefill/decode dropped from 211.51/203.92 ms to 28.46/27.15 ms; vendor PyTorch is 24.43/22.96 ms, giving latency ratios of 1.17 and 1.18.
- Acceptance threshold: #822 requires comparison against vendor PyTorch on the same physical card with the same workload on domestic accelerators, achieving a throughput ratio of no less than 0.80 or a latency ratio of no higher than 1.25. The NVIDIA side has met the threshold; no public data yet exists for domestic cards. #824 describes itself as an experimental NVIDIA inference implementation—not a default mode switch, and not a claim of full operator coverage.
- Continuity: Six PRs merged in succession from 11-15 through the early hours of 11-16—#826 compiles FlagGems source and model host, #835 brings the full native host registry into account, #836 generates eager entries with schema types, #837 generates source-call bridging, #838 preserves scalar semantics, and #839 reduces shared host overhead (+2388/-55, 15 files). This line advances in the order of “get a single model running, complete the registry, drive down overhead”—not scattered performance patches.
- Design orientation: The design doc in #823 set the route before implementation, concluding that memory allocation rather than kernel launch is what determines the ceiling, so the first phase takes memory directly from the allocator; the plan was approved by maintainers after #824 and #826 merged, with #829 and #830 carrying it forward.
- Implication: FlagGems’ eager path has always carried a “Python tax”—every call performs metadata, allocation, and launch preparation in Python. Compiling that layer away expands the cross-chip inference gap from “how well the kernels are written” to “how the host path is organized,” and the 2.3 inference performance acceptance criteria need to specify the host-side baseline as well.
I. Open Source Project Progress (GitHub Activity)
1.1 Organizational Cadence: 19 Repos Pushed, 166 Commits, 223 Merges; First Day Accounts for 80% of Merges (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Source: flagos-ai organization repositories · Merged PR search within window · FlagGems tags page · FlagTree tags page
- Totals: 19 of 64 repositories had pushes within the window; 166 new commits on default branches, 223 merged PRs, and 269 open PRs updated within the window, of which 132 were newly opened.
- Daily distribution: Commits were 125 on 11-13, 15 on 11-14, 25 on 11-15, and 1 in the early hours of 11-16; merges were 180 on 11-13, 15 on 11-14, 27 on 11-15, and 1 on 11-16. The first day accounts for 75.3% of commits and 80.7% of merges.
- Comparison with the previous period: The window expanded from 24 to 72 hours, yet commits fell from 193 to 166 (down 14.0%) while merges rose from 167 to 223 (up 33.5%). Weekdays are dominated by new commits and weekends by clearing reviews—the division of labor is clearly visible in the numbers.
- Merge distribution: FlagGems 112 (50.2%), build-infra 39 (17.5%), FlagQuantum 16 (7.2%), FlagGems-vllm 13 (5.8%), FlagQAI 10 (4.5%), Torch-FL 8, optest 6, FlagTree 5, FlagCX, vllm-plugin-FL, and FlagTrain 3 each, docs and sglang-plugin-FL 2 each, FlagGems-sglang 1.
- Commit distribution: FlagGems 50, build-infra 39, FlagQuantum 19, FlagGems-vllm 15, FlagQAI 12, Torch-FL 8, optest 6, docs 4, FlagTree 4, FlagCX 3, FlagTrain 3, FlagGems-sglang 2, sglang-plugin-FL 1.
- Open PR structure: Of the 132 newly opened, FlagGems 44, FlagGems-vllm 38, FlagGems-sglang 20, FlagTree 10, FlagAttention 4, with the rest spread across 10 repositories including docs, vllm-plugin-FL, FlagTrain, sglang-plugin-FL, and build-infra.
- Pushed but no mainline commits: Six repositories—vllm-plugin-FL, FlagFFT, FlagOmni, FlagSparse, Trident, and release-info—had pushed_at timestamps within the window but no commits on their default branches within it. FlagOmni and Trident activity is on feature branches and PRs, release-info is a gh-pages publishing page, and FlagFFT and FlagSparse are branch pushes; none are counted in this period’s commit totals.
1.2 Version Milestones: The 2.3 FEP Merge Deadline Falls on the First Day of the Window; No New Releases or Tags Within the Window (11-13)
Date: 2026-11-13 Source: community 2.3 release schedule · FlagGems tags page · FlagTree tags page · docs #584
- Hard deadline reached: The community 2.3 schedule sets 2026-11-13 as the FEP merge deadline; FEPs targeting 2.3 must be merged by this date, and those not merged are marked Deferred and carried to the next version without retroactive acceptance. The first day of the window falls exactly on this gate.
- Subsequent milestones: Feature freeze and rc0 cut on 11-30, requiring consistency across tag, manifest, and package metadata, with the RM using the rc0 manifest for a full-module build that week; rc1 on 12-07; rc2 on 12-14; test validation ends and the release is sealed on 12-25; GA on 12-30, with the RM issuing a written report 48 hours before GA and the TSC voting in the issue.
- No new versions within the window: Atomic checks of Releases across 31 repositories and Tags across 8 repositories found no additions within the window; the two most recent are FlagGems v5.5.0.dev1 (11-11 18:25) and FlagTree 0.7.1+xpu3.6 (11-12 19:28), both covered in the previous report.
- Community repo quiet: The community repository’s last push was 11-12 00:49, with no merges within the window; no new FEP entry records appeared on the FEP merge deadline day itself, meaning the 2.3 proposal surface had already closed before the deadline.
- Documentation side: docs merged two manual changes—#584 removed the release guide page and updated the model list, and #585 updated the model list again by platform; three model-list bot PRs from ModelScope, HuggingFace, and aihuanxin are also under review.
1.3 build-infra: Four Backend 2.3.0 Image Lines Land in the Same Week; PPU Builder and MUSA 5.3.0 Enter (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Source: build-infra #1789 (PPU builder) · build-infra #1792 (MUSA 5.3.0 base) · build-infra #1770 (910C SDK) · build-infra #1780 (AMD daily row) · build-infra repository
- Scale: 39 merges correspond one-to-one with 39 commits; tengqm led the manual changes, with github-actions backfilling changelog dates. This is the second-heaviest repository in the window, after FlagGems.
- Four backends complete 2.3.0 version data: Ascend ascend-cann9.2.1-910c (#1770 fetches the SDK from the in-node Nexus, #1776 builds runtime the same way, #1772 and #1777 backfill the changelog, #1778 adds a description page); AMD amd-rocm7.2.4 (#1780 enables flagtree’s amd daily row, #1781 pre-seeds the changelog, #1782 backfills it, #1783 reorders the Chinese); T-Head thead-ppu2.1.0 (#1789 PPU builder and daily row, #1790 pre-seeds, #1791 backfills); Moore Threads mthreads-musa5.3.0 (starting with #1792, including base, runtime, flagcx packaging, and toolkit handling—over ten commits in total).
- T-Head PPU builder: The vendor wheel 0.7.0+ppu3.6 lacks Project-URL, making the build commit untraceable when problems arise. The builder reuses amd’s two-stage approach—bare ubuntu 24.04 plus pre-staged LLVM and triton dependencies, compiling third_party/amd without laying down the PPU SDK; the first main wheel is A/B tested against the vendor wheel in the thead-ppu2.1.0 runtime, with the nine-operator sweep and torch.compile maxdiff fully matching.
- Moore Threads MUSA 5.3.0: The base image uses ubuntu 24.04 (MUSA 5.3.0, MCCL 2.5.0, muDNN 3.5.0), retaining 24.04’s newer libstdc++ rather than falling back to 22.04 as the other two backends do; this line only has cp312 vendor wheels, so the config specifies python 3.12, driver 5.3.0-server, and the torch 2.13 stack. The container toolkit layer is disabled (#1817): when the node driver is already 5.3.0, mtml 2.4.2 is a single libmtml.so, while mt-container-toolkit 2.1.0 still looks for the old multi-file layout—injection reports no error but the container starts empty, which would be misread as a broken image.
- Packaging line: The flagcx builder adds a wheel path with two-stage validation—building the wheel in the builder and verifying the deployed set in the runtime (#1806, #1808 through #1811); FlagTree adds a cp312 wheel row (#1807); #1812 pins FlagTree and flagcx for the MUSA runtime; #1814 makes flagtree’s precompiled inputs check against upstream.
- Closing details: The days backend drops corex4.4.0 from the application layer (#1771), and iluvatar’s LLVM is pinned to v0.7.0 (#1819). gendoc gets five consecutive hardening commits—reject rendering on an empty version set (#1796), lower the retry cap from 50 to 10 (#1798), preserve the collection set across retries (#1799), fail upon reaching the cap (#1800), and report backends not collected (#1801); #1802 makes the record commit rebase separately without touching the entire working tree.
- Under review: #1784 enables RPM rows for libtriton_jit on Fedora 43 and openEuler 24.03; #1820 switches sunrise’s precompiled inputs to FlagTree main’s new cache keys (llvm-ff67ef55, sunriseTritonPlugin_v0.7.0.1), otherwise the daily wheel would fall back to a vendor object store unreachable from the node.
1.4 FlagGems: 112 Merges; KernelGen’s 25 and 62 Test Alignments Run in Parallel (11-13)
Date: 2026-11-13 Source: FlagGems repository · FlagGems #9118 (82 platform test packages) · FlagGems #7910 (PPU grouped_gemm_w8a8) · FlagGems #8874 (int8/uint8)
- Scale: 112 merges and 50 commits, accounting for half of all organizational merges. Author concentration is high—lukatao 62, Yukun-Cui 14, youseeyouonedayday 9, 103yiran 6, bin913 4, and rikyaa, llaboon, rainlyxp, sunge666-ui 3 each.
- Test alignment batch: lukatao’s 62 merges landed consecutively within 52 minutes from 11-13 15:17 to 16:09, themed around aligning vendor branch test cases to master—the days side’s FP64 slogdet, Complex128 reference, and skip scope for row_stack and replication_pad; MetaX’s LU solve and FP32 reference; T-Head’s Bessel K0, Cholesky, SVD, Lanczos, and special_multigammaln; Hygon’s geqrf and put. This batch produces no new operators but determines whether “which is correct, the vendor branch or mainline” can be judged automatically.
- KernelGen 25: Moore Threads 14 (convolution_double_backward, conv_tbc, upsample_nearest3d_backward and grad_input, greater, grid_sampler_3d, convolution_overrideable, index_select_backward, convolution_mode, hardtanh_backward, multiply, _weight_norm, _adaptive_avg_pool3d_backward, upsample_trilinear3d_backward); NVIDIA 8 (_cast_Double, _sparse_csr_sum, cholesky_inverse precision, cross_attention’s TLE dependency, mixed_dtypes_linear shared memory overflow, fake_quantize rounding, fused_moving_avg asymmetric quantization, Hermite float64 pointer); Enflame 3 (torch.gcu device lookup, float_power’s scalar-tensor and tensor-tensor paths).
- Kunlunxin 9: upsample backward, trapezoid weighted fast path, mixed_dtypes_linear and margin_ranking_loss, geqrf and grid_sampler tuning, linalg_polar/tensorinv/lstm, diagflat’s detach_copy, gradient skipping fp64, view_as_real_copy.
- Ascend 4: index_put keeps the size parameter at runtime so single-element indexing compiles (#6701); gelu’s pow literal (#6719); ones/ones_like empty-tensor guard (#6728); lift_fresh does not launch a zero grid on empty tensors (#6802).
- T-Head: #9029 fixes upsample_bilinear2d for non-contiguous inputs; #7910 adds grouped_gemm_w8a8—an unfilled grouped NT GEMM with signed INT8, FP32 per-row and per-output-channel scales, BF16 output, INT32 full-K accumulation, K capped at 131071, registered only on the PPU backend (+3435/-0, 7 files).
- Other: #8592 tiles LayerNorm by row on NVIDIA and returns FP32 statistics; #8874 adds int8/uint8 support for acosh, amin, atanh, atan2, arctan2, and arcsin; #9079 splits out metadata for deprecated fused operators separately; #9085 adds a CI check for torch computation calls in operator code.
- KernelGen pilot launch: #9118 merges 82 reviewed platform pytest packages (57 operator IDs) into kernelgen-dev—335 files, +51687 lines; content addressing preserves platform variants of shared files, with each manifest binding master and dev revisions, test and helper hashes, review reference hashes, and summaries of two passing preparation reports.
1.5 FlagGems-vllm: Hygon W8A8 MoE and NVIDIA Hopper Pre-Attention Land in a Single Persistent Kernel (11-13)
Date: 2026-11-13 to 2026-11-15 Source: FlagGems-vllm #1020 (Hygon MegaMoE) · FlagGems-vllm #1007 (Hopper MegaPreAttention) · FlagGems-vllm #1204 (Ascend A5) · FlagGems-vllm #1202 (CI self-contained)
- Scale: 13 merges and 15 commits; 59 open PRs, of which 38 are newly opened—the repository with the heaviest review backlog this week.
- Hygon W8A8 DUSHMEM MegaMoE: #1020 adds expert-parallel megamoe_int8 for gfx936, exposed through the Hygon backend and the public namespace. A single persistent FlagTree/TLE kernel runs routing packing, DUSHMEM peer publish and aggregation, W8A8 L1, dynamic INT8 qA2, W8A8 L2, return, and TopK merge; DUSHMEM initialization, symmetric allocation, epoch management, and destruction are held by the caller—the operator only consumes an explicit context and creates no implicit global communication state (+4478/-31, 14 files).
- NVIDIA Hopper MegaPreAttention: #1007 fuses tensor-parallel AllGather, gated QKV projection, Gemma RMSNorm, interleaved MRoPE, and KV cache update into a single persistent kernel, overlapping communication and computation with tile-ready signals; BF16, supporting TP2/TP4/TP8, head_dim 256, rotary_dim 64, with grid and communication metadata determined by the device’s actual SM count (+2250/-0, 4 files).
- Ascend A5: #1204 moves ascend_kda’s optimized implementation into runtime/backend/_ascend/a5—chunk_kda and fused_recurrent_kda under FLA, causal_conv1d_fn and causal_conv1d_update under ops; selected on Ascend 950/A5 via architecture detection, with other devices retaining the original implementation, and four operators get beta 0.3 metadata in operators.yaml (+3817/-19, 13 files).
- Other merges: #1015 adds the qsa_prepare baseline toolchain and dispatches by vendor tuning; #1205 optimizes the forward intra of chunk gated delta rule on the 950 NPU; #1215 relaxes mhc_post’s d-chunk to 4096 elements; #1217 removes the three dispatch registrations enable, only_enable, and use_gems; #1202 makes CI self-contained and adds triage, daily, release, and triton-test lines; #1156 removes hugo-site’s workflow_run trigger.
- Under review: Hygon MegaPreAttention, NVIDIA MegaMHC and MegaGate,
select_candidate_blocks_cooperative for DeepSeek-V4.1 decode, MetaX quantization and routing optimizations, and others—all fused operators and decode paths.
1.6 Torch-FL: Eager Host Compiler Ships Eight Commits in Eight Days; Commit Overhead Cut from 211 ms to 28 ms (11-14 to 11-16)
Date: 2026-11-14 to 2026-11-16 Source: Torch-FL #824 · Torch-FL #826 · Torch-FL #835 · Torch-FL #837 · Torch-FL #838
- Scale: 8 merges and 8 commits, all between 11-14 12:48 and 11-16 00:17—the most continuous line over the weekend; all by lvyufeng alone, of which #825 is a CI fix pointing the NVIDIA channel at the A100 host label, with the other seven centered on the host compiler.
- Path: #824 generates a guarded native host program while retaining the original device kernel; #826 compiles FlagGems source together with the model host; #835 makes native host registry overrides verifiable; #836 generates schema-typed eager entries; #837 generates source call bridging; #838 preserves scalar semantics; #839 reduces shared host overhead. The seven revolve around one goal: making a hot eager call do no Python work at all.
- Performance: For full Qwen3-0.6B inference, prefill/decode hot-start CPU submission drops from 211.51/203.92 ms to 28.46/27.15 ms, versus 24.43/22.96 ms for vendor PyTorch—latency ratios of 1.17 and 1.18.
- Boundaries: Currently only NVIDIA-side data exists; the implementation describes itself as experimental and requires explicit enablement; numerical kernels still come from pinned FlagGems and FlagTree wheels, with no hand-written Triton kernels, Inductor kernels, or model capture allowed, nor may host compiler limitations be disguised as platform operator gaps.
- Assessment: Treating “FlagGems’ Python submission overhead” as the number-one problem marks a departure from past efforts focused solely on operator kernels; if domestic cards can also meet the 80% threshold, the cross-chip inference comparison baseline should shift to “vendor PyTorch plus the same workload.”
1.7 FlagTree: TLE Explicit Layout, Thrive cumsum, Moore Threads KV Physical Layout (11-13 to 11-14)
Date: 2026-11-13 to 2026-11-14 Source: FlagTree #1536 (TLE explicit layout) · FlagTree #1649 (Thrive cumsum) · FlagTree #1450 (KV physical layout) · FlagTree #1650 (nvidia3.5 build)
- 5 merges, 4 commits: #1639 fixes fork PR author and PR lookup; #1536 makes TLE respect explicit layouts between memory and structured control flow (+1850, 32 files); #1649 adds TLE cumsum and tile operators to the Thrive backend (via ConvertTleToThvTile); #1450 makes KV physical layout shared across SQMMA views (+340, 3 files); #1650 fixes the nvidia3.5 build.
- Details of #1536: TLE kernels use explicit layouts to control memory access and inter-thread value distribution; compilation previously conflated such contracts with the intrinsic layout of view operations or another signature of scf.while, causing conversion failures, invalid IR, or errors propagated across layout boundaries. The fix inserts conversions at relevant use sites, preserves explicit memory and consumer layouts, and handles scalar and tensor reductions, scan and gather, views, and warp counts.
- Details of #1450: When the same staged KV allocation is shared by wgmma(q, tile, trans_b=True) and wgmma(p, tile), it was rejected because logical operand metadata was propagated onto the physical allocation root. The fix tracks physical root requirements separately from logical contracts, retains logical contracts only on the final view, and requires exact equality of the root’s MemDescType before reuse—without inferring equivalence of arbitrary physical address mappings.
- Under review: 21 PRs, of which 10 are newly opened; this period’s merges are all concentrated on TLE and backend builds, with no version number changes.
1.8 Quantum Line: FlagQuantum Merges 16, Advancing Hubbard Graphs and One-Body Density; FlagQAI 10 (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Source: FlagQuantum #1176 (Hubbard graph training) · FlagQuantum #1179 (one-body density) · FlagQuantum #1180 (stabilizer sampling) · FlagQuantum #1084 (FTQC integration, under review)
- FlagQuantum 19 commits, 16 merges: mixed-axis Pauli sector training (#1170), bounded training of the H2 spin sector (#1171) and cross-native-encoding training (#1172), Hubbard model assembly on explicit graphs (#1174) and graph-pair training under three encodings (#1176), OpenFermion gap audit and next delivery plan (#1178), one-body density tensor measurement (#1179) and capability registry (#1181), compiled DEM sampling (#1157), noiseless stabilizer sampling (#1180), and FP64 statevector training performance on Torch-FL (#1164).
- Under review: #1084 is an integration draft for native fault-tolerant compilation and reversible phase execution—1822 files, +285308/-566, including G1 slicing and G2 arithmetic and binary-field modules, reviewed in batches by commit; #1182 adds two-body density tensors.
- FlagQAI 12 commits, 10 merges: On the web side, binding research metrics to saved experiment versions, comparing research results with resource costs, browsing anonymous circuits by ordered gate intervals, preserving SDK synthesis attribution, and displaying full state-vector probabilities; on the CLI side, fixing DeepSeek inference replay across protocols, message receipts persisting across service restarts, model routing recovery after session interruption, and macOS native file read convergence.
- Assessment: The quantum line remains the fastest-moving direction within FlagOS, but its core is simulation and algorithms, and FlagQAI merely turns the quantum experiment workflow into an operable agent interface; it should not be read as release-grade.
1.9 Frameworks and Communication: FlagCX Hygon uniRunner, vllm-plugin-FL Days vLLM 0.32 Lands (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Source: FlagCX #713 (uniRunner) · FlagCX #717 (IPC path) · vllm-plugin-FL #674 (days 0.32) · vllm-plugin-FL #742
- FlagCX 3: #710 adds templated NVIDIA memory operations and CUDA traits tests to UIL; #713 makes CUDA and Hygon share the uniRunner device kernel and GEMM helper, enables kernel-backed uniRunner on Hygon with FLAGCX_SIMT_WIDTH set to a 64-thread wavefront, and adds perf-experimental-test (hygon) covering IpcRingAR, IPC buffer rebinding, SlicedAR, RingAG, GroupedAG, P2pAG, AllGatherGemm, and GemmReduceScatter; #717 removes a redundant IPC pointer path.
- vllm-plugin-FL 3: #674 adapts the days backend to vLLM 0.32 with a regression gate (+1808/-285, 58 files); #742 clears the remaining 0.28 paths in the V1 runner (T-Head); #738 elevates device to a first-class configuration dimension in NVIDIA CI.
- Review queue: vllm-plugin-FL 15—MUSA 0.32 (#715), Hygon DCU 0.32 (#711), MetaX C550 0.32 (#693), Kunlunxin P800 0.32 (#726), Ascend Runner V2 (#703), Enflame S60 0.32 (#675), AMD W7900D graph mode (#659), Sunrise 0.32 empty and Runner V2 (#697); sglang-plugin-FL 7—Qingwei TX8110’s SGLang 0.5.18 adaptation (#187), T-Head 2.3.0-runtime CI temporary validation (#186), NVIDIA and Hygon 0.5.18 CI images (#179, #180), Kunlunxin 0.5.18 (#155), Moore Threads MiniMax-M3 (#158, #165).
- Training frameworks: Megatron-LM-FL #225 is the fourth commit in the DeepSeek-V4 sync, migrating SBHD compressed sparse attention and DSv4 hybrid attention to core fused kernels (+13113/-5163, 30 files), depending on the previous DSA migration branch; TransformerEngine-FL #160 runs the full plugin test suite on all platforms and fixes portability issues in Triton, NPU, Enflame, and FlagOS backends (+3328/-1132, 69 files), with still-skipped cases individually documented with reason, owner, and restoration condition.
- Evaluation: optest 6—automatic baseline data update (#54), sglang baseline update (#55), enabling Enflame S60 tests for FlagAttention (#52), more hardware backends for FlagTrain (#50), cleaning up redundant ops-test branch inputs (#51), and fixing FlagTrain and FlagGems-vllm workflows (#53).
1.10 Competition and New Projects: 17 Track1 Operators Under Review, FlagOmni’s First LightX2V Operators (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Source: FlagGems-sglang #370 (Track1 operator table) · FlagGems-sglang #372 (Track1 under review) · FlagOmni #1 (LightX2V operators, under review) · Trident #11 (guard tree, under review)
- Competition Track1 (KV cache quantization): FlagGems-sglang opened 20 new PRs within the window, 17 of them labeled FlagOS Competition-Track1, mostly around SGLang’s FP8 KV cache—assign_extend_cache_locs_uniform, set_mla_kv_buffer_fp8_quant, dequantize_k_cache_paged, act_quant, and a dozen other kernels.
- Competition Track1 (MLA and linear-layer quantization): Another batch around MLA and matrix multiplication—mla_kv_pack_quantize_fp8, triton_sparse_mla_fwd, per_tensor_quant_mla_fp8, absorbed_bmm_concat_cast_q_fp8, and others. In addition, #371 is a performance optimization for AMD gfx1100, and #385 is an Enflame migration.
- Competition entry: FlagGems-sglang #370 has merged, filling in the 17 Track1 operators missing from operators.yaml—competition submissions now have a unified operator table to check against, and reviewers can judge coverage by the table.
- FlagOmni: #1 is under review, moving LightX2V’s numerical operators into FlagOmni—RMS norm, packed SwiGLU, QK norm and rotary, residual fusion, reference blend, rotary tables, and SageAttention; portable implementations live in lightx2v/ops and hardware implementations in runtime/backend; SageAttention has a single public entry dispatching by platform, with BF16 QKV taking the portable path and FP8 PV taking the SM90 implementation (+5818/-16, 77 files).
- lightx2v-plugin-FL: #8 is under review, wiring H3 operators on top of global policies and dispatch; the default H3 launcher leaves host-side FFN, rotary, scheduler, and VAE unchanged, with fusion enabled explicitly only via –enable-fusions; the Kunlunxin P800 packaged launcher has been validated (plugin commit 5d34048, FlagOmni commit 431a7e6).
- Trident: #11 is under review, refactoring guard dispatch with a guard tree to lay groundwork for shared checks among candidates (+5627/-3795, 90 files). Trident itself is an experimental bridge that compiles Python functions containing Triton and Torch-style computation to MLIR/LLVM, then executes via tvm_ffi; the repository was created on 10-12, and this is its only activity this period.
- FlagAttention: 5, all newly opened and under review—Ascend chunked GLA frontend and backend (#135, +4295, 12 files, including fixed-length and packed varlen, fusing A computation into the output kernel at D=128), Hygon GLA dispatch and validation simplification (#136), NVIDIA baseline and optest performance comparison (#137), Ascend compatibility fix (#138), and GDN prefill and recurrent decode optimization (#113).
II. News Coverage and Ecosystem
2.1 Channel Verification: 24 Query Sets, 31 Candidates, Zero Hits for FlagOS and Component-Name Queries (11-13 to 11-15)
Date: 2026-11-13 to 2026-11-15 Sources: Google News RSS · HN Algolia · CSDN FlagOS Column · BAAI Community
- gnews results: 24 Chinese and English queries, deduplicated over a 72-hour window, yielded 31 candidates. Queries for FlagOS, FlagGems, FlagScale, FlagTree, FlagCX, KernelGen, Qingwei Intelligence, MetaX open source, Hygon open-source operators, BAAI FlagOS, and flagos-ai all returned zero hits; the hits were 13 for Enflame, 12 related to BAAI (7 open source, 3 chips, 2 large-model open-source stacks), 4 for domestic compute, and 1 each in Chinese and English for Moore Threads.
- Removal breakdown: Of the 13 Enflame items, 10 were stock-price and earnings pieces (shares up over 86% in 8 trading days, intraday gains above 9%, among the week’s top ten gainers, etc.) and were removed as a batch of market coverage; 2 concerned executive departures and 1 was the Q3 earnings briefing. The BAAI Community items were model, paper, and event pieces (MiniMax image generation, ERNIE large model, Andrew Yao team paper, Tesla and SpaceX) and unrelated to the FlagOS project. On the English side, the single Moore Threads query actually hit a report on the Horizon Robotics–Infinigence AI partnership, unrelated to Moore Threads, and was reclassified as ecosystem background for 2.2.
- HN: 4 hits for FlagOS but 0 direct matches; all other component queries returned 0, all tokenization false positives.
- CSDN column: The latest item on the homepage was created at 11-13 11:44, but the original publication date shown on the card is 2026-08-27 (Qwen3.8-Flash-Next, LightX2V, etc.), indicating redistribution of historical articles; no new original content within the window.
- BAAI Community: The search page is client-rendered and items could not be scraped; the homepage carries model and popular-science content with no FlagOS project updates.
- Conclusion: The news side remains quiet for a consecutive period; this issue’s facts are sourced primarily from GitHub. Zero hits for component-name queries have now persisted across multiple windows and are treated as normal rather than as a “collection failure.”
2.2 Ecosystem Background: Horizon Robotics Sunrise Chips Adopted by RLinf and Other Open-Source Embodied Tools; Enflame Co-Founder and Chief Ecosystem Officer Departs (11-13)
Date: 2026-11-13 Sources: 42HOW (Horizon Robotics–Infinigence AI partnership) · Robothot (same partnership) · Tencent News (Enflame executive change) · NetEase (Leiphone exclusive)
- Horizon Robotics Sunrise: On 11-13, D-Robotics and Infinigence AI announced a strategic partnership to adapt RLinf and three other open-source robotics tools to Sunrise chips. This chip line already has a corresponding backend in FlagOS — the build-infra configuration includes a sunrise row (hardware Sunrise SR-SUN-S2-X1, driver 0.24.0, TangRT 1.2.0, flagtree 0.7.0+sunrise3.6), and vllm-plugin-FL has a Sunrise on vLLM 0.32 adaptation under review (#697). The partnership itself does not involve FlagOS and is recorded only as ecosystem background.
- Enflame personnel: Leiphone reported exclusively on 11-13 that co-founder and Chief Ecosystem Officer Li Xingyu had departed; Southern Metropolis Daily followed the same day, and the company had not responded as of publication. Enflame is the vendor behind FlagOS’s Enflame backend; its code activity in this window consisted mainly of three KernelGen fixes in FlagGems and an under-review portability fix in TransformerEngine-FL, with no change in pace observed.
- MetaX funding: Xixuan Chuangzhi, jointly initiated by UBTech and MetaX, completed a RMB 500 million angel round at a post-money valuation of RMB 4 billion (reported 11-13). MetaX is the vendor behind FlagOS’s MetaX backend; this is a capital-level matter. On the code side this issue, its activity is test-criteria alignment and a 0.32 adaptation for C550 under review.
- Handling policy: All three items above are treated as ecosystem background, not counted as project updates, and not used as evidence of FlagOS progress. Industrial moves by member organizations are recorded separately from project code activity.
III. Deep Dive into Member Units and Backend Vendors
3.1 Hygon: MegaMoE Folds Routing and Two-Level W8A8 into a Single Persistent Kernel (11-13)
Date: 2026-11-13 Source: FlagGems-vllm #1020 · FlagCX #713 · FlagGems #8918 (under review) · FlagAttention #136 (under review)
- Inference fusion: #1020 is Hygon’s heaviest contribution this period (+4478/-31). A single persistent kernel performs routing packing, DUSHMEM peer publish and aggregation, two-level W8A8, and TopK merging, while handing the communication context’s lifecycle back to the caller. For multi-card inference, this changes usability far more than adding another batch of pointwise operators.
- Communication: FlagCX #713 lets CUDA and Hygon share the uniRunner device kernel and GEMM helper; the Hygon side sets FLAGCX_SIMT_WIDTH per 64-thread wavefront and brings Hygon’s experimental performance tests into CI, covering 8 collective communication cases.
- Under review: FlagGems #8918 fixes the Hygon optimization of add on master, and #9045 optimizes linear projection dispatch and kernels; FlagAttention #136 simplifies Hygon GLA dispatch and validation; sglang-plugin-FL #180 uses the SGLang 0.5.18 CI image.
- Testing: FlagGems’ 62 alignment PRs include the Hygon portions of put and geqrf (#8953, #8954), pulling vendor-branch acceptance criteria back to mainline.
3.2 Iluvatar: vLLM 0.32 Adaptation Merged, corex4.4.0 Retired from the Application Layer (11-13~11-15)
Date: 2026-11-13 ~ 2026-11-15 Source: vllm-plugin-FL #674 · build-infra #1771 · build-infra #1819 · FlagGems #8940 (test scope)
- vLLM 0.32 lands: #674 adapts the Iluvatar backend to vLLM 0.32 with a regression gate (+1808/-285, 58 files), merged 11-15 22:11. This makes Iluvatar the third vendor to complete 0.32 landing, after NVIDIA (#737) and T-Head (#706).
- Build side: The application layer drops corex4.4.0 (#1771), and iluvatar’s LLVM is pinned to v0.7.0 in the flagtree precompiled inputs (#1819)—with the SDK baseline moved forward to corex5.0.1, the old build layer begins to retire.
- Test side: Iluvatar cases in FlagGems are largely “scope-restricted” rather than relaxed this period—the FP64 cast skip for row_stack, the Complex128 skip for replication_pad1d_backward, and the mixed-precision skip for triplet loss are all changed to apply only to Iluvatar (#8924, #8940 to #8942), and the reference implementations for Cholesky and slogdet are fixed in the same batch.
- Assessment: This round for Iluvatar is “framework migration plus test-criteria convergence,” not a push for new operator volume; after 0.32 lands, its position in 2.3 depends on the regression gate’s pass rate.
3.3 Ascend: CANN 9.2.1-910C Image Becomes a Full Line, A5 Splits KDA and Causal Convolution into Separate Directories (11-13)
Date: 2026-11-13 Source: build-infra #1770 · build-infra #1776 · FlagGems-vllm #1204 · FlagGems #9068 (Ascend sharded tests)
- 910C becomes a full line: #1770 and #1776 solve the 910C’s problem of fetching the SDK from the in-node Nexus (on both base and runtime sides), #1772 and #1777 backfill the changelog, and #1778 adds a description page; 910C goes from a single base row to a complete line with docs, runtime, and version data.
- A5 directory split: #1204 moves the optimized implementations of KDA and causal convolution into runtime/backend/_ascend/a5 and performs architecture detection by Ascend 950/A5. This means Ascend goes from being a single backend in FlagOS to being “split into directories by generation.”
- Operators and CI: Four FlagGems PRs fix empty-tensor and literal issues in index_put, gelu, ones_like, and lift_fresh; #9068 adds sharded operator tests for Ascend and produces coverage artifacts; FlagGems-vllm #1168 and #1169 switch Ascend performance tests to the Benchmark class; #1215 relaxes the d-chunk of mhc_post.
- Under review: FlagAttention #135’s Ascend chunked GLA frontend and backend (+4295), and Megatron-LM-FL #198’s native and Triton Ascend SwiGLU.
3.4 T-Head: PPU 2.1.0 Runtime and FlagTree PPU Builder Land as a Pair on the Same Day (11-13)
Date: 2026-11-13 Source: build-infra #1789 · build-infra #1791 · FlagGems #7910 · FlagGems #9029
- Paired landing: #1789 adds the FlagTree PPU builder and daily row, and #1791 records the runtime changelog for thead-ppu2.1.0; both complete on the same day, marking the first time PPU’s compile and runtime sides share the same batch of version data.
- Significance of the builder: The vendor wheel 0.7.0+ppu3.6 lacks Project-URL, so build commits cannot be located when problems arise; with the self-built builder, the daily wheel and vendor wheel match on maxdiff across the nine-operator scan and torch.compile, effectively adding “traceability.”
- Operators: FlagGems #7910’s grouped_gemm_w8a8 (+3435) is registered only on the PPU backend, with signed INT8, FP32 dual scales, BF16 output, INT32 full-K accumulation, and a K upper limit of 131071; #9029 fixes non-contiguous input for upsample_bilinear2d.
- Framework: sglang-plugin-FL #186 performs T-Head hardware validation on the 2.3.0-runtime CI image (temporary), and vllm-plugin-FL #742 removes the V1 runner’s 0.28 path.
- Terminology: PPU in this section refers to T-Head’s AI accelerator card product line and is unrelated to DAMO XuanTie’s RISC-V processor line; thead is the upstream branch name T-Head continues to use and must not be written as “DAMO XuanTie.”
3.5 Moore Threads: MUSA 5.3.0 Becomes a Full Line and a Batch of 14 Dedicated Operators (11-13~11-15)
Date: 2026-11-13 ~ 2026-11-15 Source: build-infra #1792 · build-infra #1805 · build-infra #1817 · FlagTree #1450
- MUSA 5.3.0 becomes a full line: base (#1792), runtime (#1805), flagcx builder and wheel (#1806, #1808 to #1811), FlagTree cp312 wheel (#1807), FlagTree and flagcx pinning (#1812), and toolkit layer handling (#1816, #1817)—more than ten PRs complete within the same week; this is the most complete new SDK line in build-infra this period.
- Driver details: The 5.3.0 driver deb can be installed on 24.04, so this line retains 24.04’s newer libstdc++; the container toolkit layer is disabled due to layout differences between mtml 2.4.2 and mt-container-toolkit 2.1.0 (#1817), otherwise an empty container startup would be misjudged as a broken image.
- Operators: KernelGen’s 14 Moore Threads dedicated operators are merged in batches from the morning to the afternoon of 11-13, covering convolution double backward, conv_tbc, upsampling backward and its grad_input, grid_sampler_3d, index_select backward, hardtanh backward, _weight_norm, adaptive_avg_pool3d backward, and trilinear3d backward.
- Compilation: FlagTree #1450 lets KV’s physical layout be shared across SQMMA views, directly corresponding to Moore Threads’ wgmma path.
- Under review: sglang-plugin-FL #158 and #165 (MUSA support and dynamic INT8 for MiniMax-M3), vllm-plugin-FL #715 (MUSA on 0.32).
3.6 Kunlunxin, Tsingmicro, MetaX, Enflame, Horizon Robotics: Gap-Filling, Migration, and the sunrise Row (11-13~11-15)
Date: 2026-11-13 ~ 2026-11-15 Source: FlagGems #8646 (Kunlunxin) · FlagGems #8758 (Tsingmicro) · sglang-plugin-FL #187 (Tsingmicro TX8110) · build-infra #1820 (sunrise)
- Kunlunxin: 9 FlagGems PRs (upsample backward, trapezoid, mixed_dtypes_linear, geqrf and grid_sampler, linalg_polar/tensorinv/lstm, diagflat, gradient fp64 skip, view_as_real_copy, plus one code-style cleanup); lightx2v-plugin-FL #8’s P800 packaging launcher is verified; vllm-plugin-FL #726 (P800 on 0.32) and sglang-plugin-FL #155 (0.5.18) remain under review.
- Tsingmicro: FlagGems #8758 merges the out-of-bounds tile skip for tx81/tx82’s GSL persistent-loop; sglang-plugin-FL #187 brings TX8110 (TXDA) to SGLang 0.5.18 (under review, +3173/-139); FlagTree’s 0.7.1 version remains at a1.
- MetaX: FlagGems’ MetaX cases are aligned this period (#8927’s RNN ReLU and FP32 reference, #8931’s LU solve); vllm-plugin-FL #693 (C550 on 0.32), #635 (Qwen3.8-Flash-Next), #713 (HY4 sparse MLA and pipeline parallelism), and #541 (DeepSeek-V4 Flash W8A8) are all under review.
- Enflame: Three FlagGems KernelGen fixes (#9028’s torch.gcu device lookup, #9050 and #9051’s two float_power paths); TransformerEngine-FL #160 fixes Enflame backend portability (under review); vllm-plugin-FL #675 (S60 on 0.32) is under review.
- Horizon Robotics: build-infra’s sunrise row (Sunrise SR-SUN-S2-X1, TangRT 1.2.0, flagtree 0.7.0+sunrise3.6) has its precompiled input cache key fixed by #1820 this period; vllm-plugin-FL #697’s Sunrise on 0.32 empty is under review.
IV. Summary and Trend Observations
- 2.3 reaches the eve of rc: The 11-13 FEP merge deadline has passed, and the next gate is the 11-30 feature freeze and rc0. Within the window, build-infra completes the 2.3.0 version data for four backends—910C, ROCm 7.2.4, PPU 2.1.0, and MUSA 5.3.0—with the closing unit still “one backend, one PR.”
- The performance narrative shifts from “operator count” to “host path”: Torch-FL brings FlagGems’ Python submission overhead down to 1.17 to 1.18 times that of vendor PyTorch, entering #822’s 80% threshold. If reproduced on domestic cards, cross-chip inference comparisons must also provide a host baseline, otherwise the numbers are not comparable.
- Fused operators are the main battleground on the inference side: Hygon MegaMoE, NVIDIA MegaPreAttention, and Ascend A5’s KDA and causal convolution all bind communication, projection, normalization, and KV updates into a single persistent kernel. What this kind of work changes is multi-card usability, not single-point compute.
- The weight of testing and CI rises: FlagGems’ 62 alignment PRs, 82 platform pytest packages (#9118), FlagGems-vllm’s CI self-containment (#1202), and five gendoc hardening PRs all point to the same thing—2.3’s bottleneck is shifting from “writing code” to “acceptance and traceability.”
- The weekend is not downtime: 11-13 saw 180 merges and 125 commits in a single day, while 11-14 and 11-15 saw 15 and 27 merges respectively. Reviews accumulated on weekdays are cleared over the weekend, while genuinely new development remains concentrated in the daytime of 11-13.
- The news side and the code side continue to diverge: gnews component queries return zero hits for consecutive periods; the only external reports with informational value this period are Horizon Robotics’ Sunrise chip embodied-tool adaptation and Enflame’s personnel changes, both of which are background. The daily briefing’s criteria remain first-hand data from commits, PRs, and the packaging matrix.
Appendix: Source Verification Table and Full Source List
Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org: flagos-ai repository list | 64 repos, 19 pushed within the window; pushed_at checked repo by repo (UTC converted to Beijing time) |
| GitHub commit search within window | 166 commits (11-13 06:00 to 11-16 06:00), classified by repo and vendor tag |
| GitHub merged PR search | 223 PRs; another 269 open PRs had updates within the window, of which 132 were newly opened |
| GitHub Releases and Tags | Atomic check of Releases in 31 repos and Tags in 8 repos; no new entries within the window; the two most recent both predate the window |
| 2.3 schedule | Read directly from community’s schedule_CN.md: FEP deadline 11-13, rc0 11-30, rc1 12-07, rc2 12-14, code freeze 12-25, GA 12-30 |
| build-infra packaging matrix | Read directly from configs.yaml, including the sunrise row (Sunrise SR-SUN-S2-X1, TangRT 1.2.0, flagtree 0.7.0+sunrise3.6) |
| Google News RSS (24 query sets) | 31 candidates after deduplication; component-name queries returned zero hits; hits concentrated in member-company stock-market pieces and popular-science articles |
| HN Algolia (FlagOS, FlagGems, FlagScale, FlagTree, BAAI, etc.) | All 4 FlagOS hits were tokenization mismatches; 0 for the rest |
| CSDN FlagOS column | Latest item on the homepage was created 11-13 11:44; the card’s original date is 2026-08-27, indicating redistribution of an older article; no new articles within the window |
| BAAI Community | Search page is client-rendered; homepage carries model and popular-science content; no FlagOS project updates |
Full Source List
- [1] Torch-FL generates native FlagGems hosts (#824) — https://github.com/flagos-ai/Torch-FL/pull/824
- [2] Torch-FL reduces shared host overhead (#839) — https://github.com/flagos-ai/Torch-FL/pull/839
- [3] Torch-FL feature request (#822) — https://github.com/flagos-ai/Torch-FL/issues/822
- [4] FlagOS Project 14 (eager FlagGems hosts) — https://github.com/orgs/flagos-ai/projects/14
- [5] Torch-FL operator host compiler design doc (#823) — https://github.com/flagos-ai/Torch-FL/pull/823
- [6] Torch-FL compiles FlagGems source and model hosts (#826) — https://github.com/flagos-ai/Torch-FL/pull/826
- [7] Torch-FL native host registry coverage (#835) — https://github.com/flagos-ai/Torch-FL/pull/835
- [8] Torch-FL schema-type eager entries (#836) — https://github.com/flagos-ai/Torch-FL/pull/836
- [9] Torch-FL source-call bridging (#837) — https://github.com/flagos-ai/Torch-FL/pull/837
- [10] Torch-FL preserves scalar semantics (#838) — https://github.com/flagos-ai/Torch-FL/pull/838
- [11] community 2.3 release schedule — https://github.com/flagos-ai/community/blob/main/release/2.3/schedule_CN.md
- [12] FlagGems v5.5.0.dev1 — https://github.com/flagos-ai/FlagGems/releases/tag/v5.5.0.dev1
- [13] FlagTree 0.7.1+xpu3.6 — https://github.com/flagos-ai/FlagTree/releases/tag/0.7.1%2Bxpu3.6
- [14] docs removes release guide and updates model list (#584) — https://github.com/flagos-ai/docs/pull/584
- [15] docs updates model list by platform (#585) — https://github.com/flagos-ai/docs/pull/585
- [16] build-infra T-Head PPU builder (#1789) — https://github.com/flagos-ai/build-infra/pull/1789
- [17] build-infra PPU runtime 2.3.0 (#1791) — https://github.com/flagos-ai/build-infra/pull/1791
- [18] build-infra MUSA 5.3.0 base (#1792) — https://github.com/flagos-ai/build-infra/pull/1792
- [19] build-infra MUSA runtime changelog (#1805) — https://github.com/flagos-ai/build-infra/pull/1805
- [20] build-infra flagcx builder and wheel (#1806) — https://github.com/flagos-ai/build-infra/pull/1806
- [21] build-infra MUSA runtime pinning (#1812) — https://github.com/flagos-ai/build-infra/pull/1812
- [22] build-infra disables container toolkit layer (#1817) — https://github.com/flagos-ai/build-infra/pull/1817
- [23] build-infra Ascend 910C SDK (#1770) — https://github.com/flagos-ai/build-infra/pull/1770
- [24] build-infra 910C runtime build (#1776) — https://github.com/flagos-ai/build-infra/pull/1776
- [25] build-infra 910C description page (#1778) — https://github.com/flagos-ai/build-infra/pull/1778
- [26] build-infra opens AMD flagtree daily row (#1780) — https://github.com/flagos-ai/build-infra/pull/1780
- [27] build-infra AMD runtime changelog (#1782) — https://github.com/flagos-ai/build-infra/pull/1782
- [28] build-infra Iluvatar application layer removes corex4.4.0 (#1771) — https://github.com/flagos-ai/build-infra/pull/1771
- [29] build-infra Iluvatar LLVM pinned to v0.7.0 (#1819) — https://github.com/flagos-ai/build-infra/pull/1819
- [30] build-infra gendoc rejects rendering on empty version set (#1796) — https://github.com/flagos-ai/build-infra/pull/1796
- [31] build-infra gendoc retry limit and failure determination (#1798 to #1801) — https://github.com/flagos-ai/build-infra/pull/1798
- [32] build-infra git-ops single-commit rebase (#1802) — https://github.com/flagos-ai/build-infra/pull/1802
- [33] build-infra libtriton_jit Fedora and openEuler RPMs (#1784, open) — https://github.com/flagos-ai/build-infra/pull/1784
- [34] build-infra sunrise precompiled input keys (#1820, open) — https://github.com/flagos-ai/build-infra/pull/1820
- [35] FlagGems pytest packages for 82 platforms (#9118) — https://github.com/flagos-ai/FlagGems/pull/9118
- [36] FlagGems PPU grouped_gemm_w8a8 and upsample_bilinear2d (#7910, #9029) — https://github.com/flagos-ai/FlagGems/pull/7910
- [37] FlagGems NVIDIA LayerNorm (#8592) — https://github.com/flagos-ai/FlagGems/pull/8592
- [38] FlagGems six int8/uint8 operators (#8874) — https://github.com/flagos-ai/FlagGems/pull/8874
- [39] FlagGems Ascend index_put (#6701) — https://github.com/flagos-ai/FlagGems/pull/6701
- [40] FlagGems Kunlunxin fixes (#8646, #8715, #8960, #8962, #8976) — https://github.com/flagos-ai/FlagGems/pull/8646
- [41] FlagGems-vllm Hygon W8A8 MegaMoE (#1020) — https://github.com/flagos-ai/FlagGems-vllm/pull/1020
- [42] FlagGems-vllm Hopper MegaPreAttention (#1007) — https://github.com/flagos-ai/FlagGems-vllm/pull/1007
- [43] FlagGems-vllm Ascend A5 KDA and causal convolution (#1204) — https://github.com/flagos-ai/FlagGems-vllm/pull/1204
- [44] FlagGems-vllm CI self-contained (#1202) — https://github.com/flagos-ai/FlagGems-vllm/pull/1202
- [45] FlagTree TLE explicit layout (#1536) — https://github.com/flagos-ai/FlagTree/pull/1536
- [46] FlagTree Thrive cumsum (#1649) — https://github.com/flagos-ai/FlagTree/pull/1649
- [47] FlagTree KV physical layout sharing (#1450) — https://github.com/flagos-ai/FlagTree/pull/1450
- [48] FlagTree nvidia3.5 build (#1650) — https://github.com/flagos-ai/FlagTree/pull/1650
- [49] FlagQuantum Hubbard graph-pair training (#1176) — https://github.com/flagos-ai/FlagQuantum/pull/1176
- [50] FlagQuantum monomer density tensor (#1179) — https://github.com/flagos-ai/FlagQuantum/pull/1179
- [51] FlagQuantum stabilizer sampling (#1180) — https://github.com/flagos-ai/FlagQuantum/pull/1180
- [52] FlagQuantum compiled DEM sampling (#1157) — https://github.com/flagos-ai/FlagQuantum/pull/1157
- [53] FlagQuantum fault-tolerant compilation integration (#1084, open) — https://github.com/flagos-ai/FlagQuantum/pull/1084
- [54] FlagQAI research metrics bound to experiment versions (#36) — https://github.com/flagos-ai/FlagQAI/pull/36
- [55] FlagQAI state-vector probability display (#40) — https://github.com/flagos-ai/FlagQAI/pull/40
- [56] FlagCX Hygon uniRunner and performance CI (#713) — https://github.com/flagos-ai/FlagCX/pull/713
- [57] FlagCX removes redundant IPC pointer path (#717) — https://github.com/flagos-ai/FlagCX/pull/717
- [58] FlagCX UIL NVIDIA templated memory operations (#710) — https://github.com/flagos-ai/FlagCX/pull/710
- [59] vllm-plugin-FL Iluvatar adaptation for vLLM 0.32 (#674) — https://github.com/flagos-ai/vllm-plugin-FL/pull/674
- [60] vllm-plugin-FL clears 0.28 paths (#742) — https://github.com/flagos-ai/vllm-plugin-FL/pull/742
- [61] vllm-plugin-FL Sunrise on 0.32 (#697, open) — https://github.com/flagos-ai/vllm-plugin-FL/pull/697
- [62] vllm-plugin-FL Enflame S60 on 0.32 (#675, open) — https://github.com/flagos-ai/vllm-plugin-FL/pull/675
- [63] sglang-plugin-FL Tsingmicro TX8110 on 0.5.18 (#187, open) — https://github.com/flagos-ai/sglang-plugin-FL/pull/187
- [64] FlagGems-sglang completes Track1 operator table (#370) — https://github.com/flagos-ai/FlagGems-sglang/pull/370
- [65] FlagGems-sglang Track1 operators (#372 to #389, open) — https://github.com/flagos-ai/FlagGems-sglang/pull/372
- [66] FlagOmni LightX2V operators and SageAttention (#1, open) — https://github.com/flagos-ai/FlagOmni/pull/1
- [67] lightx2v-plugin-FL H3 operators (#8, open) — https://github.com/flagos-ai/lightx2v-plugin-FL/pull/8
- [68] Trident guard tree (#11, open) — https://github.com/flagos-ai/Trident/pull/11
- [69] FlagAttention Ascend chunked GLA (#135, open) — https://github.com/flagos-ai/FlagAttention/pull/135
- [70] Megatron-LM-FL DeepSeek-V4 sync, 4th PR (#225, open) — https://github.com/flagos-ai/Megatron-LM-FL/pull/225
- [71] TransformerEngine-FL all-platform plugin tests (#160, open) — https://github.com/flagos-ai/TransformerEngine-FL/pull/160
- [72] optest baseline auto-update (#54) — https://github.com/flagos-ai/optest/pull/54
- [73] optest FlagAttention Enflame S60 tests (#52) — https://github.com/flagos-ai/optest/pull/52
- [74] Horizon Robotics and Infinigence AI partnership (42HOW) — https://ai.42how.com/en/article/c49aa0f4-49c0-489b-a9af-bbdc08de8ca6
- [75] Horizon Robotics and Infinigence AI partnership (Robothot) — https://www.robothot.com/d-robotics-infinigence-ai-partner-on-open-source-embodied-ai-infrastructure
- [76] Enflame co-founder and chief ecosystem officer departs (Tencent News, Southern Metropolis Daily) — https://new.qq.com/rain/a/20261113A0B55000
- [77] Enflame co-founder and chief ecosystem officer departs (Leiphone, NetEase) — https://m.163.com/news/article/LBK5J92Q05118HA4.html
- [78] MetaX joint venture closes RMB 500 million angel round (Google News) — https://news.google.com/rss/articles/CBMiZkFVX3lxTE0xMTczMlUzZWdMcG9FR2xLcEg0TU9BYjFuQVZkLXRPYWF6bUI2NVVUczdvNEc0dklXRlVlZVdoV1BGczNmMGVPNE04YlhxdWh1YS1yazZ2OEMzZmJqWjFhdXgwT25WQQ?oc=5