FlagOS Daily Intelligence Report (2026-11-18)
Research window: 2026-11-17 06:00 ~ 2026-11-18 06:00 (Beijing time, 24 hours; continuing from the cutoff of the previous daily report).
Sources: GitHub (full verification of 64 repositories under org: flagos-ai, 146 commits within the window, 201 merged PRs, 279 updated in-review PRs within the window, pushes to 20 repositories, atomic verification of Releases in 31 repositories and Tags in 8 repositories), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia, CSDN FlagOS column, BAAI Community, original-site reports from member organizations (see the appendix source list for details).
This Issue’s Index
- Today’s Focus: FlagTree 0.7.1 official line fills in T-Head and Enflame in one evening; Enflame’s two tops lines fully replaced (11-17~11-18)
- Today’s Focus: vLLM 0.32.0 application layer opens five backends on the same day; Ascend and T-Head each blocked by one upstream dependency (11-17)
- I. Open Source Project Progress (GitHub Activity)
- 1.1 Organizational Rhythm: 20 repos pushed, 146 commits, 201 merges; 20 in a single hour at 17:00 (11-17)
- 1.2 build-infra: Enflame’s two retired lines swapped for integration lines; flagcx builder adjusts dependencies overnight (11-17~11-18)
- 1.3 FlagTree: Ascend 3.6.x backend port touches 1,286 files; Enflame TritonGCU backend follows to 0.7.1 (11-17~11-18)
- 1.4 FlagGems: 55 MUSA reference alignments merged within eight minutes; 285B model profile added to benchmarks (11-17)
- 1.5 FlagGems-vllm: DeepSeek-V4.1 decode block selection and Hygon INT8 output projection land the same day (11-17)
- 1.6 FlagGems-sglang: Seventeen operator skeletons for batch12 enter the repo; 16 competition FP8 operators merged (11-17)
- 1.7 Sparse, Vision and Attention: Three NCIC kernel optimizations enter FlagSparse; FlagCV adds five-tier evaluation (11-17)
- 1.8 Quantum Line: FlagQuantum connects embedded Stim sampling; FlagQAI builds research pipeline with 27 commits (11-17)
- 1.9 Framework and Communication: Torch-FL Ascend native host pilot passes all 27 items; FlagCX adds MUSA platform (11-17)
- 1.10 Review Queue: 104 newly opened; Mega-series fused large operators and SiliconFlow six-backend batch enter review (11-17)
- II. News Coverage and Ecosystem
- 2.1 Channel Verification: 24 query groups yield 16 candidates; all component names return zero hits (11-17)
- 2.2 People’s Daily commentary places Open3D-PIMC alongside CANN and MUSA; JRJ publishes interpretation piece the same day (11-17)
- 2.3 Member Units: Moore Threads adds training-side adaptation; MetaX and Cambricon both onboard SenseNova Loop-9B the same day (11-17)
- 2.4 CSDN Column: Import timestamp is new, but the article’s original date is late August (11-17)
- III. Member Units and Backend Vendors In Depth
- 3.1 Enflame: Compiler, build, and communication lines all move on the same day (11-17)
- 3.2 Ascend: Two single-core fused large operators deliver 1.34x and 1.18x (11-17)
- 3.3 T-Head: INT8 MegaMoE completes the full expert-parallel workflow with one persistent kernel (11-17)
- 3.4 Moore Threads and MetaX: mtml floor set at 2.5.0; MetaX pauses one bare-metal CI (11-17)
- 3.5 Hygon and Kunlunxin: BW1000 reuses FP8 weights as INT8; Kunlunxin checks off operators one by one (11-17)
- IV. Summary
- Appendix: Source Verification Table and Complete Source List
Today’s Focus: FlagTree 0.7.1 stable line adds T-Head and Enflame in one night; both Enflame tops lines fully upgraded (11-17~11-18)
Date: 2026-11-17 ~ 2026-11-18 Sources: FlagTree 0.7.1+ppu3.6 · FlagTree 0.7.1+enflame3.6 · build-infra #1852 · FlagCX #712
- Two stable tags: FlagTree tagged
0.7.1+ppu3.6at 18:23 on 11-17 and0.7.1+enflame3.6at 01:16 on 11-18. Both Release notes contain only a single line,pypi 0.7.1+…, marking publication to PyPI rather than serving as changelogs. - Lineage changes: T-Head had been stuck at
0.7.1.dev1+ppu3.6(dev line), and Enflame at0.7.0+enflame3.6; the two commits together lift both backends into the 0.7.1 stable line. The 0.7.1 stable release now covers four backends: ascend3.5, xpu3.6, ppu3.6, and enflame3.6. On the pre-release side, the staged rollout pattern persists, with a1’s six backends and a2’s two backends coexisting. - Enflame upgrade: build-infra #1852 swaps two retired lines for their replacements —
enflame-tops1.9.10(Xiaoxiao C200 and S60) becomesenflame-tops1.9.29, andenflame-tops1.10.6(L600) becomesenflame-tops4.1.6. The changes touch the new deb and new driver packages in the Base Containerfile, the torch-gcu and pyefml dependencies inconfigs.yaml, the runner renames inbuild-config.yml(tops1106 for 1.9.29, tops416 for 4.1.6), and the setting of tops4.1.6 as the vendor’sdefault_for_vendorinbackends.yaml. - One symlink: The 4.1.6 image adds a new
/opt/triton_gcupointer because FlagTree’stoolkit.pylooks for the compiler at that path. The same commit corrects the topsaten version from 3.8.20261111 on the working branch to 4.1.20261111, per the Containerfile. - Communication-side interlock: FlagCX #712 adds an ECCL device API backend for Enflame (device API implementation, communication traits, platform traits, ECCL and TOPS adapters, build scripts; +5629/-41 across 21 files). The single-node cases
test_device_api_intraandtest_device_ir_intrapass; the cross-node cases are explicitly noted as “implemented but untested.” On the build-infra side, #1854 pins flagcx to 0.14.1 for both Enflame lines, and #1870 then declares the ECCL version per line, avoiding a shared ECCL number across the two tops lines. - Scheduling implications: The description of #1852 states that follow-up work is out of scope for this PR — first ship FlagCX wheels for the two backends, then move from base to runtime images, with the application layer deferred. In other words, this Enflame line switch is currently in an intermediate state: “compiler and build in place, runtime image pending.”
- Observation: The real cost of a line switch lies not in version numbers but in packages and symbols. Path conventions like
/opt/triton_gcu, which file governs the topsaten version, and per-line ECCL version declarations are all implicit contracts for “running the same code on multiple chips.” Three such adjustments landing in a single night suggests this backend line is shifting from a follower of vendor SDKs to a community-maintained build route.
Today’s Focus: vLLM 0.32.0 Application-Layer Five Backends Open the Gate on the Same Day; Ascend and T-Head Each Blocked by an Upstream Dependency (11-17)
Date: 2026-11-17 Source: build-infra #1851 · build-infra #1863 · build-infra #1872 · vllm-plugin-FL #703
- Scope of the gate opening: #1851 adds
deps_app.vllm0.32.0to the five rows nvidia-cuda13.3, amd-rocm7.2.4, ascend-cann9.2.1, iluvatar-corex5.0.1, and thead-ppu2.1.0, with the corresponding application image tags unified as2.3.0-0.5.0.dev0_g814b6d3.d20261117; #1863 separately createsstatus_matrix.vllm0.32.0.yaml, with all five backends in an unverified state, and lowers the validation script’s memory-usage parameter from 0.6 to 0.4 (insufficient node memory). - One hard dependency: The vLLM 0.32.0 line uses the 0.5.0-dev plugin, whose FlagGems backend hard-imports
flaggems_vllmwithout declaring it, so all five rows explicitly carry
flaggems_vllm==0.3.0.dev20261117—a version published to the various vendor indexes only that day. The nvidia row additionally addsnumpy==1.26.4, because the native plugin wheel’s install chain pulls numpy to 2.x, triggering the matrix integrity gate. - Ascend blocker: #1872 records the blockage in the ascend changelog: the
0.3.0.dev20261117build lacks Ascend’s V2 sampling exports, while the 0.5.0-dev plugin unconditionally installs the Ascend V2 sampling shims on import on npu, requiring the three symbolsv2_apply_penalties,v2_apply_temperature, andv2_gumbel_rng; these three exist only in FlagGems-vllm’scodex/ascend-v2-samplingbranch, i.e., the still-under-review #1234, whereas the daily wheel builds from main. The author’s conclusion is direct: this is not a problem that waiting for a newer daily package can solve; the Ascend application image must wait for #1234 to be merged. - T-Head blocker: #1863 explains that the thead row’s validation failure is because the plugin-side thead vendor adaptation layer’s attention implementation file hard-imports
flash_attn_3._C, while the thead runtime does not include that module, requiring a fix on the vllm-plugin-FL side; on the same day, #1868 first addsflash_attn_3to thead’s vllm 0.32.0 application dependencies—a two-step approach of “add the dependency first, then wait for the plugin to change the import.” - Tags already recorded: The four application image tags amd-rocm7.2.4, iluvatar-corex5.0.1, thead-ppu2.1.0, and nvidia-cuda13.3 have been written to the changelog by the bot (#1864, #1865, #1869, #1874); the ascend row has only a blockage record, no tag.
- Observation: The benefit of opening the application-layer gate is that it breaks “can it run” into a matrix that can be filled in cell by cell; the cost is that behind every cell hangs an upstream branch. Neither of today’s two blockers is a problem with FlagOS’s own code, but rather a symbol contract between the plugin and the operator library—such issues can only be explicitly recorded in the matrix, not masked by retries.
1. Open-Source Project Progress (GitHub Activity)
1.1 Organizational Cadence: 20 Repos Pushed, 146 Commits, 201 Merges, 20 Commits in the 17:00 Hour (11-17)
Date: 2026-11-17 Source: flagos-ai organization · FlagGems · FlagGems-vllm
- Totals: 20 of 64 repos had pushes within the window; commit search returned 146 hits, 201 merged PRs, 279 open PRs updated within the window (104 newly opened that day), and 100 organization events (66 PushEvent, 13 CreateEvent, 10 DeleteEvent, 4 PR-related).
- Commit distribution: build-infra 36, FlagQAI 27, FlagGems-sglang 24, FlagGems-vllm 17, FlagGems 16, FlagQuantum 12, FlagSparse 4, FlagTree 2, FlagAttention 2, and 1 each for optest, vllm-plugin-FL, Trident, FlagCX, Torch-FL, and FlagTrain.
- Merge distribution: FlagGems 71, build-infra 36, FlagQAI 27, FlagGems-sglang 22, FlagGems-vllm 17, FlagQuantum 12, vllm-plugin-FL 4, FlagTree 2, FlagAttention 2, and 1 each for the remaining eight repos (optest, Megatron-LM-FL, Trident, FlagTrain, sglang-plugin-FL, Torch-FL, FlagSparse, FlagCX).
- Time distribution: By commit timestamp, 17:00 peaked at 20 commits, 11:00 and 13:00 had 14 each, and 14:00 had 13; 5 commits landed after midnight (four acceptance fixes and late-night wrap-up in build-infra), indicating the build line kept moving overnight.
- New and retired repos: No new repos in the window; the ten
DeleteEvents were scattered across temporary branch cleanup in Trident, FlagQuantum, build-infra, and FlagGems, a normal wind-down.
1.2 build-infra: Two Enflame Retirement Lines Swapped for Handoff Lines, flagcx Builder Tunes Dependencies Overnight (11-17~11-18)
Date: 2026-11-17 ~ 2026-11-18 Source: build-infra #1852 · build-infra #1854 · build-infra #1870 · build-infra #1878
- Shape: 36 merges, close to a fifth of the whole organization’s total that day, all falling into three buckets—Enflame line swap (#1852, #1854, #1855, #1870, #1873, #1875), vLLM 0.32.0 application layer (#1850, #1851, #1863, #1864, #1865, #1868, #1869, #1872, #1874), and builder and documentation fixes.
- Swap details: See the first item in today’s highlights; worth noting is that #1852 also deleted the changelogs and generated description pages for four old lines and added four changelogs pending gating—an intermediate state of “old pages offline, new pages to be filled.”
- flagcx builder: Four consecutive fixes on the Enflame row—#1859 installs torch and torch-gcu separately with
--no-deps, #1860 pins setuptools back to 69.5.1 before installing torch, #1862 drops--preand merges the two into a single call, and #1873 pins that row’s own triton version into the builder. Together these show one thing: Enflame 4.1.6’s dependency closure does not match upstream PyPI’s default resolution and must be explicitly pinned item by item. - Acceptance fixes: Three late-night commits (#1877, #1878, #1880) all fix “the verification step misreporting itself”—inaccurate application image verification output, flagcx wheels should be verified in the builder rather than at runtime, and step 2’s payload truncated by double quotes. Such fixes don’t change artifacts but determine whether the matrix reads true values or illusions.
- Release records: The bot logged two wheel releases for vllm-plugin and vllm-plugin-native (#1848, #1849) and the flagcx wheel release (#1876); the Moore Threads musa row pointed to the musa530 runner at 08:25 the same day (#1838).
- Moore Threads toolkit layer: #1841 re-declares the mthreads toolkit layer and sets a floor of
mtml >= 2.5.0. The root cause is specific—the host’s mtml 2.4.2 has only a barelibmtml.so, while the runtime looks forlibmtml_musa.so; not finding it isn’t fatal, and the container starts with an empty device table, looking as if the image is broken; mtml 2.5.0 adds that file, and the same toolkit 2.1.0 can then inject eight devices. Verification record: on worker33024 (driver 5.3.0, mtml 2.5.0, toolkit 2.1.0),--runtime mthreadsyields 8/dev/mtgpu*,mthreads-gmireports 8 MTT S5000s,torch.musa.device_count()is 8, matmul runs, andMTHREADS_VISIBLE_DEVICES=0,1narrows to 2. #1845 subsequently scopes this floor to driver 5.3.0.
1.3 FlagTree: Ascend 3.6.x Backend Port Touches 1,286 Files, Enflame TritonGCU Backend Follows to 0.7.1 (11-17~11-18)
Date: 2026-11-17 ~ 2026-11-18 Source: FlagTree #1637 · FlagTree #1619 · FlagTree #1675 · FlagTree #1676
- Ascend port: #1637 migrates the entire Ascend NPU backend to Triton 3.6.x, +307349/-294 across 1,286 files, including the compiler, driver, runtime, and interpreter layers plus supporting pieces such as hint handler, dtype guard, op dtype config, and npu_utils; the compiler side carries a dozen-plus passes including SplitDataflow, ComputeBlockOpt, DynamicCVPipeline, TritonToLinalg, TritonToGraph, PlanComputeBlock, and AllocMultiCache, and the test side adds 200+ operator-level unit tests and UB tuner cases.
- Enflame backend: #1619 updates the TritonGCU backend, +20069/-16757 across 114 files—aligning lib and include for triton_gcu300 and 400, backend and TLE tests, removing legacy gcu_intrinsics, adding Combine38.td, restoring TLE-Raw’s tops runtime, efshmem, and
test_tle_*cases, and switching the build side to multi-Python-version MLIR staging. This commit and the early-morning0.7.1+enflame3.6tag are two steps of the same line. - Toolchain under review: #1675 proposes unified configuration and environment switches (centralized parameters in
config.env, switchable PyTorch Profiler and Eager modes), refactors shell scripts into a two-layer “main flow / abstract methods” structure, and addsgetshape.pyaggregating CPU/GPU time by operator, shape, and dtype (+1503/-0, 14 files). - Other open items: #1676 splits Iluvatar’s SME row-stride support out of NVIDIA async copy; #1679 adds a tutorial for TLE fused scaled matmul-plus-reduce-scatter; #1673 enables the FlagGems-vllm v0.2.0 baseline on hcu3.6; #1677 leaves the MUSA debugger dependency to the runtime; #1678 bumps git-auto-commit-action from v4 to v7.
- Community governance: community #168 (FEP-0168, FlagTree multi-platform support and Triton API compatibility) still had updates at 11-17 17:34, and its corresponding branch
codex/fep23-flagtree-device-enablement-20261105was pushed multiple times within the window.
1.4 FlagGems: 55 MUSA Reference Alignments Merged in Eight Minutes, 285B Model Profile Enters Benchmarks (11-17)
Date: 2026-11-17 Source: FlagGems #9334 · FlagGems #4916 · FlagGems #8526 · FlagGems #9160
- Scale and concentration: 55 of the 71 merges were landed by lukatao within eight minutes from 11:06 to 11:14, all MUSA-side test reference alignments; the author uses “CPU reference” as the unified approach—
linalg_inv_ex,linalg_tensorinv, the Cholesky series,slogdet,geqrf,orgqr(Householder reference),determinant,solve_ex, and others are filled in batch across the linear algebra domain, while quantization and convolution handle FP8 comparison,quantized_max_pool3doutput fixtures, and the unsupported case forconv_transpose3d. - Model profile enters benchmarks: #9334 turns the operator profiling results of a 285B model into a benchmark shape profile (
285B.yaml), covering 11 registered operators (mm, bmm, add, copy_, fill_scalar, mul, zeros, rms_norm, silu, silu_and_mul, softmax), each with prefill and decode GPU time share and call counts, plus a new--model-profileswitch and a “Time Share (prefill/decode)” column in results; data comes from a profiling report collected on 11-16 on 8× H100 80GB with TP=8. - Operators and backends: #8526 refreshes GEMM test shapes and layout coverage for T-Head PPU; #9160 adds the Triton autograd backward for
log10; #7730 and #7401 add NVIDIA-side_cast_Floatand_sparse_coo_tensor_with_dims; #9067 fixes launch-parameter mismatch betweenaddmmandact_quanton cuda; #9335 fixes W8A8 INT8 following MetaX’s updated TLE capability API; #9278 optimizes Ascend’s mul scalar and contiguous loads. - Performance and CI: #4916 optimizes LayerNorm backward at large row counts; #8788, #9084, #9294 fix Kunlunxin’s nested tensors,
masked_selectbackward, andmax_pool3dprecision, and addthnn_differentiable_lstm_cell_backward,row_stack, andpadded_dense_to_jagged; #9340 moves the daily pipeline to 01:30 CST; #9341 renamesmodel_shapestomodels_shapes. - Observation: Within a single day, “test baselines” and “real model shapes” advanced together, showing that this library’s acceptance criteria are shifting from “is the operator semantics correct” to “is it worth optimizing given its time share under real workloads”—that is precisely the significance of the 285B profile.
1.5 FlagGems-vllm: DeepSeek-V4.1 Decode Block Selection and Hygon INT8 Output Projection Land the Same Day (11-17)
Date: 2026-11-17 Source: FlagGems-vllm #1197 · FlagGems-vllm #1192 · FlagGems-vllm #1177 · FlagGems-vllm #1280
- Decode block selection: #1197 adds the candidate block selection operator
select_candidate_blocks_cooperative, porting vLLM’s DeepSeek-V4.1 decode candidate block selection into a Triton implementation (+1175, 7 files)—each row scores by intra-block maximum (NaN-propagating), pins the block containing the latest token, takes top-k within the valid prefix, pads with -1, and does not sort. Layering matches upstream: two Triton kernels (scoring, finalization) are ported directly, row-prefix top-k preferentially reuses this repo’spersistent_topk(the author notes it is the faster Triton backend on NVIDIA currently), falling back tocooperative_topk; supported shapes are fp32 logits, no more than 64 rows, k of 512/1024/2048, and block counts divisible by 4; out-of-bounds paths throw directly with no torch fallback, with 35 functional tests and benchmarks. - Hygon: #1192 adds
dsv41_o_projfor BW1000—routing FP8wo_athrough an INT8, in-place BF16 path, i.e., changing the algorithm on existing FP8 weights rather than changing precision. - Fusion and communication: #1177 adds TLE fused matmul-plus-reduce-scatter, dispatched by
symm_mem; #1282 and #1227 add fused operators for DeepSeek-V4’s qnorm, rope, and kv-rope full-cache FP8 insertion (multi-platform). - Per-backend fixes: #1239 changes Ascend TLE’s
mhc_preto mix-major layout and cleans up tests; #1250 adds min-p filtering and adjusts Hopper’s TLE layout; #1252 addsapply_thinking_budget(fused marker cache update); #1254 fixes kernel launch-parameter mismatch on Triton 3.7 and above; #1237 adds FP8-state GDN decode and prefill for Moore Threads; #1280 optimizescooperative_topkfor T-Head PPU-ZW810E; #1222 adds themhc_pre_broadcastoperator; #1265 adds 285B profiling shapes and vLLM API validation for 8 operators; #1268 and #1276 add model-profile options and documentation; #1277 pauses the metax-maca3720 bare-run CI; #1281 skips thinking_budget collection when the vLLM API is missing. - Observation: What these merges share is “taking the fast path by model shape”—decode block selection tiers by row count, o_proj switches algorithms by weight precision, and thinking budget fuses by cache layout. Cross-silicon-stack differentiation is shifting from “supported or not” to “which path the same model takes on our card.”
1.6 FlagGems-sglang: Seventeen Operator Skeletons from batch12 Land, 16 Competition FP8 Operators Merged (11-17)
Date: 2026-11-17 Source: FlagGems-sglang #392 · FlagGems-sglang #390 · FlagGems-sglang #371 · FlagGems-sglang #385
- Skeletons land: #392 migrates batch12’s 17 operators from
sglang-ops’docs/tasks/batch12into this repo’s conventions, each with a reference implementation, a correctness test, a benchmark, and aconf/operators.yamlentry, 52 files total. Coverage spans dsa (7, includingact_quant,dequantize_k_cache_fast,quantize_k_cache_fast,triton_sparse_mla_fwd), kvcache (3), attention (2), quantization (2), moe (2), and speculative (1). The author explicitly notes this commit excludessrcimplementations: before Triton implementations land,get_candidateraises ImportError andget_opreturns None, with tests and benchmarks skipped; it also removes the software FP8 shims for Ascend, Kunlunxin, and Enflame from the original repo. - Competition channel: Between 17:37 and 18:05, 16 competition PRs were merged back-to-back—12 from c2flowDS (absorbed BMM concat conversion, activation quantization, paged dequantize k cache, gather-dequant-requant FP8 paging, MLA KV packed quantization, per-token-group quantized MLA mask, quantize k cache fast and slow versions, reshape-and-cache-flash, set MLA KV buffer FP8 quantization, swiglu OAI MXFP8 quantization, sparse MLA forward), 4 from yzw1128 (dequantize k cache, FP8 per-token to per-tensor, per-tensor quantized MLA, quantize k cache split version), plus one task180 competition submission (#372).
- Three AMD performance items: #371, #390, and #391 speed up gfx1100—
fused_softmax_pool,per_group_transpose,ltx2_split_rotary,dsv3_fused_a_gemm,embedding_lora_a, theapply_token_*series, and MoE GEMM extended with silu and a shrink-recompute path. - Enflame merge: #385 merges the
enflame_flagosbranch, in the same direction as the same-day Enflame line swap in build-infra.
1.7 Sparse, Vision, and Attention: Three NCIC Kernel Optimizations Enter FlagSparse, FlagCV Adds Five Evaluation Tiers (11-17)
Date: 2026-11-17 Source: FlagSparse #121 · FlagAttention #142 · FlagTrain #89 · FlagDNN #26
- FlagSparse: #121 merged by the NCIC-AlphaSparse account (+3371/-39, 21 files), with commits
cuad 105 opt,cuda sddmm opt, andascend spgemm opt—targeting sampling and sparse matmul on CUDA, and sparse general matmul on Ascend, respectively. - FlagCV: lukatao opened five KernelGen evaluation profiles in a row (bitwise_and, bitwise_not, flip, subtract, multiply), all evaluation configurations rather than operator implementations.
- FlagAttention: #142 restructures the repo (merged 11-18 00:03), and #144 sets up NVIDIA-side benchmark dependencies; two open items have clear direction—#141 adds a MoBA backend for Ascend, and #143 adds Inkling FA4 relative-position attention support for T-Head.
- FlagTrain: #89 aligns
bias_geglu_implwith the native two-parameter interface; five more KernelGen items are under review (Transformer Engine’s L2 norm backward, Megatron’s MTP loss auto-scaling, multimodal rotary position embedding, weighted SwiGLU, fused Sinkhorn), plus one Ascendmulti_tensor_lionandmulti_tensor_adam. - FlagDNN: #26 adds RPM packaging for Fedora and openEuler (under review).
1.8 Quantum Line: FlagQuantum Connects Embedded Stim Sampling, FlagQAI’s 27 Commits Build a Research Pipeline (11-17)
Date: 2026-11-17 Source: FlagQuantum #1190 · FlagQuantum #1195 · FlagQAI #96
- FlagQuantum scale: 12 merges and 12 commits, advancing continuously from 10:32 to 19:58, along two main lines—compiler module reorganization by responsibility (#1192 and #1193 split out canonical rewriting and logical scheduling), and error correction and simulation.
- Error correction and simulation: #1190 packages Sparse Blossom batch decoding, #1197 switches the batch pipeline to native Sparse Blossom, and #1201 batches AI pre-decoder evaluation; #1195 uses embedded Stim for sampling with position-dependent noise, and #1199 runs external Stim text experiments on the embedded engine; #1194, #1196, #1198, #1200, and #1202 cover offline qfunc export and deployment packaging, preflight and semantics of synthetic Bell artifacts, targeted Bell artifacts, and deployment preflight for bound artifacts, respectively.
- FlagQAI: All 27 merges were advanced by the FlagQuantum account from 10:31 to 21:20, themed “research pipelining”—Agent transcript reading-position restore, unread reply markers, numeric-only experiment result display, Python completion output, circuit source line display, Agent code preview, comparison slots for QAOA results, active-round elapsed time, project baselines managed by Jobs, candidate revision tiering and review, local validation of reviewed candidates (#96), and KMX circuit reset-slot reuse.
- Observation: The quantum line is turning “Agent generation → human review → local validation” into a stateful workflow, a different cadence from the operator library iteration on the compute stack: the former wins by closing the functional loop, the latter by coverage and performance.
1.9 Frameworks and Communication: Torch-FL Ascend Native Host Pilot Passes All 27, FlagCX Adds MUSA Platform (11-17)
Date: 2026-11-17 Source: Torch-FL #846 · FlagCX #712 · FlagCX #722 · sglang-plugin-FL #155
- Torch-FL Ascend host: #846 adds experimental Ascend adaptation (+2232/-98, 35 files), using the shared native host ABI with the released FlagTree NPU loader and stock FlagGems kernels. The difficulty is in the ABI—Ascend uses packed structs rather than CUDA’s pointer arrays: booleans occupy 32 bits, FP16/BF16 scalars use float wire format, logical grids and workspace and locks and FFTS all depend on compiler metadata, plus ACL current-stream ownership issues. In testing, all 27 consistency cases passed with the wheel installed on 910C (29.90 seconds, zero skips, zero errors); the CI job on
85b1fd8efailed due to exhausted retries downloading FlagTree and needs a rerun; full-model logits and performance no lower than 0.8× the vendor PyTorch remain tracked in #833. - FlagCX Enflame: #712’s ECCL device API has been merged; see the first item in today’s highlights.
- FlagCX MUSA: #722 is under review, adding MUSA platform support (default backend) to the FlagCX device API.
- sglang-plugin-FL: #155 upgrades the Kunlunxin backend to SGLang 0.5.18.
- vllm-plugin-FL: Four merges—#703 adapts to vLLM 0.32’s Runner V2 and enables full CI (Ascend side), #694 fixes Enflame’s M-RoPE interleaved segment mask to rotary_dim/2 and supports non-block FP8 checkpoints, #758 replaces the end-to-end smoke test with an adaptation gate, and #769 cleans up stale references in
cuda.yaml. - Other: Megatron-LM-FL #221 adds a nightly supplemental coverage suite (covering megatron/core and megatron/plugin); Trident #6 supports user-supplied LLVM_ROOT; optest #49 upgrades the backend FlagTree daily wheel for operator tests.
1.10 Review Queue: 104 Newly Opened, Mega-Series Fused Large Operators and SiliconFlow Six-Backend Batch Enter Review (11-17)
Date: 2026-11-17 Source: FlagGems-vllm #1286 · FlagGems-vllm #1287 · FlagGems-vllm #1274 · FlagGems-vllm #1283
- Totals: 279 open PRs updated within the window, 104 newly opened that day; by distribution, FlagGems 127, FlagGems-vllm 44, FlagTree 23, vllm-plugin-FL 17, FlagTrain 17, sglang-plugin-FL 10, FlagCX 10, FlagCV 5, FlagQuantum 4, with the rest scattered.
- Mega series: Two single-kernel fused large operators for Ascend (MegaGate, MegaMHC; see 3.2) and one INT8 MegaMoE for T-Head (see 3.3); all three share the approach of folding multiple stages of “routing, gating, expert compute, normalization, communication” into one persistent kernel, reusing launches and workspace via pre-bound context.
- SiliconFlow batch: winfan1314 opened seven PRs in a row with the
[SiliconFlow]prefix—triton_ihc_precovering six backends while preserving the originalihc_preinterface, grouped Gemma RMSNorm reused across row counts, cold-compile time compression for Ascend and decode/prefill paths, and reduced CPU reference contention forgate_sigmoid_muland packed decode. This batch’s hallmark is treating “CPU reference contention in tests and benchmarks” as a performance problem rather than test noise. - Operator gaps: Under review on the FlagGems side are T-Head
conv2d, Ascendrouter_gemm, Iluvatar SDPA forward, T-Head replacing the removed TLE pipeline iterator withtl.range, Hygon reducing BLAS launch overhead without CUDA Graph, and Kunlunxin routing the SDPA math path through two-stage Triton kernels. - Platform side: vllm-plugin-FL #766 brings Tsingmicro onto vLLM 0.32’s 0.5.0-dev line; FlagBLAS #162 stores H100 grouped GEMM baselines for T-Head and Hygon; flagtree-cpu #12 and #13 port the TLE minimal subset to Triton 3.7 and add CPU libdevice hypot and atan2.
II. News Coverage and Ecosystem
2.1 Channel Verification: 24 Query Sets, 16 Candidates, Zero Hits on All Component Names (11-17)
Date: 2026-11-17 Sources: Google News RSS · Hacker News · CSDN FlagOS Column
- Queries and hits: 24 Chinese and English queries yielded 16 deduplicated candidates within the 24-hour window. Queries for FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen, BAAI FlagOS, flagos-ai, and domestic compute open-source software stacks all returned zero hits; the hits were 7 items related to BAAI, 3 on Enflame, 2 on Moore Threads, 2 on MetaX, 1 on Tsingmicro, and 1 on Hygon operators.
- Exclusion criteria: Three categories of noise were removed based on title semantics—stock market and sector coverage (e.g., “compute chip concept fluctuates and retreats”), same-word false hits unrelated to BAAI (the “源” in “source-side DC distributed solar-storage-grid intelligent control system”), and model vendors’ own release announcements. Usable items after exclusion were merged into 2.2 and 2.3.
- HN Algolia: FlagOS returned 8 hits, FlagGems 2, and BAAI 30, but neither titles nor body text directly matched this project—these were broad keyword hits.
- Conclusion: News coverage remains dominated by member organizations’ capital and product news; FlagOS’s own technical narrative continues to be carried by GitHub and the column.
2.2 People’s Daily Commentary Places Open3D-PIMC Alongside CANN and MUSA; JRJ Publishes Analysis Same Day (11-17)
Date: 2026-11-17 Sources: JRJ analysis · Xinhua Open3D-PIMC report (background)
- Commentary framing: A recent People’s Daily commentary on bringing AI to small and medium-sized enterprises placed three domestic hardware-software technology ecosystems side by side—CANN, Open3D-PIMC, and MUSA—arguing that with hardware breakthroughs such as compute-in-memory and 3D integration, plus continued open-source software ecosystem development, the barrier to AI adoption will keep falling.
- Analysis content: JRJ’s 11-17 10:44 article clarified the relationship among the three—CANN belongs to Huawei, MUSA to Moore Threads, while Open3D-PIMC is the world’s first programming model and software framework for 3D compute chips, jointly released by Tsingmicro and BAAI and open-sourced in September 2026 at the China Computing Conference; it is also one of FlagOS’s primary solutions for next-generation 3D chips. The article also gave FlagOS’s scale figures: adapted to 20 vendors and 30-plus AI chips, covering GPU, NPU, DSA, RISC-V and other architectures.
- Technical logic: Open3D-PIMC folds compilation optimization into the framework; developers only describe data distribution and task partitioning, and each vendor’s private instruction set is integrated as a plugin. Men Chunlei, head of AI systems research at BAAI, judges that with 3D chips nearing mass production, lacking matching compilation capability would lead to “advanced hardware, inefficient software.”
- Caveat: Open3D-PIMC’s open-sourcing itself dates back to September; this item is re-exposure driven by the commentary and analysis, not a new release event. The “20 vendors, 30-plus chips” figure is the reporting framing and does not share the same denominator as component-level statistics (the number of vendors and models covered by operator libraries).
2.3 Member Organizations: Moore Threads Adds Training-Side Adaptation; MetaX and Cambricon Both Support SenseNova Loop-9B Same Day (11-17)
Date: 2026-11-17 Sources: Yicai (Moore Threads) · Yicai (MetaX) · NBD (Cambricon) · SenseNova-Loop repo
- Moore Threads: After Moonshot AI open-sourced the 2.8-trillion-parameter Kimi K3.1 Pro, Moore Threads completed training adaptation based on MTT S5000 and the MUSA software stack, building a native FSDP2 multimodal training path and connecting the full chain from model loading and sharding scheduling to gradient synchronization (Yicai, 11-17 19:24). Note the distinction: Day-0 inference adaptation for the same model was announced on 11-06; today’s addition is the training side.
- MetaX: SenseTime’s SenseNova Loop-9B was released and open-sourced on 11-17; MetaX announced Day 0 adaptation and claimed to be the first domestic compute platform to complete adaptation for the model (Yicai, 11-17 22:44).
- Cambricon: Announced Day 0 adaptation for the same model at 23:44 the same day (NBD).
- Model-side facts: SenseNova-Loop-9B has roughly 9B parameters, using 24 decoder layers executed twice in a loop to expand 32 physical layers into 56 layer computations in a single forward pass; it uses dual-stream mHC for residual-stream routing and fusion, alternating linear and full attention in groups of four layers, and a 512K context. The official claim is that 256K multimodal inference requires about 15GiB of VRAM with 4-bit weights plus 8-bit KV cache, and an AWQ 4-bit quantized version was offered on day one.
- Comparative observation: Hyper-connection structures like dual-stream mHC belong to the same structural class as the
mhc_pre,mhc_postoperator work FlagGems-vllm is currently doing and Ascend’s MegaMHC; whether the two sides have integrated is not stated in repos or public reports—this is noted only as an architectural trend comparison, not an adaptation claim.
2.4 CSDN Column: Import Timestamps Are New, but Article Dates Are Late August (11-17)
Date: 2026-11-17 Source: CSDN FlagOS Column
- Page status: The column homepage has 43 article links, with timestamps including recent values such as 11-17 20:08; the latest three are “FlagOS Runs the ‘Niulai’ Model on More Domestic Chips: GLM-5.3-Flash Day0 Adaptation for 9 AI Chips,” “Qwen3.8-Flash-Next Adapted for 8 AI Chips on Launch Day,” and “Video Generation Inference Framework LightX2V Officially Integrates with FlagOS.”
- Assessment: The original dates on the three cards are 2026-08-27 to 08-28—these are historical articles reordered and republished by topic and are not counted as same-day developments. The column’s recent practice is consistent: import time and original publication time are not synchronized, and when the two conflict, the original publication time governs.
III. Deep Dive into Member Units and Backend Vendors
3.1 Enflame: Compiler, Build, and Communication All Move on the Same Day (11-17)
Date: 2026-11-17 Sources: FlagTree 0.7.1+enflame3.6 · build-infra #1852 · FlagCX #712 · vllm-plugin-FL #694
- Three things: The
0.7.1+enflame3.6tag was cut at 01:16; starting at 18:08, build-infra switched two tops lines to the handoff line; at 10:26 FlagCX merged the ECCL device API. These three correspond to compiler, build, and communication respectively—three layers, top to bottom, of the same backend line. - Supporting details: On the flag side,
packaging/flagcx/backends.yamlsets tops4.1.6 as the vendor default, with ECCL versions declared per line; vllm-plugin-FL #694, submitted by an Enflame engineer, fixes the M-RoPE interleaved-segment mask and non-block FP8 checkpoints, showing the application side is keeping pace. - Risk surface: Two tops lines run in parallel (1.9.29 for Xiaoxiao C200 and S60, 4.1.6 for L600), meaning the build matrix has twice the combinations; #1852 explicitly defers the application layer, first getting the FlagCX wheel and base-to-runtime images working. This order is correct—application-layer validation is meaningless when the communication library doesn’t work.
3.2 Ascend: Two Single-Core Fused Mega-Operators Deliver 1.34x and 1.18x (11-17)
Date: 2026-11-17 Sources: FlagGems-vllm #1286 · FlagGems-vllm #1287 · FlagGems-vllm #1239 · FlagTree #1637
- MegaGate: #1286 adds a FlagTree TLE single core for 910B, fusing BF16 MoE routing projection with expert selection—the Cube core computes FP32-accumulated routing logits, the Vector core does scoring, optional bias correction, and group selection, returning top-k expert IDs and normalized routing weights;
prepare_mega_gatebinds the launcher and workspace to input shapes, routing config, and NPU stream, then reuses them on each call. Across 5 models and 353 shapes, average speedup is 1.336x over ACL Graph and 1.408x over Eager, with DeepSeek V4 Flash’s 106 shapes at 1.553x. - MegaMHC: #1287 fuses mHC post, mHC pre, and RMSNorm into a single core, supporting four residual streams and both normal and shifted execution; the shifted path uses the carried
pre_mixto launch RMSNorm while the Cube computes the next projection. Across 63 shapes, average speedup is 1.175x over ACL Graph and 1.182x over Eager with pre-bound context, with DeepSeek V4 Pro at 1.773x for a single shape; Eager without pre-binding is only 0.648x, showing the gain comes from launch and workspace reuse rather than pure kernel arithmetic. - Supporting: #1239 switches
mhc_preto a mix-major layout and cleans up tests; #1637’s Triton 3.6.x backend port provides the compiler foundation for this class of TLE single cores. - Observation: Both changes focus on “single-core fusion + context pre-binding” and both include ACL Graph and Eager numbers. For Ascend, the gap between graph mode and eager mode is itself a product-line signal—the same kernel differs by over 20% between the two modes, so the deployment side’s choice of execution mode directly affects the payoff.
3.3 T-Head: INT8 MegaMoE Runs the Entire Expert-Parallel Pipeline in One Persistent Kernel (11-17)
Date: 2026-11-17 Sources: FlagGems-vllm #1274 · FlagGems #8526 · FlagGems-vllm #1280 · FlagTree 0.7.1+ppu3.6
- Single-core scope: #1274 adds INT8 expert-parallel MegaMoE for the PPU backend (+4691/-8, 10 files), completing token dispatch, two expert GEMMs, SwiGLU, INT8 requantization, cross-card writeback, and merging within one resident FlagTree/Triton kernel; inputs are INT8 activations and expert weights plus FP32 scales and routing weights, output is on-card BF16. The caller provides a reusable context, symmetric peer buffers, and stream barriers.
- Scheduling strategy: Eight-card workloads use whole-block dispatch; large-scale two-card and four-card workloads use partial return when every card satisfies the fast path. The EP group agrees on the return layout during context initialization, before CUDA Graph capture; when workspace alignment is inconsistent, the whole group falls back to per-route return, preventing cards from interpreting the same peer slot differently.
- Same-day supporting changes: #8526 refreshes T-Head’s GEMM test shapes and layout coverage; #1280 optimizes
cooperative_topkfor PPU-ZW810E; the early-morning0.7.1+ppu3.6tag promotes the backend from the development line to the release line. - Observation: Packing dispatch, two GEMM stages, activation, quantization, writeback, and merging into one kernel saves a large amount of intermediate materialization and launch overhead, at the cost of requiring the caller to guarantee consistency of context state and peer buffers—which is why this change repeatedly emphasizes “layout aligned before capture.”
3.4 Moore Threads and MetaX: mtml Floor Set at 2.5.0, MetaX Pauses a Bare-Metal CI (11-17)
Date: 2026-11-17 Sources: build-infra #1841 · build-infra #1845 · FlagGems #9335 · FlagGems-vllm #1277
- Moore Threads: mtml 2.5.0 becomes the minimum requirement at the toolkit layer, with eight cards filled in one pass (see 1.2); the musa line is redirected to the musa530 runner; FlagGems-vllm #1237 adds FP8-state GDN decode and prefill.
- MetaX: #1277 pauses the metax-maca3720 bare-metal CI test, a convergence on unstable nodes; FlagGems #9335 fixes W8A8 INT8 imports against MetaX’s updated TLE capability API; also under review are the MetaX backend for ple_conv (#1279) and pure-Triton block-level top-k (#1289).
- Observation: Both vendors make explicit declarations about “environment floors”—one on the mtml version, the other on node trustworthiness. Much of the daily friction across chip stacks comes from this kind of environment difference; writing the floor into the build config is cheaper than troubleshooting after the fact.
3.5 Hygon and Kunlunxin: BW1000 Reuses FP8 Weights via INT8, Kunlunxin Clears Operators One by One (11-17)
Date: 2026-11-17 Sources: FlagGems-vllm #1192 · FlagGems #8788 · FlagGems #9294 · sglang-plugin-FL #155
- Hygon: #1192’s
dsv41_o_projruns INT8 on BW1000 for FP8’swo_awith in-place BF16 output, effectively switching the compute path without changing the weight format; #9338 under review attempts to reduce BLAS launch overhead without using CUDA Graph. - Kunlunxin: Three fixes merged in the evening—#8788 fixes XPU precision for
nested_tensor_from_mask,masked_selectbackward, andmax_pool3d, #9084 addsthnn_differentiable_lstm_cell_backward, #9294 addsrow_stack,padded_dense_to_jagged, and Sobol engine coverage; the plugin side is upgraded to SGLang 0.5.18; #9330 under review submits five XPU operator precision fixes and matmul backward performance in one go, and #9325 routes the SDPA math path through two-stage Triton kernels. - Observation: The rhythm of this Kunlunxin line is “clearing operators one by one”—three to five per day, each covering one operator family; this work produces no headlines, but it determines whether models can actually run.
IV. Summary
- Version line: The FlagTree 0.7.1 release line filled in T-Head and Enflame in one night (
0.7.1+ppu3.6at 18:23,0.7.1+enflame3.6at 01:16 the next day), with the release version covering four backends: ascend3.5, xpu3.6, ppu3.6, and enflame3.6; Enflame’s two tops lines switched generations on the same day, withtops4.1.6becoming the vendor default. - Build line: vLLM 0.32.0’s application layer opened for five backends simultaneously, with four backends already recording image tags; Ascend is stalled due to missing V2 sampling symbols (FlagGems-vllm #1234 under review), and T-Head is stalled due to the plugin hard-importing
flash_attn_3; neither is a problem in FlagOS’s own code. - Operator line: Of FlagGems’ 71 merges, 55 are MUSA test reference alignments, and the 285B model profile enters the benchmark; FlagGems-vllm’s 17 cover DeepSeek-V4.1 decode block selection, Hygon INT8 output projection, and various backend fixes; FlagGems-sglang turns batch12’s seventeen operators into a skeleton and merges 16 competition FP8 operators.
- Frontend line: Torch-FL’s Ascend native host pilot passes all 27 consistency cases on 910C, but CI needs a rerun due to a FlagTree download failure; FlagCX adds the Enflame ECCL device API, with MUSA platform support under review.
- Ecosystem side: The news front continues to be supplied by member units’ product and capital announcements—Moore Threads adds training-side adaptation for Kimi K3.1 Pro, and MetaX and Cambricon both pick up SenseTime’s SenseNova Loop-9B on the same day; a People’s Daily commentary lists Open3D-PIMC alongside CANN and MUSA, giving FlagOS’s co-build路线 another mention in mainstream media.
- Risks and to-dos: The Ascend application image awaits #1234, T-Head awaits the plugin import change, the Enflame application layer awaits the FlagCX wheel and runtime image, and Torch-FL’s CI needs a rerun; all four are on others’ branches or subsequent builds, and the daily report records them as “blockers accounted for, awaiting downstream.”
Appendix: Source Verification Table and Full Source List
Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org (flagos-ai) | Full check of 64 repos; 20 repos pushed within the window; 146 commits, 201 merges, 279 in-review updates, 104 newly opened |
| Release / Tag atomic check | Release scan of 31 repos, Tag scan of 8 repos; two new official tags within the window (0.7.1+ppu3.6, 0.7.1+enflame3.6) |
| Google News RSS (24 query sets, via proxy) | 16 candidates; zero hits on all component names; usable items were member-company product and capital news |
| HN Algolia | 8 hits for FlagOS, 2 for FlagGems, 30 for BAAI; no direct title matches |
| CSDN FlagOS column | Import timestamps of 11-17 present; article original dates are 2026-08-27 to 08-28, indicating redistribution |
| BAAI community hub | Homepage and search pages accessible; no FlagOS component entries found |
| Member-company original sites | All four product-side items (Moore Threads, MetaX, Cambricon, Tsinghua Unigroup/清微) were checked against the original sites for dates and key points |
| People’s Daily commentary and interpretation | The 11-17 Jrj article fully quotes the commentary’s wording; Open3D-PIMC background verified against Xinhua’s September report |
Limitations: The GitHub side of this issue was cross-checked across three channels—commit search, merged-PR search, and org events; merge times of in-review PRs are not counted. The descriptions for the two FlagTree Releases are PyPI markers only and contain no changelog; version contents are inferred from the corresponding PRs (#1637, #1619) and the tag lineage. Zero hits on component names on the news side is normal and should not be taken to imply “no news means no progress.” The CSDN column’s import times do not match the original article dates; where they conflict, the original publication time is used. The comparison between model architectures and operator libraries in today’s highlights is a trend observation only and does not mean the two sides have completed integration.
Full Source List
- [1] FlagTree 0.7.1+ppu3.6 (Release) — https://github.com/flagos-ai/FlagTree/releases/tag/0.7.1%2Bppu3.6
- [2] FlagTree 0.7.1+enflame3.6 (Release) — https://github.com/flagos-ai/FlagTree/releases/tag/0.7.1%2Benflame3.6
- [3] build-infra #1852 (Enflame tops line generation change) — https://github.com/flagos-ai/build-infra/pull/1852
- [4] build-infra #1851 (vllm 0.32.0 five backends enter application layer) — https://github.com/flagos-ai/build-infra/pull/1851
- [5] build-infra #1863 (vllm 0.32.0 status matrix) — https://github.com/flagos-ai/build-infra/pull/1863
- [6] build-infra #1872 (Ascend vllm 0.32.0 upstream blockage) — https://github.com/flagos-ai/build-infra/pull/1872
- [7] build-infra #1841 (mthreads toolkit layer and mtml baseline) — https://github.com/flagos-ai/build-infra/pull/1841
- [8] build-infra #1845 (mtml baseline limited to driver 5.3.0) — https://github.com/flagos-ai/build-infra/pull/1845
- [9] build-infra #1854 (Enflame flagcx pinned to 0.14.1) — https://github.com/flagos-ai/build-infra/pull/1854
- [10] build-infra #1870 (ECCL version declared per Enflame row) — https://github.com/flagos-ai/build-infra/pull/1870
- [11] build-infra #1868 (thead vllm application dependency adds flash_attn_3) — https://github.com/flagos-ai/build-infra/pull/1868
- [12] build-infra #1878 (flagcx wheel verified inside builder) — https://github.com/flagos-ai/build-infra/pull/1878
- [13] FlagTree #1637 (Ascend Triton 3.6.x backend port) — https://github.com/flagos-ai/FlagTree/pull/1637
- [14] FlagTree #1619 (Enflame TritonGCU backend update) — https://github.com/flagos-ai/FlagTree/pull/1619
- [15] FlagTree #1675 (unified configuration and profiling toolchain) — https://github.com/flagos-ai/FlagTree/pull/1675
- [16] FlagTree #1676 (Iluvatar SME row stride split) — https://github.com/flagos-ai/FlagTree/pull/1676
- [17] FlagGems #9334 (285B model profile enters benchmark) — https://github.com/flagos-ai/FlagGems/pull/9334
- [18] FlagGems #4916 (large-row-count LayerNorm backward optimization) — https://github.com/flagos-ai/FlagGems/pull/4916
- [19] FlagGems #8526 (T-Head GEMM test shape refresh) — https://github.com/flagos-ai/FlagGems/pull/8526
- [20] FlagGems #9160 (Triton autograd backward for log10) — https://github.com/flagos-ai/FlagGems/pull/9160
- [21] FlagGems #9294 (Kunlunxin row_stack and other coverage) — https://github.com/flagos-ai/FlagGems/pull/9294
- [22] FlagGems #9335 (MetaX W8A8 follows TLE API) — https://github.com/flagos-ai/FlagGems/pull/9335
- [23] FlagGems-vllm #1197 (DeepSeek-V4.1 decode block selection) — https://github.com/flagos-ai/FlagGems-vllm/pull/1197
- [24] FlagGems-vllm #1192 (Hygon BW1000 INT8 output projection) — https://github.com/flagos-ai/FlagGems-vllm/pull/1192
- [25] FlagGems-vllm #1177 (TLE fused matrix multiply-add reduce-scatter) — https://github.com/flagos-ai/FlagGems-vllm/pull/1177
- [26] FlagGems-vllm #1280 (T-Head PPU-ZW810E cooperative_topk) — https://github.com/flagos-ai/FlagGems-vllm/pull/1280
- [27] FlagGems-vllm #1239 (Ascend mhc_pre mix-major layout) — https://github.com/flagos-ai/FlagGems-vllm/pull/1239
- [28] FlagGems-vllm #1286 (Ascend MegaGate, in review) — https://github.com/flagos-ai/FlagGems-vllm/pull/1286
- [29] FlagGems-vllm #1287 (Ascend MegaMHC, in review) — https://github.com/flagos-ai/FlagGems-vllm/pull/1287
- [30] FlagGems-vllm #1274 (T-Head INT8 MegaMoE, in review) — https://github.com/flagos-ai/FlagGems-vllm/pull/1274
- [31] FlagGems-vllm #1283 (SiliconFlow triton_ihc_pre, in review) — https://github.com/flagos-ai/FlagGems-vllm/pull/1283
- [32] FlagGems-vllm #1234 (Ascend Runner V2 sampling export, in review) — https://github.com/flagos-ai/FlagGems-vllm/pull/1234
- [33] FlagGems-vllm #1277 (pause metax-maca3720 bare-run CI) — https://github.com/flagos-ai/FlagGems-vllm/pull/1277
- [34] FlagGems-sglang #392 (batch12 seventeen-operator skeleton) — https://github.com/flagos-ai/FlagGems-sglang/pull/392
- [35] FlagGems-sglang #390 (AMD gfx1100 performance) — https://github.com/flagos-ai/FlagGems-sglang/pull/390
- [36] FlagSparse #121 (NCIC cuad, sddmm, and spgemm optimization) — https://github.com/flagos-ai/FlagSparse/pull/121
- [37] FlagAttention #142 (repository restructuring) — https://github.com/flagos-ai/FlagAttention/pull/142
- [38] FlagAttention #141 (Ascend MoBA backend, in review) — https://github.com/flagos-ai/FlagAttention/pull/141
- [39] FlagAttention #143 (T-Head Inkling FA4 relative attention, in review) — https://github.com/flagos-ai/FlagAttention/pull/143
- [40] FlagTrain #89 (bias_geglu_impl aligned to two-parameter interface) — https://github.com/flagos-ai/FlagTrain/pull/89
- [41] FlagQuantum #1190 (package Sparse Blossom batch decoding) — https://github.com/flagos-ai/FlagQuantum/pull/1190
- [42] FlagQuantum #1195 (embed Stim for position noise sampling) — https://github.com/flagos-ai/FlagQuantum/pull/1195
- [43] FlagQAI #96 (local validation of review candidates) — https://github.com/flagos-ai/FlagQAI/pull/96
- [44] Torch-FL #846 (Ascend native host pilot) — https://github.com/flagos-ai/Torch-FL/pull/846
- [45] FlagCX #712 (Enflame ECCL device API) — https://github.com/flagos-ai/FlagCX/pull/712
- [46] FlagCX #722 (MUSA platform support, in review) — https://github.com/flagos-ai/FlagCX/pull/722
- [47] vllm-plugin-FL #703 (Ascend adaptation to vLLM 0.32 Runner V2) — https://github.com/flagos-ai/vllm-plugin-FL/pull/703
- [48] vllm-plugin-FL #694 (Enflame M-RoPE and FP8 checkpoint fix) — https://github.com/flagos-ai/vllm-plugin-FL/pull/694
- [49] sglang-plugin-FL #155 (Kunlunxin upgraded to SGLang 0.5.18) — https://github.com/flagos-ai/sglang-plugin-FL/pull/155
- [50] Megatron-LM-FL #221 (nightly supplementary coverage suite) — https://github.com/flagos-ai/Megatron-LM-FL/pull/221
- [51] FlagDNN #26 (Fedora and openEuler RPM packaging, in review) — https://github.com/flagos-ai/FlagDNN/pull/26
- [52] community #168 (FEP-0168 FlagTree multi-platform support) — https://github.com/flagos-ai/community/pull/168
- [53] Jrj: Tsinghua Unigroup Open3D-PIMC and People’s Daily commentary interpretation — https://finance.jrj.com.cn/2026/11/17104459165997.shtml
- [54] Xinhua: Open3D-PIMC fills out domestic computing software ecosystem (September open-source background) — http://www.news.cn/finance/20260914/0aa70ce06a464f7fab31403fb884da96/c.html
- [55] Yicai: Moore Threads completes Kimi K3.1 Pro training adaptation — https://www.yicai.com/news/103444014.html
- [56] Yicai: MetaX completes SenseNova Loop-9B Day 0 adaptation — https://www.yicai.com/news/103444279.html
- [57] NBD: Cambricon completes SenseNova Loop-9B adaptation — https://www.nbd.com.cn/articles/2026-11-17/4650764.html
- [58] SenseNova-Loop repository README (model structure and deployment requirements) — https://github.com/OpenSenseNova/SenseNova-Loop/blob/main/README_CN.md
- [59] CSDN FlagOS column — https://flagos.csdn.net/
- [60] flagos-ai organization homepage — https://github.com/flagos-ai