FlagOS Daily Intelligence Report (2026-11-17)
Research window: 2026-11-16 06:00 ~ 2026-11-17 06:00 (Beijing time, 24 hours; continues from the cutoff of the previous daily report).
Sources: GitHub (full audit of 64 repos under org: flagos-ai, 131 commits within the window, 135 merged PRs, 297 updated open PRs within the window, pushes to 20 repos, atomic audit 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).
Index of This Issue
- Today’s Highlights: The 2.3.0 runtime image begins to be accounted for by vendor; Hygon dtk26.10 and FlagTree 0.7.1a1 complete version interlocking (11-16)
- I. Open Source Project Progress (GitHub Activity)
- 1.1 Organizational cadence: 20 repos pushed, 131 commits, 135 merges, 22 commits in the single hour at 18:00 (11-16)
- 1.2 build-infra: flagtree precompiled input keys changed for four backends, Moore Threads cp312 line moves to the daily line (11-16)
- 1.3 FlagGems: 44 merges, Kunlunxin’s 19 and Moore Threads’ 12 KernelGen batches land in the afternoon (11-16)
- 1.4 FlagGems: exponential_ replay random stream and clone’s FakeTensor crash fixed on the same day (11-16)
- 1.5 FlagGems-vllm: Ascend 950DT’s first BNSD FlashAttention forward, sampling and speculative decoding operators completed (11-16)
- 1.6 FlagTree: TLE tracks shared byte ranges across calls, AMD atomic-order barrier ported to Triton 3.9 (11-16~11-17)
- 1.7 Quantum line: FlagQAI’s 22 web features land in a series, FlagQuantum connects H2 UCCSD three encodings (11-16)
- 1.8 Framework and communication: Torch-FL isolates native host startup ABI, FlagCX adds Iluvatar GDR and PXN hooks (11-16)
- 1.9 Megatron-LM-FL: DeepSeek-V4 sync third batch, DSA support packages THD and context parallelism (11-16)
- 1.10 Review queue: 297 under review, 93 newly opened that day, Kunlunxin 287 operator fix enters review in one go (11-16)
- II. News Coverage and Ecosystem
- 2.1 Channel audit: 24 query sets yield 13 candidates, zero hits for FlagOS and component names (11-16)
- 2.2 BAAI: REAL Physical AI Open Day scheduled for November 24 (11-16)
- 2.3 Member side: MetaX earnings call and Enflame Q3 report (11-16)
- III. Deep Dive into Member Units and Backend Vendors
- 3.1 Moore Threads: 12 dedicated foreach family operators added to the repository in one go (11-16)
- 3.2 Kunlunxin: 19 backend patches land within one hour (11-16)
- 3.3 Hygon: LLVM and FlagTree version interlocking in dtk26.10 (11-16)
- 3.4 Iluvatar CoreX: FlagCX adds GDR device family and PXN topology queries (11-16)
- 3.5 Ascend: 950DT’s a5 architecture and KMCompiler silent wrong-value fix proceed in parallel (11-16)
- IV. Summary
- Appendix: Source audit table and complete source list
Today’s Focus: 2.3.0 runtime images begin per-vendor bookkeeping; Hygon dtk26.10 and FlagTree 0.7.1a1 complete version interlock (11-16)
Date: 2026-11-16 Source: build-infra #1831 · build-infra #1832 · build-infra #1833 · build-infra #1834
- Bookkeeping: All 13 of build-infra’s merges today revolve around a single thread—four bot-generated documentation records (#1824 and #1835 update runtime, #1833 updates base, #1834 adds the changelog date for hygon-dtk26.10) write the 2.3.0 tag into the description pages of each vendor’s base and runtime images. On the Hygon side, two tags are directly visible:
flagos-base-hygon-dtk26.10:2.3.0andflagos-runtime-hygon-dtk26.10:2.3.0. The images pinflagtree==0.7.1a1+hcu3.6,flag_gems==5.4.0,flagcx==0.14.1+dtk2610, andtorch==2.13.0+dtk2610, with Triton 3.7.0 retained as an optional switchable item. - Reason for pairing: The description in #1831 spells out the root cause precisely—hygon 26.10’s clang-18 switched from bare
dtk_llvm.runtodtk_llvm-1.0.0+dtk2604.2609221416.g5f9d5e, and thenumRecordsparameter ofllvm.amdgcn.make.buffer.rsrcis i32 in the old package but i64 in the new one; FlagTree emitted i32 before 0.7.0 and i64 starting from 0.7.1a1. When the two are mismatched, every kernel compilation reportsIntrinsic has incorrect argument type—which is exactly the failure the 26.10 line had previously hit on the F path. This change locks aillvm and FlagTree to the same version pair, and on hygon25 verifies with a locally built cp312 wheel that the add and reduction kernels produce values consistent with torch on the FlagTree path. - Registration: #1832 adds the index record for the cp312 wheel of
flagtree-0.7.1a1+hcu3.6inflagos-pypi-hygon—that index previously had only the cp310 record for 0.7.0, while the 26.10 runtime now pins 0.7.1a1, so the missing piece was exactly the wheel this line needs to install. - Schedule: Per the community 2.3 timeline, the FEP merge deadline is 11-13, feature freeze and rc0 cut is 11-30, rc1 is 12-07, rc2 is 12-14, final freeze is 12-25, and GA is 12-30. The 2.3.0 tags written today belong to the pre-release line and do not equal an official release; the daily will henceforth record in two layers: “images land first, release follows.”
- Implication: The real fragility of a cross-chip stack often lies not in operator semantics but in the binary conventions between the compiler and driver packages—a single change in an intrinsic’s parameter width can break compilation for an entire vendor line. Only by writing such pairings into build-infra’s contract checks (
check_flagtree_contract.py --rows hcu-dtk2610reports no drift) is the premise of “the same code running on multiple chips” truly pinned down.
I. Open Source Project Progress (GitHub Activity)
1.1 Organizational Cadence: 20 Repos Pushed, 131 Commits, 135 Merges, 22 Commits in the 18:00 Hour (11-16)
Date: 2026-11-16 Source: flagos-ai organization · FlagGems · FlagQAI
- Totals: 20 of 64 repos had pushes within the window; 131 commits on default branches, 135 merged PRs, and 297 open PRs updated within the window (93 of them newly opened that day). Merging was concentrated in the afternoon: rising hour by hour after 13:00, with 26 merges and 22 commits in the 18:00 hour on the 16th, the peak of this window.
- Commit distribution: FlagGems 44, FlagQAI 29, FlagGems-vllm 16, build-infra 13, FlagTree 9, FlagQuantum 8, FlagAttention 4, Torch-FL 2, and 1 each for FlagCX, FlagScale, FlagTrain, FlagSparse, optest, and Trident.
- Merge distribution: FlagGems 44, FlagQAI 29, FlagGems-vllm 16, build-infra 13, FlagTree 10, FlagQuantum 8, FlagAttention 4, Torch-FL 2, Megatron-LM-FL 2, and 1 each for the remaining 7 repos.
- New repos: No new repos in the window; the most recent, FlagMRI-3D (created 11-06) and FlagOmni (created 11-09), are still within ten days and are the result of recent line expansion.
- Observation: No releases and no new tags in this window; all version-side activity is at the mirror and index layer, consistent with the preparation period for the 2.3.0 pre-release line.
1.2 build-infra: flagtree Precompiled Inputs Re-keyed for Four Backends, Moore Threads cp312 Row Moves to Daily Line (11-16)
Date: 2026-11-16 Source: build-infra #1820 · build-infra #1829 · build-infra #1830 · build-infra #1825
- Three re-keying PRs: #1820 switches the LLVM and Triton plugin in the sunrise build stage to the cache keys that FlagTree
mainnow looks up (llvm-ff67ef55,sunriseTritonPlugin_v0.7.0.1), becausesunrise.pyrenamed the key fromsunrise_llvm22_dev_releasetosunrise_llvm-<package>; with the old key unavailable, the daily wheel falls back to a vendor ks3 bucket unreachable from the runner. #1822 and #1823 do the same for Moore Threads and Kunlunxin respectively. All three require the corresponding line incheck_flagtree_contract.pyto report no drift. - Daily line expansion: #1829 releases the cp312 row of
mthreads-musa5.3.0fromdaily.enabled=false, placing it alongside the existing cp310 row inflagos-pypi-daily; cache keys are switched to the v0.7.1 pair, and the LLD assertion and directory carried over from the v0.6.1 era are removed. The author notes that once this row’s builder stage lands, it must be republished, otherwise the daily build will still warm on the old LLVM. - Version derivation: #1830 removes the
guess-next-devheuristic from the FlagGems daily-line base version and derives it directly from release records, avoiding daily package version drift. - Contract check moved earlier: #1825 makes the flagtree contract check a gate on flagtree PRs, catching “a backend’s precompiled inputs no longer match” before merge.
1.3 FlagGems: 44 Merges, Kunlunxin 19-PR and Moore Threads 12-PR KernelGen Batches Land in the Afternoon (11-16)
Date: 2026-11-16 Source: FlagGems #9157 · FlagGems #9186 · FlagGems #7333 · FlagGems #9102
- Scale: 44 merges, matching the commit count, one third of the whole organization. Author concentration is very high: Yangyang0906C 18, Yukun-Cui 12, 30 combined; the rest are CI and scattered fixes.
- Kunlunxin batch: Between 18:00 and 18:19, Yangyang0906C merged 18 KernelGen backend fixes in nineteen minutes, covering linear algebra (
linalg_inv,linalg_multi_dot,linalg_lu_solve), reductions (_aminmax,unique_dim_consecutive), copy and view (detach_copy,view_as_real_copy,col_indices_copy), special functions (scaled modified Bessel K0), quantization (weight_int8pack_mm), attention (_addmm_activationincluding the out variant, fixing GELU and improving throughput), the backward gradient of nearest-neighbor 3D upsampling, nested tensor metadata, andfloat_power. sunge666-ui added two more:split_copyperformance and anative_batch_normbenchmark defect. - Moore Threads batch: Between 14:43 and 15:18, Yukun-Cui merged 12 dedicated operators in thirty-five minutes —
true_divideandtrue_divide_,_foreach_abs_,_foreach_acos_, five variants of_foreach_add_(Scalar, ScalarList, List, Scalar, ScalarList), and three variants of_foreach_addcdiv_(Scalar, ScalarList, Tensor). The_foreach_family is a hot spot for batched optimizer updates; this batch consolidates the scattered multi-tensor elementwise paths into dedicated kernels. - NVIDIA sparse batch: Four KernelGen operators — compressed sparse index argument validation, COO sparse tensor argument validation,
_sparse_sparse_matmul, andreshape_as. - Three CI PRs: #9153 bumps the backend image from 2.2.0 to 2.3.0; #9152 upgrades Kunlunxin’s flagtree to
0.7.1a1+xpu3.6; #9192 fixes daily test efficiency and handles Ascend job cancellation. - In-flight branches: The Moore Threads week1107 batch has opened several branches (CPU reference for inverse, Sobol sampling device, geqrf core cases, fake-quantize tensor qparams, quantized max-pool3d output), corresponding to 8 open test-alignment PRs; a
kernel_detectbranch is also pushing kernel detection.
1.4 FlagGems: exponential_ Random-Stream Replay and clone FakeTensor Crash Fixed the Same Day (11-16)
Date: 2026-11-16 Source: FlagGems #8141 · FlagGems #8893 · FlagGems #9266
- Random-stream replay: #8141 notes that vLLM samples via the exponential-race form
argmax(probs / q),q ~ Exp(1), soexponential_must return i.i.d. samples and advance the generator between calls. The f32unroll8kernel takes one philox counter per output element, but the wrapper only advances the generator byN/8, so each call replays the previous call’s 7/8 of the random stream — in a 2^20-scale sampling,q2[:-N/8]andq1[N/8:]coincide exactly across 917504 elements. The fix reserves counters according to the programs actually launched (including tail lanes that are masked but still sampling). - Log fast-path distortion:
safe_fast_log_f32splitsln(u)intoexponent*ln2 + p(mantissa); whenuis slightly below 1, both terms are about 0.69 and cancel, leaving the polynomial’s ~1e-5 absolute error directly in the result and distorting the left tail. Fixed in the same PR. - Crash under compile: #8893 addresses
x.clone()reporting “Cannot access data pointer of Tensor (e.g. FakeTensor…)” insidetorch.compilewith FlagGems enabled —aten::cloneis a composite operator whose Meta implementation goes through the default backend kernel, so shape propagation lands inflag_gems/ops/clone.pyand callsdata_ptr()on a FakeTensor. The fix detectsis_metaorfake_modeinputs and resolves the output layout directly without launching a kernel; 3 new cases covering memory formats times transposed/regular were added, andtest_clone.pypasses 64 items. AMD’s vllm-plugin-FL adaptation is pinned to this PR. - Benchmark defect: #9266 fixes the hardcoded backend dispatch in the
router_gemmbenchmark so it picks the kernel for the current backend.
1.5 FlagGems-vllm: First BNSD FlashAttention Forward for Ascend 950DT, Sampling and Speculative Decoding Operators Completed (11-16)
Date: 2026-11-16 Source: FlagGems-vllm #1245 · FlagGems-vllm #1194 · FlagGems-vllm #1115 · FlagGems-vllm #1163
- 950DT forward: #1245 adds a BNSD FlashAttention forward for Ascend 950DT (dense and causal branches), registered as
flash_attention_fwdunder_ascend/a5/opsand dispatched viaSpecOpRegistrar’sa5architecture key, +2215 lines, 7 files. The interface is dense-formq[B,Hq,S,D],k/v[B,Hkv,S,D], fp16,D ∈ {64,128},Hq % Hkv == 0,Sq == Skv, kept distinct from the existing varlen/paged paths. - Implementation approach: The kernel uses plain
tl.dotplus online softmax (dense and causal sharebuild_tile_order,_max_nan, and_COMPILE_OPTS), delegating Cube/Vector pipeline splitting to the compiler (--set-cv-pipeline-mode) rather than hand-writingcann.scopeandsync_block_*scheduling; the causal branch no longer peels the diagonal block out of the pipelined KV loop, and masking becomes an in-tile row/column comparison. The mask uses a true-infinstead ofqk + (-1e6)— whenq·k·scaleapproaches 1e6 the finite bias is overwhelmed; the author notes this was reproduced against the vendor implementation. - T-Head PPU: #1194 adds
cast_sigmoid(elementwise sigmoid with explicit output and compute dtype, for vLLM’s grouped top-k gating scores); sigmoid is written as1/(1+exp2(-x*log2(e)))so the exponential maps to an SFU instruction andlog2(e)folds into the parameter, with the FP32 result converted only once on store; autotune candidates are registered in both T-Head and NVIDIA tune_configs. - Sampling and speculative decoding: five consecutive merges complete the draft-to-sample chain — draft slot padding (#1171), watermark mask (#1112), draft sampling collection (#1113), sampling output (#1158) and sampled/rejected counts (#1115), under the operator names
pad_trailing_draft_slots,draft_watermarking_mask,gather_draft_sampled,gather_sampler_output
andget_num_sampled_and_rejected. - Sampling counter on H100: #1115 changes H100 to process at most 128 requests per Triton program (previously one program per request), adds a T-Head backend implementation, and auto-selects it on PPU devices.
- Extension scaffolding: #1163 adds the optional
flaggems_vllm.extensionsnamespace (+1288 lines, 44 files), dispatching arch → vendor → generic in descending order and raisingNotImplementedErrorwhen an implementation is missing rather than falling back to torch computation; it is kept separate fromflaggems_vllm/ops/because the latter is pulled into_FULL_CONFIGbyenable()and patched into ATen. This is scaffolding only, with no operators yet. - Other merges: #1229 adds exact-FP32
cooperative_topkfor Ascend 910B4 and Tianshu BI-V150 (148 passed, 8 skipped); #1225 aligns CI naming and structure with FlagGems; #1253 makes benchmarks use the vLLM reference when available.
1.6 FlagTree: TLE Tracking Shares Byte Ranges Across Calls, AMD Atomic-Ordering Barrier Ported from Triton 3.9 (11-16~11-17)
Date: 2026-11-16 ~ 2026-11-17 Source: FlagTree #1537 · FlagTree #1654 · FlagTree #1626 · FlagTree #1662
- TLE alias analysis: #1537 makes TLE’s barrier analysis track physical byte ranges and borrowed shared memory — shared buffers may alias via byte offsets or be passed as borrowed arguments, with the callee’s scratch landing in the translated call frame; the old analysis only looked at callee exit accesses and thus missed ordering between caller and callee and between atomic return broadcasts. The change propagates call entry and exit summaries (with frame offsets) and checks atomic scratch in the operator’s owning function (+1043/-11, 8 files, native compilation changes guarded by
__TLE__). - Verification strength: All 5 new TLE MLIR regression files pass, including a cross-call atomic case that fails on the original implementation; 15 GPU cases pass on H800, and the 4-warp owned-alias case reproduces 64/128 wrong values on
main; the full TLE MLIR suite passes 121, with the same 7 failures asmain. - AMD atomic ordering: #1654 ports Triton v3.9’s
Insert CTA barriers for atomic memory semanticsand a structured-binding fix inemitFence. The observed failure is stark — on AMD, release atomics lower only tobuffer_wbl2+s_waitcntcovering the current wave, so with more than 4 warps per program, other warps’ writes are not visible before the barrier count advances: 4-warp and 8-warp runs read stale data 0-2 and 1-2 times per 300 launches, and 16-warp runs 300/300. FlagGems’linalg_matrix_powergrid-sync kernel hits exactly this, withtest_linalg_matrix_power.pyfailing 30-32/506 cases per run on Instinct. - Non-MFMA dot: #1626 fixes AMD Backend BlockPingpong behavior on non-MFMA dot, porting the corresponding Triton v3.7 change.
- Backend list: #1662 adds the Intellifusion backend to the readme and updates the TLE support list (-922 lines are documentation reflow).
- Three CI PRs: #1659 makes smart-checkout survive dropped connections; #1647 wires FlagGems-vllm tests into the shared FlagGems CI; #1663 notifies PR authors when Code-Format-Check fails; #1669 and #1670 are duplicate PRs clearing
/dev/shmbefore all-backend Qwen benchmarks.
1.7 Quantum Line: FlagQAI Lands 22 Web Features in a Series, FlagQuantum Completes H2 UCCSD Across Three Encodings (11-16)
Date: 2026-11-16 Source: FlagQAI #51 · FlagQAI #66 · FlagQuantum #1188 · FlagQuantum #1191
- FlagQAI scale: 29 commits, 22 merges, all driven by the FlagQuantum account between 12:10 and 22:51, with highly regular commit titles (feat/fix/ci prefix plus module name), showing a continuous toolchain-style progression.
- Feature surface: On the web side, from displaying distributions for saved samples (#42), experiment diff summaries (#43), synthesis API preflight (#45), local or SSH compute environment selection (#47), applying simulator execution profiles (#48), pinning official synthesis examples and comparing evidence (#49, #55, #56), nested model folders (#51), distinguishing synthesis from gate-level Python (#52), and wider circuit previews (#53), to selecting saved circuits directly in Execution (#57), viewing compiler source in Quantum (#59), and viewing gate sequences in Jobs (#60); on the Agent side, current-round activity display (#62), collapsible tool output (#64), failure explanations (#61), and transcript scroll sync (#68).
- CI side: #65 and #66 route admin PRs to dedicated runners, and #67 pauses Windows runner jobs — related to the QAI platform’s own web build costs.
- 8 FlagQuantum merges: Native measurement of two-body density tensors (#1182), projecting measured densities onto fixed-particle-number constraints (#1185), binding execution inputs at runtime while preserving training state (#1184), and training UCCSD for four-mode H2 under Jordan–Wigner, parity, and Bravyi–Kitaev encodings (#1188), followed by tapering for the three encodings (#1191) and tapered training within the JW spin sector (#1189).
- Numerical results: #1188 reports that every path yields two particles and a nonzero initial gradient, with final energies within 1e-6 Ha of the two-particle bounded reference -1.1372701747 Ha; 78 unit, integration, CLI, and isolation tests pass, with no dependency on OpenFermion or Cirq.
1.8 Frameworks and Communication: Torch-FL Isolates the Native Host Launch ABI, FlagCX Adds Tianshu GDR and PXN Hooks (11-16)
Date: 2026-11-16 Source: Torch-FL #841 · Torch-FL #844 · FlagCX #714 · vllm-plugin-FL #659
- Precision routing: #841 configures native precision routing for eager inference (+4115/-232, 59 files). The unmodified #839 wheel fails all 513 full-logit comparisons, so this PR generates six dtype-scoped precision exceptions against the published vendor numerical libraries and compiles a testable replacement
mul.Tensorhost from the same published FlagGems wheel; the generated backend config explicitly declares these native dtype routes, and registering native modules must not override ordinary FlagGems config entries. After installation, wheel re-verification passes 513/513 across three full native processes, with a maximum logit error of 0.0. - Launch ABI: #844 isolates the native host backend launch ABI (+1413/-401, 18 files) — CUDA metadata, descriptors, scratch parameters, and enqueue calls are collected into an NVIDIA adapter, and the shared host compiler and launcher reuse Torch-FL’s device, stream, and allocation abstractions; the build, preflight, and activation steps all verify that the selected adapter matches the packaged source, the published compiler driver, and the compilation ABI. Only NVIDIA is admitted for now; native-host adapters for other vendors are unavailable until separately implemented and verified.
- Tianshu communication: FlagCX #714 adds Tianshu’s GDR device family and registers GPU memory registration capability through the IB adapter, adds the IXCUDA peer-access query needed for PXN topology discovery, and cleans up GPU allocation when GDR pointer setup fails while preserving existing write-flush behavior; previously quarantined SendRecv and PXN AlltoAll cases are re-enabled. The PR reports local checks passing 11/11, 9/9, and 3/3; hardware results on CoreX GPU with IB need separate confirmation outside the merge process.
- AMD adaptation: vllm-plugin-FL #659 enables graph mode for vLLM 0.32 on AMD W7900D (gfx1100) (+2615/-121, 36 files), with a runtime of ROCm 7.2.3 container, RCCL 2.27.7, PyTorch 2.13.0, and FlagCX 0.15.0.dev52; the prebuilt image
v0.32.0-amd-pr659-5f3f669has been published and pulled by digest on a second W7900D host, and after 104/104 images pass it enters the 0.5.0-dev mainline.
1.9 Megatron-LM-FL: Third DeepSeek-V4 Sync Batch, DSA Supports Packed THD and Context Parallelism (11-16)
Date: 2026-11-16 Source: Megatron-LM-FL #218 · Megatron-LM-FL #222 · Trident #11
- Scale: #218 is the third batch syncing DeepSeek-V4 upstream (+10506/-2082, 25 files), extending DSA from fixed-length SBHD inputs to packed THD and all-gather context parallelism, and adding core cuDNN/FlashMLA backends while retaining the PyTorch fallback.
- Technical points: Packed THD requires handling sequence boundaries, global token positions, causal masks, and effective top-k lengths; the unfused path no longer builds the full dense QK score matrix, instead using blocked sparse gathering with FP32 accumulation. The backend config replaces
apply_dsa_kernel_fusionwithdsa_kernel_backend="none"|"cudnn", defaulting tocudnnfor DSv4 hybrid andnonefor other attention variants, and validates unsupported combinations such as dependency, hardware support, projection dtype, and SM90 dense indexer loss. - Continuity: The first two batches were [1/N] on 11-03 (DSv4 and mHC precision) and [2/N] on 11-04 (mHC expert parallelism); this batch follows them and closes out the same sync line.
- CI: #222 makes
run_ci_test.shrespect a caller-suppliedENABLE_LIGHTWEIGHT_MODE, preventing functional tests from being forced into lightweight mode. - New project: Trident #11 introduces a guard tree to support shared candidate checking. Trident’s positioning is stated in its README — compiling Python functions containing Triton/Torch-style computation into MLIR/LLVM and executing them via tvm_ffi, another compilation-path exploration alongside FlagTree (the Triton route) and TileLang.
1.10 Open Queue: 297 Open PRs, 93 Opened That Day, Kunlunxin 287-Operator Fix Enters Review in One Shot (11-16)
Date: 2026-11-16 Source: FlagGems #9198 · FlagGems #9297 · FlagGems #9288 · FlagGems-vllm #8526 (T-Head PPU)
- Totals: 297 open PRs updated within the window, 93 newly opened that day. Distribution: FlagGems 167 (88 new), FlagGems-vllm 42, vllm-plugin-FL 19, FlagTree 17, FlagCX 11, FlagTrain 10, sglang-plugin-FL 6, FlagAttention 5, and 1 to 3 each for the rest.
- Kunlunxin mega-PR: #9198 is titled “[kunlunxin] 287 ops fix”, +53022/-1494, 203 files, opened at 14:20 that day; the PR description is empty, so reviewers can only judge coverage from the diff — a scale that needs close attention.
- SDPA dispatch: #9297 routes
scaled_dot_product_attentionand_scaled_dot_product_cudnn_attentionthrough the backend’s existing blocked attention kernel instead of materializing the full(B*H, Q, K)score buffer; the author’s motivation is that the widest-shape benchmark still produced no result after 1225 seconds, while after the change official accuracy is 96/0/0 with an aggregate of 1.159. - Moore Threads test batch: #9288, #9289, #9291, #9298, #9302, #9304, #9305, and #9307 were opened back-to-back by the same author between 20:05 and 23:20, aligning per-channel and tensor-qparams fake quantize, the geqrf CPU Householder reference, the quantized max-pool3d output fixture, and Sobol sampling inputs to the active backend.
- Other open PRs: GEMM test shape and layout coverage for T-Head PPU (#8526), Kunlunxin operator coverage for rnn/kl_div/row_stack and others (#9285, #9290, #9292, #9293, #9294), flash-attention forward
l_iinitialization (#9295), and KernelGen autograd backward for logaddexp/min/logaddexp2 (#9299, #9300, #9301).
II. News Coverage and Ecosystem
2.1 Channel Verification: 24 Query Sets Yield 13 Candidates; Zero Hits for FlagOS and Component Names (11-16)
Date: 2026-11-16 Sources: Google News RSS · Hacker News · CSDN FlagOS Column
- Query results: 24 Chinese and English query sets, deduplicated within a 24-hour window, yielded 13 candidates.
FlagOS,FlagGems,FlagScale,FlagTree,FlagCX,KernelGen,flagos-ai,清微智能, andTsingmicroall returned zero hits; component-name queries have now returned zero hits for several consecutive windows, consistent with FlagOS’s reach running mainly through technical communities rather than mass media. - Hit composition: 6 for Enflame (4 of which were sector market and capital-flow pieces), 4 for BAAI, 1 for MetaX, 1 for domestic compute, and 1 for the BAAI large-model open-source stack. After removing gambling-related and pure market-recap pieces by title semantics, the usable industry-side items were merged into Sections 2.2 and 2.3.
- HN: Within a 7-day window,
FlagOSreturned 7 hits but with no direct match in titles or body text; all other components returned 0. - CSDN column: The page shows two imports, at 11-16 10:59 and 14:18. The latest item is “GLM-5.3-Flash Day0 adaptation for 9 AI chips,” but the publish time inside the card is 2026-08-28, making it a redistribution of a historical article and thus not counted as same-day activity; this is consistent with the column’s recent practice of reordering and republishing old WeChat posts by theme.
2.2 BAAI: REAL Physical AI Open Day Set for November 24 (11-16)
Date: 2026-11-16 Source: Chaoqi Net report (gnews redirect)
- Event: BAAI will hold the “Measuring Intelligence by Reality: BAAI REAL Physical AI Open Day” on November 24, with registration opening on 11-16. REAL is BAAI’s vehicle for physical AI, related to embodied intelligence training and sim-to-real transfer.
- Relationship to FlagOS: FlagOS-Robo is the organization’s embodied intelligence toolchain, and embodied and robotics items have appeared repeatedly on the feed side; however, this public agenda does not mention any FlagOS component, so it is recorded only as ecosystem background without inferring that the two are already linked.
- Observation: BAAI’s public activity over the past two weeks has centered on physical AI and embodied evaluation (including the Ant Lingbo embodied large-model challenge’s 48-hour real-machine contest reported on 11-16). This runs in parallel with the cadence of FlagOS’s quantum and operator lines, indicating that the lead organization is building out both the “stack” and the “scenarios” ends simultaneously.
2.3 Member Organizations: MetaX Earnings Call and Enflame Q3 Report (11-16)
Date: 2026-11-16 Sources: Cailian Press report (MetaX earnings call) · Sina Finance report (Enflame Q3) · Sohu Finance report (sector market)
- MetaX: Held an earnings briefing on 11-16, where management judged that the GPU industry landscape is shifting and that the company will accelerate gains in domestic compute market share. The same-day code-side action is the MetaX MoE top-k routing extension in FlagGems-vllm (see the adjacent item in Section 3.1), with hardware and software moving in the same direction.
- Enflame: The Q3 report shows a loss of RMB 754 million for the first three quarters and RMB 6.1 billion raised in its September listing; on the day, the AI compute chip sector fell 6.86% overall, with Enflame leading the decline and net main-fund outflows of RMB 3.788 billion. That evening there was also an industry piece on “the STAR Market reinforcing the DSA-architecture compute puzzle.”
- Caveat on framing: The above is public information at the listed-company and sector level and has no direct correspondence to FlagOS code-side progress on the Enflame backend; there were no Enflame-related merges in the FlagOS org this period, so this is recorded only as member-organization background.
III. Deep Dive into Member Units and Backend Vendors
3.1 Moore Threads: 12 Dedicated foreach Family Operators Merged in One Batch (11-16)
Date: 2026-11-16 Source: FlagGems #9186 · FlagGems #9288 · build-infra #1829 · FlagSparse #120
- Operator surface: The foreach family gains 12 dedicated operators in one pass, covering two unary variants of
absandacos, five binary variants ofadd(Scalar, ScalarList, List, Scalar, ScalarList), three three-segment variants ofaddcdiv(Scalar, ScalarList, Tensor), and two forms oftrue_divide. What these operators share is that their inputs are lists of tensors; a generic elementwise implementation must expand over the list, whereas a dedicated kernel can fill the work items in one pass. - Build line: build-infra #1829 releases the cp312 row of
mthreads-musa5.3.0into the daily line, alongside the cp310 row; the author notes that this row’s builder stage must be republished once, otherwise daily builds will still use the old LLVM. MUSA 5.3.0 thus moves from “single py version” to dual py versions running together. - Test batch in flight: Eight MUSA week1107 test-alignment PRs are under review (fake quantize for per-channel and tensor qparams, a CPU Householder reference for geqrf, output fixtures for quantized max-pool3d, Sobol sampling inputs, coverage for kl_div and row_stack, etc.), while branch
fix/musa-week1107-inverse-cpu-referenceand fivetj/week1107-musa-*branches are being pushed. - Sparse line: FlagSparse #120 merges from
NCIC-AlphaSparse/main, titled “musa stable again” — that is, the sparse operators on the Moore Threads side are stable once more.
3.2 Kunlunxin: 19 Backend Fixes Landed Within an Hour (11-16)
Date: 2026-11-16 Source: FlagGems #9240 · FlagGems #9246 · FlagGems #9198 · FlagGems #9297
- Cadence: 18 merges within nineteen minutes from 18:00 to 18:19, plus two performance and benchmark fixes, for 20 in total; submitted consecutively by the same account, this is a single batch backfill rather than scattered fixes.
- Coverage: linear algebra (
linalg_inv,linalg_multi_dot,linalg_lu_solve,gradientnumerical precision), reductions (_aminmax,unique_dim_consecutive), copy and views (detach_copy,view_as_real_copy,col_indices_copy,split_copy), special functions (scaled modified Bessel K0), quantization (weight_int8pack_mm), attention (GELU and throughput for_addmm_activation), upsampling backward, and nested tensor metadata. - Large PR under review: #9198 “287 ops fix” at +53022/-1494 across 203 files, opened that day with an empty description. If merged, Kunlunxin’s operator coverage will change by an order of magnitude; review cost rises accordingly, so whether it gets split bears watching.
- Attention path: #9297 switches SDPA to a tiled kernel, motivated by the widest shape’s benchmark producing no result after 1225 seconds; after the change, official accuracy is 96/0/0 with an aggregate of 1.159.
- Build side: build-infra #1823 switches Kunlunxin’s flagtree precompiled inputs to the key currently looked up by
main, complementing FlagGems #9152’s upgrade to0.7.1a1+xpu3.6; the Kunlunxin line as a whole follows the 0.7.1 series.
3.3 Hygon: LLVM and FlagTree Version Interlock for dtk26.10 (11-16)
Date: 2026-11-16 Source: build-infra #1831 · build-infra #1832 · FlagAttention #136 · build-infra runtime docs
- Pairing details: clang-18 in the 26.10 base image is now provided by
dtk_llvm-1.0.0+dtk2604.2609221416.g5f9d5e. That build widens thenumRecordsparameter ofllvm.amdgcn.make.buffer.rsrcfrom i32 to i64, while FlagTree only emits i64 starting from 0.7.1a1; when the two are mismatched, every kernel compilation reportsIntrinsic has incorrect argument type. The runtime therefore pins FlagTree to0.7.1a1+hcu3.6and adds that version’s cp312 wheel to theflagos-pypi-hygonindex (previously only the cp310 record for 0.7.0 existed). - Image contents:
flagos-runtime-hygon-dtk26.10:2.3.0contains Python 3.12,
FlagTree0.7.1a1+hcu3.6, FlagGems5.4.0, FlagCX0.14.1+dtk2610, torch2.13.0+dtk2610, and flash_attn2.8.4+dtk2610; host driver 6.4.2-V2.1.0, chip model Hygon BW1000. Inside the container you can switch between FlagTree and Triton 3.7.0, and environment variables can re-pin FlagTree and FlagGems versions before startup. - Operator side: FlagAttention #136 simplifies dispatch and validation for Hygon GLA.
- Observation: The keyword for the Hygon line is “pairing” rather than “addition” — only by locking the driver package, LLVM, FlagTree, and the wheel in the index to the same set of versions does reproducibility become possible.
3.4 Iluvatar CoreX: FlagCX Adds GDR Device Family and PXN Topology Query (11-16)
Date: 2026-11-16 Source: FlagCX #714 · FlagGems-vllm #1229
- Capability surface: #714 adds Iluvatar’s GDR device family to FlagCX’s PAL layer and advertises GPU memory registration capability through the IB adapter, supplying the IXCUDA peer-access query needed for PXN topology discovery; it also fixes allocation cleanup when GDR pointer setup fails, preserves existing write-flush behavior, and re-enables the previously quarantined SendRecv and PXN AlltoAll cases.
- Validation: The PR self-reports local checks as 11/11 on the Iluvatar runner script, 9/9 on adapter GPU CI checks, and 3/3 on platform configuration, with the modified GDR and IBRC code passing local syntax checks; the PR also states that hardware results for CoreX GPU and IB must be confirmed outside the merge process, with acceptance criteria being four one-sided cases with no skips, PXN AlltoAll showing relay traffic on both logical nodes, and SendRecv passing in all three modes: HYBRID P2P, forced NET, and PXN.
- Operator side: #1229’s
cooperative_topkadds an exact FP32 implementation for Iluvatar BI-V150 — a shared-memory histogram plus candidate compaction, with exact refinement on capacity overflow, preserving the six-parameter API and in-place INT32 output. - Build context: FlagGems-vllm #659’s AMD adaptation notes mention that HEAD already includes one Iluvatar conflict resolution, indicating that the Iluvatar and AMD adaptation lines share the same batch of conflict handling on vLLM 0.32.
3.5 Ascend: 950DT’s a5 Architecture and KMCompiler Silent Wrong-Value Fix in Parallel (11-16)
Date: 2026-11-16 Source: FlagGems-vllm #1245 · FlagGems-vllm #1229 · FlagGems #8886 · FlagAttention #138
- New architecture: The 950DT BNSD FlashAttention forward enters
_ascend/a5/opswitha5as the architecture key, alongside the existing varlen/paged paths in_ascend/fused/attention.py; this PR uses “generate a usable forward for new hardware first” as the entry point, delivering benchmarks and accuracy tests together. - Precision stance: The mask uses a true
-infreplacement rather thanqk + (-1e6); the author’s rationale is that a finite bias can be breached whenq·k·scaleapproaches 1e6 — this kind of implementation difference that “looks like it runs but errs at the boundary” is precisely the hardest part to catch in cross-chip alignment. - Silent wrong values: #8886 fixes silent wrong values from KMCompiler on Ascend for polygamma and GRU caused by blockify region mismatches (the title directly names silent wrong results). Silent wrong values are more dangerous than compilation failures, because downstream only sees degraded accuracy.
- Other: #1229’s Ascend 910B4
cooperative_topkuses segmented CANN sort/merge plus integer-key refinement and a masked histogram; FlagAttention #138 addresses Ascend-side compatibility issues; on the build line,0.7.1+ascend3.5was already released on 11-12, with no new additions this period.
IV. Summary
- Version line: Pre-release images for 2.3.0 are beginning to be recorded per vendor. Hygon dtk26.10 and dtk26.04 already carry 2.3.0 tags, with images pinning FlagTree 0.7.1a1+hcu3.6, FlagGems 5.4.0, and FlagCX 0.14.1+dtk2610. Per the community schedule, rc0 lands on 11-30 and GA on 12-30; this cycle is a case of “images landing first, release following later.”
- Engineering focus: Of the 135 merges throughout the day, operator and backend fixes account for the overwhelming majority (FlagGems 44, FlagGems-vllm 16, FlagTree 10), and they show a clear batch pattern—19 Kunlunxin merges clustered within 19 minutes and 12 Moore Threads merges within 35 minutes, indicating that vendor adaptation work has entered a weekly batch-processing cadence.
- New hardware: The a5 architecture of Ascend 950DT enters FlagGems-vllm’s architecture-key dispatch for the first time, delivering BNSD FlashAttention forward. On the same day, operators such as
chunk_kdaon the 950 side had already landed a batch via the a5 path on 11-13; this line took shape over two consecutive weeks. - Correctness: Two fixes this cycle deserve separate mention—the random-stream replay in
exponential_(each call repeats 7/8 of the random numbers, directly affecting the independence of vLLM sampling) and the FlagTree AMD atomic-order barrier (300/300 stale reads at 16 warps). Neither is a performance issue; both silently alter results. - Review pressure: 297 PRs under review and 93 newly opened that day, among which the Kunlunxin “287 ops fix” alone touches 203 files with +53022 lines, a significant review cost; if accumulation continues over the coming days, the 2.3 FEP and merge thresholds will come under pressure.
- News side: Component-name queries returned zero hits for multiple consecutive windows. Public coverage centers on member companies’ capital and earnings (MetaX earnings call, Enflame Q3 report). FlagOS’s external narrative still relies mainly on technical documentation and its WeChat account, and the daily report’s core information continues to be supplied by the GitHub side.
Appendix: Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org (64 repos) | 20 repos pushed within the window; 131 commits, 135 merges, 297 under-review updates (93 newly opened) |
| GitHub Release / Tag | No additions within the window; most recent are FlagGems v5.5.0.dev1 (11-11) and FlagTree 0.7.1+xpu3.6 (11-12) |
| Google News RSS (24 groups) | 13 candidates; zero hits for FlagOS and component names; usable items merged into 2.2 and 2.3 |
| HN Algolia (7 days) | 7 hits for FlagOS, 0 direct matches; 0 for other components |
| CSDN FlagOS column | Two imports on 11-16; card original dates are historical dates such as 2026-08-28, constituting redistribution and not counted as same-day activity |
| BAAI community hub | Search page is client-rendered with no items to scrape; homepage features models and popular-science content |
| build-infra runtime docs | Both hygon-dtk26.04 and hygon-dtk26.10 already have 2.3.0-tagged docs |
| 2.3 timeline | FEP deadline 11-13; rc0 11-30; rc1 12-07; rc2 12-14; code freeze 12-25; GA 12-30 |
Limitations: GitHub-side data this cycle is cross-checked across three channels—commit search, merged-PR search, and org events; merge times of under-review PRs are not included in the statistics. Zero hits for component names on the news side are normal and do not support an inference that “no news means no progress.” CSDN column publication times are taken from the original publication field within the article page, which differs from the import time on the listing page; where the two conflict, the original publication time is used.
Full Source List
- [1] flagos-ai org homepage — https://github.com/flagos-ai
- [2] build-infra #1831 (hygon 26.10 paired with FlagTree 0.7.1a1) — https://github.com/flagos-ai/build-infra/pull/1831
- [3] build-infra #1832 (flagtree hcu 0.7.1a1 wheel registered) — https://github.com/flagos-ai/build-infra/pull/1832
- [4] build-infra #1833 (base 2.3.0 record) — https://github.com/flagos-ai/build-infra/pull/1833
- [5] build-infra #1834 (runtime 2.3.0 changelog date) — https://github.com/flagos-ai/build-infra/pull/1834
- [6] build-infra #1820 (sunrise precompiled input key change) — https://github.com/flagos-ai/build-infra/pull/1820
- [7] build-infra #1829 (mthreads cp312 line enters daily pipeline) — https://github.com/flagos-ai/build-infra/pull/1829
- [8] build-infra #1830 (FlagGems daily-line version derivation) — https://github.com/flagos-ai/build-infra/pull/1830
- [9] build-infra #1825 (flagtree contract check moved earlier) — https://github.com/flagos-ai/build-infra/pull/1825
- [10] FlagGems #9157 (Kunlunxin gcd_) — https://github.com/flagos-ai/FlagGems/pull/9157
- [11] FlagGems #9240 (Kunlunxin addmm_activation GELU and throughput) — https://github.com/flagos-ai/FlagGems/pull/9240
- [12] FlagGems #9246 (Kunlunxin _aminmax reduction) — https://github.com/flagos-ai/FlagGems/pull/9246
- [13] FlagGems #9186 (Moore Threads foreach_addcdiv Tensor) — https://github.com/flagos-ai/FlagGems/pull/9186
- [14] FlagGems #7333 (NVIDIA sparse validation operators) — https://github.com/flagos-ai/FlagGems/pull/7333
- [15] FlagGems #9102 (Kunlunxin split_copy performance) — https://github.com/flagos-ai/FlagGems/pull/9102
- [16] FlagGems #8141 (exponential_ random stream and log fast path) — https://github.com/flagos-ai/FlagGems/pull/8141
- [17] FlagGems #8893 (clone FakeTensor under torch.compile) — https://github.com/flagos-ai/FlagGems/pull/8893
- [18] FlagGems #9266 (router_gemm backend dispatch) — https://github.com/flagos-ai/FlagGems/pull/9266
- [19] FlagGems #8886 (Ascend KMCompiler silent wrong values) — https://github.com/flagos-ai/FlagGems/pull/8886
- [20] FlagGems #9153 (backend image 2.2.0 to 2.3.0) — https://github.com/flagos-ai/FlagGems/pull/9153
- [21] FlagGems #9198 (Kunlunxin 287 ops fix, under review) — https://github.com/flagos-ai/FlagGems/pull/9198
- [22] FlagGems #9297 (Kunlunxin SDPA tiled dispatch, under review) — https://github.com/flagos-ai/FlagGems/pull/9297
- [23] FlagGems #8526 (T-Head PPU GEMM coverage, under review) — https://github.com/flagos-ai/FlagGems/pull/8526
- [24] FlagGems-vllm #1245 (Ascend 950DT BNSD FlashAttention) — https://github.com/flagos-ai/FlagGems-vllm/pull/1245
- [25] FlagGems-vllm #1194 (T-Head PPU cast_sigmoid) — https://github.com/flagos-ai/FlagGems-vllm/pull/1194
- [26] FlagGems-vllm #1233 (MetaX MoE top-k routing) — https://github.com/flagos-ai/FlagGems-vllm/pull/1233
- [27] FlagGems-vllm #1115 (sampling and rejection counts) — https://github.com/flagos-ai/FlagGems-vllm/pull/1115
- [28] FlagGems-vllm #1163 (multi-backend extension operator scaffolding) — https://github.com/flagos-ai/FlagGems-vllm/pull/1163
- [29] FlagGems-vllm #1229 (Ascend and Iluvatar cooperative_topk) — https://github.com/flagos-ai/FlagGems-vllm/pull/1229
- [30] FlagGems-vllm #1203 (dsv41_o_proj baseline) — https://github.com/flagos-ai/FlagGems-vllm/pull/1203
- [31] FlagTree #1537 (TLE cross-call shared byte range) — https://github.com/flagos-ai/FlagTree/pull/1537
- [32] FlagTree #1654 (AMD atomic-order barrier ported to Triton 3.9) — https://github.com/flagos-ai/FlagTree/pull/1654
- [33] FlagTree #1626 (AMD non-MFMA dot) — https://github.com/flagos-ai/FlagTree/pull/1626
- [34] FlagTree #1662 (Intellifusion backend and TLE manifest) — https://github.com/flagos-ai/FlagTree/pull/1662
- [35] FlagQAI #51 (nested model folders) — https://github.com/flagos-ai/FlagQAI/pull/51
- [36] FlagQAI #66 (dedicated runner routing for management PRs) — https://github.com/flagos-ai/FlagQAI/pull/66
- [37] FlagQuantum #1188 (H2 UCCSD three-encoding training) — https://github.com/flagos-ai/FlagQuantum/pull/1188
- [38] FlagQuantum #1191 (H2 UCCSD taper) — https://github.com/flagos-ai/FlagQuantum/pull/1191
- [39] Torch-FL #841 (eager native precision routing) — https://github.com/flagos-ai/Torch-FL/pull/841
- [40] Torch-FL #844 (native host launch ABI isolation) — https://github.com/flagos-ai/Torch-FL/pull/844
- [41] FlagCX #714 (Iluvatar GDR and PXN hooks) — https://github.com/flagos-ai/FlagCX/pull/714
- [42] vllm-plugin-FL #659 (AMD W7900D graph mode) — https://github.com/flagos-ai/vllm-plugin-FL/pull/659
- [43] Megatron-LM-FL #218 (DeepSeek-V4 sync batch 3) — https://github.com/flagos-ai/Megatron-LM-FL/pull/218
- [44] Megatron-LM-FL #222 (CI lightweight mode passthrough) — https://github.com/flagos-ai/Megatron-LM-FL/pull/222
- [45] Trident #11 (guard tree) — https://github.com/flagos-ai/Trident/pull/11
- [46] FlagAttention #136 (Hygon GLA dispatch) — https://github.com/flagos-ai/FlagAttention/pull/136
- [47] FlagAttention #137 (NVIDIA baseline vs. optest comparison) — https://github.com/flagos-ai/FlagAttention/pull/137
- [48] FlagAttention #138 (Ascend compatibility) — https://github.com/flagos-ai/FlagAttention/pull/138
- [49] FlagScale #1328 (qwenvl remote-sensing image reading) — https://github.com/flagos-ai/FlagScale/pull/1328
- [50] FlagSparse #120 (musa stable again) — https://github.com/flagos-ai/FlagSparse/pull/120
- [51] optest #56 (same-day test log coverage) — https://github.com/flagos-ai/optest/pull/56
- [52] build-infra runtime docs (hygon-dtk26.10) — https://github.com/flagos-ai/build-infra/blob/main/docs/content/zh-cn/runtime/hygon-dtk26.10.md
- [53] CSDN FlagOS column — https://flagos.csdn.net/
- [54] BAAI REAL Physical AI Open Day (Chaoqi.com, gnews redirect) — https://news.google.com/rss/articles/CBMiY0FVX3lxTE1CUE1OblE2bTNNVm5QWVNkLUZsQlFwNXp0S2NlNFFtNWhyaThzNVJiM3NPb1JTNDc2WHlvUm5FNFcxU3lVckJhcWYzeGdyaG8yMkcxZnl6eVF0WmJ6VlNtTjViMA?oc=5
- [55] MetaX earnings call (Cailian Press, gnews redirect) — https://news.google.com/rss/articles/CBMiSEFVX3lxTE9zVW9KUVQwcVhpejdaT2NfcXJTVzQyVm9aa2FSSzlBb2xVMG5hSmplSmo2bHFVRFZfdV85V1Vhd3NaU3JYaURuMw?oc=5
- [56] Enflame Q3 report (Sina Finance, gnews redirect) — https://news.google.com/rss/articles/CBMibkFVX3lxTE45ZklzMWpXaXdWY3BfSTd6Nkltdk92bDFKS2F4WEF1cHVaR2Z2SDBoWGhvSkhjT19oYXBnMWQ0RkFxRlFWNEU0eGlDbzh4ZnhqX3NOeTlOQXVzcnNCdkNHXzhLY0RPQ29xUHZ1a2Nn?oc=5