Reporting period: Monday, November 9, 2026 to Friday, November 13, 2026 — 5 working days, 5 daily reports; plus supplementary checks on the release day (November 14) covering org pushes, merged-PR searches, and the image line, capturing wrap-up changes from 11-13 06:00 to 11-14 18:00


Part 1 · Brief

This section is the week’s curated highlights, kept to 800 words by default (799 words this issue); full entries, itemized lists, and sources are archived in Part 2.

Of the 64 repos this week, 19 to 22 had pushes; across the five issues there were 1,077 commits and 1,030 merges. The open-review backlog fluctuated between 252 and 343, ending the period at 272.

  • The 2.3 roadmap opens, targeting a 12-30 release: #176 sets three main lines, with vLLM 0.32 and SGLang 0.5.18 as the primary release versions; FEP-0163 FlagMega merged the same day, requiring a 20% compile-efficiency gain for the full Qwen3.8 model and GLM5.3 Flash decode throughput 5% above vLLM, and it must land in main by 12-07.
  • The image line is settled backend by backend: T-Head PPU 2.1.0 ships its first runtime image, ROCm 7.2.4 is recorded, Tianshu corex5.0.1 replaces 4.4.0, Ascend CANN 9.2.1 adds 910C support, and the release day adds a MUSA 5.3.0 base.
  • FlagTree 0.7.1 released in batches: the official release covers only ascend3.5 and xpu3.6; the rest remain a1, a2, or dev1; Intellifusion joins as ifwa:100, and Ascend 3.6.x is under review.
  • Operator production remains the main output: FlagGems merged 591 PRs this week, 57.4% of the whole org; KernelGen peaked in a single day at 87 for Ascend, 63 for Enflame, 89 for MetaX, and 22 for Hygon, with the operator table trimmed to 141.
  • Cambricon adds queue ordering: Torch-FL #809 routes 17 operators in the Qwen3 training step through CNNL and fixes the silent defects of USE_CAMBRICON not being compiled and Triton launching on an empty queue; runtime is still 1.86× that of torch_mlu.
  • Kunlunxin completes the 618-to-256 routing: after P800 measurements, 256 operators are routed to FlagGems; vLLM 0.32.0 was pulled within two hours of listing due to a py3.10 import failure.
  • vLLM 0.32 becomes a ten-vendor migration checklist: NVIDIA and T-Head have merged, eight are under review; 17 competition-track operators merged, with 9 more entering review in the early hours of 11-13.

This week’s assessment: 2.3 has moved from collecting proposals to filling in deliveries backend by backend; the risk lies not in API renames but in platform differences and release speed. The 272-item review backlog remains the bottleneck, with only three weeks left until 12-07.


Part 2 · Archive

Below is the full text for the week, with structure and conventions consistent with previous weekly reports; every judgment in the brief can find its itemized basis and sources here.

Sources: 5 issues of this publication’s daily FlagOS reports (11-09, 11-10, 11-11, 11-12, 11-13). The 11-09 issue covers a 72-hour window across the weekend (11-06 06:00 to 11-09 06:00); the other four issues are contiguous 24-hour windows. The factual behavior interval is 11-06 06:00 to 11-13 06:00. Release-day checks are based on GitHub org commit and merge searches, repo push lists, Release and Tag atomic sources, and build-infra directory and packaging-matrix measurements.

Editorial note: Organized along six dimensions: the week’s top news, versions and releases, component and code co-building, member units and vendor adaptation, ecosystem and community, and trend observations. Summed across their respective windows, the five daily reports give 1,077 commits and 1,030 merges; taking only the four full 24-hour windows (11-10 to 11-13) gives 816 commits and 777 merges. The number of repos with pushes per issue is 21, 22, 21, 20, and 19 respectively, and the org repo count is recorded as 62, 63, 63, 64, and 64 (FlagMRI-3D was created on 11-06 and FlagOmni on 11-09, with one further count change afterward).


I. Weekly Highlights

  • The 2.3 roadmap is published, targeting a December 30 release, with model compilation elevated to stand alongside operators and frameworks (11-12): community #176 published the 2.3 roadmap in the early hours of November 12, listing proposals, tracking tickets, and owners by SIG, with checkboxes serving as release acceptance evidence. The three main tracks are model and framework coverage, compiler and runtime performance, and developer experience and delivery; on the framework side, it requires completing vLLM 0.24 acceptance and designating vLLM 0.32 as the primary release version for the twelve-vendor matrix, with SGLang upgraded to 0.5.18 covering ten vendors. On the same day, FEP-0163 FlagMega was merged and marked as in implementation by its companion PR, with tracking ticket #173 taking over delivery; the FEP requires compiling the full Qwen3.8 and GLM5.3 Flash models into MegaKernel, validating on three backends—NVIDIA, AMD, and Iluvatar CoreX 5.0—with a 20% optimization target set for Qwen3.8 and GLM5.3 Flash decode throughput required to exceed vLLM by more than 5% on four H800s. feature/flagmega must be merged into main before the FlagTree release branch is cut on December 7, with release acceptance completed by December 25.

  • FlagTree 0.7.1 official releases roll out by backend in batches, with Intellifusion becoming the first new backend in the 2.3 cycle (11-12): Two official releases, 0.7.1+ascend3.5 (15:43) and 0.7.1+xpu3.6 (19:28), were published on the same day, with the Release notes containing only one line: “pypi 0.7.1+"; the remaining backends remain at a1 (hcu, tsingmicro), a2 (iluvatar, mthreads), or dev1 (ppu). Releasing by backend in batches under the same version number is standard FlagTree practice, but this time it runs directly up against the 12-07 branch-cut deadline: backends not released before the cut can only enter 2.3 with pre-release versions. On the same day, #1598 merged the Intellifusion backend, targeting ifwa:100 and Triton 3.8, including lowering for memory access and matmul, software pipelining, async copy, LHS-to-TMEM promotion, and a separate manual wheel workflow. The Ascend Triton 3.6.x port (#1637) is still under review and, if merged, will return from the triton_v3.5.x branch to main.

  • The 2.3.0 image line and version data were filled in backend by backend over the week, with the smallest unit of release preparation being one backend, one document (11-07 ~ 11-14): T-Head PPU 2.1.0 first received its base and runtime changelogs on 11-07, completed its first runtime image delivery on 11-11, and added cp312 vendor wheels for torchvision 0.25.0 and torchaudio 2.10.0; on 11-10, the 2.3.0 change date was recorded for AMD ROCm 7.2.4; on 11-12, Iluvatar corex5.0.1 fully replaced corex4.4.0, and on the same day the 910C base and flagcx packaging lines for Ascend CANN 9.2.1 were added; the release-day check also saw build-infra #1792 add the Moore Threads mthreads-musa5.3.0 base layer and #1789 add the ppu builder and daily line. One tag per PR, with version data and changelog landing in the same batch, was build-infra’s sole working method this week.

  • FlagGems merged 591 commits in one week, as KernelGen turns “writing operators” into “reviewing generated results” (11-06 ~ 11-12): The five-period window saw 123, 166, 149, 85, and 68 merges respectively, accounting for 57.4% of the organization’s 1,030 commits—the only single repo above half. Single-day peaks occurred at three vendors: Ascend with 87 (11-06, landed by the same author within one hour), Enflame with 63 (11-09, all boundary-correctness fixes for GCU300), MetaX with 89 (11-10, the entire foreach family rolled out), and Hygon with 22 (11-12, 20 foreach operators within 100 minutes). On 11-12, #8740 deleted 7 host-side operators that violated repository conventions, reducing the experimental operator table from 142 to 141—showing that while capacity is scaling up, a reverse cleanup has begun.

  • Cambricon reclaims 17 operators in the Qwen3 training step to native implementations and fixes a silent queue-ordering defect (11-10): Torch-FL #809 has the in-place arithmetic family, three fused CNNL entry points, four view operator categories, and zeros, sqrt, and div supplied directly from the code generator to Cambricon native implementations, so the training step no longer detours through FlagGems for these operators, while also fixing the fill_ lifecycle issue. More critically, there was a build defect: torch_bindings had never been compiled with USE_CAMBRICON, so the binding that tells Triton which cnrt queue to dispatch to simply did not exist—Triton was actually running on an empty queue with no ordering guarantee relative to CNNL kernels. Such defects raise no error; they only cause occasional wrong numbers. The PR itself states that the training step still takes 1.86× as long as torch_mlu, addressing only coverage and correctness, not performance parity.

  • Kunlunxin uses two-stage measurement to narrow 618 operators down to 256 routed ones, as vLLM 0.32.0 hits its first hard blocker on py3.10 (11-11): Torch-FL #806 took over the empty list left the previous day—#789 had disabled four breakpoints on the P800 and achieved end-to-end execution, but the allowlist was empty and the config file was byte-for-byte identical to before the change, on the grounds that the first route must be backed by measured data. The next day, #806 performed two-stage measurement on the P800, filtering 256 out of the 618-operator ceiling; the config changed from “no route 1985, Kunlunxin 57” to “no route 1729, Kunlunxin 57, FlagGems 256”; the route count went from 316 to 315 to 256 across three revisions, with the author shifting focus from “listing names” to “fixing the filtering criteria.” On the same day, build-infra added vLLM 0.32.0 to deps_app in the early morning, then recorded the version’s import failure under py3.10 and withdrew it within two hours.

  • vLLM 0.32 goes from a single line on the roadmap to a migration checklist for ten vendors (11-09 ~ 11-13): Two landmark adaptations merged within a single day on 11-12—NVIDIA #737 migrated KV cache initialization from a private function removed in 0.32 to the public allocate_kv_cache and bind_kv_cache, while T-Head #706 brought PPU-ZW810E onto 0.5.0-dev using the “0.32-empty” approach (deleting removed symbols and their call sites rather than stubbing them out). Since then, eight vendors are under review: MUSA (#715), Hygon (#711), MetaX (#693), Kunlunxin P800 (#726), Ascend Runner V2 (#703), Iluvatar (#674), Enflame (#675), and AMD W7900D (#659). Framework migration has entered the execution phase, and the risk has shifted from “whether to upgrade” to “who gets stuck first.”

  • The competition track enters its final stretch: 72 problem definitions, 17 implementations, and 9 submissions opened for review in the early hours, with judging starting after 11-13 (11-09 ~ 11-13): On 11-09, 72 SGLang competition operator definitions were merged at once, along with 23 missing definitions; on 11-10, 17 implementations tagged Track1 were merged, concentrated on block-scaled matrix multiplication for fp8 and mxfp8, quantization at three granularities, KV cache mxfp8 write-back, and AWQ dequantization; on 11-12, awq_gemm was merged; between 01:07 and 01:08 on 11-13, the same author opened 9 PRs in succession. The competition runs from August 17 to November 13, with acceptance completed within one month after it ends. The repository contains only definitions and submitted code, which does not mean they have passed the performance leaderboard; this report draws no conclusions on results.


II. Versions and Release Activity

2.3.0: One PR per Backend, Image Lines Expand from 23 to Around 25 This Week

Date: 2026-11-07 ~ 2026-11-14 Source: build-infra #1675 (PPU base) · build-infra #1747 (corex5.0.1) · build-infra #1792 (musa5.3.0 base) · build-infra repo

Backend line This week’s action Date
thead-ppu2.1.0 Added backend key and directly referenced vendor PAI-DSW training image (about 25 GB, SDK at /usr/local/PPU_SDK); first delivery of runtime changelog; first image delivery plus cp312 wheels for torchvision 0.25.0 and torchaudio 2.10.0; app layer still not connected 11-07 / 11-11
amd-rocm7.2.4 2.3.0 change date recorded on this line, the fifth backend line to land in 2.3.0 in the past week 11-10
ascend-cann9.2.1 Opened daily build line and added cp312 builder; added 910C base; added flagcx packaging line and disabled it for all channels to avoid duplicate wheels when sharing pins with 910B 11-10 / 11-12
iluvatar-corex5.0.1 Added backend and retired corex4.4.0 build layer; added driver user-space, libopenmpi3t64, flagcx 0.14.1 pinning; 16 commits from one vendor 11-12
nvidia-cuda13.3 Built vLLM plugin wheels in standard mode; native wheels carry no local label 11-09
metax 2.3.0 Added mcoplib and deep_gemm dependencies, then removed same day; logged withdrawal and rebuild and restored changelog 11-07
mthreads-musa5.3.0 Release-day check added base image and changelog date record 11-14
  • Recording method: changelog and version data proceed on a “one tag, one PR” basis. Of the 28 build-infra merges on 11-12, 12 were bot date backfills and 16 were manual changes; author tengqm accounted for 16 alone.
  • Validation scope: When #1742 recorded the T-Head PPU 2.1.0 base docs, the tool pulled built-in system package versions against a 23-backend matrix, and all 23 checks passed; this number shows the matrix itself is already stable, and what changes is the per-entry version data.
  • Release state: The community release/2.3 directory still contains only the release schedule; the manifest file is planned for generation at the 11-30 feature freeze. Images running ahead of the manifest is within the planned cadence.

FlagTree 0.7.1: Only Two Backends in the Official Release, the Rest Still Pre-release

Date: 2026-11-12 Source: FlagTree 0.7.1+ascend3.5 · FlagTree 0.7.1+xpu3.6 · FlagTree #1598 (Intellifusion) · FlagTree #1637 (Ascend 3.6.x, under review)

  • Release status: ascend3.5 and xpu3.6 are official releases (11-12 15:43 and 19:28); hcu3.6 remains at 0.7.1a1+hcu3.6 (entered PyPI on 11-06); tsingmicro3.6 remains at 0.7.1a1+tsingmicro3.6 (10-29); iluvatar3.6 and mthreads3.6 are 0.7.1a2; ppu3.6 is 0.7.1.dev1+ppu3.6.
  • New backend: ifwa:100 (Intellifusion) is not in the above tag table; it was merged this week via #1598 and uses Triton 3.8.

  • Packaging refactor: deb packages are now one package per backend per backend table; the Ascend rpm is built directly from the published wheel rather than rebuilt in this repo; these two follow the same approach as the build-infra hcu wheel release line.
  • Mainline version distribution: The repo README’s mainline matrix lists 15 vendor backends; the mainline uniformly uses Triton 3.6, with Ascend on the 3.5.x branch, Arm and CPU on 3.3.x, and Cambricon on 3.2.x.
  • Assessment: Releasing official versions per backend is not delay but the inevitable result of each backend passing its own CI; with three weeks until the 12-07 branch cut, backends not yet released can only enter 2.3 as pre-releases.

Tags and Releases: Zero New Official Release Objects, Three Dev Tags

Date: 2026-11-06 ~ 2026-11-13 Source: FlagGems tags · vllm-plugin-FL tags · FlagTree tags

  • New tags: FlagGems v5.5.0.dev1 (11-11 18:25), vllm-plugin-FL v0.5.0.dev0 (11-11 14:23), FlagTree 0.7.1+ascend3.5 and 0.7.1+xpu3.6 (11-12). 0.4.0-dev had already been forward-synced into 0.5.0-dev; vllm-plugin-FL jumped directly from 0.4.0 to the 0.5 dev line, in sync with the vLLM 0.32 target for 2.3.
  • Zero Releases: Five daily briefings checked the Release atom sources of 31 repos and the Tag atom sources of 8 repos day by day; there were no new Release objects all week. The latest official versions on the GitHub Releases pages remain the 09-28 batch (FlagGems v5.4.0, FlagCX v0.14.0, FlagScale v2.1.0, vllm-plugin-FL v0.3.0).
  • Component version lines: FlagGems’ official line remains 5.4.0; FlagQuantum’s baseline remains v0.3.1 from 10-16; Torch-FL self-reports v2.10.0; the remaining components remain at the 09-28 batch.
  • Assessment: Public versions silent, dev-line tags active, and image lines advancing per backend — this is the typical shape after the 2.2 release and before 2.3 GA on 12-30.

vLLM 0.32 and SGLang 0.5.18: From Main Release Versions to a Per-Vendor Migration Matrix

Date: 2026-11-09 ~ 2026-11-13 Source: vllm-plugin-FL #706 (PPU 0.32) · vllm-plugin-FL #737 (KV cache) · sglang-plugin-FL #170 (PPU 0.5.18)

  • Merged: NVIDIA KV cache migration (#737) and T-Head PPU-ZW810E (#706); on the SGLang side, Kunlunxin #155 and #169 upgrade the backend to 0.5.18, and T-Head #170 adds empty-runtime adaptation for 0.5.18; all are under review.
  • Eight under review: MUSA #715, Hygon #711, MetaX #693, Kunlunxin P800 #726, Ascend Runner V2 #703, Enflame #674, Enflame #675, AMD W7900D #659; #742 additionally clears the residual 0.28 paths in the V1 runner.
  • Failure record: Kunlunxin 0.32.0 fails to import on py3.10, the first known blocker after 2.3 required “0.32 as the main version”; platform differences are harder to handle than API renames.
  • Version floor: build-infra #1683 evaluated vLLM 0.32.0’s floors for torch and Python, paving the way for the subsequent matrix.

2.3 Timeline and This Week’s Milestones

Date: 2026-11-12 Source: community #176 (2.3 roadmap) · FEP-0163 text

  • Timeline: FEP merge deadline 11-13, feature freeze and rc0 11-30, rc1 12-07, rc2 12-14, code lock 12-25, GA 12-30; two additional action milestones are set: FlagTree release branch cut on 12-07 and release acceptance on 12-25.
  • This week’s completion: The roadmap and the first implementation-type proposal landed on the same day, indicating that 2.3 has moved from “collecting proposals” to “collecting while implementing”; proposals from the five SIGs — operators, networking, training, KernelGen, and chips — are still marked pending.
  • Deferred and new: The Hexagon experimental backend for edge-side llama.cpp is listed as a deferred candidate; the OS and packaging track proposes a self-extracting offline installer (FEP-139), starting with vLLM for the generic and NVIDIA backends.

III. Components and Collaborative Code Development

FlagGems: 591 commits, throughput and cleanup in parallel

Date: 2026-11-06 ~ 2026-11-12 Source: FlagGems repo · FlagGems #8740 (remove 7 operators) · FlagGems #5882 (int64 offset)

  • Scale: 123, 166, 149, 85, and 68 merged across five periods, totaling 591, or 57.4% of the organization-wide 1030; the 166 on 11-10 was the weekly peak, with 137 of them crammed into the two hours between 10:00 and 12:00.
  • KernelGen distribution: single-day peaks by vendor were Ascend 87 (11-06), Enflame 63 (11-09), MetaX 89 (11-10), and Hygon 22 (11-12); the 130 in the 11-11 period were MetaX 89, Hygon 20, NVIDIA 9, Moore Threads 9, and Kunlunxin 3, while the 64 in the 11-12 period shifted to a multi-vendor small-batch mix of MetaX 46, Moore Threads 9, NVIDIA 4, Kunlunxin 2, Enflame 2, and Ascend 1.
  • Correctness convergence: the 63 Enflame commits were almost all GCU300 boundary fixes (index_copy_’s clone falling within the grid.x limit, searchsorted returning int32 and skipping the out= write, nonzero_static coordinates corrupted by int64 lane arithmetic, cummin propagating NaN, one_hot rejecting out-of-range categories); Kunlunxin added non-atomic overwrite for hash_tensor and histogram; Iluvatar changed linalg_det to a single-warp-launched panel/blocked kernel; Tsingmicro fixed a cache_results import error.
  • Cleanup and hardening: #8740 deleted 7 host-side operators that violated repo conventions (expand_as, alias, make_dep_token, _reshape_alias, unflatten, t, _unsafe_view), dropping the experimental operator table from 142 to 141, and synchronized vendor specializations, tests, benchmarks, operators.yaml, and the Chinese and English docs; #5882 changed the large-tensor index offset to int64, fixing overflow above 2^31 elements; #8735 fixed a kernel launch argument mismatch on Triton 3.7 and above.
  • Engineering standards: added a full-width punctuation pre-commit hook, fixed an unused-variable warning in Moore Threads’ linalg_tensorinv, and cleaned up trailing commas and import order in the metax and mthreads operator files.

FlagTree: shifting from “adding backends” to “packaging per backend, gating per backend”

Date: 2026-11-06 ~ 2026-11-12 Source: FlagTree #1496 (FlagPrism connects MetaX) · FlagTree #1450 (MTHREADS SQMMA) · FlagTree #1592 (PH1 and SQMMA)

  • Weekly volume: twelve on 11-06, nine on 11-10, eight on 11-11, seven on 11-12, plus five merges verified on release day.
  • Compiler correctness: #1532 preserves TLE’s local pointer proofs in read/write lowering; #1531 preserves shared arena geometry across pointer and buffer views; #1401 lets mixed-precision SQMMA reuse the FP32 accumulator; #1592 fixes PH1’s warp specialization and SQMMA synchronization; #1450 lets the KV physical layout be shared across SQMMA views.
  • Performance switches: #1517 adds four optional switches to the compilation pipeline—prefetch hoisting, wgmma carry, index prefetch, and load merging—turning performance tuning into configurable items; #1063 selects prefetch depth by shape on the FlagGems-vllm side.
  • Backend details: Iluvatar added set_layout and bumped LLVM to 0.7.0 to fix the attention path; Enflame updated the backend plugin and isolated the LLVM cache by package revision number, avoiding cross-contamination when different revisions share a cache; on HCU, amdgcn code generation for tle.raw now goes through dcc instead of aillvm; Ascend #1637 ports the NPU backend to Triton 3.6.x (under review).
  • Packaging: #1417 ships one deb per backend, #1461 makes the Ascend rpm use the published wheel, #1382 makes the runtime depend on Python development headers, #1369 strips the debug sections from libtriton.so, and #1368 drops the bundled CUPTI static library.

Framework side: FlagGems-vllm, vllm-plugin-FL, and sglang-plugin-FL

Date: 2026-11-09 ~ 2026-11-12 Source: FlagGems-vllm #752 (Marlin MoE) · FlagGems-vllm #1032 (qsa 0.32) · vllm-plugin-FL repo

  • FlagGems-vllm: ten on 11-10, twelve on 11-11, twenty on 11-12; the 12th’s merges centered on SiliconFlow’s 14 (triton_ihc_post, GLM5Next pad_indexer_heads, triton_reshape_and_cache_flash, fused_indexer_k_norm, with the rest fixing KDA/GDN benchmarks and backend tuning), plus 1 from T-Head and 4 combined from Ascend and NVIDIA. 47 are under review, including 13 MetaX quantization and routing optimizations, Hygon MegaPreAttention, and NVIDIA Hopper MegaMHC/MegaGate.
  • T-Head Marlin MoE: #752 adds a unified fused_marlin_moe backend covering four weight formats—W4A16 INT4, W4A16 MXFP4, W8A16 INT8, and W8A16 FP8; the W8 path uses a 64-element K tile to reduce decode footprint, and MXFP4 scale conversion is folded into CUDA Graph capture.
  • Interface alignment: #1032 optimizes qsa_sparse_paged_attention and aligns with the vLLM 0.32.0 interface; the KV cache migration in #737 closed the gap left by #615, which changed only CI/config and missed the runtime; #1170 adds a registry-driven multi-backend install script, consolidating “pick dependencies by vendor” into one place.
  • Plugins and tests: #1135 adds a check—operators that are exported but not registered in the operator manifest are blocked by CI; #725 restores NVIDIA H20 end-to-end coverage on the replaced runner; #731 supports running two 4-card runners on a single 8-card machine.

FlagCX: Ascend VMM and multi-node CI connected; Iluvatar and Enflame under review

Date: 2026-11-07 ~ 2026-11-12 Source: FlagCX #706 (Ascend VMM and RMA) · FlagCX #714 (Iluvatar GDR/PXN, under review) · FlagCX #670 (CoreX unit tests)

  • Ascend: #706 implements CANN-side VMM-backed memory allocation, shareable handle exchange, flat and multi-segment mapping, and resource reclamation, completing SymMem’s full lifecycle; signal addresses still rejected by aclrtValueWrite after IPC import are switched to stream-ordered async copies.
  • Test scope: removed the single-node-only coverage for Ascend RMA and SymMem, switching to shared three-stage multi-node IPC, NET, and VMM runs, and made GTest fail outright when “all cases were skipped”—this matters more than the feature itself, preventing green CI with zero actual coverage.
  • Other backends: #670 adds a unit-test pipeline and CI image for Iluvatar CoreX; #702 restores Corex’s barrier initialization; #711 fixes a duplicate host-device helper definition in the Iluvatar build (which directly causes link failures); #709 enables Ascend Torch API tests and Python coverage; #697 lets UniRunner support fused collectives and DSL.
  • Under review: #712 adds an ECCL device API backend for Enflame; #714 adds a GDR device family and PXN capability hooks for Iluvatar, advertising memory registration through the IB adaptor and adding an IXCUDA peer-access query for topology discovery.

Torch-FL: rename, routing in practice, and two native paths

Date: 2026-11-08 ~ 2026-11-12 Source: Torch-FL #806 (618 to 256 routing) · Torch-FL #809 (Cambricon training step) · Torch-FL repo

  • Positioning and rename: Torch-FL is the repo formerly named PyTorch-Plugin-FL; the README now reads torch-fl, exposing a single flagos device that routes each operator among four implementation classes: local kernels, compiler kernels, vendor-native implementations, and explicit CPU fallback.
  • Routing: #806 routes 256 of 618 operators to FlagGems on P800; #819 adds fast-launch replay for FlagGems routing, amortizing repeated launch preparation; #814 makes TileOPs’ MetaX gating read the parser switch, which lets the Hygon path run.
  • Native paths: Cambricon #809 and #813 (softmax, zero_, and backward go native), #770 (eight operators made native plus batched matmul), #767 (CNNL code generator), #771 (silu goes through the activation kernel), #769 (strided gather stays on device), #772 (torch.compile tier); on the MUSA side, #821 fixes pip cache, the MPS gate, and mudnn cast, while #802 adds argmin and max.dim/min.dim on MPS.
  • Scale: fourteen on 11-10, seven on 11-11, three on 11-12; open issues peaked at 102.

FlagQuantum and FlagQAI: 107 commits to the quantum stack in one week

Date: 2026-11-06 ~ 2026-11-14 Source: FlagQuantum #1150 (Hubbard ring) · FlagQuantum #1145 (216 native programs) · FlagQAI repo

  • Scale: 16, 5, 10, 6, and 70 commits across five periods, totaling 107, with 63 merged; the 70 commits on 11-12 ranked first organization-wide for a single day, and its 29 merges ranked second.
  • Error-correction chain: from integrating Stim and PyMatching on 11-07, to six measurement record formats and a weighted matching workflow on 11-09, to converting measurement files into detector records and streaming out six native formats on 11-11—within one week, “integrating a decoder” advanced to “custom record formats and matching strategies.”
  • Fermionic models: three model classes were completed in a single day on 11-12—spin-periodic Hubbard rings with UCCSD training, spinless t-V chains and rings, and molecular UCCSD (native FCIDUMP input); diagonal fermionic parity symmetry and commuting Pauli symmetries were also discovered, advancing training into parity sectors.
  • De-dependency and synthesis: #1145 adds native implementations for all 216 quantum programs (five categories: Algorithms, Applications, Community, Functions, Tutorials); the chemistry side now reads MolecularData’s HDF5 directly instead of depending on OpenFermion, forming two paths alongside its optional conversion interface under review.
  • FlagQAI: the QuantumAI Agent repo (TypeScript) had 8 merges verified on release day, centered on web-side saved-state probability display, SDK synthesis attribution and per-gate-interval navigation, and CLI session restoration.

build-infra: image lines, the distribution layer, and the wheel directory

Date: 2026-11-09 ~ 2026-11-14 Source: build-infra #1687 (TE-FL pure-Python wheel) · build-infra #1705 (wheel directory refactor) · build-infra #1721 (CANN 9.2.1 daily line)

  • Distribution layer: #1687 adds a pure-Python wheel release line for TransformerEngine-FL, building a py3-none-any package with TE_FL_SKIP_CUDA=1 and pushing a single build to the PyPI and daily indexes of all non-NVIDIA vendors, with versions like 2.17.0+fl.0.3.0; NVIDIA is excluded because its native path is upstream TE itself. This is the first time “one line, many vendors” has landed as an actual artifact.
  • Wheel governance: #1705 changes the wheel-info directory to “one record per wheel, with platform as the row identity”; #1697 specifies that NVIDIA native wheels carry no local label; #1696 lets a backend select only one native record, avoiding wrong package selection when multiple backends share; #1701 makes repack auditing carry extra-requested dependencies and strip torch and triton from them.
  • Build lines: #1721 opens the daily build line for Ascend CANN 9.2.1, adds a cp312 builder and target options, and preinstalls libzstd for the 9.2.1 LLVM tree; after #1687 came #1690’s transformer_engine directory entry and #1688’s clone hardening.
  • Bug fixes: #1724 fixes the Ascend builder’s ref parameter being read as a pip version number; #1720 makes official releases produce clean Ascend wheel version numbers; #1686 requires the flagcx wheel to be checked out by tag rather than the resolved sha.

Other components: operator libraries, training evaluation, and toolchain

Date: 2026-11-09 ~ 2026-11-13 Source: FlagBLAS #161 (Ascend L1) · FlagOmni repo · FlagTrain repo

  • FlagOmni: repo created on 11-09, positioned as a Triton-implemented omnimodal generation operator library (video, image, audio), complementary to FlagGems—the latter covers general LLM inference operators, the former covers generative workloads; LightX2V is marked in development, FastVideo and xDiT are planned, with 2 commits on the first day, Apache 2.0.
  • FlagMRI-3D: repo created on 11-06, 1 commit that day, README not yet uploaded, positioning to be observed.
  • FlagBLAS: #161 adds the BLAS L1 layer for Ascend and optimizes L2 kernel performance, meaning this backend is extending from inference operators toward numerical foundation libraries.
  • FlagTrain, FlagCV, and optest: FlagTrain 6 commits (DeepSpeed’s FP8 matmul Triton kernel, NVIDIA fused dual activation, four baselines), FlagCV 4 commits (evaluation candidates dispatched by operator), optest 1 commit plus FlagTrain support; the three bring the new libraries into a unified evaluation loop, with no external release yet.
  • FlagOS-Compressor and FlagSparse: Compressor 2 commits fixing MiMo calibration input retention and unifying GPTQ inverse Hessian across backends; Sparse 2 commits merging PRs from the external AlphaSparse project plus one MUSA sparse optimization.
  • FlagPrism, docs, and OpenCourse: FlagPrism #19 connects the Hygon HIP debugger and DTK profiler, #20 turns vendor adaptation requirements into checkable items; docs #580 adds about 390 lines each in Chinese and English, rewriting the adaptation spec into 6 pure virtual interfaces and a six-step collection flow; OpenCourse #15 adds the FlagCompressor Quantization Lab (merged 11-13 10:17).
  • Trident and FlagTensor: Trident #4 introduces a Monad intermediate layer for computation composition, #8 is compatible with old and new TVM FFI header layouts; FlagTensor 5 commits fix PPU device and baseline detection and merge PPU work into the main branch; FlagCICD #19 updates references after the repo rename (formerly FlagOps).

IV. Member Units and Vendor Adaptation

T-Head: “Key, Image, Delivery” in Two Days, Then Framework Adaptation

Date: 2026-11-06 ~ 2026-11-12 Source: build-infra #1675 (PPU base) · build-infra #1738 (PPU runtime closure) · vllm-plugin-FL #706 (PPU on 0.32)

  • Distribution constraint: T-Head does not provide a PPU SDK installer package, nor does it place source code in the filestore; the SDK exists only inside its PAI-DSW training image (about 25 GB, /usr/local/PPU_SDK). The base layer can therefore only reference that image directly, and changing versions means changing a single FROM line; the backend’s self-hosted runner is registered at the organization level and is that PAI-DSW node itself. The node has no GitHub egress, images are built directly on the node, and the persistent disk is CPFS 685 GB.
  • Delivery chain: On 11-07 the backend key thead-ppu2.1.0 and the runtime changelog landed; on 11-11 the runtime image was delivered for the first time, with a closure of torch 2.10.0, numpy 1.26.4, FlagGems 5.4.0, default compiler FlagTree 0.7.0+ppu3.6, and the vendor triton 3.5.0 installed at /opt/triton as a fallback; as of 11-12 the app layer is still not integrated.
  • Known limitation: FlagGems’ release mapping has no PPU backend, so the image does not carry a cpp build of the FlagGems wheel, and operator capability depends on the FlagTree compilation path—a substantive gap versus mature lines such as NVIDIA and AMD.
  • Frameworks and operators: #706 uses “0.32-empty” to bring PPU-ZW810E onto vLLM 0.32; #752 adds unified fused_marlin_moe; #8508 optimizes LayerNorm row scheduling on PPU while retaining FP32 statistics; #1148 adds a thead install entry that directly installs the vendor vLLM wheel; Log-Linear Attention in FlagAttention #134 passes 4 correctness cases on a single PPU-ZW810E card (under review).
  • Terminology: PPU in this section belongs to T-Head’s AI accelerator card product line and is unrelated to XuanTie’s (DAMO) RISC-V processor line; thead is the brand and upstream branch name T-Head continues to use.

Ascend: 87 Plus 41 Operator Commits, CANN 9.2.1 Adds 910C, Long-Sequence Attention Optimization

Date: 2026-11-06 ~ 2026-11-12 Source: FlagGems #8842 (D256 attention) · build-infra #1760 (CANN 9.2.1 910C) · FlagGems-vllm #967 (sparse FlashAttention)

  • Operators: On 11-06, 87 unique operators were merged in one batch (math, normalization, pooling, and copy categories, all Triton, auto-generated by KernelGen); on 11-09 another 41 commits were merged (kl_div, floor, inverse, etc.); on 11-12, three commits.
  • Performance: #8842 optimizes FlashAttention forward and backward for causal FP16/BF16 workloads with head dimension 256; the target workload is BF16, batch 4, sequence 8192, 6 query heads to 1 KV head, no dropout. The approach includes compile-time specialization of shapes and strides, enlarged UB-safe tiling, hoisting the forward PV matrix multiply ahead of the independent softmax update, splitting dK and dV to reduce L0C pressure, and retaining FP32 softmax normalization and gradient accumulation. #8567 and #8573 take the KMCompiler route to hand-optimize linear and log_sigmoid_backward.
  • Images: CANN 9.2.1 covered only 910B on 11-03; on 11-12 the 910C base was added; #1761 synchronously adds the flagcx packaging line and disables it for all channels, because the two 910 generations share the single pin flagcx 0.14.1+cann9.2.1—enabling it would produce duplicate wheels, while omitting the line would cause matrix drift and set-matrix failures across all channels.
  • Three layers in parallel: Operators (FlagGems), numerical libraries (FlagBLAS #161’s L1), and communication (FlagCX #700’s IPC and GPU MR) advanced in the same week.
  • Compilation: FlagTree 0.7.1+ascend3.5 stable has been released; if the Triton 3.6.x port under review in #1637 is merged, Ascend will return from the 3.5.x branch to mainline.

Hygon: 62 Dedicated Operators Across Three Batches in One Week, foreach and bmm Both Filled In

Date: 2026-11-09 ~ 2026-11-12 Source: FlagGems #8963 (bmm trio) · FlagGems #9011 (or_tensor) · FlagTree #1597 (BW1100 test label)

  • Volume: The 20 dedicated operators from 11-09 moved from all-under-review to merged; another 20 were merged on 11-10; on 11-12 alone there were 27 commits tagged Hygon, of which 22 were FlagGems (20 foreach-family plus linalg_ldl_factor_ex and or_tensor), 1 bmm/addmm/baddbmm optimization, 1 test-scope commit, 2 build-infra, and 1 FlagGems-vllm.
  • Time density: The 20 foreach operators were merged consecutively within 100 minutes, from 17:38 to 19:21, with author bwbwzzz submitting 22 commits alone; foreach is a high-frequency entry point for optimizers and quantized training scripts, and filling it in will noticeably reduce Hygon fallbacks on torch optimizer paths.
  • Compiler and CI: FlagTree in the same batch adds Hygon FP64 matrix multiply-add support, so the FP64 path no longer relies solely on emulation; the hcu3.6 and BW1100 combination adds Qwen test labels, bringing the new-generation chip into routine regression.
  • Debugging surface: FlagPrism #19 hooks up the HIP debugger and DTK profiler; vendor documentation states that the backend is called DCU in code, the SDK is called DTK, the compiler backend is called hcu, with target gfx936 and warp_size 64.
  • Under review: FlashAttention dropout (#8969) and the FlashAttention, sum, and clone optimizations on master (#8922, #8921, #8919) remain queued; on the inference side, MegaPreAttention (#1189) is under review.

Cambricon: From “It Runs” to “Zero Fallback,” Then Queue Ordering

Date: 2026-11-08 ~ 2026-11-12 Source: Torch-FL #809 · Torch-FL #813 · Torch-FL #770

  • Integration form: MLU enters Torch-FL as a vendor-native platform, adding the libcnrt runtime layer and a torch.mlu shim, running on stock torch 2.10.0+cpu, with end-to-end testing on eight MLU590-M9 cards; on 11-05 the cambricon CI lane had been withdrawn on the grounds that no runner could host it.
  • Coverage: #770 natively serves eight more operators and changes matrix multiply to batched processing, #771 routes silu directly to the CNNL activation kernel, #769 and #768 keep strided gather, softmax, and negate on device; #813 on 11-12 routes softmax, zero_, and backward through native implementations as well.
  • Queue defect: torch_bindings was not compiled with USE_CAMBRICON, so Triton launched on an empty cnrt queue with no ordering guarantee relative to CNNL kernels; #809 also fixes the fill_ lifetime issue.
  • Scope: Training step time remains 1.86× that of torch_mlu; the PR explicitly covers only coverage and correctness, not performance parity; both Cambricon neuware 4.4.3 and 4.7.2 image layers fall within 2.3.0 coverage, and FlagTree’s Cambricon backend remains on Triton 3.2.x, the oldest version in the matrix.

Kunlunxin: Four Breakpoints Closed, 256 Routes Landed, 0.32 Still Stuck on py3.10

Date: 2026-11-09 ~ 2026-11-12 Source: Torch-FL #789 (four breakpoints) · Torch-FL #806 (256 routes) · build-infra #1737 (withdraw 0.32.0)

  • Breakpoints: The Kunlunxin wheel was in fact already compiled into FlagGems, but the backend configuration routed zero operators to it, so all covered operators fell back to CPU on Kunlunxin; the four breakpoints were the config generator’s inability to route (the documented escape hatch being a silent no-op), the absence of a registration file for the FlagGems slot, the missing runtime glue needed to run FlagGems under CPU-only torch, and the native-versus-FlagGems priority configuration not naming the platform. #789 closes all of them and verifies end-to-end on P800, but the allowlist remains empty.
  • Routing: #806 uses two-stage empirical screening to select 256 routes, changing the FlagGems channel in the configuration from 0 to 256; another PR under review routes all 316 operators empirically measured on P800.
  • Operator gap-filling: 25 KernelGen commits plus several fixes merged during the week—atomic-free overwrites for hash_tensor and histogram, .out precision for adaptive_avg_pool2d, foreach_lgamma parsing, special-bessel and thnn recurrent backward coverage, weight_norm backward switched to flat addressing, an XPU-safe kernel for quantized_max_pool3d, and mH non-contiguous input precision; 8 commits covering linear algebra, sampling, and pointwise paths are under review.
  • 0.32 blocker: vLLM 0.32.0 fails to import on py3.10 and was withdrawn within two hours of listing, with no recovery record seen this period; vllm-plugin-FL #726 (P800 on 0.32) remains under review.

MetaX: foreach Family Rolled Out Across the Board, Work Begins on Shared Code

Date: 2026-11-09 ~ 2026-11-12 Source: FlagGems #8822 (foreach_tanh_) · FlagGems #8414 (layernorm) · FlagTree #1496

  • Volume: Two commits in the 11-09 period (addmm, baddbmm with router GEMM specialization, cross_attention), 89 KernelGen commits in the 11-10 period, 50 in the 11-11 period (46 of them KernelGen), and 46 KernelGen commits in the 11-12 period plus three structural changes.
  • Whole-family completion: The 89 commits cover tensor, scalar, list, and in-place variants of the add, mul, and div families, plus maximum, minimum, log, log2, log10, log1p, abs, acos, neg, and addcdiv; 11-11 added sin, tan, sqrt, rsqrt, sigmoid, trunc, reciprocal, pow, sub, neg, tanh, and special functions, linear algebra, and sampling categories.
  • Structural changes: #8414 adds and optimizes layernorm, #8591 improves tuning and parameter handling for small-size FP32 Linear, #7981 widens index offsets before stride multiplication to avoid overflow, and #8857 fixes _metax operator code style—none of the four is bulk single-operator rollout, but rather shared defects and tuning.
  • Debugging and build: FlagTree #1496 enables FlagPrism support for MetaX, landing the code on the same day the roadmap called for “expanding to MetaX”; build-infra #1745 pre-provisions the MACA SDK slice for the metax builder, fixing the nightly metax daily failures after 0.7.1 enabled FlagPrism by default.

Moore Threads: Second Tier for KernelGen, Day-0 Goes Through the Native Stack Instead

Date: 2026-11-06 ~ 2026-11-12 Source: FlagGems #8539 (FP8 TopK) · vllm-plugin-FL #715 (MUSA 0.32, under review) · Moore Threads Kimi K3.1 Pro report

  • Operators: Nine commits on 11-06, nine on 11-09 plus one baddbmm backend optimization, nine on 11-11 (igamma, kl_div, uniform, two special polynomial types, fractional max pooling backward, 3D replication-pad backward, moving-average quantization helper, fake_quantize quantization cache mask), and five dedicated operators on 11-12 still under review.
  • Performance and testing: #8539 optimizes FP8 TopK using exact-codebook histogram selection; #1089 removes the synchronization wait in the indexer_k_quant_and_cache test, preventing the test from masking async defects with synchronization; FlagTree #1624 updates the ref.txt values for qwen3.6.
  • Frameworks: vllm-plugin-FL #715 pushes MUSA adaptation to vLLM 0.32.0 (under review); no new records on the SGLang side.
  • Independent progress: On 11-06 Moore Threads announced Day-0 adaptation of Kimi K3.1 Pro based on MTT S5000 and the MUSA software stack, via SGLang-MUSA carrying upstream, the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler; the announcement does not mention FlagOS or FlagGems, making this independent progress by a member unit that does not count toward FlagOS stack results.

Iluvatar CoreX: corex5.0.1 Fully Replaces 4.4.0, CI Advances on Three Fronts

Date: 2026-11-09 ~ 2026-11-12 Source: build-infra #1746 (replace base 4.4.0) · build-infra #1747 (corex5.0.1 backend) · FlagScale #1322 (BI training CI)

  • Version generational change: 16 build-infra commits make corex5.0.1 the default for both the base and runtime lines, retiring the corex4.4.0 build layer; they add the driver user space, libopenmpi3t64, and the flagcx 0.14.1 pin, and backfill the changelog item by item. This is not an incremental patch but a wholesale forward move of the SDK baseline by one version.
  • CI three fronts: FlagScale #1322 brings BI training into the shared CI framework, Megatron-LM-FL #213 adds the BI-V150 pipeline and development image, and FlagCX #711 fixes duplicate host-device helper definitions in the Iluvatar build.
  • Operators and compilation: Ten BI-V150 dedicated operators on 11-06 (three float_power variants, avg_pool1d, expand_copy, greater_, etc.); #8310 on 11-09 changes linalg_det to a single-warp launch; FlagTree adds set_layout and upgrades LLVM 0.7.0.

Enflame, Tsingmicro, Sunrise, and Intellifusion: Quality Phase and Small-Step Progress

Date: 2026-11-09 ~ 2026-11-12 Source: FlagGems #8310 (single-warp fix) · FlagGems #8785 (GCU300) · FlagTree #1598 (Intellifusion)

  • Enflame: All 63 commits on 11-09 carry KernelGen and Enflame tags, with titles pointing directly at GCU300 hardware behavior—correctness convergence under boundary conditions rather than new operators; #8781 and #8783 on 11-11 fix cummaxmin backward and float_power tensor-scalar compilation issues; #8785 and #8968 on 11-12 fix the foreach family and asin_, erfc_; FlagCX #712’s ECCL device API is under review.
  • Tsingmicro: #8625 on 11-09 fixes a cache_results import error; on 11-11 the tsm260610 backend is upgraded to FlagTree 0.7.0 plus tsingmicro3.6; sglang-plugin-FL #133 makes txda tolerate the qwen3_asr registration differences in transformers 5.15 and above and fixes a token_to_kv_pool leak in multi-segment prefill; FlagGems #8758 remains under review, and the FlagTree side is at 0.7.1a1.
  • Sunrise: FlagTree #1623 updates the Sunrise backend plugin, and #1625 isolates the LLVM cache by package revision number; #1488 syncs the XPU cluster source and upgrades the SDNN object to v0.3.6.9.1.
  • Intellifusion: FlagTree #1598 merges the Intellifusion backend (ifwa:100, Triton 3.8), including memory access and matmul lowering, software pipelining, async copy, LHS-to-TMEM promotion, IFWA Toolkit build checks, a CModel example, and a separate manual wheel workflow; it is not yet registered in the README mainline matrix, making it an addition rather than an expansion.

BAAI and the Industry Side: MDC 2026 Dated, T-Head IPO, Two Models at Day-0

Date: 2026-11-06 ~ 2026-11-12 Source: Shanghai Securities News (MDC 2026) · Operator talent report (Sohu) · Shanghai open source system (Sina Shanghai)

  • MDC 2026: The Moore Threads developer conference is set for December 29–30 in Beijing, with more than 10 new products expected, the S6000 intelligent compute card and the “Lushan” series based on the “Huagang” architecture as focal points, 26 forums, and a 3,000-square-meter exhibition area; the news broke on the evening of 11-11 and is an industry development outside the window.
  • Capital side: 36Kr discussed the motivations for a T-Head independent IPO on 11-12, a capital-level matter with no observed impact on this week’s code and image cadence; secondary-market news such as Enflame’s 76% share price rise over seven days and market cap above RMB 240 billion is uniformly excluded.
  • Model Day-0: On 11-11 MiniMax open-sourced the image generation model H3-Image (officially claimed to generate 4K images in 10 seconds), and both MetaX’s Xiyun C series and Moore Threads claimed same-day adaptation; both used “first” phrasing, cannot corroborate each other, and no FlagOS component participation is seen; the Kimi K3.1 Pro Day-0 on 11-06 ran on the MUSA native stack and likewise does not count toward FlagOS results.
  • Industry judgment: A commentary centered on the operator talent gap cites the judgment of Liu Guang, head of the BAAI systems research group, at the 2026 BAAI Conference—operator development is the hardest, most foundational work in AI infrastructure, and talent scarcity has reached a critical point; the competition mentioned is China Telecom’s Xingchen Cup (177 universities, 539 teams, 28 advancing, with sparse FlashAttention backward as the final problem), which is not the same event as FlagOS’s SGLang track, and this report does not conflate them.
  • Regional ecosystem: In December 2025 Shanghai issued an implementation plan to strengthen the open source system, targeting 1 to 2 open source communities with international influence, 100 open source commercialization enterprises, and more than 3 million developers by 2027; on October 30 the Open Source China headquarters settled in Zhangjiang, Pudong, and the Moli Fangzhou base opened the same day, with the compute side already adapted to Ascend, MetaX, and Iluvatar CoreX.

V. Ecosystem and Community

Competition Track 1: From Definition Closure to Implementation Merge, Awards Begin After the Finale

Date: 2026-11-09 ~ 2026-11-13 Source: FlagGems-sglang #329 (72 definitions) · FlagGems-sglang #324 (awq_gemm) · FlagGems-sglang repo

  • Definition side: #329 on 11-09 added 72 SGLang competition operator definitions in one go, #326 added 23 more, #325 added competition batch labels, and #330 and #328 clarified the contracts for quantization, routing, ep_gather, and rmsnorm_scale; 17 implementations were merged on 11-10, one day after the definitions.
  • Implementation side: seventeen merges on 11-09 (grouped_gemm, four rmsnorm variants in the gemma series, kda_gate_chunk_cumsum, chunk_kda_scaled_dot_kkt, swiglu_oai_split, etc.); seventeen on 11-10 (fp8 and mxfp8 block-scaled matrix multiplication, three quantization granularities — per-tensor, per-token, and per-token-group — KV cache mxfp8 write-back, AWQ dequantization); awq_gemm merged on 11-12; 9 more merged in the early hours of 11-13.
  • Evaluation alignment: #336 merged SGLang’s native baseline workloads for reuse by optest, and #342 then migrated the aligned baseline into optest, so competition results and the evaluation tooling share the same baseline.
  • Caveat: the repository only reflects code being merged, not that performance leaderboards have been passed; the competition runs from August 17 to November 13, with acceptance completed within one month after it ends.

Community Governance: FEP-0163 Merged and the 2.3 Proposal Matrix

Date: 2026-11-12 Source: community #163 (FEP-0163) · community #173 (tracking issue) · community #176 (roadmap)

  • Lifecycle: FEP-0163 was merged on 11-12, after which a docs commit marked the implementation status as in progress (#174), and #175 then removed the non-standard implementation-status field and restored the template-required Status: Provisional, with implementation progress consolidated under tracking issue #173.
  • Compiler-side proposals: besides FlagMega, six proposals await review: TLE distributed primitives and communication-computation fusion (extended to Iluvatar, Hygon, Kunlunxin, and DAMO), TLE Raw compiling and linking vendor device code on five backends, FLIR optimization for Ascend and MetaX, FlagTree compilation optimization across twenty operator-and-device combinations, and FlagTree platform enablement plus Triton 3.6/3.7 compatibility.
  • Framework-side proposals: Torch-FL extended to seven accelerators with evaluation of FSDP2 optimization, an Iluvatar edition of Torch-FL, three-platform training for Megatron-LM-FL, TransformerEngine migrating from a fork back upstream with a standalone plugin, three quantization paths (GPTQ/AWQ/AutoRound) in FlagOS-Compressor, and LightX2V integration with Qwen-Image-2.1, MiniMax H3, and Wan 2.1/2.2.
  • FlagPrism expansion: 2.3 requires FlagPrism to extend from source-and-IR correlated debugging to Enflame, MetaX, Kunlunxin, T-Head, and Hygon, providing execution-data export and an optional instrumented build.

Teaching and Documentation: Quantization Lab Enters the Course, Adaptation Spec Becomes an Interface Contract

Date: 2026-11-10 ~ 2026-11-13 Source: OpenCourse #15 (quantization lab) · docs #580 (FlagPrism spec) · FlagPrism repo

  • Course: OpenCourse #15 adds the FlagCompressor Quantization Lab, turning the quantization workflow into a course lab and merging it into the teaching-materials branch, forming a teaching entry point alongside the FlagOS-Compressor component; the same repo had 4 other merges in the window (including #11 updating the hook documentation).
  • Spec: docs #580 rewrites FlagPrism’s adaptation requirements from a one-pager into a full specification, structured in four parts: goals, Profiler adaptation interfaces, Debugger integration, and zero impact when disabled. It clarifies that of the three data categories — basic instrumentation info, numerical summaries, and chip performance info — only the third depends on the vendor Profiler; it defines 6 pure virtual interfaces and fixes runtime collection into six steps: generate plan, configure collector, start/stop operators, flush, stop, and import. Onboarding a new chip to FlagPrism goes from modifying code to implementing a set of interfaces.
  • Other documentation items: the docs repo also had a ModelScope model-list bot and bot auto-updates in the window, which do not constitute project activity.

Channel Check: Component Names Score Zero Hits for Consecutive Windows

Date: 2026-11-09 ~ 2026-11-12 Source: Google News RSS · CSDN FlagOS column · BAAI Hub

  • Component names: seven Chinese queries (FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, KernelGen) plus four English queries (FlagOS, FlagGems, BAAI FlagOS, flagos-ai) returned zero hits across all five windows — now the thirteenth or more consecutive quiet window.
  • Hit composition: candidates across the 24 queries ranged from 9 to 38 items, all from member organizations and industry terms; Enflame was the largest source in every window (stock price, Dragon-Tiger list, shareholder count, market cap) and was excluded throughout.
  • Channel side: the latest CSDN FlagOS column post dates to 2026-08-28, with no new articles in the window; the BAAI Hub homepage featured model evaluation, embodied AI, and quantum computing explainers, with no FlagOS project activity; HN Algolia hits were all tokenization mismatches.
  • Assessment: on GitHub, 19 to 22 repos pushed and 817 to 1030 merges per week (by window definition), in stark contrast to zero component-level hits on the news side. FlagOS’s external narrative still relies on the CSDN column, BAAI Hub, and conferences; third-party media coverage of component-level progress has yet to take shape.

VI. Trend Observations

  • 2.3 release preparation has shifted down to “backend × component”: FlagTree gating by backend, 2.3.0 adding image docs per backend, and vLLM 0.32 closing out per vendor — none of these are one-shot actions. This means that until 12-30, the weekly report’s unit of observation is no longer the version number but “whether a given backend has been gated on a given component.”
  • Model compilation has become a third mainline alongside operators and frameworks: FEP-0163 sets two verifiable targets, 20% and 5%, tied to Qwen3.8, GLM5.3 Flash, and three backends. The previous mainlines were operator coverage and backend adaptation; from now on there is an additional layer — “whether whole-model compilation meets the target” — which is also the most uncertain item before the 12-07 cutoff.
  • With operator throughput peaking, focus shifts to correctness and clearing the books: all 63 Enflame merges are GCU300 boundary fixes, Kunlunxin added non-atomic overwrite, Iluvatar switched to single-warp launch, FlagGems removed 7 non-compliant operators and cut the experimental table from 142 to 141, and another 11 merges converged test drift. This work generates no news, but it determines whether “which is right, the vendor branch or mainline” can be decided automatically.
  • Version-migration risk has shifted from APIs to platform differences: Kunlunxin’s vLLM 0.32.0 failed to import on py3.10 and was pulled two hours after listing; T-Head used “0.32-empty” to delete symbols outright rather than stub them. For the same upstream version, each vendor must recompute against its own Python and driver stack, with the hidden cost of maintaining two validation environments at once.
  • Review remains the sole bottleneck: the review backlog stands at 252, 343, 317, and 272 items, consistently above half the current merge volume; 120 new items were opened on 11-12 alone. KernelGen has scaled up the code-writing end, but merge speed is limited by review and hardware validation environments — which explains why the day of “less writing, more deciding” matters more.
  • The distribution layer is converging; the gap remains in operator artifacts: TransformerEngine-FL ships a single pure-Python wheel for all non-NVIDIA vendors, FlagTree’s deb is one package per backend, and the rpm pulls published wheels directly; but T-Head PPU still lacks a cpp wheel because the FlagGems release mapping omits that backend, so operator capability can only go through the compilation path. Once distribution is unified, the next gap is “whose operators have prebuilt artifacts.”
  • The quantum stack moves noticeably faster than other components, but should not be read as release-grade: 107 commits and 63 merges in a week advanced Hubbard and t-V condensed-matter models, parity symmetry, and native implementations of 216 programs simultaneously; the value of this work is making simulators trainable and verifiable, but it remains far from productization, and the AI for Science section in 2.3 is still marked TBD.

Appendix: Sources and Verification

Source List

Date Daily Report Research Window Nature
11-09 (Monday) FlagOS Community Daily Update 11-06 06:00 ~ 11-09 06:00, 72 hours Covers Saturday and Sunday, continuing from the previous issue’s cutoff
11-10 (Tuesday) FlagOS Community Daily Update 11-09 06:00 ~ 11-10 06:00, 24 hours Back-to-back
11-11 (Wednesday) FlagOS Community Daily Update 11-10 06:00 ~ 11-11 06:00, 24 hours Back-to-back
11-12 (Thursday) FlagOS Community Daily Update 11-11 06:00 ~ 11-12 06:00, 24 hours Back-to-back
11-13 (Friday) FlagOS Community Daily Update 11-12 06:00 ~ 11-13 06:00, 24 hours Final period
11-14 (Saturday) Release-Day Verification 11-13 06:00 ~ 11-14 18:00 Supplementary verification of commits, merges, open PRs, and the mirror line

Verification Notes

  • GitHub side: All five daily reports performed push verification across every repository in the flagos-ai org, cross-validated with three independent queries: commit search, merged PR search, and open PR search; the 11-13 issue additionally re-checked the Release atom source for 31 repos and the Tag atom source for 8 repos day by day, finding zero new Release objects for the entire week. Branch attribution and default-branch commits were separately re-verified at the atom level to avoid misrecording PR branches or gh-pages deployments as default-branch activity.
  • Release-day verification: Within the window from 11-13 06:00 to 11-14 18:00, the org had 130 commits, 185 merged PRs, and 227 open PRs with updates; merged counts by repository were FlagGems 64, FlagQAI 8, build-infra 6, FlagTree 5, FlagQuantum 5, and FlagGems-vllm 5. During this period, build-infra added mthreads-musa5.3.0 base and ppu builder daily lines, Torch-FL #824 added an eager inference path with a native FlagGems host, and FlagTree #1450 made the KV physical layout shared across SQMMA views.
  • Versions and dates: The layer names, version numbers, and dates for 2.3.0 are taken from the build-infra PR body and changelog; the 2.3 schedule is taken from the community roadmap and release timetable; FlagTree backend tags were re-verified against the tags atom source, with only two official releases: ascend3.5 and xpu3.6.
  • Vendor activity: Entries for member organizations are based on actual commits and PRs in flagos-ai repositories; industry-side figures are taken from public reports with the nature of the source noted, and earnings, awards, listings, and stock-price content are explicitly labeled as company-side activity and not counted as project progress.
  • Scope note: The 11-09 issue covers a 72-hour window and cannot be directly added to or compared with the other four 24-hour windows; when giving totals, this report lists both the five-day cumulative figures (1077 commits, 1030 merges) and the four full 24-hour windows (816 commits, 777 merges). Competition entries record only commits landing in the repository, without performance results.