Research window: 2026-11-05 04:30 ~ 2026-11-06 04:30 (Beijing time, past 24 hours; directly contiguous with the previous issue’s cutoff at 11-05 04:30).

Sources: GitHub key repository and organization push verification (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler; repository-by-repository verification of in-window commits, PRs, Releases, and tags for key repositories, with branch-level atom fallback; upstream forks filtered by RISC-V / RVV / XuanTie keywords), ruyi package manager Releases, packages-index, openRuyi packaging repository SPECS and merged PRs, RuyiSDK developer community forum, ruyisdk WeChat article repository (biweekly progress issue 080), ruyiai-daily daily intelligence archive, three-page snapshot of the RuyiAI official homepage, openRuyi distribution news page, Google News RSS targeted queries in Chinese and English (via proxy), arXiv search.

This Issue’s Index

  • Today’s Highlights: triton-riscv merged into Triton 3.9.0 plugin refactor, patch set cleared to zero, standalone distribution package retired (11-05)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 triton-riscv: build entry consolidated into Makefile, backend installed into Triton package as plugin directory (11-05)
    • 1.2 triton-riscv: release cadence changed to daily nightly, weekend release, pytest concurrency set to 16 (11-05)
    • 1.3 triton-riscv: stable ABI wheel covers 3.12 to 3.15, fixes manylinux toolchain detection (11-05)
    • 1.4 pytorch: zero commits on riscv branch, fifty-five upstream syncs on main with no RISC-V items (11-06 status)
    • 1.5 tilelang-riscv and llvm-project: zero commits in window, LLVM latest still at 11-04 (11-06 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi: LibreOffice packaging 1/n and 2/n merged, forty-one commits in a single day (11-05)
    • 2.2 openRuyi: libvirt 12.9.0 security update, util-linux 2.42.4, gnupg splits out scdaemon (11-05)
    • 2.3 ruyi: 0.54 scheduled for around November 11, zero commits on main repo, three draft fixes not yet promoted (11-05)
    • 2.4 packages-index zero commits, software source adds K1/K3 toolchain, October monthly release still missing (11-06 status)
    • 2.5 OpenJDK: JEP 544 AOT port passes tests and merges into JDK mainline (11-05)
  • III. Vector Extensions and Operator Optimization
    • 3.1 LLVM: six P-ext and XTHeadBb upstream PRs merged, Ztt matrix extension v0.6 initial version lands (11-05)
    • 3.2 GCC: Zba/Zbs/Zbkb three optimizations enter mainline, SPEC CPU 2026 benchmark preparation begins (11-05)
    • 3.3 V8: six patches including scalar move preserving vector tail, enabling VLEN=1024 (11-05)
    • 3.4 VecIntrinBench, IntrinTrans, FBGEMM three repos remain silent (11-06 status)
  • IV. Buddy Compiler Updates
    • 4.1 runtime adds shared serving generation controller, supporting device-side word retrieval (11-05)
    • 4.2 VIR scalable vector lowering patch: broadcast switched to vector<[2]xT>, kernel downgraded via llc option (11-05)
    • 4.3 Nightly build advances to 0.1.4.dev20261105 (11-05)
  • V. Community, Events, and Official Website
    • 5.1 Biweekly progress issue 080 finalized: all sections merged, preface and official website docs two sections still empty (11-05)
    • 5.2 Official website three pages unchanged, two blog posts, twenty-three documentation chapters (11-06 status)
    • 5.3 No new forum posts in window, latest still the 11-03 ruyi installation guide (11-06 status)
    • 5.4 ruyiai-daily 11-05 issue archived, six sections including PyTorch and RISC-V core news (11-05)
    • 5.5 External news six query groups zero hits, no new arXiv papers in window (11-06 status)
  • VI. Trend Observations
    • 6.1 From maintenance branch to upstream residency: pluginization trades patch cost for version binding
    • 6.2 Desktop office gap-filling on the packaging side and the silent parallel of the AI stack
    • 6.3 Issue 080’s increments land in the compiler and runtime, not the application-layer frontend

Today’s Focus: triton-riscv Merges Triton 3.9.0 Plugin Refactor — Patch Set Cleared, Standalone Distribution Package Retired (11-05)

Date: 2026-11-05　Source: PR #98, commit list, PR #92

At 11:13 on 11-05 (Beijing time), trdthg’s PR #98 “Bump Triton to main (3.9.0)” was merged: 5 commits, 66 files, 1,806 lines added, 3,868 lines deleted. The deletions are roughly twice the additions, and that ratio defines the nature of this refactor — it is not adding another layer of patches on top of an existing branch, but deleting the practice of patching altogether.

The first change is the backend form. Triton’s TRITON_PLUGIN_DIRS mechanism installs backend/ as a triton_shared backend plugin, registers the entry point, links the C++ plugin into libtriton.so, and copies triton-shared-opt into the package as well. The standalone triton-riscv distribution package disappears accordingly, and imports in the code are uniformly redirected through this plugin name. The backend itself now locates buddy-opt, buddy-llc, mir-translate, and triton-shared-opt via paths.py; the CPU driver declares itself active only when no GPU backend is present.

The second change is clearing the patch set. The patches/ directory and the scripts that applied patches are all deleted. The patch set previously existed for two reasons: to let Triton build in environments without NVIDIA/AMD toolchains, and to absorb frequent LLVM API changes. Now that buddy-mlir’s LLVM distribution ships with NVPTX/AMDGPU targets and lld, and Triton’s GPU dependency paths are pointed at a stub directory, neither reason holds, so the patch set no longer has any value.

The third change is consolidating the packaging path. release.sh no longer assembles a second wheel: Triton’s plugin support already places the backend and triton-shared-opt inside Triton’s wheel, so the script only strips tools and runs bdist_wheel directly. Makefile becomes the unified entry point for setup, build, test, lit, and clean, with test depending on lit.

The cost is written into the same PR’s commit message. Triton 3.9’s frontend reads fields such as backend_name from the backend; without it, every tl.atomic_* call throws AttributeError, causing 383 FlagGems test cases to fail on riscv64. Adding the three fields backend_name, default_dot_input_precision, and deprecated_fp8_dot_operand_dtypes this time is what closes out the failures. There are concessions on the example side as well: upstream rejected the frontend patches for asin/acos/atan/atan2, so the examples switch to libdevice, which lowers them to the same __nv_* extern_elementwise, which triton-shared then converts into MLIR math operators.

On the same path, PR #92 “Rebase integration onto the official Triton v3.8.0 release” was closed unmerged that day. The two-route dispute recorded in the previous issue (#92 favoring stability at 3.8.0, #98 following the 3.9.0 mainline) has now converged on the latter.

I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Porting)

1.1 triton-riscv: Build Entry Point Consolidated into Makefile, Backend Installed into Triton Package as Plugin Directory (11-05)

Date: 2026-11-05　Source: Commit list, PR #98

The backend’s import path within the Triton package is triton.backends.triton_shared. The installation surface is narrower than before: there are no longer two distribution packages, and no assumption of a sibling buddy-mlir checkout. The documentation’s narrative about “two packages,” the patching steps, and the sibling-checkout assumption have all been removed. In setup.py, shutil.which("lit") now prefers the pip-installed lit console script over the one bundled with the LLVM distribution. The latter calls main() without an if __name__ == "__main__" guard, and Python 3.14’s forkserver multiprocessing re-runs it when importing the main module, causing the lit target to fail at 80-way concurrency.

The backend’s path resolution for local tools has also changed: buddy-opt, buddy-llc, mir-translate, and triton-shared-opt are now uniformly resolved by paths.py. CPU driver activation is determined by “no GPU backend present,” which differs from the previous environment-variable-based approach and behaves more predictably on machines with multiple coexisting backends.

1.2 triton-riscv: Release Cadence Changed to Daily Nightly, Weekend Release; pytest Concurrency Set to 16 (11-05)

Date: 2026-11-05　Source: PR #99, Release list

PR #99 aligns the release cadence with buddy-mlir: the nightly job publishes nightly/<date>-<sha> pre-releases, and the weekend job publishes release/<date>-<sha> marked as the latest version; runs are skipped when there are no new commits, releases remain in draft state until all wheels are uploaded, and workflow_dispatch provides an out-of-band dev/patch option. Two nightly builds were produced that day under this cadence: nightly/20261105-48bde12 (19:50) and nightly/20261105-849a180 (22:48).

On the testing side, the pytest suite was switched to 16 processes. This number is empirically grounded: on a 207-test slice, 1 process took 360 seconds, 8 processes 89 seconds, 16 processes 67 seconds, and 24 processes 67 seconds, while 32 and 64 were actually slower—each test case is a serial chain of “compile, triton-shared-opt, llc, execute, compare,” and beyond a certain point more processes contend for the shared compilation cache. Previously, when run serially, the flaggems suite took 1 hour 14 minutes. setup.sh installs pytest-xdist, and the count can be overridden with PYTEST_WORKERS.

1.3 triton-riscv: Stable ABI Wheel Covers 3.12 through 3.15, Fixes manylinux Toolchain Detection (11-05)

Date: 2026-11-05　Source: PR #100, PR #101

PR #100 builds wheels following upstream Triton’s cp312-abi3 approach: scripts/env.sh and scripts/release.sh set TRITON_STABLE_ABI=ON, nanobind is compiled against the limited API (Py_LIMITED_API=0x030C0000), and setup.py tags the wheel as cp312-abi3. A single wheel covers all interpreters after 3.12 (3.12/3.13/3.14/3.15), reducing the release matrix from three Pythons to one. The source remains unpatched—upstream itself provides this switch.

The same branch also fixes two build-script issues. First, release.sh hardcoded /opt/rh/gcc-toolset-14/enable and checked for it before dnf install clang, but the manylinux image is a rolling :latest that already shipped gcc-toolset-15 at the time, so the enable script never ran and clang could not find crtbeginS.o, -lstdc++, -lgcc (riscv64), or libstdc++’s <version> header (x64), causing both architectures to fail to build; now clang is installed first, then all toolchains are enabled by presence with the newest last. Second, with only one abi3 wheel, the matrix’s python: ["3.12", "3.13", "3.14"] would simply build the same wheel three times. PR #101 was opened that day and remains unmerged, addressing a release-workflow failure during the checkout cleanup stage.

1.4 pytorch: Zero Commits on riscv Branch, Fifty-Five Upstream Syncs on main with No RISC-V Items (11-06 status)

Date: 2026-11-06 (status)　Source: riscv branch commit list, main branch commit list

The riscv branch had zero commits in the window, with the head commit still [SyncBots] Integrate PyTorch at 3afdbd3 from 11-04 03:00 (Beijing time). The main branch had fifty-five upstream syncs in the window; after checking each one, none involve RISC-V or RVV changes—the content centers on XPU/oneDNN, CUDA component loading, Mooncake transport, and ROCm fallbacks.

1.5 tilelang-riscv and llvm-project: Zero Commits in Window, LLVM Latest Still at 11-04 (11-06 status)

Date: 2026-11-06 (status)　Source: tilelang-riscv commit list, llvm-project riscv branch

The tilelang-riscv main branch had zero commits, with the head commit stuck at 11-03 06:13 UTC; the open PRs remain #4 (RVV masking and max reduction) and #2. The llvm-project riscv branch had zero commits in the window, with the head commit still the BOSCAME commit from 11-04 15:40 (Beijing time) highlighted in the previous issue; the main branch’s latest is at 11-02 14:48 UTC. In other words, this issue’s RuyiAI stack commits are all concentrated in the single triton-riscv repository.

II. RuyiSDK Toolchain and Distributions

2.1 openRuyi: LibreOffice packaging 1/n and 2/n merged, forty-one commits in a single day (11-05)

Date: 2026-11-05　Source: commit list, PR #1595, PR #1596

The openRuyi packaging repo saw forty-one commits in the window, thirty-seven of them belonging to the LibreOffice dependency chain, merged in two batches.

The first batch, PR #1595 (1/n, 25 files, 1383 lines added), adds twenty-one packages at once: librevenge and the document import filters built on it, covering WordPerfect (libwpd, libwps), Visio (libvisio), Publisher (libmspub), PageMaker (libpagemaker), QuarkXPress (libqxp), Zoner Draw (libzmf), StarOffice and legacy Office binary formats (libstaroffice, libmwaw, libe-book), plus liborcus for spreadsheet import, libepubgen for Writer’s EPUB export, the OpenDocument generator libodfgen, libetonyek (the companion library for font-embedding parser libeot), the tag language library liblangtag, and the base container mdds.

The second batch, PR #1596 (2/n, 17 files, 4633 lines added), fills in the font toolchain: Draw uses AFDKO’s tx and makeotf to rebuild embedded fonts when importing PDFs, so python-afdko and its dependencies (defcon, fontparts, fontpens, ufoprocessor, ufonormalizer, mutatormath, the Python bindings for zopfli) are packaged together with antlr4-cpp-runtime and libeot; the same batch also updates python-tmt 1.80.0, python-testcloud 0.12.0, python-pbr, python-munch, and python-fedora-distro-aliases.

This chain is not finished yet. PR #1600 (4/n, 14 files, 1094 lines added) packages Firebird and its parser generator btyacc and interface generator cloop for the Base component, and is still under review; PR #1602 (6/n, 16 files, 1360 lines added) adds PDFium, ZXing, SANE, and Box2D, and is still under review; PR #1603 (7/n, 47 files, 3342 lines added) covers Pentaho, HSQLDB 1.8, BeanShell, and a JavaScript engine, and was closed without merging. PR #1595’s description notes AI-assisted contribution (Assisted-by: GPT-6-Astra xhigh), consistent with openRuyi’s AI contribution policy requirements.

2.2 openRuyi: libvirt 12.9.0 security update, util-linux 2.42.4, gnupg splits out scdaemon (11-05)

Date: 2026-11-05　Source: libvirt 12.9.0, util-linux 2.42.4, gnupg split

Four routine commits outside LibreOffice: libvirt bumped to 12.9.0 and tagged security-fixes; util-linux bumped to 2.42.4; clang-wrap bumped to 20261105; gnupg splits scdaemon into a separate subpackage, reducing dependencies for smart card and hardware key scenarios. Three other security-side PRs are still under review: #1580, #1581, and #1582 handle vulnerability backports, model-related packages, and network-related package security fixes respectively.

2.3 ruyi: 0.54 slated for around November 11, zero commits in the main repo, three draft fixes not yet promoted (11-05)

Date: 2026-11-05　Source: issue 080 draft, ruyi open PRs, ruyi Releases

Issue 080 gives a definite timeline: RuyiSDK 0.54 will be released around November 11, with the package manager version also at 0.54.0. The headline feature of this release is ABI compatibility detection between software packages and the current system, initially supporting any Linux distribution whose C runtime is glibc or musl; second, download requests returning non-2xx status codes are now correctly judged as failures, and the download failure troubleshooting guide adds two categories of causes: server issues and terminal network environment. On the engineering side, dependency versions were updated, including TLS certificate-related security fixes since October; because pygit2 has not yet adapted to Python 3.15, CI does not use 3.15 for now.

The main repo had zero commits and zero releases that day; the latest pre-release is still 0.54.0-beta.20261102 from 11-02, and the latest stable is still 0.53.0 from 09-24. The three fix PRs opened on 11-04 (#511 ignore relative XDG base directories, #513 handle trailing empty tables in manifests, #515 expose repo.disabled via config get) remain in draft status and have not been promoted.

2.4 packages-index zero commits, software source adds K1/K3 toolchain, October monthly release still missing (11-06 status)

Date: 2026-11-06 (status)　Source: packages-index commit list, PR #231, openRuyi news page

packages-index had zero commits in the window, and none of the five open PRs were updated in the window; the latest is #231 from 10-29 (openwrt-sifive-unmatched bumped to 25.12.6). Issue 080 discloses that the substantive change to the software source this period is the addition of toolchain adaptations for SpacemiT K1/K3 under toolchain/gnu-ruyisdk, plus ABI compatibility information added for some packages.

The openRuyi news page’s monthly release is still stuck at 2026.09 (released 09-30), and the 2026.10 monthly release remains missing; there have been no new entries since the two announcements on 10-19.

2.5 OpenJDK: JEP 544 AOT port passes tests and merges into JDK mainline (11-05)

Date: 2026-11-05　Source: issue 080 draft, JDK PR #32311, JEP 544

The biggest single item recorded in issue 080 is the RISC-V port of JEP 544 (Ahead-of-Time Code Compilation): functional tests pass, performance gains meet expectations, and it has been merged into the JDK mainline (PR #32311, 8390102 Support AOT Code Cache). Nine other mainline PRs were reviewed and merged in the same period, covering enabling high general-purpose registers in C1 (#32496), implementing the CRC32C intrinsic with the Zbc extension (#32061), switching logical right shift plus mask combinations to shift pairs (#32046), switching C2 pointer constants to movptr2 (#32629), and more; on the LTS side, the out-of-bounds read fix for the string_indexof_char intrinsic was backported to the 25u, 21u, and 17u lines.

III. Vector Extensions and Operator Optimization

3.1 LLVM: Six P-ext and XTHeadBb Upstream PRs Merged, Ztt Matrix Extension v0.6 Initial Version Lands (11-05)

Date: 2026-11-05　Source: Issue 080 article, LLVM PR 235306, LLVM PR 237815

Issue 080 lists six upstream merges, concentrated on packed SIMD and bit manipulation: the P extension synthesizes more pm*add* plus subsequent-add patterns into pm*adda*-class instructions (#235306), improves code generation for 64-bit funnel shifts on RV32 (#235245), adds packed multiply-accumulate intrinsics with widening horizontal add (#234450), and adds signed, unsigned, and mixed-sign forms of packed byte horizontal multiply-accumulate intrinsics to Clang (#234484); on the GlobalISel side, under XTHeadBb it first widens narrow-typed G_CTLZ_ZERO_POISON (#237815), and widens G_CTTZ to XLEN before going through G_CTLZ to select th.ff1 (#238214).

RuyiSDK’s LLVM branch has two additional changes: preliminary RV32/RV64 matrix extension support based on the Ztt specification and intrinsics documentation v0.6, covering assembly and disassembly, Clang types and intrinsics, LLVM IR and code generation, as well as M/ACC register allocation, copying, and spill handling (#169); and GitHub Actions CI for the AME branch, automatically building LLVM/Clang for RISC-V targets and running CodeGen, assembly, and semantic tests (#170).

3.2 GCC: Three Zba/Zbs/Zbkb Optimizations Enter Mainline, SPEC CPU 2026 Evaluation Preparation Begins (11-05)

Date: 2026-11-05　Source: Issue 080 article, GCC Zba commit

Three RISC-V backend optimizations entered the GCC development mainline on October 27 (r17-5638 through r17-5640): on the Zba side, improved pattern recognition for combinations of shift-add with integer division and remainder; on the Zbs side, leveraging bclr’s own shift-index semantics to eliminate redundant mask computation, improving RV32/RV64 bitmap operations; on the Zbkb side, correcting the packh cost model and match priority, resolving the issue where the better instruction could not be selected when Zbb and Zbkb are enabled simultaneously.

Supporting toolchain CI also progressed: RISE-CI connected the Hertner container service with GitHub Actions, added multilib build scenario checks, and plans to integrate an LLM to evaluate and analyze CI errors. On the performance side, preparation began for real-hardware RISC-V evaluation based on SPEC CPU 2026; a memory bottleneck has been identified for running the full SPECspeed 2026 on the candidate K3 platform, and other hardware options are being evaluated.

3.3 V8: Six Patches Including Preserving Vector Tail on Scalar Moves and Enabling VLEN=1024 (11-05)

Date: 2026-11-05　Source: Issue 080 article, V8 CL 8714039

The V8 section of Issue 080 lists six commits submitted and merged this period and six reviewed and merged. The former include fixing trusted pointer tag range checks, zeroing out-of-bounds trusted pointers on tag underflow, Maglev saving zero-valued constants via the zero register, deriving CPU features from the GN build configuration, preserving the NaN sign bit when converting Float64 to Float32, and Liftoff moving temporary register allocation for subword atomic RMW out of the loop. The latter include preserving the vector tail on scalar moves (CL 8714039), enabling VLEN=1024 (CL 8585179), implementing the missing RV64 Add643 operation, and generating only ebreak for stop() on hardware.

3.4 VecIntrinBench, IntrinTrans, FBGEMM Repos Remain Silent (11-06 status)

Date: 2026-11-06 (status)　Source: VecIntrinBench, IntrinTrans, FBGEMM

All three vector-related repos had zero commits during the window, with their latest commits stopping at 02-13, 06-25, and 07-13 respectively. This period’s vector and operator increments still appear only in the article and upstream compiler commits, and have not returned to these three benchmark and transpilation repos.

IV. Buddy Compiler Updates

4.1 Runtime adds a shared serving generation controller to support device-side token sourcing (11-05)

Date: 2026-11-05　Source: PR #1018, commit c2fd2e1

PR #1018 adds a model-agnostic serving generation controller: 18 files, 2045 new lines, letting runtime adapters share a single set of generated-token budgets, stop conditions, usage accounting, and streaming cleanup. Callers can either use host sampling via LLMSession or a token source that supplies tokens directly from the device side.

A few implementation details are worth noting: string stop conditions are matched on raw bytes first and then UTF-8 decoded, avoiding misjudgment at multi-byte boundaries; the token budget counts both the first prediction and the EOS/stop token; a fixed random seed supports a value of zero; and repetition penalty is applied once per token ID after deduplication within the request’s history window. The first token comes from prefill, so throughput statistics attribute the first selection to prefill and subsequent selections to decode, reporting zero throughput when there is zero decode work. The generation paths of the existing CLI and server entry points remain unchanged, with golden tests using fixed logits and the production DeepSeek tokenizer and sampler to protect outputs. Follow-up adapter integrations are tracked separately under DeepSeek R1 (#1021), Llama 3.1 TTNN (#1022), and Qwen3-VL, among others.

4.2 VIR scalable vector lowering patches: broadcast goes through vector<[2]xT>, kernels downgraded per llc options (11-05)

Date: 2026-11-05　Source: PR #1030, PR #1031

Both midend changes target scalable vectors. The lowering of vir.shuffle is split into several stages: broadcast first goes through vector<[2]xT> and then takes the first half (vector.scalable.extract) to obtain vector<[1]xT>, because when vector.broadcast is applied directly to vector<[1]xT>, LLVM generates a shufflevector with mask [0], which createOrFold folds away as an identity operation, leaving the remaining vscale-1 lanes all poison; llc turns the insertelement into a vfmv.s.f that writes only lane 0, and once this value is stored to a register instead of being merged into a .vf instruction, the result is wrong. After switching to [2]xT, llc uses vfmv.v.f to complete the broadcast at the LMUL of [1]xT, with no increase in instruction count and one fewer instruction in the loop. The other change downgrades kernels according to options such as llc’s -mattr, and the test vir-broadcast-scalable-one.mlir covers both the strip-mining path and the set_vl-with-result path.

4.3 Nightly build advances to 0.1.4.dev20261105 (11-05)

Date: 2026-11-05　Source: Release list

The 0.1.4.dev20261105 nightly was released at 10:58 (Beijing time) that day, replacing the previous issue’s 0.1.4.dev20261103. There were three commits in the window: the runtime controller (4.1) and the two VIR lowering changes (4.2).

V. Community, Events, and Official Homepage

5.1 Biweekly Progress Issue 080 Finalized: All Sections Merged, Foreword and Official Docs Sections Still Empty (11-05)

Date: 2026-11-05　Source: Issue 080 draft, article repo commit list

Issue 080 merged eight commits between 11-05 09:59 and 15:25 (one commit titled biweekly 80). Previously only the base C library and V8 sections were drafted; this round added the package manager, IDE (two commits), development board support matrix, OpenJDK, GCC, Go, and ruyipkg sections. Technical highlights are covered in 2.3, 2.5, 3.1, 3.2, and 3.3. On the board side, new Chinese and English test reports were added for Milk-V Jupiter2 with Bianbu 4.0.1 Desktop and SpacemiT K3 Pico-ITX with Bianbu 4.0.8 LXQt, plus an updated Ubuntu 24.04.5 report for VisionFive 2 Lite; the examples repo added Dhrystone and SiFive Kernel Library examples for HiFive Premier P550, a llama-server deployment example for Jupiter2, and ROS 2 course entries for K3 CoM260 and K3 Pico-ITX.

Three sections remain unfilled: the foreword, version testing and known issues, and official homepage & docs; the QEMU section is also empty. The “GCC” subheading appears twice in the draft, an editorial oversight.

5.2 Official Homepage Unchanged Across Three Pages; Two Blog Posts, Twenty-Three Doc Chapters (11-06 status)

Date: 2026-11-06 (status)　Source: RuyiAI official homepage

Snapshots of the index, blog, and docs pages match the previous issue, with no new blog posts or doc chapters. The blog index currently lists two posts (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices), and the docs index lists twenty-three chapters.

5.3 No New Forum Posts in Window; Latest Still the 11-03 ruyi Install Guide (11-06 status)

Date: 2026-11-06 (status)　Source: RuyiSDK community latest topics, 11-03 post

Zero new posts in the window. The latest is still the 11-03 17:50 “Guide to installing and upgrading ruyi via script”; although issue 080 was finalized in the article repo, it has not yet been posted to the forum, and the latest biweekly progress thread on the forum remains issue 079 (10-26).

5.4 ruyiai-daily 11-05 Issue Archived, Six Sections Including PyTorch and RISC-V Core News (11-05)

Date: 2026-11-05　Source: 11-05 issue archive

The daily intelligence repo archived the day’s issue at 11-05 10:01, structured as today’s headlines, today’s index, and six body sections: PyTorch ecosystem core updates, latest LLVM/MLIR progress, Triton and TileLang technical updates, RISC-V core news, major AI industry news, and summary and trend observations. The repo is auto-generated by a pipeline and serves as cross-reference for this daily report, not as a primary source.

5.5 Six External News Queries Return Zero Hits; No New arXiv Papers in Window (11-06 status)

Date: 2026-11-06 (status)　Source: Google News, arXiv RISC-V search

All six Chinese and English queries (e.g., RISC-V, RuyiSDK in Chinese and English, RuyiAI, Ruyi community open source, Ruyi RISC-V AI) returned zero hits within the 24-hour window; generic RISC-V topic crawls returned only numerous personal repos and entries unrelated to the Ruyi community, such as Infineon AURIX RC1 and Quintauris EM-Altair. No new papers appeared in the arXiv search window, with the latest entry dated 11-03.

VI. Trend Observations

6.1 From maintaining a fork to piggybacking on upstream: pluginization trades patch cost for version coupling

Date: 2026-11-05　Source: PR #98, PR #92

In a single day, triton-riscv went from “maintaining a patched Triton fork” to “a plugin for upstream Triton,” netting over two thousand lines deleted and eliminating the patch directory entirely. The payoff is straightforward: future upstream version bumps no longer require replaying patches, packaging collapses to a single wheel, and the install surface drops from two distribution packages to one.

The cost is equally clear, and it surfaced the same day: Triton 3.9’s frontend reads a backend_name field from the backend, and its absence caused 383 FlagGems test cases to fail. Parasitizing on upstream extension points means upstream frontend implicit contracts become interfaces that must be satisfied—and these contracts are often undocumented, discoverable only by hitting them in tests. Whether this refactor succeeds should not be judged by how many lines were deleted today, but by the cost of replaying it at the next upstream version bump, and by whether downstream suites like FlagGems still need a dedicated person watching versions.

6.2 Desktop office gap-filling on the packaging side and the silent parallel of the AI stack

Date: 2026-11-05　Source: openRuyi commit list, RuyiAI-Stack organization

openRuyi saw forty-one commits in a single day, thirty-seven of them on LibreOffice’s dependency chain: first twenty-one document import filters, then the AFDKO font toolchain, followed by in-review PRs for Firebird, PDFium, ZXing, SANE, and Box2D. This is not scattered updating but a complete desktop office suite packaging effort advanced component by component in batches.

On the same day, the RuyiAI stack was silent apart from triton-riscv: zero commits on pytorch’s RISC-V branch, zero on tilelang-riscv, zero on llvm-project’s riscv branch. The two tracks are driven by different people, different repositories, and different cadences; at the daily-report level they are worth viewing separately, and activity on the packaging side should not be mistaken for progress on the AI stack.

6.3 Issue 080’s increments land in the compiler and runtime, not the application-layer frontend

Date: 2026-11-05　Source: Issue 080 article

The four most substantive items in Issue 080 are all at the low level: three GCC Zb* optimizations into mainline, six LLVM P-ext and XTHeadBb merges, the OpenJDK JEP 544 AOT port merge, and V8’s vector tail and VLEN handling. The application-layer frontend (PyTorch, TileLang) saw no corresponding capability integration this period.

This distribution points to the same thing as 6.1: the focus at this stage is making existing operators and language runtimes run stably and correctly on RISC-V, not expanding frontend coverage. For downstream users, perceptible near-term changes will appear in compiler output and JDK behavior, not in the list of newly supported frameworks.

Source Verification Table

Source Verification result
RuyiAI-Stack key repos triton-riscv: eight commits, three PR merges (#98, #99, #100), two nightly builds; pytorch: zero commits on riscv branch, fifty-five upstream syncs on main with no RISC-V items; tilelang-riscv, llvm-project: zero commits in window; remaining repos: no commits in window
ruyisdk organization ruyi, packages-index: zero commits, zero releases; wechat-articles: eight commits (all Issue 080 sections merged); ruyi-tuner, docs, support-matrix, website repo: no commits in window
openRuyi-Project Forty-one commits: LibreOffice 1/n (twenty-one packages) and 2/n (seventeen files, 4633 lines added) merged, four routine updates; PRs #1595, #1596, #1594 merged; 2026.10 monthly release still missing
buddy-compiler buddy-mlir three commits (one runtime controller, two VIR lowering); nightly 0.1.4.dev20261105
RuyiSDK forum Zero new posts; latest is the 11-03 ruyi script installation guide; forum biweekly progress still stuck at Issue 079
RuyiAI official homepage index, blog, docs pages all unchanged; blog 2 posts, docs 23 chapters
ruyiai-daily 11-05 issue archived
Google News Zero hits across six Chinese and English queries
arXiv No new papers in window; latest stops at 11-03

Complete Source List