RuyiAI Daily Intelligence Report (2026-11-09)
Research window: 2026-11-06 04:30 ~ 2026-11-09 04:30 (Beijing time, the past 3 days, covering Saturday and Sunday; directly contiguous with the cutoff of the previous report on 11-06).
Sources: GitHub key repositories and organization push verification (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler; repository-by-repository verification of commits, PRs, Releases, and tags within the window, with branch-level atom as a fallback; upstream forks filtered by RISC-V / RVV / XuanTie keywords), Ruyi Firmware firmware repository, ruyi package manager Release, packages-index, openRuyi packaging repository SPECS, RuyiSDK developer community forum, ruyiai-daily daily intelligence archive, three-page snapshot of the RuyiAI official homepage, openRuyi distribution news page, Google News RSS targeted Chinese and English queries (via proxy), arXiv search.
This Issue’s Index
- Today’s Highlights: Ruyi Firmware releases on two consecutive days, HiFive Unmatched enters the support list (11-07, 11-08)
- Today’s Highlights: Models compiled for an exact VLEN silently compute incorrectly on a different VLEN; buddy-mlir switches to load-time validation (11-08)
- I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
- 1.1 triton-riscv: release wheels gain backend registration; release workflow switches to the same checkout action as CI (11-06)
- 1.2 triton-riscv: add the riscv64-missing codegen library for the AMD backend; nightly builds now produce both x86_64 and riscv64 wheels (11-07)
- 1.3 pytorch: over a hundred upstream syncs on main in three days, zero commits on the riscv branch (11-09 status)
- 1.4 tilelang-riscv and llvm-project: zero commits in the window (11-09 status)
- II. RuyiSDK Toolchain and Distributions
- 2.1 openRuyi: kernel bumped on both lines to 7.3.3 and 6.18.62, patch sets and configs refreshed in sync (11-08)
- 2.2 openRuyi: glibc bumped to 2.44 and full test suite run on x86_64; rpm config gains the x86_64 architecture macro (11-08)
- 2.3 openRuyi: Clonezilla packaged, tmux bumped to 3.8, llvm-defaults splits out python-clang (11-06)
- 2.4 ruyi and packages-index: zero commits and zero releases in the window; three draft PRs still not promoted (11-09 status)
- III. Vector Extensions and Operator Optimization
- 3.1 buddy-mlir: model libraries built for an exact VLEN gain load-time validation; on K3, mistakenly running on an X100 core is blocked immediately (11-08)
- 3.2 buddy-mlir: detect the RISC-V host’s VLEN when lowering VIR, fixing riscv64 CI bit-level failures since #1030 (11-08)
- 3.3 Vector benchmarks and the three transpilation repos remain silent (11-09 status)
- IV. Buddy Compiler Updates
- 4.1 AME: BOSCAME backend adds SPM optimization, merging two hundred and nine files at once (11-06)
- 4.2 Three AME and XTAME changes under review, plus a llama-bench-style RAX benchmark (11-08 status)
- 4.3 Nightly builds still at 0.1.4.dev20261105 (11-09 status)
- V. Community, Events, and Official Website
- 5.1 Official website unchanged across three pages; two blog posts, twenty-three documentation chapters (11-09 status)
- 5.2 Forum gains a repost on Infineon RISC-V automotive MCUs; an hwprobe baseline discussion reply on 11-07 (11-07)
- 5.3 ruyiai-daily adds two archived issues, 11-07 and 11-08 (11-08)
- 5.4 Site contributor index auto-updates daily; three records fall within the window (11-09)
- 5.5 Six groups of external news queries return zero hits; arXiv latest stops at 11-03 (11-09 status)
- VI. Trend Observations
- 6.1 The firmware line moves from single-board to multi-board: the scenarios Unmatched and PXE boot point to
- 6.2 The cost of exact-VLEN specialization: silent miscalculation is harder to spot than a crash
- 6.3 Division of labor in the weekend window: the AI stack goes quiet while packaging and firmware advance
Today’s Focus: Ruyi Firmware Ships Two Releases on Consecutive Days, HiFive Unmatched Enters the Support List (11-07, 11-08)
Date: 2026-11-07, 2026-11-08 Source: Ruyi Firmware 20261107, Ruyi Firmware 20261108, Firmware repo commit list
openRuyi’s UEFI firmware project shipped two releases in two nights. 20261107 was published at 22:21 (Beijing time) on 11-07, and 20261108 at 22:47 on 11-08, marked as a pre-release; each release carries four build artifacts.
20261107 is the feature release for this window, and its biggest change is expanding support from SpacemiT K3 to the SiFive HiFive Unmatched: adding SPI boot, JTAG recovery, NVMe and USB boot, and persistent UEFI variables for that board. Alongside this expansion, the release notes list four supporting adjustments: separate DEBUG and RELEASE packages for each board; PXE and HTTP ISO boot controls added to Platform Settings (off by default); UEFI driver emulation now enabled by default, changeable in Setup or settable as the image default via a configuration script; and both boards now share the Ruyi logo and boot countdown bar, with a file selector added to the Firmware Update screen. The K3 side also gets two usability fixes: improved responsiveness of the graphical console, and reading the RTC on P1 during the UEFI boot phase.
20261108 changes only the K3 line: USB Fastboot switches to EDK2’s shared implementation, usable for firmware updates, readback, and booting EFI applications from memory; it also reports the EDK2 firmware version via fastboot getvar version-bootloader. This change folds a vendor-proprietary flashing channel into the standard EDK2 path — a directly usable interface for production lines and after-sales service, and for developers it means no dedicated tool is needed to confirm the firmware version.
Both releases carry the same compatibility statement: Linux mainline support is provided on a best-effort basis; known framebuffer issues have workarounds given in the release notes. Judging by the three releases in a row from 20261013 to 20261015 in October (CPU mode selection, System Information page), this firmware line has settled into a daily iteration cadence, and the addition of Unmatched is its first step beyond a single chip vendor.
Today’s Highlight: Models compiled for an exact VLEN silently compute wrong results on other VLENs; buddy-mlir switches to load-time validation (11-08)
Date: 2026-11-08 Source: PR #1032, commit 4063c48
PR #1032 addresses a class of failure mode that is worse than an error: no error at all, just wrong results. A RISC-V model built for an exact VLEN has vector code that is only valid for that one VLEN—llc is explicitly told the VLEN (+zvlNb and -riscv-v-vector-bits-max), and the scalable vectors in decode attention are sized accordingly (#1028, #1030). On a different VLEN, the model does not crash; it simply produces incorrect results. The commit message gives a concrete example: on K3, the DeepSeek-R1 rax built with prefill_ime (BUDDY_RISCV_VLEN=1024) lands on an X100 core (VLEN 256) when /proc/set_ai_thread has not been executed, producing wrong results with no warning whatsoever.
The fix moves validation forward to load time. runtime/rvv/BuddyRvvVlen.c is linked into the model library as a constructor; as soon as the library is loaded, it reads vlenb, and if it does not match the exact VLEN used at build time, it prints a message and terminates the process with _exit(1). Using _exit rather than exit is deliberate: running destructors while the loader holds its lock would deadlock. The constructor form covers all loaders, including command-line tools, servers, and tests; generic RVV builds (without an exact VLEN) and other targets are unaffected.
A second commit followed review, changing the check from “based on build variables” to “based on the final llc options.” The original approach had two problems: first, native builds may specify a VLEN via the model’s own LLC_ATTRS, and if BUDDY_RISCV_VLEN disagrees with it, the validated value would not match the actual code; second, it was only linked into the third stage, missing both Qwen3-VL shims (vision and decode) and the split-link path. Now the build script extracts the exact VLEN from the final llc options (the same place VIR lowering reads it), and the third stage, both shims, and split linking all carry it; if the options contradict each other at configure time, the build fails immediately.
Testing is the other half of this change: test_buddy_rvv_vlen.py builds the model library for each VLEN from 128 to 65536; only the host’s VLEN can load it, and all others exit with code 1.
test_buddy_rvv_vlen_cmake.py exercises the configure script across combinations of the build VLEN variable, LLC attributes, prefill selection, and zvl parameters. On-device validation was completed on both X100 and A100 cores of K3, and the repository’s self-checks passed 239 plus 409 items.
I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Adaptation)
1.1 triton-riscv: Release wheels gain backend registration; release workflow switches to the same checkout action as CI (11-06)
Date: 2026-11-06 Source: PR #101, commit d8c2b3e
PR #101 fixes two release-chain issues exposed after #100 landed, touching two files with seventeen added lines. The release workflow now uses the same checkout action version as CI; scripts/release.sh gains a backend registration step so that wheels produced by release carry the same plugin registration information as wheels built by CI. Previously the release path only packaged, and the plugin directory was never registered into the wheel—after installation the backend was invisible. This class of problem is invisible in CI because CI goes through the install path rather than the wheel path.
1.2 triton-riscv: Add missing riscv64 codegen library for the AMD backend; nightlies now ship x86_64 and riscv64 wheels (11-07)
Date: 2026-11-07 Source: PR #102, nightly aa51ac2, nightly 6895885
PR #102 is titled “Add AMD support,” but its actual content is supplying a missing library rather than adding a new backend. The upstream AMD backend requires libtriton_amd_codegen.so: either a prebuilt copy is downloaded, or it is built against a separately pinned LLVM version—and neither path has a riscv64 variant. The library itself is small and can be compiled in seconds with the LLVM already present in the repo. setup.sh therefore builds it locally and places it at the agreed path when the library is detected missing, and env.sh points TRITON_AMD_CODEGEN_PATH at it—the previous GPU dependency stub redirected all GPU downloads to an empty directory, which also skipped upstream’s fetch of this library, while plugin compilation still reads that path. Without it, the AMD backend refuses to compile anything.
This change also clarifies the repo’s positioning after the plugin refactor: the same tree both produces riscv64 Triton and hosts, at build time, the components other backends need—multi-backend coexistence no longer relies on branch switching. Two nightlies were produced on the new cadence within the window (11-06 12:08 and 11-07 07:52), each with two artifacts: a cp312-abi3 x86_64 wheel (~320 MB) and a riscv64 wheel (~337 MB). Previous nightlies had five to six artifacts; the drop in artifact count is consistent with the stable-ABI single-wheel, backends-folded-into-the-Triton-package direction.
1.3 pytorch: Over a hundred upstream syncs on main in three days, zero commits on riscv branch (11-09 status)
Date: 2026-11-09 (status) Source: riscv branch commit list, main branch commit list
This fork’s default branch is riscv, not main, so the branch must be specified explicitly when reading commits. The riscv branch had zero commits in the window; its head commit remains the batch of [SyncBots] Integrate PyTorch syncs from 11-04 03:00 (Beijing time). The main branch had over a hundred in the window (the API returns a full single page of one hundred; the earliest is 11-06 11:07 UTC), all authored by PyTorch MergeBot, and filtering by the keywords riscv, rvv, XuanTie, and rv64 yields zero. The three days’ content centers on CUDA and MPS operator migrations, C++ FakeTensor, the XCCL communication component, and inductor’s CPU reductions and CI adjustments.
1.4 tilelang-riscv and llvm-project: Zero commits in the window (11-09 status)
Date: 2026-11-09 (status) Source: tilelang-riscv commit list, llvm-project commit list
tilelang-riscv’s main branch head commit is stuck at 11-03 06:13 UTC; llvm-project’s riscv branch remains at the 11-04 BOSCAME commit, and main at 11-02 14:48 UTC. Neither repo had commits in the window; all of the RuyiAI stack’s additions this period come from triton-riscv and buddy-mlir.
II. RuyiSDK Toolchain and Distributions
2.1 openRuyi: Both kernel lines advance to 7.3.3 and 6.18.62, patch sets and configs refreshed in sync (11-08)
Date: 2026-11-08 Source: linux 7.3.3, linux-lts 6.18.62, commit list
Both kernel lines move forward at once: the rolling line’s linux rises to 7.3.3, and the long-term line’s linux-lts rises to 6.18.62, with both commits noting refreshed patch sets and refreshed configs. These two are the most substantive changes in the packaging repo since the 11-05 batch of LibreOffice dependencies, because kernel version bumps in openRuyi usually come with a re-check of config options—RISC-V platform peripherals, power management, and filesystem options are all maintained inside %ifarch riscv64 conditional blocks, so the config refresh is more noteworthy than the version numbers themselves.
2.2 openRuyi: glibc bumped to 2.44 with full test run on x86_64, rpm config gains x86_64 arch macros (11-08)
Date: 2026-11-08 Source: glibc bump, glibc full test, x86_64 macros
glibc is bumped to 2.44, pointing at commit c5c6fc6f, and the full test suite is run on x86_64; in the same window, rpm-config-openruyi gains x86_64 arch macros and tune configuration. For a distribution whose sole target is riscv64, these two taken together show the build system shoring up the host side: cross-building, running tests on x86 machines, and native bootstrapping on x86 all require rpm to recognize this architecture. Running glibc tests on the host does not mean giving up riscv64 validation, but it does indicate that the current process places part of the regression checking on the build machine.
2.3 openRuyi: Clonezilla packaged, tmux bumped to 3.8, llvm-defaults splits out python-clang (11-06)
Date: 2026-11-06, 11-08 Source: Clonezilla, python-clang, tmux 3.8
Beyond four routine commits, two in this window are worth noting. Clonezilla enters SPECS, adding a tool for full-machine backup and image distribution; llvm-defaults splits out a python-clang subpackage so users who only need the Python bindings don’t have to pull in the entire compiler. There is also a tmux bump to 3.8, a build-error fix for qt6-qtwebengine, and an early 11-08 dbus dependency cleanup (removing %{?_isa} from dbus-common’s Requires to avoid being pinned to the interpreter architecture during multi-arch installs).
2.4 ruyi and packages-index: zero commits and zero releases in the window, three draft PRs still not promoted (11-09 status)
Date: 2026-11-09 (status) Source: ruyi Release, ruyi open PRs, packages-index commit list, issue 080 draft
The ruyi main repo saw zero commits and zero releases in the window; the latest pre-release remains 0.54.0-beta.20261102 from 11-02, and the stable release remains 0.53.0 from 09-24. The three fix PRs opened on 11-04 (ignoring relative XDG base directories, handling trailing empty tables in manifests, and exposing repo.disabled via config get) are unchanged and still drafts. Issue 080 stated that 0.54 would be released around November 11, making midweek the point to watch. packages-index likewise saw zero commits, with its head commit stuck at the 11-02 ABI information completion.
III. Vector Extensions and Operator Optimization
3.1 buddy-mlir: Model library built for exact VLEN plus load-time validation; on K3, accidental execution on X100 cores is blocked immediately (11-08)
Date: 2026-11-08 Source: PR #1032, commit 4063c48
The change spans twelve files, adding 547 lines and deleting 86, with the core being a C file of just over ten lines: runtime/rvv/BuddyRvvVlen.c. It is linked into the model library as a constructor; on load it reads vlenb and compares it against the exact VLEN used at build time, printing a message and exiting on mismatch. The accompanying build-script changes are more involved than the runtime code—buddy_rvv_vlen.cmake now derives VLEN from the final llc options (sharing the same read logic as VIR lowering), buddy_model.cmake compiles that object file once the spec and options are determined, and the third stage, the two Qwen3-VL shims, and the layered linking all carry it explicitly.
The documentation was updated in tandem for the A210 and K3 DeepSeek-R1 deployment notes, clarifying the boundary between “which builds use exact VLEN and which use generic RVV”: the A210 kernels use fixed-length vectors with no VIR, and llc can lower for any VLEN no smaller than 128, making it a generic build; K3’s prefill_ime requires 1024, making it an exact-VLEN build. This boundary previously existed only in build-script comments; it is now enforced at runtime.
3.2 buddy-mlir: Detect the RISC-V host’s VLEN during VIR lowering, fixing the bit-exact failure in riscv64 CI since #1030 (11-08)
Date: 2026-11-08 Source: PR #1037, commit 79ffafd
The second commit by the same author on the same day fixes an error in the opposite direction. When neither rvv-vlen is given nor scalable vectors are enabled, -lower-vir-to-vector sizes fixed-length vectors by host register width, but its host detection covered only x86 (AVX2, AVX-512) and SVE: RISC-V hosts with the V extension always fell back to the generic default (128 bits, 16 registers), regardless of the actual VLEN.
The consequences showed up in tests. Since #1030, the kernels pipeline lowers VIR by host when no llc options are given; the riscv64 CI host has VLEN 256 but was treated as 128-bit: the squared norms of q and k were split into vector<32xf32> segments for summation, in a different order from the bit-by-bit comparison against the vector kernels, so vir decode qk_norm ... same bits evaluated to false. This problem was present in every riscv64 run on main; x86 (AVX-512) got a width of 1024, so it always passed. The author first added a change that pinned rvv-vlen=1024 use-scalable=false in tests only, then reverted it—that would have created a second place to set VLEN outside the llc options, conflicting with the direction of #1032. The final fix detects the RISC-V host during lowering (reading host CPU features and hwprobe) and sizes by the real VLEN, with the test file replaced by a new case in host form.
3.3 Vector benchmarks and the three translation repos remain silent (11-09 status)
Date: 2026-11-09 (status) Source: VecIntrinBench, IntrinTrans, FBGEMM
The three vector-related repos had zero commits in the window, with their latest commits stopping at 02-13, 06-25, and 07-13 respectively. This issue’s vector progress remains concentrated on buddy-mlir’s VIR and runtime validation, with no return to these three benchmark and translation repos.
IV. Buddy Compiler Updates
4.1 AME: BOSCAME backend adds SPM optimization, 209 files merged in one go (11-06)
Date: 2026-11-06 Source: PR #1033, commit f9d7189
This is the largest change in this window: 209 files, over 46,000 lines added, and more than 2,300 lines removed. It adds SPM (scratchpad memory) optimization to the BOSCAME backend, covering the contract, planning, and emission paths, along with the supporting dialect, runtime, and FPGA toolchain. On the documentation side, four new notes were added: kernel contract pipeline, private DDR, zero accumulator, and KernelContract. On the examples side, two sets of test cases were added for the BOSCAME dialect and FPGA, covering matrix multiplication, transpose, broadcast, and elementwise operations. The second commit fixes tool discovery in CI: the buddy tools are now exposed via PATH and BUDDY_OPT instead of hardcoding the build directory, so that external LLVM caches can find llc, clang, and buddy-opt.
4.2 Three AME and XTAME changes under review, plus a llama-bench-style RAX benchmark (11-08 status)
Date: 2026-11-08 (status) Source: PR #1038, PR #1039, PR #1040, PR #1035
After #1033, three more AME-line changes are under review, all updated on the same day, 11-08: adding descriptor offsets for AME tile addressing; a new lowering of the matmul family from Linalg to XTAME; and XtameBench, a benchmark that evaluates the effectiveness of AME matmul lowering by instruction count. Separately, #1035 provides a llama-bench-style RAX model benchmark, which remains under review after its 11-06 update. Taken together, these changes show AME shifting from “it runs” to “it’s correct and comparable.”
4.3 Nightly build still at 0.1.4.dev20261105 (11-09 status)
Date: 2026-11-09 (status) Source: Release list
The three commits in this window (SPM, VLEN validation, VIR host sizing) have not yet produced a new nightly build. The latest is still 0.1.4.dev20261105 from 11-05 10:58 (Beijing time), with eight build artifacts. Based on the previous cadence, nightly builds are published the day after new commits land. Both VLEN changes in this issue landed over the weekend, so whether they ship with the next nightly build depends on jobs running after Monday.
V. Community, Events, and Official Homepage
5.1 Official homepage’s three pages unchanged; two blog posts, twenty-three documentation chapters (11-09 status)
Date: 2026-11-09 (status) Source: RuyiAI Official Homepage
The snapshots of the index, blog, and docs pages are consistent with the previous issue, with no new blog posts or documentation chapters. The blog index currently has two posts (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices), and the documentation index has twenty-three chapters.
5.2 Forum gains a repost on Infineon’s RISC-V automotive MCU; hwprobe baseline discussion reply on 11-07 (11-07)
Date: 2026-11-06, 11-07 Source: RuyiSDK Community Latest Topics, 11-06 repost
At 12:05 on 11-06 (Beijing time), a new repost was added covering Infineon’s first RISC-V automotive MCU (AURIX RC1); on 11-07, a user replied under the post “Outputting the ISA baseline in /proc/cpuinfo (hwprobe introducing rva23u64 baseline behavior),” which has ninety-two views. Although the biweekly progress issue 080 was finalized in the article repository on 11-05, it has still not been posted to the forum; the latest forum post remains issue 079 (10-26).
5.3 ruyiai-daily adds two archived issues for 11-07 and 11-08 (11-08)
Date: 2026-11-07, 2026-11-08 Source: 11-07 issue, 11-08 issue
The daily intelligence repository published two issues as usual over the weekend, each with six sections: PyTorch ecosystem, LLVM/MLIR, Triton and TileLang, core RISC-V news, major AI industry news, and summary and trend observations. The 11-07 issue covered triton-riscv’s #101 and #102 build changes, the SPM merge in buddy-mlir, and Linux-side patches including Metanoia MT2824 SoC support and the StarFive JH7110 display subsystem; the 11-08 issue covered Ruyi Firmware 20261107 and updates to NEMU and riscv-arch-test. This repository is auto-generated by a pipeline and can serve as cross-reference, but does not constitute a primary source.
5.4 Site contributor index auto-updates daily; three records fall within the window (11-09)
Date: 2026-11-07 to 11-09 Source: ruyiai-stack.github.io commit list
The organization site repository made three automated commits within the window (11-07 05:16, 11-08 03:53, 11-09 04:12, Beijing time), all with the content chore: update contributors.json, belonging to a scheduled task maintaining the contributor index, with no page content changes.
5.5 Six external news queries return zero hits; latest arXiv entry remains at 11-03 (11-09 status)
Date: 2026-11-09 (status) Source: Google News, arXiv RISC-V search
All six Chinese and English query groups (RuyiAI, RuyiSDK, Ruyi RISC-V, openRuyi, Buddy Compiler, etc.) returned zero hits within the three-day window; general crawls on the RISC-V topic returned only entries unrelated to the Ruyi community, such as personal repositories, ESP32 development boards, and Infineon AURIX RC1. No new papers appeared in the arXiv search window; the latest entry remains at 11-03 (silent degradation of bf16/fp16 auto-vectorization on RISC-V), which belongs to the same category of topics as the VLEN issue discussed in buddy-mlir within this window.
VI. Trend Observations
6.1 The firmware line moves from single-board to multi-board: what Unmatched and PXE boot point to
Date: 2026-11-07 Source: Ruyi Firmware 20261107
Ruyi Firmware had previously iterated almost entirely around SpacemiT K3, with all three October releases filling in K3’s CPU mode selection, system information page, and flashing path. 20261107 is the first to bring SiFive’s HiFive Unmatched into support, and not as a minimal bring-up: SPI boot, JTAG recovery, NVMe and USB boot, and persistent UEFI variables are all provisioned at once — taking a development board from “it lights up” to “it works as a workstation.”
Two accompanying settings make the target scenario even clearer: PXE and HTTP ISO boot controls (off by default), and separate DEBUG and RELEASE packages per board. The former is the entry point for data centers and batch deployment; the latter is the division of labor between production lines and development. Open-source RISC-V platforms have long lacked a firmware solution aligned with EDK2 and reusable across vendors; the addition of Unmatched gives this line its first sample from a second chip vendor. The next question is who supplies the third board.
6.2 The cost of exact VLEN specialization: silent miscalculation is harder to catch than a crash
Date: 2026-11-08 Source: PR #1032, PR #1037
K3’s dual-cluster structure lays RISC-V’s VLEN problem bare on a single machine: the A100 core has VLEN 1024 and the X100 core has VLEN 256, and the same executable run on the wrong cluster won’t crash — it will just compute wrong results. This used to be handled by deployment documentation reminders (write /proc/set_ai_thread first); now it is caught by load-time validation. For a toolchain that delivers models, failing to start is preferable to producing wrong results — that is the value of #1032.
On the same day, #1037 proves the same point from the other side: the host-defined default (128 bits, 16 registers) happens to be equivalent to a large VLEN on x86, but on riscv64 it changes the summation order, and bit-for-bit comparison fails immediately. One patch covers build time and the other run time; what they share is that neither assumes any longer that “VLEN is whatever the compiler decides.” Scalable vectors are supposed to be RVV’s standard answer, but in inference kernels that require bit-level reproducibility, exact VLEN is still the better trade; the cost is that such validation must follow the final llc options rather than build variables.
6.3 Division of labor over the weekend window: the AI stack goes quiet while packaging and firmware advance
Date: 2026-11-09 Source: RuyiAI-Stack organization, openRuyi commit list
Over the three-day window, PyTorch’s RISC-V branch, tilelang-riscv, and llvm-project’s riscv branch all had zero commits; the only substantive AI-stack changes were two build-script commits in triton-riscv and three VLEN- and SPM-related commits in buddy-mlir. Over the same period, the openRuyi packaging repo saw ten commits and the firmware repo shipped two releases — a noticeably denser cadence.
This division of labor relates to the two sides’ release cycles: packaging and firmware ship daily or nightly and must track any upstream version change, while the AI stack makes structural adjustments (pluginization, VLEN validation) between version windows and does not aim for daily output. For those following the Ruyi community, the current low-effort reading is: AI stack Monday through Friday, distro and firmware on weekends.
Source Verification Table
| Source | Verification Result |
|---|---|
| RuyiAI-Stack key repos | triton-riscv: two commits (#101, #102), two nightly builds, two abi3 wheels each; pytorch: over a hundred upstream syncs on main, zero commits on the riscv branch; tilelang-riscv, llvm-project: zero commits; remaining repos have no commits in the window |
| ruyisdk organization | ruyi, packages-index: zero commits and zero releases; ruyi: three draft PRs not yet promoted; wechat-articles: no commits in the window; remaining repos have no updates |
| openRuyi-Project | Packaging repo: ten commits (kernel 7.3.3 and 6.18.62, glibc 2.44, x86_64 macros, Clonezilla, tmux 3.8, etc.); firmware repo: two releases (20261107, 20261108); news page: no additions, monthly edition still at 2026.09 |
| buddy-compiler | buddy-mlir: three commits (SPM, VLEN validation, VIR host-defined sizing); nightly build still at 0.1.4.dev20261105; four PRs under review |
| RuyiSDK Forum | New 11-06 Infineon repost thread; 11-07 reply in the hwprobe baseline discussion; biweekly progress forum still at issue 079 |
| RuyiAI official site | index, blog, and docs pages all unchanged; 2 blog posts, 23 documentation chapters |
| ruyiai-daily | 11-07 and 11-08 issues archived |
| Site repo | contributors.json auto-updated three times (11-07, 11-08, 11-09) |
| Google News | Zero hits across six groups of Chinese and English queries |
| arXiv | No new papers in the window; latest one still dated 11-03 |
Complete Source List
- [1] Ruyi Firmware 20261107 — https://github.com/openRuyi-Project/ruyifirmware/releases/tag/20261107
- [2] Ruyi Firmware 20261108 — https://github.com/openRuyi-Project/ruyifirmware/releases/tag/20261108
- [3] Ruyi Firmware commit list — https://github.com/openRuyi-Project/ruyifirmware/commits/main
- [4] buddy-mlir PR #1032 — https://github.com/buddy-compiler/buddy-mlir/pull/1032
- [5] buddy-mlir commit 4063c48 — https://github.com/buddy-compiler/buddy-mlir/commit/4063c4894681e376e6d031cf6e8092aeba91a3c4
- [6] buddy-mlir PR #1037 — https://github.com/buddy-compiler/buddy-mlir/pull/1037
- [7] buddy-mlir commit 79ffafd — https://github.com/buddy-compiler/buddy-mlir/commit/79ffafd8cd823ba4522d589d5f4ba43e08d4917f
- [8] triton-riscv PR #101 — https://github.com/RuyiAI-Stack/triton-riscv/pull/101
- [9] triton-riscv PR #102 — https://github.com/RuyiAI-Stack/triton-riscv/pull/102
- [10] triton-riscv commit list — https://github.com/RuyiAI-Stack/triton-riscv/commits/main
- [11] triton-riscv nightly aa51ac2 — https://github.com/RuyiAI-Stack/triton-riscv/releases/tag/nightly%2F20261106-aa51ac2
- [12] triton-riscv nightly 6895885 — https://github.com/RuyiAI-Stack/triton-riscv/releases/tag/nightly%2F20261106-6895885
- [13] pytorch riscv branch commit list — https://github.com/RuyiAI-Stack/pytorch/commits/riscv
- [14] pytorch main branch commit list — https://github.com/RuyiAI-Stack/pytorch/commits/main
- [15] tilelang-riscv commit list — https://github.com/RuyiAI-Stack/tilelang-riscv/commits/main
- [16] llvm-project commit list — https://github.com/RuyiAI-Stack/llvm-project/commits/main
- [17] openRuyi linux 7.3.3 — https://github.com/openRuyi-Project/openRuyi/commit/04ed27d11450dd8c0fe59679430444063d06570a
- [18] openRuyi linux-lts 6.18.62 — https://github.com/openRuyi-Project/openRuyi/commit/48fa7e07da48ed2a588a8b3d91dfa486ff8dacf8
- [19] openRuyi glibc version bump — https://github.com/openRuyi-Project/openRuyi/commit/1207d672b83cd64b5c799c03d911314f573010a1
- [20] openRuyi glibc full test — https://github.com/openRuyi-Project/openRuyi/commit/fe6bd73219a0efeabdf1598f5b54f96c8767a3e4
- [21] openRuyi x86_64 architecture macros — https://github.com/openRuyi-Project/openRuyi/commit/637c87262219dee70efc2b02a2a12941134a6de1
- [22] openRuyi Clonezilla — https://github.com/openRuyi-Project/openRuyi/commit/eae4fe33d24cabd842a507740ad3e507f72525e0
- [23] openRuyi python-clang — https://github.com/openRuyi-Project/openRuyi/commit/bedeb334d9d8ed87376b2fc237c4d79f7d872071
- [24] openRuyi tmux 3.8 — https://github.com/openRuyi-Project/openRuyi/commit/c57cf038d6cf359bd0e43efd523ebed47181dac9
- [25] openRuyi commit list — https://github.com/openRuyi-Project/openRuyi/commits/main
- [26] openRuyi news page — https://openruyi.cn/zh-Hans/news/
- [27] ruyi Release list — https://github.com/ruyisdk/ruyi/releases
- [28] ruyi open PR list — https://github.com/ruyisdk/ruyi/pulls
- [29] packages-index commit list — https://github.com/ruyisdk/packages-index/commits/main
- [30] RuyiSDK Biweekly Progress Issue 080 — https://github.com/ruyisdk/wechat-articles/blob/main/20261105-ruyisdk-biweekly-80.md
- [31] RuyiSDK community latest topics — https://ruyisdk.cn/latest
- [32] Forum 11-06 repost thread — https://ruyisdk.cn/t/topic/2903
- [33] ruyiai-daily 11-07 issue — https://github.com/RuyiAI-Stack/ruyiai-daily/blob/main/daily-news-2026-11-07.md
- [34] ruyiai-daily 11-08 issue — https://github.com/RuyiAI-Stack/ruyiai-daily/blob/main/daily-news-2026-11-08.md
- [35] RuyiAI official site — https://www.ruyiai.org/
- [36] ruyiai-stack.github.io commit list — https://github.com/RuyiAI-Stack/ruyiai-stack.github.io/commits/main
- [37] buddy-mlir PR #1033 — https://github.com/buddy-compiler/buddy-mlir/pull/1033
- [38] buddy-mlir PR #1038 — https://github.com/buddy-compiler/buddy-mlir/pull/1038
- [39] buddy-mlir PR #1039 — https://github.com/buddy-compiler/buddy-mlir/pull/1039
- [40] buddy-mlir PR #1040 — https://github.com/buddy-compiler/buddy-mlir/pull/1040
- [41] buddy-mlir PR #1035 — https://github.com/buddy-compiler/buddy-mlir/pull/1035
- [42] buddy-mlir Release list — https://github.com/buddy-compiler/buddy-mlir/releases
- [43] arXiv RISC-V search — https://arxiv.org/list/cs.AR/recent