Research window: 2026-10-23 04:30 ~ 2026-10-26 04:30 (Beijing time, past 3 days, covering Saturday and Sunday), directly contiguous with the previous issue (cutoff at 10-23 04:30).

Sources: Full-repository audit of GitHub organizations (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler, and riscv/riscv-arch-test, which was found in this issue to have been relocated; commits, PRs, and Releases within the window were re-verified item by item; upstream forks were filtered by keywords such as RISC-V / RVV / XuanTie, and unrelated upstream commits were excluded), commit-by-commit counts and author distribution for the main openRuyi repository, Release and PyPI version verification for the ruyi package manager, the ruyi-imager commit line, the openRuyi distribution news page, the RuyiSDK developer community forum, the ruyiai-daily daily intelligence archive, three-page snapshots of the official RuyiAI homepage, IREE upstream, arXiv search, and targeted Chinese and English queries via Google News RSS (through a proxy).

This Issue’s Index

  • Today’s Highlights: BUDDY MLIR 0.1.2 released; openRuyi metadata correction PART I lands 102 commits in one go (10-25)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 llvm-project: riscv branch rebased for the third time on 10-25, baseline advanced by 474 commits (10-25)
    • 1.2 llvm-project upstream: seven dense P-extension merges, spacemit-a100 tuning enters mainline (10-23)
    • 1.3 pytorch: riscv branch has no commits for three consecutive days, whole-branch replay cadence interrupted (10-25 status)
    • 1.4 tilelang-riscv and triton-riscv: ten days and seven days of silence (10-25 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi main repo: 124 commits, metadata correction PART I lands 102 commits at once (10-23)
    • 2.2 openRuyi PRs: 13 merged, 16 Go dependency packages for SQL-migrate brought into the repo (10-24)
    • 2.3 ruyi package manager: holds at 0.53.0, only five days left in the end-of-month window for 0.54.0 (10-25 status)
    • 2.4 ruyi-imager: adds SpacemiT K3 UFS and UEFI flashing support (10-23)
    • 2.5 packages-index and other ruyisdk repos: zero commits in the window (10-25 status)
  • III. Vector Extensions, Architecture Testing, and Operator Optimization
    • 3.1 riscv-arch-test moves to the riscv organization: 30 commits, suite consolidation and H-extension completion (10-25)
    • 3.2 AME and BOSC AME: LLVM side replayed as-is with the rebase, AME-operator silent (10-25 status)
    • 3.3 IREE and vector repos: five IREE commits with no RISC-V-specific work (10-25 status)
    • 3.4 arXiv: no new RISC-V AI preprints in the window (10-25 status)
  • IV. Buddy Compiler Updates
    • 4.1 BUDDY MLIR 0.1.2 released, stable release line resumes progress (10-25)
    • 4.2 Three commits on VIR vectorized representation, PR number passes one thousand (10-26)
    • 4.3 Downstream anchoring: triton-riscv nightly still locked to the 10-17 build (10-25 status)
  • V. Community, Events, and Official Homepage
    • 5.1 Official homepage unchanged across three pages, two blog posts, twenty-three documentation chapters (10-25)
    • 5.2 ruyiai-daily publishes technical intelligence briefings for three consecutive days (10-25)
    • 5.3 ruyisdk.cn forum has no new posts in the window, summit recap remains the latest (10-25 status)
    • 5.4 Organization site contributors.json auto-updates daily (10-26)
    • 5.5 External news and distribution announcements: targeted queries return zero hits, 2026.10 release not yet out (10-25 status)
  • VI. Trend Observations
    • 6.1 Both replay lines stop together: one doesn’t rebase, the other rebases without adding anything
    • 6.2 Upstream P extension shifts from “can compile” to “compiles well”
    • 6.3 Test suites trade consolidation for maintenance cost
    • 6.4 Distributions enter pre-release quiet period

Today’s Highlights: BUDDY MLIR 0.1.2 released; openRuyi metadata correction PART I lands 102 commits in one go (10-25)

Date: 2026-10-25　Sources: BUDDY MLIR 0.1.2, openRuyi PR #1401 metadata correction PART I, llvm-project riscv branch two-round head comparison

This issue covers a three-day window, and three things are most worth noting, conveniently belonging to three organizations.

First, buddy-mlir ends a five-issue pause, releasing BUDDY MLIR 0.1.2 at 06:18 on 10-25. The release notes are empty and auto-generated, so the actual changes must be inferred from commit history: the last three commits in the window all land on VIR (vectorized intermediate representation), generalizing “how to determine vector size” from a single loop to nested loops, with RVV’s vlen as the target upper bound; the 1000th PR lands exactly the day after the 0.1.2 release.

Second, the openRuyi main repo sees 124 commits and 13 merged PRs. Of these, 102 are concentrated in a single time span from 17:00 to 17:15 on 10-23, matching the merge time of PR #1401 “Correct package metadata and source declaration (PART I)”; the content adds source file checksums to a whole swath of specs and switches to %autochangelog; another 16 land at 00:54 on 10-24 with PR #1366, packaging the entire Go dependency chain of SQL-migrate into RPMs. 10-24 and 10-25 see zero commits.

Third, the two most closely watched fork branches both have no new features. pytorch’s riscv branch has zero commits for three days, with head still at 10-22; llvm-project’s riscv branch does rebase at 23:00 on 10-25, but after comparing all 26 local patches one by one, they are completely unchanged—only the upstream baseline is advanced by 474 commits. Nearly all new technical content this issue comes from the upstream mainline—seven P-extension commits migrated in, and spacemit-a100’s LoadStore scheduling preference entering mainline.

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

1.1 llvm-project: riscv branch rebased for the third time on 10-25, baseline advanced by 474 commits (10-25)

Date: 2026-10-25　Sources: riscv branch commit list, new head 29edf281, head-to-head comparison of the two rounds

The riscv branch completed its third full-branch rewrite at 23:00 on 10-25, with the head moving from 58d62a47 to 29edf281; the head title remains “[RISCV] Add A100 HP/SP and pack/unpack lowering with calibrated scheduling (#31)”. Comparison against the previous head yields diverged: 474 commits ahead, 26 behind, 300 files differing. The 26 behind exactly equals the number of local patches; a patch-by-patch comparison confirms the titles and order of all 26 patches are unchanged — from #31’s A100 HP/SP and pack/unpack scheduling, through BOSC AME’s #14, #16, #18, #24, #29, the Buckyball vendor extensions #20 and #23, IME builtins to A100 dot #30, down to the CI and MLIR-JIT fixes in #11, #6, #4, #3, #1. In other words, this rebase added no new local functionality.

The new baseline brings in two pieces of engineering infrastructure directly related to RISC-V: a new .github/workflows/riscv64-ci.yml (+121 lines) adding a native riscv64 CI job to the image repo; and on the BOLT side, a new RISCVLongJmpPass (+174 lines) plus a series of split-functions tests, handling branches and long jumps across fragments of split functions. The rebase span also returns from the previous round’s 275 commits to 474, close to the routine catch-up magnitude of the 1769-commit rebase on 10-21.

Assessment: the rebase cadence has stabilized from “as scheduled” to once every two or three days, but the zero change in the patch set shows that output during this period is concentrated on the upstream baseline side. Tracking this branch should rely on the set of patch titles, not on SHAs or commit counts.

1.2 llvm-project upstream: seven dense P-extension merges, spacemit-a100 tuning lands in mainline (10-23)

Date: 2026-10-23 to 10-25　Sources: Packed Widening MAC intrinsics a94e3ab3, mul/macc isel patterns bbbc1cb9, Packed Slide b0e82457, mulu isel matching 99e46680, spacemit-a100 scheduling preference fbdefa77

Of the 474 upstream commits brought in by the rebase, 14 are RISC-V related, 7 of which belong to the P (packed SIMD) extension — a higher density than in the previous several windows.

Intrinsics layer: Packed Widening Multiply Accumulate intrinsics merged at 00:47 on 10-24 (#226505); scalar mqwacc and mqrwacc intrinsics merged at 14:55 on 10-23 (#234560); packed byte horizontal multiply-accumulate intrinsics merged at 09:59 on 10-23 (#234484). Instruction selection layer: isel patterns added for mul.h01, mul.h11 and macc.h01, macc.h11 at 11:02 on 10-24 (#234976); isel matching added for mulu.h01, mulu.h11 at 21:34 on 10-24 (#235226); Packed Slide supported at 13:17 on 10-24 (#234629). Validation layer: immediate range checks added for sati and usati intrinsics at 23:32 on 10-23 (#234721).

The remaining RISC-V changes lean toward backend quality and machine model: hasMultipleConditionRegisters made to return true for vector registers at 10:58 on 10-25 (#234876); multiplication judged high-cost when the M extension is absent at 07:30 on 10-25 (#235363); jump table range checks preserved under cf-protection-branch at 00:11 on 10-25 (#234733); TunePreferAscendingLoadStore enabled for spacemit-a100 at 21:31 on 10-23 (#234532); BOLT support for cross-fragment branches of split functions at 11:03 on 10-23 (#223958).

Assessment: the P extension had previously been stuck at “compiles” — instructions and intrinsics merged piecemeal; this batch fills in isel patterns and immediate validation, pushing usability toward “compiles well and doesn’t miscompile”. The spacemit-a100 scheduling preference shows that domestic commercial cores have entered the machine-model tuning stage in LLVM, not just instruction addition.

1.3 pytorch: riscv branch has no commits for three consecutive days, full-branch replay cadence interrupted (10-25 status)

Date: 2026-10-25 (status)　Sources: riscv branch commit list, branch head 29259d90, PR list

The previous issue recorded the riscv branch’s fourth consecutive day of full-branch rewrites at 03:00; in this window that cadence stopped. The branch head is still 29259d90 “Add riscv64 image locally (#51)” from 03:00 on 10-22, with zero commits on 10-23, 10-24 and 10-25.

The seven open PRs (#23, #35, #36, #38, #43, #46, #48) had no update time at all during the window — a search with an updated-time start of 04:30 on 10-23 returns zero hits. Among them, #46 “RISC-V Vector Extension support for ATen” and #48 “Add RISC-V Vector CPU capability” remain the two main proposals for upstreaming RVV capability, with status unchanged from 10-22; the local patch series (riscv64 local image #51, FLOAT fix #44, inductor blocklist #50, ULP tolerance #42 and the CI series) is also unchanged.

Assessment: the previous issue grouped pytorch and llvm as “the same kind of daily ops”; this window overturns half of that judgment — llvm is still rebasing, pytorch has stopped. Three days of silence is not enough to call it stalled, but it shows that replay on this branch is triggered manually or by a scheduled job, not a steady daily pipeline.

1.4 tilelang-riscv and triton-riscv: ten days and seven days of silence (10-25 status)

Date: 2026-10-25 (status)　Sources: tilelang-riscv commit list, tilelang-riscv PR #4, triton-riscv commit list, nightly build nightly/20261018-a305fb6

The last commit on tilelang-riscv’s main branch is still from 10:52 on 10-16, ten consecutive days with no commits; open PR #4 “Add RVV vector mask and max reduction lowering” was last updated on 09-28, and PR #2 on 06-16. triton-riscv’s last commit was at 00:57 on 10-18, and the latest nightly build is still nightly/20261018-a305fb6, released at 10:01 on 10-19 — seven days with no new build, and no new activity on PR #85.

The two repos can be summed up in one sentence: within the RuyiAI stack, both the TileLang and Triton RISC-V backend lines are waiting on upstream — TileLang on buddy-mlir’s VIR and lowering paths, Triton on a new buddy-mlir nightly build. The only progress this period happened on the buddy-mlir side (see 4.1, 4.2), so the recovery signal for these two lines should be watched via when buddy-mlir restarts its nightly builds.

II. RuyiSDK Toolchain and Distributions

2.1 openRuyi main repo: 124 commits, metadata-correction PART I lands 102 at once (10-23)

Date: 2026-10-23　Sources: commit list, PR #1401 metadata and source declaration corrections PART I, systemd homed subpackage 54ab6aa3, orbit2 memory allocation fix ea03990f

All 124 commits in the window carry the SPECS: prefix, from ten authors — Jingwiw 98, cl2t 14, woqidaideshi 2, chihuo2104 2, sunyuechi 2, Z572 2, and one each from DingliZhang, KimmyXYC, wzssyqa, and jchzhou.

The time distribution is the most notable point this issue: all 124 commits landed on 10-23, with 102 of them within a single 17:00–17:15 Beijing-time slot, matching the 17:15 merge of PR #1401; 10-24 and 10-25 saw zero commits. By content, 108 are metadata and source-declaration corrections (adding source-file checksums, switching to %autochangelog for changelog generation, removing unused signatures), and 16 are SPECS: Add new Go dependency packages.

A few substantive changes are worth naming: systemd splits systemd-homed.service into a separate homed subpackage (54ab6aa3); libisoburn fixes a conflict between xorriso’s %post and the man-page alternatives (e335ac44); orbit2 fixes excessive memory allocation during address splitting (ea03990f); plus checksum additions for ltrace, logrotate, libwebp, libunwind, libspng, libsndfile, and others.

Assessment: this is the sequel to the 54-commit sweep of 10-22, with coverage expanded significantly from the previous issue’s luksmeta-to-xmltoman range. It is invisible to ordinary users, but it extends two properties across the whole repo: “build inputs are verifiable” and “changelogs are generated from commit history.”

2.2 openRuyi PRs: 13 merged, 16 Go dependency packages for SQL-migrate enter the repo (10-24)

Date: 2026-10-23 to 10-24　Sources: PR list, PR #1366 SQL-migrate prerequisites, PR #1501 openjdk-25 security release, PR #1470 llvm-snapshot 20261019

13 merged in the window, with 71 open (down six from 77 last issue).

The merged list falls into two categories. The first is routine version tracking and security fixes: openjdk-25 updated to 25.0.5.8 with the security-fixes tag (#1501, 10-23 13:21), llvm-snapshot updated to 20261019 (#1470), postgresql disabling the systemd startup timeout (#1504), openruyi-linux-build fixing an empty linux-devel package (#1491), and libusb adding a systemd-udev recommended dependency (#1507). The openjdk-25 security release listed in last issue’s open queue has now moved to merged; openjdk-latest (#1500) remains queued.

The second is desktop and tooling fill-in: new ydotool (#1457) and new virglrenderer (#1415), which together with the qemu GUI test feature mentioned last issue point to the same thing — running a full desktop on RISC-V and validating it automatically.

There is also one small but clearly directional item: PR #1366 “SPECS: Add SQL-migrate prerequisites.” merged at 10-24 00:54, bringing 16 Go module packages (go-github-gobuffalo-packd, go-github-gobuffalo-logger, go-github-denisenkom-go-mssqldb, go-github-ziutek-mymysql, go-gopkg-gorp.v1, go-howett-plist, etc.). This is a textbook dependency-closure fill-in: before a database migration tool written in Go can enter the repo, the entire Go dependency chain must first be packaged as RPMs.

2.3 ruyi package manager: 0.53.0 holds, only five days left in the end-of-month window for 0.54.0 (10-25 status)

Date: 2026-10-25 (status)　Sources: commit list, Releases, PyPI ruyi

The last commit on main is still the 09-25 merge commit; the latest stable release on both PyPI and GitHub Releases is 0.53.0 (released 09-24); the latest pre-release tag is 0.54.0-alpha.20260918; branches only go up to v0.53.x, with no v0.54.x.

The “0.54.0 slated for end of October” recorded last issue saw no corresponding engineering activity in this window: no new beta or rc tag, and no version branch. If release follows the established cadence at month-end, only five days of build and regression window remain, and a beta must appear first before the release process can begin. This item goes on next issue’s watch list.

2.4 ruyi-imager: adds SpacemiT K3 UFS and UEFI flashing support (10-23)

Date: 2026-10-23　Sources: commit list, K3 UFS/UEFI support 732a9214, PR #14 CLI charset and JSON

4 commits in the window, centered on device support and the catalog model. At 10:30, UFS and UEFI flashing support for SpacemiT K3 was added (732a9214), bringing the new-generation SpacemiT board’s storage media and firmware paths into the image-writing flow; at 12:15, all image-target associations are retained in the catalog and GUI (97d387dd), fixing multi-target images showing only one entry; at 13:31, repository messages are parsed only once per catalog load (fcf0ac7e), a performance and duplicate-rendering fix; at 09:56, PR #14 merged, aligning CLI output with the console charset and ensuring JSON is UTF-8 (6f3e2159).

Assessment: K3 support is the only hardware-specific forward step in this window, and it lands at the “firmware plus storage media” layer — RuyiSDK’s imaging tool is beginning to cover K3’s UFS flashing path, not just SD cards.

2.5 packages-index and the rest of the ruyisdk repos: zero commits in the window (10-25 status)

Date: 2026-10-25 (status)　Sources: packages-index commit list, linux-xuantie-kernel, QEMU p-dev branch, QEMU ame-0.6 branch

packages-index’s last commit is still 10-13, with zero commits in the window; the ISCAS source sync anomaly recorded last issue shows no evidence of recovery this issue. The rest of the ruyisdk org is equally quiet: AME-operator, ruyi-tuner, support-matrix, ruyisdk-vscode-extension, ruyici, and linux-xuantie-kernel all had zero commits in the window; qemu’s main branch is stalled at 09-27, and the two key working branches, p-dev (P extension v0.6) and ame-0.6 (DAMO AME v0.6), last saw commits on 10-21 and 10-16 respectively. wechat-articles had zero commits, with no preparation for a new issue after biweekly report 079.

III. Vector Extensions, Architecture Testing, and Operator Optimization

3.1 riscv-arch-test Migrates to the riscv Organization: 30 Commits, Suite Consolidation and H Extension Completion (10-25)

Date: 2026-10-23 to 10-25　Source: Commit list, PMP suite merge #2996, Sv suite merge #2981, hypervisor CSR and instruction suite #2982, Svinval H and SmH groups #2997

During this period’s review, we found that the original address riscv-arch-test/riscv-arch-test now returns 404, and the repository has moved to riscv/riscv-arch-test (latest push 10-25 18:56 UTC). The previous issue’s conclusion that “the latest commit stopped at 09-08” is therefore void—the migrated repository has 30 commits within the window, making it one of the most active upstreams in the entire window. The RuyiSDK support-matrix and the compliance judgments for various chips are all based on this test suite, so this signal deserves a separate note.

The content splits into two lines. The first is suite consolidation to reduce maintenance cost: merging the scattered PMP suites into one (#2996), merging the F, S, Sm, U, Zaamo, Zalrsc, Zc, and Zicbo groups of Exceptions into a unified Exceptions suite, merging Sv, SvSm, SvZicbo, and ExceptionsSv into a unified Sv suite (#2981), with the CI side simultaneously changed to one exclude_extensions entry per line (#3001). The second is filling gaps in privilege levels and virtualization: the H extension CSR and instruction suite (#2982), merging hypervisor M-mode tests into the Sm group (#2983), VS, VU, and VS-level interrupts in S and H modes (#2911), the H and SmH groups of Svinval (#2997), counter enable and htimedelta tests for H_Zicntr and Sm_Zicntr_H (#2998), six S-state constraints including Shcounterenw (#2987), four basic constraints Ziccif, Ziccrse, Ziccamoa, and Ziccamoc (#2939), plus a fix for the xepc[1] read mask when misa.C is 0 (#2994). The main authors are davidharrishmc and jordancarlin.

Assessment: Doing consolidation and completion at the same time shows that the goal of this reorganization is to first straighten out the structure, then fill content into that structure. The direction is consistent with the openRuyi side pushing RVA23 forward—both are filling in the server and virtualization line.

3.2 AME and BOSC AME: LLVM Side Replayed As-Is with Rebase, AME-operator Silent (10-25 status)

Date: 2026-10-25 (status)　Source: AME instruction update #29, BOSC AME backend design #24, AME-operator, QEMU ame-0.6 branch

Of the AME trio (LLVM backend, QEMU emulation, AME-operator operator library), only the SHA of the first changed in this window. The 10-25 rebase replayed the four BOSC AME-related commits (#14 instructions and builtins, #16 extension support, #24 backend design, #29 instruction update) and the xtame design update (#27) as-is onto the new baseline, with no content changes; AME-operator had zero commits in the window, and the last commit on the QEMU ame-0.6 branch remains 10-16.

Assessment: The AME line entered stagnation after completing the trio loop on 10-22, and this issue has no new evidence to change that status. What is actually moving is the upstream P extension (see 1.2); the AME side must wait for the next actual functional commit.

3.3 IREE and Vector Repos: Five IREE Commits with No RISC-V-Specific Work (10-25 status)

Date: 2026-10-25 (status)　Source: IREE commit list, VecIntrinBench, IntrinTrans

IREE had 5 commits in the window, with zero hits after filtering by the riscv, riscv-v, rvv, and vector keywords, and did not touch the RISC-V backend. RuyiAI-Stack’s VecIntrinBench (last commit 02-13) and IntrinTrans (06-25) remain silent, and the vector builtin testing and conversion tools have no new material this issue.

3.4 arXiv: No New RISC-V AI Preprints in the Window (10-25 status)

Date: 2026-10-25 (status)　Source: arXiv RISC-V search, ATLAS preprint

Searching by RISC-V combined with compiler and vector, there are no new entries in the window. The two most recent are still the 10-22 ATLAS (a multi-agent framework that lifts binaries to canonical compiler IR) and the 10-21 long-vector processor FlashAttention refactor (already covered in the previous issue, with measured 1.84 to 2.08x speedup).

IV. Buddy Compiler Dynamics

4.1 BUDDY MLIR 0.1.2 Released, Stable Release Line Resumes Progress (10-25)

Date: 2026-10-25　Source: BUDDY MLIR 0.1.2, Releases list, main branch commit list

buddy-mlir released BUDDY MLIR 0.1.2 on 10-25 at 06:18 (tag release/v0.1.2). The “zero commits, zero releases within the window, five consecutive stalled periods” noted in the previous issue ends here: the version line advances from 0.1.1 (10-18) to 0.1.2, a seven-day interval that is closer to a regular cadence than the two-day gap between 0.1.0 and 0.1.1.

As with 0.1.1, the 0.1.2 release notes are empty, generated by an automated process, so the changes must be inferred from the commit history (see 4.2). The main branch saw 4 commits within the window: three VIR-related (#996, #999, #1000) and one CI addition of a YOLO26 example test (#808, 10-25 20:22). Nightly builds have not been produced since 0.1.1.dev20261017; of the two channels—stable and nightly—only the former resumed this period.

4.2 Three VIR Vectorization Representation Commits, PR Numbers Pass One Thousand (10-26)

Date: 2026-10-25 to 10-26　Source: vir.from_vector and reduce reassoc #996, RVV vlen as vector size target #999, count live vectors in nested loops #1000

The three commits at the end of the window all land on VIR (vectorization intermediate representation), all in the early hours of Beijing time. [midend][VIR] Add vir.from_vector and reduce reassoc (#996) at 10-25 23:33 adds a from-vector construction entry operation to VIR and reduces reduction reassociation; [midend][VIR] Add (rvv-vlen) the RISC-V Vector target of the vector size (#999) at 10-26 00:31 sets the vector size target to RVV’s vlen; [midend][VIR] Count the vectors live in nested loops for the vector size (#1000) at 10-26 01:09 switches to counting live vectors within nested loops to determine vector size.

Judging from the titles, all three do the same thing: generalize “how to determine vector size” from a single loop to nested loops, with RVV vlen as the target upper bound. This is also buddy-mlir’s 1000th PR—only one day separates the 0.1.2 release from the number passing one thousand, indicating that this intermediate representation line is accelerating.

4.3 Downstream Anchoring: triton-riscv Nightly Still Pinned to the 10-17 Build (10-25 status)

Date: 2026-10-25 (status)　Source: triton-riscv commit list, buddy-mlir nightly 0.1.1.dev20261017

The last commit on the triton-riscv main branch is still the 10-18 commit anchoring the downstream dependency to buddy-mlir’s 10-17 nightly (0.1.1.dev20261017). buddy-mlir has released 0.1.2, but there is no new nightly tag, so the two sides still lack a build they can interface with.

Whether to switch the anchor to 0.1.2 is a step worth watching next issue: only if the three VIR changes are adopted by triton-riscv will they truly enter the operator generation pipeline of the RuyiAI stack.

V. Community, Events, and Official Homepage

5.1 Official homepage: three pages unchanged; two blog posts, twenty-three documentation chapters (10-25)

Date: 2026-10-25 (status)　Source: RuyiAI official homepage, Blog, Documentation

The sha256 of all three pages matches the previous issue: 3 entries in the homepage index, 2 blog posts (Triton’s RISC-V port, PyTorch RISC-V CI practices), and 23 documentation chapters. No new blog posts or documentation chapters have appeared for several consecutive windows; official content remains stalled.

5.2 ruyiai-daily publishes technical intelligence briefings for three consecutive days (10-25)

Date: 2026-10-23 to 10-25　Source: ruyiai-daily repository, 10-25 issue commit, 10-24 issue commit

The ruyiai-daily repository under the organization received one commit on each of the three calendar days in the window (10-23, 10-24, 10-25), with files archived as daily-news-YYYY-MM-DD.md. The content is the “Codex Technical Intelligence Daily,” covering five tracks: the PyTorch ecosystem, LLVM and MLIR, Triton and TileLang, the RISC-V software stack, and the AI industry.

The RISC-V section of the 10-25 issue corroborates this issue’s verification: jump table bounds checking under control-flow protection (corresponding to upstream #234733 on 10-25), the P extension packed slide (corresponding to #234629 on 10-24), and the architectural test semantics of Sscofpmf and Sdtrig. Archiving has continued since 08-21, with a few missing days in between.

5.3 ruyisdk.cn forum: no new posts in the window; summit recap remains the latest (10-25 status)

Date: 2026-10-25 (status)　Source: RuyiSDK developer community, Summit recap post #2893

The two latest posts on the forum homepage are still the 10-22 recap of the 6th RISC-V China Summit (#2893) and the lecture notes for Session 02 of the RISC-V Linux practice course (#2894); there were no new posts in the window. The most recent biweekly report on the forum is still issue 078; issue 079 has not yet been published as a post, consistent with the previous issue’s status.

5.4 Organization site contributors.json auto-updates daily (10-26)

Date: 2026-10-24 to 10-26　Source: ruyiai-stack.github.io commits

The official site repository received 3 commits in the window, all bot-driven daily updates to contributors.json (one each on 10-24, 10-25, and 10-26), with no manual content changes. Such commits only indicate that the site’s automated tasks are still running and do not constitute content updates.

5.5 External news and distribution announcements: zero hits on targeted queries; 2026.10 release not yet out (10-25 status)

Date: 2026-10-25 (status)　Source: Google News RSS, openRuyi distribution news page

Targeted Google News queries in Chinese and English for RuyiSDK, RuyiAI, and 如意 RISC-V returned zero hits in the window; there was no media coverage on the community side. The latest item on the openRuyi distribution news page is still the 2026.09 release announcement (9-30), and the 2026.10 release had not shipped as of the cutoff. Judging by the cadence of previous months (6-30, 7-23, 8-27, 9-30), the October release will most likely land at the end of October, overlapping the time window of ruyi 0.54.0.

VI. Trend Watch

6.1 Two rebase lines pause together: one keeps its base, the other rebases without adding anything

The clearest signal this period is that “rebase” itself is carrying less information. The pytorch riscv branch has had zero commits for three days; the llvm riscv branch rebased on 10-25, but its 26 local patches are unchanged one by one—only the upstream baseline moved forward by 474 commits.

Together this shows that the maintenance mode of these two fork branches is “keep the baseline close to upstream,” not “keep adding features.” Real new features come from upstream merges (see 6.2); the local patch set has long stayed on the same set of vendor extensions.

The practical implication for readers: to track these two branches, look at the set of patch titles, not the SHA or the commit count. A rebase rewrites committer dates uniformly and inflates the commit count with upstream, but the patch content may not change for a week—citing any downstream by SHA will be stale the same day.

6.2 Upstream P extension moves from “compiles” to “compiles well”

In three days there were 14 upstream RISC-V-related commits, 7 of them for the P extension: three add builtins (multiply-add wide, scalar accumulate, byte horizontal multiply-add), three add instruction selection patterns (high/low nibble combinations of mul, macc, mulu and packed slide), and one adds immediate range checking for sati and usati.

These three layers correspond exactly to the sequence by which an extension goes from “has instructions” to “safe to use”: first builtins so the compiler can express it, then isel so code can be generated, and finally validation so bad input is rejected. Last issue’s judgment that the P extension “lacks application-side benchmarks” is unchanged this period, but the foundation has been filled in enough to start performance comparisons. The spacemit-a100 LoadStore scheduling preference and “treat multiplication as high-cost when there is no M extension” are machine-model calibrations in the same direction.

6.3 Test suites trade merging for maintenance cost

After riscv-arch-test migrated to the riscv org, there were 30 commits in three days, a substantial share of them merging scattered suites into one (PMP, Exceptions, Sv). Such changes do not increase test coverage; they reduce the maintenance surface: with fewer suites, CI configuration, exclusion lists, and report aggregation all get simpler.

The H extension tests added at the same time are a different matter—hypervisor-related CSR and interrupt semantics were previously thinly covered, and this period filled in Svinval, Shcounterenw, H_Zicntr, and others in one go. Doing merging and filling in at the same time shows that the goal of this reorganization is to straighten out the structure first and then fill content into it, rather than simply slimming down.

6.4 Distributions enter a pre-release quiet period

All 124 openRuyi commits are concentrated in the dozen or so hours of 10-23, followed by two days of zero commits; openRuyi 2026.10 has not been released, and the news page still shows the 2026.09 announcement from 9-30. On the ruyi side, 0.54.0 has only the 9-18 alpha tag, and the main branch has had no commits since 9-25.

Following the pattern of previous months, there is a “cleanup and freeze” quiet period before a monthly release: push metadata, dependencies, and security updates all at once, then cut the version. This window’s 102 metadata corrections and 16 Go dependency packages fall exactly in the cleanup phase of this pattern. What to watch next period is which comes out first, 2026.10 or 0.54.0.

Source verification table

Source Verification result
RuyiAI-Stack org (all repos) pytorch: riscv branch has zero commits in three days, head stays at 29259d90 (10-22), seven open PRs with no update time; llvm-project: riscv branch rebased for the third time on 10-25 (head 58d62a47 replaced by 29edf281, ahead 474, behind 26, 300 files), the 26 local patches unchanged one by one; triton-riscv and tilelang-riscv have zero commits and zero releases; ruyiai-daily three issues; ruyiai-stack.github.io three bot commits; FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, harness, torq-tile, pytorch-ci, dsh-osc have no commits in the window
ruyisdk org (all repos) ruyi-imager 4 commits (K3 UFS and UEFI flashing, image target association, catalog message parsing, CLI charset PR #14); ruyi zero commits and zero releases (0.53.0 maintained, consistent with PyPI); packages-index, AME-operator, wechat-articles, linux-xuantie-kernel, ruyi-tuner, support-matrix, ruyisdk-vscode-extension, ruyici all have no commits in the window; qemu main branch stopped at 09-27, p-dev stopped at 10-21, ame-0.6 stopped at 10-16
openRuyi-Project org openRuyi main repo 124 commits (Jingwiw 98, cl2t 14, the other eight people 1 to 2 each) + 13 PRs merged + 71 open; homepage zero commits
buddy-compiler org buddy-mlir 4 commits (three VIR, one CI) + 1 formal release BUDDY MLIR 0.1.2; nightly still at last period’s 0.1.1.dev20261017
riscv/riscv-arch-test The original riscv-arch-test org path now returns 404; the repo has moved to the riscv org; 30 commits in the window, mainly suite merges and H extension fill-ins
RuyiSDK developer community forum No new posts; latest are #2893 and #2894 from 10-22; the forum edition of the biweekly report is still issue 078
openRuyi distribution news page No new articles; the 2026.09 release announcement is still the latest; 2026.10 not released
ruyi package manager / PyPI 0.53.0 maintained, no 0.54.0 beta or rc; main branch has had no commits since 09-25
Official homepage (ruyiai.org) Three pages unchanged: homepage index 3 items, blog 2 posts, docs 23 chapters; sha256 identical to last period
IREE upstream 5 commits in the window, no RISC-V-specific changes
arXiv (RISC-V) No new items in the window; most recent are ATLAS on 10-22 and long-vector FlashAttention on 10-21
Google News RSS (multiple Chinese and English queries) Zero hits for targeted queries on RuyiSDK, RuyiAI, and 如意 RISC-V

Full source list