Research window: 2026-10-21 04:30 ~ 2026-10-22 04:30 (Beijing time, past 24 hours), contiguous with the previous issue.

Sources: Full-repository verification across four GitHub organizations (commits, pull requests, and releases in RuyiAI-Stack, ruyisdk, openRuyi-Project, and buddy-compiler, with every item in the window reviewed individually; upstream forks filtered by RISC-V keywords, with unrelated upstream commits excluded), per-commit counts and author distribution for the main openRuyi repository, ruyi package manager and PyPI version verification, the openRuyi distribution news page, the RuyiSDK developer community forum and biweekly report draft repository, three-page snapshots of the official RuyiAI homepage, upstream activity in IREE and riscv-arch-test, official RISE announcements, arXiv searches, and targeted Google News RSS queries in Chinese and English (via proxy).

This Issue’s Index

  • Today’s Highlights: IREE mainline lands RVV f16 multiply-add MMA, LLVM riscv branch rebased and rewritten after eight days of silence (10-21)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 pytorch: third consecutive day of full-branch replay, head switched to 3c1ea315 (10-21)
    • 1.2 llvm-project: eight days of silence end, riscv branch rebased to b3b5e33 (10-21)
    • 1.3 tilelang-riscv: no commits for five consecutive days (10-21 status)
    • 1.4 triton-riscv: nightly build remains at 10-19, no new activity on PR #85 (10-21 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi main repo: 80 commits, Go dependency closure wrapping up and uv backend completed (10-21 to 10-22)
    • 2.2 openRuyi PRs: five merged, nineteen security updates under review (10-21)
    • 2.3 ruyi package manager: 0.53.0 maintained, 0.54.0 alpha still not out (10-21 status)
    • 2.4 ruyi-imager: three Gradle build refactors, nightly build not yet following (10-21)
    • 2.5 Kernel and firmware line: linux-xuantie-kernel, edk2-ruyi, ruyifirmware silent (10-21 status)
  • III. Vector Extensions and Operator Optimization
    • 3.1 IREE: RVV f16 vfmacc and vfwmacc plus Zvfhmin fallback land on mainline (10-21)
    • 3.2 IREE: scalable vector distribution tiling aligned to base times vscaleMax (10-21)
    • 3.3 This stack’s vector repos and upstream validation line: VecIntrinBench, IntrinTrans, riscv-arch-test silent (10-21 status)
  • IV. Buddy Compiler Updates
    • 4.1 buddy-mlir: zero commits, zero releases in window (10-21 status)
    • 4.2 Downstream anchoring: triton-riscv nightly build still locked to 0.1.1.dev20261017 (10-21 status)
  • V. Community, Events, and Official Homepage
    • 5.1 Biweekly Report Issue 079 draft checked in: twelve OpenJDK merges, JEP 544 porting initially complete (10-22)
    • 5.2 Ecosystem: RISE releases Android 17 open-source developer builds for K3 Pico-ITX and A210 EVB (10-20)
    • 5.3 Official homepage three pages unchanged (2 blog posts, 23 documentation chapters) (10-21 status)
    • 5.4 Forum zero new posts, latest topic still the 10-15 daily news (10-21 status)
    • 5.5 External news and arXiv: zero hits for targeted queries on RuyiSDK, RuyiAI, openRuyi (10-21)
  • VI. Trend Observations
    • 6.1 Dependency closure enters wrap-up phase: Go shifts from “de-vendoring” to “filling in import paths”
    • 6.2 Two riscv branches fully rewritten on the same day: replay has become routine operations, review remains the bottleneck
    • 6.3 RVV backend fills in f16 before discussing throughput: the signal from IREE’s precision coverage ordering

Today’s Focus: IREE Mainline Lands RVV f16 Multiply-Accumulate MMA; LLVM riscv Branch Rebased and Rewritten After Eight Days of Silence

Date: 2026-10-21　Sources: IREE commit e7490c3b, IREE PR #25000, LLVM riscv branch commit list, two-head comparison

At 16:48 on 10-21, IREE merged the RISC-V V f16 inner_tiled MMA lowering into mainline (PR #25000, 946 lines changed, of which over 800 are tests). This is the second precision tier after the f32 vfmacc.vf path in v3.13.0 released on 10-19: the native f16 enum had already landed in mainline in an earlier commit, and this change adds the two lowering and widening paths—f16 times f16 yielding f16 goes through vfmacc.vf, while f16 times f16 yielding f32 goes through vfwmacc.vf; with only +zvfhmin, these degrade respectively to “widen the accumulator to f32, run vfmacc, then truncate back” and “widen both panels to f32 and run vfmacc”. Tile N is vlen/8, and at VLEN=256 the cost model counts the three paths as 14, 7, and 6 respectively.

The performance data offers something more useful than “supported”: on X60 (VLEN=256), single-threaded inner_tiled under both native f16 and vfwmacc encodings is already faster than the mmt4d ukernel; large shapes at 4 and 8 threads are still slower. In other words, instruction selection on a single core is already in place, and the remaining gap lies in tiling and threading; that tracking item is #24932.

On the same day, RuyiAI’s own LLVM branch ended eight days of silence. Yesterday’s report listed “whether the LLVM branch will restart on a new upstream baseline” as an observation point, and this window provides the answer: the riscv branch was rewritten wholesale at 23:00 on 10-21, with head moving from 0fb2da86 to dd05e5b5; SyncBots advanced the upstream integration to b3b5e33, and the comparison with the old head is diverged (1769 ahead, 26 behind, 300 files differing). The patch series within the branch is fully preserved: BOSC AME extension (#16), Buckyball vendor extension (#20), alignment with the Pebble core ISA (#23), AME path design (#24), xtame design update (#27), AME instruction update (#29), IME builtins to SpacemiT A100 dot instructions (#30), and A100 HP/SP and pack/unpack scheduling (#31, #32). The conclusion should be stated clearly: this is a rebase-and-replay, not new patches; the two FPGA PRs (#25, #28) had their timestamps refreshed in the same batch at 00:59 on 10-22 and remain open.

Placing the two side by side, today’s message is “upstream is moving forward, local work is waiting for review”: on the IREE side, one precision step from f32 to f16 was completed in two days, while on the RuyiAI side the compiler patches stay usable only by rebasing, and the rate at which the review queue is drained remains the real metronome of this chain.

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

1.1 pytorch: Third Consecutive Day of Whole-Branch Replay, Head Switched to 3c1ea315 (10-21)

Date: 2026-10-21　Sources: riscv branch commit list, new head 3c1ea315, two-head comparison, PR list

The riscv branch underwent its third whole-branch rewrite at 03:00 on 10-21, with the head switched from 7f49d33f to 3c1ea315; the head commit title remains “Add riscv64 image locally (#51)”, and the committer timestamps of all visible commits on the branch were uniformly rewritten to the same instant. The comparison against the previous head is diverged: 67 ahead, 24 behind, 208 files differing, with the delta concentrated in the upstream baseline sync range.

The local patch series is unchanged, item by item: riscv64 local image (#51), bytes_to_scalar floating-point fix (#44), inductor opinfo blocklist (#50), ULP tolerance restriction for linear cross entropy (#42), plus the CI series (#20 core/full split, #40 continue parsing after failure, #41 verdict bypass). Seven open PRs (#23, #35, #36, #38, #43, #46, #48) had their update timestamps refreshed in the same batch at 09:03 on 10-21, with no new review comments and no merges. Three consecutive days of replay indicate that the “daily baseline update, static patches” mechanism has become fixed; the cost is that any downstream reference by SHA breaks daily.

1.2 llvm-project: Eight-Day Silence Ends, riscv Branch Rebased to b3b5e33 (10-21)

Date: 2026-10-21　Sources: riscv branch commit list, new head dd05e5b5, IME builtin to A100 dot instruction #30, two FPGA PRs #25, #28

The riscv branch, silent for eight days since 10-13, was rewritten wholesale at 23:00 on 10-21, with the head changed from 0fb2da86 to dd05e5b5 and the title still “[RISCV] Add A100 HP/SP and pack/unpack lowering with calibrated scheduling (#31)”. The rebase was performed by SyncBots, advancing the upstream integration to LLVM b3b5e33; compared with the old head it is 1769 ahead, 26 behind, with 300 files differing, the bulk of the delta coming from the new baseline itself.

The branch contents, in order, are: BOSC AME extension support (#16), fix extension-name printing test (#18), Buckyball vendor extension (#20), ISA alignment with the Pebble core (#23), BOSC AME backend path design (#24), xtame design update (#27), AME instruction update (#29), IME builtin lowered to SpacemiT A100 dot instruction (#30), A100 HP/SP and pack/unpack lowering with scheduling calibration (#31, #32). All of these patches were merged in early October; there are no new merge records in this window.

The two FPGA backend PRs (#25 “BOSCAME Add FPGA backend support”, #28 “Feat/xboscame fpga”) had their update timestamps refreshed in the same batch at 00:59 on 10-22 and remain open; this is a side effect of the rebase, not review progress.

1.3 tilelang-riscv: Five Consecutive Days Without Commits (10-21 Status)

Date: 2026-10-21 (status)　Sources: commit list, PR #4 RVV mask and reduction

The last commit on the main branch remains the batch of five changes from 18:52 on 10-16 (codegen formatting conventions, upstream operator runtime semantics, TileOps runtime adaptation, Buddy MLIR release CI, building the Python extension in CI); there have been no new commits for five consecutive days since. Open PR #4 “Add RVV vector mask and max reduction lowering” has been stalled since 09-28, and PR #2 since 06-16. The repo remains in a lull where “CI and runtime scaffolding are in place, operator coverage still pending”; nothing in this window changes that assessment.

1.4 triton-riscv: Nightly Stays at the 10-19 Build, No New Activity on PR #85 (10-21 Status)

Date: 2026-10-21 (status)　Sources: Releases, Nightly 2026-10-18-a305fb6, PR #85 softmax and matmul optimization

No commits and no new releases in the window; the latest nightly is still nightly/20261018-a305fb6 from 10:01 on 10-19 (based on the 10-18 dependency upgrade, pinning Buddy MLIR nightly 0.1.1.dev20261017). The previous version was 20261011-fe554ff from 10-12; even after resuming releases, it has not returned to a daily cadence. Open PR #85 was last updated at 07:55 on 10-19 and remains the only active operator optimization proposal; the other open PRs (#83, #82, #58, #56, #55, #54) all last saw activity in mid-to-late September.

II. RuyiSDK Toolchain and Distributions

2.1 openRuyi main repo: 80 commits, Go dependency closure finalized and uv backend completed (10-21 to 10-22)

Date: 2026-10-21 to 2026-10-22　Source: commit list, umoci de-vendor commit 3012dbc8, sonobuoy de-vendor commit b5c8450b, rust-uv-build-backend commit 2413dc83, python-uv-build 0.13.4 commit 2c522b06

All 80 commits in the window carry the SPECS: prefix, from four authors—Leetfs 29, KimmyXYC 23, purofle 18, cl2t 10. They fall into four time batches: 10-21 16:58 (10 commits), 17:01 (18), 17:21 (8), 20:46 (15), plus 29 commits at 10-22 02:26.

The Go side accounts for 51 commits, along two lines. The first is filling in base dependencies: go-github-beevik-etree, ghodss-yaml, golang-plus errors/testing/uuid, googleapis-gnostic, liggitt-tabwriter, ma314smith-signedxml, texttheater-golang-levenshtein, tredoe-osutil (corresponding to #1361); followed by anthropic-sdk-go, discordgo, argo-go, bytedance’s gopkg and sonic-loader, ergochat’s irc-go and readline, vektah-gqlparser-v2, standard-webhooks, go-vcr, go-fyne-systray, go-google-api’s transport, etc. (purofle). The second is wrapping up de-vendoring: go-k8s-api, go-k8s-apimachinery, go-k8s-client-go add the import paths required after the vendored copies were removed, go-github-moby-sys adds the repo import path, go-github-smarty-assertions adds the legacy import path, and finally sonobuoy, umoci, go-github-vbatts-go-mtree land as “build without vendored dependencies”.

The Rust side’s 29 commits are concentrated at 10-22 02:26: 23 rust-uv-* packages (build-backend, cache-key, distribution-filename, pep440, pep508, platform-tags, pypi-types, small-str, static, toml, version 0.13, windows, warnings, etc., mostly at version 0.0.96), plus archive-trait 0.0.15, astral-async-zip 0.0.22, astral-mail-headers 0.0.1, tar-codec 0.0.15, tar-framing 0.0.15, and python-uv-build bumped to 0.13.4.

This batch lands on the Python build chain: with uv as a pip alternative frontend, once its Rust dependencies are complete in the distribution, installing Python packages on K1/K3 no longer needs to fall back to source builds, and completing uv-build-backend means the uv build backend in pyproject can be resolved natively.

2.2 openRuyi PRs: five merged, nineteen security updates under review (10-21)

Date: 2026-10-21　Source: PR #1466 umoci, PR #1464 sonobuoy, PR #1361 kubecomps, PR #1136 picoclaw dependencies, PR #1486 python-uv-build, PR #1470 llvm-snapshot

Five merged: #1361 kubecomps base dependencies (13 files, +1017), #1136 picoclaw dependencies first batch (19 files, +832), #1464 sonobuoy de-vendor (9 files, +268/-19), #1466 umoci de-vendor (20 files, +897/-13, including 15 commits and the full Apex Log dependency graph), #1486 python-uv-build 0.13.4 (58 files, +5640/-201). Yesterday’s report noted #1464 and #1466 were still in the review queue; both merged in this window, turning de-vendoring from proposal into mainline fact.

The 19 under review fall into three groups. Security updates dominate: go 1.27.3, nodejs and brotli, php 8.5.12, nss 3.131.1, ca-certificates-mozilla 2.92, rustup 1.29.2, python-pip 26.2.2, python-certifi 2026.10.19, perl-Mozilla-CA, rustls-webpki 0.103.16, webpki-roots 1.0.10 (two PRs). One toolchain PR, #1470, bumps llvm-snapshot to 20261019 (upstream hash 91c6a053) and adds -DLLVM_RAM_PER_LINK_JOB=30000 to limit link parallelism; #1471 bumps libudev-zero 1.0.5. The dependency and CVE batch continues: #1364 kubecomps network and database dependencies, #1468 and #1472 two CVE fixes plus package upgrades, #1469 new and upgraded packages, #1440 sequoia. No unmerged PRs were closed in the window.

2.3 ruyi package manager: 0.53.0 holds, 0.54.0 still without an alpha (10-21 status)

Date: 2026-10-21 (status)　Source: ruyi Releases, PyPI ruyi

Both GitHub Release and PyPI remain at 0.53.0 (released 09-24), and the last commit on main is from 09-25 (merging #503); no new version or commits in the window. The latest pre-release on GitHub is 0.54.0-alpha.20260918 from 09-18, with no alpha since. Per biweekly report issue 078 of 10-13, 0.54.0 is scheduled for end of October due to a release-cycle adjustment, with in-progress items being ABI compatibility checks between packages and the system, macOS packaging for more officially maintained packages, and SpacemiT K1/K3 toolchain adaptation; the package manager section of biweekly report 079’s draft is still empty.

2.4 ruyi-imager: three Gradle build refactors, nightly build not yet following (10-21)

Date: 2026-10-21　Source: commit list, nightly channel

All three commits in the window are by Glavo: 16:55 centralizes Gradle dependency versions into a shared version catalog (226d2732), 17:08 upgrades dependencies and bundled tools to stable maintained versions (313b56fb), 18:54 bumps TOMLJ to 2.2.0 and simplifies parsing dependencies (731e52b2). These are build engineering refactors and do not change image-writing logic.

The release side has not followed: the nightly channel’s asset is still
ruyi-imager-1.0.0-nightly.20261019T221503Z.7f62f9d (source commit 7f62f9dd, built 10-19 22:15, 10 assets). With three pushes to the code on 10-21 and no nightly rebuild, the channel’s automated build either did not trigger after the pushes or is still queued. For users validating RISC-V image writing with nightly builds, there is no new version available this period.

2.5 Kernel and firmware lines: linux-xuantie-kernel, edk2-ruyi, ruyifirmware quiet (10-21 status)

Date: 2026-10-21 (status)　Source: linux-xuantie-kernel commit list, edk2-ruyi commit list, ruyifirmware Releases

The kernel line linux-xuantie-kernel (linux-6.6.36 branch) still has its last commit at 10-16 11:14, the cpufeature and hwprobe change batch; the firmware line ruyifirmware’s latest release is still “Ruyi Firmware 20261015” from 10-15 17:53; edk2-ruyi is stuck at the K3 system information page from 10-15 15:22. All three lines have been quiet for a sixth consecutive day, continuing to contrast with the packaging peak in 2.1: this window’s effort again went entirely into userspace packages, with no new activity on kernel or firmware.

III. Vector Extensions and Operator Optimization

3.1 IREE: RVV f16 vfmacc and vfwmacc plus Zvfhmin fallback land on main (10-21)

Date: 2026-10-21　Source: commit e7490c3b, PR #25000, performance tracking #24932, f32 version PR #24933

This commit adds a lowering path for RISC-V V f16 inner_tiled MMA, reusing the lowerRiscvVFmaccLike introduced in #24933 and the hoisted accumulator legalization, with Tile N set to vlen/8. Encoding selection is split into three tiers by extension availability: f16 × f16 → f16 with +zvfh uses vfmacc.vf; f16 × f16 → f32 uses the widening vfwmacc.vf; with only +zvfhmin, the two paths respectively go through “extf the accumulator to f32, f32 vfmacc, truncf back” and “extf both panels, then f32 vfmacc”. When both tiers are legal, the native path is taken by enum order. At VLEN=256 the cost model reports intrinsics_m 14 (native), 7 (vfwmacc), and 6 (both zvfhmin fallbacks).

On the performance side there is only a directional conclusion: on X60 (VLEN=256), single-threaded inner_tiled with native f16 and vfwmacc is already faster than the mmt4d ukernel, while large shapes at 4 and 8 threads remain slower. Tests land with the commit, including a 187-line encoding materialization case, a 414-line lowering case, and build changes for the e2e matmul, giving directly regressable coverage.

3.2 IREE: scalable-vector distribution tiling aligned to base × vscaleMax (10-21)

Date: 2026-10-21　Source: commit 9ea0f2f2, PR #24849

The LLVMCPU backend has another scalable-vector change the same day: aligning scalable-vector distribution tiles to base * vscaleMax. This kind of alignment is routine legalization on the VLA (vector-length-agnostic) path — tile sizes must be divisible by an integer multiple of the hardware’s actual VLEN, otherwise tail-block degradation occurs on narrow-VLEN hardware. This change and the f16 MMA in 3.1 both belong to the CPU backend, and together they constitute the current RISC-V V progress in IREE.

3.3 This stack’s vector repos and upstream validation lines: VecIntrinBench, IntrinTrans, riscv-arch-test quiet (10-21 status)

Date: 2026-10-21 (status)　Source: VecIntrinBench, IntrinTrans, riscv-arch-test commit list, XiangShan commit list

This stack’s vector benchmarks and translation toolchain remain static: VecIntrinBench is stuck at 02-13, IntrinTrans at 06-25, the FBGEMM fork at 07-13, harness at 10-10, cpuinfo at 05-20, and torq-tile at 04-09, with no commits in the window. The upstream validation lines are equally inactive: riscv-arch-test and the main XiangShan repo (kunminghu-v3 branch) have zero commits in the window. All substantive vector-side progress this period comes from upstream IREE; this stack’s own benchmarks produced no new data.

IV. Buddy Compiler Updates

4.1 buddy-mlir: zero commits, zero releases in the window (10-21 status)

Date: 2026-10-21 (status)　Source: commit list, Releases

The last commit on main is still “[ci] Add the publish-models workflow (#925)” at 10-16 09:35, with 0 commits and 0 releases in the window. The version line remains the combination of the stable release BUDDY MLIR 0.1.1 (10-18 06:26) plus the last nightly 0.1.1.dev20261017 (10-17 07:24), with no new nightly builds for four consecutive days since. After the two stable releases 0.1.0 and 0.1.1 two days apart, things have stalled, a state that has now persisted for three reporting periods.

4.2 Downstream anchoring: triton-riscv nightly builds still pinned to 0.1.1.dev20261017 (10-21 status)

Date: 2026-10-21 (status)　Source: triton-riscv commit list, buddy-mlir nightly 0.1.1.dev20261017

The last commit on triton-riscv main (10-18 08:57, PR #88) anchors the downstream dependency to nightly/v0.1.1.dev20261017, and the nightly/20261018-a305fb6 released on 10-19 was built on that upgrade. buddy-mlir has had no new nightly since 10-17, so triton-riscv’s nightly builds have no new upstream to follow; the two sides are still waiting on each other: restoring daily output on this chain depends on when buddy-mlir restarts its nightly workflow.

V. Community, Events, and Official Homepage

5.1 Biweekly Report Issue 079 Draft Committed: Twelve OpenJDK Merges, JEP 544 Porting Initially Complete (10-22)

Date: 2026-10-22　Source: Biweekly Report 079 Draft, OpenJDK Update Commit #380, Base C Library Update Commit #381, Development Board Update Commit #382

Four commits landed in the biweekly report repository during the window: #379 initialized biweekly-79 via the workflow (10-21 10:20), #380 added OpenJDK content, #381 added base C library content, and #382 added development board content at 10-22 00:19. The committed file 20261022-ruyisdk-biweekly-79.md is dated 2026-10-22, and several sections of the body remain blank (foreword, package manager, GCC, LLVM, V8, Go, QEMU pending), indicating it is still in the pre-publication fill-in stage.

Three substantive sections have been written. OpenJDK lists the twelve mainline PRs reviewed and merged in this issue, covering the Zalasr extension (RVWMO memory model support, added to the template interpreter and C1/C2), dedicated vector mask logic instructions, a memory out-of-bounds fix for string_indexof_char plus removal of one branch, zext.h/zext.w for AndL on 16/32-bit masks, Zicond optimizations for zero-result integer cmove and heap object encode/decode, a QEMU user-mode test case fix, a JSR166 test case regression fix, and cleanup of ICache synchronization calls common to all platforms. The RISC-V port of JEP 544 (AOT static compilation) is initially complete and has entered functional testing and performance validation. On the base C library side, glibc ported sinh, sinhf, and sinpi into the existing libmvec framework, while newlib ported acosf, logf, asinf, asinhf, asinpif, and atanf into the vector math framework.

The development board section records existing items: the RT-Thread Smart 5.0.0 factory firmware test report for RDC2022-D1s (Allwinner D1s) (PR #410, submitted 10-16, still under review), and the ROS 2 chapters 9 and 10 course handouts and lab manuals for the SpacemiT K3 CoM260 Kit (submitted 10-14). Both predate this window; this issue merely compiles them into the publication.

5.2 Ecosystem: RISE Releases Android 17 Open-Source Developer Builds for K3 Pico-ITX and A210 EVB (10-20)

Date: Released 2026-10-20, discussion appeared on Hacker News at 10-21 05:51　Source: RISE Official Announcement, RISE Android 17 v1.0 Release, Hacker News Discussion

RISE announced the first Android 17 build running on RVA23 hardware, targeting the SpacemiT K3 Pico-ITX and Alibaba A210 EVB, project number RP024, funded and led by RISE, with SpacemiT and Alibaba DAMO providing platforms and hardware and BayLibre handling engineering implementation (building on its earlier K1 Android port). The release includes binary images for both boards, a pinned source manifest, checksums, flashing tools, and rebuild instructions; the boot chain is based on U-Boot and OpenSBI, the graphics stack uses the upstream Imagination kernel driver plus Mesa, and source history and authorship information are preserved.

The announcement explicitly states this is a developer build, not a production or certified version. Phase two will automate builds, hardware testing, and releases: first connecting to the A210 already in the RISE Board Farm, with K3 joining within weeks, after which both boards will run the same Android CI; RISE also plans follow-up RFPs for Bionic, BoringSSL, ART, V8, Chrome, and multimedia libraries, and will push vendors to upstream platform support to the Linux kernel, U-Boot, and OpenSBI.

The connection to this topic lies in the platform line: K3 is the platform whose course content appears in the biweekly report 079 draft (CoM260 Kit), and its firmware and kernel upstreaming demands align with the direction of RuyiSDK’s two silent lines on kernel and firmware. The Android 17 K3 image itself does not fall within RuyiSDK’s distribution scope, but it makes the platform’s current boot chain public as a reproducible baseline, serving as a comparable reference for teams subsequently building Debian-family images.

5.3 Official Homepage: Three Pages Unchanged (2 Blog Posts, 23 Documentation Chapters) (10-21 Status)

Date: 2026-10-21 (status)　Source: RuyiAI Official Homepage, Blog, Documentation

The three-page snapshot is identical to the previous issue: index.html at 7246 bytes with 3 index entries; blog.html at 5632 bytes with 2 index entries (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices); docs.html at 10843 bytes with 23 index entries. No new blog posts and no new documentation chapters this round.

5.4 Forum: Zero New Topics, Latest Still the 10-15 Daily News (10-21 Status)

Date: 2026-10-21 (status)　Source: RuyiSDK Community Forum, Latest Topic #2892

Zero new topics in the forum during the window; the latest post remains “2026-10-15 Open Source Technology Hot News” (#2892) from 10-15 03:30. Biweekly Report Issue 079 has not yet been published on the forum, and the latest issue on the community side remains Issue 078 (#2891) from 10-13. After the summit (10-18 to 10-20) concluded, no post-summit summary thread has appeared on the forum, and public output remains concentrated on the two fixed lines of daily news and biweekly reports.

5.5 External News and arXiv: Zero Hits for Targeted Queries on RuyiSDK, RuyiAI, openRuyi (10-21)

Date: 2026-10-21　Source: Google News RSS, arXiv cs.AR Recent List

Targeted Google News queries in Chinese and English for RuyiSDK, RuyiAI, and openRuyi returned zero hits during the window (the bare query channel works normally; same-batch RISC-V queries return results). Queries of the “Ruyi RISC-V distribution” type returned five results, all old articles, the most recent being a 10-20 XuanTie shipment piece unrelated to the distribution. The latest RISC-V-related preprint on arXiv is from 10-20, on context-action intent verification for industrial cyber-physical systems; there are no new papers in the window on RISC-V AI compilers or vector optimization.

VI. Trend Observations

6.1 Dependency closure enters its endgame: Go shifts from “de-vendoring” to “filling in import paths”

The package-addition threads in the previous three issues were laid out separately: Python filling out the full fsspec cloud-storage family, Rust filling out the crate closure, and Go starting on de-vendoring. This issue, the center of gravity on the Go side’s 51 commits has clearly shifted later—first filling in base dependencies (etree, yaml, gnostic, tabwriter, and about ten others), then filling in the import paths for the k8s trio plus moby-sys and smarty-assertions, and only finally landing the three genuinely “build without vendored dependencies” commits for sonobuoy, umoci, and go-mtree.

The ordering itself shows that the completion condition for de-vendoring is not deleting the vendor directory, but making every depended-upon module resolvable in the distribution under its original import path; without filling in the missing import paths, deleting vendor simply causes an immediate compile failure. So “filling in import paths” is the last mile of this round. The direct implication for RISC-V users is that container image tools and conformance testing tools go from “packaged with the source” to “packaged per distribution dependencies,” improving both build reproducibility and the propagation path for security updates—future CVEs only require upgrading a single Go module, not rebuilding the whole package.

6.2 Two riscv branches fully rewritten on the same day: replay has become routine operations, review remains the bottleneck

This issue, the riscv branches of pytorch and llvm-project each underwent a full-branch rewrite on the same day: pytorch at 03:00 (the third consecutive day), llvm at 23:00 (a rebase after eight days of silence). The common features are uniformly rewritten committer timestamps, an unchanged set of patch titles, and SHA values that expire daily; the difference is cadence—pytorch is a daily-updated baseline, llvm rebases on a chosen schedule, and after rebasing the diff against the old head reaches 300 files, almost all from the new upstream.

The engineering significance of this approach is decoupling “tracking upstream” from “maintaining local patches,” at the cost of untraceable history. Any downstream that references by SHA (image builds, CI caches, issue reproduction) will break daily or after each rebase; these two branches should be versioned by their set of patch titles rather than by SHA. More noteworthy is that the patches themselves added nothing in this window: pytorch’s seven open PRs only refreshed timestamps, and llvm’s two FPGA PRs likewise only refreshed timestamps. The replay mechanism is already solid; the actual metronome has become the digestion rate of the review queue.

6.3 RVV backend fills in f16 before discussing throughput: the signal from IREE’s precision-coverage ordering

Looking at the 10-19 f32 vfmacc.vf (v3.13.0) together with the 10-21 f16 vfmacc/vfwmacc (#25000), IREE climbed one precision step in two days on the RISC-V V inner_tiled MMA path, reusing the same lowerRiscvVFmaccLike and the same accumulator legalization both times. The order is to first guarantee a correct path (f32), then pick the fastest one according to extension availability (native f16, widened f16, +zvfhmin fallback), with the cost model pricing the three paths separately.

Measurements on the X60 make the bottleneck location clear: single-threaded is already faster than the mmt4d ukernel, while 4-thread and 8-thread are still slower, indicating that once single-core instruction selection is in place, what remains is tiling and threading. For tilelang-riscv and triton-riscv, this is a roadmap that can be compared item by item—the Tile N = vlen/8 derivation and the cost model’s tiering are both worth comparing on the same basis in their own lowering, rather than just checking “whether f16 is supported.”

Source Verification Table

Source Verification Result
RuyiAI-Stack org (all repos) pytorch: riscv branch’s third full-branch replay (head 7f49d33f replaced by 3c1ea315, compare ahead 67/behind 24, 208 files), seven open PRs refreshed together at 10-21 09:03; llvm-project: riscv branch rebased at 10-21 23:00 after eight days of silence (head dd05e5b5, ahead 1769/behind 26, 300 files), two FPGA PRs refreshed at 10-22 00:59; triton-riscv and tilelang-riscv zero commits and zero releases; ruyiai-daily 1 issue (10-21); ruyiai-stack.github.io 1 contributor update; harness, FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, torq-tile, dsh-osc, skills, SyncBots-DeepAgents no in-window commits
ruyisdk org (all repos) ruyi: 0 commits, 0 releases (0.53.0 maintained); ruyi-imager: 3 Gradle build refactors, nightly build still at 20261019T221503Z; wechat-articles: 4 commits (biweekly 079 draft); packages-index, ruyisdk-website, support-matrix, linux-xuantie-kernel, qemu, ruyici, ruyi-tuner, ruyi-backend all 0 commits; llvm-project AME branch stalled at 10-19
openRuyi-Project org (all repos) openRuyi main repo 80 commits (Go 51, Rust 29, four authors) + 5 PRs merged + 19 under review, 0 closed unmerged; ruyifirmware and edk2-ruyi no in-window commits (stalled at 10-15 17:53 and 10-15 15:22 respectively)
buddy-compiler org buddy-mlir main branch 0 commits, 0 releases; latest stable 0.1.1 (10-18), latest nightly 0.1.1.dev20261017 (10-17)
RuyiSDK community forum Zero new topics in window, latest is #2892 (10-15); biweekly 079 not yet published, latest remains issue 078
openRuyi distribution news page No new articles, the 2026.09 release announcement remains latest, the 2026.10 page is not yet live
ruyi package manager / PyPI 0.53.0 remains latest (released 09-24), no new versions or commits in window
Official homepage (ruyiai.org) Three pages unchanged: homepage index 3 items, blog 2 posts, docs 23 chapters
Tech intelligence repo (ruyiai-daily) Pushed one issue on 10-21, RISC-V section includes RISE’s K3 Pico-ITX and A210 EVB Android 17 images
IREE upstream 4 commits, including RVV f16 vfmacc/vfwmacc with Zvfhmin fallback (#25000, 946 lines) and scalable tiling alignment (#24849); no new official release in window
riscv-arch-test / XiangShan Both repos zero commits in window
arXiv (RISC-V) Latest RISC-V preprint is 10-20, no new papers on compilers or vector optimization in window
Google News RSS (multiple Chinese and English queries) RuyiSDK / RuyiAI / openRuyi all zero hits; five Ruyi-type queries all returned old articles

Complete Source List