Research window: 2026-10-20 04:30 ~ 2026-10-21 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, the ruyiai-daily intelligence repository, three-page snapshots of the official RuyiAI homepage, upstream releases of IREE and riscv-arch-test, arXiv searches, and targeted Google News RSS queries in Chinese and English (via proxy; results in 5.5).

This Issue’s Index

  • Today’s Focus: openRuyi logs 137 commits in a single day — the full fsspec cloud storage backend family is completed, and Go de-vendoring lands on main (10-20)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 pytorch: replayed again today, head switched to 7f49d33f, upstream baseline advances with main (10-20)
    • 1.2 llvm-project: riscv branch silent for eight days since 10-12, two FPGA PRs stalled (10-20 status)
    • 1.3 tilelang-riscv: no commits for four consecutive days after five on 10-16 (10-20 status)
    • 1.4 triton-riscv: nightly stays at the 10-18 build, no activity on main or the PR queue (10-20 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi main repo: 137 commits, 42 each across the Rust / Go / Python lines (10-20)
    • 2.2 openRuyi PRs: ten merged, the large CVE PR closed without merging, two Go de-vendoring PRs awaiting review (10-20)
    • 2.3 ruyi package manager: stays at 0.53.0, main has had no commits since 09-25 (10-20 status)
    • 2.4 Images and distributions: ruyi-imager nightly updated, openRuyi 2026.10 not yet released (10-20)
    • 2.5 Kernel and firmware line: linux-xuantie-kernel, ruyifirmware, and edk2-ruyi all silent (10-20 status)
  • III. Vector Extensions and Operator Optimization
    • 3.1 IREE 3.13.0 stable: RISC-V V f32 vfmacc.vf MMA enters release (10-19)
    • 3.2 Verification and cache side: three riscv-arch-test merges, XiangShan CoupledL2 banked memory (10-19 to 10-20)
    • 3.3 This stack’s vector repos: VecIntrinBench / IntrinTrans / FBGEMM silent (10-20 status)
  • IV. Buddy Compiler Updates
    • 4.1 buddy-mlir: no commits and no new nightly for three days after 0.1.1 (10-20 status)
    • 4.2 Downstream pinning: triton-riscv nightly still locked to 0.1.1.dev20261017 (10-20 status)
  • V. Community, Events, and Official Site
    • 5.1 Official site’s three pages unchanged (2 blog posts, 23 doc chapters) (10-20 status)
    • 5.2 Forum: zero new posts, latest topic still the 10-15 daily briefing (10-20 status)
    • 5.3 Industry: DAMO XuanTie E908A automotive-grade CPU IP released, C950A targets 5x performance (10-19)
    • 5.4 Industry: SpacemiT V100 three machine types and compatible distribution list, openRuyi included (10-19)
    • 5.5 External news: targeted queries for RuyiSDK / RuyiAI / openRuyi return zero hits (10-20)
  • VI. Trend Observations
    • 6.1 Dependency closure work proceeds in parallel across three language lines: Rust adds a uv backend, Go de-vendors, Python adds cloud storage
    • 6.2 The PyTorch branch has entered a daily replay rhythm: baseline updated daily, patch series unchanged
    • 6.3 What server-class platforms bring to distributions is a compatibility list, not new porting work

Today’s Focus: openRuyi Lands 137 Commits in a Single Day — Full fsspec Cloud Storage Backend Family Completed, Go De-vendoring Merged to Main

Date: 2026-10-20　Sources: openRuyi commit list, fsspec 2026.6.0 commit 9bc134bc, s3fs 2026.6.0 commit 6b61df84, gcsfs commit 80b1a07a, minio de-vendoring commit ca13b2cf, PR #1459, PR #1355

The openRuyi main repository received 137 commits within the window (author timestamps 10-20 12:46 to 20:24), a noticeably more dispersed structure than the previous edition’s “one-sided Rust crate imports”: Rust 42, Go 42, Python 42, and 11 others. Four authors — Leetfs 94, HNO3Miracle 26, KimmyXYC 12, cl2t 5. None of the commit messages mention riscv; the day was about building out dependency closures rather than architecture-specific backports.

Of the 42 Python commits, 28 land in fsspec and the various cloud vendor backends: fsspec 2026.6.0, s3fs 2026.6.0, gcsfs, plus adlfs (Azure), ocifs (OCI), dropboxdrivefs, smbprotocol and pyspnego (SMB), and cloudpathlib; the accompanying auth and API dependencies are added alongside, including azure-identity, msal, msal-extensions, the aio extension of azure-storage-blob, google-cloud-storage and its control interface, google-auth-oauthlib, google-cloud-core, google-api-core, google-resumable-media, grpc-google-iam-v1, proto-plus, crc32c, and circuitbreaker. This layer has direct implications for RuyiAI: PyTorch DataLoaders and training scripts commonly read remote datasets through the fsspec abstraction, and with the backends in place, running real training on K1/K3 no longer requires on-the-spot pip compilation.

The 42 Go commits continue the wrap-up of minio de-vendoring and fill in observability: minio-mc, console, madmin-go-v3, prometheus, prom2json, alertmanager, openzipkin, applicationinsights, yunion’s s3cli and log, the Alibaba Cloud darabonba client, k8s structured-merge-diff, and go-otel 1.46.0. The 42 Rust commits cluster into three families: the windows-* 0.61 family (core, future, result, strings, threading, numerics, collections), zerocopy 0.8.63 with the bytecheck/rkyv family, and cap-std 4 with the io-lifetimes/io-extras/fs-set-times family; together with the same day’s python-uv-build, this appears to be paving the way for the dependency closure of the uv build backend.

Ten PRs merged during the day: #1459 minio de-vendoring (20 files +973/-43), #1355 first batch of kubecomps Go dependencies, #1128 Prometheus 3.13 module (DAG L4-12), #1344 fsspec CVE fix (32 files +1156/-24), #1342 CVE Part 1, #1461 graphql-core 3.3.0, #1462 buildx, #1372 moniker, #1463 dropwatch, #1458 uwsgi. #1341 “CVE Vulnerability Fixes (Part 0)” was closed with 50 files +997/-4788 but not merged; the same direction is continued by two new tickets, #1468 and #1469.

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

1.1 pytorch: Replayed again same day, head switched to 7f49d33f, upstream baseline advances with main (10-20)

Date: 2026-10-20　Sources: riscv branch commit list, new head 7f49d33f, two-head comparison, PR list

Within the window, the riscv branch was again fully rewritten and force-pushed at 10-20 08:32, with refs/heads/riscv updated from 44826095 to 7f49d33f; all branch commit author timestamps were uniformly rewritten to 10-20 03:00, and the head commit title remains “Add riscv64 image locally (#51)”. Counted by commit date, there were 0 new additions in the window; this is still a replay rather than new patches.

Compared with the previous head: GitHub compare reports diverged (60 ahead, 24 behind, 121 files changed), with the delta concentrated in the upstream baseline sync range — the new head contains a batch of new upstream commits including #201423 and #201490, while several upstream commits from the old head are no longer retained in the new tree. The basis for judging that the replay did not alter local work is the patch sequence: #51 (riscv64 local image), #44 (bytes_to_scalar floating-point fix), #50 (inductor opinfo blocklist), #42 (linear cross entropy ULP tolerance), and the CI series (#20 core/full split, #40 continue parsing after Jenkins failure, #41 verdict bypass) are consistent item by item across commit titles at each layer. The seven open PRs (#23, #35, #36, #38, #43, #46, #48) all had their update timestamps refreshed in the same batch at 10-20 08:32, with no new review comments and no merges.

1.2 llvm-project: riscv branch silent for eight days since 10-12, two FPGA PRs stalled (10-20 status)

Date: 2026-10-20 (status)　Sources: riscv branch commit list, PR list, PR #25 BOSCAME FPGA backend, PR #28 xboscame FPGA

The riscv branch head remains 0fb2da86 “[RISCV] Add A100 HP/SP and pack/unpack lowering with calibrated scheduling (#31)” (10-12 23:00), with no new pushes for eight consecutive days since 10-13. The two FPGA backend PRs (#25, #28) last updated in the early hours of 10-13, with no review activity in the window. The main branch continues its daily upstream sync as usual (a batch of gn build and SPIRV/GlobalISel changes at 10-20 14:42), not itemized here per the filtering rules.

Yesterday’s report raised “whether the LLVM branch will restart with a new upstream baseline” as an observation point; there was no restart in this window, indicating that RuyiAI’s LLVM investment remains in a phase of awaiting review and scheduling a replay.

1.3 tilelang-riscv: No commits for four consecutive days after the five commits on 10-16 (10-20 status)

Date: 2026-10-20 (status)　Sources: commit list, PR #4 RVV mask and reduction

The last commit on the tilelang-riscv main branch remains the batch of five changes at 10-16 18:52 (codegen formatting conventions, upstream operator runtime semantics, TileOps runtime adaptation, Buddy MLIR release CI, building Python extensions in CI), with no new commits for four consecutive days since. Open PR #4 “Add RVV vector mask and max reduction lowering” has been stalled since 09-28, and PR #2 stalled on 06-16. The repo remains in a lull where “CI and runtime skeleton are in place, operator coverage pending”.

1.4 triton-riscv: Nightly stays at the 10-18 build, no activity on main branch or PR queue (10-20 status)

Date: 2026-10-20 (status)　Sources: Releases, Nightly 2026-10-18-a305fb6, PR #85 softmax and matmul optimization

triton-riscv had no commits and no new releases in the window; the latest nightly remains nightly/20261018-a305fb6, released 10-19 10:01 (corresponding to the 10-18 dependency upgrade on main, Buddy MLIR nightly v0.1.1.dev20261017). The previous one was 20261011-fe554ff on 10-12; since resuming releases it has not yet returned to a daily cadence. Open PR #85 has had no new activity since 10-19 07:55 and remains the only active operator optimization proposal.

II. RuyiSDK Toolchain and Distributions

2.1 openRuyi main repo: 137 commits, 42 each across Rust / Go / Python (10-20)

Date: 2026-10-20　Source: commit list, python-fsspec 2026.6.0 commit 9bc134bc, python-opentelemetry-api 1.44.0 commit e19d4e69, rust-zerocopy 0.8.63 commit f3e9db3b, python-uv-build commit

By committer time, the 137 commits in the window are distributed across 12:46 (26 commits), 15:10 to 15:29 (66), 16:08 (30), 19:12 to 19:21 (13), and 20:23 to 20:24 (2). The symmetric structure of 42 commits per language is the most striking feature of this issue: of the 221 commits in the previous issue, Rust accounted for 211; this issue Rust drops to 42, with Go and Python each filling in 42.

On the Python side, besides cloud storage backends, there are also langchain 1.3.9 and langchain-core 1.4.6, Django 5.2.17, setuptools 83.0.0, pip 26.2, pymongo 4.18.1, h2 4.4.1, soupsieve 2.9, webob 1.8.11 and other upgrades tagged with security fixes. On the Go side, besides minio and the Prometheus family, there are nvidia go-nvml, google go-tdx-guest and go-sev-guest, module declarations for aws-sdk-go-v2’s appconfig/ssm, and internal module declarations for azure-sdk-for-go — declarative changes aimed at “letting downstream packages resolve dependencies by module.” On the Rust side, the only version updates to non-new packages are zerocopy and zerocopy-derive to 0.8.63, two commits; the remaining 40 are all new package imports.

The other 11 commits that day are: uwsgi, dropwatch, the minio de-vendoring rework, buildx, and a batch of 7 security updates at 15:10 — containerd 2.3.5, harfbuzz 14.0.0, exiv2 0.28.9, open-iscsi 2.1.13, ppp 2.5.4, avahi 0.9~rc5, gi-docgen 2025.5.

2.2 openRuyi PRs: ten merged, large CVE PR closed without merge, two Go de-vendoring PRs pending review (10-20)

Date: 2026-10-20　Source: PR #1341, PR #1342, PR #1344, PR #1466 umoci, PR #1464 sonobuoy, PR #1468, PR #1469, PR #1440 sequoia

A total of 10 PRs were merged in the window, in chronological order: #1128 (Prometheus 3.13 modules, 04:46), #1342 (CVE Part 1, 07:10), #1461 (graphql-core 3.3.0, 07:12), #1355 (first batch of kubecomps Go dependencies, 07:29), #1462 (buildx, 08:02), #1344 (fsspec CVE, 08:08), #1459 (minio de-vendoring, 11:12), #1372 (moniker, 11:21), #1463 (dropwatch, 12:23), #1458 (uwsgi, 12:24).

The large CVE PR #1341 “CVE Vulnerability Fixes (Part 0)” was closed without merge at 50 files, +997/-4788 — a clear disposition compared with the still-under-review status reported yesterday; the same day saw newly opened #1468 “Fix CVEs and upgrade some packages” (34 files, +2023/-594) and #1469 “Add packages and upgrade packages” (9 files, +223/-21), indicating the original large PR was split up and resubmitted.

In the pending queue, de-vendoring continues to advance: #1466 umoci (22 files, +971/-13) and #1464 sonobuoy (9 files, +268/-19) replace the bundled vendor archives of the container image and conformance testing tools with distribution-packaged dependencies; #1467 fixes opencv’s hardcoded /home build path; #1440 sequoia (21 files, +2799/-78) and #1456 picoclaw 0.3.1 (43 files, +2080/-42) remain queued.

2.3 ruyi package manager: 0.53.0 holds, no commits on main since 09-25 (10-20 status)

Date: 2026-10-20 (status)　Source: ruyi Releases, PyPI ruyi, Biweekly Report No. 078

Both GitHub Release and PyPI remain at 0.53.0 (released 09-24), and the last commit on main is stuck at 09-25 (merging #503); no new version and no new commits in the window. According to Biweekly Report No. 078 of 10-13, 0.54.0 is scheduled for late October due to a release cycle adjustment, with in-development items being ABI compatibility detection between packages and the system, macOS packaging of more officially maintained packages, and SpacemiT K1/K3 toolchain adaptation. With about ten days left in the month, the watch point remains whether an alpha version ships first.

2.4 Images and distributions: ruyi-imager nightly build updated, openRuyi 2026.10 not yet released (10-20)

Date: 2026-10-20　Source: ruyi-imager nightly build, openRuyi distribution news page, packages-index commit list

The ruyi-imager nightly channel was updated at 10-20 06:22. Its package version is
1.0.0-nightly.20261019T221503Z.7f62f9d, source commit 7f62f9dd, with 10 assets attached. packages-index (last commit 10-13), ruyisdk-website (10-19 07:54), support-matrix (10-16 11:46), and the AME branch of ruyisdk/llvm-project (PR #170 CI workflow at 10-19 13:32) all saw no new commits in the window.

The openRuyi distribution news page has no new articles; the 2026.09 release announcement remains the latest, and the 2026.10 page is not yet online. Following the monthly cadence, the October release has entered its end-of-month window, while the main repo was still bulk-adding packages that day — the release timing is the next watch point.

2.5 Kernel and firmware lines: linux-xuantie-kernel, ruyifirmware, edk2-ruyi all silent (10-20 status)

Date: 2026-10-20 (status)　Source: linux-xuantie-kernel commit list, ruyifirmware Releases, edk2-ruyi commit list

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

III. Vector Extensions and Operator Optimization

3.1 IREE 3.13.0 Official Release: RISC-V V f32 vfmacc.vf MMA Enters Release (10-19)

Date: 2026-10-19　Source: IREE v3.13.0 Release, PR #24933 RISC-V V inner_tiled MMA

IREE v3.13.0 was released on 10-19 at 19:26. The CPU backend adds an inner_tiled MMA for RISC-V V f32 vfmacc.vf (PR #24933), mapping vector floating-point multiply-add directly onto RISC-V V’s fused multiply-accumulate. Meanwhile, LLVMCPU scalable vectorization now requires --iree-llvmcpu-vscale-range to be specified, taking the larger of the two lower bounds when the +zvl*b feature is present. On the model side, PT2E per-tensor, per-channel-group, and per-token quantized import into LinalgExt Q/DQ is retained.

The significance for this stack lies in the path comparison: RuyiAI’s tilelang-riscv and triton-riscv take the Tile-level DSL to RVV route, while IREE takes MLIR→LLVMCPU scalable vectorization. On the same RVV, both frontends simultaneously select f32 multiply-add as an optimization target, indicating that this instruction form has become a consensus landing point for RISC-V AI backends.

3.2 Verification and Cache Side: Three riscv-arch-test Merges, XiangShan CoupledL2 Banked SRAM (10-19 to 10-20)

Date: 2026-10-19 to 2026-10-20　Source: riscv-arch-test PR #2413 Sdtrig, PR #2791 S_F/S_V/S_VF, PR #2858 Vector Instruction SEW Coverage, XSCache PR #36, XiangShan PR #6738

Three merges landed within the RISC-V architecture test window: Sdtrig’s A-extension trigger coverage points and Sm/S/U three-mode icount coverage (10-20 05:53); adding S_F/S_V/S_VF to the S suite to cover the privilege levels actually used by operating systems (10-20 02:53); and extending vector instructions to all valid SEWs while filling in SEW=8 vrgatherei16 and fault-only-first LMUL selection (10-19 21:55). Together, the three push the verification surface toward “S-mode + full SEW,” providing a usable regression baseline for teams running RVV operators on K1/K3.

On the XiangShan side, CoupledL2 splits directory and data SRAM banks by set index, restricting MCP2 stalls to consecutive accesses within the same bank, and pauses acceptance of new prefetches when RequestBuffer occupies 2 slots (PR #36, merged 10-19 15:12; default bank split value is 2, revertible to 1); the main repository subsequently updated the XSCache pointer to incorporate this change (PR #6738). Cache parallelism and prefetch admission are adjusted together—a verifiable microarchitectural change.

3.3 This Stack’s Vector Repos: VecIntrinBench / IntrinTrans / FBGEMM Quiet (10-20 Status)

Date: 2026-10-20 (status)　Source: VecIntrinBench, IntrinTrans, FBGEMM fork

None of this stack’s three vector benchmark and transpilation toolchain repos had commits within the window; their latest activity stopped at 02-13 (VecIntrinBench), 06-25 (IntrinTrans), and 07-13 (FBGEMM) respectively; harness stopped at 10-10, cpuinfo at 05-20, and torq-tile at 04-09. The only external progress on the operator side within the window is the release of IREE 3.13.0’s RVV MMA (see 3.1); this stack’s own vector benchmarks produced no new data.

IV. Buddy Compiler Updates

4.1 buddy-mlir: No Commits, No New Nightly Builds for Three Days After 0.1.1 (10-20 Status)

Date: 2026-10-20 (status)　Source: Commit List, Releases

The last commit on the buddy-mlir main branch remains “[ci] Add the publish-models workflow (#925)” at 10-16 09:35, with 0 commits and 0 releases within the window. The version line is: after the official BUDDY MLIR 0.1.1 (10-18 06:26), the last nightly build was 0.1.1.dev20261017 (10-17 07:24), with no new nightly builds for three days thereafter. The two official releases from 0.1.0 to 0.1.1 were two days apart (10-16, 10-18), after which activity stalled.

4.2 Downstream Anchoring: triton-riscv Nightly Build Still Locked to 0.1.1.dev20261017 (10-20 Status)

Date: 2026-10-20 (status)　Source: triton-riscv 10-18 Dependency Upgrade Commit, buddy-mlir Nightly 0.1.1.dev20261017

The last commit on the triton-riscv main branch at 10-18 08:57 is “Update dependency buddy-compiler/buddy-mlir to nightly/v0.1.1.dev20261017,” anchoring the downstream dependency to that nightly build; the nightly/20261018-a305fb6 released on 10-19 was built on this upgrade. Since buddy-mlir has had no new nightly builds since 10-17, triton-riscv’s nightly builds lack a new upstream to follow, leaving both sides waiting on each other—whether this chain resumes daily builds depends on when buddy-mlir restarts its nightly workflow.

V. Community, Events, and Official Homepage

5.1 Official homepage’s three pages unchanged (2 blog posts, 23 documentation chapters) (10-20 status)

Date: 2026-10-20 (status)　Source: RuyiAI Official Homepage, Blog, Documentation

The snapshots of the official homepage’s three pages are consistent with the previous issue: index.html is 7246 bytes with 3 index entries; blog.html is 5632 bytes with 2 index entries (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices); docs.html is 10843 bytes with 23 index entries. This round has no new blog posts and no new documentation chapters.

5.2 Forum has zero new posts; latest topic remains the 10-15 daily news (10-20 status)

Date: 2026-10-20 (status)　Source: RuyiSDK Community Forum, Latest Topic #2892

The forum window has zero new topics; the latest post is still “2026-10-15 Open Source Technology Hot News” (#2892) at 03:30 on 10-15. Going back in order: Biweekly Report Issue 078 (#2891, 10-13), Summit participation announcement (#2889, 10-13), practice course handouts and environment experiments (#2883, #2885). After the summit (10-18 to 10-20) ended, no post-summit recap post appeared on the forum; the community’s public output remains concentrated on the two fixed lines of daily news and biweekly reports.

5.3 Industry side: DAMO XuanTie E908A automotive-grade CPU IP released, C950A targets 5x performance (10-19)

Date: 2026-10-19　Source: ITHome: XuanTie E908A

At the 10-19 RISC-V China Summit, DAMO XuanTie released the automotive-grade enhanced CPU IP E908A, claiming it is the first RISC-V CPU IP to simultaneously achieve ASIL-B and ASIL-D, with single-point safety metric SPFM exceeding 90% for the first time, targeting safety islands, domain controllers, and automotive gateways; it also presented for the first time a product roadmap for autonomous driving and embodied intelligence, with the next-generation C950A targeting 5x the performance of existing RISC-V automotive-grade products. The summit also mentioned that C950 improves by more than 10% over mainstream traditional-architecture processors on four types of Agentic workloads, with Qwen3.8-Flash-Next inference reaching 43 token/s.

The distance from this topic depends on the software stack: the deployment of automotive-grade IP relies on functionally safety-certified toolchains and RTOS/kernel combinations, whereas the current RuyiSDK toolchain adaptation is concentrated on general-purpose platforms such as K1/K3; safety-certification-related compiler qualifications, test suites, and documentation systems have not appeared in the public roadmap. E908A is a demand signal, not a current task.

5.4 Industry side: SpacemiT V100’s three machine types and compatible distribution list, with openRuyi included (10-19)

Date: 2026-10-19　Source: EETrend: SpacemiT V100, LinuxGizmos: 46-core RISC-V server CPU

Yesterday’s report covered the release of the V100 chip itself; reports in this window fill in the system-side information: the V100 is a 46-core hybrid architecture (6 XiangShan Kunminghu V2R2 plus 40 X100), up to 2.4GHz, dual-socket coherent interconnect up to 92 cores, single-chip 8-channel DDR5 and 52 PCIe 5.0 lanes; the accompanying machines are three types—C2201 compute, G5201 intelligent compute, and S5236 storage—among which the G5201 has validated domestic accelerator cards such as Kunlunxin M100 and supports RDMA cluster networking.

The most critical sentence is in the distribution list: the report states that the V100 passes all BRS specification tests and supports direct installation, boot, and operation of mainstream distributions including openEuler, openRuyi, openKylin, Debian, Ubuntu, Fedora, Red Hat, and Rocky. openRuyi being included in the vendor’s server platform compatibility list has not appeared in this topic’s previous public sources; however, the openRuyi main repository still had no server-platform-related commits that day, and whether the list can be fulfilled depends on the 2026.10 release and subsequent firmware-line actions.

5.5 External news: targeted queries for RuyiSDK / RuyiAI / openRuyi return zero hits (10-20)

Date: 2026-10-20　Source: Google News RSS, arXiv cs.AR recent list

Google News targeted queries in Chinese and English for RuyiSDK, RuyiAI, and openRuyi (including “Ruyi Community RISC-V” and “Ruyi RISC-V distribution”) all returned zero hits, with the channel functioning normally (the same batch of RISC-V queries returned 62 results, indicating both the proxy and parsing are usable). One hit for “Ruyi RISC-V distribution” was actually a XuanTie shipment article, unrelated to distributions. The latest RISC-V-related preprint on arXiv in the past seven days is from 10-15 (FPGA efficiency, constant-time compilation pipelines, etc.); there are no new papers in the window directly related to this stack’s compiler or vector optimization.

Summit-related Chinese reports are concentrated on the hardware and capital sides: XuanTie shipments exceeding 6 billion units and pushing C950 into AI servers, XuanTie E908A automotive-grade IP, SpacemiT V100’s three machine types, and Tenstorrent’s $1.5 billion financing ($6.2 billion valuation). None involve specific solutions for RuyiSDK or openRuyi.

VI. Trend Observations

6.1 Dependency closure work proceeds in parallel along three language-specific tracks: Rust fills in the uv backend, Go removes vendoring, Python fills in cloud storage

The 42/42/42 structure of this period’s 137 commits is no coincidence; it reflects three independent package-completion tracks advancing on the same day. On the Rust side, work centers on three families—windows-* 0.61, zerocopy/rkyv, and cap-std/io-lifetimes—together with python-uv-build, pointing to the Rust dependencies of Python build tools such as uv. On the Go side, the work is mainly minio and Prometheus/Kubernetes components, together with three de-vendoring PRs that change “bundle with source” into “bundle by dependency.” On the Python side, all major cloud backends for fsspec are filled in at once. The common goal of all three tracks is to complete the dependency closure so that Python AI and cloud-native deployments on RISC-V do not need to compile on site.

Worth noting is how this layer relates to RuyiAI: PyTorch’s DataLoader, checkpoint, and dataset tooling generally access remote storage through fsspec, so whether the backends are complete directly determines whether training scripts can run on K1/K3. This portion of dependencies that the distribution fills in for the upper layers matters more to “whether it can be used” than any single operator optimization.

6.2 The PyTorch branch enters a daily-replay rhythm: baseline updated daily, patch series unchanged

The same pattern has been observed for two consecutive days: a full-branch rewrite push at 09:14 on 10-19 and another at 08:32 on 10-20, with committer timestamps across the whole branch uniformly rewritten, local patch titles unchanged one by one, and the only change being the sync range of the upstream baseline. This approach decouples “tracking upstream” from “maintaining patches,” at the cost of an untraceable branch history—any downstream that references by SHA (CI caches, image builds, issue reproduction) will break every day. If this rhythm becomes fixed, the branch should be versioned by its “set of patch titles” rather than by SHA when read.

Meanwhile, the state of the three fork lines continues to diverge: PyTorch replays daily, LLVM has been silent for eight days since 10-12, and Triton nightly builds resumed with one version on 10-19 and then stopped. The three lines correspond to three states of investment, and also suggest that the current bottleneck of the RuyiAI stack is not the compilation backend itself but the sustained supply of review and release capacity.

6.3 What a server-class platform brings to the distribution is a compliance checklist, not a new porting task

Coverage of the SpacemiT V100 lists openRuyi among the available distributions for the server platform. If this holds, the implication is not “openRuyi needs to do a new port” but a set of specification-compliance requirements: RVA23, the server SoC/platform specification, BRS, and the security model—five items—plus the UEFI/ACPI/SMBIOS firmware chain. What the distribution needs to do is mostly align its existing user-space packages, installer, and image pipeline with the specifications, rather than rewrite components.

The current evidence remains incomplete: the openRuyi main repo had zero server-related commits that day, the firmware line (ruyifirmware, edk2-ruyi) has been silent since 10-15, and the 2026.10 release has not yet shipped. There are three things to watch: whether the openRuyi 2026.10 release notes mention the server platform; whether ruyifirmware/edk2-ruyi resume commits targeting the V100; and whether a V100 entry appears in the support matrix.

Source Verification Table

Source Verification Result
RuyiAI-Stack organization (all repos) pytorch: riscv branch had 1 rewrite-style replay (0 new commits by commit date; compare shows 60 ahead/24 behind) + 7 open PRs with baseline refresh, main upstream sync as usual; llvm-project: riscv branch had no commits for eight days, two FPGA PRs stalled at 10-13; triton-riscv: 0 commits, 0 releases (nightly still the 10-18 build); tilelang-riscv: 0 commits; ruyiai-daily: 1 issue (10-20); harness, skills, caffe, mxnet, pytorch-ci, FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, torq-tile, dsh-osc had no commits in the window
ruyisdk organization (all repos) ruyi: 0 commits, 0 releases; llvm-project: AME branch 0 commits (stalled at PR #170 from 10-19); packages-index, ruyisdk-website, support-matrix, linux-xuantie-kernel all 0 commits; ruyi-imager: 1 nightly build update (10-20 06:22)
openRuyi-Project organization (all repos) openRuyi main repo: 137 commits (Rust/Go/Python 42 each, 11 other) + 10 PRs merged; CVE mega-PR #1341 closed without merge; open queue includes #1464 and #1466 (two de-vendoring PRs) and #1468 and #1469 (two CVE follow-ups); ruyifirmware, edk2-ruyi had no commits in the window (stalled at 10-15 17:53 and 10-15 15:22 respectively)
buddy-compiler organization buddy-mlir main branch: 0 commits, 0 releases; latest stable 0.1.1 (10-18), latest nightly 0.1.1.dev20261017 (10-17), no new builds since
RuyiSDK community forum Zero new topics in the window; latest is #2892 (10-15 daily news); biweekly report issue 078 (#2891) remains the latest issue
openRuyi distribution news page No new posts; the 2026.09 release announcement remains the latest; the 2026.10 page is not yet online
ruyi package manager / PyPI 0.53.0 remains the latest (released 09-24); no new versions or commits in the window
Official homepage (ruyiai.org) Three pages unchanged: homepage index 3 items, blog 2 posts, docs 23 chapters
Tech intelligence repo (ruyiai-daily) One issue pushed on 10-20; the RISC-V section covers the XuanTie E908A and C950A roadmaps, three types of SpacemiT V100 complete machines, the K1 bootph-pre-ram patch v2, three riscv-arch-test merges, and the XiangShan CoupledL2 banked memory; also includes IREE 3.13.0’s RVV MMA
IREE / riscv-arch-test upstream IREE v3.13.0 officially released (10-19 19:26), including RISC-V V f32 vfmacc.vf inner_tiled MMA (PR #24933); riscv-arch-test three merges (#2413, #2791, #2858)
arXiv (RISC-V) Latest in the past seven days is 10-15; no new RISC-V compiler or vector optimization preprints in the window
Google News RSS (multiple Chinese and English queries) RuyiSDK / RuyiAI / openRuyi all zero hits; one Ruyi-type query returned a XuanTie shipment article, unrelated to the topic; industry coverage concentrated on XuanTie, SpacemiT, and Tenstorrent

Complete Source List