Reporting period: October 12, 2026 (Monday) to October 16, 2026 (Friday), 5 working days in total

Sources: 5 issues of this publication’s daily RuyiAI updates (2026-10-12 to 2026-10-16, each covering the window from 04:30 the previous day to 04:30 of the current day); plus a 168-hour repository audit of three organizations — RuyiAI-Stack, ruyisdk, and buddy-compiler — a supplementary audit of key repositories in openRuyi-Project, and a dual-source cross-check of the RuyiSDK developer community and the openRuyi news page (coverage through the morning of 10-17)

Editorial note: This week’s main technical threads are openRuyi’s K3 firmware platform line (three releases in three days, 10-13 to 10-15) and the landing of the Ztt matrix extension toolchain on both GCC and LLVM backends. Beyond the daily-report material, the weekly adds two new audit findings: first, a single batch of 20 Ztt commits to the ztt-dev branch of riscv-gcc on the morning of 10-15 (not captured in that day’s daily report); second, the revival of tilelang-riscv and the release of Buddy MLIR 0.1.0 on 10-16 (both after the daily-report window).

I. Weekly Highlights

  • openRuyi launches the K3 firmware platform line, shipping three firmware releases in three days from 10-13 to 10-15: On 10-13, edk2-ruyi was created (a 35-commit batch first pushing the EDK2 platform code for SpacemiT K3 Pico ITX) along with ruyifirmware (5 commits establishing the packaging and release chain), and the first release “Ruyi Firmware 20261013” was published at 18:46 that day; on 10-14, the second release introduced CPU mode selection within UEFI (X100 only / X100+A100 mixed / A100 only, including RVV and H extension switches) with persistent storage; on 10-15, the third release added a read-only System Information page, x64 and ARM64 UEFI driver emulation (disabled by default, enabled via Platform Settings), and an experimental AMD ARM64 GOP 1.68 pilot. Sources: edk2-ruyi, Ruyi Firmware 20261013, 20261014, 20261015

  • Ztt matrix extension toolchain completes the LLVM side, forming a GCC + LLVM dual-backend (10-12): On the morning of 10-12, ruyisdk/llvm-project created the ztt-dev branch and completed a 26-commit full-stack first push of LLVM / Clang patches (author LukeZhuang, covering basic Ztt extension support, function-level profiling, CSR and matrix register groups, instruction assembly and disassembly, IR types and intrinsics, Clang matrix register types, pseudo-instruction expansion after register allocation, M register and ACC spill handling, with the final commit aligning tests with the AME diagnostic system); on the GCC side, the ztt-dev branch of riscv-gcc, established on 09-30, has since landed 36 commits on 10-08, 51 on 10-11, and 20 on 10-15 (the last batch identified by this week’s review). Combined with the Ztt intrinsic documentation V0.6 PR in attached-matrix-extension, the “specification—documentation—dual toolchain” chain is active in sync. Sources: llvm-project ztt-dev, head aeb52e1a, riscv-gcc ztt-dev

  • Buddy MLIR 0.1.0 officially released, with RISC-V cross-compilation and CI rolled out comprehensively (10-16): buddy-compiler/buddy-mlir saw 13 commits and 3 releases this week (16 commits and 4 releases including boundary-day items in the audit window), of which 0.1.0 was released in the early hours of 10-16 Beijing time (tag release/v0.1.0), after which the nightly line moved to the 0.1.1.dev series (dev20261016, dev20261017). RISC-V activity was intense: adding RISC-V board cross-compilation for buddy-cli (#987), fixing the Makefile’s RISC-V target (#986), simplifying the cross-compilation process (#975), adding RISC-V server cross-compilation for buddy-server (#989), and installing the riscv64 version of torch via the Ruyi package index in the manylinux riscv64 build while adding libsndfile (#990, #937, #988). Sources: buddy-mlir, BUDDY MLIR 0.1.0, PR #989

  • RuyiAI’s tilelang-riscv ends nearly a month of silence: 5 RISC-V commits connect to the Buddy release chain (10-16): The repository’s last commit was on 09-20; this week five commits landed in succession—code formatting, support for upstream operator runtime semantics (lane-local state, ragged copy, dynamic control flow, FP8/FP4 conversion fixes), upgraded TileOps runtime adaptation (submodule bumped to 67250e81, getting the paged-cache gather and varlen GQA sinks kernels working), and a new Buddy MLIR release CI plus building the Python extension in Buddy CI. This means the RISC-V backend of TileLang is beginning to run continuous integration against Buddy MLIR release artifacts. Sources: tilelang-riscv commit list, first batch of 5, 32fce5ed

  • Major single-week maintenance for the openRuyi distribution: ~90 Go module commits, KDE Plasma 6.7.5 across 39 packages, LLVM 24, mesa 26.2.4, and PowerVR backports (10-12 ~ 10-14): On 10-12, the main repository saw 40 commits, 39 of which upgraded the entire KDE Plasma desktop stack to 6.7.5 (plasma-workspace / kwin / systemsettings / powerdevil / breeze, etc.); on 10-13, 103 commits in a single day—about 90 were bulk updates of Go ecosystem modules (the k8s 0.37.1 series, gRPC 1.84, OpenTelemetry Collector 1.67, cloud vendor SDKs, etc.), plus llvm-snapshot upgraded to LLVM 24, audit 4.2.1 enabling RISC-V support, and a full batch update of the GnuPG stack; on 10-14, three commits turned to the graphics stack: mesa 26.2.4 and the first round of PowerVR optimization backports. On 10-16, SPECs for 6 more Rust RISC-V related crates including rust-riscv-0.16 were added. Sources: openRuyi commit list, PR #1410, PR #1347, PR #1399

  • ruyisdk/llvm-project merges XTHeadVector indexed load/store mixed-width fix (10-13, PR #168): Fixes th.vlxe / th.vxse handling when data and index element widths differ (preserving signed offsets, memory element size, mask passthrough, and zero-length semantics), extends intrinsic lowering to mixed widths of integer and floating-point vectors, and adds codegen coverage for RV32 / RV64 with masked, zero-length, negative-offset, and large-index cases. The last substantive commit on this rebase-17.0.6 branch was in June; this marks the XuanTie vector toolchain line shifting from long-term maintenance to targeted fixes. Sources: PR #168, issue #136

  • Dual-track kernel progress: openRuyi linux opens ruyi/7.3.y with a first batch of 31 commits; linux-xuantie-kernel unifies vendor keys as XUANTIE and adds hwprobe extensions (10-15 ~ 10-16): ruyi/7.3.y was created on 10-13 and landed its first 31 commits on 10-15 (K3 GPU nodes and display power domains, Type-C Port A / USB3 PHY / FUSB301, Ethernet MAC bound to EEPROM, the spacemit DP fix series, amdgpu disabling dynamic PCIe rate switching on K3, UFS FASTAUTO, SG2044 SRD3-12 board support, etc.); on 10-15, linux-xuantie-kernel renamed the T-Head vendor extension keys wholesale to the XUANTIE form (XTHEADVDOT → XUANTIE_VDOT, etc.), and on 10-16 added 4 more commits—incorporating Ziccamoa / Ziccif / Ziccrse / Za64rs / Zicclsm into cpufeature and hwprobe, carving out a private band in the ISA extension id space, and aligning extension ids with upstream. Sources: ruyi/7.3.y, linux-xuantie-kernel

  • Biweekly Report 078 goes live, themed “RuyiSDK to Appear at the RISC-V China Summit”; the summit runs 10-18 to 10-20 in Shenzhen (10-13): Issue 078 was finalized through an 8-commit editorial pipeline (#371 ~ #378) and synced to the forum. Highlights include the package manager 0.54.0 expected by month’s end (ABI compatibility detection, official macOS packages, SpacemiT K1/K3 toolchain adaptation in development), a round of IDE polish, porting of 6 vector math functions each for glibc and newlib, GCC’s ZTT backend improvements and GEMM/GEMV validation scheme (compile time reduced by about 5.05% ~ 6.97%), LLVM’s RISC-V P extension intrinsics, V8 conducting RVV Wasm SIMD tests on real hardware for the first time and merging the hwprobe fallback patch, OpenJDK’s JEP 483 / 515 AOT porting, and Go’s riscv64 atomic operation fixes. The summit agenda includes keynotes, poster sessions, and demos at booth D15. Sources: Issue 078 file, forum #2891

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

Window Overview: The RuyiAI-Stack organization (18–19 repos) saw activity this week concentrated in pytorch (one full-stack rebuild each of four mornings), llvm-project (paused after a 10-12 nightly replay), triton-riscv (a new nightly build), tilelang-riscv (five commits reviving it on 10-16), ruyiai-daily (five consecutive issues from 10-12 to 10-16), and the site repo (routine contributor-data commits); harness, mxnet, caffe, FBGEMM, IntrinTrans, VecIntrinBench, cpuinfo, torq-tile, dsh-osc, pytorch-ci, and SyncBots-DeepAgents had no pushes during the window.

PyTorch fork: one full-stack rebuild each of four mornings, 8 open PRs refreshed daily

  • Rebuild cadence: The riscv branch completed a full-stack rebuild on each of four mornings—10-12, 10-13, 10-14, and 10-15—with 23 local commits re-landed under new committer timestamps each round, and the head advancing across rounds as 6ff4eefe → a17584a3 → b870a71a → 73ef01df5d (same title “[SyncBots] Integrate PyTorch at e188cec”, SHA refreshed with each rebuild). The stack composition held to the existing sequence (#44 bytes_to_scalar fix, #50 mask-table expansion, #42 ULP tolerance, three SyncBots integration commits, #40 / #41 CI parsing, #20 / #21 CI split, #18 cpuinfo replacement, #12 commit-hash pinning, #11 / #4 / #1 mask-table series). The 10-15 round differs from the prior round’s head by 101 commits, including upstream commits from 10-13 to 10-14 merged in.
  • Review queue: The 8 open PRs (#23 Enable XNNPACK, #35 test PR, #36 timeout relaxation, #38 temporary DataLoader disable, #43 patch slimming, #46 ATen RVV support, #48 RVV CPU capability, #51 riscv64 local image) had their baselines batch-refreshed each morning at the same time as the rebuild, and remained open all week with no merges or closures.
  • Post-window follow-up: At check time (morning of 10-17), another rebuild batch was seen with committer timestamps in the early hours of 10-17, in which “Add riscv64 image locally (#51)” had landed on the branch with the stack—i.e., the CI local-image completion beyond RVV has entered the code line; recommend adding this to next week’s watch list. Sources: riscv branch, PR list

Compiler fork (llvm-project): no further activity for three consecutive nights after the 10-12 replay

  • The riscv branch completed a 26-commit patch-stack replay late on 10-12 (head 0fb2da86, still #31 “A100 HP/SP and pack/unpack lowering and scheduling calibration”), with the base synced to that day’s upstream snapshot; afterward, no new round appeared on the nights of 10-13, 10-14, or 10-15, and as of the 10-16 check the head remained at the 10-12 version.
  • The two FPGA backend PRs (#25 “BOSCAME FPGA backend”, #28 “xboscame FPGA”) remain open, with last updates stuck at 10-12; the buckyball branch (head from 10-06) saw no activity. Upstream sync on the main branch continued as usual (daily around 15:00 this week; generic commits are not itemized per the filtering rules).
  • Watch points: whether the overnight replay returns, and whether its division of labor with the ruyisdk/llvm-project maintenance line (the XTHeadVector fix has been merged into rebase-17.0.6) changes. Sources: riscv branch, PR #25, PR #28

triton-riscv: dependency-upgrade chain closed; goes quiet after first nightly artifact ships

  • On the morning of 10-12 the nightly channel produced “Nightly 2026-10-11-fe554ff”—based on the head following the Buddy MLIR dependency upgrade merged on 10-10 (PR #84, upgraded to nightly v0.0.11.dev20260925)—providing 6 wheels across CPython 3.12 / 3.13 / 3.14 and both x86_64 and riscv64 architectures. This is the first closed-loop artifact of the “patch adaptation (#86) → dependency upgrade (#84) → nightly artifact” chain.
  • Since then the main branch has had no new commits (head remains fe554ff), and the nightly channel has not produced another issue; the community optimization PR #85 “softmax and matmul optimization” has remained open since 10-08 with no updates during the window. Sources: Nightly 2026-10-11-fe554ff, PR #85

tilelang-riscv: five-commit revival on Friday, hooked into the Buddy release chain

  • Five consecutive RISC-V commits on 10-16 (spaced tens of minutes apart, constituting a single grouped push): first a standalone clang-format pass so subsequent changes are readable; then support for upstream operator runtime semantics (lane-local state across serialization stages, safe ragged copy, dynamic control flow, plus general IEEE and bitwise ops, FP8/FP4 conversion fixes, unsigned memref storage and configurable shared-memory reuse, and explicit rejection of unsupported cooperative while loops); then an upgrade of the TileOps submodule to 67250e81 with paged-cache gather and varlen GQA sinks kernels passing; and finally a new Buddy MLIR release CI plus “build Python extensions in Buddy CI”.
  • Significance: after nearly a month of silence since 09-20, this week’s grouped commits advance the TileLang RISC-V backend from “operator coverage” to “continuous integration based on Buddy MLIR versioned artifacts”, meshing with the release cadence of Buddy MLIR 0.1.0. Source: tilelang-riscv commit list

Technical intelligence and site

  • ruyiai-daily published 5 consecutive issues from 10-12 to 10-16 (pushed daily around 06:1x, with the self-described window running from the previous morning to the current morning), maintaining the five-chapter structure of PyTorch ecosystem / LLVM-MLIR / Triton and TileLang / RISC-V / AI industry; RISC-V items collected this week include the Buddy Compiler running the Qwen3-VL-30B-A3B text decoder on K3, the ggml-spacemit dual backend merged into llama.cpp, the XuanTie T-SBI privileged test protocol layer, RISE SpeexDSP RVV, XiangShan StoreQueue and L2 load-to-use optimizations, and a source snapshot of the SpacemiT Bianbu Agent Memory.
  • The site repo ruyiai-stack.github.io saw only routine contributor-data commits this week (one per day, no code changes); no new posts on the official site or blog. Source: ruyiai-daily

III. RuyiSDK Toolchain and Distributions (Firmware, Compilers, Kernel, openRuyi)

K3 Firmware Platform Line: Three Releases in Three Days, from End-to-End Connectivity to Configurable, Simulatable

  • 10-13 Repo Opening and First Release: edk2-ruyi was created at 11:07 the same day, with 35 commits pushed as a single batch (11:55 ~ 16:21, single author RevySR), covering the SpacemitPlatformPkg platform package and shared resources / clocks, power domains, DMA high-address non-coherency, board-level EEPROM and Ethernet MAC allocation, SEC-phase memory discovery and device tree publication, boot selection and hotkeys, boot logo and countdown, NOR-backed UEFI runtime variables, DDR training records, the full Fastboot suite (image loading, NOR/NVMe/UFS firmware updates, version query, verified updates, USB device-side transfer), UFS high-speed link, QSPI NOR, EQoS and RTL8127 SFP, DPU / Inno DP / GOP display chain, USB and PCIe controllers; the companion ruyifirmware was created at 12:01, with 5 notes establishing the coreboot + OpenSBI + EDK2 packaging chain (make package with two variants, PCI Option ROM / GOP, CI release pipeline), and the first firmware release was published at 18:46 the same day. kiwi-image added 3 commits the same day wiring up the GitHub Action and local build scripts for riscv64 images.
  • 10-14 Second Release: Selectable CPU Modes: UEFI setup added CPU mode selection with persistence, and fastfetch can now identify three tiers: k3 / k3-x100-a100 / k3-a100 — X100-only (default, 8 cores, RVV VLEN 256, H extension retained), X100 + A100 mixed (16 cores, both RVV and H disabled), A100-only (8 cores, VLEN 1024, H disabled). The same day, OpenSBI k3/port landed 13 commits (6 FROMLIST + 7 RUYI), coreboot k3/port 2 commits (SPI configuration selecting CPU tier, device tree fixup moved earlier), and edk2-ruyi 1 commit, forming a same-day loop of “base → platform → packaging → release”.
  • 10-15 Third Release: Information Surface and Simulation Layer: Added a read-only System Information page (CPU, memory, PCIe devices and link speeds, storage details; SMBIOS reports DRAM type and configured speed), early PCI graphics initialization and a fix for Setup graphics console re-entry, IMG GPU core frequency fixed at 1228.8 MHz; x64 and ARM64 UEFI driver emulation both included in DEBUG and RELEASE packages (disabled by default), CPU mode and emulation switches merged into Platform Settings and stored in a dedicated SPI sector (with CRC and read-back verification); experimental AMD ARM64 GOP 1.68 wired in for early GCN / Polaris / Vega M GPUs (no fallback on replacement failure). Fixed interrupt mapping in A100-only mode; the first-install documentation switched to raw SPI flashing. That day: edk2-ruyi 6 commits, coreboot 2 commits, ruyifirmware build 8 commits, MultiArchUefiPkg 2 commits.
  • Watch points: performance and compatibility trade-offs of the three CPU modes under real workloads, behavior of settings persistence across firmware upgrade paths, and the pace of expansion of the simulation layer and device list. Sources: edk2-ruyi, opensbi k3/port, coreboot k3/port, ruyifirmware

Compilers and Toolchain: Ztt Dual Backends, XTHeadVector Fix, AME Tutorial Kickoff

  • GCC side (riscv-gcc ztt-dev): After a single batch of 51 commits on 10-11, another 20 landed on the morning of 10-15 (new findings from review) — topics extended from descriptor-system refinement to coordination with other extensions: preserving Ztt state across RVV operations and RVV tail segments, preserving Ztt state across scalar floating-point tail segments / CPU atomics / Zfa silent comparisons, removing interleaved Ztt reloads and dead spill saves between RVV operations, skipping irrelevant register groups when Ztt Md is invalid, Ztt integer zero-absorption and AND-NOT combination folding, direct evaluation of constant leaves, on-demand validation of public builtins (LTO), pagination of the public declaration table, etc. The batch header was pushed at 10-15 11:07.
  • LLVM side (ruyisdk/llvm-project): ztt-dev was created on 10-12 with an initial 26 commits (see this week’s headlines); on 10-13, a mixed-width fix for XTHeadVector’s th.vlxe / th.vxse was merged into rebase-17.0.6 (#168).
  • Matrix extension support: AME-operator created the c-tutorial branch on 10-15 and initialized the “AME GCC notebook tutorial”; the Ztt intrinsic documentation V0.6 PR #1 in attached-matrix-extension remained open this week (last updated 10-12).
  • Review and release: riscv-binutils ztt-dev is still stalled at 09-30; riscv-gnu-toolchain’s ame-experimental had no commits in the window. Sources: riscv-gcc ztt-dev, PR #168, AME-operator

openRuyi Main Repository: Bulk Dependency Updates and Security Response in Parallel

  • 10-12: 40 commits, of which 39 were a bulk upgrade of the KDE Plasma 6.7.5 desktop stack (plasma-workspace / plasma-desktop / kwin / systemsettings / powerdevil / kscreen / kscreenlocker / libplasma / kdecoration / breeze / discover / plasma-nm / plasma-pa, etc.); the other was the addition of AGENTS.md (repository collaboration notes for AI coding agents).
  • 10-13: 103 commits in a single day — about 90 were bulk updates of Go ecosystem modules (k8s cri-api / streaming / component-base etc. to 0.37.1, gRPC 1.84, OpenTelemetry Collector 1.67 / 0.161, google-cloud and google-api series, Hashicorp / AWS / Azure SDKs), with the rest including llvm-snapshot to LLVM 24, audit 4.2.1 enabling RISC-V support, and a bulk update of the GnuPG stack (gnupg 2.5.24, libgcrypt 1.12.4, libksba 1.8.1, libgpg-error 1.61).
  • 10-14: 3 commits shifting to the graphics stack — mesa 26.2.4, a new mesa-26.2.4-pvr-backport.patch starting the first round of PowerVR optimization backports, and adjustment of mesa-gallium’s dependency recommendation on mesa-vulkan-drivers.
  • 10-15: 8 commits — five security fixes in a row (cups backports fixing CVE-2026-87876 and two GHSAs, python-nltk fixing CVE-2026-81726, capstone to 5.0.9 with two CVE backports, llama-cpp to 0.4.1 fixing CVE-2026-21869 and CVE-2025-53630, multipath-tools to 0.15.1), plus three maintenance commits (spdk installing bdevperf, systemd providing a commented .conf template, rdma-core avoiding a hard dependency on udev).
  • 10-16: 6 commits — new SPECs for rust-riscv-0.16, rust-riscv-macros-0.4, rust-riscv-types-0.1, rust-embedded-hal-1, rust-volatile-register-0.2, rust-vcell-0.1; Rust embedded and RISC-V related crates are beginning to enter the distribution in groups. Sources: openRuyi commit list, PR #1347, PR #1399, PR #1422

Kernel Line: ruyi/7.3.y Kickoff and This Cycle’s hwprobe Extensions

  • ruyi/7.3.y: Created on 10-13, with the first 31 commits landing on 10-15, covering K3 GPU nodes and display power domains, Type-C Port A / USB3 PHY / FUSB301 autonomous controller, Ethernet MAC bound to Pico ITX EEPROM, the spacemit DP display fix series (hot-plug interrupt recovery, connector workqueue quiescing, MOT, AUX replies, stopping register access after DPU clock failure), disabling dynamic PCIe rate switching for amdgpu on K3, UFS FASTAUTO, shmem GEM writecombine, Lichee Pi 4A and Milk-V Meles USB, SG2044 SRD3-12 board support; topics overlapping with 7.2.y are migration replays onto the new branch.
  • linux-xuantie-kernel: On 10-15, three commits renamed the T-Head vendor extension detection keys to the XUANTIE form throughout and migrated XuanTie Ssdtso into the XuanTie vendor extension framework; on 10-16, four more landed — Ziccamoa / Ziccif / Ziccrse / Za64rs and Zicclsm brought into cpufeature and hwprobe, a private band carved out in the ISA extension id space, and extension ids aligned with upstream. The consolidation of naming and id space facilitates consistent references on the hwprobe and toolchain sides. Sources: ruyi/7.3.y, linux-xuantie-kernel

ruyi Package Manager and Peripheral Tools

  • ruyi: The main repository was quiet all week (last commit 09-25); PyPI and Releases both remain at 0.53.0 (uploaded 09-24), and the latest pre-release is still 0.54.0-alpha.20260918; biweekly report 078 previews 0.54.0 for release at the end of the month.
  • IDE: VS Code extension #263 (low-version ruyi prompt) remains open after being filed on 10-10; #264 (package install errors moved into inner details) and #265 (error reporting added to package index refresh) were merged on the afternoon of 10-12, with 0.1.7 still the latest on the subscription side.
  • Build and tuning: ruyici PR #13 was merged, adding native macOS (Apple Silicon) toolchain builds (10-12); ruyi-tuner connected aarch64 to the CSiBE benchmark pipeline on 10-12, and on 10-14 completed “writing per-project actual compilation comparison results to txt” and merged v1.x into main.
  • Images and software sources: packages-index fix merged (#226 SiFive Unmatched Ubuntu server image URL, thanks to weilinfox) and #228 newly opened (telemetry uploads no longer fail the workflow); ruyi-imager nightly channel refreshed routinely, with one commit on 10-15 aligning UI colors and button hierarchy with the RuyiSDK homepage.
  • Support matrix: Two reports merged into main on 10-16 — RT-Thread Smart (Milk-V Duo S) and Zephyr (Milk-V Duo S). Sources: ruyi Releases, VS Code PR #264, ruyici PR #13, ruyi-tuner, support-matrix

IV. Vector Extensions and Operator Optimization (VecIntrinBench / IntrinTrans / FBGEMM)

Target repositories: silent all week; substantive progress on adjacent workstreams

  • VecIntrinBench (last commit 02-13), IntrinTrans (06-25), and the FBGEMM fork (07-13) had no commits all week; cpuinfo (05-20) and torq-tile (04-09) from the same group were likewise silent. The open PR counts for the two classic frameworks were unchanged: MXNet 15, Caffe 3, with no updates during the window.
  • Actual work on the vector and matrix front was concentrated in: the Ztt line on both the GCC / LLVM sides (see Section III), the TileOps runtime adaptation and FP8 / FP4 conversion fixes in tilelang-riscv (see Section II), and the tutorial and validation line of AME-operator.

External developments: RISE’s SpeexDSP RVV optimization and the XuanTie C950 matrix extension narrative

  • RISE SpeexDSP RVV optimization (10-14): Introduced a Meson build, Checkasm tests, and hand-written RVV kernels for SpeexDSP, validated on K1 (VLEN 256) and K230 (VLEN 128) — K1 floating-point resampling reached roughly 3.71x / 3.59x the Clang 18 / GCC 13 baselines respectively. This work pushes RVV optimization into foundational multimedia libraries such as VoIP and audio resampling, and is directly relevant to the SpacemiT platform. Source: RISE blog post
  • DAMO XuanTie C950 interview (10-15): The RISC-V International website published an interview with XuanTie chief scientist Jianyi Meng — the team ran Qwen3.8-27B locally on the C950’s 64-core target configuration, achieving throughput above 30 tokens/s without relying on a discrete GPU or NPU; reaching that throughput requires roughly 400 GB/s-class memory movement, depending on high-bandwidth memory and the team’s self-developed matrix extension validated in silicon; the team subsequently joined with Qualcomm, IBM, VRULL, ISCAS, BOSC, and others to launch the AME (Attached Matrix Extension) task group to drive standardization. Key points: prefill is dominated by dense matrix computation, while decode shifts to a bandwidth bottleneck; separate matrix registers are used (INT4 / INT8 to INT32, FP4 / FP8 / FP16 / BF16 to FP32 mixed-precision accumulation) to avoid competing with RVV registers; the specification still needs to cover data layout, accumulators, rounding and saturation, context save/restore, exceptions and virtualization, and compiler and ABI mapping. Source: RISC-V International interview
  • Documentation side: The Ztt intrinsic documentation V0.6 (PR #1) for attached-matrix-extension remains open, echoing the code progress on the two ztt-dev lines for GCC / LLVM.

V. Buddy Compiler Updates (buddy-mlir Periodic Summary)

Window Overview: In the buddy-compiler organization this week, buddy-mlir was the only active repository—13 commits and 3 releases this week (16 commits and 4 releases including 10-11 and 10-17 boundary items in the audit window), with themes centered on “RISC-V cross-compilation and CI enablement,” “model runtime (RAX / buddy-server / buddy-cli) polish,” and “release pipeline (0.1.0 and the nightly line).”

Releases: 0.1.0 Official Release and Nightly Line Switchover

  • Release sequence from 10-15 to 10-16: Nightly 0.0.12.dev20261015 (Beijing time 10-15 21:53) → BUDDY MLIR 0.1.0 (10-16 00:53, tag release/v0.1.0) → Nightly 0.1.1.dev20261016 (10-16 07:11) → Nightly 0.1.1.dev20261017 (10-17 07:24, outside the window). The 0.1.0 tag points to 2860a0e (manylinux release build fix), meaning this version centers on “cross-architecture release artifacts being usable.”
  • Interlock with the RuyiAI side: triton-riscv’s dependency upgrade (PR #84) had already switched to the v0.0.11 line, and this week tilelang-riscv added a Buddy MLIR release CI—Buddy’s release cadence has become the upstream anchor for both of RuyiAI’s operator frontend repositories.

RISC-V Cross-Compilation and CI: From “Can Build” to “Routinely Verified” Within a Week

  • buddy-cli: #975 simplified the RISC-V cross-compilation flow, #986 fixed the Makefile’s RISC-V target, #987 cross-compiled buddy-cli for RISC-V boards; #977 made buddy-cli load the RAX runtime dependencies before loading the model.
  • buddy-server: #989 added RISC-V cross-compilation for the server (extracted RaxDependenciesHandle as a shared runtime library, added a riscv-server Make target, and added unit and integration coverage for ordered RTLD_GLOBAL dependency loading); #982 unified buddy-server’s plugin discovery mechanism.
  • CI and releases: #937 added cross-compilation checks, #990 fixed the manylinux release build (switched to the Ruyi package index to install torch on riscv64, added libsndfile to the riscv64 manylinux image), #988 made the Ruyi package index effective only on riscv64, #925 added the publish-models workflow. These commits form the infrastructure for “riscv64 wheels being releasable and regression-testable.”
  • Runtime: #985 added incremental UTF-8 decoding for text streaming output (byte-level BPE splits multi-byte characters across multiple decode steps; previously streaming them directly caused an LLVM JSON serialization assertion and terminated buddy-server with SIGABRT); #980 added a profiling mode to the model thread pool.

Models and Kernels: Qwen3-VL on K3 and w4g32 for Zhixin A210

  • 10-12: Ran the Qwen3-VL-30B-A3B text decoder on SpacemiT K3 (int4 MoE expert kernels, 13.05 tok/s for short-context decoding)—this result was covered in the 10-13 issue of the tech intelligence repository and is the most concrete measurement this week of “running a large MoE on a RISC-V board.”
  • 10-11: Added w4g32 kernels for Zhixin A210 (#979), and changed the w4g32 weight reuse key to be based on weight layout (#978), pointing toward reuse of low-bit quantization kernels across different backends.
  • 10-13: Fixed the load order of the RAX runtime shared library; the streaming UTF-8 fix on 10-14 also falls under “model inference pipeline robustness.”
  • Watch points: API stability after 0.1.0, feedback on riscv64 wheel usability on real boards, and the deployment form of buddy-cli / buddy-server on K1 / K3. Sources: buddy-mlir commit list, 0.1.0, PR #985, PR #987

VI. Community and Events (Biweekly Progress, Forum, Support Matrix)

Biweekly Report 078 and Summit Cadence

  • Issue 078 (10-13): The editorial pipeline’s 8 commits (#371 ~ #378) were completed during the day on 10-13; the final commit “add summary, remove sections with no progress” added the summary and removed sections without progress. The theme was “RuyiSDK to appear at the RISC-V China Summit” (10-18 to 10-20, Shenzhen Convention & Exhibition Center Futian, keynote + poster presentation + D15 booth demo). Content highlights are in this week’s top news.
  • External teasers: Media outlets such as 21 Caijing published summit previews on 10-13 (expected 3,000+ offline attendees, 150+ exhibitors, 800,000 online reach, seven parallel sub-forums); the RuyiSDK-side communication nodes were the forum participation announcement (#2889) and the biweekly report launch post (#2891).

Forum: News Posts and Hands-On Posts Interleaved, No New Additions After 10-15

  • 6 new topics this week: #2887 “WordCraft for riscv” (a pure-Rust open-source Word alternative failed on LPi4A and succeeded on VisionFive 2, with screenshots), #2888, #2890, and #2892 as three daily “open-source tech hot news” posts, #2889 the summit participation announcement, and #2891 the biweekly report 078 launch post. No new topics after 10-15; the community entered a quiet preparation period before the summit.
  • The previous active topic remains the 10-08 hands-on notes on installing and upgrading ruyi via script (#2886). Sources: ruyisdk.cn, #2887, #2892

Support Matrix and Source Status

  • support-matrix merged two reports on 10-16 (RT-Thread Smart and Zephyr on Milk-V Duo S); the previously pending queue (Zephyr Milk-V Duo S, ESP32-C6 RT-Thread, etc.) converged over the course of this week.
  • External sources: Google News returned zero relevant hits all week across multiple Chinese and English queries for RuyiSDK / RuyiAI / openRuyi / “如意社区” (generic RISC-V queries returned normally during the same period, so the channel is not anomalous); arXiv had no new entries in RISC-V compilation, vector, or edge AI directions; the openRuyi distro news page still shows the 2026.09 release announcement as the latest (the 2026.10 release page is not yet online); the RuyiAI website and blog had no new posts.

VII. Trend Observations

  • K3 became the week’s only systematic main thread: firmware (three versions in three days), kernel (the first 31 commits of ruyi/7.3.y), and images (kiwi-image automation) all moved within the same week, and the firmware has progressed from “link bring-up” to “selectable CPU modes + simulatable drivers.” The watch point is whether these capabilities can form a demonstrable complete-device form during the 10-18 summit week, and the measured data for the three CPU modes.
  • Matrix expansion shifted from “single-backend advancement” to parallel “dual-backend + specification + operator validation”: the three GCC ztt-dev batches (36 → 51 → 20 commits) have moved from feature rollout to correctness in coordination with other extensions (state preservation across RVV / Zfa / atomics); the LLVM-side ztt-dev first push fills in the other half, with documentation V0.6 and the AME-operator tutorial following up; meanwhile the upstream AME task group is advancing standardization under XuanTie’s leadership—the in-repo draft and the specification line are starting to stay in sync.
  • The cadence divergence between the two fork maintenance lines is widening: the PyTorch riscv branch maintains a 23-commit rebuild every morning with 8 PRs refreshed in sync; the LLVM riscv branch has not replayed for three consecutive nights since 10-12. The former indicates a high degree of automation in the upstream-following mechanism; the latter may mean the LLVM-side focus has shifted to the ruyisdk/llvm-project maintenance line (the XTHeadVector fix is a result of that line).
  • Buddy Compiler has become one of RuyiAI’s de facto upstreams: buddy-mlir completed the 0.1.0 release and riscv64 cross-compilation / CI / manylinux release-chain fixes within a week, while triton-riscv’s dependency upgrade and tilelang-riscv’s new CI both anchor on Buddy’s versioned artifacts—the chain “MLIR compilation framework → operator frontend → nightly artifacts” closed the loop on riscv64 to a releasable wheel for the first time this week.
  • Distro maintenance has entered a dual cadence of “batch processing + security batches”: openRuyi’s single-day 90 Go modules, 39 KDE packages, the full GnuPG stack, and five security packages in quick succession show that dependency following has been engineered; the implicit constraint is the coupling between full-rebuild cost and image release cadence, and the timing of 0.54.0 and the 2026.10 images is worth watching.
  • Watch list: 0.54.0 release at month-end and ABI compatibility checks; when ruyi/7.3.y enters the image channel; triton-riscv #85 review; the two LLVM FPGA PRs (#25 / #28); PyTorch #46 / #48 (RVV) and #51 (local image already merged into the branch); attached-matrix-extension documentation PR #1; whether any versions or results are released during the summit (10-18 ~ 10-20).

Appendix: Sources and Verification Notes

Sources (5 issues)

  • 2026-10-12 issue: window 10-11 04:30 ~ 10-12 04:30.
  • 2026-10-13 issue: window 10-12 04:30 ~ 10-13 04:30.
  • 2026-10-14 issue: window 10-13 04:30 ~ 10-14 04:30.
  • 2026-10-15 issue: window 10-14 04:30 ~ 10-15 04:30.
  • 2026-10-16 issue: window 10-15 04:30 ~ 10-16 04:30.
  • Developments outside the period boundary (daytime 10-16 through the morning of 10-17) were added after organization repository verification, including the 10-15 batch of riscv-gcc ztt-dev (not covered in the daily reports), the 10-16 tilelang-riscv and buddy-mlir commits, the rust-riscv crate batch in openRuyi, four hwprobe commits in linux-xuantie-kernel, and the merge of two support-matrix reports.

Verification Method and Scope

  • Organization verification (168-hour window): covers commits, PRs, and releases across three organizations — RuyiAI-Stack, ruyisdk, and buddy-compiler (via both Atom feeds and web pages, not reliant on API rate limits); commits in upstream forks (PyTorch riscv, LLVM riscv, LLVM main, PyTorch main) were included after filtering by RISC-V / RVV / vector extension / XuanTie-related keywords.
  • Supplementary verification of key repositories: openRuyi, edk2-ruyi, ruyifirmware, opensbi, coreboot, linux (ruyi/7.3.y), and MultiArchUefiPkg in openRuyi-Project; riscv-gcc (ztt-dev), llvm-project (ztt-dev / rebase-17.0.6), linux-xuantie-kernel, support-matrix, and wechat-articles in ruyisdk; and buddy-mlir in buddy-compiler.
  • Dual community sources: the RuyiSDK developer community forum (latest topics list) and the openRuyi distribution news page.
  • Time basis: repository commits use committer time converted to Beijing time; “daily” descriptions follow consecutive calendar days; the period covers five working days from 2026-10-12 to 2026-10-16.

Main References

Data sources: daily RuyiAI dynamics reports (5 issues, 2026-10-12 to 2026-10-16) and this week’s supplementary verification (organization repositories, dual community sources, periodic activity in key repositories); the reporting period is 2026-10-12 to 2026-10-16.