Research window: 2026-11-13 04:30 ~ 2026-11-16 04:30 (Beijing time, the past 3 days; the Monday window covers the weekend and connects end-to-end with the cutoff of the previous report on 11-13).

Sources: GitHub push verification for key repositories and organizations (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler; repository-by-repository checks of commits, PRs, and Releases within the window, with commit counts re-verified via the API), the full text of the openRuyi packaging repository SPECS and the python-torch-ruyiai spec, kernel-team-tools configuration and tags, commit details for ruyifirmware and ruyi-imager, the RuyiSDK developer community forum, snapshots of three pages on the official RuyiAI homepage, the openRuyi distribution news page, targeted Chinese and English Google News RSS queries (via proxy), and arXiv searches.


This Issue’s Index

  • Today’s Highlights: BUDDY MLIR 0.1.4 officially released, with nineteen of thirty commits landing in the VIR middle end (11-15)
  • Today’s Highlights: openRuyi picks up python-torch-ruyiai, bringing RuyiAI’s PyTorch snapshot into the distribution (11-13)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V porting)
    • 1.1 pytorch riscv branch: CI criteria changed twice in one day, return codes replaced with statistics files (11-13)
    • 1.2 Branch rebased and force-pushed on 11-14: twenty-six commits given a unified timestamp, only two new ones added (11-14)
    • 1.3 triton-riscv: #116 and #117 merged in succession, three nightly builds in three days (11-14)
    • 1.4 tilelang-riscv, llvm-project and other key repos quiet (11-16 status)
  • II. RuyiSDK Toolchain and Distribution
    • 2.1 openRuyi packaging repo sees forty-five commits: besides torch, Go 1.27.4, thirty-odd Go packages and KDE completion (11-13)
    • 2.2 openRuyi kernel refreshes patchset again, kernel-team-tools pushes tags to 7.3.4-6 and 6.18.63-7 over the weekend (11-15)
    • 2.3 llvm22/23/24 enable PIE and generate clang resource dir, clang-wrap and cmake follow suit (11-13)
    • 2.4 ruyifirmware eight commits: p550 uses OpenOCD 0.12 for memory recovery, board kernels unified to 7.3.4-3 (11-13)
    • 2.5 ruyi-imager: pause download and full-image progress, Gradle wrapper bumped to 9.9.1 (11-15)
    • 2.6 ruyi and packages-index see zero commits, 0.54.0 remains the latest stable release (11-16 status)
  • III. Vector Extensions and Operator Optimization
    • 3.1 ABI mismatch in bf16 conversion helper: llc reads return value from floating-point register, launcher returns via integer (11-14)
    • 3.2 Five batches of fifty-four FlagGems operator examples, five hundred eighty-seven pass, forty-one bf16 strictly pending (11-13)
    • 3.3 After explicit target switches to -O3, flash attention object gains three vector divisions (11-14)
    • 3.4 wheel slimmed by forty percent with strip-all, triton-shared-opt drops from four hundred seven MB to one hundred eighteen MB (11-13)
    • 3.5 VecIntrinBench, IntrinTrans and FBGEMM still quiet (11-16 status)
  • IV. Buddy Compiler Updates
    • 4.1 0.1.4 shares the same tree as the 11-13 nightly, no new commits on main in the two days before release (11-15)
    • 4.2 Open PR #1059: replacing memref copies with vector moves, sample drops from sixty-six ms to two point eight ms (11-13 status)
    • 4.3 Downstream Renovate still pinned to dev20261113, has not yet followed to the stable release (11-16 status)
  • V. Community, Events and Official Homepage
    • 5.1 Forum has one routine hot thread on 11-13, no new posts over the weekend, biweekly progress still at issue 080 (11-13)
    • 5.2 Official homepage three pages unchanged, two blog posts, twenty-three documentation chapters (11-16 status)
    • 5.3 ruyiai-daily gap reaches its eighth day (11-16 status)
    • 5.4 External news and arXiv: no RuyiAI-related entries (11-16 status)
  • VI. Trend Observations
    • 6.1 The packaging layer has become the new battleground for RuyiAI adoption
    • 6.2 The stable release is just that nightly tree
    • 6.3 The performance ledger has turned to the copy page
    • 6.4 CI criteria are still being trialed

Today’s Focus: BUDDY MLIR 0.1.4 Officially Released, Nineteen of Thirty Commits Land in the VIR Middle End (11-15)

Date: 2026-11-15 Source: BUDDY MLIR 0.1.4, 0.1.3 to 0.1.4 diff, PR #1059

At 08:01 on 2026-11-15 (Beijing time), buddy-mlir shipped the official 0.1.4 release, two weeks after 0.1.3 (10-31). From 0.1.3 to 0.1.4 there are thirty commits in total, nineteen of which carry the middle-end prefix, and they share a single theme: taking the scalable-vector (VIR) lowering path from “can run a handful of operators” to “covers a class of shapes.”

These nineteen commits fall into four groups. The first group is shape coverage: #1017 fixes narrowing in scalable-vector tiling counts; #1056 completes all reduction kinds for integers and floats; #1057 lowers rank-0 memref linalg.fill to a single store; #1058 supports integer matrix multiplication and fixes dynamic dimensions; #1041 routes convolution through VIR as well. The second group is VLEN handling: #1028 lowers set_vl for an exact rvv-vlen; #1029 can also generate code for any VLEN when there is no rvv-vlen; #1032 and #1037 respectively prevent “a model compiled for one exact VLEN” from being moved to a different VLEN, and lower according to the host’s V extension width; #1046 validates dynamic vector types and prints the scaling factor as m2; #1048 preserves the scaling factor when cloning dynamic vector types. The third group is tiling and composition: #1031 lowers scalable-vector shuffle in several slices; #1047 merges vector slices for associative reductions; #1045 gives vir.set_vl a single-block region; #1050 documents which operators strip-mined set_vl supports; #1051 handles full-lane reductions of vector<[1]xT>. The fourth group is the seven commits from 11-12 already covered in the previous report (the masking, libm, bf16, fast-math, reduction, rank-0, and integer matmul items among #1052 to #1058), which all enter the official release this time.

Beyond VIR there are eleven side branches: on the model side, #1042 adds YOLOv8 detection and pose on SpacemiT K3, and #1044 changes the A100 llc parameters to be taken from a100_llc_args; on the runtime side, #1018 adds a shared serving generation controller; on the compiler side, #1025 bumps LLVM to bef335e2143c, and #1027 ships LLVM with NVPTX, AMDGPU, and lld; on the testing side, #1026 adds PyTorch 2.15 to the matrix; on the backend side, #1033 adds SPM optimization for BOSCAME; on the tooling side, #1030 lets the kernel lower VIR to the target corresponding to the llc options.

Worth noting is the release timing: the last commit on main is still #1058 at 14:44 on 11-12, and there were no new commits between then and the 11-15 release. In other words, 0.1.4 and the 11-13 nightly are the same tree; the official release effectively gives that tree two extra days of observation rather than stacking another layer of changes.

Today’s Highlight: openRuyi Adopts python-torch-ruyiai, RuyiAI’s PyTorch Snapshot Enters the Distribution (11-13)

Date: 2026-11-13 Source: openRuyi commit add python-torch-ruyiai, python-torch-ruyiai.spec, PyTorch riscv branch

At 22:58 on 11-13 (Beijing time), the openRuyi packaging repo landed a new package, python-torch-ruyiai. This is the first time RuyiAI appears in openRuyi as a distribution package: version 2.15.0~a0+git20260921, source tracking the RuyiAI-Stack riscv branch (commit 6d667e1e), CPU-only, built with the clang toolchain, and it cannot coexist with the distribution’s existing python-torch.

Three packaging difficulties can be read from the spec file. The first is BLAS: OpenBLAS’s Fortran cdotu_/cdotc_ return complex values in registers, while PyTorch’s non-CBLAS path takes the result from the first argument per f2c convention, so dot and vdot on complex tensors silently return 0; patch 2001 enables the CBLAS path and makes PYTORCH_BLAS_USE_CBLAS_DOT=ON in the spec take effect under native builds as well (upstream only reads this variable during cross-compilation). The second is RVV gating: the __riscv_v_intrinsic macro is also defined under the rva20 baseline, so the old condition selected deep-convolution vector kernels that should never be used; patch 2011 switches to checking __riscv_vector. The third is the build backend: the RuyiAI snapshot requires scikit-build-core 1.0 or later, while openRuyi still ships 0.12.x and python-pyvex pins an upper bound, so patch 2012 simply bundles this backend as source, while also removing the license file glob that cannot be satisfied after trimming the third-party tree.

Dependency handling follows the “use system libraries where possible” approach: a batch of switches such as USE_SYSTEM_CPUINFO, USE_SYSTEM_ONNX, USE_SYSTEM_SLEEF, USE_SYSTEM_PYBIND11, and USE_SYSTEM_PTHREADPOOL are enabled, USE_CUDA, USE_ROCM, USE_MKLDNN, and USE_XNNPACK are disabled, and distributed and TensorPipe are retained. The third-party directory keeps only the components actually compiled (gloo, flatbuffers, cpp-httplib, kineto, tensorpipe, miniz), and tensorpipe’s bundled libuv and libnop are also unpacked. The check phase does not run the fork’s full test suite (most are blocklisted on riscv64, and the remaining failures are recorded in fork issues 33 and 34); instead it runs a functional smoke test: matrix multiplication, autograd, two hundred training steps, plus a guard that complex dot/vdot must not be zero—exactly covering the path fixed by patch 2001.

The same night, the packaging repo also landed llvm22/23/24 with PIE enabled and clang resource dir generation, clang-wrap bumped to 20261113, and cmake bumped to 4.4.5; judging by timing, these fill in the installation layout for the in-distribution clang build.


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

1.1 pytorch riscv branch: CI criterion revised twice in one day, return code replaced by stats file (11-13)

Date: 2026-11-13 Source: PR #56, PR #57

At 00:25 on 11-13, PR #56 was merged, adding a criterion to the riscv64 CI yaml: when a new job builds successfully but executes no test cases (0 failures, 0 passes), the parent job previously recorded a pass, so a check was added to fail on any non-zero Jenkins return code. Seventeen hours later (17:04), PR #57 was merged, removing that return-code check entirely and instead fetching pytorch-ci-stat.json via curl --fail first, then judging by the statistics values. The two commits total +35 −20 lines, both touching only .github/workflows/ci-riscv64.yml.

Revising the same criterion twice in one day shows the return-code approach is inherently unstable: Jenkins exit codes can still be 0 when the build succeeds but all test cases fail, so the return code alone cannot block false greens. After switching to reading the stats file, the criterion shifts from “did the job die” to “did the test cases run”—the same direction of tightening as the “zero test cases passed” criterion in the previous report.

1.2 Branch reordered and re-pushed on 11-14: twenty-six commits share a unified timestamp, only two are new (11-14)

Date: 2026-11-14 Source: pytorch riscv branch commit list

At 19:00 on 11-14, the last push to the riscv branch carried twenty-six RISC-V-related commits, all with committer timestamps falling in the same second. Checking commit by commit, only #56 and #57 are genuinely new in this window; #41 (use GitHub Actions to judge CI pass/fail), #42 (restrict linear cross-entropy ULP tolerance to RISC-V), #44 (fix bytes_to_scalar for float/complex), #50 (add inductor opinfo tests to the blocklist), and #51 (local riscv64 image) are all old commits merged between July and October, reordered this time with timestamps uniformly refreshed. The same day also saw three upstream syncs (SyncBots aligning to upstream 3afdbd3, 2760264, and 1154a55 respectively).

This pipeline now works as “RuyiAI branch + scheduled upstream sync,” so monthly commit counts in reports get inflated by branch housekeeping; judging new work must go back to PR merge times, not commit times.

1.3 triton-riscv: #116 and #117 merged in succession, three nightlies in three days (11-14)

Date: 2026-11-14 Source: PR #116, PR #117, nightly 2026-11-14

The two PRs still under review in the previous report were both merged in this window: #116 (strip debug sections from packaging) merged 11-13 20:44, and #117 (explicit targets use -O3, add +zvfbfmin to the default feature string) merged 11-14 15:19. Three nightlies followed in succession: 11-12-02759d7, 11-13-72176d5, and 11-14-e99c326, all published between midnight and 7 a.m. Beijing time.

A noteworthy boundary condition is documented in #117: the native RISC-V path (without a target triple) still skips opt and keeps llc at -O0, with code comments recording miscompilation at higher levels; the IME path uses buddy-llc because standard opt does not recognize IME intrinsics; only the cross-compilation and explicit-target paths switch to -O3.

1.4 tilelang-riscv, llvm-project, and other key repos quiet (11-16 status)

Date: 2026-11-16 Source: tilelang-riscv, llvm-project (RuyiAI fork)

tilelang-riscv’s latest commit is still 11-03, and the llvm-project fork is stopped at 11-02; pytorch-ci, FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, harness, torq-tile, and dsh-osc all had zero commits in the window. All activity in the RuyiAI-Stack organization this window was concentrated in three repos: pytorch, triton-riscv, and ruyiai-daily.

II. RuyiSDK Toolchain and Distributions

2.1 openRuyi packaging repo: 45 commits — beyond torch, Go 1.27.4, 30+ Go packages, and KDE additions (11-13)

Date: 2026-11-13 Source: openRuyi commit list

Apart from the torch package covered in today’s highlights, the openRuyi packaging repo saw 45 commits in the window. The largest batch was 30+ Go language packages added all at once at 17:04 on 11-13: Go itself was bumped to 1.27.4, followed by the pion family (webrtc, dtls, srtp, sctp, turn, stun, ice, etc.), the modernc family (libc, memory, mathutil, fileutil, sqlite), mautrix and mau-util (Matrix protocol bridges), and SDKs such as larksuite, line-bot, and github-copilot. Together these packages point to one purpose: running a complete Matrix and bot ecosystem on openRuyi.

The same day brought three desktop and toolchain changes: on the KDE side, bluedevil and kf6-bluez-qt were added (the Bluetooth panel and its Qt bindings); clang-wrap was bumped to 20261113 and cmake to 4.4.5; on the performance-tools side, lmbench received a patch fixing getsummary.

2.2 openRuyi kernel patchsets refreshed again; kernel-team-tools pushes tags to 7.3.4-6 and 6.18.63-7 over the weekend (11-15)

Date: 2026-11-15 Source: linux SPECS refresh, linux-lts SPECS refresh, kernel-team-tools commit list

At 15:13 on 11-14, the linux and linux-lts specs in openRuyi were each refreshed again: linux’s patchset_release went from 2 to 3 (config_version reset from 1 to 0, effectively opening a new configuration baseline for the new patchset), and linux-lts went from 3 to 4, with upstream tarball checksums updated on both sides. This is the second refresh in four days; the previous report covered the 11-12 one.

The kernel repo kernel-team-tools kept moving over the weekend, with at least six tags discernible among its 35 commits: 7.3.4-4, 7.3.4-5, 7.3.4-6 and 6.18.63-5, 6.18.63-6, 6.18.63-7, while the accompanying 7.2.15 and 7.0.14 lines also advanced to -8 and -14. Configuration-side changes include enabling WERROR and making the config follow that switch, raising x86_64 NR_CPUS to 512, turning on STACKTRACE_BUILD_ID, changing the DA9063 driver to a module, making r8169 and the Realtek PHY modules with LED support enabled, and dropping ERRATA_SIFIVE_H_0_6_1. The gap between the two sides indicates the specs currently point to 7.3.4-3 and 6.18.63-4; the next refresh will catch up to this weekend’s tags.

2.3 llvm22/23/24 enable PIE and generate the clang resource dir; clang-wrap and cmake follow (11-13)

Date: 2026-11-13 Source: llvm PIE commit, clang-wrap commit

All three LLVM package versions enable PIE in the same commit and add clang’s resource dir. These two actions address the “compile and link with the distribution’s clang” scenario: without a resource dir, clang cannot find its own headers and runtime; without PIE, the distribution’s default hardening options fail at link time. clang-wrap was bumped to 20261113 in step with this change.

2.4 ruyifirmware: 8 commits — p550 uses OpenOCD 0.12 for memory recovery, board kernels unified on 7.3.4-3 (11-13)

Date: 2026-11-13 Source: p550 RAM recovery, ruyifirmware commit list

ruyifirmware saw 8 commits in the window: adding an OpenOCD 0.12.0-based memory recovery procedure for the SiFive P550 and simplifying that board’s UART and JTAG recovery instructions; unifying the firmware kernel source for all boards to Linux ruyi/v7.3.4-3; adding PLL1 definitions to K3’s OpenSBI; merging a RISC-V TCG fix for Unicorn; tidying up emulator-related documentation and the readable text console; and trimming the MS2130 capture card notes.

2.5 ruyi-imager: pausable downloads and whole-image progress, Gradle wrapper bumped to 9.9.1 (11-15)

Date: 2026-11-15 Source: PR #22, ruyi-imager commit list

The image flashing tool merged a feature change at 09:42 on 11-15: downloads can be paused, and the progress bar now reflects the whole image rather than individual chunks. A separate maintenance commit at 09:34 on 11-13 bumped the Gradle wrapper to 9.9.1. The change from the previous report that identifies local images by catalog metadata (nearly a thousand lines) had already landed earlier.

2.6 ruyi and packages-index: zero commits, 0.54.0 still the latest stable release (as of 11-16)

Date: 2026-11-16 Source: ruyi releases, packages-index

The ruyi package manager and the packages-index software package index both had zero commits in the window; the stable release remains 0.54.0 from 11-02, with no new beta. The openRuyi distribution news page also has no additions, the latest still being the two posts from 10-19 (the ecosystem infrastructure long-read and the Ruyi Firmware release); the 2026.11 monthly release has not yet shipped.

III. Vector Extensions and Operator Optimization

3.1 ABI mismatch in bf16 conversion helper functions: llc reads the return value from a floating-point register, but the launcher returns it as an integer (11-14)

Date: 2026-11-14 Source: PR #121, issue #114

A new PR fixing bf16 was opened on 11-14. In the launcher, two conversion helper functions were originally declared with integer prototypes: the truncation function returned uint16_t, and the extension function’s parameter was also uint16_t; but the call sites generated by llc pass bfloat through floating-point registers — on x86-64 the return value lands in xmm0, on riscv64 in fa0, and the caller reads the low 16 bits of the register. Integer returns go through eax/a0, so every kernel that converts an fp32 result to bf16 was storing zeros or garbage left over from the previous floating-point operation.

The PR’s verification was quite thorough: on an x86 host, disassembly shows the call site as callq __truncsfbf2 followed by vmovaps %xmm0 and then vpextrw, with eax never read; cross-compiling to riscv64 (with +zvfbfmin removed from the feature string) shows fa0 passed in and fmv.x.w reading back from fa0, while the helper’s integer return is in a0. The fix is to change the prototypes to float so both sides land in the same register class, with the bf16 bits still in the low 16 bits; __bf16 cannot be used because the launcher is compiled with g++, and g++ 13 on Ubuntu 24.04 does not have that type on riscv64. The newly added test_bf16_convert.py compares bit-for-bit against torch, covering multiple round-to-nearest-even edge cases; four of them failed before the fix.

3.2 Five batches of fifty-four FlagGems operator examples: 587 passing, 41 bf16 cases strictly xfailed (11-13)

Date: 2026-11-13 Source: PR #120

Another PR opened on 11-13 adds FlagGems operator examples to triton-riscv in batches: five independent batches combined into one integration sequence, 13 commits in total, 54 operator examples, covering unary math (acosh, arcsin, log1p, sinc, etc.), binary operations (bitwise_xor, copysign, fmax, nextafter, etc.), activations (hardshrink, mish, relu6, etc.), reductions (amin, argwhere, frobenius_norm, nansum, etc.), window functions and losses (bartlett, blackman, hamming, kaiser, bce, huber_loss, kl_div, l1_loss).

The five test groups on the x86 backend passed 77, 128, 114, 126, and 142 items respectively, 587 in total; another 41 bf16 cases were written as strict xfails, precisely because of the launcher ABI defect in 3.1 — an ungated riscv64 CI run showed all 41 bf16 tests failing. These xfails were initially scoped to x86 hosts only, then changed to apply unconditionally on all hosts, to be removed once #121 merges.

3.3 After the explicit target goes through -O3, the flash attention object gains three vector divides (11-14)

Date: 2026-11-14 Source: PR #117

A quantifiable change brought in by #117 is recorded in the test changes: the assembly assertions for the two attention kernels previously looked for scalar fdiv.s; after the middle-end runs at -O3 the division is vectorized, so the assertions were changed to accept vfdiv.vv/vfdiv.vf. Taking the flash attention kernel as an example, the object goes from 1 vfdiv.vf plus 1 fdiv.s and 3 vfmul to 4 vfdiv.vf and 36 vfmul. This is the most visible difference between the two build paths, “middle-end not run” and “run at -O3”.

3.4 wheel slimmed by 40% with strip-all: triton-shared-opt drops from 407 MB to 118 MB (11-13)

Date: 2026-11-13 Source: PR #116

#116 changes the strip applied to the bundled LLVM and buddy tools at packaging time from --strip-unneeded to --strip-all. Previously the single triton-shared-opt executable carried nine debug sections and 407 MB; after the change it is 118 MB, and applying the same treatment to the other six tools shrinks the entire wheel payload from about 970 MB to about 576 MB. The PR also records a rejected approach: pointing bin/llc at buddy-llc (the two produce identical output), but buddy-llc does not exit after writing its output, so it cannot replace the ordinary llc invoked by the non-IME path.

3.5 VecIntrinBench, IntrinTrans, and FBGEMM remain silent (11-16 status)

Date: 2026-11-16 Source: VecIntrinBench, IntrinTrans

The three repos for vectorization benchmarking and conversion tools still had no commits during the window: VecIntrinBench last at 02-13, IntrinTrans at 06-25, FBGEMM at 07-13. All vector-side work in this window took place in triton-riscv and buddy-mlir.

IV. Buddy Compiler Updates

4.1 0.1.4 Shares the Same Tree as the 11-13 Nightly; No New Commits on Main in the Two Days Before Release (11-15)

Date: 2026-11-15 Source: Nightly 0.1.4.dev20261113, Main commit list

Nightly 0.1.4.dev20261113 was built at 09:22 on 11-13 (Beijing time), and the official 0.1.4 was built at 08:01 on 11-15; both correspond to commit #1058 from 14:44 on 11-12. With no commits on main during the intervening two days, this release followed a “ship the nightly, observe for two days, then tag the official release” process rather than “release immediately after merge.” The previous report’s observation that “the nightly contains only the first commit of the day” saw no new counterexamples during these two days.

4.2 Open PR #1059: Replacing memref Copies with Vector Moves, Cutting the Sample from 66 ms to 2.8 ms (11-13 Status)

Date: 2026-11-13 Source: PR #1059

#1059 is still under review, but the data is worth recording now. It addresses a regression in finalize-memref-to-llvm: only memrefs with identity layouts or statically contiguous shapes can use llvm.memcpy; all other memref.copy operations degrade into memrefCopy calls that copy one element at a time—and the masked loads and stores in Triton kernels (dynamic lengths with offsets, matrix tiling) are exactly this kind of copy. When profiling on K3, such copies accounted for 44% to 99% of kernel time for add, exp, softmax, and matmul after lowering through VIR.

The fix lowers memref.copy with unit stride in the last dimension into vector loads and stores using vir.set_vl on the last dimension, with a loop over rows; in the fixed-vector path, the main and tail loops change from affine.for to scf.for, because the vl computed inside the loop is not a valid affine operand. Results on K3: add on X100 (VLEN 256) drops from 66.1 ms to 2.76 ms, and on A100 (VLEN 1024) from 138.9 to 2.26; across 1811 FlagGems kernels, memrefCopy calls drop from 3694 to 819 (affected kernels from 1499 to 317); the 4043-test suite still passes 4036 tests on SG2044 and K3, with the other 7 failing under the original pipeline as well.

4.3 Downstream Renovate Still Pinned to dev20261113, Not Yet Moved to the Official Release (11-16 Status)

Date: 2026-11-16 Source: PR #115

The Renovate dependency update for triton-riscv is still under review, bumping buddy-mlir from nightly/v0.1.4.dev20261103 to dev20261113. Now that the official 0.1.4 is out, this dependency should be re-pinned to release/v0.1.4, but no new PR has followed up yet. Given that 0.1.4 and dev20261113 share the same tree, which tag is pinned makes no functional difference—the difference lies in traceability.

V. Community, Events, and Official Homepage

5.1 Forum: One routine hot-topic post on 11-13, no new posts over the weekend, biweekly progress still at issue 080 (11-13)

Date: 2026-11-13 Source: 2026-11-13 Open Source Technology Hot Topics, RuyiSDK Biweekly Progress Issue 080

Within the window, the RuyiSDK community forum gained only one open source technology hot-topics post on 11-13; there were no new posts on 11-14 or 11-15. Biweekly progress remains at issue 080 from 11-11 (three tools released new versions); issue 081 has not appeared, and based on the usual cadence, the next issue will land in late November. The Ruyi Imager beta-test thread received one reply on 11-11, with no new additions this week.

5.2 Official homepage: three pages unchanged, two blog posts, twenty-three documentation chapters (11-16 status)

Date: 2026-11-16 Source: RuyiAI Official Homepage, Blog Page

The sha256 values of the official homepage’s index, blog, and docs pages match the previous issue, with no changes: the blog still contains the two posts “Triton’s RISC-V Port” and “PyTorch RISC-V CI in Practice,” and the documentation index has twenty-three chapters. Over the three-day window, official channels published no new announcements, and community-side information was concentrated mainly on GitHub.

5.3 ruyiai-daily: eighth day of hiatus (11-16 status)

Date: 2026-11-16 Source: ruyiai-daily commit list

The daily-report repository’s last entry is still 11-08 06:21 (corresponding to the 2026-11-08 issue), with no new commits from 11-09 through 11-16—eight days in total. The hiatus has now exceeded the five days recorded in the previous report, and this repository is currently unsuitable for use as a source.

Date: 2026-11-16 Source: arXiv cs.AR latest, openRuyi News Page

Six groups of Chinese and English Google News queries returned no hits for RuyiAI, RuyiSDK, openRuyi, or the Ruyi community within the window; arXiv also had no new papers in the RISC-V AI compiler direction. On the general RISC-V news side, a few items serve as background: FURSOY released an open-source RISC-V core that can boot Linux and synthesizes at 500 MHz; Purism’s RISC-V phone entered crowdfunding; and FIORA5 opened a funding round for its edge-AI RISC-V processor. None of the three is directly related to the RuyiAI stack and is recorded only as external context.

VI. Trend Observations

6.1 The packaging layer has become the new battleground for RuyiAI adoption

The two heaviest items in this window are not about algorithms: one is turning a PyTorch snapshot into a distro package, the other is cutting wheel size by forty percent. Of the six patches in the torch package, three are build and ABI issues (CBLAS complex dot, system fmt and concurrentqueue, submodule checks), one is a macro definition gated on the wrong condition (__riscv_v_intrinsic vs __riscv_vector), and only one is a genuine feature adaptation. This shows the RuyiAI stack has moved from “can it compile” to “can it be installed, can it be linked, can other packages depend on it”—the pitfalls at this layer are more tedious than vectorization in the kernel, and they matter more to whether a user’s first attempt succeeds.

6.2 The stable release is the same tree as the nightly

The way 0.1.4 was released is worth noting: the last batch of VIR changes landed on 11-12, a nightly came out on 11-13, the stable tag was cut on 11-15, with zero commits in between. This effectively defines “release” as confirmation of an existing nightly rather than another round of integration. That is good for downstream—the nightly pinned by Renovate is the same tree as the stable release, so switching over brings no functional difference, just a more traceable version number.

6.3 The performance ledger has turned to the copy page

The data in #1059 shows that the biggest cost in the VIR path right now is not arithmetic but data movement: masked loads and stores degrade into element-by-element copies, accounting for 44% to 99% of kernel time, which makes the simplest operator like add forty times slower than the non-VIR path (66.1 vs 1.60 ms). After switching to vector moves a gap remains (2.76 vs 1.60), but it is back in the same order of magnitude. This thread follows directly from last issue’s “shape coverage”: once shapes are filled in, the next gains come from pushing down peripheral overhead such as copies and layout transforms.

6.4 CI criteria are still being trial-and-errored

The pytorch riscv64 CI changed its criteria twice in one day: first adding the Jenkins return code, then switching to reading a stats file. Such back-and-forth is common in small teams, but it exposes a non-trivial problem—a pipeline that “builds successfully but executes zero test cases” renders every downstream green badge meaningless. Moving the criterion from job status to test-case count is the right direction;真正稳定下来还需要把统计文件的口径固化下来。


Source Verification Table

Source Verification Result
RuyiAI-Stack key repos pytorch riscv branch gained two CI changes (#56, #57), plus twenty-six rebase-and-repush commits and three upstream syncs on 11-14; triton-riscv merged two (#116, #117), three nightlies, three newly opened PRs (#120, #121, and #115 still under review); tilelang-riscv, llvm-project, pytorch-ci, FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, harness, torq-tile, dsh-osc had zero commits
ruyisdk org ruyi zero commits, zero releases (0.54.0 still latest); packages-index zero commits; ruyi-imager two commits (pause download and full-image progress, Gradle 9.9.1)
openRuyi-Project Packaging repo forty-five commits (including the new python-torch-ruyiai, Go 1.27.4 and thirty-plus Go packages, llvm PIE, clang-wrap, cmake, KDE Bluetooth, lmbench, linux and linux-lts refresh); kernel-team-tools thirty-five commits, six-plus tags; ruyifirmware eight commits; opensbi, edk2-ruyi, coreboot, MultiArchUefiPkg had pushes but no new commits within this window
buddy-compiler buddy-mlir zero commits this window; stable 0.1.4 and nightly 0.1.4.dev20261113 released, both on the same tree; open PR #1059 still under review
RuyiSDK Forum One routine hot-topic post on 11-13, no new posts over the weekend; biweekly progress still at issue 080
RuyiAI official homepage index, blog, docs pages unchanged; 2 blog posts, 23 doc chapters
ruyiai-daily Stalled at 11-08, eight issues from 11-09 to 11-16 not generated
openRuyi distro news page No additions, latest still the two posts from 10-19; the 2026.11 monthly release not published
Google News Zero hits across six groups of targeted Chinese and English queries within the window, no external RuyiAI-related coverage
arXiv No new papers on RISC-V AI compiler directions within the window

Full Source List

Limitations

This issue’s window is three days (11-13 04:30 to 11-16 04:30), covering the weekend, with Monday as usual. On the GitHub side, anonymous APIs and Atom feeds were cross-checked; API quota is limited, so commit counts for some repos follow the Atom basis; pytorch fork commit counts are affected by the 11-14 branch rebase, and the body distinguishes new from old by PR merge time. On the external news side, Google News was scraped via proxy, and all six query groups returned no hits; coverage not indexed by Chinese-language media cannot be ruled out.