RuyiAI Daily Intelligence Report (2026-11-20)
Research window: 2026-11-19 04:30 ~ 2026-11-20 04:30 (Beijing time, past 24 hours; directly contiguous with the cutoff of the previous 11-19 report).
Sources: Verification of pushes across key GitHub repositories and organizations, covering 18 RuyiAI-Stack repos, 29 ruyisdk repos, 8 buddy-compiler repos, and the openRuyi-Project packaging repo. Commits, PRs, and Releases within the window were checked repo by repo; additionally reviewed were the descriptions of open PRs in triton-riscv and buddy-mlir, merge records in the llvm-project fork, and SPECS changes in openRuyi. On the community side, reviewed were the RuyiSDK biweekly report repo, the support matrix, the latest topics on the developer forum, and snapshots of three pages of the official homepage. External leads came from the ruyiai-daily daily-update repo, the openRuyi distribution news page, targeted Chinese and English queries via Google News RSS (through a proxy), and arXiv searches.
Issue Index
- Today’s Highlights: Software repository lists QEMU 11.1.0 AME and RVP packages; Biweekly Report 081 published the same day (11-19)
- Today’s Highlights: buddy-mlir merges ZTT v0.6 and LLVM matrix extensions; triton-riscv backend tuning follows up (11-19)
- I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
- 1.1 triton-riscv nightly 20261119 catches up with #125, zero commits on mainline during window (11-19)
- 1.2 FlagGems examples triple: 63 + 43 + 40 operators across three branches pending merge (11-19)
- 1.3 RISC-V backend tuning PR #129 opened, depends on Buddy pass not yet merged (11-19)
- 1.4 pytorch riscv branch has zero commits, main continues upstream sync (11-19 status)
- 1.5 tilelang-riscv still stuck at 11-16 (11-19 status)
- II. RuyiSDK Toolchain and Distributions
- 2.1 packages-index 13 commits: QEMU adds AME and RVP, batch image updates (11-19)
- 2.2 Biweekly Report 081 finalized: ruyi 0.55.0 expected to slip to end of November (11-19)
- 2.3 ruyi-imager adds flash-write window-close confirmation, site adds German (11-19)
- III. P Extension, Vector, and Runtime Optimization
- 3.1 GCC delivers 3 P extension examples, 13 commits to mainline (11-19)
- 3.2 LLVM 14 P extension PRs, 13 merged (11-19)
- 3.3 V8 on K3 JetStream 3 total score +3.5% (11-19)
- 3.4 OpenJDK and Go enter final stretch before 1.28 freeze (11-19)
- 3.5 VecIntrinBench, IntrinTrans, and FBGEMM still silent (11-19 status)
- IV. Buddy Compiler Updates
- 4.1 AME dialect replaced with ZTT v0.6, 138 instruction ops merged (11-19)
- 4.2 LLVM bumped to 163d6a7, XSAI’s ZAMES and ZMASYNC enter backend (11-19)
- 4.3 EmbeddingGemma-300M speeds up 23 to 38x on K3 (11-20)
- 4.4 GPT-2 decoding fix opened, polynomial approximation math pass pending review (11-19)
- V. Community, Events, and Official Homepage
- 5.1 Official homepage three pages unchanged, 2 blog posts, 23 doc chapters (11-20 status)
- 5.2 Forum latest topic still stuck at 11-07, openRuyi news page has no new version (11-20 status)
- 5.3 openRuyi 84 commits in one day: 61 Go dependency packaging commits, grpc migrates to system OpenSSL 4 (11-19)
- VI. Trend Observations
- 6.1 Matrix extensions connect from simulator to backend in one line
- 6.2 Software repository filling in both P and matrix extension chains
- 6.3 Distribution side doing Go and OpenSSL 4 migration
Today’s Highlights: Software repository lists QEMU 11.1.0 AME and RVP packages; Biweekly Report 081 published the same day (11-19)
Sources: Biweekly Report Issue 081. qemu-user-riscv-ame commit. qemu-system-riscv-rvp commit.
From 10:33 to 14:42 on 11-19, packages-index landed 13 consecutive commits. Four of them add four simulator software packages: one set each for qemu-system and qemu-user, with AME and RVP support respectively.
All four packages are based on upstream QEMU 11.1.0. The AME package version string is 11.1.0-ruyi.20261016, and the RVP package is 11.1.0-ruyi.20261027. Both extension sets correspond to the v0.6 draft.
Biweekly Report 081 states that AME support was contributed by DAMO Academy. This is the first time matrix extensions have entered the RuyiSDK software repository as installable software packages.
The same biweekly report was finalized on 11-19. The wechat-articles repo landed 9 commits during the window, filling in from skeleton and body to foreword in sequence (#411 to #418).
The package manager’s cadence has been adjusted. The biweekly report states that ruyi 0.55.0 is expected to be released at the end of November, with the current latest still 0.54.0 (11-10). The official reason given is a release cycle adjustment.
The remaining updates are batch image follow-ups. Armbian’s VisionFive 2 and Orange Pi RV2 minimal images are bumped to 26.11.0-trunk.120. LicheeRV Nano’s Buildroot SDK is bumped to 0.20261109.0.
FreeBSD riscv64 mini-live is bumped to 14.5, OpenBSD riscv64-live is bumped to 8.0, and 7.6.0, no longer provided upstream, is removed. OpenWrt’s SiFive Unmatched image is bumped to 25.12.6.
CanMV K230 fixes a broken download link. On the CI side, another commit describes telemetry upload as non-blocking for workflows.
Two more image PRs remain unmerged. Bianbu K3 bump to 4.0.8 (#250) and removal of the old VisionFive 2 image (#249) are both still open.
Today’s Highlights: buddy-mlir merges ZTT v0.6 and LLVM matrix extensions; triton-riscv backend tuning follows up (11-19)
Sources: buddy-mlir PR #1066. PR #1067. triton-riscv PR #129.
From 19:20 to 19:21 on 11-19, buddy-mlir merged two PRs in succession. #1067 bumps the LLVM submodule from 542365c9cd38 to 163d6a753b58.
The new pointer points to the #33 merge point of the llvm-project fork. That merge adds XSAI’s ZAMES matrix instructions and ZMASYNC synchronization instructions. Coverage includes matrix and synchronization intrinsics, sizeless target extension types, assembly and disassembly, and RV64 SelectionDAG lowering.
The support scope has boundaries. Matrix copy, spill and reload, and cross-function ABI are still not covered. The PR notes that only the submodule pointer is changed, with +1 −1 lines.
#1066 replaces the retired AME dialect entirely with ZTT v0.6, still using the ame namespace. It defines 138 typed instruction ops, with custom assembly syntax and an intrinsic bridge.
The lowering path uniformly goes through lower-ame, landing on llvm.riscv.ztt.* calls. The PR includes 8×8 and 4×4 example sets, plus validation of two freestanding workloads on gem5. The change touches 30 files, +8349 −1277.
Downstream caught up the same day. triton-riscv opened #129 at 18:20, tuning the buddy-opt lowering parameters for the RISC-V target. The change only takes effect when the target is RISC-V, across 2 files, +137 −12.
The tuning has three parts. Math functions switch to polynomial approximation, covering 13 operators including exp, log, tanh, erf, sin, cos, atan. Vectors switch to scalable vectors with the target’s minimum VLEN, with avl loop epilogue. The vector width for matrix multiplication is changed to 2·VLEN/32.
VLEN is read from the target’s +zvlNb. When only +v is given, it counts as 128. The author compared parameters tuned for triton-cpu on a VLEN-256 RISC-V board.
This PR depends on buddy-mlir #1068, which has not yet been merged. The PR notes that using an older buddy-opt will fail directly at unknown pass or option.
I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Porting)
1.1 triton-riscv nightly 20261119 catches up with #125, zero commits on main within the window (11-19)
Source: nightly 20261119. PR #125.
At 09:44 on 11-19, triton-riscv released nightly 20261119-85d3bae. The tag points to commit 85d3bae, i.e. #125 “Strip the copies the wheel actually packages”.
#125 was merged at 18:43 on 11-18. The previous day’s nightly 20261118-e9c34b7 was built at 09:59 on 11-18, earlier than this merge. So 20261119 is the first nightly to carry the wheel slimming fix.
Main had no new commits within the window. The latest is still 85d3bae, timestamped 11-18 18:43. Renovate’s #115 remains open, with the buddy-mlir dependency still pinned to nightly/v0.1.4.dev20261113.
The nightly has no release notes; the body is empty. The only verifiable information is the commit and timestamp the tag points to.
1.2 FlagGems examples land in three stacked PRs: 63 + 43 + 40 operators pending merge (11-19)
Source: PR #126. PR #127. PR #130.
FlagGems examples stacked into three branches within a single day. #126 covers 63 aten operators, #127 adds 43 torch.special operators, and #130 adds another 40 activation, dropout, and loss operators.
The three are stacked serially. #127 sits on top of #126, #126 on top of #120, and #130 on top of #127. Each PR description notes that while predecessors are unmerged, the diff counts predecessor commits as well.
Per their respective descriptions, the sizes relative to main are 239, 325, and 405 files. #130’s own portion is 4 commits and 81 files; the rest belongs to preceding PRs.
Test counts rise layer by layer. #126 reports 6416 passing, #127 reports 6782, and #130 reports 7091. Newly added cases are 309 and 366 respectively.
#127 covers special functions. The Bessel family, four kinds of Chebyshev plus shifted variants, two kinds of Hermite, Laguerre, and Legendre are all included, along with log_softmax and logsumexp.
Intrinsics that the toolchain cannot lower are replaced with closed forms. erfcx uses a three-segment hybrid, with an fp32 segment, an fp64 segment, and an asymptotic series each handling one range. The tail of log_ndtr is computed as −t² + log(erfcx(t)/2), avoiding ndtr underflow.
#130 leans toward neural-network operators. Backward formulas for GELU and GLU, fused conv/addmm/prelu, dropout and randomized-ReLU variants, and a loss group including CTC forward and backward.
CTC’s lattice dynamic programming uses an fp64 Triton kernel. The backward pass keeps the beta scan on the host side while sanitizing the incoming alpha. The author notes that aten’s CPU implementation leaves unreachable lattice cells uninitialized.
The cudnn epilogue of convolution is routed through aten. Triton keeps only the bias, Z addition, and ReLU. Triplet margin follows pairwise_distance semantics, with eps broadcast onto the difference vector.
All three PRs limit their scope to python/examples/flaggems. The descriptions state that the backend, CI, and other PRs’ files are untouched.
1.3 RISC-V backend tuning PR #129 opened, depends on a Buddy pass not yet merged (11-19)
Source: PR #129. PR #128. buddy-mlir PR #1068.
#129 was opened at 18:20 with the same title as #128. #128 was opened at 10:19 and closed at 10:20, then reopened as #129 with identical content.
The change centers on one function in the Triton-to-LLIR path. Arguments are appended only when the target is RISC-V; other targets are unaffected. The author states no binding to a specific VLEN.
The arguments fall into three categories. Math functions enable polynomial approximations, covering exp, expm1, log, log2, log1p, tanh, erf, erfc, sin, cos, atan, atan2, asin, acos, cbrt.
Vectorization switches to scalable vectors, with width taken from the target’s minimum VLEN, and enables avl loops. Matrix multiply vector width becomes 2·VLEN/32, versus the previous constant 4.
The accuracy basis is on the Buddy side. #1068’s description gives 1 to 7 ULP for f32 and at most 1 ULP for f16 and bf16. The exceptions are subnormal results for exp and erfc.
The dependency is a hard constraint. #1068 is not yet merged, and buddy has no release carrying these two features. The PR description states that older buddy-opt versions will stop at the unknown pass or option.
1.4 pytorch’s riscv branch has zero commits; main continues upstream sync (11-19 status)
Source: riscv branch commit list. main branch commit list.
The fork has 9 branches in total. Besides main and riscv, there are riscv-base, improve-riscv64-ci, agent/riscv-linear-cross-entropy-ulp, and others.
Within the window, both riscv and improve-riscv64-ci have zero commits. The main sample hit the 100-entry cap, and the content is still upstream traffic.
The sampled commits center on oneDNN type replacements, XPU distributed test porting, ROCm CI labels, and Inductor’s libdevice calls. No RuyiAI RISC-V changes are visible.
1.5 tilelang-riscv still stuck at 11-16 (11-19 status)
Source: tilelang-riscv commit list.
tilelang-riscv has zero commits within the window. The latest is stuck at 11-16 14:56, marking four consecutive days without new commits.
The repo sits under RuyiAI-Stack and handles TileLang’s RISC-V porting. It is outside this round’s FlagGems expansion in triton.
II. RuyiSDK Toolchain and Distributions
2.1 packages-index: 13 commits — QEMU adds AME and RVP, batch image updates (11-19)
Sources: qemu-user-riscv-rvp commit. qemu-user-riscv-ame commit. qemu-system-riscv-ame commit.
The four emulator packages landed system first, then user, alternating between AME and RVP. Commits ran from 14:34 to 14:42 — all done within nine minutes.
Package names and versions map as follows. qemu-user-riscv-ame is 11.1.0-ruyi.20261016, and qemu-system-riscv-ame is the same version. qemu-user-riscv-rvp and qemu-system-riscv-rvp are both 11.1.0-ruyi.20261027.
The AME packages correspond to Matrix extension draft v0.6. The biweekly report states the support was contributed by DAMO Academy. The RVP packages correspond to P extension draft v0.6.
Image updates landed in a burst between 11:04 and 11:11. The two Armbian minimal images were upgraded first, followed by LicheeRV Nano, FreeBSD, and OpenBSD.
OpenBSD 8.0 and the removal of 7.6.0 were two separate commits. The removal rationale states that upstream no longer supports or provides that version.
Download links for all three CanMV K230 board types were fixed together. The OpenWrt SiFive Unmatched image was bumped separately to 25.12.6.
The earliest commit was a CI adjustment. Telemetry upload failures no longer fail the entire workflow.
2.2 Biweekly Report 081 finalized: ruyi 0.55.0 expected to slip to end of November (11-19)
Sources: Biweekly Report No. 081. ruyi release list.
Biweekly Report 081 received 9 additional commits between 11-19 10:58 and 15:57. The last commit of the day was the opening-remarks update (#418), following earlier additions to the documentation and ruyipkg sections.
The package manager section notes a release cadence adjustment. 0.55.0 is expected at the end of November; the current latest stable release is 0.54.0 (11-10). The ruyi repo saw zero commits during the window and no new Release.
There are five in-development items. Greater tolerance for malformed software sources, ignoring relative-path configuration per the XDG spec, exposing repo.disabled in repo get, an updated troubleshooting guide for download failures, and official support for Python 3.15.
The two new packages on the software-source side were listed in the previous section. The biweekly report groups them under packages-index resource updates and gives the ruyi update method for fetching them.
The IDE was updated as well. Each of the three Linux platforms now bundles the ruyi program, the plugin switched to the release build on OpenVSX, and an unused button on the news page was removed.
2.3 ruyi-imager adds close-window confirmation during flashing; site adds German (11-19)
Sources: ruyi-imager #25. ruyisdk-website commit f9de316. Commit 8a55dcf.
ruyi-imager had one commit in the window. It adds a confirmation prompt before closing the window during a flashing operation (#25), corresponding to commit 21e2c86.
The website repo had 13 commits in the window. The largest block is German localization, filled in across the translation directory, pages, docs, and news; commit f9de316 explicitly states that the German locale is enabled.
Two UI changes were made. The color-mode toggle moved below the back-to-top button, and another commit made the toggle actually take effect. Icons on the download cards and the image-writing entry point were switched to lucide glyphs.
On the ops side, a scheduled job was added to auto-update submodules (#618). On the docs side, the Ruyi Imager sidebar labels were also localized.
The kernel repo had one additional commit. linux-xuantie-kernel merged a sign-extension fix for riscv/atomic, covering the unsigned path of arch_cmpxchg (commit da94ef7).
The support-matrix, board-docs, and docs repos saw zero commits in the window. The most recent commit in the support matrix remains the 11-18 Lichee Pi 4A test report.
III. P Extension, Vector, and Runtime Optimization
3.1 GCC Delivers 3 P Extension Examples, 13 Commits Land in Mainline (11-19)
Source: Biweekly Report No. 081. GCC mainline commit fb66e0c.
The three P extension examples run directly on RV32 and RV64. The scenarios are image matching, Q15 fixed-point dot product, and Q15 digital filtering; the repository includes a prebuilt toolchain and two interactive tutorials.
In the image example, each packed instruction compares 4 pixels on RV32 and 8 on RV64. The program iterates over 289 candidate positions and recovers a displacement of 5 pixels right and 3 pixels up.
Dynamic instruction reductions are measured with QEMU at -O3 with unaligned memory access enabled. Image matching drops 75.7% on RV32 and 80.0% on RV64. Q15 dot product drops 70.7% and 74.9% respectively, and Q15 filtering drops 69.5% and 73.2%.
On the GCC side, 13 commits land in mainline. Eight improve bit manipulation and rotation, four support vector dot product, and one fixes a general loop optimization.
Gains on K1 hardware are concentrated in scalar bit manipulation. Setting, clearing, and flipping 32-bit masks all cut execution cycles by about 27%. The RV64 single-bit set function shrinks from 3 instructions to 2.
Other microbenchmarks range from positive to flat. Shift-add followed by rotate left or right saves about 19.8% of cycles; when followed by division or modulo, dynamic instructions drop 8.30% with cycles unchanged. Concatenating two 32-bit words on gem5 saves 32.77% of cycles, while SQLite varint decoding regresses by 0.63%.
The Zvdota and Zvbdota vector dot products have landed in mainline. C/C++ can call integer, BF16, and FP8 dot products plus batched FP32 dot product.
On the Binutils side, Shlcofideleg v2 was submitted. RISE’s pre-commit CI ran a manual dry run for the first time, with three configurations that automatically compare patch test results against the production baseline.
3.2 LLVM: 14 P Extension PRs, 13 Merged (11-19)
Source: Biweekly Report No. 081. LLVM PR #243813.
Fourteen P extension PRs were submitted this period, 13 of which are merged upstream. The remaining one is still in review.
The merged set has broad coverage. Legalization of widening shifts is restricted to left shifts, fixing the issue where right shifts were mis-selected as left-shift instructions.
New intrinsics include scalar saturating shift-add, scalar comparison, saturating and rounding shifts, high/low halfword multiply-accumulate, scalar zip/unzip, and narrowing clip.
There is also a batch of fixes. zext_invec gains a v8i8 to v2i32 match pattern, packed shifts return the original operand directly on zero, and RV32 i64 shift-amount legalization no longer fails an assertion.
Crashes in WADDA and WSUBA with a zero accumulator are also fixed. The cause was the zero accumulator being selected as X0_Pair; register pairs are now built with buildGPRPair.
On the GlobalISel side, KCFI type information is preserved. Indirect calls can therefore generate checks, matching SelectionDAG behavior.
Still in review is #243813. It folds constant shifts exceeding the element bit width per the P extension’s signed shift semantics, producing zero or an equivalent fill plus most-significant-bit extraction.
3.3 V8 on K3: JetStream 3 Total Score +3.5% (11-19)
Source: Biweekly Report No. 081. CL 8698757.
This period’s V8 gains come from a batch of performance patches. The biweekly report states that these patches were located with tools such as riscvlint, drafted with AI assistance and benchmark-validated, then merged after manual review by ISCAS maintainers.
The largest single item is short jumps after the trampoline pool. Previously 8- to 12-byte auipc plus jr, these now emit 4-byte jal, with a new far-jump pool ensuring targets are reachable.
On K3 hardware, JetStream 3 total score improves 3.5%. Of 75 items, 46 exceed noise; gaussian-blur alone improves 27%, and wasm items improve 17% to 22%. Generated code shrinks by 1.7 MB, about 2.2%.
Zfa’s fli also lands in TurboFan and Liftoff. Each floating-point constant goes from 3 instructions to a single fli.d, eliminating 248 occurrences in JetStream 3.
There are two more cleanups. PushAll and PopAll with empty register lists return early, removing about 90,000 useless instructions from Maglev. Int32 comparisons compare registers directly, cutting instruction count by a further 0.47%.
The built-in simulator also benefits. After IsShortInstruction is inlined into the header, richards drops from 28.9 s to 19.0 s, a 34% reduction; ML drops from 40.2 s to 27.3 s, a 32% reduction.
Caching of debug checks saves another tenth. Six checks are skipped when debugging is off, cutting richards by 27% and ML by 28%. After optimizing the decode path, richards drops a further 20% and ML a further 22%.
Thirteen patches were merged and 15 reviewed this period. Topics include heap number deduplication, sign extension of Liftoff stack arguments, modeling of fence and AMO memory ordering, and CAS address computation order.
3.4 OpenJDK and Go Wrap Up Before the 1.28 Freeze (11-19)
Source: Biweekly Report No. 081. OpenJDK PR #32635. OpenJDK PR #32727.
OpenJDK lists 10 PRs this period, mostly RISC-V related. They fall into four categories: hardware support, code generation, correctness fixes, and redundancy cleanup.
Hardware support adds vendor feature detection for XuanTie processors (#32727). On code generation, conditional arithmetic is optimized with Zicond (#32635), and outer barriers are removed when atomic instructions already provide memory ordering (#31955).
Correctness fixes include PC recording for JNI fast field access, mismatching of weak memory-ordering instructions, support conditions for half-precision FMA, and branch offset checks now also taking effect in release builds.
Cleanup removes an unreachable non-acquire atomic match rule, a redundant null check, and the broken VerifyStackAtCalls debug option.
On the Go side, only one week remains before the 1.28 feature freeze, and changes are mostly bug fixes. New CLs cover exponential function optimization on riscv64 (864281) and switching cgocallback to ABIInternal (864283).
There is also a blocker fix disabling fixedbugs/issue80271.go on riscv64 (864243). AES zvkned block encryption/decryption has been merged (765200).
Three test-side items were opened. Strict allocation checks for database/sql, TestUnsafePoint, and assembly tests for Zfa instructions are all in progress.
3.5 VecIntrinBench, IntrinTrans, and FBGEMM Remain Silent (11-19 Status)
Source: VecIntrinBench commit list. IntrinTrans commit list. FBGEMM commit list.
All three repositories remain at zero commits. VecIntrinBench is stalled at 02-13, IntrinTrans at 06-25, and FBGEMM at 07-13.
The vector intrinsic benchmark and translation toolchain have been untouched for three quarters. The vector performance data currently available comes mainly from triton-riscv operator tests and the microbenchmarks in the biweekly reports.
IV. Buddy Compiler Updates
4.1 AME dialect replaced with ZTT v0.6, 138 instruction ops merged (11-19)
Source: buddy-mlir PR #1066.
#1066 was merged at 19:21, touching 30 files, +8349 −1277. It replaces the retired AME dialect wholesale with ZTT v0.6, keeping the ame namespace.
The new dialect defines 138 typed instruction ops. Each op comes with custom assembly syntax, an intrinsic bridge, and matrix/accumulator target extension types.
Lowering goes uniformly through lower-ame, landing on llvm.riscv.ztt.* calls. The intrinsic bridge retains a path that translates directly to LLVM IR.
Verification is two-layered. One set covers lowering, diagnostics, all intrinsics, mixed dialects, side effects, and backend selection. The other is two freestanding payloads run on AME_gem5 in atomic syscall emulation mode.
The 4×4 signed i32 payload passes bit-for-bit comparison against an independent scalar reference. The PR description states that the old ISA contract is replaced, with no promise of numerical behavior compatibility with the old implementation.
Documentation is not in this PR. README and standalone docs are left to a follow-up PR, and the BOSCAME and XTAME implementations are unchanged.
4.2 LLVM bumped to 163d6a7, XSAI’s ZAMES and ZMASYNC land in the backend (11-19)
Source: buddy-mlir PR #1067. llvm-project PR #33.
#1067 was merged at 19:20, changing only the LLVM submodule pointer, +1 −1. The old pointer was 542365c9cd38, the new one 163d6a753b58.
The new pointer corresponds to the merge commit of #33 in the llvm-project fork. That merge adds XSAI’s ZAMES matrix instructions and ZMASYNC synchronization instructions to the RISC-V backend.
Coverage includes matrix and sync intrinsics, sizeless target extension types, assembly and disassembly, and RV64 SelectionDAG lowering.
The boundaries are equally clear. Matrix copy, spill and reload, and cross-function ABI support are not yet supported.
This update is a prerequisite for native ZAMES support in Buddy. The PR description states that native instructions can only be wired up on the Buddy side once the LLVM backend is in place.
4.3 EmbeddingGemma-300M speeds up 23–38× on K3 (11-20)
Source: buddy-mlir PR #1070.
#1070 was opened at 01:56 on 11-20, touching 19 files, +3072. It builds on #835, fixing model import, RAX packing, and tokenizer compatibility.
The optimizations are IME and RVV. A single forward pass on K3 is 23.24–38.24× faster than the F32 RAX baseline and 7.11–19.03× faster than native Hugging Face F32.
Memory usage drops in step. Peak process RSS is down 71.6%–93.5% versus the RAX baseline. All nine payloads and 81 formal runs pass the embedding accuracy threshold.
The import-side fix switches to a whole-graph approach. The bidirectional additive mask is constructed in the model wrapper layer and imported in one pass through pooling, two dense layers, and L2 normalization, replacing the old version’s textual rewriting of generated MLIR.
Batch shapes and sliding attention are also fixed. Import uses eager attention, passing the full mask and sliding mask separately, no longer assuming B1.
4.4 GPT-2 decode fix opened, polynomial-approximation math pass under review (11-19)
Source: buddy-mlir PR #1069. PR #1068. PR #1065.
#1069 was opened at 23:57, fixing a numerical error in GPT-2 decoding. The problem was that batched matmul with transposed B was treated by decode vectorization as a non-transposed layout.
The linalg layer has no separate transposed-B op. Both share linalg.batch_matmul and are distinguished only by indexing_maps. In the decode pipeline two passes run back to back, and the first vectorizes all batch_matmul ops under the default layout.
The fix adds a guard. Only default indexing_maps enter vectorization; transposed B is left to the dedicated pass that follows, and such ops are marked dynamically legal. 2 files changed, +127.
#1068 is the math approximation itself and is still under review. The new pass replaces vector math such as exp, tanh, and log with polynomials, keeping computation in vector registers instead of calling libm per lane.
Upstream MLIR’s approximations have edge-case problems on six ops. log and log2 clamp subnormal inputs to FLT_MIN, sin and cos single-constant reduction distorts on large arguments, and atan2’s quadrant determination is also off. This pass writes separate patterns for them.
sin and cos use three-stage Cody-Waite reduction, with arguments beyond 1e5 going through 192-bit Payne-Hanek. The branch is only taken when such lanes appear in the vector.
The loop tail is also changed. The avl loop runs the remainder in a single active-length iteration instead of splitting in half and adding a scalar loop.
#1065 was closed at 11:10 the same day without merging. It implemented cache-friendly tiling for linalg.transpose and was created on 11-18.
V. Community, Events, and Official Homepage
5.1 Official homepage: three pages unchanged; blog has 2 posts, docs have 23 chapters (11-20 status)
Sources: RuyiAI official homepage. Blog page. Docs page.
The hashes of all three pages match the previous issue. index.html is 7246 bytes, blog.html is 5632 bytes, and docs.html is 10843 bytes; all are marked unchanged.
The blog index still has two posts: Triton’s RISC-V port and PyTorch RISC-V CI practices. The docs index has 23 chapters, with no new chapters.
The latest commit in the official site repo ruyiai-stack.github.io is dated 11-18 22:50, with no commits during the window.
5.2 Forum’s latest topic still stuck at 11-07; no new release on the openRuyi news page (11-20 status)
Sources: Latest topics on the RuyiSDK community. openRuyi news page.
The latest topic on the forum homepage is still the 11-07 /proc/cpuinfo ISA baseline thread. There were no new threads during the window, and no new biweekly report posts; the biweekly report on the forum is still at issue 079 (10-26).
The openRuyi news page has no additions. The latest is still the two posts from 10-19, one on public infrastructure and one on Ruyi Firmware.
The monthly release is still 2026.09 (09-30). Neither 2026.10 nor 2026.11 has been released; on the distro side there are currently only rolling updates in the packaging repo.
5.3 openRuyi: 84 commits in one day; 61 for Go dependency packaging, grpc migrated to system OpenSSL 4 (11-19)
Sources: openRuyi commit list. grpc commit b4d00ef. grub commit 6000ecc.
The openRuyi packaging repo saw 84 commits during the window, covering 36 SPECS packages. The time span runs from 11-19 09:38 all the way to 11-20 01:17.
The largest block is Go dependency packaging, with 61 commits. Newly added packages include gofumpt, unparam, misspell, structtag, paho.golang, azure-go-amqp, zerolog, testify, and others.
This batch also fixed the build dependencies of existing Go packages. go-opentelemetry, go-google-protobuf, go-golang-x-sys, go-golang-x-sync, and go-golang-x-mod are all included.
The second block is the networking and RPC stack. grpc was upgraded to 1.85.3 and switched to linking against system OpenSSL 4; protobuf was upgraded to 36.2, with protobuf-c following to the new field API. re2 was upgraded to 2025.11.05, and abseil-cpp to 20260817.0.
The third block is the boot chain. grub added kernel-install integration and began shipping /etc/default/grub with the package.
The rest is routine version follow-ups. ffmpeg was upgraded to 9.0.2; postfix to 3.11.7 with dynamic lookup modules enabled; openvpn to 2.7.9; and tzdata to 2026f.
Miscellaneous items also include socket_wrapper 1.5.2, python-libvirt 12.9.0, and python-sip 6.17.3. There are also qca 2.3.12, harfbuzz 15.1.0, and meson 1.12.1.
The opensc update fixes the build under OpenSSL 4.0.0. The four sequoia packages were uniformly given the default-feature dependency on bindgen-0.72.
VI. Trend Observations
6.1 Matrix extensions connect from simulator to backend in one line
Four layers of activity appeared in the same window. The software sources picked up QEMU’s AME package, the LLVM backend merged XSAI’s ZAMES, Buddy switched to the ZTT v0.6 dialect, and the triton side opened RISC-V lowering tuning.
This is the first time in a single window that the simulator, compiler, dialect, and operator backend have all moved together. AME and XSAI are still two separate naming schemes, and downstream users will have to pick a side based on their scenario.
6.2 Software sources are filling out both the P and matrix extension chains
QEMU’s RVP package fills in the simulator link. GCC’s three examples and LLVM’s 13 PRs fill in the compiler link, and the examples repo fills in the hands-on path.
Both chains are still at draft v0.6. The toolchain and simulator will change as the draft is updated, so current results can only be treated as interim records.
6.3 The distro side is migrating to Go and OpenSSL 4
Of openRuyi’s 84 commits in one day, 61 are Go dependencies, and another group wires grpc to system OpenSSL 4. The opensc update is likewise a build fix for OpenSSL 4.
These two types of work do not appear in feature lists, but they determine whether upper-layer software can compile on the new baseline. The distro side’s focus is shifting from adding packages to swapping out the foundation.
Appendix
Source Verification Table
| Source | Verification Result |
|---|---|
| RuyiAI-Stack (18 repos) | pytorch’s riscv and improve-riscv64-ci branches had zero commits; main sampled at the full 100-commit cap for upstream syncs; triton-riscv mainline had zero commits, nightly 20261119 released, new #127/#129/#130 opened; tilelang-riscv and the new repo both had zero commits |
| ruyisdk (29 repos) | packages-index 13 commits, wechat-articles 9 commits (biweekly 081 finalized), ruyisdk-website 13 commits, ruyi-imager 1 commit, linux-xuantie-kernel 1 commit; ruyi, support-matrix, board-docs, docs, ruyici had zero commits |
| buddy-compiler (8 repos) | buddy-mlir merged #1066 and #1067, opened #1068, #1069, #1070; #1065 closed without merging; the other 7 repos had zero commits |
| openRuyi-Project (packaging repo) | 84 commits, 36 SPECS packages; 61 commits for Go dependency packaging, grpc migrated to system OpenSSL 4, plus ffmpeg 9.0.2, postfix 3.11.7, openvpn 2.7.9 |
| RuyiAI official homepage | Hashes unchanged for the index, blog, and docs pages; 2 blog posts, 23 documentation chapters |
| ruyiai-daily | Issue 11-19 generated (11-19 06:24); issue 11-20 not yet out |
| openRuyi distro news page | No additions; monthly release remains 2026.09, with 2026.10 and 2026.11 unreleased |
| RuyiSDK forum | Latest topic still 11-07; no new posts in the window |
| Google News | Zero hits across five queries: RuyiSDK, Ruyi Community RISC-V, openRuyi, RuyiAI RISC-V, Buddy Compiler MLIR |
| arXiv | Latest RISC-V compilation and vector retrieval paper is 11-15; no new entries in the window |
Full Source List
- [1] Biweekly Report No. 081 original text — https://github.com/ruyisdk/wechat-articles/blob/main/20261119-ruyisdk-biweekly-81.md
- [2] Biweekly 081 preface update (#418) — https://github.com/ruyisdk/wechat-articles/pull/418
- [3] Biweekly 081 ruyipkg section (#416) — https://github.com/ruyisdk/wechat-articles/pull/416
- [4] Biweekly 081 documentation section (#417) — https://github.com/ruyisdk/wechat-articles/pull/417
- [5] qemu-user-riscv-rvp 11.1.0-ruyi.20261027 — https://github.com/ruyisdk/packages-index/commit/0e383ae
- [6] qemu-user-riscv-ame 11.1.0-ruyi.20261016 — https://github.com/ruyisdk/packages-index/commit/cd8a950
- [7] qemu-system-riscv-rvp 11.1.0-ruyi.20261027 — https://github.com/ruyisdk/packages-index/commit/92e820c
- [8] qemu-system-riscv-ame 11.1.0-ruyi.20261016 — https://github.com/ruyisdk/packages-index/commit/7417f17
- [9] packages-index commit list — https://github.com/ruyisdk/packages-index/commits/main
- [10] packages-index PR #242 VisionFive 2 image 26.11.0-trunk.120 — https://github.com/ruyisdk/packages-index/pull/242
- [11] packages-index PR #237 OpenBSD 8.0 — https://github.com/ruyisdk/packages-index/pull/237
- [12] packages-index PR #250 Bianbu K3 4.0.8 (under review) — https://github.com/ruyisdk/packages-index/pull/250
- [13] packages-index PR #249 remove old VisionFive 2 images (under review) — https://github.com/ruyisdk/packages-index/pull/249
- [14] ruyi release list — https://github.com/ruyisdk/ruyi/releases
- [15] ruyi-imager PR #25 flash-write close-window confirmation — https://github.com/ruyisdk/ruyi-imager/pull/25
- [16] ruyisdk-website commit f9de316 enable German — https://github.com/ruyisdk/ruyisdk-website/commit/f9de316
- [17] ruyisdk-website commit 8a55dcf submodule auto-update — https://github.com/ruyisdk/ruyisdk-website/commit/8a55dcf
- [18] linux-xuantie-kernel commit da94ef7 — https://github.com/ruyisdk/linux-xuantie-kernel/commit/da94ef7
- [19] support-matrix commit list — https://github.com/ruyisdk/support-matrix/commits/main
- [20] GCC mainline commit fb66e0c — https://github.com/gcc-mirror/gcc/commit/fb66e0c3e24743e415c50eff069e61a935134e92
- [21] GCC vector dot-product commit series — https://github.com/gcc-mirror/gcc/compare/fb66e0c3e24743e415c50eff069e61a935134e92...6b58ef48245e22dcb1edfa75137ff0b426c9299b
- [22] RVP-Example sample repository — https://github.com/pz9115/RVP-Example
- [23] LLVM PR #243813 out-of-bounds shift folding (under review) — https://github.com/llvm/llvm-project/pull/243813
- [24] LLVM PR #242773 P extension widening shift legalization — https://github.com/llvm/llvm-project/pull/242773
- [25] LLVM PR #241979 P extension narrowing truncation — https://github.com/llvm/llvm-project/pull/241979
- [26] V8 CL 8698757 trampoline pool short jumps — https://chromium-review.googlesource.com/c/8698757
- [27] V8 CL 8703878 inline IsShortInstruction in simulator — https://chromium-review.googlesource.com/c/8703878
- [28] OpenJDK PR #32635 Zicond conditional arithmetic — https://github.com/openjdk/jdk/pull/32635
- [29] OpenJDK PR #32727 XuanTie vendor feature detection — https://github.com/openjdk/jdk/pull/32727
- [30] OpenJDK PR #31955 redundant barriers for atomic instructions — https://github.com/openjdk/jdk/pull/31955
- [31] Go CL 864281 riscv64 exponential function optimization — https://go-review.googlesource.com/c/go/+/864281
- [32] Go CL 765200 AES zvkned — https://go-review.googlesource.com/c/go/+/765200
- [33] triton-riscv nightly 20261119 — https://github.com/RuyiAI-Stack/triton-riscv/releases/tag/nightly/20261119-85d3bae
- [34] triton-riscv PR #125 wheel strip — https://github.com/RuyiAI-Stack/triton-riscv/pull/125
- [35] triton-riscv PR #126 63 FlagGems operators — https://github.com/RuyiAI-Stack/triton-riscv/pull/126
- [36] triton-riscv PR #127 43 torch.special operators — https://github.com/RuyiAI-Stack/triton-riscv/pull/127
- [37] triton-riscv PR #130 40 activation and loss operators — https://github.com/RuyiAI-Stack/triton-riscv/pull/130
- [38] triton-riscv PR #129 RISC-V backend tuning — https://github.com/RuyiAI-Stack/triton-riscv/pull/129
- [39] triton-riscv PR #128 same-name close record — https://github.com/RuyiAI-Stack/triton-riscv/pull/128
- [40] triton-riscv PR #123 backend toolchain decoupling — https://github.com/RuyiAI-Stack/triton-riscv/pull/123
- [41] pytorch fork riscv branch commit list — https://github.com/RuyiAI-Stack/pytorch/commits/riscv
- [42] pytorch fork main branch commit list — https://github.com/RuyiAI-Stack/pytorch/commits/main
- [43] tilelang-riscv commit list — https://github.com/RuyiAI-Stack/tilelang-riscv/commits/main
- [44] buddy-mlir PR #1066 AME switch to ZTT v0.6 — https://github.com/buddy-compiler/buddy-mlir/pull/1066
- [45] buddy-mlir PR #1067 LLVM submodule update — https://github.com/buddy-compiler/buddy-mlir/pull/1067
- [46] buddy-mlir PR #1068 polynomial approximation math and avl loops — https://github.com/buddy-compiler/buddy-mlir/pull/1068
- [47] buddy-mlir PR #1069 transposed-B batched matmul fix — https://github.com/buddy-compiler/buddy-mlir/pull/1069
- [48] buddy-mlir PR #1070 EmbeddingGemma-300M on K3 — https://github.com/buddy-compiler/buddy-mlir/pull/1070
- [49] buddy-mlir PR #1065 transpose tiling (closed) — https://github.com/buddy-compiler/buddy-mlir/pull/1065
- [50] llvm-project PR #33 XSAI ZAMES and ZMASYNC — https://github.com/RuyiAI-Stack/llvm-project/pull/33
- [51] openRuyi commit list — https://github.com/openRuyi-Project/openRuyi/commits/main
- [52] openRuyi commit b4d00ef grpc 1.85.3 — https://github.com/openRuyi-Project/openRuyi/commit/b4d00ef
- [53] openRuyi commit 8cb3ece protobuf 36.2 — https://github.com/openRuyi-Project/openRuyi/commit/8cb3ece
- [54] openRuyi commit 6000ecc grub kernel-install — https://github.com/openRuyi-Project/openRuyi/commit/6000ecc
- [55] openRuyi distro news page — https://openruyi.cn/zh-Hans/news/
- [56] RuyiAI official homepage — https://www.ruyiai.org/
- [57] RuyiAI blog page — https://www.ruyiai.org/blog.html
- [58] RuyiAI docs page — https://www.ruyiai.org/docs.html
- [59] Official website repo commit list — https://github.com/RuyiAI-Stack/ruyiai-stack.github.io/commits/main
- [60] ruyiai-daily commit list — https://github.com/RuyiAI-Stack/ruyiai-daily/commits/main
- [61] RuyiSDK community latest topics — https://ruyisdk.cn/latest
- [62] arXiv cs.AR latest — https://arxiv.org/list/cs.AR/recent
Limitations
This issue’s window is 24 hours (11-19 04:30 to 11-20 04:30). The GitHub anonymous API used roughly 42 calls during collection; pytorch main’s commit count was sampled at the 100-commit cap and not classified commit by commit.
The biweekly report’s figures were measured and reported by the team itself and were not re-tested on riscv64 hardware for this issue. Reductions such as V8’s 3.5% and 27%, and GCC’s cycle savings, are recorded as stated in the biweekly report.
triton-riscv #129 depends on buddy-mlir #1068, which has not yet been merged. The 1–7 ULP accuracy figures in the text are recorded per the PR description; actual behavior awaits re-verification after #1068 is merged.
The QEMU AME and RVP packages correspond to draft v0.6. Instruction encodings and behavior may change with the draft; the 20261016 and 20261027 strings in package version numbers are build dates and do not indicate that the extensions are finalized.
openRuyi’s 84 commits are counted per the API response; the Go dependency packaging batch may extend into the next window. On the external news side, five Google News queries returned zero hits, though Chinese-language media may carry reports that are not indexed.