RuyiAI Daily Intelligence Report (2026-11-13)
Research window: 2026-11-12 04:30 ~ 2026-11-13 04:30 (Beijing time, past 24 hours; directly contiguous with the cutoff of the previous 11-12 report).
Sources: GitHub key repository and organization push verification (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler, with per-repository checks of commits, PRs, Releases, and branch head commits within the window; commit counts and times re-verified via API), ruyi-imager and ruyifirmware commit details, openRuyi packaging repository SPECS, post-by-post verification of the RuyiSDK developer community forum, snapshots of three pages of the official RuyiAI homepage, the openRuyi distribution news page, Google News RSS targeted queries in Chinese and English (via proxy), and arXiv searches.
This Issue’s Index
- Today’s Focus: buddy-mlir merges seven VIR commits in one day, bringing masks, bf16, reductions, and integer matrix multiplication into the vector path category by category (11-12)
- Today’s Focus: triton-riscv takes in the BOSCAME FPGA SPM path, plus an open PR cutting wheel size by 40% and switching explicit targets to -O3 (11-12)
- I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
- 1.1 pytorch riscv branch: CI adds two criteria—Jenkins return code and zero test cases passed (11-13)
- 1.2 triton-riscv nightly build merged before BOSCAME, description empty (11-12)
- 1.3 tilelang-riscv and llvm-project: zero commits in window (11-13 status)
- II. RuyiSDK Toolchain and Distributions
- 2.1 ruyi-imager: identifies local images via catalog metadata, one commit of nearly a thousand lines (11-13)
- 2.2 openRuyi: llvm23 bumped to 23.1.5, mainline and LTS kernels each get a patchset and config refresh (11-12)
- 2.3 openRuyi completes both ends of fwupd: gcab packaged, OVMF uses FMP capsule as cab payload (11-12)
- 2.4 ruyifirmware: fu740 recovery flow supports OpenOCD 0.12 (11-12)
- 2.5 ruyi and packages-index zero commits, stable release remains 0.54.0 (11-13 status)
- III. Vector Extension and Operator Optimization
- 3.1 i1 masks read/written byte-wise; previously 179 of 325 test files were blocked on this (11-12)
- 3.2 bf16 falls back to integer bit operations without Zvfbfmin, two paths compared input by input (11-12)
- 3.3 Reductions cover all kinds, int8 matrix multiplication emits vwmacc.vx (11-12)
- 3.4 Three test-surface numbers: 3980 up to 4006, 393 down to 0, 2 down to 0 (11-12)
- 3.5 Vector three repos still silent (11-13 status)
- IV. Buddy Compiler Updates
- 4.1 Nightly 0.1.4.dev20261112 contains only the day’s first commit (11-12)
- 4.2 Open PR #1059: memref copies switched to vector transfers (11-12 status)
- 4.3 Downstream pins nightly into dependencies, features string synced with #1054 (11-12)
- V. Community, Events, and Official Homepage
- 5.1 Forum has one routine hot post on 11-12, biweekly progress still at issue 080 (11-12)
- 5.2 Official homepage three pages unchanged, two blog posts, twenty-three documentation chapters (11-13 status)
- 5.3 ruyiai-daily missing for fifth day (11-13 status)
- 5.4 External news and arXiv: no additions in window (11-13 status)
- VI. Trend Observations
- 6.1 What was filled in this day is shape coverage, not performance
- 6.2 Numerical validation has reached the scale of exhaustive inputs
- 6.3 The release artifact itself has become an engineering metric
Today’s Highlight: Seven VIR Merges Land in buddy-mlir in One Day — Masks, bf16, Reductions, and Integer Matmul Each Wired into the Vector Path (11-12)
Date: 2026-11-12 Sources: PR #1052, PR #1053, PR #1054, PR #1055, PR #1056, PR #1057, PR #1058
From 08:12 to 14:44 (Beijing time) on 11-12, the buddy-mlir mainline merged seven consecutive middle-end changes, numbered #1052 through #1058, all authored by zhanghb97. What the seven share is that operator forms previously confined to scalar loops are now wired, category by category, into VIR’s vector lowering path: i1 masks, libm-only math functions, bf16, fast-math flags, all reduction kinds, rank-0 fill, and integer matmul. Each comes with lit tests and quantitative data, and most also include measurements on SG2044 and SpacemiT K3.
The first, #1052, addresses the threshold for memory access. When a memref element type is i1, scalable vectors were previously rejected outright (with the message a memref of i1 needs fixed vectors), because a memref occupies one byte per element while an i1 vector occupies one lane per bit — only fixed-length vectors have a per-lane lowering. Triton’s load/store masks are exactly i1 memrefs: on the pure VIR pipeline with scalable vectors, 179 of 325 Triton-RISCV test files were stuck on this error. The fix reads and writes by byte — contiguous accesses use llvm.load/store on i8 vectors, while non-contiguous strip-mined regions use vp.strided.load/store with all lanes enabled; it also sizes i1 lanes as one byte, where previously a region whose first dynamic vector was i1 was sized at 1 bit, yielding under VLEN 256 a scale like vector<[512]xi1> — sixteen register groups.
#1053 handles math functions LLVM refuses to lower on scalable vectors (erf, erfc, cbrt, powf, atan2, acosh, etc.) by expanding lane by lane: for vector<[N]xT>, extract each of the N times vscale lanes, call the scalar function, then insert. The trigger was that convert-math-to-libm treats scalable vectors like fixed-length ones, computing only the first N and silently zeroing the rest — a scalable erf under VLEN 128 (vscale 2) has half its lanes wrong. The effect shows directly in test coverage: on the Triton-RISCV FlagGems suite of 4045 items on SG2044 (VLEN 128), passing tests rose from 3980 to 4006, with the 26 fixed being test_erf and test_gelu, and no new failures.
#1054 is the most heavily validated of the batch. RVV only places bf16 in vector registers with Zvfbfmin; without that extension LLVM refuses to lower even a single load (Scalarization of scalable vectors is not supported). The change adds an rvv-mattr parameter to the lowerer, defaulting to assuming Zvfbfmin is present (Buddy’s several RVV targets — K3’s X100 and A100, and SG2044 — all have it); when -zvfbfmin is given explicitly, it instead treats values as integer bits, reading and writing bf16 vectors as i16 vectors, using integer operations for extf and truncf (truncf rounding to nearest even), and handling absf and negf via the sign bit. The two paths were compared input by input on both SG2044 and K3: truncf over all 2^32 f32 values, and extf, absf, negf, and extf-mulf-truncf over all 2^16 bf16 values, with bit-identical results.
#1055 fixes fast-math flags being dropped in strip-mined lowering: an arith.mulf with contract plus arith.addf was rebuilt as plain fmul and fadd, which LLVM will not fuse. Flags are now propagated with the same-kind operators to the vector loop and epilogue, so f32 multiply-add yields vfmacc and bf16 yields vfwmaccbf16 under +zvfbfwma; the author compared fmaf expansion bit for bit over 4M lanes on K3’s X100 and A100, with matching results. This is the only one of the seven that directly changes instruction selection.
#1056 extends vir.reduce from only floating-point add and maxnum to all vector.reduction kinds: add, mul, minsi, minui, maxsi, maxui, and, or, xor, minnumf, maxnumf, minimumf, maximumf, with integer support as well, and a verifier checking that kind and element type match. Previously Triton’s integer sums, min/max, products, and all/any (along with argmin, argmax, count_nonzero, bitonic sort, and radix sort that depend on them) stayed in loops. The change adds 570 lines across 9 files; the scalar epilogue uses vector::makeArithReduction’s combiner, and vector<[1]xT> takes yesterday’s #1051 LLVM intrinsic path.
#1057 and #1058 close out two remaining categories. #1057 lowers linalg.fill on rank-0 memrefs (mostly reduction accumulators) to a single memref.store, cutting residual linalg.fill in FlagGems’ 1811 kernels from 393 to 0. #1058 enables integer matmul lowering: when A and B are the same integer type and no wider than C, sign- or zero-extend before multiply-add (without fma), so on RVV int8 matmul yields vsext.vf2 plus vwmacc.vx, and unsigned yields vzext.vf2 plus vwmaccu.vx; the same PR also fixes an assertion failure in dynamic-shape matmul — M and K were originally taken inside the vir.set_vl region and erased while still in use by a later pass. After the change, none of the 1811 kernels contain linalg.matmul (previously 2, namely the int8 one).
Today’s Highlights: triton-riscv lands the BOSCAME FPGA SPM path, plus open PRs for a 40% wheel slim-down and explicit targets moving to -O3 (11-12)
Date: 2026-11-12 Sources: PR #104 BOSCAME FPGA, PR #116 wheel slimming, PR #117 explicit target -O3
#104 merged at 17:40 (Beijing time), wiring the BOSCAME FPGA SPM path into triton-riscv: AOT contract export, lit and integration tests, kernels under examples/BOSCAMEFPGA, launch and makefile, with entry documentation in docs/BoscameAOT.md and the example README. The first line of buddy-hash.txt pins the buddy-mlir commit that lands SPM, ensuring the example and compiler versions match. Author caihanyi111 opened this PR on 11-09; self-testing covered only formatting and lint, with the full build left to CI on both x64 and riscv64.
Two more open engineering changes landed the same day, both aimed at the release artifacts themselves. #116 notes that the LLVM and buddy tools bundled into the wheel previously used –strip-unneeded, leaving debug sections intact: a single triton-shared-opt carried nine .debug sections at 407 MB; switching to –strip-all brings it down to 118 MB, and the combined payload of the seven tools drops from about 970 MB to about 576 MB, roughly 60% of the original size. The author also records a rejected approach — pointing bin/llc directly at buddy-llc (which is LLVM llc plus buddy patches) — but buddy-llc does not exit after writing its output, so it cannot replace the plain llc invoked on the non-IME path.
#117 changes the optimization level on the cross-compilation path: when a RISC-V target triple is given explicitly (cross-compiling, or enabling the cross-compile switch on a RISC-V host), the build previously skipped the LLVM middle end, llc used its default optimization level, and loops that MLIR left in scalar form stayed scalar in the object file. The change runs opt -O3 and llc -O3 first for such builds, with the triple and features matching llc; the native RISC-V path without a triple (llc -O0, with a source comment noting miscompiles at higher levels) and the IME path (via buddy-llc) are untouched. The same PR also adds +zvfbfmin to the default llc features — otherwise scalable bf16 aborts outright in llc, while buddy-mlir’s #1054 assumes the extension is present by default, so the two sides must stay in sync.
I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Adaptation)
1.1 pytorch riscv branch: CI adds two criteria — Jenkins return code and zero-test-case pass (11-13)
Date: 2026-11-13 Source: PR #56
The riscv branch merged #56 at 00:25 (Beijing time), changing only one file, .github/workflows/ci-riscv64.yml (16 lines added, 7 deleted), by author yuzibo. The trigger was that in the community CI, PytorchCi’s 39th build showed “build succeeded, CI job failed, parent job still recorded as passed” — one such case being statistics where failed was 0 and pass was also 0.
The change adds two criteria: a non-zero Jenkins return code fails the step directly; and a statistics count with PASSED equal to 0 is also judged a failure — only PASSED greater than 0 and FAILED equal to 0 counts as a pass. For a branch that runs nightly packages every day, this is more worth noting than a feature change — if a green mark allows zero test cases to pass, regressions become invisible.
The previous batch of commits was the SyncBots integration at 11-10 23:23 (#51 local riscv64 image, #44 float and complex fixes for bytes_to_scalar, #50 masking torchinductor’s opinfo test cases, #42 ULP tolerance for linear cross entropy).
1.2 triton-riscv nightly released before BOSCAME merge, empty release notes (11-12)
Date: 2026-11-12 Source: nightly 2026-11-12-9ded1d4
The nightly was packaged at 09:35 (Beijing time), earlier than the merge time of #104 (17:40); the release notes are empty, and the attachments are, as usual, build artifacts for each platform.
The main branch gained one new commit in the window (#104); 19 PRs are open, of which three have changes: #115 pins the buddy-mlir nightly 0.1.4.dev20261112 into dependencies, and #116 and #117 are covered in today’s highlights. The rest, such as #109’s plugin rename (unifying triton_shared as riscv) and #103’s tutorial additions, remain in their original state.
1.3 tilelang-riscv and llvm-project: zero commits in the window (11-13 status)
Date: 2026-11-13 (status) Source: tilelang-riscv, llvm-project
tilelang-riscv’s head commit is stuck at 11-03 06:13 and llvm-project’s at 11-02 14:48 (UTC); there were no commits or releases in the window. The RuyiAI stack’s increments this period are concentrated in pytorch’s CI and triton-riscv’s one merge and two open changes.
II. RuyiSDK Toolchain and Distributions
2.1 ruyi-imager: Identifying local images via catalog metadata, a nearly thousand-line commit (11-13)
Date: 2026-11-13 Source: Commit 4ab1789
ruyi-imager received a commit at 02:22 (Beijing time) with 978 lines added and 78 deleted, authored by Glavo, titled “Identify and prepare local images using catalog metadata.” The changes span both GUI and SDK: MainWindow gains an entry point, Chinese and English i18n each get new strings, and on the SDK side new ImageCatalogService, LocalImageMatcher, LocalImageMatch, and DistfileIntegrity components work together with RuyiDistfileDownloader, plus a JavaFX smoke test.
Judging from the file composition, this matches existing local image files against catalog entries by metadata and performs integrity verification on downloaded artifacts—previously the flashing tool relied more on users selecting the right file themselves. The nightly build’s source commit from 11-12 02:28 (UTC) is caf379b, which predates this commit, so this build does not include the feature.
2.2 openRuyi: llvm23 bumped to 23.1.5; mainline and LTS kernels each get a patchset and config refresh (11-12)
Date: 2026-11-12 Source: openRuyi commit list
Five commits landed in the packaging repo window; the most substantial bumps llvm23 to 23.1.5 (author Su Yunqiang). This continues the 11-11 clang-wrap update to 20261111: the distribution’s LLVM mainline package and daily snapshot package each proceed independently, with the former advancing by upstream minor versions.
Two kernel maintenance commits follow: linux and linux-lts each get a patchset and config refresh, 2 lines changed each, authored by Gao Han. These wrap up the 11-11 moves that bumped mainline to 7.3.4 and LTS to 6.18.63, with version numbers unchanged.
2.3 openRuyi completes both ends of fwupd: gcab packaged, OVMF uses FMP capsule as cab payload (11-12)
Date: 2026-11-12 Source: gcab commit, OVMF patch commit
fwupd itself landed on 11-11; today’s commits supply the other two pieces it needs. First, a new gcab package (72-line spec, author Wang Xiang), the tool fwupd uses to handle cab-format firmware archives; second, a patch for ovmf, 2000-OvmfPkg-RiscVVirt-Use-FMP-Capsule-as-cab-payload-for.patch (42 lines added, 2 deleted), which makes the RiscVVirt platform use FMP capsule as the cab payload.
Together these three form a complete firmware update chain: device firmware is published as cab archives, fwupd parses and delivers them within the system, and the UEFI side completes the write via capsule. For RISC-V desktops and development boards, this chain was previously missing.
2.4 ruyifirmware: fu740 recovery flow supports OpenOCD 0.12 (11-12)
Date: 2026-11-12 Source: ruyifirmware commit
The ruyi branch of ruyifirmware received “fu740: Support recovery with OpenOCD 0.12” at 16:22 (Beijing time). The recovery flow is the trickiest part of the firmware repo to get right—after a bad flash, the board must be rescued via debugger, and any OpenOCD version change requires corresponding script and config updates. During the same period, the opensbi, edk2-ruyi, and coreboot repos saw coordinated push activity, but their branch heads remain at 11-11 with no new commits in the window.
2.5 ruyi and packages-index see zero commits; stable release remains 0.54.0 (11-13 status)
Date: 2026-11-13 (status) Source: ruyi releases, packages-index
Neither the ruyi package manager nor the package index repo saw any commits in the window; the stable release remains 0.54.0 from 11-10, and the biweekly update states the next version is due at the end of November. The ruyi-imager nightly build shipped on 11-12 with ten assets, covering linux riscv64, aarch64, and x86_64 architectures plus macOS and Windows.
III. Vector Extensions and Operator Optimization
3.1 i1 masks read/written byte-wise; 179 of 325 test files previously failed on this (11-12)
Date: 2026-11-12 Source: PR #1052
Scalable vector access for i1 previously only had per-lane lowering for fixed-length vectors, so load/store on memrefs with i1 elements was rejected outright—and Triton’s masks are exactly i1 memrefs. Running the pure VIR pipeline, 179 of 325 test files failed because of this—the widest-impact change among the seven in this batch.
Lowering is now done byte-wise: contiguous access uses llvm.load/store of i8 vectors; non-contiguous access falling in strip-mined regions uses vp.strided.load/store (mask fully enabled, length fully supplied). There is also a size fix: when the first dynamic vector in a region is i1, element width was previously counted as 1 bit, yielding vector<[512]xi1> at VLEN 256—sixteen register groups when read as bytes; i1 lanes now count as one byte each.
3.2 bf16 lowered to integer bit operations without Zvfbfmin, both paths compared input-by-input (11-12)
Date: 2026-11-12 Source: PR #1054
The bf16 path previously had only one possibility—”Zvfbfmin present”; without that extension LLVM does not even lower loads for scalable bf16. The change passes rvv-mattr as a parameter to the lowerer, defaulting to assuming Zvfbfmin (K3’s X100 and A100, and SG2044 all have it); with -zvfbfmin it switches to integer bits: bf16 vectors are read/written as i16, extf and truncf use integer operations, absf and negf flip the sign bit, and NaN takes the canonical value.
Verification is exhaustive input-by-input: truncf over all 2^32 f32 values; extf, absf, negf, and extf-mulf-truncf over all 2^16 bf16 values, bit-compared against the Zvfbfmin path on both SG2044 and K3 cores. If f8 types or (without Zvfbfmin) bf16 appear in other operators or in memrefs of non-integer memory spaces, it falls back to fixed-length vector lowering.
3.3 Reduction covers all kinds; int8 matmul emits vwmacc.vx (11-12)
Date: 2026-11-12 Source: PR #1056, PR #1058
vir.reduce previously recognized only floating-point add and maxnum; integer sum, min/max, product, and all/any all stayed in loops. It now accepts integers and floats and all kinds of vector.reduction; the verifier checks that kind matches element type, and scalar tails are finished element-by-element with a combiner. Integer and min/max reductions are exact; floating-point sum and product remain ordered except under reassoc.
Integer matmul is likewise filled in: when A and B share the same integer type and are no wider than C, extend first then multiply-accumulate; on RVV, int8 yields vsext.vf2 and vwmacc.vx (unsigned: vzext.vf2 and vwmaccu.vx). The same change fixes an assertion failure in dynamic-shape matmul—M and K are taken within the vir.set_vl region and erased during use by later passes, reporting “expected that op has no uses”.
3.4 Three test-surface numbers: 3980 to 4006, 393 to 0, 2 to 0 (11-12)
Date: 2026-11-12 Source: PR #1053, PR #1057, PR #1058
Each of the three changes provides a verifiable quantitative result, all measured on Triton-RISCV’s FlagGems suite: the math-functions change raised passing tests from 3980 to 4006 out of 4045 on SG2044 (VLEN 128), with the 26 fixed being test_erf and test_gelu; the rank-0 fill change reduced residual linalg.fill across 1811 kernels from 393 to 0; the integer matmul change reduced residual linalg.matmul from 2 to 0.
Taken together, the three numbers say one thing: VIR as the sole pipeline is no longer at the “can it run” stage, but at converging on “which operators still remain in linalg or scalar form”.
3.5 Three vector repos still silent (11-13 status)
Date: 2026-11-13 (status) Source: VecIntrinBench, IntrinTrans, FBGEMM
The head commits of the three vector-related repos remain at 02-13, 06-25, and 07-13—no activity for months. Vector-side investment continues to concentrate in the compiler: all seven VIR changes this issue determine what vector instructions are generated, while the measurement tool and transpiler repos have no corresponding updates.
IV. Buddy Compiler News
4.1 Nightly 0.1.4.dev20261112 contains only the day’s first change (11-12)
Date: 2026-11-12 Source: Nightly 0.1.4.dev20261112
The nightly was packaged at 10:33 (Beijing time) with empty release notes. Ordered by merge time, it only had time to include #1052 (08:12); the six changes from #1053 onward will land in the next package. The official release remains BUDDY MLIR 0.1.3 from 10-31.
4.2 Open PR #1059: copy memrefs with vectors instead of memrefCopy (11-12 status)
Date: 2026-11-12 (status) Source: PR #1059
#1059 is titled “Copy memrefs with vectors instead of memrefCopy”, was updated on the evening of 11-12, and remains open. Its direction matches the day’s seven changes: convert whole-block copies still done by memrefCopy into vector reads/writes, reducing the last few non-vector paths. An earlier #1049 (group set_vl vectors exceeding eight register groups into four-register groups) is also still under review.
4.3 Downstream pins the nightly into dependencies; features string synced with #1054 (11-12)
Date: 2026-11-12 Source: triton-riscv PR #115, PR #117
triton-riscv’s #115 writes buddy-mlir nightly 0.1.4.dev20261112 into dependencies, while #117 adds +zvfbfmin to the default llc features. The two are two ends of the same matter as upstream #1054: the compiler defaults to assuming the hardware has Zvfbfmin, so the runtime’s llc parameter string must keep up—otherwise scalable bf16 aborts directly at kernel compilation. This “default assumption—parameter sync” coupling is exactly what fork-type downstreams most easily miss.
V. Community, Events, and Official Website
5.1 Forum has one routine hot-topic post on 11-12; biweekly progress still at issue 080 (11-12)
Date: 2026-11-12 Source: RuyiSDK community latest topics, 11-12 hot-topic post
One new topic appeared in the forum window: the routine “open-source tech hot news” aggregation post on 11-12. Biweekly progress remains at issue 080 from 11-11; issue 081 has not appeared. The Ruyi Imager beta post received one reply on 11-11, part of ongoing Q&A. No new announcements on the community side this issue.
5.2 Three official homepage pages unchanged; two blog posts, twenty-three doc chapters (11-13 status)
Date: 2026-11-13 (status) Source: RuyiAI official homepage, Blog page
The hashes of the index, blog, and docs pages match the previous inspection; the blog still has two posts (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices), and the docs index has twenty-three chapters. There is no corresponding article on the site for this day’s batch of VIR changes.
5.3 ruyiai-daily gap reaches day five (11-13 status)
Date: 2026-11-13 (status) Source: ruyiai-daily commit list
The last commit in the daily intelligence repo remains 11-08 06:21 (Beijing time); five issues from 11-09 to 11-13 were not generated. The repo previously produced one issue per day; the gap has now lasted five days.
5.4 External news and arXiv: no additions in window (11-13 status)
Date: 2026-11-13 (status) Source: Google News, arXiv RISC-V search
Six groups of Chinese and English queries—RuyiSDK, RuyiAI, 如意 RISC-V, openRuyi, buddy-mlir, triton riscv—had zero hits in the window, returning only old pieces from months ago. The latest in arXiv’s RISC-V search remains the 11-09 MCU delegation-protection paper; no new papers in the window. The openRuyi distro news page still shows the two posts from 2026-10; the 2026.11 monthly release has not yet shipped.
VI. Trend Observations
6.1 What this day filled in was shape coverage, not performance
Of the seven patches, only #1055 (fast-math flags) directly changes instruction selection; the other six bring forms previously left in loops into the vector path: i1 masks, libm-only math functions, bf16, reduction kinds, rank-0 fill, and integer matrix multiplication. The right yardstick for this kind of change is “how much linalg or scalar path remains,” not benchmark scores—393 down to 0, 2 down to 0, and 179 failing files going to zero all use that yardstick.
Comparing this with yesterday’s #1050 strip-mined remainder narrowing shows the division of labor across the two days: first saturate the vector loops with lanes (performance), then let every operator enter the vector loop (coverage). Doing it in the reverse order would be wasted work.
6.2 Numerical validation has reached the scale of exhaustive inputs
#1054 ran all 2^32 f32 values through truncf and all 2^16 bf16 values through conversions and sign operations; #1055 compared the fmaf expansion bit by bit across 4M lanes. This level of validation cost shows that bf16 is now treated as a shipping path, not an “optimize when there’s time” item.
This echoes the 11-02 paper discussing LLVM’s silent degradation when auto-vectorizing bf16/fp16 on RISC-V—the same risk point: wrong results without an error. The validation approach in this batch of PRs is a direct response to this kind of silent error.
6.3 The release artifact itself has become an engineering metric
Three changes on the same day concern “what gets shipped”: the triton-riscv wheel shrank from about 970 MB to about 576 MB (triton-shared-opt alone from 407 MB to 118 MB), buddy-mlir nightlies are packaged per batch, and pytorch’s CI criterion blocks false greens from zero test cases passing. For RISC-V users, downloading a wheel costs far more than on x86, so the size and trustworthiness of release artifacts are becoming metrics on par with features.
Source Verification Table
| Source | Verification result |
|---|---|
| RuyiAI-Stack key repos | pytorch riscv branch: one merge (#56); triton-riscv: one merge (#104), one nightly, three open PRs changed within the window (#115, #116, #117); tilelang-riscv, llvm-project, FBGEMM, cpuinfo, VecIntrinBench, IntrinTrans, harness, torq-tile, pytorch-ci all had zero commits |
| ruyisdk org | ruyi: zero commits, zero releases (0.54.0 still latest); ruyi-imager: one local image recognition commit (978 lines) and the 11-12 nightly; packages-index: zero commits; ruyisdk-website, riscv-isa-sim, sail-riscv had push activity but no new commits within the window |
| openRuyi-Project | Packaging repo: five commits (llvm23 23.1.5, linux and linux-lts each with a patchset refresh, gcab, OVMF capsule patch); ruyifirmware: one commit (fu740 recovery flow); opensbi, edk2-ruyi, coreboot had pushes but branch heads remain at 11-11, with no new commits within the window |
| buddy-compiler | buddy-mlir: seven merges (#1052 to #1058) and nightly 0.1.4.dev20261112; two open PRs (#1059, #1049); official release still 0.1.3 |
| RuyiSDK Forum | One routine hot-topic post on 11-12; biweekly progress still issue 080, issue 081 not yet out |
| RuyiAI official homepage | index, blog, and docs pages unchanged; 2 blog posts, 23 documentation chapters |
| ruyiai-daily | Stalled at 11-08; five issues from 11-09 to 11-13 not generated |
| openRuyi distro news page | Latest still the two posts from 2026-10; the 2026.11 monthly release not published |
| Google News | Zero hits within the window across six targeted Chinese and English queries |
| arXiv | No new papers within the window; latest stalled at 11-09 |
Complete Source List
- [1] buddy-mlir PR #1052 Scalable vector access for i1 memref — https://github.com/buddy-compiler/buddy-mlir/pull/1052
- [2] buddy-mlir PR #1053 Per-lane expansion of libm-only math functions — https://github.com/buddy-compiler/buddy-mlir/pull/1053
- [3] buddy-mlir PR #1054 Scalable bf16 and Zvfbfmin — https://github.com/buddy-compiler/buddy-mlir/pull/1054
- [4] buddy-mlir PR #1055 Preserve fast-math flags in strip-mined regions — https://github.com/buddy-compiler/buddy-mlir/pull/1055
- [5] buddy-mlir PR #1056 All reduction kinds and integer support — https://github.com/buddy-compiler/buddy-mlir/pull/1056
- [6] buddy-mlir PR #1057 Lower rank-0 linalg.fill to store — https://github.com/buddy-compiler/buddy-mlir/pull/1057
- [7] buddy-mlir PR #1058 Integer matrix multiplication and dynamic dimension fixes — https://github.com/buddy-compiler/buddy-mlir/pull/1058
- [8] buddy-mlir PR #1059 Switch memref copy to vectors — https://github.com/buddy-compiler/buddy-mlir/pull/1059
- [9] buddy-mlir PR #1049 Group set_vl vectors in four-register groups — https://github.com/buddy-compiler/buddy-mlir/pull/1049
- [10] buddy-mlir nightly 0.1.4.dev20261112 — https://github.com/buddy-compiler/buddy-mlir/releases/tag/nightly%2Fv0.1.4.dev20261112
- [11] buddy-mlir commit list — https://github.com/buddy-compiler/buddy-mlir/commits/main
- [12] triton-riscv PR #104 SPM path for BOSCAME FPGA — https://github.com/RuyiAI-Stack/triton-riscv/pull/104
- [13] triton-riscv PR #115 Pin dependency to buddy-mlir nightly — https://github.com/RuyiAI-Stack/triton-riscv/pull/115
- [14] triton-riscv PR #116 Fully strip packaging tools — https://github.com/RuyiAI-Stack/triton-riscv/pull/116
- [15] triton-riscv PR #117 Explicit target -O3 and default +zvfbfmin — https://github.com/RuyiAI-Stack/triton-riscv/pull/117
- [16] triton-riscv nightly 2026-11-12-9ded1d4 — https://github.com/RuyiAI-Stack/triton-riscv/releases/tag/nightly%2F20261112-9ded1d4
- [17] triton-riscv open PR list — https://github.com/RuyiAI-Stack/triton-riscv/pulls
- [18] pytorch (RuyiAI fork) PR #56 riscv64 CI yaml — https://github.com/RuyiAI-Stack/pytorch/pull/56
- [19] pytorch riscv branch commit list — https://github.com/RuyiAI-Stack/pytorch/commits/riscv
- [20] tilelang-riscv commit list — https://github.com/RuyiAI-Stack/tilelang-riscv/commits/main
- [21] llvm-project (RuyiAI fork) commit list — https://github.com/RuyiAI-Stack/llvm-project/commits/main
- [22] ruyi-imager local image recognition commit — https://github.com/ruyisdk/ruyi-imager/commit/4ab1789a4637e347a413ab7b352fc3ba48c2bcd0
- [23] ruyi-imager nightly — https://github.com/ruyisdk/ruyi-imager/releases/tag/nightly
- [24] ruyi 0.54.0 release — https://github.com/ruyisdk/ruyi/releases/tag/0.54.0
- [25] packages-index commit list — https://github.com/ruyisdk/packages-index/commits/main
- [26] openRuyi packaging repo commit list — https://github.com/openRuyi-Project/openRuyi/commits/main
- [27] openRuyi llvm23 bumped to 23.1.5 — https://github.com/openRuyi-Project/openRuyi/commit/6399422ddf42581e63c83c09fa54d465d827aaca
- [28] openRuyi new gcab package — https://github.com/openRuyi-Project/openRuyi/commit/9135f9200fc4cfda87c8ecb264bdde731c1e937c
- [29] openRuyi OVMF FMP capsule patch — https://github.com/openRuyi-Project/openRuyi/commit/6d66f2ddc419caca63e9323d3cad27ab4fcff75f
- [30] openRuyi linux and linux-lts refresh — https://github.com/openRuyi-Project/openRuyi/commit/3053bd8781c74f81943cbaaea1e33cd856643594
- [31] ruyifirmware fu740 recovery flow — https://github.com/openRuyi-Project/ruyifirmware/commit/21d415d5d8a0e038006e873b831ad6ef7af9298b
- [32] openRuyi distro news page — https://openruyi.cn/zh-Hans/news/
- [33] RuyiSDK community latest topics — https://ruyisdk.cn/latest
- [34] Forum 11-12 open-source tech hot-topic post — https://ruyisdk.cn/t/topic/2907
- [35] RuyiSDK biweekly progress issue 080 — https://ruyisdk.cn/t/topic/2906
- [36] RuyiAI official homepage — https://www.ruyiai.org/
- [37] RuyiAI blog page — https://www.ruyiai.org/blog.html
- [38] ruyiai-daily commit list — https://github.com/RuyiAI-Stack/ruyiai-daily/commits/main
- [39] arXiv RISC-V search — https://arxiv.org/list/cs.AR/recent