Research window: 2026-11-18 04:30 ~ 2026-11-19 04:30 (Beijing time, past 24 hours; directly contiguous with the cutoff of the previous 11-18 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. Also reviewed: pending PR descriptions in triton-riscv and buddy-mlir, merge records in the llvm-project fork, and SPECS changes in openRuyi. On the community side: the RuyiSDK biweekly repo, the support matrix, the latest topics on the developer forum, and three-page snapshots of the official homepage. External leads came from the ruyiai-daily daily repo, the openRuyi distro news page, targeted Google News RSS queries in Chinese and English (via proxy), and arXiv searches.


Index

  • Today’s Highlights: FlagGems operator set adds another 63, triton-riscv example repo expands to 239 files (11-18)
  • Today’s Highlights: XSAI matrix and sync interfaces land in the fork, Buddy side simultaneously swaps AME implementation to ZTT v0.6 (11-18)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V porting)
    • 1.1 riscv branch silent for a third day, main continues pulling upstream traffic (11-18 status)
    • 1.2 triton-riscv: three operator example sets under review simultaneously, v3.8.0 rebase PR closed (11-18)
    • 1.3 Nightly wheels drop 668 MB of redundancy, strip and packaging paths aligned (11-18)
    • 1.4 llvm-project merges a batch of XSAI, RV32 explicitly unsupported (11-18)
    • 1.5 tilelang-riscv next commit still stuck at 11-16 (11-18 status)
  • II. RuyiSDK Toolchain and Distro
    • 2.1 Biweekly Report Issue 081 in assembly, V100 board docs and vector math porting drafted (11-18)
    • 2.2 openRuyi 36 commits in a day: curl and nginx security updates, openssl 4.1.0 with RISC-V assembly patches (11-18)
    • 2.3 RuyiSDK site and docs: LicheePi 4A test report, site spacing fix (11-18)
  • III. Vector Extensions and Operator Optimization
    • 3.1 FlagGems suite fills libdevice gaps with closed-form replacements, 6416 items pass (11-18)
    • 3.2 VecIntrinBench, IntrinTrans, and FBGEMM still silent (11-18 status)
  • IV. Buddy Compiler Updates
    • 4.1 LLVM submodule bumped to 542365c, PR opened to swap AME dialect to ZTT v0.6 (11-18)
    • 4.2 transpose tiling pass cuts transpose bucket from 199.6G cycles to 31.5G (11-18)
    • 4.3 Model sample queue warms up, BGE-M3 RVV chain updated again (11-18)
  • V. Community, Events, and Official Homepage
    • 5.1 Official homepage three pages unchanged, 2 blog posts, 23 doc chapters (11-19 status)
    • 5.2 Forum adds StarFive BMC mass-production thread, course thread gets one more reply (11-18)
    • 5.3 ruyiai-daily stuck at the 11-18 issue, zero external news hits (11-18 status)
  • VI. Trend Observations
    • 6.1 Operator validation sets are expanding rapidly
    • 6.2 Matrix extensions now have a three-layer stack, top to bottom
    • 6.3 Toolchain enters the pre-release wrap-up phase

Today’s Highlight: FlagGems operator set adds 63 more, triton-riscv example repo expands to 239 files (11-18)

Date: 2026-11-18 Source: PR #126. PR #120.

At 22:32 on 11-18, triton-riscv opened #126 “Add FlagGems operators (63 ops)”. It moves 63 aten operators into python/examples/flaggems/ as Triton examples, each paired with a test against the aten CPU reference implementation.

This branch builds on top of the still-open #120. Per the PR description, of the 24 commits and 239 files relative to main, 111 files belong to #120 and 128 are new here. Once #120 is merged, this PR will be rebased down to its own commits.

The operators fall into six groups by purpose. The first four are scalar elementwise, comparison and bit-shift, 12 ops; elementwise mid-range, view and factory, 14 ops; numeric, logical and ldexp, 11 ops; unsafe indexing, amp and quantization helpers, 9 ops. The last two are masked softmax, pdist and linear packing, 9 ops; foreach and fused optimizers, 8 ops. The last two groups include _foreach_binary with 81 wrappers and _foreach_unary with 60 entries.

This toolchain cannot lower some libdevice builtins, and buddy-opt rejects the calls outright. The author substituted closed-form equivalents and measured error empirically against fp64 and torch. tan goes through sin/cos, with relative error no more than 1.2e-7. log2 becomes log times 1/ln2, no more than 3.2e-7. lgamma uses Lanczos g=7 plus the reflection formula, no more than 6.4e-4. sinh, cosh and tanh go through exponential forms, with error between 2.4e-5 and 6.3e-5.

erfc gets the most careful treatment. In fp32, computing 1-erf directly is identically zero for x greater than or equal to 3.5, so below 2.85 it uses 1-erf and above that switches to a 12-term asymptotic series; worst case at the seam is 6.4e-4, and above it no more than 1.3e-4. The four boundary values plus infinity and NaN match torch exactly.

Semantics follow torch, and wherever upstream FlagGems diverges, each case is pinned down with a test. _fused_adam applies weight decay in L2 form (g plus wd times p), and amsgrad takes the max of the uncorrected v. _fused_adamw_ does decoupled decay first, then the step. _fused_adagrad_ and _fused_sgd_ skip the update only when found_inf is exactly 1.

Validation records the full FlagGems suite passing 6416 cases, run locally on x86_64, excluding three pre-existing native-crash files. Tolerances leave 1.6x to 10x headroom over the measured worst-case error.

Today’s Focus: XSAI Matrix and Sync Interfaces Land in Fork, Buddy Side Swaps AME Implementation for ZTT v0.6 (11-18)

Date: 2026-11-18 Sources: llvm-project PR #33. buddy-mlir PR #1066. buddy-mlir PR #1064.

At 02:05 on 11-19 (Beijing time), a merge commit 163d6a7 appeared on the riscv branch of the llvm-project fork. It corresponds to PR #33 “Add XSAI ZAMES and ZMASYNC vendor extensions”, a 5150-line patch covering matrix and sync builtins, target extension types, register classes, instruction encoding, and assembly/disassembly.

The implementation uses RV64 SelectionDAG, including config materialization before register allocation and expansion after allocation; it tracks matrix memory operands, validates builtin immediate ranges, and emits clear diagnostics during RV32 code generation. Matrix tile types live in the riscv.xsai.zames.tile.* namespace, and sync values in riscv.xsai.zmasync.sync.

The author lists three scope gaps: matrix copy, spill and reload, and cross-function ABI. Vector exp2, MXFP, and XX8 conversion extensions are explicitly out of scope for this change. Validation consists of 28 dedicated tests passing, plus 11 smoke scenarios.

At 23:14 the same day, buddy-mlir merged #1064 “Bump LLVM to 542365c9cd38”, which only changes the llvm submodule pointer. The LLVM update being followed is exactly the XAME/ZTT v0.6 backend in the fork, i.e. the merge recorded in the previous daily briefing.

Then at 23:36, #1066 “Replace legacy AME with ZTT v0.6 operations” was opened: 30 files, +8349 −1277. It replaces the retired AME dialect, defines 138 typed instruction operations under the existing ame namespace, and unifies lowering of both interfaces to llvm.riscv.ztt.*.

#1066 ships with two sets of test cases and two freestanding gem5 validation workloads. The author states no commitment to numerical compatibility with the old ISA contract. A 4×4 signed integer workload passes exact comparison on AME_gem5: matrix A set to 1 through 16, B all ones, C initialized to zero; all 16 results match an independent scalar reference, the printed row sums are 10, 26, 42, 58, and both guest processes exit with code 0.

The three threads point in the same direction: LLVM provides backend capabilities, the compiler middle end connects the dialect and lowering paths, and the Triton side validates with operator examples.


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

1.1 riscv Branch Quiet for a Third Day, main Continues to Pull Upstream Traffic (11-18 Status)

Date: 2026-11-18 Source: riscv branch commit list. main branch commit list.

The latest commit on the riscv branch of RuyiAI-Stack/pytorch is still #57 “Drop JENKINS_RC”, committed at 11-16 03:00 (Beijing time). There has been no new branch rebase or new porting patch in these three days.

The repo’s main serves only as an upstream mirror and continued to receive PyTorch upstream merges during the window, with the latest landing at 11-18 18:15 UTC (11-19 02:15 Beijing time). The content is mostly routine fixes to inductor, aotautograd, export, and xpu, unrelated to RISC-V, and is not covered separately.

Compared with the “no rebase for two consecutive nights” noted in the 11-18 issue, this is now the third day. To judge branch status, look directly at the commit time of riscv; do not mistake main traffic for porting progress.

1.2 triton-riscv: Three Sets of Operator Examples Under Review Simultaneously, v3.8.0 Rebase PR Closed (11-18)

Date: 2026-11-18 Source: PR #120. PR #126. PR #115. PR #92.

Operator examples are now split across three places: #120 is the six batches of 54 operators opened on 11-13, with 111 files, still updated as of 11-18 09:51; #126 is 63 newly added operators, showing 24 commits and 239 files against main; the two file sets do not overlap.

#115 is a Renovate dependency update bumping buddy-mlir from nightly/v0.1.4.dev20261103 to dev20261113, touching a single file, +2 −2. It was opened on 11-12, updated on 11-18 18:44, and remains unmerged; the pinned version recorded in the previous issue is still dev20261103.

At the same 21:49 update timestamp, #92 “Rebase integration onto the official Triton v3.8.0 release” is in a closed state with no merge record. This PR was opened on 10-31; its approach was to rebase the integration onto Triton’s v3.8.0 release commit and reorder eight patches, changing 16 files, +437 −138. The PR page does not state the reason for closure.

1.3 Nightly Wheel Sheds 668 MB of Redundancy, strip Aligned with Packaging Path (11-18)

Date: 2026-11-18 Source: PR #125. Nightly 20261118.

On 11-18 18:43, #125 “Strip the copies the wheel actually packages” was merged, three files, +37 −31. Previously, the strip step in release.sh operated on artifacts in the source tree, while the wheel packaged a different copy in the build directory, so three large items were missed.

Per the numbers given in the PR: triton-shared-opt is 398 MB in the wheel, about 118 MB after stripping. libtriton.so is 441.5 MB, about 144 MB after stripping. libtriton_amd_codegen.so exists in both amd/lib and amd/codegen/build, at 84 MB each. Together that is 668 MB of avoidable members in a 502 MB wheel.

The fix has four steps: strip the copies that are actually packaged, remove the duplicate cmake build tree, fail outright rather than silently skip when strip does not take effect, and check the final size before release. The PR also has the Release Wheels workflow print the final size on every PR without publishing.

The previous issue recorded an earlier attempt where strip-all cut wheel weight by forty percent; this change fills in the missed copies. The most recent nightly before the merge is nightly/20261118-e9c34b7 (released 11-18 09:59 Beijing time), whose size has not yet been rebuilt under the new process; the next nightly build will be the first under the new accounting.

1.4 llvm-project Merges XSAI Batch, RV32 Explicitly Unsupported (11-18)

Date: 2026-11-18 Source: PR #33. Merge commit 163d6a7. PR #28.

#33 was submitted by xyyy1420, opened on 10-26, with review spanning more than three weeks. The patch went through several rounds of renaming: first unifying the native interface prefix to XSAI, then splitting the type namespaces for matrix and synchronization, with no legacy aliases retained.

Architecture feature names, assembly spellings, and instruction encodings remained unchanged throughout. The author states this is a pure interface and implementation change with no change to numerical behavior, and that representative integer, floating-point, and synchronization codegen text is byte-for-byte identical.

The XBOSCAME line in the same fork is still open (#28, opened 09-23); together with XAME, merged on 11-17, three vendor extensions now coexist in the same fork without overriding one another.

1.5 tilelang-riscv Next Commit Still Stuck at 11-16 (11-18 Status)

Date: 2026-11-18 Source: Commit list.

Zero commits during the window; the last is still merge commit #6 from 11-16 14:56. Neither #4 (RVV masking and max reduction) nor #2 (Codex-compatible operator workflow) under review saw any new activity.

II. RuyiSDK Toolchain and Distributions

2.1 Biweekly Report Issue 081 in Assembly, V100 Board Documentation and Vector Math Ports Drafted (11-18)

Date: 2026-11-18 Source: Biweekly repo PR #406. Biweekly Report 081 original text. board-docs PR #58. support-matrix commit 5dd057a4.

Biweekly Report Issue 081 was initialized at 10:23 on 11-18, with the title date reading 2026-11-19, i.e., published today. Four commits within the window are filling in its content, the latest at 01:28 on 11-19.

Four categories of entries are now drafted. On the IDE side, useless buttons on the news page were removed and the l10n tools documentation was updated. The support matrix adds Chinese and English test reports for RevyOS 20260504 on the Sipeed Lichee Pi 4A (PR #413), and fixes the CI build dependency installation issue (PR #414). The board examples repo adds Chinese and English board documentation plus HelloWorld and CoreMark examples for the SpacemiT V100-C2201 (PR #58). The same batch also opens entries for Chapters 16 through 19 of the ROS 2 course for the K3 CoM260 Kit.

The base C libraries saw two sets of changes. glibc ports tanh, tanhf, tanpi, and tanpif to the existing libmvec framework. newlib ports a batch of functions including hypotf, lgammaf, log10f, log1pf, powf, sincosf, and sinf to the vector math framework, with sinhf, sqrtf, tanf, tanhf, and tanpif following in the same batch.

The V100 documentation is the only board entry this issue with hands-on testing. The PR description states that both toolchains built and ran HelloWorld on the V100-C2201, and both CoreMark runs passed CRC verification. The Chinese and English examples use the same commands and sample output.

2.2 openRuyi 36 Commits in a Day: curl and nginx Security Updates, openssl 4.1.0 with RISC-V Assembly Patches (11-18)

Date: 2026-11-18 Source: Packaging repo commit list. nginx 1.31.9. curl 8.23.0. openssl 4.1.0.

The packaging repo saw 36 commits in the window, mainly security updates and dependency follow-ups. nginx was bumped to 1.31.9, bringing in 10 CVE fixes; curl was bumped to 8.23.0, bringing in 20 CVE fixes. Both spec changes are small: +3 −2 for the former, +6 −13 for the latter.

The openssl bump to 4.1.0 is the largest change in this batch and directly touches the RISC-V portion. The change removes the old alignment-crash workaround and clang format patch, replacing them with a RISC-V ChaCha20 dot-product assembly implementation. It also adds three GHASH multi-block aggregation patches, including the Zvbc and Zvkg vector crypto extension paths.

The rest are routine follow-ups: tpm2-tss 4.2.2, libinput 1.32.0, lighttpd 1.4.85, libp11 0.4.21, acpica 20260930. Two batches of new Go and yunion-related packages were also added. There are two small RISC-V-specific fixes. libspng enables the generated %check on riscv64, removing 5 lines of skip logic. bash-completion drops %{?_isa} since it is itself noarch.

2.3 RuyiSDK Site and Docs: LicheePi 4A Test Report, Site Spacing Fixes (11-18)

Date: 2026-11-18 Source: site commit 5dd057a4. Site commit e1356f02.

The support matrix’s LicheePi 4A test report was updated to RevyOS 20260504, with both Chinese and English READMEs synchronized and a desktop screenshot attached; the same repo also has a CI change switching build dependencies to direct apt installation.

ruyisdk-website saw over ten commits in the window, all layout fixes: positioning of the back-to-top button, footer copyright occlusion, spacing alignment of docs and standalone pages with the homepage, glass fade on the mobile sidebar, and mount position of the download modal. No functional or data changes.

ruyi and packages-index had zero commits in the window, with 0.54.0 (released 11-10) still the latest stable release. ruyi-imager likewise had no direct commits.

III. Vector Extensions and Operator Optimization

3.1 FlagGems Suite Fills libdevice Gaps with Closed-Form Substitutions, 6416 Items Pass (11-18)

Date: 2026-11-18 Source: PR #126.

The value of #126 goes beyond operator count. It lists, one by one, the libdevice builtins this toolchain currently cannot lower, and provides testable substitute expressions—effectively a catalog of backend capability gaps.

The rejection points mentioned in the list are concentrated in buddy-opt. pow can only go through the libdevice version; ** in the AST is unsupported. round is changed to explicit half-to-even. erfc is handled piecewise due to insufficient fp32 precision. Each substitute expression comes with relative error and boundary behavior, making later comparison easier when swapping back to native builtins.

The testing methodology is also worth noting: all tolerances leave 1.6x to 10x margin over the worst measured error, and seeds are process-independent. This approach makes regression results for new operators reproducible rather than dependent on luck.

3.2 VecIntrinBench, IntrinTrans, and FBGEMM Still Silent (11-18 Status)

Date: 2026-11-18 Source: VecIntrinBench. IntrinTrans. FBGEMM.

The three repos still have no commits: VecIntrinBench last active 02-13, IntrinTrans 06-25, FBGEMM 07-13. The vector intrinsic benchmark and conversion toolchain have been untouched for three quarters, so performance data can only be read indirectly from triton-riscv operator tests.

IV. Buddy Compiler Updates

4.1 LLVM submodule bumped to 542365c; PR opened to replace the AME dialect with ZTT v0.6 (11-18)

Date: 2026-11-18 Source: PR #1064. PR #1066.

#1064 changes only one line of gitlink: it moves the llvm submodule from f346011a to 542365c9, the latest commit on the riscv branch of the RuyiAI LLVM fork. This commit includes the XAME/ZTT v0.6 backend, target extension types, builtins, and two profiles.

#1066 then replaces the entire old AME implementation on the Buddy side. It defines 138 typed instruction ops, along with custom assembly syntax, builtin bridging, and matrix/accumulator target extension types; both interfaces are lowered to llvm.riscv.ztt.* via --lower-ame, and the retired AME ops and examples are removed.

The acceptance criteria match the backend merge on 11-17: two workloads — 4×4 signed integers and a parameterized kernel — are run on gem5 and pass exact comparison. The author states this is a functional check, with no performance testing and no claim of numerical-behavior compatibility with the old ISA.

The main-branch nightly is still at nightly/v0.1.4.dev20261113; no separate release was cut for this submodule update.

4.2 Transpose tiling pass cuts the transpose bucket from 199.6G cycles to 31.5G (11-18)

Date: 2026-11-18 Source: PR #1065.

#1065 adds the linalg-transpose-tile pass: it rewrites linalg.transpose on memrefs into scf.parallel loops tiled by tile-size × tile-size, changing only loop order, with bit-identical results.

The default lowering moves elements one by one along full-row strides on the swapped axis, repeatedly flushing the cache on large matrices. Measured on x86 with a graph of 240 materialized weight transposes, the author reports the transpose bucket dropping from 199.6G cycles to 31.5G, an 84% reduction.

The PR checklist is not fully ticked: the existing-tests and documentation items are empty, and the code-style item is unchecked. These need to be completed before merge.

4.3 Model example queue warms up slightly; BGE-M3 RVV chain updated again (11-18)

Date: 2026-11-18 Source: PR #934. PR #1061. PR #1059.

The older PR #934, “BGE-M3: RVV pass chain + A100 hybrid offload on K3”, was updated during the window; it combines the RVV pass chain with A100 hybrid offload on K3. It did not make the daily previously, so it is noted here.

#1061, #833, #945, #993, and #1059 all remain open with no merge activity. The front end is long and the middle end is waiting; the queue structure is unchanged from the previous issue.

V. Community, Events, and Official Website

5.1 No changes across the three official site pages; 2 blog posts, 23 doc chapters (11-19 status)

Date: 2026-11-19 Source: Official website. Blog page. Docs page.

The hashes of the three site pages match the previous round: index.html, blog.html, and docs.html are all unchanged. The blog still has 2 posts (Triton’s RISC-V port, PyTorch RISC-V CI practices), and the docs still have 23 chapters covering compilers, operator programming languages, and operator libraries.

In the website repo, the only change during the window was a contributor-file update, which is routine maintenance.

5.2 Forum gains a StarFive BMC mass-production post; course thread gets one more reply (11-18)

Date: 2026-11-18 Source: Latest community topics.

One new post appeared on 11-18, reposting StarFive’s official announcement of mass-production delivery of its RISC-V server BMC chip; no replies yet. The 11-16 xv6-riscv course Part 3 preview post got one more reply on 11-17, still about the file-system portion.

No other activity on remaining topics; biweekly report issue 080 is still sitting with no replies. The issue 081 forum post has not yet been published.

5.3 ruyiai-daily stops at the 11-18 issue; zero external news hits (11-18 status)

Date: 2026-11-18 Source: ruyiai-daily commit list.

The latest issue in the daily repo is 11-18; the 11-19 issue has not yet been generated. The earlier eight-issue gap has not recurred.

On the external news side, zero hits. Google News queries for RuyiAI, RuyiSDK, openRuyi, and 如意 RISC-V returned no relevant items during the window, and RISC-V topic news contained nothing about the Ruyi community. The openRuyi distro news page has no new entries; the latest are still the two from 10-19, and the 2026.11 monthly release has not shipped. arXiv searches for RISC-V AI compilation and vector optimization likewise returned no new entries.

VI. Trend Observations

6.1 The operator validation suite is expanding fast

Within a week, triton-riscv’s FlagGems examples grew from 54 operators to 117 across two batches, and the file count rose from 111 to 239. That pace shows the backend can now cover the backbone of common aten operators; the remaining work centers on precision patches and edge-case behavior.

6.2 Matrix extensions now have a three-layer stack, top to bottom

The backend is in LLVM, the dialect and lowering are in Buddy, and validation is in Triton and gem5. All three landed separately this week, with interface names unified on ZTT v0.6, while XSAI starts a separate native namespace. The two can coexist without conflict, but downstream users will have to pick a side.

6.3 The toolchain is in pre-release cleanup

Wheel slimming, dependency pinning, dialect replacement, board documentation — none of these are new features; they are debts that must be cleared before release. Biweekly report issue 081 putting V100 on-board measurements into the draft shows the documentation side is doing the same cleanup.


Appendix

Source Verification Table

Source Verification Result
RuyiAI-Stack (18 repos) triton-riscv merged #125, opened #126, closed #92; llvm-project merged #33 (5,150-line patch); pytorch’s riscv branch had zero commits, main continues pulling from upstream; tilelang-riscv had zero commits
ruyisdk (29 repos) wechat-articles four commits (biweekly 081), support-matrix two commits, board-docs one commit, site a dozen-plus commits; ruyi, packages-index, and ruyi-imager had zero commits
buddy-compiler (8 repos) buddy-mlir merged #1064, opened #1066 and #1065; #934 updated; #1059, #1061, #833, #945, #993 still open
openRuyi-Project (packaging repo) 36 commits; two security updates for nginx 1.31.9 and curl 8.23.0, openssl 4.1.0, libspng enables %check on riscv64; monthly release still not published
RuyiAI official homepage Hashes unchanged across index, blog, and docs pages; 2 blog posts, 23 documentation chapters
ruyiai-daily Latest issue is 11-18; the 11-19 issue has not yet been generated
openRuyi distro news page No new items; latest remains the two posts from 10-19; 2026.11 not released
Google News Zero hits within the query window for all four groups: RuyiAI, RuyiSDK, openRuyi, Ruyi RISC-V
arXiv No new results for RISC-V AI compilation and vector optimization searches
RuyiSDK forum One new repost thread on StarFive BMC mass production, plus one additional reply on a course thread

Full Source List

Limitations

This issue’s window is 24 hours (11-18 04:30 to 11-19 04:30). The GitHub anonymous API hit its rate limit during collection, so PR statuses and descriptions were cross-checked against pages and Atom feeds instead; some counts (such as the number of openRuyi PRs under review) could not be obtained and are not cited in the main text.

The 6,416 passing items in triton-riscv PR #126 are the author’s records from a local x86_64 machine and were not re-tested on riscv64 hardware in this issue. Both llvm-project’s XSAI and Buddy’s ZTT v0.6 are provisional-version extensions; instruction encodings and profiles may change in the future. This issue records only the state and verification results at the time of merge, and does not imply acceptance by upstream LLVM.

Biweekly Report Issue 081 had only a skeleton and partial entries within the window; the main text faithfully records the four commits as of 11-18, and the final published content is subject to the official version. On the external news side, all four Google News queries returned zero hits; it cannot be ruled out that Chinese-language media carried reports that were not indexed.