Research window: 2026-11-04 04:30 ~ 2026-11-05 04:30 (Beijing time, past 24 hours; directly contiguous with the previous issue’s cutoff at 11-04 04:30).

Sources: GitHub key repository and organization push verification (RuyiAI-Stack, ruyisdk, openRuyi-Project, buddy-compiler; repository-by-repository verification of in-window commits, PRs, Releases, and tags for key repositories, with branch-level atom fallback; upstream forks filtered by RISC-V / RVV / XuanTie keywords), ruyi package manager Releases, packages-index, openRuyi packaging repository SPECS and merged PRs, RuyiSDK developer community forum, ruyisdk WeChat article repository (biweekly progress draft), ruyiai-daily daily intelligence archive, three-page snapshot of the RuyiAI official homepage, openRuyi distribution news page, Google News RSS targeted Chinese and English queries (via proxy), arXiv search.

This Issue’s Index

  • Today’s Focus: AME backend pins i32 matrix accumulators to AccReg, configurable tile operations deferred until expansion (11-04)
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 llvm-project: riscv branch adds BOSCAME commit, supplementing register classification and tile scheduling (11-04)
    • 1.2 pytorch: riscv branch has zero commits, main syncs twenty upstream commits (11-05 status)
    • 1.3 triton-riscv: PR #98 proposes upgrade to Triton 3.9.0, running in parallel with the 3.8.0 rebase in #92 (11-04)
    • 1.4 tilelang-riscv: zero commits, two old PRs still open (11-05 status)
  • II. RuyiSDK Toolchain and Distributions
    • 2.1 openRuyi: ten commits in the window, mainly browser and Python package updates, CI switches signing method (11-04)
    • 2.2 ruyi main repo: zero commits, zero tags, 0.54.0-beta.20261102 remains the latest pre-release (11-05 status)
    • 2.3 packages-index: zero commits, ABI side-note list unchanged (11-05 status)
    • 2.4 openRuyi 2026.10 monthly release still missing, news page stuck at the September release (11-05 status)
  • III. Vector Extensions and Operator Optimization
    • 3.1 Three vector repos remain silent, latest commits stuck at 02-13, 06-25, 07-13 (11-05 status)
    • 3.2 Biweekly Progress 080 draft: glibc libmvec and newlib vector math gain functions (11-04)
    • 3.3 arXiv: bf16/fp16 silently degrade under auto-vectorization on RVV (11-02, outside window)
  • IV. Buddy Compiler Updates
    • 4.1 VIR’s set_vl semantics tightened: exact VLEN and arbitrary VLEN split into two paths (11-04 to 11-05)
    • 4.2 CI: pytorch 2.15 enters the test matrix, LLVM release enables NVPTX/AMDGPU and lld (11-04)
    • 4.3 Nightly build stuck at 0.1.4.dev20261103, no new releases in the window (11-05 status)
  • V. Community, Events, and Official Homepage
    • 5.1 Biweekly Progress 080 enters the repo: published 11-05, only the basic C library and V8 sections finalized (11-04)
    • 5.2 Official homepage’s three pages unchanged, two blog posts, twenty-three documentation chapters (11-05 status)
    • 5.3 No new forum posts in the window, summit post and issue 079 still on the homepage (11-05 status)
    • 5.4 ruyiai-daily 11-04 issue archived, RISC-V section covers K3 fixes and firmware patches (11-04)
    • 5.5 Four external news queries return zero hits, no new arXiv papers in the window (11-05 status)
  • VI. Trend Observations
    • 6.1 Backend moves first, frontend and runtime stand still
    • 6.2 Routine packaging maintenance coexists with release stagnation
    • 6.3 Vector math reinforcements appear in the draft first

Today’s Focus: AME backend pins i32 matrix accumulators to AccReg, configurable tile operations deferred until expansion (11-04)

Date: 2026-11-04　Source: commit f346011, PR #35

At 15:40 on 11-04 (Beijing time), the riscv branch of RuyiAI-Stack/llvm-project gained a new commit f346011, touching 8 files, +412/−1, authored by caihanyi111 with caohanghang as co-author. The change targets the LLVM backend implementation of the RISC-V AME matrix extension (recorded in code as BOSCAME), in two parts.

The first part is register classification. Loop-carried accumulator PHIs take their default register class from the type table, while nxv32i32 was previously registered only as TileReg; the accumulator operands and results of the accumulate instructions mlce32.m, mqma.*, and msce32.m are all confined to AccReg, so every loop edge had to generate a copy between the tile and accumulator register banks, and copyPhysReg could not expand such cross-bank copies into a single AME instruction. The fix registers nxv32i32 as AccReg as well; for targets that also enable the V extension, RVV re-registers the type, leaving default behavior unchanged. Same-bank whole-register copies are now expanded using mmve64.t.t and mmve64.a.a; cross-bank copies remain a hard error and are no longer downgraded to per-lane indexed moves.

The second part is scheduling. Each tile configuration and its consumers are kept in an ordered, side-effecting pseudo-instruction until expansion just before emit; matrix registers become allocatable, MMA accumulators are tied, zero-accumulator definitions do not read old values, and branch relaxation accounts for the fully expanded size. The commit also adds three tests: ordering of configurable load/MMA/store, standalone accumulators and zero-value definitions, and both RVV and non-RVV selections, plus the relaxed branches.

The commit message states the motivation: the upper-level TTIR-to-BOSCAME lowering requires reduction loops to keep i32 accumulators in AccReg throughout. Per the context in the commit message, this chain corresponds to the matrix operator path on the TileLang side; what the LLVM side fills in is the register allocation and scheduling layer, not new instructions.

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

1.1 llvm-project: New BOSCAME Commit on the riscv Branch, Adding Register Classification and Tile Scheduling (11-04)

Date: 2026-11-04　Source: commit f346011, riscv branch commit list

Beyond the two changes covered in today’s highlights, the test surface of this commit deserves separate mention: xboscame-ordered-tiles.ll covers the ordering and machine verification of configurable load/MMA/store, xboscame-matrix-copy.ll covers same-heap copy expansion, and xboscame-matrix-copy-cross.ll pins down the behavior that cross-heap copies must error out. The verification criteria stated in the commit message are that all RUN lines in xboscame-ordered-tiles.ll pass, including machine verification and object file generation.

On the timeline, this is the first code commit on the riscv branch since the CI workflow change at 03:57 on 11-03; the earlier [RISCV] Lower IME intrinsics to SpacemiT A100 dot instructions and [RISCV] Add A100 HP/SP and pack/unpack lowering on the branch remain at 11-02. The main branch saw zero commits in the same period (latest at 14:48 UTC on 11-02).

1.2 pytorch: Zero Commits on the riscv Branch, Twenty Upstream Syncs on main (11-05 Status)

Date: 2026-11-05 (status)　Source: riscv branch commit list, main branch commit list, open PR list

Zero commits on the riscv branch during the window. The last full-branch replay push occurred at 03:00 on 11-04 (Beijing time), falling at the end of the previous window, with head commit 6c89022 ([SyncBots] Integrate PyTorch at 3afdbd3); this issue only records status, and the patch set is identical to the previous issue.

The main branch had twenty upstream sync commits during the window, concentrated in inductor/AOTI, cuDNN frontend, oneDNN for macOS, MPS, and dynamo; after checking each one, none involve RISC-V or RVV changes. The eight open PRs (#36, #38, #43, #46, #48, #51, #23, #35) remain open, with no merges seen during the window.

1.3 triton-riscv: PR #98 Proposes Bumping to Triton 3.9.0, Running in Parallel with the 3.8.0 Rebase in #92 (11-04)

Date: 2026-11-04　Source: PR #98, PR #92, commit list

Zero commits on the main branch, with the latest still the commit at 05:15 UTC on 10-31. The real new activity is on the PR side: trdthg opened #98 “Bump Triton to main (3.9.0)” at 17:25 on 11-04, last updated at 00:11 on 11-05; on the same day, jason-hue’s #92 “Rebase integration onto the official Triton v3.8.0 release” was also updated.

The two take different routes. #92 switches triton-hash.txt from the mainline dev pin 5889527 to the v3.8.0 release commit c01b6774; the PR description gives the gap: the original pin is 417 dev commits ahead of and 25 release commits behind the release; the rebase is on top of #55 (asin), test_inverse_trig_dtype.py passes 30/30, and CPU and RISC-V pipeline behavior remains unchanged. #98, by contrast, pushes the anchor directly to the 3.9.0 mainline. One favors stability, the other follows upstream; the final landing point depends on review trade-offs.

1.4 tilelang-riscv: Zero Commits, Two Older PRs Still Open (11-05 Status)

Date: 2026-11-05 (status)　Source: commit list, PR #4, PR #2

Zero commits on the main branch during the window, with the head commit stuck at 57c548c (the merge result of PR #5) at 06:13 UTC on 11-03. Two PRs remain open: #4 “Add RVV vector mask and max reduction lowering” (zxfj12138) and #2 “Add Codex-compatible RISC-V operator workflow” (zzwwji-lab). When the previous issue recorded the merge of PR #5, it noted that RVV masks and max reduction were not yet covered; in this issue that path is still stuck at #4, unmerged.

II. RuyiSDK Toolchain and Distributions

2.1 Ten commits in the openRuyi window: browser and Python package updates dominate, CI switches signing method (11-04)

Date: 2026-11-04　Source: commit list, firefox 159.0, expat 2.10.0, CI signing change

All ten commits landed on 11-04 and fall into three groups.

Package updates: firefox bumped to 159.0; expat bumped to 2.10.0 and tagged security-fixes; four Python packages updated the same day — python-paste 3.11.0, python-multidict 7.2.1, and python-hydra-core 1.3.8 all tagged security-fixes, python-aiobotocore bumped to 3.10.0, plus python-nltk 3.10.4.

CI: the update-package workflow was changed to run inside the openruyi container (21:43), reverted two hours later (23:48); then apt-key was replaced with a scope-restricted OBS signing key (23:50). PR #1591 (containerization) and #1584 (expat security update) were also merged the same day, along with #1544 updating firefox. No commit all day touched the RuyiAI trio, and no new AI-related spec was added.

2.2 ruyi main repo: zero commits, zero tags; 0.54.0-beta.20261102 remains the latest pre-release (11-05 status)

Date: 2026-11-05 (status)　Source: ruyi Release list, commit list

Zero commits in the main repo during the window, and no new tags. The latest Release entry is still 0.54.0-beta.20261102 (11-02 18:37 Beijing time); the latest stable release remains 0.53.0, published 09-24. 0.54.0 has gone from alpha on 09-18 to beta on 11-02, with no stable release yet.

2.3 packages-index: zero commits, ABI annotation list unchanged (11-05 status)

Date: 2026-11-05 (status)　Source: packages-index commit list

Zero commits during the window; the latest remains 11-02 13:26 (Beijing time). The ABI annotation coverage issue discussed last issue (only 3 of 327 package version description files carry annotations) is unchanged this period, with no new determination data on the index side.

2.4 openRuyi 2026.10 monthly release still missing; news page stuck at September release (11-05 status)

Date: 2026-11-05 (status)　Source: openRuyi news page

No new items on the news page during the window. The latest release-type announcement is still 2026.09; October had only two feature articles (RISC-V software ecosystem public infrastructure, Ruyi Firmware open UEFI firmware solution), with no monthly release announcement. The absence of 2026.10 carries over from last issue.

III. Vector Extensions and Operator Optimization

3.1 The three vector repos remain silent, with latest commits stuck at 02-13, 06-25, and 07-13 (11-05 status)

Date: 2026-11-05 (status)　Source: VecIntrinBench, IntrinTrans, FBGEMM

VecIntrinBench’s latest commit is still 02-13, IntrinTrans is stuck at 06-25, and FBGEMM is stuck at 07-13; all three had zero commits within the window. The vector intrinsic benchmarking and transpilation tooling side has seen no activity for several consecutive issues. The actual operator-related progress this issue lies in buddy-mlir’s VIR and the vector math functions covered in the issue 080 draft.

3.2 Biweekly progress 080 draft: glibc libmvec and newlib vector math function additions (11-04)

Date: 2026-11-04　Source: Issue 080 draft, article repo commit list

In the 11-04 update of the RuyiSDK WeChat article repo, the two “Base Components” sections of the issue 080 draft have been written. Base C library part: glibc ports sinpif, sqrt, sqrtf, and tanf into the existing libmvec framework; newlib ports a total of 14 functions—atanhf, atan2f, atanpif, atan2pif, exp2f, cosf, coshf, cospif, exp10f, expf, log2f, log2p1f, rsqrtf, and expm1f—into the existing vector math framework.

These functions are the foundation for low-precision inference and activation computation on RVV: when the vector math library lacks functions, operators can only fall back to scalar paths, and no amount of compiler-side RVV lowering can make up for it. The remaining sections of the draft (package manager, IDE, GCC, LLVM, OpenJDK, Go, QEMU, etc.) are still empty.

3.3 arXiv: bf16/fp16 silently degrade under auto-vectorization on RVV (11-02, outside window)

Date: 2026-11-02 (outside window, past three days)　Source: arXiv 2611.03917

There were no new RISC-V-related papers on arXiv within the window; one study from the past three days directly relevant to RVV vectorization is worth noting. Authors Vaishnavi Kouru and Robert Chun compare LLVM’s RISC-V auto-vectorization using multiply-add loops at four precisions—fp32, int8, bf16, and fp16: under the common baseline of -march=rv64gcv, int8 and fp32 vectorize cleanly, while bf16 and fp16 degrade into extensive scalar fallbacks, with fp16 even degrading into per-element software conversion calls unless zvfbfmin/zvfbfwma and zvfhmin are additionally enabled.

Dynamic instruction counts under QEMU user mode give the magnitude: the baseline bf16 and fp16 kernels execute 17.0× and 75.9× the instruction counts of the extension-enabled versions, respectively. The paper also notes that compiler diagnostics do report this degradation, but the wording does not distinguish between “solvable by adding a compiler flag” and “inherent conversion cost,” so users get no direct signal pointing to the fix. For the RuyiAI stack, this is two ends of the same problem as the vector math function additions in issue 080: one end lacks library functions, the other lacks visibility into feature flags.

IV. Buddy Compiler Updates

4.1 VIR’s set_vl semantics tightened: exact VLEN and arbitrary VLEN split into two paths (11-04 to 11-05)

Date: 2026-11-04 to 11-05　Source: PR #1028, merge commit dd57775, PR #1029, merge commit 61ba191

Two PRs from zhanghb97 landed in sequence within the window, clarifying the semantics of set_vl in the VIR dialect.

#1028 (merged 11-04 23:54) makes rvv-vlen denote the target’s exact VLEN: vscale = rvv-vlen / 64, a vector<[N]xT> holds N times vscale lanes, and splitting is done on full vectors. Previously rvv-vlen was treated as the minimum VLEN, with masks all true yet still left in the code—llc kept vid and vmslt, memory accesses carried masks, RoPE’s rotate_half became slides plus masked selects, and masks even overflowed in decode attention. Kernels on K3 built for a 1024-bit vector width are the typical case for this path.

#1029 (merged 11-05 01:32) handles the generic build without rvv-vlen: here the code must hold for any VLEN, strip-mined regions step by N times vscale at runtime, and vectors are sized to the minimum 128 bits; a set_vl with a result is rejected outright with “requires rvv-vlen, i.e. exact VLEN,” whereas this pass previously only reported an error. Fixed vectors remain as-is without rvv-vlen, sized to the host; declaring a minimum VLEN above 128 is not allowed.

Together, the two mean: semantics no longer rely on “masks happening to be all true” as a fallback—at compile time you must state whether this code is compiled for a specific VLEN or for any VLEN.

4.2 CI: pytorch 2.15 enters the test matrix, LLVM release enables NVPTX/AMDGPU and lld (11-04)

Date: 2026-11-04　Source: PR #1026, PR #1027, PR #1025

#1026 adds pytorch 2.15 to the test matrix, covering 2.11 through 2.15. The change was prompted by a real regression: under 2.15 the latest full check failed at tests/Python/AtenOps/test_resize_enlarge.py, and the failure occurred before the importer—AOT autograd’s functionalization now rejects enlarging resize_ on an already-functionalized tensor, and torch.compile fails on the same function too, so the graph never reaches ResizeOp. The fix replaces the test with the out-of-place aten.resize, which goes through the same ResizeOp handler and produces the same MLIR, passing across the full 2.11 to 2.15 matrix.

#1027 gives CI-released LLVM the NVPTX/AMDGPU targets and enables lld; #1025 bumps LLVM to bef335e2143c. All three are infrastructure, but the regression recorded in #1026 is worth noting: upstream tightening of functionalization shows up as “the graph never reaches the backend” rather than a backend error.

4.3 Nightly stuck at 0.1.4.dev20261103, no new release in the window (11-05 status)

Date: 2026-11-05 (status)　Source: buddy-mlir Release list

No new Release in the window; the latest is still Nightly 0.1.4.dev20261103 (11-03 18:23 Beijing time); the stable release is still BUDDY MLIR 0.1.3 from 10-31. The five commits merged today have not yet produced a new nightly package.

V. Community, Events, and Official Homepage

5.1 Biweekly Progress 080 Enters the Repository: Drafted 11-05, Only Basic C Library and V8 Sections Completed (11-04)

Date: 2026-11-04　Source: Issue 080 draft, commit list

The article repository had four commits on 11-04: draft creation at 09:56 (#393), LIBC addition at 16:01 (#394), and two updates to the Issue 080 file at 18:02 and 19:35 (#395, #396). The filename is 20261105-ruyisdk-biweekly-80.md, with the title reading “Issue 080 · November 05, 2026,” meaning publication is tomorrow.

Judging by the degree of completion, only two sections currently have content: Basic C Library (porting of glibc libmvec and newlib vector math, see 3.2) and V8 (6 RISC-V patches submitted and merged this issue, plus 6 more reviewed and merged, including “allow vlen=1024” and “preserve vector tail on scalar moves”). The Package Manager, IDE, Version Testing, packages-index, Development Board Matrix, Official Documentation, GCC, LLVM, OpenJDK, Go, and QEMU sections remain empty. On the forum, Issue 079 (10-26) is still the latest issue, and Issue 080 has not yet been posted.

5.2 Official Homepage: Three Pages Unchanged, Two Blog Posts, Twenty-Three Documentation Chapters (11-05 status)

Date: 2026-11-05 (status)　Source: RuyiAI official homepage, blog page, docs page

The snapshots of the three pages are consistent with the previous issue: index.html 7,246 bytes, blog.html 5,632 bytes, docs.html 10,843 bytes, all with status unchanged. The blog index still has two posts (Triton’s RISC-V adaptation, PyTorch RISC-V CI practices), and the documentation index still has twenty-three chapters.

5.3 No New Forum Posts Within the Window, Summit Posts and Issue 079 Still on the Front Page (11-05 status)

Date: 2026-11-05 (status)　Source: RuyiSDK community latest topics, Issue 079

Among the latest topics sorted by creation time, there are no new posts within the window. The latest on the front page is still the 11-03 “Instructions and Usage Guide for Installing and Upgrading ruyi via Script,” followed by three summit-related posts from 10-29 (RuyiAI appearing at the 6th RISC-V China Summit, the XiangShan Ruyi Ecosystem Conference, and a repost of Bao Yungang’s talk), and Issue 079 from 10-26.

5.4 ruyiai-daily 11-04 Issue Archived, RISC-V Section Covers K3 Fix and Firmware Patch (11-04)

Date: 2026-11-04　Source: ruyiai-daily 11-04

The archive commit occurred at 11-04 06:37 (Beijing time). This issue covers 11-03 05:07 to 11-04 06:00. The RISC-V section contains several items relevant to this stack: a per-token decoding error in llama.cpp’s IME2 flash-attention on the K3 Pico-ITX, whose root cause was that M[QLEN] initialized only the first element while the remaining heads used zero as the initial maximum; a reply within the window confirmed the fix was verified on K3 and generation and performance tests were run with Qwen2.5-7B Q4_K_M, with no significant performance regression observed; the U-Boot OpenSBI FULL FIT v3 patch series was proposed, completing the full FIT loading path under RAM SPL; Sail merged support for indirect CSR access via Smcsrind/Sscsrind; Spike fixed the issue where hip.VSTIP was not cleared when henvcfg.STCE is zero.

5.5 External News: Four Query Groups Return Zero Hits, No New arXiv Papers Within the Window (11-05 status)

Date: 2026-11-05 (status)　Source: Google News RSS, arXiv

The four query groups “Ruyi RISC-V,” “RuyiSDK,” “RuyiAI,” and “Ruyi Community RISC-V” all returned no results within the window; arXiv also had no new papers in the RISC-V compiler direction within the window. The only relevant item in the past three days is the 11-02 bf16/fp16 auto-vectorization study, already noted in 3.3.

VI. Trend Observations

6.1 Backend Moves First, Frontend and Runtime Stand Still

Date: 2026-11-05　Sources: llvm-project riscv branch, buddy-mlir commit list

All of the day’s incremental work landed in “compilation and lowering”: LLVM’s AME backend added register classification and tile scheduling, while buddy-mlir’s VIR tightened the VLEN semantics of set_vl. Meanwhile, all three frontend repos — pytorch, triton-riscv, and tilelang-riscv — had zero commits. Progress in the stack is concentrated on reliably lowering existing operators to the low level, rather than integrating new frontend capabilities.

6.2 Routine Packaging Maintenance Coexists with Stalled Releases

Date: 2026-11-05　Sources: openRuyi commit list, ruyi Release list

openRuyi had ten commits in a day, covering firefox, expat, and four Python packages, plus experiments with CI signing methods and workflow containerization (the containerization was reverted the same day); ruyi and packages-index had zero commits, and the 2026.10 monthly release is still missing. Routine distribution maintenance is moving forward; releases aimed at the AI stack are not. The previous issue summarized this state as “the bottleneck is in releasing, not in development,” and it still holds on the distribution side this issue.

6.3 Vector Math Reinforcement Appears First in Drafts

Date: 2026-11-05　Sources: Issue 080 draft, VecIntrinBench

The most substantive content in the Issue 080 draft is the porting of vector math functions for glibc libmvec and newlib, yet the corresponding benchmark and translation repos (VecIntrinBench, IntrinTrans) remain silent. This indicates that vectorization work on low-level libraries is progressing, but it does not appear as commits in the key RuyiAI-Stack repos — the RuyiSDK foundational component line (GCC, glibc, newlib, V8) and the key RuyiAI stack repos are not worked on by the same people or in the same repositories. It will be worth comparing the pace of the two sides next issue.

Source Verification Table

Source Verification Result
RuyiAI-Stack key repos pytorch: zero commits on the riscv branch, twenty upstream syncs on main, eight open PRs unmerged; triton-riscv: zero commits, PR #98 newly opened, #92 updated; tilelang-riscv: zero commits, PR #4 and #2 still open; llvm-project: one BOSCAME commit on the riscv branch, zero commits on main; no commits in the window for the remaining repos
ruyisdk organization ruyi, packages-index: zero commits and zero releases; wechat-articles: four commits (Issue 080 draft); ruyi-tuner, docs, support-matrix, and the website repo: no commits in the window
openRuyi-Project Ten commits: firefox 159.0, expat 2.10.0, four Python packages, two CI changes; merged PRs #1591, #1584, #1544; the 2026.10 monthly release is still missing
buddy-compiler buddy-mlir: five commits (two VIR, two CI, one LLVM upgrade); zero releases in the window
RuyiSDK forum Zero new posts; the latest is the ruyi script installation guide from 11-03; Issue 079 remains the latest biweekly progress
RuyiAI official homepage No changes on the index, blog, or docs pages; 2 blog posts, 23 documentation chapters
ruyiai-daily The 11-04 issue has been archived
Google News Zero hits across four groups of Chinese and English queries
arXiv No new papers in the window; one bf16/fp16 auto-vectorization study on 11-02

Complete Source List