TileLang Daily Intelligence Report (2026-11-10)
Research window: 2026-11-09 07:00 ~ 2026-11-10 07:00 (Beijing time, 24 hours; previous issue cutoff at 11-09 07:00).
Sources: Per-repo inspection of the GitHub tile-ai organization (30 repos, 4 pushed within the window) + ground-truth search of merged and in-review items (21 merged, 45 in-review updates, 6 new issues within the window) + per-commit inspection of main branches (tilelang 1, TileOPs 18, TileFoundry 1, tilelang-musa 1) + release page inspection (tilelang / TileOPs / TileFoundry / tilelang-musa, no new versions within the window) + two adopter repos + two sets of Google News queries in Chinese and English (via proxy) + Hacker News + arXiv. The 24-hour special inspection of the four RISC-V sources showed no changes.
Issue Index
-
Today’s Highlights: TileOPs merges 18 PRs in three batches of 6 each; 10 of the 11 performance items self-report as below target; TileFoundry reorders 1,491 files per source-code conventions; Ascend main branch has no merges for three days, with in-review work concentrated on toolchain additions
- I. Core Project Progress
- 1.1 Mainline: dense views restore precise logical regions, cross-row dependencies refined to the row (#3890) (11-09)
- 1.2 Mainline: 0.1.16 version bump PR in review, no new tag within the window (11-09)
- II. Multi-Backend Adaptation
- 2.1 Ascend: no merges on main branch, in-review work concentrated on torch.compile, L2 flush, and L0C dequantization (11-09)
- 2.2 Ascend: six chunk-series operators in review, catlass upgraded to v2.2.0 (11-09)
- 2.3 MetaX: docs push the SDK-matched official release to v0.1.12+musa.5, glibc threshold raised to 2.34 (11-09)
- 2.4 Hygon: no merges within the window, stalled at 11-06’s scaled cast and lane swap (11-09)
- 2.5 mlir-ascend: no merges, two harness-hardening PRs in review (11-09)
- III. Ecosystem and Adopters
- 3.1 TileOPs: 18 merges in three batches of 6, 10 performance items self-report as below target (11-09)
- 3.2 TileOPs: three boundary-input fixes — K=0 empty contraction, variable-length GQA hang, GLA backward shared memory (11-09)
- 3.3 TileOPs: Softmax keeps entire rows in registers, one exponential per element, instruction count down 21% to 23% (11-09)
- 3.4 TileOPs: RMSNorm narrow blocks request residency limits from the compiler, short-row warp activity stops declining (11-09)
- 3.5 TileFoundry: 1,491 files reorganized per source-code conventions, rules enforced in pre-commit (11-09)
- 3.6 Third party: Ascend TileLang operator repo ops-tilelang appears in the CANN community (11-09)
- IV. Community, Tutorials, and Events
- 4.1 News side: zero hits for both Chinese and English query sets within the window, 8 same-name noise items on HN (11-09 to 11-10)
- 4.2 New issues: float scalar parameter inlining loses 1 ulp, roofline upper bound coupled to CTA count (11-09)
- V. Trend Observations
Today’s Highlights
This window has only 21 merges, the fewest in nearly half a month, but the structure is clear: 18 in TileOPs, 1 in TileFoundry, 1 in the tilelang main repo, and 1 in tilelang-musa. TileOPs’ 18 merges fall into three batches by time, exactly 6 each, landing within 3 to 14 minutes (03:10 to 03:13, 05:41 to 05:55, 14:46 to 14:58), indicating that the overnight benchmark-to-merge queue is still running automatically. The cost is the pass rate: of the 11 performance PRs, 10 self-label Result: below bar, and only #2882’s BMM ping-pong is labeled bar met. The value of this batch of commits is not in the scores but in the clearly written root causes — register spills, TMA zero-fill, excessive dispatch counts — each of which the next author can pick up and fix.
The second change is structural. TileFoundry used a single #312 to reorder 1,491 files (+41324 to −42379), codifying five categories of rules — imports, exports, naming, mutable state, layering — as documents S1 to S11 and C1 to C11, and enforcing them in pre-commit via lint scripts, ruff, and import-linter. In the same window, the tilelang main repo saw the #3887 0.1.16 version bump PR. A major refactor and a version bump appearing together is usually a pre-release combination: rules land first, then the version follows.
The third item is that Ascend’s rhythm has broken. The latest merge on tilelang-ascend’s default branch ascendc_pto is still the seven commits from 11-06, with zero commits within the window; but there are 20-plus Ascend PRs in review across the main repo and tilelang-ascend, all updated in a burst on 11-09, with themes shifting from “adding operators” to “adding toolchain interfaces”: torch.compile support, L2 flush API, L0C accumulator dequantization, catlass upgraded to v2.2.0. Review bandwidth, not development volume, is the bottleneck on this line.
Sources: https://github.com/tile-ai/TileOPs/pull/2882
https://github.com/tile-ai/TileFoundry/pull/312
https://github.com/tile-ai/tilelang/pull/3887
I. Core Project Progress
1.1 Mainline: dense views restore precise logical regions, cross-row dependencies refined to the row (#3890)
After #3805, multi-element pointer accesses through explicit-stride views all fall back to the entire buffer, even when the view is dense row-major. The consequence is concrete: reading row 0 of a (2, 128) BF16 view is treated as reading both rows, creating a spurious V-to-MTE2 dependency on a subsequent DMA write to row 1. In TileKernels’ engram_gate_bwd, this erased the distinction between two slots of a BF16 view over FP32 storage, the precise-region assume_no_conflict hint failed to match, and synchronization stalled the next slot’s DMA behind the previous head’s vector computation.
#3890 first proves that the explicit stride indeed describes a dense row-major view, then reuses #3805’s offset-aware reconstruction (including cross-row carry, with a conservative fallback when it cannot be expressed), leaving physical lifetime and reuse analysis untouched and limiting the change to logical dependency precision. Indirect accesses remain conservative: gather/scatter indices and vld2 with offsets can reach beyond the compact access_ptr range, so regardless of whether the buffer is a dense view, an implicit layout, or one-dimensional, the entire logical buffer is retained in front of such consumers, preserving the V-to-MTE2 synchronization required before DMA covers indirectly reachable rows. New regression tests distinguish “overlapping slots” from “adjacent slots”.
Source: https://github.com/tile-ai/tilelang/pull/3890
1.2 Mainline: 0.1.16 version bump PR in review, no new tag within the window
#3887 ([Release] Bump version to 0.1.16) was updated on 11-09 11:07 and remains in review. None of the four repos has a new version within the window: tilelang is still v0.1.15 from 09-30, TileOPs is v0.1.0 from 10-27, TileFoundry is v0.0.4rc1 from 10-25, and tilelang-musa is v0.1.15+musa.1 from 09-30. If 0.1.16 lands this week, it will be the first feature release since 09-30 and worth watching in the next issue.
Source: https://github.com/tile-ai/tilelang/pull/3887
II. Multi-Backend Adaptation
2.1 Ascend: No merges on the main branch; torch.compile, L2 flush, and L0C dequantization queued for review
The default branch of tilelang-ascend is ascendc_pto, with the latest merges stopping at seven commits on 11-06 (#1959 bitwise workspace shape, #1953 fla_chunk_o, #1802 chunk_o_bwd, #1869 Sparse MLA forward SeeSaw, #1866 Chunk O forward, etc.), with zero commits during the window. The review queue is concentrated in the main tilelang repo: #3892 adds torch.compile support for AscendC and PTO; #3893 directly dequantizes the int32 accumulator during the L0C-to-GM copy, saving one transfer; #3891 uses aclrtDeviceL2CacheFlush to flush the L2 cache in benchmarks, replacing the previous workaround; #3894 fixes the PTO VMI descriptor and removes a duplicate logic coverage check.
Source: https://github.com/tile-ai/tilelang/pull/3892 , https://github.com/tile-ai/tilelang/pull/3893
2.2 Ascend: Six chunk-family operators under review; catlass upgraded to v2.2.0
The tilelang-ascend review queue saw a concentrated update on 11-09: #1952 fla_chunk_intra_token_parallel, #1994 fla_chunk_inter, #1799 Chunk Delta H, #1884 sparse_mla_topk_reducesum, #1917 indexer_topk_reducesum; additionally, #1989 upgrades catlass to v2.2.0, #2000 fixes the uint8 wrapping and src1 validation for PTO bitwise operations, and #1999 and #1998 enable automatic synchronization for copy test kernels. This batch continues the same FLA chunk-family migration as the chunk_o forward and backward merged on 11-06, indicating that operator coverage on the Ascend side is being filled in toward long-context attention.
Source: https://github.com/tile-ai/tilelang-ascend/pull/1989
https://github.com/tile-ai/tilelang-ascend/pull/1952
2.3 MetaX: Docs push the SDK companion release to v0.1.12+musa.5, raise the glibc threshold to 2.34
#11 only touches four documents (24 additions, 24 deletions), but it clarifies the channel relationships. First, the MUSA SDK companion release is upgraded from v0.1.12+musa.2 to v0.1.12+musa.5. Second, the C runtime requirement is raised from glibc 2.28 to 2.34, because the companion Moore Threads TVM FFI wheel requires it. Third, the MUSA SDK requirement changes from “recommended 5.2.0 and above, compatible with 4.3.8” to “matched with the GPU, driver, and torch_musa,” no longer specifying a fixed version number. The docs also emphasize that the SDK companion package and the open-source release are two separate channels and cannot substitute for each other: the open-source release page still uses v0.1.12+musa.1 and v0.1.13+musa.1 as references, while the latest repo tag is already 0.1.15+musa.1 (09-30). The three version numbers need to be checked separately; looking only at the major version is not enough.
Source: https://github.com/tile-ai/tilelang-musa/pull/11
2.4 Hygon: No merges during the window; stalled at the 11-06 scaled cast and lane swap
The main branch’s latest is still #33 (scaled cast) and #34 (HCU reduce lane swap layout derivation) from 11-06, with no commits, releases, or new issues during the window. The issue noted in the previous window—gfx928 being declared supported but the backend lacking a branch—has seen no new progress.
Source: https://github.com/tile-ai/tilelang-hygon/pull/34
2.5 mlir-ascend: No merges; two harness-hardening PRs under review
The default branch of mlir-ascend is latest at #209 from 11-04 (moving the BishengIR toolchain adapter into contrib), with no merges during the window. Two PRs under review were updated on 11-09: #219 hardens the execution and evaluation pipeline and adds knowledge-base and search@2 dependencies to the harness; #217 hardens knowledge recording and cooldown validation. These are engineering hardening efforts and do not involve frontend operator coverage.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/219
III. Ecosystem and Adopters
3.1 TileOPs: 18 merges, 6 in each of three batches, 10 performance items self-reported as below target bar
In the window, TileOPs merged 18 PRs, 86% of the organization’s 21: 11 performance, 6 bug and test, and 1 maintenance (#2881, trimming narrative comments left by the variable-length RoPE and target-cache fix). In time, three batches of 6 each, each batch landing within 3 to 14 minutes, matching the overnight pipeline from benchmark runs to the gap queue. All 11 performance PRs carry the three lines Op, Benchmark, Result; 10 of them write Result: below bar (target bar not met), and only #2882 (BMM ping-pong sizing bandwidth by tiles per CTA rather than M, N, so square 512 also goes ping-pong) writes Result: bar met. Meeting the bar is not a precondition for merging; stating the root cause and the A/B is—consistent with the previous issue.
Source: https://github.com/tile-ai/TileOPs/pull/2882 , https://github.com/tile-ai/TileOPs/pull/2881
3.2 TileOPs: three boundary-input fixes—K=0 empty contraction, variable-length GQA hang, GLA backward shared memory
The common shape of the three bug fixes is “does not occur under default parameters, only blows up on boundary inputs.” #2878: the K=0 empty contraction previously raised OpNotAvailableError on SM89, because the cp.async kernel’s narrow-row rule guards the innermost-dimension load in 4-byte units, while with K=0 the main loop never runs the K steps at all; after the fix, test_gemm.py -m "smoke or sm89" goes from 63 passed 6 failed to 66 passed 4 failed, the remaining four being SM90-only cases. #2876: the fused RoPE in variable-length GQA forward hangs when the rotary half-width is padded to a value not divisible by the thread count (rotary width 10, 5 chunks padded to 6, 128 threads); the root cause is that CI’s TileLang places the barrier inside the guard, so threads on the padding chunk skip it; a new split_reads puts “read plus rotate” and “write back” in two rounds, enabled only for the variable-length GQA kernel. #2880: float32 GLA backward at chunk 64 needs 114688 bytes of shared memory, while SM89 has only 101376; the fix assembles the dv state term after the dA product, at v’s last use, letting the decayed key take v’s place, bringing it down to 98304 bytes, with gradients bit-identical to the default order.
Source: https://github.com/tile-ai/TileOPs/pull/2878 , https://github.com/tile-ai/TileOPs/pull/2876 , https://github.com/tile-ai/TileOPs/pull/2880
3.3 TileOPs: Softmax keeps the whole row in registers, one exponential per element, instruction count down 21% to 23%
#2886 makes SoftmaxClusterSplitKernel also serve 16-bit rows of up to 8 shares of 16KB: each CTA keeps the share’s whole row in registers, 128 threads per CTA, one exponential per element. The original segmented program did two exponentials per element and routed half the shares through shared memory; 16-bit rows had 52% more instructions than quack, and 65536-element rows lagged by 3%. #2871 replaces the exponentials of the two register-resident softmaxes with a bare MUFU.EX2 plus one row reciprocal, adding a shared primitive exp_shifted_scaled that first multiplies by 2 to the 24th power, ensuring terms roundable to nonzero float32 remain normal numbers, with the reciprocal’s scaling factor canceling automatically. Executed instructions drop 23% at 512 columns and 21% at 2048; bf16 at 512 columns is 16.67 microseconds versus quack’s 16.35, and at 2048 columns 62.99 versus 62.21.
Source: https://github.com/tile-ai/TileOPs/pull/2886 , https://github.com/tile-ai/TileOPs/pull/2871
3.4 TileOPs: RMSNorm narrow block asks the compiler for the residency cap, short-row warp activity stops dropping
#2884’s paired RMSNorm 64-thread block (one row of 1024 16-bit elements, two vectors per thread) now requests the SM’s 32-block residency cap from the compiler; without the cap, ptxas gives 40 registers per thread, and one H200 SM holds only 24 of the 32-block cap, dropping warp activity on short-row loads to 64%. The wider paired block keeps its register configuration unchanged to avoid affecting large-row throughput.
Source: https://github.com/tile-ai/TileOPs/pull/2884
3.5 TileFoundry: 1491 files reorganized per source spec, rules enforced in pre-commit
#312 is the window’s only large change: 1491 files, +41324 to −42379, closing #296. The approach is to write down as rules what previously had no enforced constraint: a new docs/spec/code-organization.md states the S1–S11 and C1–C11 Python source rules and the layering table, with scripts/code_organization_lint.py, ruff, and import-linter enforced in pre-commit; the rules apply to every .py file, the only configuration being src and per-file exemptions, and __init__.py only re-exports, with no __all__. Structurally, ir/ is split into expr, types, dim, utils, core, pattern, clause, hir, tir; visitor_registry and evaluator are merged into seven pieces under interfaces—valuation, typeinfer, verify, cost, access, evaluator, schedule; the root compile.py moves into compiler/, script.py and module.py into parser/, and assembly logic into bootstrap.py; Go-style qualified names (types.TensorType) are adopted, and Options are split by phase after clang’s CompilerInvocation. The implication for downstream: TileFoundry’s internal import paths change in this version, and scripts referencing private symbols need to follow.
Source: https://github.com/tile-ai/TileFoundry/pull/312
3.6 Third party: Ascend TileLang operator repo ops-tilelang appears in the CANN community
ops-tilelang under the Atlas-Ascend-Fork organization (created 10-20, 88 public repos) describes itself as “the CANN community’s high-performance operator repo for Ascend NPU built on TileLang,” uniformly managing kernel implementations, automated test cases, standard workloads, and performance baselines, with the goal of making operators reusable, verifiable assets. The repo was created 10-20, had one push on 11-09 19:57, and has no stars yet. Its positioning differs from the official tilelang-ascend: the official repo delivers the compiler and operator migration, while this repo delivers operator assets and baselines—a signal of a third party treating TileLang as a delivery format, worth watching to see whether it diverges from the official repo.
Source: https://github.com/Atlas-Ascend-Fork/ops-tilelang
IV. Community, Tutorials, and Events
4.1 News side: zero hits in the window for both Chinese and English query sets, 8 same-name noise items on HN
Google News returned 29 items for Chinese “TileLang” and 25 for English “TileLang OR tile-ai”; by publication time, the latest hits remain the 10-28 Caifuhao article (TileLang, DeepEP, DeepGEMM, FlashMLA coordination between DeepSeek and Ascend 950) and the 10-21 SemiAnalysis piece (The Huawei-DeepSeek Symbiosis), with zero hits in the window. All 8 Hacker News hits are same-name noise (Emma real-time video persona, Vibe Slopped Cryptography, etc.), unrelated to TileLang. The latest TileLang-related item on arXiv is the 11-03 KV compression paper v2, also outside the window.
Source: one Chinese and one English Google News query set each (via proxy), Hacker News, arXiv; no relevant entries in the window.
4.2 New issues: float scalar parameter inlining loses 1 ulp, roofline upper bound coupled to CTA count
Six new issues were opened in the window. #3885 reports that float scalar parameters are inlined to 7 significant digits, 1 ulp off from float32—a silent precision loss. TileFoundry #313 argues the roofline upper bound should be independent of the mesh’s CTA count, or else explicitly state that it depends on it—an analyzer modeling-convention issue. There are also four analyzer gaps in TileFoundry #308 through #311 (split inputs of multi-wave grids priced only once, residency CTA and cluster pricing, writes when scheduled storage is reshaped, and four gaps in ping-pong batched GEMM), closed on 11-09 between 15:05 and 15:06.
Source: https://github.com/tile-ai/tilelang/issues/3885 , https://github.com/tile-ai/TileFoundry/issues/313
V. Trend Observations
First, capacity and pass rate in the performance queue have decoupled. Three batches of 6 each, 18 in a day — capacity is stable; but 10 marked below bar shows that the current pipeline’s goal is not “get faster this round” but “localize the gap to a fixable degree this round.” The root cause in #2887 is the most typical: the base tile splits B into 256-byte fragments read across rows, and DRAM streaming efficiency is lower than with fewer rows and longer runs, so it switches to 64-row blocks with a deeper loop. Even if such changes miss the score, they pin down the direction for the next step. To judge the maturity of a performance PR, readers should look at the Result line first, then check whether the root cause is reproducible.
Second, registers and dispatch have become the common bottleneck in this batch. #2884 is register quota capping the number of resident blocks (40 registers means the 32-block ceiling only holds 24 blocks), #2870 is 6272 128-thread blocks versus 1568 512-thread blocks — a fourfold difference in dispatch count, and #2885 is a 192-wide tile with higher fill rate but row-mask storage dividing by 192 overflows registers. All three point to the same thing: on SMs like the H200, operator performance is no longer determined by arithmetic volume but by the second-order parameter set of “how many registers per thread, how many blocks per SM, how many dispatches per block.”
Third, the bottleneck on the Ascend line has shifted from development volume to review bandwidth. The default branch has had no merges for three days, yet more than 20 PRs are stacked in review, and the topics have shifted from operator migration to toolchain interfaces (torch.compile, L2 flush, L0C dequantization). The dense-view fix in #3890 points the same way: it adds no new capability but makes the existing dependency analysis hold on more shapes. For users, improved usability of the Ascend path will first show up as “more shapes compile with correct dependencies,” not as a longer list of new operators.
Fourth, versions and structure are moving in the same window. tilelang’s 0.1.16 version-number PR, TileFoundry’s 1491-file reorganization, and the MetaX docs pushing the SDK companion package to musa.5 and raising the glibc threshold are all “pre-release preparation.” Downstream teams tracking versions need to check three places at once: the main repo tag, the backend repo’s musa.* or ascendc_pto branch, and build-time dependencies (glibc, apache-tvm-ffi). A previous window had the pitfall of “upgrading tilelang_musa without upgrading tvm-ffi”; this period’s docs write it up as an explicit rule.
Finally, the RISC-V line. The four-source patrol shows no change in 24 hours: the two kubuds- branches of kubuds/tilelang-riscv-kubuds, the MLIR route of RuyiAI-Stack/tilelang-riscv, the index page of TongmingLAIC/tilelang-riscv, and the SOPHGO branch of TongmingLAIC/tilelang-tpu all remain at their original references — this is now the third consecutive window (archived on 11-07, 11-09, 11-10) with no change. Contrasted with the GPU side’s “three batches a day” cadence, the RISC-V side is still on a manually driven weekly cadence; the gap is not in engineering effort but in whether there is a loop that automatically discovers gaps, automatically verifies, and automatically queues.
Appendix: Sources and Verification Notes
Source Verification Table
| Source | Verification Result |
|---|---|
| tile-ai organization (30 repos) | 4 repos pushed within the window: tilelang, TileOPs, TileFoundry, tilelang-musa |
| tilelang main repo | 1 merge to main (#3890); 24 open updates; 1 new issue (#3885) |
| TileOPs | 18 merges to main (11 performance, 6 bug and test, 1 maintenance); 4 open updates; no new issues |
| TileFoundry | 1 merge to main (#312, 1491 files); 5 new issues (#308 to #311 closed, #313 open) |
| tilelang-ascend | No merges on default branch ascendc_pto within the window (latest 11-06); 15 open updates |
| tilelang-hygon | No merges within the window (latest 11-06 #33, #34); no release, no new issues |
| tilelang-metax | No pushes within the window (main branch only syncs upstream, latest #3760) |
| tilelang-musa | 1 documentation change merged (#11); no new version on release page |
| tilelang-mlir-ascend | No merges within the window (latest 11-04 #209); 2 open updates |
| Release pages (tilelang / TileOPs / TileFoundry / musa) | No new versions within the window: v0.1.15 (09-30), v0.1.0 (10-27), v0.0.4rc1 (10-25), v0.1.15+musa.1 (09-30); 0.1.16 version-number PR under review |
| Adopters TileKernels / FlashQLA | No pushes within the window (10-14 / 10-30) |
| Third-party ops-tilelang | Repo created 10-20, pushed 11-09 19:57, no stars yet |
| TileLang×RISC-V four-source inspection | No change (all four upstream references match the previous snapshot, 24-hour window) |
| Google News (two queries, EN and ZH) / Hacker News / arXiv | Zero hits on the news side within the window (latest relevant items 10-28 and 10-21); all 8 HN items were same-name noise; latest relevant arXiv item 11-03 |
Full Source List
- [1] tilelang Ascend dense view restores exact logical region (#3890) — https://github.com/tile-ai/tilelang/pull/3890
- [2] tilelang 0.1.16 version number (#3887) — https://github.com/tile-ai/tilelang/pull/3887
- [3] tilelang Ascend torch.compile support (#3892) — https://github.com/tile-ai/tilelang/pull/3892
- [4] tilelang Ascend L0C accumulator dequantization (#3893) — https://github.com/tile-ai/tilelang/pull/3893
- [5] tilelang Ascend L2 flush API (#3891) — https://github.com/tile-ai/tilelang/pull/3891
- [6] tilelang Ascend PTO VMI descriptor fix (#3894) — https://github.com/tile-ai/tilelang/pull/3894
- [7] tilelang-ascend catlass v2.2.0 upgrade (#1989) — https://github.com/tile-ai/tilelang-ascend/pull/1989
- [8] tilelang-ascend fla_chunk_intra_token_parallel (#1952) — https://github.com/tile-ai/tilelang-ascend/pull/1952
- [9] tilelang-musa documentation advanced to v0.1.12+musa.5 (#11) — https://github.com/tile-ai/tilelang-musa/pull/11
- [10] tilelang-hygon HCU reduce lane swap (#34) — https://github.com/tile-ai/tilelang-hygon/pull/34
- [11] tilelang-mlir-ascend harness hardening (#219) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/219
- [12] TileOPs BMM ping-pong bandwidth (#2882) — https://github.com/tile-ai/TileOPs/pull/2882
- [13] TileOPs maintenance comment cleanup (#2881) — https://github.com/tile-ai/TileOPs/pull/2881
- [14] TileOPs K=0 empty contraction (#2878) — https://github.com/tile-ai/TileOPs/pull/2878
- [15] TileOPs variable-length GQA RoPE padding hang (#2876) — https://github.com/tile-ai/TileOPs/pull/2876
- [16] TileOPs float32 GLA backward shared memory (#2880) — https://github.com/tile-ai/TileOPs/pull/2880
- [17] TileOPs Softmax cluster full-row register residency (#2886) — https://github.com/tile-ai/TileOPs/pull/2886
- [18] TileOPs Softmax single MUFU.EX2 (#2871) — https://github.com/tile-ai/TileOPs/pull/2871
- [19] TileOPs RMSNorm register cap (#2884) — https://github.com/tile-ai/TileOPs/pull/2884
- [20] TileFoundry source-spec reorganization (#312) — https://github.com/tile-ai/TileFoundry/pull/312
- [21] ops-tilelang Ascend operator repo — https://github.com/Atlas-Ascend-Fork/ops-tilelang
- [22] tilelang floating-point scalar inline precision issue (#3885) — https://github.com/tile-ai/tilelang/issues/3885
- [23] TileFoundry roofline upper bound and CTA count issue (#313) — https://github.com/tile-ai/TileFoundry/issues/313
Verification Notes
This issue’s window is 24 hours (Tuesday). The GitHub organization-level API hit the anonymous rate limit (60 requests per hour per IP) on its first call this issue, so the organization push list was completed via repository-level atom feeds and web scraping; main-branch commit titles, merge times, and PR bodies all come from atom and .patch originals, with no secondhand paraphrase. Merge counts, open-update counts, and new-issue counts are taken from GitHub search ground truth (the is:merged, is:open, and created query classes) and cross-checked against per-repo commit counts, which matched: 21 merges = tilelang 1 + TileOPs 18 + TileFoundry 1 + tilelang-musa 1. The Result judgments for performance items are taken directly from the Result: lines in each PR body, with no secondary interpretation. The two Google News queries on the news side went through a proxy; returned items were filtered by publication time and were zero within the window, so this issue has no news items and the coverage centers on repository activity. The RISC-V four-source inspection script produced empty output, which is recorded as “no change” per the established convention and is not treated as a collection failure.