TileLang Daily Intelligence Report (2026-11-12)
Research window: 2026-11-11 07:00 ~ 2026-11-12 07:00 (Beijing time, 24 hours; the previous issue’s cutoff was 11-11 07:00).
Sources: Repository-by-repository inspection of the GitHub tile-ai organization (30 repositories, 7 pushed within the window, 2 of which were automated pushes from the documentation site) + organization-level search for merged and under-review ground truth (6 merged, 57 under-review updates, 35 newly opened issues within the window) + commit-by-commit inspection of the main branch (tilelang 3, TileOPs 0, tilelang-hygon 0, mlir-ascend 1) + inspection of non-default branches (hygon distributed, mlir-ascend master) + release page inspection (tilelang / TileOPs, no new versions within the window) + two repositories on the third-party RISC-V line (TileLang–Weft and its benchmark repository) + Google News Chinese and English queries + Hacker News + arXiv. The special inspection of the four RISC-V sources showed no changes over 24 hours.
This Issue’s Index
-
Today’s Highlights: All three mainline merges are in Ascend; the PTO target promotes tvm_ffi to the default execution backend; the Ascend frontend yields 30 issues in a single day of review
- I. Core Project Progress
- 1.1 Ascend: PTO target’s default execution backend switches to tvm_ffi; the entire module is compiled and cached during the device build phase (11-11)
- 1.2 Ascend: AutoSchedule tightens the cycle lower bound with three necessary conditions, cutting Z3 solve time by 10.04% (11-11)
- 1.3 Ascend: Integer atomic operations adapted to signless carriers; 12 PTO test cases cover 32-bit and 64-bit (11-11)
- 1.4 Mainline: All three window merges are in Ascend; CUDA and CPU side remain in the review queue, version still at v0.1.15 (11-11)
- 1.5 TIRx: Two patches under review for ceildiv and split div/mod with negative scale, sharing the same root cause as the nested floor division error (11-11)
- II. Multi-Backend Adaptation
- 2.1 Hygon: Distributed trio merged into the distributed branch, with Block-Push and Block-Pull communication paths (11-11)
- 2.2 Hygon: ds_read_vector Alt2 and LDS layout inference enter review (11-11)
- 2.3 mlir-ascend: CI consolidated into a single PR gate for Ascend 950 AscendC (11-11)
- 2.4 mlir-ascend: Harness knowledge base hardening merged; tl.or resubmitted under a new number (11-11)
- 2.5 Ascend AscendC repo: Ten patches under review including Gated MLA and the fla_chunk trio; merges still at zero (11-11)
- 2.6 CUDA and Metal: SIMD-group out-of-bounds fix, 256-bit vectorization, and SM90 shared transpose under review (11-11)
- II·V. RISC-V Backend and Community (Special Focus)
- 2.7 Four-source inspection: No changes for the fifth consecutive window; TVM and MLIR lines remain at their original references (11-12)
- 2.8 Beyond the four sources: TileLang–Weft releases weft-parity, with 553/553 within 1.03x of Triton (11-11)
- III. Ecosystem and Adopters
- 3.1 TileOPs: Zero commits on mainline; two stacked W4A16 patches cut chunked-prefill-tile from 37.3 to 31.6 microseconds (11-11)
- 3.2 TileOPs: FP8 grouped GEMM for DeepSeek-V3 prefill moves A scale to TMA (11-11)
- 3.3 TileOPs: Softmax two-CTA cluster meets target; long-K NT GEMM selector change falls short (11-11)
- 3.4 TileOPs nightly: Perf Tracker wiki refreshed at 19:52 that day (11-11)
- IV. Community, Tutorials, and Events
- 4.1 Ascend frontend systematic audit: 30 issues in one day, 26 directly targeting Ascend (11-11)
- 4.2 Fixes under review echo the audit: VMI mask validation, SIMD fill restoration, E4M3 ND/NZ copy (11-11)
- 4.3 News side: Zero hits across five Chinese and English query windows; no related entries on HN or arXiv either (11-11)
- 4.4 Correction: Old Triton artifacts in two K3 Q1 comparisons overflow when negating -128 (11-11)
- V. Trend Observations
Today’s Highlights
This window saw only 6 merges in the organization, another third down from the previous period’s 9, but the structure matters more than the number: 3 of the 6 landed in the main repo, all on Ascend, with not a single one on CUDA or CPU. The largest is #3516 (20 files, +1549/−182), which promotes tvm_ffi to the default execution backend for the pto target—previously target="pto" went through the Cython launcher; now execution_backend="auto" resolves directly to tvm_ffi, with shared host codegen generating MakePackedAPI and C host, and the Ascend runtime launching the aibin compiled from PTODSL via aclrtLaunchKernelWithHostArgs, matching the path used by AscendC kernels. Alongside is a module-level toolchain callback: the entire PTODSL module is compiled in one pass within the device build step (ptodsl → ptoas → bisheng fatobj-link → extract .aicore_binary), with results keyed by ptoas parameters, ptoas version, and CANN identity. The old path is retained as execution_backend="pto", with bit-identical output from both.
The second, #3776, is a tightening on the solver side. AutoSchedule previously started its bisection from a fixed period lower bound; this change takes the maximum of three necessary conditions: the single-task bound uses the maximum of each task’s initiation interval; the resource-group bound sorts pairwise-conflicting tasks by residual value and derives from adjacent differences and head-to-tail differences that the period is no less than the sum of initiation intervals; the self-dependency bound solves S_u - S_u >= wait - d*P backwards to obtain a period no less than ceil(wait/d), with version-dependent distances further scaled by num_stages. No additional lower-bound probing was introduced, and the scheduling model is unchanged. Empirical results on the weekly build of Ascend 950PR with CANN 9.3.0: with compilation caching disabled and one warm-up plus three measured runs alternating, all Ascend examples compile successfully with unchanged scheduling results, and cumulative Z3 solving time drops 10.04%.
The third item is not in the main repo. Hygon’s distributed trio #36 (6 files, +2022/−0) merged into the distributed branch rather than main at 16:20, comprising three compute-communication fused operators: a Block-Push version of GEMM + AllReduce, a Block-Push version of GEMM + ReduceScatter, and a Block-Pull version of AllGather + GEMM, with performance data and implementation notes in the README. The same day, #37 opened HCU’s ds_read_vector Alt2 and LDS layout inference. This line runs from the 10-29 push-wave reduce-scatter example all the way to the fused operator list—the first time the Hygon repo has done distributed work systematically.
The fourth item is a concentrated burst on the quality side: of the 35 new issues in the window, 30 landed in the main repo, 26 of them aimed squarely at Ascend. Two accounts each filed a dozen-plus between afternoon and late night, highly consistent in form—not “it ran wrong” but “it should have errored yet passed silently”: T.Kernel accepts zero or negative block counts, T.alloc_l1 accepts a 2 GiB allocation, T.Persistent accepts negative waves and stages, T.unroll accepts −1 and emits #pragma unroll -1, and negative tensor dimensions crash the compiler outright. Another group is semantic mismatch: hf32’s nearest_zero compiles to NEAREST_AWAY, the GEMM input dtype table and codegen disagree in both directions, the reduce_bitand family emits types undefined in the Ascend templates, and same-shape ND-to-NZ scatter falls into an infinite loop because mask counting does not fall back. Most of these issues remain open, but corresponding fixes—VMI mask validation, SIMD fill restoration, and others—are already under review as of the same day.
Sources: https://github.com/tile-ai/tilelang/pull/3516 , https://github.com/tile-ai/tilelang/pull/3776 , https://github.com/tile-ai/tilelang-hygon/pull/36 , https://github.com/tile-ai/tilelang/issues/3920
I. Core Project Progress
1.1 Ascend: PTO target default execution backend switched to tvm_ffi; whole module compiled and cached at device build stage (#3516)
#3516 (20 files, +1549/−182) is the largest change this period. The substance of the change is migrating the pto target from a “dedicated Cython launcher” to “the same host path as AscendC”: when target="pto" is paired with execution_backend="auto", it resolves to tvm_ffi, the shared host codegen handles MakePackedAPI and C host construction, and the Ascend runtime module launches the aibin produced by PTODSL via aclrtLaunchKernelWithHostArgs.
On the device build side, a new module-level toolchain callback compiles the entire PTODSL module in one pass: ptodsl frontend output, ptoas lowering, bisheng fatobj linking, then .aicore_binary extraction, with artifacts cached by aibin. The cache key incorporates ptoas parameters, ptoas version, and CANN identity together—this point is critical, since previously such keys omitted the toolchain version, causing cross-CANN upgrades to hit stale artifacts.
Validation coverage is broad, with four new test groups: auto resolving to tvm_ffi, callee-allocated outputs, dynamic shapes and dynamic outputs, multi-device kernels, dynamic shared memory, cube/vector mixed GEMM, source accessor, exported libraries, host-side loop launches, and host-side profiling ranges; additionally, equivalence tests bit-compare tvm_ffi against the old pto on three kernels—vecadd, multi-kernel, and dynamic shared memory; two cache test groups cover binary caching and kernel caching respectively, including cache hits skipping the toolchain, corrupted entries raising errors, and hits in a fresh process. The old path is retained as planned, to be retired after coverage is confirmed.
Source: https://github.com/tile-ai/tilelang/pull/3516
1.2 Ascend: AutoSchedule tightens cycle lower bound with three necessary conditions, cutting Z3 solve time by 10.04% (#3776)
#3776 (1 file, +20/−1) changes only the initial value, not the algorithm: the original bisection started from a fixed lower bound, whereas now three necessary conditions are computed first and the maximum is taken. The single-task bound retains the hard lower bound on each task’s initiation interval; the resource-clique bound targets cliques of pairwise resource conflicts, sorts task residual values in ascending order, and derives from the sum of adjacent-difference constraints and first-last-difference constraints that the cycle is no less than the sum of initiation intervals within the clique, with the clique itself greedily enumerated using each task as a seed; the self-dependency bound is solved back from S_u - S_u >= wait - d*P, with version-dependent distance bounded above by num_stages.
The empirical environment is specified in detail: Huawei Ascend 950PR, CANN 9.3.0 weekly build 20261009.01, compile cache disabled, one warm-up and three measurements alternating per configuration. There are two results: all Ascend examples compile successfully with scheduling results identical to before the change (indicating only the lower bound was tightened, not the solution), and cumulative time in the Z3 solve phase dropped 10.04%. The bisection bounds and solve-state handling are unchanged, and no additional lower-bound probing flow was introduced.
The value of such a change to upstream lies not in single-point performance but in the fact that as solver time grows with operator scale, the waste from the bisection starting point amplifies linearly—writing necessary conditions into the initial value is equivalent to saving a meaningless interval on every solve.
Source: https://github.com/tile-ai/tilelang/pull/3776
1.3 Ascend: Integer atomics adapted to signless carriers, 12 PTO cases covering 32-bit and 64-bit (#3794)
#3794 (4 files, +52/−6) fixes the fallout from #3512. That change moved signed integer types from signless i32/i64 to si32/si64, while PTO’s atomic interface requires signless carriers, so integer add/min/max atomics in SIMT kernels failed to compile outright.
The fix performs local adaptation at the generated PTODSL call boundary: pointers and operands are reinterpreted as signless at the same width, preserving the pointer’s address space, and return values are restored to the logical type; min/max retain their original signed or unsigned comparison semantics. This way the logical type system introduced by #3512 is not overturned, and the adaptation is confined to the atomic layer.
Regression cases land in test_simtvf_atomics.py under testing/ascend/language/simtvf, covering 32-bit and 64-bit add/min/max, negative operands, returning the old value, values exceeding 32 bits, and unsigned initial values with the high bit set; the 12 new PTO cases all pass on Ascend. #3792, merged in the same batch, fixes another facet of the same family of issues: T.atomic_add/max/min failing to compile on PTO due to signed inputs.
Source: https://github.com/tile-ai/tilelang/pull/3794
1.4 Mainline: All three merges in the window are Ascend; CUDA and CPU sides entirely in the review queue; version still at v0.1.15
Breaking down the three merges in the main repo window by direction, the conclusion is straightforward: #3516, #3776, and #3794 all land on Ascend, with zero merges on the CUDA and CPU sides. The previous period still had three CUDA instruction-level primitives entering the mainline; this period has none—not because the CUDA side has stalled, but because everything is piled up in the review queue: #3809 (SM90 shared transpose lowered to ldmatrix/stmatrix), #3826 (int16x16/uint16x16 256-bit vectorization), #3723 (SIMT im2col hoisted before warp specialization), #3923 (pipelined GEMM operands live until completion wait), and #3898 (async GEMM shared operand lifetime) all have updates within the window but are unmerged.
The version side remains pending. The main repo release page still shows v0.1.15 from 09-30, with the tag sequence unchanged; the 0.1.16 version-number PR was not updated within the window. The TileOPs release page still shows v0.1.0 from 10-27. In other words, the 1549 lines of Ascend execution backend changes merged this period will not enter downstream distribution packages until at least the next tag.
Source: https://github.com/tile-ai/tilelang/releases , https://github.com/tile-ai/tilelang/pull/3809 , https://github.com/tile-ai/tilelang/pull/3826
1.5 TIRx: Two in-review changes for ceildiv and split div/mod with negative scale, same root as nested floor division bug (#93, #94)
On the tile-ai/tvm side, two changes are in review within the window: #94 ([Fix][TIRx] Make ceildiv exact for signed divisors) and #93 ([Arith] Guard negative scales in canonical split div/mod). Both fix two ends of the same class of problem—once signs enter integer division and modulo, the preconditions for canonicalization rewrites no longer hold, so “exactness” and “safety” must be handled separately: ceildiv requires results to remain exact with signed divisors, while split div/mod must reject canonicalization under negative scales.
This points to the same place as #3937 newly opened in the main repo that day ([BUG] Nested floor division produces wrong integers under negative multipliers): once negative coefficients enter the rewrite chain of arithmetic simplification, boundary conditions are easily missed. The difference is that #3937 is a user-side reproduction, while #93/#94 are proactive hardening on the compiler side.
Source: https://github.com/tile-ai/tvm/pull/94 , https://github.com/tile-ai/tvm/pull/93 , https://github.com/tile-ai/tilelang/issues/3937
II. Multi-Backend Adaptation
2.1 Hygon: Distributed Trio Merged into the distributed Branch, Two Communication Paths Block-Push and Block-Pull (#36)
#36 (6 files, +2022/−0) was still open in the previous issue and was merged this issue at 16:20, landing on the distributed branch rather than main—the repo’s default branch main still has its latest commit at 11-06, so this change will not appear in the default branch’s commit list and can only be seen by querying the branch. The landing-point difference needs to be spelled out: main shares the same reference as v0.1.15-release, while distributed has separately carried forward the 11-03 signal primitives, the 10-29 push-wave reduce-scatter example, and this issue’s trio.
The content is three operators: the Block-Push version of GEMM + AllReduce, the Block-Push version of GEMM + ReduceScatter, and the Block-Pull version of AllGather + GEMM. In Push, the compute side actively pushes results to the peer; in Pull, the compute side pulls the peer’s allgather results. Which path to choose depends on the communication domain size and the overlap window; the README provides a performance comparison and implementation notes. On the validation side, the Torch performance baseline for the GEMM reduce-scatter benchmark was already added on 11-06 and was not changed again this issue.
Source: https://github.com/tile-ai/tilelang-hygon/pull/36
2.2 Hygon: ds_read_vector Alt2 and LDS Layout Inference Opened for Review (#37)
On the same day at 11:21, #37 ([HCU] Support ds_read_vector Alt2 and inferred LDS layouts) was opened, aiming to complete two instruction forms: the second encoding form of ds_read_vector, and LDS layouts that can be derived from constraints. This is on the same layout-inference line as #34 merged on 11-06 (which uses DPP and lane swap to optimize the XOR shuffle for reduction and propagates known layout constraints before free candidates); the difference is that #34 handles lane swapping on the reduction side, while #37 handles the encoding and layout realization of vector reads.
The Hygon repo currently has 15 branches, including six in the feat/hcu-* series and one distributed branch, indicating that this backend’s work has expanded from single-card instructions to both communication and layout.
Source: https://github.com/tile-ai/tilelang-hygon/pull/37
2.3 mlir-ascend: CI Narrowed to a Single Ascend 950 AscendC PR Gate (#222)
#222 (five workflow files, +7/−9) was merged into the master branch at 10:27. The change is deliberately small: all job definitions, conditions, and steps are left untouched; only the on: block of each workflow file is changed—the region least prone to concurrent editing, thereby minimizing the merge-conflict surface.
The motivation is stated plainly: previously every PR expanded across the entire upstream matrix (self-hosted NVIDIA, AMD gfx942, macOS, and an Ascend leg that first builds CANN and then runs examples/ascend and testing/ascend twice), whereas the only backend under development on master is Ascend 950; the other legs are essentially noise and are still occupying scarce hardware, and no one is currently pushing the PTO backend either. After the change, the PR gate has only two items: clean-code lint, and the Ascend 950 AscendC line.
The implication for downstream is that the admission baseline of the mlir-ascend repo is now a single hardware generation, and regressions on older Ascend generations will no longer be a merge condition.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/222
2.4 mlir-ascend: Harness Knowledge-Base Hardening Merged, tl.or Resubmitted Under a New Number (#219, #225)
#219 (merged at 10:03) is hardening of the harness branch: the execution and evaluation pipeline, the analysis/development knowledge base, and the search@2 dependency queue are all advanced together. This is the same sequence as #217 from 11-10 (knowledge-entry retention and approval-number deduplication), indicating that the repo treats “reliably retaining knowledge accumulated during exploration” as serious engineering rather than leaving it scattered across discussions.
Three other proposals came and went within the window: #223 (adding the four address spaces L0A/L0B/L0C/L1 to tl.alloc, covering 4 of the 21 functions required by FYP) and #224 (the tl.or dialect operator) were both closed without merging, while #225 was resubmitted with the same content and remains open; #225’s implementation path follows the add/sub/mul template, adding TL_OrOp, verifyElementwiseBinary, the conversion of ConvertBinary to hivm::VOrOp, and dispatch on the codegen side, along with dialect round-trip and TL→HIVM test groups; local ctest passes 14/14. #220 and #221 are two skill-side items: an Ascend950 simulation trace workflow (msprof collection, MindStudio Insight viewing, npusim SQLite export) and operator design guidance (stage decomposition, SIMD/SIMT selection, storage budget, Cube MFU, and a conditional Roofline calculator).
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/219 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/225
2.5 Ascend AscendC Repo: Ten-Plus Items Under Review Including Gated MLA and the fla_chunk Trio, Still Zero Merges (11-11)
The tilelang-ascend repo continues its “in-only, no-out” pattern this issue: none of the organization-level 6 merges in the window belong to it, while the review queue grew another layer, with more than ten items updated that day. On the operator side, the most notable are #1970 (adding multi-head Gated MLA forward for both the PTO and AscendC lines), #1994 / #2004 / #1952 (the fla_chunk trio of inter, intra, and intra_token_parallel), #1605 (Kimi Delta Attention decode and chunked prefill forward, targeting 910B), and #1989 (upgrading catlass to v2.2.0).
On the example and documentation side, operator benchmarks are being added: #1558 (rms_norm, 20 cann-bench cases plus performance optimization), #1748 (NMS, 20 cases), #1607 (GroupNorm, 20 cases), #1677 (refining sin/cos docstrings and adding tests and API docs), #1732 (in-place T.tile.silu/sigmoid supported on both backends), and #2000 (PTO uint8 wrapper fix for bitwise operations and src1 validation).
On the testing side there is one hard data point that day: the daily test report in #2003 shows a run at 11-11 07:58 with all 3060 items passing, a 100% pass rate. With the scale at 3060 items and merges at zero for consecutive windows, the bottleneck is indeed review, not development.
Source: https://github.com/tile-ai/tilelang-ascend/pull/1970 , https://github.com/tile-ai/tilelang-ascend/issues/2003
2.6 CUDA and Metal: SIMD-group Out-of-Bounds Fix, 256-bit Vectorization, SM90 Shared Transpose Under Review (11-11)
Non-Ascend backends are all in the review queue this issue, but the direction is clear. On the CUDA side there are three items: #3809 lowers 8-bit and 16-bit shared-memory transpose to ldmatrix/stmatrix (available from SM90), #3826 adds int16x16/uint16x16 to enable 256-bit vectorization, and #3723 moves SIMT im2col scheduling ahead of warp specialization. On the runtime side there are two: #3867 launches on the device where the input resides when the kernel and current device differ, and #3882 raises the dynamic shared-memory limit of cython kernels on each device actually launched.
On the Metal side, #3912 addresses the silent out-of-bounds from the previous issue: when copying an 8x8 simdgroup matrix to an 8x1 output, the compiler emitted a full store with a row stride of 1, overwriting the seven elements after the output. The fix is not a runtime check but rather requiring the arithmetic analyzer to prove, before emitting each SIMD-group store, that the full footprint falls within the destination buffer and the requested region; if it cannot prove this, a readable error is raised during lowering. Also that day, #3804 handles region offsets for 2D simdgroup fill.
Source: https://github.com/tile-ai/tilelang/pull/3809 , https://github.com/tile-ai/tilelang/pull/3912 , https://github.com/tile-ai/tilelang/pull/3826
II-V. RISC-V Backend and Community (Special Focus)
2.7 Four-Source Inspection: Fifth Consecutive Window with No Changes; TVM and MLIR Lines Remain at Previous References (11-12)
A reference-by-reference comparison of the four established sources shows no changes across all of them within the 24-hour window:
- The two
kubuds-branches ofkubuds/tilelang-riscv(TVM backend route) remain atkubuds-riscv-karenfc62939 andkubuds-riscv-mahiru33e0c89;mainis only an upstream sync; RuyiAI-Stack/tilelang-riscv(MLIR rewrite route,target="riscv")mainremains at 57c548c;TongmingLAIC/tilelang-riscv(community index page)mainremains at 4499c2a, with no additions or removals in the table;TongmingLAIC/tilelang-tpu(Sophgo SOPHGO TPU)mainremains at 303759f,feature/sg2260e-rv-supportremains at 122bc59, and the other 19 branch references are likewise unchanged.
This is now the fifth consecutive window with no changes (archived on 11-07, 11-09, 11-10, 11-11, and 11-12). Per the established convention, an empty script output is recorded as “no changes” and is not treated as a collection failure.
Source: GitHub four-source reference-by-reference comparison (no changes within the window); no individual landing links.
2.8 Beyond the Four Sources: TileLang–Weft Releases weft-parity, 553/553 Within 1.03× of Triton (11-11)
The CPU line recorded in the previous issue now has a conclusive version. Tianchen-ai/tilelang-weft and its benchmark repository each pushed a commit on the same day (08:51 and 08:52), changing the release tag from v0.1.15-weft.2 to weft-parity and rewriting the results table and results handbook.
The new methodology has two layers. The first is a paired comparison against Triton: all 553 items have a matching native Triton control, and each item satisfies “TileLang native time no more than 1.03× the corresponding Triton time” — SG2044 133/133, K1 137/137, K3 A100 136/136, and x86 147/147, totaling 553/553 across the four platforms. Candidate and control are measured alternately on the same batch of inputs, and the candidate is the same native library exported after passing source-level JIT numerical checks. The second layer is the ratio against external reference libraries: 543 cases have a matching reference, of which 468 are faster than the reference, with platform median ratios of SG2044 1.3598, K1 1.3176, K3 A100 1.6569, and x86 1.2917.
These two sets of numbers must be read separately from the previous issue. The previous issue recorded the old table for the same batch of cases: 41 cases faster than the reference, with an SG2044 median of 0.0580×. The two commits on the same day replaced that table wholesale, for the reason stated in the commit title — “performance inheritance from the shared compiler across four platforms.” According to the results handbook, TileLang native TIRx now imports the same Canonical compiler used by Triton–Weft, and the full packed workload preserves input preparation, encoding, integer partials, scale/min branches, and output updates. In other words, the jump comes from the measured object changing from “author-written programs” to “shared-compiler output,” not from the same implementation getting twenty times faster in a single day. This must be stated clearly when citing; the new table must not be used to overwrite the old conclusion.
There is also one correction worth noting: two Q1 controls on K3 previously used an old Triton artifact that overflowed when negating −128 — the boundary input yielded −2048, when the correct result should be 0. After the shared compiler’s exact I8 reduction was fixed and regenerated, the artifacts from both frontends pass that input; triton_historical_us retains the old timing, and triton_control_note flags the corrected comparison. This is the only disclosure in this issue of “the control itself was wrong,” and it says more about how seriously this benchmark takes reproducibility than any score does.
The technical routes still need to be kept distinct: Weft follows a CPU route that imports the TileLang native AST early into its own Canonical IR, which is not the same implementation as the TVM backend continued by the kubuds- prefixed branches or the MLIR Linalg/SCF rewrite in RuyiAI-Stack/tilelang-riscv.
Source: https://github.com/Tianchen-ai/tilelang-weft , https://github.com/Tianchen-ai/tilelang-weft-bench
III. Ecosystem and Adopters
3.1 TileOPs: Zero commits on mainline, two stacked W4A16 PRs push chunked-prefill-tile from 37.3 to 31.6 microseconds (#2895, #2896)
TileOPs had zero commits on the main branch this period, but opened five performance PRs, two of which are stacked members of the same family. #2895 (3 files, +183/−3) converts the W4A16 tile split along K into a cluster of paired CTAs: each CTA holds the rows of one consumer warpgroup, receives the FP32 partials for those rows in the peer’s shared memory via st.async, and then writes them out, so these tiles no longer launch a split-K reduction or round-trip through the FP32 workspace. On H200 at 1500 MHz, chunked-prefill-tile drops from 37.3 microseconds to 34.7 microseconds (torch dequantized matmul is 37.5 microseconds), with output bit-identical to the reduction path.
#2896 (3 files, +525/−5) stacks another layer on top: the 128-row 128-column W4A16 tile switches to a hand-written warp-specialized main loop that decodes the next K tile while the previous WGMMA is still running, and releases the ring segment only after the WGMMA reading that segment retires; since the cost model has no term for this, a calibration rule is used instead to place these tiles on four 128-wide K segments. The same benchmark drops to 31.6 microseconds, 1.19× torch dequantized matmul, with output bit-identical to the pipelined main loop.
Both PRs belong to the performance checklist in #2130, and #2896 explicitly states it is stacked on #2895 and that its commits will show alongside until the earlier PR merges.
Source: https://github.com/tile-ai/TileOPs/pull/2895 , https://github.com/tile-ai/TileOPs/pull/2896
3.2 TileOPs: FP8 grouped GEMM for DeepSeek-V3 prefill routes A scale through TMA (#2894)
#2894 (1 file, +2/−2) changes only two config entries: stage_scales for ds-v3-ep8-prefill-up and ds-v3-ep8-prefill-down goes from 1 back to 0, so their 128-row tiles fetch A scale from the TMA box added in #2865 rather than relaying through the producer. Small as the change is, it completes a whole chain: it replaces the remainder of #2839, the rest of which was absorbed by #2864, #2865, and #2885 respectively.
Numbers measured on H200 SXM, SM 1500 MHz, CUDA 13.2, PyTorch 2.13, TileLang 0.1.12, device_busy_ms, interleaved over three runs against mainline 9bf4bf56 taking the median: up projection 269.3 → 262.0 microseconds (ratio to DeepGEMM’s 287.1 microseconds rises from 1.066 to 1.096), down projection 142.2 → 137.2 microseconds (ratio 1.117 → 1.157). The other two contiguous config rows are unchanged. On the test side, tests/ops/test_moe_grouped_gemm_fp8.py passes 53 and skips 25 on H200.
Source: https://github.com/tile-ai/TileOPs/pull/2894
3.3 TileOPs: softmax two-CTA cluster meets the bar, long-K NT GEMM selector change falls short (#2898, #2897)
#2898 (SoftmaxFwdOp, marked bar met) moves two 32 KB softmax rows from a cluster of four 128-thread CTAs to a cluster of two 256-thread CTAs: on SM90 every CTA slot can hold a two-CTA cluster, whereas a four-CTA cluster leaves empty slots. Two register-side changes accompany it—row maxima are fetched as packed 16-bit pairs, and each float16 element uses a single FFMA to shift by the tile’s scaled maximum; bfloat16 and float32 still subtract the max first, because their max times log2(e) would overflow fp32 range. The root cause is stated in the description: the benchmark runs H200 at 1500 MHz, and the register scheme executes 17% more instructions than quack and is issue-rate limited.
#2897 (GEMMFwdOp, marked below bar) goes the other way and missed. It targets the selector’s decision: beyond 64 K steps, a 176- or 192-wide ping-pong tile replaces the wider coop2 tile only when it covers more “output per byte of operand,” so ds-v3-prefill-o-proj switches from a 128x176/144 ping-pong to a 128x256/240 coop2. The problem is that this row runs under H200’s 700 W power wall, where time tracks energy per call: the 176-wide tile uses 2.7% fewer cycles than DeepGEMM but moves 21% more operand bytes into the SM, so it runs at a lower clock. The PR honestly marks itself as below bar; its value is in quantifying that cycle counts are not a valid criterion under a power wall.
Source: https://github.com/tile-ai/TileOPs/pull/2898 , https://github.com/tile-ai/TileOPs/pull/2897
3.4 TileOPs nightly: Perf Tracker wiki refreshed at 19:52 that day (11-11)
The push timestamp for the TileOPs-nightly repo this period comes from a wiki change rather than code: at 19:52 that day the repo wiki’s Home page was created, corresponding to the nightly sync of the Perf Tracker dashboard. Such pushes get counted as “pushes within the window” when auditing repo by repo across the org, and need to be distinguished from real code commits—this window saw zero commits on the repo’s main branch, and the latest push to the snapshots branch is still 11-09.
In the same period, TileOPs.github.io’s gh-pages had a push on 11-10, outside this window. Automated writes like the docs site and nightly are no longer listed as separate entries this period; this note records the convention.
Source: https://github.com/tile-ai/TileOPs-nightly
IV. Community, Tutorials, and Events
4.1 A Systematic Audit of the Ascend Frontend: 30 Issues in One Day, 26 Directly Targeting Ascend (11-11)
During the window, 35 new issues were opened at the organization level, of which 30 landed in the main repo, 3 in TileFoundry, and 2 in tilelang-ascend. Of the 30 in the main repo, 26 had titles directly prefixed with Ascend, concentrated across two accounts: 12 from cn-sys (submitted consecutively from 15:19 to 16:14) and 13 from JINITAIMEI121 (08:31 to 10:39); the remaining one was a PTO multi-region duplicate-physical-segment report.
What these issues share is not “wrong results” but “things that should not pass yet do.” One class is parameter validation: T.Kernel accepts zero or negative block counts, T.Persistent accepts negative wave_size and num_stages, T.Pipelined accepts negative num_stages and invalid order/stage metadata, T.unroll accepts −1 and emits #pragma unroll -1, loop constructors accept zero or negative extents, negative tensor dimensions crash the compiler outright (a native crash rather than a Python exception), and T.alloc_l1 accepts a 2 GiB allocation with no capacity check whatsoever. Another class is semantic mismatch: hf32’s nearest_zero compiles to NEAREST_AWAY, the GEMM input dtype table and codegen disagree in both directions, reduce_bitand/bitor/bitxor emit types undefined in the Ascend templates, batch reduction degrades to a scalar AllReduce entry point with no batch overload, the pad_value of a GM-to-L1 copy is routed into an L1 fill that only accepts zero, a same-type ND-to-NZ scatter loops forever because the mask count does not fall back, and partial-VL rows produce a nonexistent PAT_VL<N> pattern. A third class is toolchain: vsts_pack_quarter has no 16-bit branch, so PK4_B32 stores on b16 vectors are silently dropped; module serialization discards the device source, leaving cached kernels with no source to inspect when launch fails; and the msprof backend’s autotuning necessarily fails because it calls signal.signal() from a worker thread.
This batch is not scattered grumbling; in form it resembles an organized audit: the same account filed them in thematic runs within half an hour, with titles uniformly of the form “[BUG] Ascend: …” plus a one-line root-cause summary. The significance for upstream is that the Ascend frontend’s “soundness” gaps were laid out on the table all at once—3060 daily tests all passing does not mean illegal inputs are rejected, because the tests cover the legal paths.
Sources: https://github.com/tile-ai/tilelang/issues/3941, https://github.com/tile-ai/tilelang/issues/3928, https://github.com/tile-ai/tilelang/issues/3947
4.2 In-Review Fixes Echo the Audit: VMI Mask Validation, SIMD Fill Restoration, E4M3 ND/NZ Copy (11-11)
Among the in-review PRs updated that day, a batch lines up precisely with the gaps above. #3861 and #3917 are two halves of the same thing: the VMI unary and vector-scalar factories previously only checked whether a mask was “present,” not its dtype or lane count, so passing an ordinary vector as a predicate would still make it all the way to codegen—seven operators (vabs, vneg, vrelu, vexp, vln, vsqrt, vnot) are built on this factory, and the vector-scalar side’s vadds, vmuls, vmaxs, vmins, vshls, vshrs have weak forms too.
#3795 fixes another side effect of #3512: when T.clear is applied to an int32 UB buffer inside T.SimdVF(), the generated vector broadcast only supports floating point and PTO compilation fails—because the PTO pipeline replaces the original SIMD pass with a dedicated Parallel pass, incidentally dropping the fill lowering. The fix adds a LowerSimdVFFill after buffer-version normalization and before PTO Parallel validation, reusing vdup/vsts for contiguous and complete vector regions while the rest still takes the old path; it also preserves the numeric conversion of the floating-point tl.simd.vdup, fixing the case where “int32 7 is bit-broadcast into roughly 9.809e-45.”
#3639 adds E4M3 ND/NZ UB-to-L1 copies, #3926 makes T.Kernel reject non-positive block counts and negative Pipelined num_stages, #3854 fixes division by zero for sub-byte elements in SimdVF lowering, and #3668 adds interleaved strided UB-to-GM writes within T.PerCoreTask() on AIV cores. Separately, #3915 and #3916 handle AssertStmt error kinds and in-place assertions for condition checks.
Sources: https://github.com/tile-ai/tilelang/pull/3861, https://github.com/tile-ai/tilelang/pull/3795, https://github.com/tile-ai/tilelang/pull/3926
4.3 News Side: Zero Hits Across Five Chinese and English Query Windows; No Relevant Items on HN or arXiv Either (11-11)
Google News queries in Chinese and English returned zero hits during the window, with the latest hits still stuck on older articles; Hacker News hits were all same-name noise (items unrelated to TileLang), and the latest all:"TileLang" results on arXiv remain outside the window. The TileLang-related content currently indexable by search engines is still the ICLR 2026 paper page and the Microsoft Research publication page—long-standing entries rather than news.
Zero hits on the news side for several consecutive days is normal for TileLang: the project’s reach runs mainly through GitHub and developer communities, not news indexes, and the reporting subject remains repo activity.
Sources: Google News Chinese and English queries (via proxy), Hacker News, arXiv; no relevant items within the window.
4.4 Methodology Correction: The Old Triton Artifact in Two K3 Q1 Comparisons Overflows on Negating -128 (11-11)
This gets its own entry because what it affects is not the score but comparability. In the same batch of commits that published weft-parity, the benchmark repo disclosed that the two K3 Q1 comparisons had previously used the retained Triton Canonical program, which overflows when negating −128—the boundary input yields −2048 when the correct result should be 0. After the shared compiler’s exact I8 reduction fix was regenerated, the artifacts from both frontends pass that input; the historical timing is retained in the triton_historical_us field, and the correction note is recorded in triton_control_note.
An erroneous control implementation is a common trap for benchmark projects: once the reference side is wrong on a boundary input, the tested side “underperforming the reference” may just mean the reference is faster at being wrong. Read together with the paired methodology in 2.8, this disclosure shows that this CPU line is moving from “the author’s program versus an external library” to “paired comparison of artifacts from the same compiler”—more credible methodologically, but the two tables cannot be mixed.
Source: https://github.com/Tianchen-ai/tilelang-weft-bench/blob/main/docs/results.zh-CN.md
V. Trend Observations
First, Ascend’s weight on the mainline keeps rising, and the focus is shifting from “adding operators” to “adding execution paths and solvers.” All three main-repo merges this issue are Ascend: switching the default execution backend, tightening the solver’s lower bound, and adapting atomics to the type system. None of the three adds to the operator list; all change “how the same program gets compiled, run, and how long compilation takes.” Against the record since October, Ascend-side merges have moved from filling in AscendC operators to three layers: interface semantics, execution backend, and scheduling solver.
Second, the CUDA side has entered a long review period. Last issue still had three CUDA instruction-level primitives landing on main; this issue has zero merges and six in review (ldmatrix/stmatrix, 256-bit vectorization, im2col scheduling, pipelined operand lifetimes, async GEMM lifetimes, PASS config namespace). A falling merge rate together with a thickening review queue usually means review bandwidth has become the bottleneck, not that development has stalled.
Third, quality auditing is starting to run in parallel with feature development. The shape of 26 Ascend issues in one day shows that someone in the community is already doing systematic illegal-input auditing, and the corresponding fix PRs were in the queue the same day (mask validation, negative-parameter rejection, fill lowering restoration). For a project spreading its backend across multiple domestic chips, this is a good thing: 3060 daily tests all passing only proves the legal paths work; the ability to reject illegal paths depends on audits like these.
Fourth, the four RISC-V sources have been silent for five consecutive windows, but the CPU line outside those four has reached an interim conclusion. The TVM route, the MLIR route, the community index page, and the Sophgo TPU—all four upstreams remain at their original citations; meanwhile TileLang–Weft, using the shared Canonical compiler, brought all 553 items to within 1.03× of Triton, with median ratios of 1.29 to 1.66 across the four platforms. Note that this is a result after a methodology switch and cannot be compared directly with last issue’s 0.058×. The real things to watch: whether this “same compiler, two frontends” approach can make its way upstream in reverse, and whether the RISC-V RVV/IME path gets proven out first by projects outside the four sources.
Limitations: This issue’s GitHub anonymous API quota was exhausted early in collection; PR bodies and merge details were filled in via the search API and the commit atom streams of non-default branches. For a few PRs, file and line counts are search-API return values and the diffs were not reviewed line by line; the hygon and mlir-ascend merges landed on non-default branches and were checked per branch. The news side had only zero hits, with no external coverage for cross-validation.
Appendix: Sources and Verification Notes
| Source | Verification Result |
|---|---|
| tile-ai org, repo by repo (30 repos) | 7 pushed within the window: tilelang, TileOPs, TileOPs-nightly, tilelang-hygon, tilelang-mlir-ascend, tilelang.github.io, TileOPs.github.io |
| Main repo tile-ai/tilelang | 3 merges (#3516, #3776, #3794), all Ascend; 30 new issues |
| Non-default branches | hygon distributed merged #36; mlir-ascend master merged #222; both repos’ main had zero commits |
| Review queue | 57 org-level PRs updated within the window, 33 in the main repo |
| Releases page | tilelang latest v0.1.15 (09-30), TileOPs latest v0.1.0 (10-27), no new tags within the window |
| Daily tests | tilelang-ascend #2003: 3060 items run on 11-11 07:58, all passed |
| TileLang×RISC-V four-source sweep | No change (fifth consecutive window) |
| Third-party RISC-V line | TileLang–Weft and the benchmark repo, 1 each; released weft-parity |
| News side | Zero hits within the window across Google News (Chinese and English), Hacker News, and arXiv |
Full Source List
- [1] tilelang #3516 (tvm_ffi execution backend for PTO targets) — https://github.com/tile-ai/tilelang/pull/3516
- [2] tilelang #3776 (AutoSchedule cycle lower bound) — https://github.com/tile-ai/tilelang/pull/3776
- [3] tilelang #3794 (PTO integer atomics) — https://github.com/tile-ai/tilelang/pull/3794
- [4] tilelang #3792 (signed inputs for atomic add/max/min) — https://github.com/tile-ai/tilelang/pull/3792
- [5] tilelang releases page — https://github.com/tile-ai/tilelang/releases
- [6] tilelang #3809 (SM90 shared-memory transpose) — https://github.com/tile-ai/tilelang/pull/3809
- [7] tilelang #3826 (256-bit vectorization) — https://github.com/tile-ai/tilelang/pull/3826
- [8] tilelang #3912 (Metal SIMD-group out-of-bounds) — https://github.com/tile-ai/tilelang/pull/3912
- [9] tilelang #3861 (VMI mask validation) — https://github.com/tile-ai/tilelang/pull/3861
- [10] tilelang #3795 (PTO SIMD fill restoration) — https://github.com/tile-ai/tilelang/pull/3795
- [11] tilelang #3926 (reject non-positive block and negative num_stages) — https://github.com/tile-ai/tilelang/pull/3926
- [12] tilelang #3937 (nested floor division) — https://github.com/tile-ai/tilelang/issues/3937
- [13] tilelang #3920 (negative block count in T.Kernel) — https://github.com/tile-ai/tilelang/issues/3920
- [14] tilelang #3928 (zero or negative extent in loop constructors) — https://github.com/tile-ai/tilelang/issues/3928
- [15] tilelang #3941 (infinite loop in ND-to-NZ scatter) — https://github.com/tile-ai/tilelang/issues/3941
- [16] tilelang #3947 (msprof autotuning always fails) — https://github.com/tile-ai/tilelang/issues/3947
- [17] tile-ai/tvm #93 (split div/mod with negative scale) — https://github.com/tile-ai/tvm/pull/93
- [18] tile-ai/tvm #94 (ceildiv with signed divisor) — https://github.com/tile-ai/tvm/pull/94
- [19] tilelang-hygon #36 (distributed trio) — https://github.com/tile-ai/tilelang-hygon/pull/36
- [20] tilelang-hygon #37 (ds_read_vector Alt2) — https://github.com/tile-ai/tilelang-hygon/pull/37
- [21] tilelang-mlir-ascend #222 (CI gate consolidation) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/222
- [22] tilelang-mlir-ascend #219 (harness hardening) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/219
- [23] tilelang-mlir-ascend #225 (tl.or dialect op) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/225
- [24] tilelang-ascend #1970 (Gated MLA forward) — https://github.com/tile-ai/tilelang-ascend/pull/1970
- [25] tilelang-ascend #2003 (daily tests, 3060 items) — https://github.com/tile-ai/tilelang-ascend/issues/2003
- [26] TileOPs #2894 (A scale via TMA) — https://github.com/tile-ai/TileOPs/pull/2894
- [27] TileOPs #2895 (cluster reduction for split-K) — https://github.com/tile-ai/TileOPs/pull/2895
- [28] TileOPs #2896 (W4A16 main loop) — https://github.com/tile-ai/TileOPs/pull/2896
- [29] TileOPs #2897 (long-K NT GEMM selector) — https://github.com/tile-ai/TileOPs/pull/2897
- [30] TileOPs #2898 (softmax two-CTA cluster) — https://github.com/tile-ai/TileOPs/pull/2898
- [31] TileOPs-nightly — https://github.com/tile-ai/TileOPs-nightly
- [32] TileLang–Weft (weft-parity) — https://github.com/Tianchen-ai/tilelang-weft
- [33] TileLang–Weft benchmark repo — https://github.com/Tianchen-ai/tilelang-weft-bench
- [34] TileLang–Weft results handbook — https://github.com/Tianchen-ai/tilelang-weft-bench/blob/main/docs/results.zh-CN.md