Research window: 2026-11-17 07:00 ~ 2026-11-18 07:00 (Beijing time, 24 hours; previous issue cutoff at 11-17 07:00).

Sources: Repository-by-repository check of the GitHub tile-ai organization (30 repos, 10 pushed within the window) + organization-level merged and in-review ground-truth search (20 merged, 77 in-review updates, 15 newly opened issues within the window) + commit-by-commit check of main branches (TileOPs 12, TileFoundry 3, tilelang 2, topa 1, docs site 2) + release page check (tilelang remains at v0.1.15, no new tags within the window) + Google News Chinese and English queries + Hacker News + arXiv + TileLang×RISC-V four-source dedicated inspection.


This Issue’s Index

  • Today’s Highlights: TileOPs merges 12 PRs in one day; pilot model interfaces move from registration to operator-by-operator delivery (11-17)
  • Today’s Highlights: Ascend frontend contributor files 5 dtype audit issues in a row, covering roughly 300 operator-dtype combinations (11-17)

  • I. Core Project Progress
    • 1.1 Main repo fixes CuTeDSL unsigned buffer read: uint8 200 interpreted as -56 (11-17)
    • 1.2 Main repo Ascend benchmark switches to device-side L2 flush, replacing the old practice of reading a temporary buffer (11-17)
    • 1.3 GEMM adds float32 output, serving three narrow-N shapes: router, compressor, and LM head (11-17)
    • 1.4 Three new operators in the operator library: Embedding resubmission, Concat, and RoPE with caller-supplied cos/sin (11-17)
    • 1.5 Five items across normalization, attention, and linear attention: zero-centered weights, irregular head dims, state slot indexing (11-17)
    • 1.6 Benchmark timing switches to nvbench cold measurement; nightly sweep budget raised from 6 to 10 hours (11-17)
    • 1.7 Under review and goals: narrow-N prefill GEMM still has a shape at 0.39x of the fastest library (11-17)
  • II. Multi-Backend Adaptation (Ascend / MetaX / Hygon / Moore Threads / Sunrise / MLIR Ascend)
    • 2.1 Ascend main repo: two synchronization primitives and predicate width fixes under review (11-17)
    • 2.2 Proposal for a standalone Ascend nightly wheel pipeline: with real-hardware smoke tests and rolling releases (11-17)
    • 2.3 MLIR Ascend: tl.print and tl.and merged; roughly 30 TLIR operators under batch review (11-17)
    • 2.4 Backend repos: Ascend reports a new 1D strided scatter write defect; Hygon and Sunrise progress still under review (11-17)
  • II·V. RISC-V Backend and Community (Special Focus: TVM Route / MLIR Route / SOPHGO TPU / Community Index Page)
    • 2.5.1 Four-source inspection shows no changes; citations for both routes match the previous issue (11-18)
  • III. Ecosystem and Adopters (TileFoundry / topa / Site / TileKernels / FlashQLA)
    • 3.1 TileFoundry changes 303 files in one pass: terminology consolidated into a single spec, layout switched to CuTe algorithms (11-17)
    • 3.2 Two TileFoundry fixes: logical window for local views and copy parameters paired by coordinate (11-17)
    • 3.3 topa round mechanism lands; 6 new issues form the next work queue (11-17)
    • 3.4 Two automated updates to the docs site; no commits in either adopter repo during the window (11-17)
  • IV. Community, Tutorials, and Events
    • 4.1 News side quiet: no topical coverage in Chinese or English search, Hacker News, or arXiv (11-17)
    • 4.2 Organization activity: 20 merges and 77 under-review updates in a single day; main repo at 8,992 stars (11-17)
  • V. Trend Observations

  • Appendix: Sources and Verification Notes
    • Source Verification Table
    • Full Source List

Today’s Highlights: TileOPs Merges 12 PRs in One Day; Pilot Model Interfaces Move from Registration to Operator-by-Operator Delivery

On 11-17, the TileOPs main branch landed 12 merges, 7 of which came from operator proposals in the TileFoundry round. The direction follows directly from the previous day: after 11-16 registered the interfaces actually called by the pilot models (Qwen3.6, DeepSeek-V4-Flash, GLM-5.3-Flash) into the manifest, 11-17 began filling in kernels one shape at a time. Newly merged operators that day include Embedding, Concat, and RoPE with “caller-supplied cos/sin,” plus zero-centered weight support for normalization, non-standard head dims for GQA variable-length attention, and state slot indexing for GDN and causal convolution—covering five segments of the pilot model inference chain: indexing, normalization, positional encoding, attention, and linear attention. Each proposal ships with a reference implementation and benchmark row, and most self-assessments remain “below bar,” indicating that merge criteria and performance criteria are separate matters: first ensure semantic correctness and interface alignment, then squeeze performance round by round. Another addition, GEMM float32 output (#3000), extends the three narrow-N shapes—router, KV compressor, and LM head—from bf16 to float32 accumulation output, and is the only change in this round of interface delivery that directly alters the main kernel structure. A newly opened goal issue, #2999, provides the counter-list: of 53 scanned shapes, 5 still lag, with the worst row at only 0.39x of the fastest library.

Sources: https://github.com/tile-ai/TileOPs/pull/3000 , https://github.com/tile-ai/TileOPs/issues/2999

Today’s Highlights: Ascend Frontend Contributor Files 5 dtype Audit Issues in a Row, Covering Roughly 300 Operator-Dtype Combinations

Starting at 20:13 on 11-17, within one minute, contributor JeffDing filed 4 Ascend frontend defects against the main repo (#4391 through #4394), then added a fifth half an hour later (#4395), forming a systematic dtype coverage audit. The core finding is that T.simd unary and binary operators perform no dtype validation at all before being passed into compiler-internal calls: CANN 9.2.1’s CCE vector interfaces provide only partial overloads (for example, absolute value has only five integer and half-precision, single-precision overloads, with no bf16; XOR has only two 32-bit integer variants), yet the frontend lets 70 of 225 (operator, dtype) combinations through directly, resulting in either compile-time failures or silent miscalculation. The cross-lane reduction family likewise lacks a whitelist, with 27 of 63 combinations let through. Two others are more specific: the part parameter of T.simd.vcvt defaults to 0, which lands on the “even part” and converts only half the elements; the documentation claims support for 40 dtype pairs, while the frontend actually implements 33. The fifth issue notes that six T.ascend_* synchronization primitives all fail inside T.SimdVF(), and that 6 of the error messages do not even mention the API the user called. The reported environment is Ascend 950PR (dav-3510), CANN 9.2.1, openEuler 24.03—a frontend quality problem under a new-hardware-plus-new-toolchain combination; all five carry reproduction steps and suggestion labels (bug, python, ascend, UX), and all remain open as of the end of the window.

Sources: https://github.com/tile-ai/tilelang/issues/4391 , https://github.com/tile-ai/tilelang/issues/4393 , https://github.com/tile-ai/tilelang/issues/4395


I. Core Project Progress

1.1 Main repo fixes unsigned buffer reads in CuTeDSL: uint8’s 200 treated as -56

#3875 fixes the issue where values read from uint8 / uint16 / uint32 buffers on the CuTeDSL target were treated as signed numbers. The root cause is that MLIR integers have no unsigned bit; A.iterator.dtype is already Int8 for unsigned tensor parameters, and cute.recast_ptr and cute.make_tensor likewise drop the unsigned information, with the element type retained only at the tensor wrapper layer, while code generation starts from the iterator type on every access to global and shared memory. The author used 256 uint8 values as a probe on an RTX PRO 6000: five formulations — Int32(x), x > 127, x >> 7, x // 3, and four-lane vector loads — each got 128 values wrong on the main branch. The fix is to have tl.make_tensor and make_tensor_at_offset accept dtype= and reattach the type to the view, with code generation explicitly passing cutlass.UintN for each view of an unsigned buffer; the earlier #3764 only added a single cast after loading, which did not cover comparisons, division, shifts, and vector loads.

Source: https://github.com/tile-ai/tilelang/pull/3875

1.2 Main repo Ascend benchmark switches to device-side L2 flush, replacing the old approach of reading a scratch buffer

#3891 changes the L2 eviction method in the Ascend benchmark from “reading a scratch buffer” to calling aclrtDeviceL2CacheFlush, executed before each NPU event or msprof sample. The old approach’s cache state depended on the replacement policy and could be inconsistent across rounds, thereby affecting the comparability of kernel latency; the new approach places the synchronous flush outside timing and runtime estimation, ensuring a deterministic cache state within the measurement window. This interface requires runtime and device support, and existing test stubs have been updated accordingly; the author honestly notes in the description that tests were not rerun for this version. This is in the same direction as the benchmark timing rework in 1.6: both are making “the numbers measured” more reproducible.

Source: https://github.com/tile-ai/tilelang/pull/3891

1.3 GEMM adds float32 output, serving three narrow-N shapes: router, compressor, and LM head

#3000 makes GEMMFwdOp set the output type to float32, i.e., computing a float32 accumulation from bf16 or fp16 inputs, handled respectively by three kernels: TMA, narrow-batch Gemv, and cp.async. Except for the coop2_swap structure, all TMA structures directly move the float32 accumulator; coop2_swap’s store instruction only supports 16 bits, so it explicitly rejects float32 output. Configuration selection adds tiling estimated for 4-byte output, with the 16-bit selection unchanged across 23,040 sampled shapes. Validation uses tensor-core accumulation bounds, with absolute error taken as the larger of 1e-4 and 2e-3 times (K/2048) to the 1.5 power, and relative error 1e-5. Benchmarks compare against cuBLASLt, DeepGEMM, and QuACK; FlagGems’ matmul has no float32 output and exits the comparison on the float32 row. This PR is part of #2927 and shares the same batch of shapes with the newly opened SOTA issue #2999 the next day.

Source: https://github.com/tile-ai/TileOPs/pull/3000

1.4 Three new operators in the operator library: Embedding resubmitted, Concat, and RoPE with custom cos/sin

Embedding, which was mistakenly merged the previous day and reverted that same day, is resubmitted and merged in a formal round this time: #2992 adds EmbeddingFwdOp, supporting int32 and int64 indices and five weight tables (fp16/bf16/fp32/int32/int64), with the kernel using TileFoundry’s tile placement — one CTA takes an entire row block or a column tile of a wide row, each thread lands its own vector into shared memory via cp.async, L2 uses the evict-last hint, then streams out; the max_norm parameter is rejected directly before kernel construction, ensuring weights are not modified in place. #2987 adds ConcatFwdOp, treating the output as “rows times total column width”, with a 128-thread CTA where each thread copies one 16-byte vector; both row count and row width are dynamic, so one compilation serves all shapes with the same vector width and dtype. #2988 adds RoPECosSinFwdOp, applying rotation per the caller-provided per-token cos/sin, covering both NeoX and interleaved layouts, partial rotation, dense and packed, single offset or per-sequence offset, and in-place writes with strided views; the same PR also establishes a general rule: if any tensor being written (output or modified input) has two elements landing at the same address (e.g., an expanded view), all operators must throw CheckError before running. All three operators meet the bar against torch and torch.compile, with self-assessments all below the acceptance line.

Source: https://github.com/tile-ai/TileOPs/pull/2992 , https://github.com/tile-ai/TileOPs/pull/2987 , https://github.com/tile-ai/TileOPs/pull/2988

1.5 Five items on normalization, attention, and linear attention: zero-centered weights, irregular head dims, state slot indexing

On the normalization side, #2986 makes RMSNormFwdOp support Qwen3.6’s six “zero-centered weight” rows (prefill, decode, and q-norm for 27B and 35B-A3B), and FusedAddRMSNormFwdOp supports four of them; all three kernel bodies and the fused variant compute 1+w in float32 at the weight multiplication, the generated code under ordinary weights is byte-for-byte identical to the main branch, and they meet the bar against flashinfer, vLLM, transformers, QuACK, and FlagGems. On the attention side, #2989 makes the variable-length GQA kernel accept its own head dims on SM90 in 8-channel steps: head dims 16, 32, 48, 80, 96, and 112 migrate from the generic kernel to the warp-specialized kernel, with each group faster on prefill, vision, and serving shapes; it also supports the 72-head-dim, 16-group non-causal shapes of Qwen3.6 vision rows without zero-padding tensors. On the linear attention side, #2993 makes GDNFwdOp support state_indices (decode and packed prefill read/write each sequence’s state by pool slot), CausalConv1dFwdOp support cache_indices, and its update kernel support conv_state_indices; each state-holding kernel looks up slots itself based on build flags, builds without a pool generate the same program as the main branch, and value-major GDN decode switches to two warps per group with each lane reading four 16-byte vectors. #2995 fixes a defect that would corrupt the previous sequence: the last-block store of packed prefill is located by “start position plus (iteration count minus one) times block length”, and when the iteration count is 0 this expression places the write start 64 rows before the empty sequence, overwriting the end of the previous sequence with uninitialized shared memory; it reproduces on H200 with lengths (2048, 64, 0) and 64 value heads, and the fix is to write the last block only when the iteration count is greater than 0. #2990 tightens the pairing condition for W4A16 decode: use dual warpgroup only when the paired K slice is longer than its own ring buffer, or when pairing adds no extra K tiles; otherwise retain the single-warpgroup tile.

Source: https://github.com/tile-ai/TileOPs/pull/2986 , https://github.com/tile-ai/TileOPs/pull/2989 , https://github.com/tile-ai/TileOPs/pull/2993

1.6 Benchmark timing switches to nvbench cold measurement; nightly scan budget raised from 6 to 10 hours

#2985 rewrites the sampling method of the CUDA benchmark: previously each sample synchronized to the host before the call, and host-side launch overhead appeared as device idling; now a blocking kernel occupies the stream, enqueueing all samples of one reading consecutively, with each sample resetting, flushing L2, and executing the call on its own, with no host synchronization in between — this follows NVIDIA nvbench’s cold-measurement approach, with the only difference being that “one block at a time” is relaxed to “one block, many samples”. Each sample records SM 0’s clock as the frequency, reports the mean within a row, and does not discard samples due to downclocking; blocks that do not fit in the queue time out and are invalidated and re-measured at a size that fits, with the block size recorded in block_samples. The preparation method is also written into the result’s l2_flush field to avoid cross-methodology comparisons in nightly history. The resulting cost is quantified in #2996: the latest nightly scan ran 20,259 seconds before being cut off by the 6-hour budget, with 5 of 61 benchmark files not run, causing Check benchmark coverage to fail; compared with the last all-green run, the shared 2,015 cases went from 15,180 seconds to 19,103 seconds (1.26x), and the 268 new cases took another 1,653 seconds. The fix is to raise the total budget from 21,600 seconds to 36,000 seconds and the job timeout from 380 minutes to 620 minutes.

Source: https://github.com/tile-ai/TileOPs/pull/2985 , https://github.com/tile-ai/TileOPs/pull/2996

1.7 Under review and goals: narrow-N prefill GEMM still 0.39x behind the fastest library on some shapes

The newly opened #2999 lays out GEMM’s performance debt: on H200, bf16 inputs, K fixed at 4096, scanning 1 to 8192 tokens across the router, KV compressor, and LM head widths of DeepSeek-V4-Flash, 48 of 53 shapes meet or exceed the fastest library, with the remaining 5 lagging; the worst is the GLM-5.3-Flash router shape (8192 by 288 by 4096), at 78.3 microseconds versus 30.6 microseconds for bf16 output, only 0.39x, and 0.38x for float32 output, with the selected tile being 64 by 128 base. The remaining lagging rows fall between 0.86x and 0.98x. Performance proposals also under review include FP8 short-row weight streaming (#3010), W4A16 single-warpgroup grouped reduction switched to cluster pairs (#3009), short-K W4A8 decode with one CTA per row block (#3005), and a kernel rewrite of FP8 paged GQA decode (#2997); on the new-operator side, MoE grouped GEMM with MXFP4 expert weights (#3001), KDA state slots and value-major state (#3002), and NSA variable-length attention with head dim 256 and 4-key blocks (#3004) are under review.

Source: https://github.com/tile-ai/TileOPs/issues/2999 , https://github.com/tile-ai/TileOPs/pull/3010 , https://github.com/tile-ai/TileOPs/pull/2997

II. Multi-Backend Adaptation (Ascend / MetaX / Hygon / Moore Threads / Sunrise / MLIR Ascend)

2.1 Ascend Main Repo: Two Synchronization Primitives and Predicate Width Fix Under Review

The Ascend changes under review in the main repo are concentrated in two places: #4396 lowers any_sync / all_sync to masked warp vote instructions, which previously had no corresponding implementation on the Ascend side; #4397 makes the backend reject unsupported SIMD predicate widths rather than reporting an error only after compilation has gone deep. In the same period, #4389 adds serial any_of / all_of lowering for the CPU backend, part of the same batch of fixes that “align frontend promises with backend capabilities.” Together with the five dtype audit issues listed in the second item of today’s highlights, the Ascend direction has formed three parallel tracks within 24 hours: “issue inventory, fixes under review, and release-channel buildout.”

Source: https://github.com/tile-ai/tilelang/pull/4396 , https://github.com/tile-ai/tilelang/pull/4397

2.2 Proposal for a Dedicated Ascend Nightly Wheel Pipeline: With Real-Hardware Smoke Tests and Rolling Releases

#3924, under review, adds a dedicated nightly build pipeline. The motivation is that the existing nightly release produces only a single fat wheel with CUDA, ROCm, and Ascend enabled, and its tag is attached to the CUDA backend (e.g., cu130): Ascend users can neither get a package reflecting their own stack nor a stable download link, and the build artifacts have never been run on real NPUs. The proposal uses cibuildwheel on hosted runners to produce an Ascend-only manylinux package, with CUDA and ROCm disabled at build time; because the Ascend side compiles against the ACL declarations shipped with the repo and a dynamically loaded ascendcl stub, the build machine does not need CANN installed, and the resulting wheel can be imported on machines without CANN. The pipeline triggers daily at 20:00 UTC, staggered from existing releases, and supports manual triggers and path-filtered PR triggers; the version tag is generated by an extension to the version script. The proposal adds 617 lines across 3 files, and the author has named two maintainers for review.

Source: https://github.com/tile-ai/tilelang/pull/3924

2.3 MLIR Ascend: tl.print and tl.and Merged; About 30 TLIR Operator Proposals Under Batch Review

The MLIR rewrite track merged two changes that day: #231 adds tl.print to the TLIR dialect and lowers it to the HIVM debug operator, covering both scalar-print and buffer-print sources, with the Python-side interface T.print, and includes round-trip, negative-verification, and lowering tests (all 5 pass); #218 adds tl.and (element-wise bitwise AND) following the existing binary-operator template, including dialect round-trip, lowering, and an expert-mode example. More notable is the review queue: within ten minutes, from 16:38 to 16:48, about 30 TLIR proposals were updated simultaneously, covering shifts, flips, concatenation, range generation, trigonometric functions, gather, transpose, vector activation, four kinds of reductions, L0A/L0B/L0C/L1 address spaces, atomic add, bitcast, cross-pipe set/wait flag, pad, sqrt and reciprocal, sort, type conversion, logical operations, and more. Individually these proposals are all small; collectively they fill in a layer of the TIR-side operator surface within the TLIR dialect, and are prerequisite work for whether the MLIR track can take on all operators of the main repo.

Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/231 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/218

2.4 Backend Repos: Ascend Reports a New 1D Strided Scatter Write Defect; Hygon and Sunrise Progress Still Under Review

The tilelang-ascend repo opened #2017 at 14:16 on 11-17, a defect report using a three-arm comparison: the environment is 910B3 with CANN 9.2.0, and the scenario is FLA’s WY fast backward, where 64 consecutive single-precision numbers in UB are scattered along the sequence dimension into GM, with the target tensor’s head dimension in the last dimension and adjacent elements’ addresses separated by the number of heads times 4 bytes (32 bytes when the number of heads is 8). The author compared three approaches—contiguous writes, direct strided writes, and strided writes relayed through GM—and traced the issue to three layers: incorrect compiler lowering parameters, a mismatch between hardware slot addressing and the compact layout, and a 32-byte slot footprint for narrow transfers that causes UB allocation to be underestimated. The contributor-credentials incident that drew attention the previous day (#2016) was closed on 11-16, with no new comments in the window. Hygon repo #27 (fixing intrinsic lowering, precision, and integer vector code generation; 592 lines, 14 files) has been continuously updated since 10-23 and remains open in the window; Sunrise repo #11 is the release merge for 0.1.15+sunrise.1.2.1, updating the three source trees TileLang, TileOPs, and TileKernels, adjusting Sunrise’s TANG/PTPU integration and public S2 continuous integration, and limiting host memory during large-input TopKSelect verification; the change totals 280k lines added and 179k lines deleted (including a full source-tree sync), and it was not merged by the end of the window, with the latest official release still 1.2.0 from 10-23.

Source: https://github.com/tile-ai/tilelang-ascend/issues/2017 , https://github.com/tile-ai/tilelang-hygon/pull/27 , https://github.com/tile-ai/tilelang-sunrise/pull/11

II·V. RISC-V Backend and Community (Special Watch: TVM Track / MLIR Track / SOPHGO TPU / Community Index Page)

2.5.1 Four-Source Inspection Shows No Change; References for Both Tracks Match the Previous Issue

The four-source inspection (the TVM track’s kubuds branch, the MLIR track’s RuyiAI repo, the community index page, and the Sophgo TPU extension) showed no new commits, no new PRs, and no index-table changes in this window: the MLIR track’s main remains at 2204c7e, the community index page’s main remains at 4499c2a, the TVM track’s mahiru branch remains at 33e0c89 and its karen branch at fc62939, and the Sophgo TPU repo’s main and feature branches have not advanced. The two tracks maintain their own pace, with no technical progress to report this issue.

Source: https://github.com/RuyiAI-Stack/tilelang-riscv , https://github.com/TongmingLAIC/tilelang-riscv

III. Ecosystem and Adopters (TileFoundry / topa / Site / TileKernels / FlashQLA)

3.1 TileFoundry changes 303 files in one pass: terminology converges to a single spec, layout switches to CuTe algorithms

#349 is the largest refactor of the quarter: it adds terms.md as the single authority for terminology and naming rules (N1 through N7), with all other spec documents, development notes, and glossary policies pointing to it; Python and C++ names are unified per the terminology (unit, topology, mesh, mesh axis, placement, etc.), and previously overloaded old terms (level, axis, participant, rank, program) are renamed across code, C++, and docs. The layout algorithm is ported function-by-function from CuTe: integer tuples, strides, layouts, composed layouts, and swizzle are ported with the same names, same parameter order, and same semantics, and cross-checked against C++ CuTe, filling in the previously missing rank, take, and right inverse; the grid itself is modeled as a layout, and cell numbering is given by the coordinate-to-index mapping. The change touches 303 files, adds 10,312 lines, and deletes 9,243 lines; it is a breaking rename, and it also means TileFoundry’s intermediate representation spec has shifted from “grown out of the implementation” to “defined first, then implemented.”

Source: https://github.com/tile-ai/TileFoundry/pull/349

3.2 Two TileFoundry fixes: logical windows in local views and copy parameters paired by coordinate

#351 fixes window misalignment in participant-local evaluation: when a layout splits a logical axis into multiple physical modes (for example, a logical shape containing 16 is represented locally as two modes, 8 by 2), the local view treats the composed mode as a scalar dimension during reshape and directly reports an unrecognized dimension; when two participants split along axis 0, the local slice also applies the global 0-to-256 window to each 128-row shard. The fix switches to local projection evaluation by logical rank, converts slices of the full-width split axis into each participant’s local length, and continues to reject partial windows that span participants rather than silently reading incomplete data. #348 fixes coordinate-based pairing of kernel parameters during copies: code generation wraps each parameter into a flattened mode in memory order, while the copy operator pairs them by CuTe indexing with “first mode fastest,” so for multi-axis views such as 2 by 4 the offsets and coordinates no longer agree, causing the second output to be misaligned and two loop-carried cases to return transposed results; the fix routes operands of the same shape through the same layout view into the copy, covering all three emission paths: regular copy, async copy, and bulk async copy.

Source: https://github.com/tile-ai/TileFoundry/pull/351 , https://github.com/tile-ai/TileFoundry/pull/348

3.3 topa lands its round mechanism; 6 new issues form the next work queue

topa merges #27: a round “before the state graph” may now open its own PR directly and register discovered problems in this repo. Six new issues (#28, #30 through #34) were then opened within the window, each a work item with a scope description: thread-scope admission and concatenation offload under the topic graph, pricing slot gather and small-CTA grids according to actual hardware behavior, pricing indexed copies and resident CTAs plus staged copy splitting, pricing and offloading fused q/k normalization with interleaved MRoPE, packing 4-bit tensors to shared memory via byte-wise copy, sum offload across CTA clusters and evaluation of the already-placed split-K window, and finalization and pricing of the gathered-tile attention kernel. The authors and mergers of this batch are all pipeline accounts, reflecting a working style in which “agent rounds open topics automatically and human review closes them out.”

Source: https://github.com/tile-ai/topa/pull/27 , https://github.com/tile-ai/topa/issues/34

3.4 Two automated doc-site updates; no commits from adopter repos in the window

The main documentation site repo had two documentation updates committed by an automated account within the window (11-17 08:17 and 11:28), consisting of site rebuilds that follow changes in the main repo and operator library, with no independent technical information. On the adopter side, DeepSeek’s TileKernels and Qwen’s FlashQLA repos had no pushes within the window (latest pushes on 10-14 and 10-30 respectively), and there were no new adoption cases this period.

Source: https://github.com/tile-ai/tilelang.github.io

IV. Community, Tutorials, and Events

4.1 A quiet news window: no topical coverage in Chinese or English search, Hacker News, or arXiv

Searches of Google News in Chinese and English (TileLang, tile-ai, and Ascend combinations), Hacker News, and arXiv over a 24-hour window turned up no reports or papers directly related to TileLang: the English side hit only one industry piece mentioning DeepSeek and Huawei, and the Chinese side hit two Ascend-related industry pieces, none containing TileLang content, so they were excluded on topical relevance; the latest arXiv paper with tilelang in its title or abstract is still the 11-03 disaggregated-inference KV compression work, which falls outside the window. The Hacker News entry in the candidate list (a Markdown flashcard app) and the third-party Hexagon backend repo hit by GitHub search (xwhzz/tilelang-hexagon, 4 stars) are keyword false matches and were not included.

Source: https://news.ycombinator.com/

4.2 Organization activity: 20 merges and 77 updated open PRs in a single day; main repo at 8,992 stars

Organization-level ground-truth queries show: 20 merges within the window (TileOPs 12, TileFoundry 3, tilelang 2, MLIR Ascend 2, topa 1), 77 open PRs updated, 15 new issues, and 10 repos with pushes. The main tilelang repo has 8,992 stars, the TileOPs operator library 203 stars, TileFoundry 147 stars, and MLIR Ascend 30 stars; the main repo’s release is still v0.1.15 from 9-30, and the operator library is still at v0.1.0 from 10-27, with no new tags in the window. The merge-to-open ratio is about 1 to 4, with the open queue concentrated in TLIR operator fill-in for MLIR Ascend and performance-related proposals for TileOPs.

Source: https://github.com/tile-ai

V. Trend Observations

  • Interface delivery phase: pilot-model-driven operator fill-in has moved from “registering interfaces” to “delivering one shape at a time,” with 8 operator proposals carrying reference implementations and benchmark rows merged in a single day; the acceptance bar is independent of the merge gate, and most proposals are merged despite self-assessed results below the acceptance bar, with further performance work queued as issues (#2999 has already provided a counterexample list at 0.39x).
  • Benchmarking as a platform: benchmark timing switches to nvbench cold measurement, preparation conventions are written into result fields, and the nightly sweep budget is doubled, showing that the operator library is treating “reproducible measurement” as part of the product rather than just a merge gate. The cost is that a single sweep now approaches 10 hours, and evaluation cost is becoming a process constraint.
  • Ascend is the most intensively hardened surface: main-repo fixes are in review, the MLIR route adds two layers of operators a day, a standalone wheel pipeline is proposed, and frontend dtype coverage is systematically inventoried (about 300 combinations found with gaps). Four lines advancing at once show Ascend moving from “it runs” to “deliverable and diagnosable.”
  • Spec first: TileFoundry uses a single 303-file refactor to converge terminology and layout algorithms into a single authority and hands layout computation to the ported CuTe algorithms; such changes raise review cost in the short term but determine in the long term whether the IR can be correctly implemented by external backends.
  • A generational shift in workflow: topa’s round mechanism lets the pipeline open its own PRs and its own issues, and all 7 new issues are executable work items. Humans shift from “finding work” to “reviewing and pricing,” consistent with the operator library’s commit cadence over the past two weeks.

Appendix: Sources and Verification Notes

Source Verification Table

Source Verification Result
GitHub tile-ai org, repo-by-repo check 30 repos, 10 pushed within window; actual main-branch additions: TileOPs 12, TileFoundry 3, tilelang 2, topa 1, docs site 2, MLIR Ascend 2 (merged)
Org-level merged-PR ground-truth search 20 merged, consistent with repo-by-repo main-branch check
Org-level open-PR search 77 open and updated; ~30 TLIR proposals in MLIR Ascend updated in a batch, the rest are performance and new-operator proposals
New issues opened 15: tilelang 5 (Ascend dtype audit), topa 6, TileFoundry 2, tilelang-ascend 1, TileOPs 1
Releases page tilelang v0.1.15 (9-30), TileOPs v0.1.0 (10-27), TileFoundry 0.0.4rc1, Sunrise 1.2.0 (10-23); no new tags within window
Adopter repos No pushes to TileKernels or FlashQLA within window
Google News (Chinese and English) No topical coverage; hits are industry pieces or false matches
Hacker News No topical entries in 24 hours
arXiv Latest paper mentioning tilelang is 11-03, outside window
TileLang × RISC-V four-source sweep No changes (no commits in TVM route, MLIR route, community index page, or Sophgo TPU)
Pending-inclusion queue No queue file for the day; no unarchived files containing TileLang entries found

Full Source List