Research window: 2026-11-06 07:00 ~ 2026-11-09 07:00 (Beijing time, 72 hours, covering Saturday and Sunday; the previous issue’s cutoff was 11-06 07:00).

Sources: Repository-by-repository check of the GitHub tile-ai organization (30 repositories, 11 pushed within the window) + ground-truth search of merged and in-review items (71 merged within 72 hours, 96 in-review updates, 50 newly opened issues) + commit-by-commit check of the main branch (tilelang 8, TileOPs 33, TileFoundry 5, topa 15, tilelang-ascend 7, tilelang-hygon 2, tilelang-mlir-ascend 1) + non-default branch check (the mlir-dialect and harness branches of mlir-ascend, the distributed branch of hygon) + release page check (tilelang / TileOPs / TileFoundry, no new versions within the window) + two adopter repositories + documentation site and nightly repository check + six groups of Google News queries in Chinese and English (via proxy) + Hacker News + arXiv. The RISC-V four-source special inspection found no changes over 72 hours.


This Issue’s Index

  • Today’s Focus: Three TileFoundry analyzer changes turn “placement semantics, TMA synchronization, and ordering of stale data before overwrite” into checkable objects; 33 TileOPs merges and 15 topa merges account for 68% of the organization’s 71 total merges, with performance work now driven by the nightly queue; Ascend sees 14 static defects in a single day coexisting with 3060 fully passing tests; on the news side, only Moore Threads’ Day-0 adaptation and one unverified V5 paraphrase
    1. Core Project Progress
      • 1.1 Mainline: CuTeDSL grid synchronization moved to a per-launch workspace, eliminating cross-talk under multi-stream concurrency (#3868) (11-08)
      • 1.2 Mainline: CuTeDSL signed integer truncation corrected to follow arith.divsi semantics (#3551) (11-08)
      • 1.3 Mainline: CuTeDSL closes example gaps across 63 files in one pass, including SM120 block-scaled MMA (#3685) (11-07)
      • 1.4 Mainline: CUDA single kernel launches can be named, making stages distinguishable in the profiler (#3816) (11-06)
      • 1.5 Mainline: Multi-source cache directories, read from any, write only to the first (#3807) (11-06)
      • 1.6 Ascend: Removed the Z3 scheduling global lock, cutting the slowest cold-compile case from 37.3s to 25.6s (#3812) (11-06)
      • 1.7 Ascend: access_ptr region fix, cross-row and transposed views no longer lose DMA dependencies (#3805) (11-06)
      • 1.8 Examples: sm_12x selects tiles based on available dynamic shared memory, RTX PRO 6000 fully passes (#3845) (11-08)
      • 1.9 New issue cluster: incorrect results remain the heaviest category, involving WS loop-carried dependencies, copy-loop atomics, and unsigned semantics (11-06 to 11-08)
    1. Multi-Backend Adaptation
      • 2.1 Ascend: Seven merges, Chunk O forward/backward and Sparse MLA SeeSaw operator migration landed (11-06)
      • 2.2 Ascend: Daily tests 3060 fully pass, 11-09 workflow job-layer failure (11-06 to 11-09)
      • 2.3 Ascend: 14 static defects in a single day, 13 fix PRs under review (11-07)
      • 2.4 Hygon: scaled cast and reduce lane swap merged, gfx928 declared supported but backend lacks branch (11-06 to 11-08)
      • 2.5 MetaX: Main branch only syncs upstream, fourth consecutive window with no proprietary changes (11-06)
      • 2.6 mlir-ascend: TLIR arithmetic operators continue to be ported, bf16 missing vadd/vmul flagged (11-06 to 11-07)
      • 2.7 Moore Threads: Repository silent, MUSA ecosystem side still lists TileLang-MUSA in migration path (11-06)
    1. Ecosystem and Adopters
      • 3.1 TileOPs: 33 merges, performance items dominate, each with A/B and target bars (11-06 to 11-08)
      • 3.2 TileOPs: SM89 completes FP8 variable-length attention, grouped GEMM, and decode sinks (11-07)
      • 3.3 TileOPs: Nightly performance gaps move from reports to wiki queue and project board (11-07)
      • 3.4 TileFoundry: Five merges, turning invariants into checkable objects; cross-CTA Partial under review (11-06 to 11-08)
      • 3.5 topa: 15 merges in two days, solidifying “one round” into three commands: hunt / round / review (11-07 to 11-08)
      • 3.6 Adopters and third parties: TileKernels, FlashQLA no pushes; Kimi K3 attention residual and V4.1 Flash snippets appear (11-08)
    1. Community, Tutorials, and Events
      • 4.1 Documentation site updated six times, nightly repo has no main-branch commits (11-06 to 11-08)
      • 4.2 News side: Moore Threads completes Kimi-K3.1 Pro Day-0 adaptation; Foro3D paraphrases DeepSeek V5 plans (11-06)
      • 4.3 Search surface: No Hacker News hits, no new arXiv papers in window (11-06 to 11-08)
    1. Trend Observations
  • Appendix: Sources and Verification Notes

Today’s Highlights

The output focus of this window is the “analyzer,” not the “generator.” TileFoundry merged five PRs in three days, three of which turned invariants that previously existed only in the author’s head into checkable objects: #291 brings shared operands of TMA copies into async-proxy semantics, pairing each TMA load with an mbarrier word, and rejects programs in cross-CTA scenarios where “one CTA performs async-proxy writes while another CTA reads the same global element without ordering”; #293 defines placement value semantics, writing the four states Split / Broadcast / Varying / Partial, the merge rule B ⊔ V = V, and the target validation for Reshard into a contract, while removing the no-longer-supported Dynamic attribute (64 files, +1165 to -536); #301 handles “regions that still need to read old data must be ordered before the schedule overwrites it,” recording the order in LivenessMetadata.orders and throwing AnalysisError when no order can be found. The other two (#290, #288) fix approximations in the analyzer itself.

The second item is pipelining. Of the 71 merges in 72 hours, TileOPs accounts for 33 and topa for 15, totaling 68%; 40 were submitted by the same author. Along with three CI changes in the same period (#2849 writes the nightly benchmark performance-gap queue into the wiki, #2852 syncs it to the Perf Tracker project, #2856 automatically moves caught-up operators to Done), performance work has formed a closed loop of “nightly benchmark → gap queue → single-operator round → review → merge”; the topa repo’s 15 PRs in two days (topa hunt, topa round start, topa round status, topa review prepare) are precisely the command-line shell of this loop. The efficiency gain is real, and so is the risk: a single-point author stacked with single-point review will degrade the pipeline into a one-person pipeline. Of the 96 PRs under review in 72 hours, 59 are still stuck in the main tilelang repo—review bandwidth is the next bottleneck.

The third item is on Ascend. On a single day, 11-07, one account opened 14 issues, with 13 fix PRs under review the same day: gemm_blockscaled’s positional parameters diverge from the public signature, causing policy to be moved into clear_accum (#3857); vcvt’s bf16-to-f16 swaps the round and sat parameters (#3829); vdupv silently maps an unknown pos to HIGHEST (#3832); fp4’s T.fill crashes with a divide-by-zero in SimdVF (#3852); the kernel cache key omits ASCEND_NPU_ARCH and the PTO toolchain, potentially reusing artifacts from one architecture on another (#3853). The common shape of these defects is “declared support, divergent implementation”: for architectures present in HCU_MCPU_SET, the C++ backend does not necessarily have a branch—the new Hygon gfx928 issue (#35) is the same shape. Meanwhile, all 3060 daily tests pass, showing that pass rates do not cover these combinations; test count cannot substitute for differential validation of shape combinations.

On the news side, only two items relate to TileLang. Moore Threads announced on 11-06 Day-0 adaptation of the 3-trillion-parameter Kimi-K3.1 Pro based on MTT S5000 and MUSA, with TileLang-MUSA explicitly listed in the migration path; the other is a 11-06 secondhand report from Foro3D claiming that DeepSeek will train an 8-trillion-parameter V5 on Ascend 950DT using the TileLang stack; DeepSeek has not officially announced this plan (the current latest is the V4.1 series), so it can only be recorded as a rumor. The four RISC-V sources show no changes in 72 hours, and both the TVM route and the MLIR route remain at citations from before 11-05.


1. Core Project Progress

1.1 Mainline: CuTeDSL grid sync moved to per-launch workspace, multi-stream concurrency no longer cross-talks (#3868)

Merged 11-08 18:00. CuTeDSL’s sync_grid() previously kept the barrier counter and generation in module-level globals (__grid_sync_ctr, __grid_sync_gen), so the same kernel launched concurrently on two streams shared one barrier: arrivals from both grids were merged, a launch could be released before all its own blocks had arrived, and leftover counts could hang subsequent launches. On H200 NVL a single launch took 17.4 ms and a pair 17.8 ms, confirming cooperative launch was indeed running concurrently. The fix replicates cooperative_groups::details::sync_grids with inline PTX: the barrier word is taken from the barrier field of the grid workspace the driver provides for each cooperative launch, with the address coming from %envreg1 (high half) and %envreg2 (low half); thread 0 of each block does one atom.add.release.gpu, block 0 adding 0x80000000 - (nblocks - 1) and the rest adding 1, then spins on ld.acquire.gpu waiting for bit 31 to flip. State is isolated per launch, so concurrent launches no longer affect each other. The regression test hangs on main (killed at 300 seconds) and passes after the fix; the author also kept the CUDA backend as a control, which was correct all along.

Source: https://github.com/tile-ai/tilelang/pull/3868

1.2 Mainline: CuTeDSL signed integer truncation fixed to arith.divsi semantics (#3551)

Merged 11-08 07:48. CuTeDSL emitted Python floor division for signed TIR DivNode, so at runtime T.truncdiv(A[i], 3) returned -3 instead of -2 for -7, and the subsequent floor/ceil corrections computed on top of that quotient, spreading the error. The fix emits tl.truncdiv (backed by MLIR arith.divsi) for signed scalars and vectors, uses exact integer arithmetic for pure Python inputs, and leaves the vector store path and unsigned lowering untouched. 25 regression cases cover int8/int16/int32/int64, scalars and vectors, negative divisors, integer extremes, and “quotient times divisor plus remainder equals dividend.” The reporter had independently verified this with a CPU reference and MLIR in the issue #2973 discussion and pointed out the boundary of pure Python %; that feedback went directly into the implementation.

Source: https://github.com/tile-ai/tilelang/pull/3551

1.3 Mainline: CuTeDSL fills example gaps across 63 files in one pass, including SM120 block-scaled MMA (#3685)

Merged 11-07 18:59, +7075 to -394, 63 files — the largest change this period. New coverage includes: warp vote, match and bit search, block-predicate barriers, explicit-width loads and stores with cache hints, NaN-propagating min/max, TMA and cluster copies (multicast, store_cluster, gather4, scatter4), SM120 block-scaled MMA (mxf4nvf4) and a matching NVFP4 GEMM example. The same change also fixes an old warp specialization problem: a T.fill placed before the pipeline was moved into the producer/consumer branches and then re-ran after every iteration even across loops or conditionals, overwriting that iteration’s T.clear result.

Source: https://github.com/tile-ai/tilelang/pull/3685

1.4 Mainline: CUDA single kernel launches can be named, stages distinguishable in the profiler (#3816)

Merged 11-06 17:45. Multiple T.Kernel in one PrimFunc previously shared an auto-generated name prefix, making stages indistinguishable in the profiler. A new optional name= parameter is supported by both T.Kernel(grid, threads=128, name="compute_stats") and T.ClusterKernel; the generated device symbol is compute_stats, with no _kernel suffix appended. The name travels through LowerOpaqueBlock via the tl.kernel_name annotation and is consumed by SplitHostDevice when generating the device function; explicit names are validated as non-empty ASCII identifiers, duplicate names and host-function conflicts go through the existing global symbol supply, and unnamed launches keep the old naming. Docs now cover naming, host call targets, invalid input, eager JIT, shared memory and cluster launches.

Source: https://github.com/tile-ai/tilelang/pull/3816

1.5 Mainline: Multi-source cache directories, read from any, write to the first (#3807)

Merged 11-06 13:31, +593 to -288, 18 files. TILELANG_CACHE_DIR now accepts multiple colon-separated directories, e.g. /cache/read/and/write:/cache/readonly/A:/cache/readonly/B; compiled artifacts can be hit from any directory but are only written to the first. This matches the common cluster pattern of “shared read-only cache + local writable layer,” letting a batch of builds share one operator cache. The author notes a boundary: the CuTeDSL backend may have problems on read-only directories.

Source: https://github.com/tile-ai/tilelang/pull/3807

1.6 Ascend: Z3 scheduling global lock removed, slowest cold-compile case drops from 37.3 s to 25.6 s (#3812)

Merged 11-06 18:17. Ascend’s Python scheduling callback serialized all Z3 solves behind one global lock, even though each call already held a private Z3 context. After removal, independent compile threads can schedule in parallel: in a cold-compile batch with 8 threads and 8 configs per kernel, SwiGLU backward dropped from 33.1 s to 30.0 s and MoE gate forward from 37.3 s to 25.6 s. Validation covered 884 existing Ascend auto-schedule CPU tests, 64 concurrent requests matching serial results, and 16 sampled SwiGLU backward and MoE gate configs producing identical schedules, source and device ELF before and after the change.

Source: https://github.com/tile-ai/tilelang/pull/3812

1.7 Ascend: access_ptr region fix, cross-row and transposed views no longer lose DMA dependencies (#3805)

Merged 11-06 12:08. When contiguous SIMD loads crossed rows or used a transposed view, the region rebuilt by access_ptr lost input DMA dependencies, so AutoSchedule could schedule the load before the copy and InsertSync would no longer insert the MTE2-to-V sync. The fix reuses conservative pointer region reconstruction: carry the start offset when crossing rows, fall back to the whole buffer for layouts that cannot be reconstructed, and keep the simple 1-D positive-stride fast path. Regression covers two cross-row cases and one transposed view, verifying the “DMA → set/wait → load” order and event ID alignment with and without AutoSchedule; 171 related host-side IR tests pass, and device execution was not tested (the author states this boundary explicitly).

Source: https://github.com/tile-ai/tilelang/pull/3805

1.8 Examples: sm_12x picks tiles by available dynamic shared memory, all pass on RTX PRO 6000 (#3845)

Merged 11-08 06:25. The dim=128 block-causal varlen case previously fixed a 128x128 tile, requiring 128 KiB for forward and about 192 KiB for dK/dV, while sm_86/89/120 cap dynamic shared memory at about 99 KiB, so launch failed outright with Failed to set the allowed dynamic shared memory size. The fix derives block_size from get_max_dynamic_shared_memory(): data-center cards keep 128x128, small-shared-memory devices use 64x128 (about 96 KiB), and block_size stays divisible by dllm_block=64. On RTX PRO 6000 (sm_120), CUDA and CuTeDSL each report 209 passed, 26 skipped, 0 failed; on H200 NVL (sm_90) the two 128x128 cases pass. The author calls this the last failing example left by #3701.

Source: https://github.com/tile-ai/tilelang/pull/3845

1.9 New issue cluster: wrong results remain the heaviest category, covering WS loop-carried state, copied-loop atomics, unsigned semantics

29 new issues were opened in the main repo during the window, the heaviest being the “computes wrong but reports no error” group. #3842: with warp specialization enabled, updates to a variable between two pipeline rounds take effect only in the consumer branch and are invisible to the producer, producing wrong results. #3870: inside a copied parallel loop, atomic operations and read-modify-writes on a fragment are performed once per copy. #3871: the tvm_ffi backend does not restore the caller’s current device after a call, so the caller subsequently launches on the wrong device. #3877 and #3881: under the NVRTC path a kernel with only one parameter cannot launch; under the cython path dynamic shared memory above 48 KB fails on the second device. #3866: Metal’s packed FP16 vector arithmetic operates on integer bit patterns; #3884 is under review (switching to half2). CuTeDSL unsigned semantics also produced three issues in a row: #3874 atomic max/min compare as signed, #3876 unsigned atomic comparison, #3875 values read from unsigned buffers should stay unsigned. The first two of these are the same type as #3808 and #3800 from the previous window: the compiler reports no error, and the generated code gives wrong results under specific combinations.

Source: https://github.com/tile-ai/tilelang/issues/3842, https://github.com/tile-ai/tilelang/issues/3870, https://github.com/tile-ai/tilelang/issues/3871, https://github.com/tile-ai/tilelang/issues/3877, https://github.com/tile-ai/tilelang/issues/3881, https://github.com/tile-ai/tilelang/issues/3866, https://github.com/tile-ai/tilelang/issues/3874, https://github.com/tile-ai/tilelang/issues/3876, https://github.com/tile-ai/tilelang/issues/3875


II. Multi-Backend Adaptation

2.1 Ascend: Seven PRs Merged, Chunk O Forward/Backward and Sparse MLA SeeSaw Operator Migration Landed (11-06)

The ascendc_pto branch merged seven PRs in a row on the afternoon of 11-06, focused on operator migration and toolchain: #1949 enables T.tile.bitwise_lshift to support tensor shift amounts on AscendC and PTO (previously it only accepted scalars); #1959 fixes the uint8 workspace view shape for PTO bitwise_xor; #1866 migrates Chunk O forward, #1802 migrates Chunk O backward, and #1869 migrates the SeeSaw implementation of Sparse MLA Forward; #1953 adds the fla_chunk_o example and tests; #1990 updates the CI notification recipients for the daily tests. The common thread across this batch is “moving FlashAttention-family operators already validated on the branch into the mainline,” rather than changing the backend foundation.

Sources: https://github.com/tile-ai/tilelang-ascend/pull/1949, https://github.com/tile-ai/tilelang-ascend/pull/1959, https://github.com/tile-ai/tilelang-ascend/pull/1866, https://github.com/tile-ai/tilelang-ascend/pull/1802, https://github.com/tile-ai/tilelang-ascend/pull/1869, https://github.com/tile-ai/tilelang-ascend/pull/1953, https://github.com/tile-ai/tilelang-ascend/pull/1990

2.2 Ascend: All 3060 Daily Tests Pass; 11-09 Workflow Job-Layer Failure (11-06 to 11-09)

Four daily test reports: #1993 (11-06) passed all 3032, #1995 (11-07) and #1996 (11-08) both passed all 3060 with a 100% pass rate; #1997 (11-09 05:50) reported Workflow test job failure, a GitHub Actions job-layer failure rather than a test-case failure, with the test-results artifact retained for 30 days and recipients assigned. Viewed alongside the 14 static defects in 2.3, the 3060 tests cover “recorded shapes” but not combinations that are “declared supported but implemented divergently.”

Sources: https://github.com/tile-ai/tilelang-ascend/issues/1993, https://github.com/tile-ai/tilelang-ascend/issues/1995, https://github.com/tile-ai/tilelang-ascend/issues/1996, https://github.com/tile-ai/tilelang-ascend/issues/1997

2.3 Ascend: 14 Static Defects in a Single Day, 13 Fix PRs Under Review (11-07)

Of the 29 new issues opened in the main repo on 11-07, 14 are for the Ascend backend, all from static audits by the same account, with 13 fix PRs under review the same day (#3834, #3835, #3836, #3843, #3844, #3849, #3850, #3851, #3854, #3860, #3861, #3862, #3863). Representative ones: gemm_blockscaled’s positional parameters diverge from the public signature, with policy moved into clear_accum (#3857); T.simd.vcvt swaps round and sat for bf16 to f16 (#3829), while also rejecting the int32/int64 to float16 conversion documented in the docs (#3833); vdupv silently falls back to HIGHEST for an unknown pos (#3832); pset accepts unsupported bit widths and emits pset_b4/pset_b64 (#3831); fp4’s T.fill divides by zero in ComputeVRegSize within SimdVF, causing a compiler SIGFPE (#3852); vgatherb declares the FP4 result register as twice the physical width (#3839); the kernel cache key omits ASCEND_NPU_ARCH and the PTO toolchain (#3853); T.gemm’s shape check uses a bare assert, which vanishes entirely under python -O (#3840); annotate_unlimit_memory rejects the "shared.ss" its own error message tells users to pass (#3846). There is also a structural issue: CodeGenTileLangPTO::VisitExpr_ is a 1441-line dispatch chain, with SIMD operator classification duplicated in seven places (#3848), a source of future maintenance cost.

Sources: https://github.com/tile-ai/tilelang/issues/3857, https://github.com/tile-ai/tilelang/issues/3829, https://github.com/tile-ai/tilelang/issues/3833, https://github.com/tile-ai/tilelang/issues/3832, https://github.com/tile-ai/tilelang/issues/3831, https://github.com/tile-ai/tilelang/issues/3852, https://github.com/tile-ai/tilelang/issues/3839, https://github.com/tile-ai/tilelang/issues/3853, https://github.com/tile-ai/tilelang/issues/3840, https://github.com/tile-ai/tilelang/issues/3846, https://github.com/tile-ai/tilelang/issues/3848

2.4 Hygon: Scaled Cast and Reduce Lane Swap Merged; gfx928 Declared Supported but Backend Lacks Branches (11-06 to 11-08)

Two feature merges: #33 adds scaled cast support, and #34 handles reduction lane swap and layout inference. Then #35 reports an architecture-level problem: K500SM_AI (gfx928, i.e., K100AI) is declared supported in HCU_MCPU_SET in tilelang/hcu/target.py, and determine_target("auto") correctly identifies mcpu=gfx928, but the C++ side in src/hcu/ (target_utils.cc, copy.h, reduce.h, scan.h, hcu_fp8.h, mls/*.hpp) generally has no gfx928 branches, so it is “declared supported but fails to compile in practice.” The reporter notes the key architectural differences: gfx928 lacks FP64 and some Matrix Core instructions and INT4, unlike gfx936. The two under review are distributed paths: #36 adds Block-Push versions of GEMM+AR and GEMM+RS, and #19 is an older PR on the Distributed branch.

Sources: https://github.com/tile-ai/tilelang-hygon/pull/33, https://github.com/tile-ai/tilelang-hygon/pull/34, https://github.com/tile-ai/tilelang-hygon/issues/35, https://github.com/tile-ai/tilelang-hygon/pull/36, https://github.com/tile-ai/tilelang-hygon/pull/19

2.5 MetaX: Main Branch Only Syncs Upstream, Fourth Consecutive Window Without In-House Changes (11-06)

The only push in the window advanced main from a35f8ddf to 64b52b4b, but the tip commit is upstream’s #3760 ([Language] Reuse macro transformation templates, committer date 11-05), i.e., an upstream sync with no in-house adaptation commits. This is the fourth consecutive window without in-house changes for MetaX; compared with Ascend and Hygon, the MetaX backend has been in maintenance rather than development during this period.

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

2.6 mlir-ascend: TLIR Arithmetic Operators Continue to Migrate; bf16 Lacking vadd/vmul Quantified (11-06 to 11-07)

The mlir-dialect branch merged #203, adding tl.sub, tl.mul, tl.div, tl.max, and tl.min to TLIR, along with TL-to-HIVM lowering and tests; also under review during the same period are #211 (tl.shl), #212 (tl.shr), #214 (tl.cast), and #218 (tl.and), migrating one operator per PR. The harness branch has three more (interactive onboard environment setup, project contract and delivery checks, project isolation and archive lifecycle). Two new issues quantify bf16 performance problems: #216 notes that bf16 lacks vadd/vmul, so every residual-add and weight-mul requires two extra fp32 passes, with FusedAddRMSNorm at only 50% versus 60% for fp16; #213 notes that bf16-to-fp16 casts are unsupported, bf16 vector operations run at half speed, and bf16 reductions are about 25% slower than fp16.

Sources: https://github.com/tile-ai/tilelang-mlir-ascend/pull/203, https://github.com/tile-ai/tilelang-mlir-ascend/issues/216, https://github.com/tile-ai/tilelang-mlir-ascend/issues/213

2.7 Moore Threads: Repo Silent, but MUSA Ecosystem Still Lists TileLang-MUSA in Migration Path (11-06)

The last push to tilelang-musa was on 11-05 (outside the window), with no commits in this window. However, in Moore Threads’ Kimi-K3.1 Pro Day-0 announcement on 11-06, the migration path explicitly states “completing the migration and integration of key compute paths for the new architecture through the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler,” indicating that TileLang-MUSA has entered the toolchain for actual model adaptation rather than existing only in the repository.

Source: https://news.qq.com/rain/a/20261106A0FH3900


III. Ecosystem and Adopters

3.1 TileOPs: 33 merges, overwhelmingly performance items, each with A/B and target bars (11-06 to 11-08)

TileOPs merged 33 PRs within 72 hours, making it the most active repository in the organization. Grouped by theme: on the elementwise and reduction side, #2869 gives float32 reciprocal two 16-byte vectors per thread; of 16 memory placements across 16 million rows, 12 previously lagged torch.compile (worst ratio 0.9917), and after the change none lag; #2867 splits a wave of comparison-row broadcasts into two rounds, halves the block count, and replaces the 16-bit ne with a single __hneu, fixing the old issue where logical_and/or read NaN as false (the author notes results fall below the target bar); #2851 removes the elementwise branch for bf16 and f32 division; #2862 makes float16 division round the way torch does on subnormal ties. On the normalization and attention side, #2866 changes the long launch of 8192-wide RMSNorm to two passes, four rows of 256 width per warp; #2846 stabilizes the statistics and variance reduction after shifting; #2848, #2854, and #2843 handle paged split residency, decode packing for SM90 variable-length GQA, and shared-memory footprint for gathered paged split, respectively. On the GEMM and quantization side, #2863 fills the SMEM ring for basic and split-K tiles, moves bias into the producer, and gives GEMV a deeper ring; #2864 reads group starts inside the kernel and selects tiles by group row count; #2865 uses TMA to load the FP8 A scale for 128-row tiles; #2853 splits K across CTA pairs for the M=128 o-proj FP8 GEMM; #2844 issues the 96- and 160-wide INT8 W8A8 tiles as a single WGMMA of the corresponding width; #2842 collapses INT8 per-tensor dequantization onto a branch-free path. On the storage and pooling side, #2868 has adjacent lanes exchange 16-byte int64 positions so each 16-byte store fills a 32-byte sector (vgg-block previously issued 900444 sectors for 501760 writes and stalled on drain at warp exit), and initializes from the first tap of the window, saving a compare and select. Maintenance also includes four cleanups and refactors: #2857, #2858, #2859, and #2850.

Sources: https://github.com/tile-ai/TileOPs/pull/2869, https://github.com/tile-ai/TileOPs/pull/2868, https://github.com/tile-ai/TileOPs/pull/2867, https://github.com/tile-ai/TileOPs/pull/2866, https://github.com/tile-ai/TileOPs/pull/2865, https://github.com/tile-ai/TileOPs/pull/2864, https://github.com/tile-ai/TileOPs/pull/2863, https://github.com/tile-ai/TileOPs/pull/2862, https://github.com/tile-ai/TileOPs/pull/2851, https://github.com/tile-ai/TileOPs/pull/2846, https://github.com/tile-ai/TileOPs/pull/2848, https://github.com/tile-ai/TileOPs/pull/2854, https://github.com/tile-ai/TileOPs/pull/2843, https://github.com/tile-ai/TileOPs/pull/2844

3.2 TileOPs: SM89 gains FP8 variable-length attention, grouped GEMM, and decode sinks (11-07)

Support for consumer and workstation cards was filled out in this window: #2855 enables the plain FP8 variable-length GQA kernel on SM89 (the warp specialization implementation remains SM90-only, declared via an explicit subclass), and the author provides full call-time and CUPTI GPU-time comparisons on RTX 6000 Ada with PyTorch 2.13.0+cu130, noting that “the candidate always runs before Torch, so execution-order effects are not controlled”; #2840 supports prefill sinks on SM89 while preserving the packed input ABI; #2845 supports dense decode sinks on SM89; #2841 supports aligned FP8 grouped GEMM on SM89. On the feature side, #2823 adds an optional bias to GEMMFwdOp, covering the dense, GEMV, and split-K paths; #2847 fixes input and RoPE precision for FP8 dense prefill; #2838 accepts TN single-element contractions with wide operand rows. These changes mean SM89 (RTX 6000 Ada, L40S and the like) is no longer an architecture that “can only run the CUDA reference implementation.”

Sources: https://github.com/tile-ai/TileOPs/pull/2855, https://github.com/tile-ai/TileOPs/pull/2840, https://github.com/tile-ai/TileOPs/pull/2845, https://github.com/tile-ai/TileOPs/pull/2841, https://github.com/tile-ai/TileOPs/pull/2823, https://github.com/tile-ai/TileOPs/pull/2847, https://github.com/tile-ai/TileOPs/pull/2838

3.3 TileOPs: nightly performance gaps go from report to wiki queue and project board (11-07)

Three CI changes turn “benchmark results” into an “actionable queue”: #2849 writes each night’s performance-gap queue into the wiki’s Perf Tracker; #2852 syncs that queue into a GitHub project board; #2856 automatically moves each operator to Done once it catches up. On the same day, #2860 is an example produced by this pipeline: when softmax input rows exceed L2, it switches to evict-last loads to avoid evicting rows still needed later. The value of this mechanism is not the automation itself, but turning “which operator is still behind by how much, and on which shape” into public, queueable, closable items, reducing performance work’s dependence on individual memory.

Sources: https://github.com/tile-ai/TileOPs/pull/2849, https://github.com/tile-ai/TileOPs/pull/2852, https://github.com/tile-ai/TileOPs/pull/2856, https://github.com/tile-ai/TileOPs/pull/2860

3.4 TileFoundry: five merges turn invariants into checkable objects; cross-CTA Partial under review (11-06 to 11-08)

In addition to the three covered in today’s highlights, there are #290 and #288: the former keeps isl relations exact over constant-product divisors, preventing scheduling analysis from misjudging due to integer approximation; the latter counts “completed partial reductions” in the cost, correcting the previous bias that priced partial reductions as incomplete. The PR under review, #292, is to support cross-CTA Partial reductions and split-K scheduling, dovetailing with the placement semantics defined in #293—#293 has confined cross-CTA Partial to allocation analysis, leaving reduction lowering, runtime reads of multiple gmem copies, and split-K scheduling to #292. If this line merges, TileFoundry will be able to express the common distributed-GEMM shape of “split a tensor by CTA, each writes a partial sum, then merge at the end.”

Sources: https://github.com/tile-ai/TileFoundry/pull/290, https://github.com/tile-ai/TileFoundry/pull/288, https://github.com/tile-ai/TileFoundry/pull/292

3.5 topa: 15 PRs in two days, crystallizing “a round” into three commands: hunt / round / review (11-07 to 11-08)

topa is a new repository in the organization, created only on 10-04, with its positioning stated in the README banner: “TileFoundry and TileOPs, a Path to Autonomous kernel engineering.” With 15 merges in two days, it mainly crystallizes the workflow: topa hunt pulls the latest night’s performance-gap queue from the wiki, topa round start <Op> starts a round of “one agent, one GPU, one operator,” topa round status shows each round’s agent and PR, and topa review prepare <PR> performs a review against the checklist. Supporting changes include: limiting the PR body to the Summary carried by the squash commit (the #7 series), attaching the operator’s prior attempt history to queue rounds (#16), reviewing the entire diff against the checklist before the PR is opened (#19), round stop --rm cleaning up the worktree created for the round (#21), and CI running unit tests and ruff on Python 3.11 and 3.13 (#14). This is not a compiler change, but it explains why TileOPs’ perf PRs could reach 33 in two days.

Sources: https://github.com/tile-ai/topa, https://github.com/tile-ai/topa/pull/5, https://github.com/tile-ai/topa/pull/16, https://github.com/tile-ai/topa/pull/19, https://github.com/tile-ai/topa/pull/21, https://github.com/tile-ai/topa/pull/14

3.6 Adopters and third parties: no pushes to TileKernels or FlashQLA; Kimi K3 attention residual and V4.1 Flash snippets appear (11-08)

The two official adopter repos saw no pushes in the window: deepseek-ai/TileKernels is stuck at 10-14, and QwenLM/FlashQLA at 10-30. On the third-party side, two new repos are worth mentioning: Index4NekoMeteor/Tilelang-AttnRes implements Kimi K3’s Attention Residual architecture in TileLang (forward only), with a test environment of Windows 11, 9800X3D, and RTX 5080; kevin3113/models-tilelang contains model-snippet examples for DeepSeek V4.1 Flash, in which L0’s “Attention output + FFN preprocessing” is chained from 7 ops-tilelang operators, performing post, GEMM with sum of squares, partial-sum reduction plus RMSNorm, and split-based mixing in sequence. There is also xwhzz/tilelang-hexagon, an upstream fork whose README is still the original TileLang text, with no verifiable Hexagon backend changes.

Sources: https://github.com/deepseek-ai/TileKernels, https://github.com/QwenLM/FlashQLA, https://github.com/Index4NekoMeteor/Tilelang-AttnRes, https://github.com/kevin3113/models-tilelang, https://github.com/xwhzz/tilelang-hexagon


IV. Community, Tutorials, and Events

4.1 Documentation site updated six times; nightly repo has no commits on main branch (11-06 to 11-08)

tilelang.github.io had six Update docs commits within the window (two on 11-06, one on 11-07, two on 11-08, plus one early-morning commit), corresponding to the kernel naming, multi-source caching, and CuTeDSL examples newly added to the main repo this round; TileOPs.github.io had one push. TileOPs-nightly had a push at 22:07 on 11-08, but no new commits on the main branch—this was a refresh of nightly benchmark artifacts.

Sources: https://github.com/tile-ai/tilelang.github.io, https://github.com/tile-ai/TileOPs.github.io, https://github.com/tile-ai/TileOPs-nightly

4.2 News side: Moore Threads completes Day-0 adaptation for Kimi-K3.1 Pro; Foro3D relays DeepSeek V5 plans (11-06)

Moore Threads announced on 11-06 that it had completed Day-0 adaptation for Kimi-K3.1 Pro (3 trillion parameters, 1 million context) based on the MTT S5000 and the MUSA software stack, covering model loading, service startup, and text/image inference pipelines; the official statement notes that the team handled the upstream implementation via SGLang-MUSA and migrated key compute paths using the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler, with the sparse attention DSA’s index selection, KV cache layout, and Prefill/Decode paths also within the adaptation scope. Public reports mention a validation configuration of 6 servers with 8 S5000 cards each, 48 cards in total; the announcement itself did not disclose throughput or concurrency figures.

On the same day, an English article from Foro3D claimed that DeepSeek plans to train an approximately 8-trillion-parameter V5 from scratch on Huawei Ascend 950DT (Atlas SuperPoD), and that its software stack uses TileLang, DeepGEMM-Ascend, and FlashMLA to raise MFU to 45%, with a procurement scale of 160,000 chips. It should be noted: DeepSeek has not officially announced a V5 plan; the latest publicly verifiable release remains the V4.1 series (V4.1-Flash and V4.1-Flash-1017 from 10-17), and third-party tracking sites also list V5 as unreleased, so this can only be recorded as a rumor, not as established fact.

Sources: https://news.qq.com/rain/a/20261106A0FH3900, https://foro3d.com/en/2026/noviembre/deepseek-prepara-su-v5-de-8-billones-de-parametros-sobre-silicio-huawe.html

4.3 Search coverage: no Hacker News hits, no new arXiv papers within the window (11-06 to 11-08)

Six Google News queries in Chinese and English returned 15 hits; after deduplication, only the two items in 4.2 were related to TileLang, with the rest being old articles on DeepSeek and Ascend or unrelated financial pieces. A Hacker News query for tilelang returned 0 hits within the window. An arXiv search for all:tilelang found the latest relevant entry to be the 11-03 KV compression paper (v2), with no new papers within the window.

Sources: https://news.google.com/, https://hn.algolia.com/, https://arxiv.org/


V. Trend Observations

First, the automation loop has taken shape, but its center of gravity has shifted from “writing code” to “defining contracts.” None of topa’s 15 commits contains a single line of kernel code—they are all round rules, review checklists, PR body formats, and worktree cleanup; most of TileOPs’ 33 commits are “change one parameter per the target bar, provide a set of A/B”; and TileFoundry’s 5 commits define what counts as “placement equality,” “TMA completion,” and “read before overwrite.” Stacked together, the three form a complete pipeline: nightly benchmarks surface gaps, the queue assigns operators, agents modify within TileFoundry’s scheduling space, review checks the numbers against the checklist, and the item is automatically closed after merge. The pipeline’s output efficiency is already visible in the numbers (71 commits in 72 hours), but it is extremely sensitive to “whether the contract is written clearly enough”—today #2867 was still merged despite being labeled “below the target bar,” and #301 chose to error out rather than guess when it could not determine an order; both are examples of the contract at work.

Second, Ascend’s bottleneck has shifted from “whether it can compile” to “whether declaration and implementation agree.” None of the 14 issues is “a feature is unsupported”—all are “declared supported but the implementation takes another path”: HCU_MCPU_SET has gfx928 while the C++ side has no branch (same for Hygon), T.simd.vcvt documents int32/int64 to f16 while the implementation rejects it, the annotate_unlimit_memory error message tells users to pass "shared.ss" while the function itself rejects it, and gemm_blockscaled’s positional parameters diverge from the public signature. The common fix for this class of defect is not adding features but putting “declaration” and “implementation” in the same table for generative checking—for example, generating from HCU_MCPU_SET the list of architectures the backend must cover, and generating from the intrinsic registry a dtype/bit-width support matrix. Until that is done, a test report with all 3060 items passing only shows that “the combinations tested are fine.”

Third, CuTeDSL is at the stage of “feature coverage approaching CUDA, correctness boundaries not yet converged.” #3685 adds 63 files in one go, filling in capabilities CUDA has long had—cluster TMA, SM120 block-scaled MMA, Philox, warp vote—showing that feature-surface alignment is nearly complete; but the defects in the same window are all at the boundaries: grid barrier under multi-stream concurrency (#3868), signed integer division (#3551), unsigned atomics and unsigned buffers (#3874/#3875/#3876). These are not “not implemented” but “implemented with semantics inconsistent with CUDA.” The practical advice for users: the CuTeDSL path is suitable for running examples and development; before switching a production kernel’s backend, run a round of element-wise diffing against the CUDA backend, especially in scenarios involving negative division, unsigned comparison, and multi-stream concurrency.

Fourth, the way performance optimization is recorded is changing. The hallmark of this batch of TileOPs PRs is that each one clearly states “which op, which benchmark, whether the result meets the bar, and what the root cause is”: #2868 gives the concrete numbers of 900444 sectors versus 501760 writes, #2869 gives the distribution of 12 memory placements with 12 lagging, and #2855 gives the median over 12 warm-up rounds with 10 calls each and proactively declares “execution-order effects not controlled.” This style lets “below the bar” still be merged (#2867), because readers can judge whether it is worthwhile. By contrast, a PR that only says “improves X%” cannot be distinguished at review between the kernel getting faster and the measurement changing.

Finally, the RISC-V line. The four-source inspection shows no change in 72 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 contrasts with the mainline’s rhythm: the agent-driven pipeline on the GPU side has compressed iteration to “days,” while the RISC-V side is still manual progress at “weeks.” The gap between the two lines is not in engineering effort but in whether there is a loop that can automatically discover gaps, automatically verify, and automatically queue.


Appendix: Sources and Verification Notes

Source Verification Table

Source Verification Result
tile-ai organization (30 repos) 11 repos pushed within the window; new repo topa (created 10-04) had 15 merges in two days
tilelang main repo 8 merges to main; 29 new issues; 59 under review (four clusters: CuTeDSL, Ascend, WS, JIT)
TileOPs 33 merges to main; 12 under review; no new issues
TileFoundry 5 merges to main; 14 new issues (mostly analyzer gaps); 1 under review (#292 cross-CTA Partial)
topa 15 merges to main; no issues; no release
tilelang-ascend 7 merges to ascendc_pto branch; of 4 issues, 3 are daily test reports and 1 is a workflow failure
tilelang-hygon 2 merges (#33, #34); 1 new issue (gfx928); 2 under review (#19, #36)
tilelang-mlir-ascend mlir-dialect merged #203; 3 merges to harness; 2 new issues; 5 under review
tilelang-metax Main branch only syncs upstream (tip is #3760); no own commits
tilelang-musa No pushes within the window (last 11-05)
Release pages (tilelang / TileOPs / TileFoundry) No new versions within the window: tilelang still v0.1.15 from 09-30, TileOPs v0.1.0 from 10-27, TileFoundry v0.0.3 from 10-21
TileKernels / FlashQLA No pushes within the window (10-14 / 10-30)
Doc sites (tilelang / TileOPs / nightly) tilelang.github.io had 6 Update docs; TileOPs.github.io had 1; nightly repo main branch had no commits
TileLang×RISC-V four-source sweep No changes (references across the four upstreams match the previous snapshot, 72-hour window)
Google News (six queries, EN/ZH) / Hacker News / arXiv Only two relevant news items (Moore Threads Day-0, Foro3D retelling); 0 HN hits; no new arXiv papers within the window

Full Source List

Verification Notes

  1. This report covers a Monday window spanning 72 hours from 2026-11-06 07:00 to 2026-11-09 07:00, including Saturday and Sunday; it does not overlap with the previous issue (cutoff 11-06 07:00).
  2. Merge ground truth is based on GitHub is:merged search; new main-branch commits are re-verified per repo via commits?since=. The tilelang-metax push is an upstream sync, and the tip commit’s committer date precedes the window start, so it is treated as “no own changes.”
  3. Only two news items relate to TileLang. The Moore Threads item is a retelling of an official announcement and includes a specific migration path; the Foro3D item is a secondhand retelling, and DeepSeek has not officially announced a V5 plan, so the report marks it “not officially confirmed” and does not use it as fact.
  4. All performance figures are taken verbatim from PR descriptions (target bars, A/B ratios, sector counts, timings) without secondary conversion; #2867 is marked “below target bar” in the original text, and the report preserves this as-is.
  5. The RISC-V four-source sweep showed no changes over 72 hours, and the relevant references match the previous snapshot, so no separate subsection is included.
  6. This report contains no local paths, proxy addresses, or intranet information; all links are publicly accessible addresses.