TileLang Weekly Report (2026-11-09 ~ 2026-11-13)
Reporting period: Monday, November 9, 2026 through Friday, November 13, 2026 — 5 business days, 5 daily reports; plus a supplementary 168-hour check of the 30 repos in the tile-ai organization on the publication date (November 14)
Part 1 · Brief
This section is a curated selection of the week’s key points, targeting 800 words or fewer by default (623 words this issue); full entries, itemized lists, and sources are archived in Part 2.
The five daily reports this week recorded 71, 21, 9, 6, and 5 merges. The main repo’s rhythm shifted from CUDA primitives early in the week to an all-Ascend cleanup by the weekend.
- PTO cleanup dominated the main repo: The eight merges across the 11-12 and 11-13 issues all landed on Ascend; the five merges in the 11-13 issue alone totaled +1903/−617, with four of them wrapping up #3512; the PTO default execution backend switched to tvm_ffi, and all-core synchronization now automatically enables batched scheduling to avoid deadlocks.
- Two audit waves pushed “silent failure” down to the instruction layer: Of the 30 issues opened on 11-12, 26 targeted Ascend; a different contributor opened eight more on 11-13, with the focus moving from parameter validation down to the
T.simdpredicate factory; a dozen or so validation PRs are under review in parallel. - ROCm became a series for the first time: Seven PRs opened within one hour on 11-13, spanning 22 files with +624/−19, covering compile parameters, launch bounds, occupancy, and register residency.
- TileOPs shifted from “three batches a day” to “awaiting review”: Merges across the five issues were 33, 18, 4, 0, and 0; most performance items fell short of their self-assessed targets, and the budget constraint was pushed down to SM89’s 101376 bytes.
- Version unchanged all week: The 0.1.16 version-number PR #3887 made no progress all week, and v0.1.15 has now been out for 44 days; TileFoundry completed a 1491-file reorganization.
- RISC-V four-source line silent for six consecutive windows: The Weft line outside the four sources closed at 553/553, but the methodology switched tables that day, so it cannot be read alongside last issue’s 0.0580×.
This week’s judgment: Ascend is now the absolute majority of main-repo merges, and the bottleneck has shifted from development volume to review bandwidth and “declared vs. implemented” mismatches; both the 1,900-line Ascend change and the 7,632-line harness content are stuck on branches, and unless 0.1.16 ships, downstream users won’t get them.
Part 2 · Archive
The following is the full text for the week, with structure and methodology consistent with previous weekly reports; every judgment in the Brief can be traced here to itemized evidence and sources.
Sources: The 5 issues of this publication’s daily TileLang community updates (11-09, 11-10, 11-11, 11-12, 11-13). The 11-09 issue covers a 72-hour window spanning the weekend (11-06 07:00 to 11-09 07:00, including Saturday and Sunday); the other four issues are contiguous 24-hour windows, so the factual interval covered by the five issues is 11-06 07:00 to 11-13 07:00; the 11-09 issue does not overlap with the previous weekly report (through 11-06 07:00). On the publication date, an additional 168-hour supplementary check of the tile-ai organization was performed (30 repos, 12 pushed within the window) to confirm each repo’s push and version status; it does not alter facts already recorded in the daily reports.
Editorial note: Organized across seven dimensions: the week’s top news, core project progress, multi-backend adaptation, ecosystem and adopters, community tutorials and events, trend observations, and appendix. Merge counts are per the windows recorded in each daily issue: the 11-09 through 11-12 issues use organization-level counts, while the 11-13 issue uses main-repo counts (that issue also had three merges on the mlir-ascend harness branch, not counted in the main-repo total). Newly opened issues across the five issues were 50, 6, 10, 35, and 18, totaling 119. Merges to non-default branches (hygon distributed, mlir-ascend harness) are noted branch by branch in the body and are not mixed into default-branch counts.
I. Weekly Highlights
-
The aftermath of the PTO vectorization overhaul ran through the entire week, with the last two batches’ eight merges all landing on Ascend (11-11 to 11-12): After #3512 replaced PTO’s vectorization and integer type system wholesale, the main repo’s merges spent four consecutive windows paying down regression debt. The 11-12 batch’s three PRs were
#3516(PTO target default execution backend switched totvm_ffi, 20 files +1549/−182),#3776(AutoSchedule tightens the cycle lower bound with three necessary conditions, Z3 solve time down 10.04%), and#3794(integer atomics adapted to signless carriers); the 11-13 batch’s five were#3614(cross-core syncmode=0forces batch scheduling, fixing deadlocks in all-core barriers and multi-stream core contention),#3894(VMI descriptor and logical coverage validation tightened, +925/−550),#3793(scalar cast facts guarded),#3796(explicit SIMD parallel loops preserved), and#3792(integer atomic sign-compatibility wrapper), totaling +1903/−617 across 22 file changes. The common shape of these four problem classes is not “new features computing wrong” but “programs that were legal before the overhaul are no longer legal,” with debugging cost high because the error location often lands on a scalar or address unrelated to the fault. Sources: https://github.com/tile-ai/tilelang/pull/3516, https://github.com/tile-ai/tilelang/pull/3614, https://github.com/tile-ai/tilelang/pull/3894 -
Two waves of systematic Ascend frontend audits pushed the “silent failure surface” from parameter validation down to instruction encoding (11-11 to 11-12): On 11-12, 30 new main-repo issues were opened in a single day, 26 of them with Ascend prefixes directly in the title, submitted in strings by two accounts (
cn-sys12,JINITAIMEI12113); the common thread is not “computing wrong” but “passing when it shouldn’t” —T.Kernelaccepts zero or negative block counts,T.alloc_l1accepts 2 GiB allocations,T.unrollaccepts −1 and emits#pragma unroll -1, negative tensor dimensions crash the compiler outright, hf32’snearest_zerocompiles to NEAREST_AWAY, and same-type ND-to-NZ scatter deadlocks because mask counts don’t fall back. On 11-13,NebulaMaoopened eight in a row within a dozen minutes, landing even lower:any_sync/all_syncleave unresolved calls after export, int64 cross-warpreduce_sumemits an undefinedHwReduce<SumOp,long>,update_mask/pintlv/pdintlvdon’t validate predicate widths,T.simd.pgequietly degrades topset_bXX, andT.ascend_pipe_barrier("PIPE_V")is silently dropped. Together, the two rounds lay out the Ascend frontend’s soundness gaps from the parameter layer all the way down to the instruction layer. Sources: https://github.com/tile-ai/tilelang/issues/3920, https://github.com/tile-ai/tilelang/issues/3992, https://github.com/tile-ai/tilelang/issues/3993 -
Seven PRs opened as a series on the ROCm side within an hour, the first systematic appearance since this publication began tracking ROCm (11-12): The seven were opened consecutively by the same author between 07:56 and 07:57, all based on
72d986e1, 22 files +624/−19:#3954makescompile_flagsactually reach hipcc (previously they only entered the cache key);#3955fixes 8-bit integer vector broadcasts emittingmake_uint2/make_uint4that hipcc rejects;#3956adds readable errors for small-tileT.gemm;#3957turns “minimum blocks per SM” on HIP from discarded into a launch bound based on waves per EU (on gfx90a an fp16 GEMM was locked to one block per CU due to 143 VGPR + 128 AGPR);#3958adds blocks-per-CU conversion to occupancy reports;#3959fixes 64-thread reductions on wave64 wastefully allocating a block of LDS that is never touched;#3960keeps fp16 fragments actually in registers (in HIPuint2/uint4are structs, and after SROA each lane eats 80 bytes of scratch). Six layers of coverage, where ROCm had long been “an occasional PR.” Sources: https://github.com/tile-ai/tilelang/pull/3960, https://github.com/tile-ai/tilelang/pull/3957, https://github.com/tile-ai/tilelang/pull/3959 -
TileOPs’ throughput and bar-met rate have completely decoupled, with shared memory budget becoming a hard constraint (all week): The five batches merged 33, 18, 4, 0, and 0 PRs respectively; the 11-10 batch’s 18 were split into three groups by time, 6 each, each group landing within 3 to 14 minutes, and of the 11 performance PRs, 10 self-rated
Result: below bar, with only#2882’s BMM ping-pong markedbar met— meeting the bar is not a prerequisite for merging; writing up the root cause and A/B is. The 11-11 batch’s four were all in DeltaNet, working around SM89’s 101376-byte shared memory for splitting and buffer reuse:#2891splits the forward pass by walk lower bound,#2890adds areuseswitch to the backward pass to compress 114944 back to 98560, and#2889splits tests by cluster and shared memory facts, cutting SM89 failures from 9 to 3. The 11-12 and 11-13 batches saw zero commits on the mainline for two consecutive periods, with only five performance PRs and onestatsbranch refresh in the window — in “awaiting review” rather than “in development.” Sources: https://github.com/tile-ai/TileOPs/pull/2882, https://github.com/tile-ai/TileOPs/pull/2890, https://github.com/tile-ai/TileOPs/pull/2891 -
TileFoundry turns invariants into checkable objects, then enforces the rules via a 1491-file reorganization (11-06 to 11-12): The 11-09 batch’s three merges respectively define placement value semantics (four states
Split / Broadcast / Varying / Partialwith the merge ruleB ⊔ V = V,#293), bring TMA copy’s shared operands into async-proxy semantics (#291), and record the ordering “old data must still be read before scheduling overwrites it” intoLivenessMetadata.orders(#301, throwingAnalysisErrorwhen no order can be derived). The 11-10 batch’s#312uses a single PR to reorder 1491 files (+41324/−42379), writing the five rule classes — imports, exports, naming, mutable state, layering — as S1 through S11 and C1 through C11, enforced in pre-commit with lint, ruff, and import-linter. The cost is that internal import paths change in this version, and scripts referencing private symbols need to follow. Sources: https://github.com/tile-ai/TileFoundry/pull/293, https://github.com/tile-ai/TileFoundry/pull/301, https://github.com/tile-ai/TileFoundry/pull/312 -
The backend matrix shifted from “one vendor advancing alone” to multiple fronts starting simultaneously, with four depths of investment coexisting (all week): Hygon’s
distributedbranch merged#36, the distributed trio (Block-Push GEMM+AllReduce, Block-Push GEMM+ReduceScatter, Block-Pull AllGather+GEMM, 6 files +2022/−0), while another branchtilelang-tileops-fixsaw ten PRs in a day adding hipRAND random numbers and cache_warmup; mlir-ascend’sharnessbranch merged three in a day +7632/−24 (Ascend950 simulation trace workflow, operator design guide and Roofline calculator, BMM and w4a8 examples), while#222in the same period narrowed the PR gate to a single Ascend 950 AscendC baseline; MetaX pushed its SDK companion release tov0.1.12+musa.5and raised the glibc floor to 2.34; third-partyxwhzz/tilelang-hexagonbrought TileLang to Qualcomm’s Hexagon NPU, with LFM2-1.2B running end-to-end on the Snapdragon 8 Elite’s v79 cDSP. Vendor repos track versions, distros track mainline, forks track their own hardware — the divergence itself is an indicator of dependency depth. Sources: https://github.com/tile-ai/tilelang-hygon/pull/36, https://github.com/tile-ai/tilelang-mlir-ascend/pull/221, https://github.com/xwhzz/tilelang-hexagon -
Version untouched for a week, with 1900 lines of Ascend changes and 7632 lines of harness content all stalled on branches (11-09 to 11-13):
#3887([Release] Bump version to 0.1.16) has seen no progress all week since its 11-09 03:07 update and remains unmerged; the main repo’s latest tag is stillv0.1.15from 09-30, now 44 days old as of this Friday. None of the four repos saw a new version in the window: TileOPs atv0.1.0from 10-27, TileFoundry atv0.0.4rc1from 10-25, tilelang-musa atv0.1.15+musa.1from 09-30. The Ascend execution backend and regression cleanup merged this period will, at the earliest, reach downstream distribution packages with the next tag. Sources: https://github.com/tile-ai/tilelang/pull/3887, https://github.com/tile-ai/tilelang/releases -
RISC-V’s four sources silent for six consecutive windows, while CPU lines beyond the four close out with a methodology switch (11-07 to 11-13):
kubuds/tilelang-riscv-kubuds’s twokubuds-branches,RuyiAI-Stack/tilelang-riscv’s MLIR route,TongmingLAIC/tilelang-riscv’s index page, andTongmingLAIC/tilelang-tpu’s SOPHGO branch all remained at their original references for six consecutive windows. Beyond the four,Tianchen-ai/tilelang-weftswitched its release tag toweft-parityon 11-11: all 553 items have matching native Triton references and satisfy “no more than 1.03x,” with the four platforms at 133/137/136/147 totaling 553/553; for the 543 exception cases referencing partial references, the median ratios are SG2044 1.3598, K1 1.3176, K3 A100 1.6569, x86 1.2917. It must be emphasized: these numbers cannot be read alongside the previous issue’s old table (41 cases faster than reference, SG2044 median 0.0580x) — the jump comes from the measured object changing from “author-written programs” to “shared Canonical compiler output.” Sources: https://github.com/Tianchen-ai/tilelang-weft, https://github.com/Tianchen-ai/tilelang-weft-bench
II. Core Project Progress (main repo tilelang)
Versions and Releases
- The main repo’s latest tag and release remain v0.1.15 (released 09-30); neither the five-period daily reports nor the release-day checks found a new tag, and as of this Friday it has been 44 days.
#3887([Release] Bump version to 0.1.16) has shown no activity since its 11-09 11:07 update, and was logged as “still pending” across all four periods on 11-10, 11-11, 11-12, and 11-13.- The organization’s other three repos saw no new versions in the window: TileOPs
v0.1.0(10-27), TileFoundryv0.0.4rc1(10-25), and tilelang-musav0.1.15+musa.1(09-30). - Structural observations related to versioning:
#3490and#3605landed CUDA primitives directly on the main branch, while#3903(process-parallel compilation) and#3897(CPU-side cumsum/cummax) are still under review — indicating that 0.1.16’s feature surface is still thickening rather than closing out.
Compiler, Language Layer, and Execution Paths
- PTO execution backend replaced (#3516, merged 11-11, 20 files +1549/−182): With
target="pto"andexecution_backend="auto", resolution now goes totvm_ffi; shared host codegen handles MakePackedAPI and the C host, and the Ascend runtime launches the aibin produced by PTODSL viaaclrtLaunchKernelWithHostArgs, matching the path used by AscendC kernels. On the device-build side, a new module-level toolchain callback compiles the entire PTODSL module in one pass (ptodsl → ptoas → bisheng fatobj-link → extract.aicore_binary), with the cache key incorporating ptoas arguments, ptoas version, and CANN identity. The old path is retained asexecution_backend="pto", and the two sets of outputs are bit-identical; validation covers four new test groups: dynamic shapes, multi-device kernels, mixed cube/vector GEMM, and cache hits plus corrupted-entry error raising. - AutoSchedule tightens the cycle lower bound (#3776, 1 file +20/−1): Instead of bisecting from a fixed lower bound, it now takes the maximum of three necessary conditions — the single-task bound uses the maximum of each task’s initiation interval; the resource-clique bound sorts pairwise-conflicting tasks by residual value and derives, from adjacent differences and the head/tail differences, that the cycle is no less than the sum of initiation intervals within the clique; the self-dependence bound solves
S_u - S_u >= wait - d*Pforceil(wait/d), with version-related distances scaled bynum_stages. Empirically, on Ascend 950PR with the CANN 9.3.0 weekly build: all Ascend examples compile and scheduling results are unchanged, while cumulative Z3 solving time drops 10.04%. No additional lower-bound probing was introduced, and the scheduling model was not changed. - Cross-core synchronization forces batch scheduling to prevent deadlock (#3614, 5 files +102/−1, of which 78 lines are regression):
mode=0means “launch as soon as cores are available,” whilemode=1means “wait until all core resources the operator needs are in place, then launch as a whole.” An operator doing all-core synchronization, if launched in normal mode, may get only some cores and run ahead to an all-core barrier and wait there — while the cores it awaits are held by another stream, forming a resource cycle and hanging the device. The fix detects “cross-core synchronization with a constant 0” on the codegen side and automatically sets__schedmode__(1), resetting the flag per function to prevent leakage. It also narrows the integermode_idof cross-core set/wait to 0/1/2/4 and documents that the three targets do not share the same mode set (910B/910C support 0/1/2, and only 950 adds 4). - VMI descriptor and logical-coverage validation closed out together (#3894, 6 files +925/−550): Fixes incorrect base address and stride when constructing descriptors for typed-view gather/scatter and single-element load forms; it also removes the frontend’s redundant logical-coverage rejection — removing the check is not a relaxation, but rather because the shared-memory detector fix (
#3805) has landed, and a second frontend block would falsely reject cross-row, transposed, and dynamic views. - Two sets of instruction-level primitives added on the CUDA side (merged 11-10):
#3490(6 files +185/−1) adds L1 prefetch, uniform L1 prefetch, and L2 prefetch, and makes eviction priority a first-class parameterEvictPriority(six levels: default, normal eviction, evict-first, evict-last, no-change, no-allocate), with non-default priorities requiring compute capability 8.0 or above;#3605(7 files +203/−1) addsT.clamp_symmetric(x, limit), which on SM86 and above clamps FP16/BF16/FP32 uniformly to[-limit, limit], preserving signed zero and subnormals, returning canonical NaN when any operand is NaN, lowering directly to a singlemin.NaN.xorsign.abs, and using the appearance of three scalar and packed forms in SASS as acceptance evidence. - Two preemptive hardening changes on the TIRx side (under review 11-11):
tile-ai/tvm’s#94([Fix][TIRx] Make ceildiv exact for signed divisors) and#93([Arith] Guard negative scales in canonical split div/mod) fix two ends of the same class of problem — once signs enter integer division and modulo, the rewrite preconditions for canonicalization no longer hold. The main repo opened#3937the same day (nested floor division produces wrong integers under negative multipliers) pointing at the same spot, except that#3937is a user-side reproduction while#93/#94are compiler-side preemption.
Correctness and Bug Fixes
- Silently wrong code remains the heaviest category in the main repo: The new issue cluster in the 11-09 period centers on “computes wrong but raises no error” —
#3842(with warp specialization enabled, variable updates between pipeline rounds take effect only in the consumer branch),#3870(in a copied parallel loop, fragment atomic operations and read-modify-write are each performed once per copy),#3871(tvm_fficalls do not restore the caller’s device on completion),#3877(single-parameter kernels cannot launch on the NVRTC path),#3881(on the cython path, dynamic shared memory over 48 KB fails on the second device); CuTeDSL’s unsigned semantics generated three issues in a row (#3874/#3875/#3876), the same type as#3808and#3800from the previous window. - CuTeDSL’s feature coverage and correctness boundary diverge in opposite directions:
#3685adds 63 files at once (+7075/−394), including warp vote, block predicate barriers, explicit-width load/store with cache hints, TMA and cluster copies, SM120 block-scaled MMA and NVFP4 GEMM examples, and incidentally fixes the long-standing problem ofT.fillplaced before a pipeline under warp specialization re-running every round; meanwhile the defects of the same period are all at the boundaries — grid barrier under multi-stream concurrency (#3868, switched to a barrier word in the grid workspace of each cooperative launch), signed integer division (#3551, now emittingtl.truncdivperarith.divsisemantics), and unsigned atomics and unsigned buffers. - Metal shifts from adding operators to fixing correctness boundaries:
#3907reports a silent out-of-bounds — when copying an 8x8metal.simdgroupmatrix to an 8x1 output, the compiler emits a fullsimdgroup_storewith row stride 1, overwriting the seven elements after the output (Apple M4,tvm_ffi, compilation cache disabled, detected with spare storage as a sentinel). The fix#3912adds no runtime check; instead, before emitting each SIMD-group store it requires the arithmetic analyzer to prove that the full 8x8 footprint falls within both the destination buffer and the requested region, and raises a readableValueErrorat lowering time if it cannot. - Structural limitations of Ascend PTO:
#3901reports duplicate physical segment errors when multiple explicit Cube/Vector regions are declared in the same kernel, with no fix PR yet;#3909extends the host-mapped buffer contract established by#3802on the CUDA side to Ascend (14 files +529), whereas previously Ascend kernels uniformly rejected CPU tensor arguments. - Three new CUDA-side issues, each representative (11-12):
#3988reports that fp16T.gemmwith M=1 and odd K still issues an illegal 2-byte transfer through pipelinedcp.async, reproducible on both 0.1.15 and main, while the regression form originally fixed by#2365passes;#3987reports thattl.LayoutInferenceandtl.LowerTileOpgrow exponentially with reused subexpressions inT.Parallel(the corresponding fix#3997is already queued; the same regression went from 34.89 seconds, over the 10-second compile limit, to passing);#3996reports thatT.Pipelinedwith explicit order/stage and copies routed through fragments raises “Fragment buffer access with non-zero index [1]”, though that index does not appear in the user program. - TileFoundry’s pricing bias repeatedly pushed back by the operator side (11-10):
#318uses elementwise clamp for placement pricing on H200, pointing out that once the CTA grid exceeds one wave, the footprint of the split input is multiplied by the number of waves, and noting that the same behavior was reported in four closed issues, all marked not planned, and still reproduces in the current build;#319lays out a 4096×7168×16384 persistent SM90 NT GEMM and gives five findings at once, the hardest being thatcheck --helpsays bf16 has 8 explicit mantissa bits with a rounding bound of 2^-9, whereas bf16 actually has 7 explicit mantissa bits and unit roundoff 2^-8, so the rtol derived from it is only half of what it should be. The analyzer’s 26% pricing bias (962.1 vs. 1213.6 GFLOP) will not slow down an individual operator, but it will cause automatic schedule selection to systematically choose wrong on boundary grids.
CI, Testing, and Toolchain
- Two overhauls of the CI feedback loop:
#3911(1 file +2/−2) removeslabeled/unlabeledevent types from the triggers — CodeRabbit applies multiple labels to a new PR at once, which previously produced several identical full CI runs within seconds, each cancelled by the concurrency group, while no CI job reads PR labels; the same change raises the ROCm runner watchdog above the workflow timeout so AMD jobs are no longer killed at 40 minutes. - Ascend daily test scale reaches 3060 items: The four reports in the 11-09 period were 3032 items all passing (
#1993) and two of 3060 items all passing (#1995,#1996), plus one workflow job-level failure (#1997, not a test-case failure); one run of 3060 items all passing at 11-11 07:58 (#2003); and one run of 3060 items all passing at 11-10 11:38 (#2001). The test scale has reached 3060 items while merges have been zero for several consecutive windows — the bottleneck is review, not development. - Alert deduplication gap:
#2006(10:23) and#2010(15:48) on 11-12 point to the same workflow run 43124767679, both with status “tests encountered an error,” indicating that alert deduplication is not keyed by run number. Compared with the 3060-item all-pass run on 11-11, this can only be logged as “the Ascend repo’s CI had one failure that day, reported repeatedly,” and the failure list is not yet readable. - Defect in the cache-hit path:
#2002([Bugfix] Restore adapter library path when loading cached kernels) is under review; it fixes the adapter library path not being restored when loading kernels from cache — a problem that only appears on the hit path, of the “works the first time, but the library can’t be found when loading from cache the second time” kind.
III. Multi-Backend Adaptation
Ascend (tilelang-ascend and PTO)
- Repo-side cadence: since the seven merges on 11-06, the default branch
ascendc_ptosaw zero merges across four windows from 11-09 to 11-12, while the review queue kept thickening (20-plus items in the 11-09 period, 15 in the 11-10 period, 9 in the 11-11 period, and more than ten in the 11-12 period). - Review themes shifted from “adding operators” to “adding toolchain interfaces and semantic consistency”:
#3892(torch.compile support for AscendC and PTO),#3893(direct dequantization of the int32 accumulator in L0C-to-GM copies),#3891(flushing the benchmark’s L2 withaclrtDeviceL2CacheFlush),#3894(PTO VMI descriptor fix); on the operator side,#1970(multi-head Gated MLA forward for both the PTO and AscendC lines),#1994/#2004/#1952(the fla_chunk inter, intra, and intra_token_parallel trio),#1605(Kimi Delta Attention decode and chunked prefill forward, targeting 910B),#1989(catlass bumped to v2.2.0). - On the examples and docs side, operator benchmarks are being added:
#1558(rms_norm, 20 cases in cann-bench),#1748(NMS, 20 cases),#1607(GroupNorm, 20 cases),#1677(sin/cos doc strings and tests),#1732(in-placeT.tile.silu/sigmoidsupported on both backends). - The seven merges from the 11-09 period remain on
ascendc_pto:#1949tensor shift amounts,#1959PTObitwise_xoruint8 workspace shape,#1866Chunk O forward,#1802Chunk O backward,#1869SeeSaw for Sparse MLA Forward,#1953fla_chunk_oexamples and tests,#1990daily test CI notification recipients.
Hygon (tilelang-hygon)
#36, the distributed trio, merged into thedistributedbranch rather thanmainat 11-12 16:20: Block-Push GEMM+AllReduce, Block-Push GEMM+ReduceScatter, and Block-Pull AllGather+GEMM, 6 files +2022/−0, with performance data and implementation notes in the README; which path to choose depends on communication domain size and overlap window.mainandv0.1.15-releaseshare the same reference, whiledistributedhas separately carried the 11-03 signal primitives, the 10-29 push-wave reduce-scatter example, and this period’s trio.#37([HCU] Support ds_read_vector Alt2 and inferred LDS layouts) opened at 11-12 11:21, adding the second encoding form ofds_read_vectorand LDS layouts inferable from constraints, on the same line as#34merged on 11-06 (XOR shuffle on the reduction side and layout constraint propagation).- The
tilelang-tileops-fixbranch landed ten commits between 11-12 15:41 and 21:14: hipRAND random numbers, the cache_warmup builtin, buffer loads rebase, tail unrolling under runtime trip counts, 8-bit integer vector broadcasting, cos/sin builtin dispatch, shuffle lowering and full-wave shuffle reduction, plus three compile-time fixes for__builtin_bit_castsize mismatch, fp64 truncation, andhexpambiguity. These ten commits are the branch content of the long-pending#27([Fix] Fix intrinsic lowering, precision, and integer vector codegen)—the Hygon side currently takes the form of “one long-lived PR carrying one development line.” - Two single-card capabilities merged in the 11-09 period:
#33scaled cast and#34reduction lane swap and layout inference; in the same window,#35reported an architecture-level issue—K500SM_AI (gfx928) is declared supported inHCU_MCPU_SETand is recognized bydetermine_target("auto"), but the C++ side undersrc/hcu/generally has no gfx928 branch, so it is “declared supported but fails to compile in practice”; the reporter notes that gfx928 lacks FP64 and some Matrix Core instructions and INT4. - The repo currently has 15 branches: six in the
feat/hcu-*series, onedistributed, and onetilelang-tileops-fix, covering three layers of work: instructions, communication, and operator compatibility.
mlir-ascend (MLIR rewrite route)
- The
harnessbranch merged three commits in one day (11-12), totaling +7632/−24 across 64 file changes, none of which landed onmainormaster:#220(14 files +1014/−12) Ascend950 simulation trace workflow (msprof collection, MindStudio Insight viewing, dumping GET/RLS events in the same run, exporting npusim SQLite for timelines; all 38 existing tests pass on WSL);#221(45 files +5178/−12) an installable version of the operator design guide (stage decomposition, SIMD/SIMT selection, storage budgeting, tile pipelining, with a Cube MFU and conditional Roofline calculator, plus snapshots of nine upstream TileLang Skill documents saved separately for offline reading);#226(5 files +1440/−0) Gemm quantization examples for BMM and w4a8, with a conv example also added in the same batch. #222(five workflow files +7/−9) merged intomasterat 10:27: it only changes theon:block of each workflow file, narrowing the PR gate to “clean-code lint + one AscendC line for Ascend 950,” with all self-built NVIDIA, AMD gfx942, and macOS legs removed. The implication for downstream is that the repo’s admission baseline now covers only one hardware generation, and regressions on older Ascend generations are no longer a merge condition.- The TL dialect added five operators in one day (11-12 17:49 to 05:45 the next day):
#228tl.xor,#229tl.pow,#231tl.print,#232tl.sort,#233tl.reshape, all built on theTL_BinaryOp<...>base class introduced by#203and rolled out following the templates of#218(and) and#225(or). By division of labor,tl.reshapeis a pure view operator (no destination operand, lowered tomemref.reinterpret_cast),tl.sortlowers tohivm.hir.vsort, andtl.printlowers tohivm.DebugOp. #217(4 files +140/−16) from the 11-10 period merged into theharnessbranch: saving evidence snapshots alongside hot knowledge records, forbidding the same approval ID from approving different entries, keeping cooled knowledge outside the temporary workspace, and prompting but allowing continued viewing when corrupted files appear in the knowledge directory; validation was 25 CPU tests passing.#219merged intoharnessin the 11-11 period;#223/#224were closed, and#225was resubmitted with identical content (localctest14/14).- The significance of filling out the dialect lies in coverage: the MLIR route must be able to absorb all common operators from upstream TIR; missing one means working around it at the TIR layer, and wherever you work around it is where behavior diverges from the main repo.
MetaX (tilelang-metax)
- The only push in the 11-09 period moved
mainfroma35f8ddfto64b52b4b, but the tip commit is upstream#3760([Language] Reuse macro transformation templates, committer date 11-05)—an upstream sync with no proprietary adaptation commit. This is MetaX’s fourth consecutive window without proprietary changes. There were no merges and no releases in the 11-10 to 11-12 window; the most recent push was 11-06 (#162onmain, merging upstreammainintodev).
Moore Threads (tilelang-musa)
tilelang-musahad no commits in the window, with the last push on 11-10; the latest tag remainsv0.1.14+musa.1(09-11). On the ecosystem side, however, it has entered the actual toolchain: in Moore Threads’ Kimi-K3.1 Pro Day-0 announcement on 11-06, the migration path explicitly states “through the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler, completing the migration and integration of key compute paths for the new architecture.”- Two documentation changes are worth noting:
#11bumps the MUSA SDK companion official version fromv0.1.12+musa.2tov0.1.12+musa.5and raises the C runtime requirement from glibc 2.28 to 2.34 (the companion TVM FFI wheel requires this), and clarifies that the SDK companion package and the open-source release are two separate channels that cannot substitute for each other;#12condenses the README installation section into a single set of commands (+10/−49), fixing the issue where “duplicate single-package install commands lead users to install only TileLang-MUSA and miss the TVM FFI.” The three version numbers (open-source tagv0.1.15+musa.1, SDK companionv0.1.12+musa.5, release page referencev0.1.13+musa.1) must be checked separately; looking only at the major version is not enough.
ROCm (HIP backend)
- See item 3 of “I. This Week’s Highlights.” Among the seven commits, the measured premise of
#3957is most worth recording: on gfx90a, a 256×128×32 fp16 GEMM compiles to 143 VGPRs plus 128 AGPRs with occupancy 1, so regardless of how much LDS remains, only one block per CU is possible; the root cause of#3960is that in HIP,uint2/uint4are structs, so vector copies into or out of local arrays degrade to memcpy, and after SROA the array still remains on the stack—on clang 18 / ROCm 6.2.4, an fp16 fragment that persists across loops and holds 2 rows of K and V per thread consumes 80 bytes of scratch per lane.
Metal
#3906([Metal] Resolve the auto execution backend to tvm_ffi) and#3804(region offset for 2D simdgroup fill) are under review, and#3912fixes the silent out-of-bounds in#3907. This is the third time Metal has appeared in the window this month, shifting from adding operators to adding correctness boundaries, which usually means someone has actually started running non-toy shapes on Apple silicon.
Sunrise (standalone distribution)
- The backend distribution repo’s
mainis still atfb7d3d62from 10-23, and the only activity in the window was the creation and deletion of a CI branch:ci/sunrise-0.1.15-1.2.1-r6created at 18:17:24 and deleted at 18:18:44, an 80-second interval. The branch name gives the version combination:0.1.15corresponds to the upstream baseline,1.2.1is bumped up from the1.2.0in the 10-19 candidate branch name, andr6is the sixth revision. This indicates the backend is still tracking the mainline and issuing revisions at its own cadence, though this period stopped at building rather than committing.
RISC-V Backend and Community (special observation)
- Four sources unchanged for the sixth consecutive window (recorded 11-07, 11-09, 11-10, 11-11, 11-12, 11-13):
kubuds/tilelang-riscv-kubuds(TVM backend route) has its twokubuds-prefixed branches parked atkubuds-riscv-karenfc62939 andkubuds-riscv-mahiru33e0c89, withmainonly upstream sync;RuyiAI-Stack/tilelang-riscv(MLIR rewrite,target="riscv") hasmainparked at 57c548c;TongmingLAIC/tilelang-riscv(community index page) hasmainparked at 4499c2a with no additions or removals in the table;TongmingLAIC/tilelang-tpu(SOPHGO TPU) hasmainparked at 303759f andfeature/sg2260e-rv-supportparked at 122bc59. Per the established convention, empty script output is recorded as “no change” and is not treated as a collection failure; in the 11-13 period, checking PRs on the index page hit anonymous API rate limiting and was downgraded to reference comparison only. - The Weft line outside the four sources (
Tianchen-ai/tilelang-weft): a fork of upstream tilelang v0.1.15 that adds the Weft CPU programming model—five categories of author information on top of the native TileLang AST: Encoding fields and bit layers, logical axes and tile domains, Levels and staged snapshots, grouped partials with finite bit widths and explicit conversions, and per-loop pipeline depth—then a shared set of physical passes covers both RISC-V and x86 (the same source construct enters either the RVV/IME or the x86 SIMD/AMX lowering path). In the 11-10 window the repo had 17 commits and the benchmark repotilelang-weft-benchhad another 23, concentrated on packed row decoding (Q1/Q2/Q4_K/Q6_K/IQ3/IQ4), IME1/IME2 record-level contraction, runtime K-dimension reduction, and dynamic rotary traversal; on 11-11 the release tag was changed toweft-parityand the results table and manual were rewritten. - The two sets of figures must be read separately: the old table from the 11-10 period is “150 operator IDs forming 553 platform cases, 543 with external references, numerical checks passing 553/553, only 41 cases faster than the reference implementation, SG2044 median 0.0580×”; the new table from 11-11 is “all 553 items have matching native Triton controls, each satisfying no more than 1.03× (SG2044 133/133, K1 137/137, K3 A100 136/136, x86 147/147), with median ratios against external references of SG2044 1.3598, K1 1.3176, K3 A100 1.6569, x86 1.2917.” The source of the jump is written in the commit title—”performance inheritance from a compiler shared across four platforms”—that is, the object under test changed from “author-written programs” to “artifacts of the shared Canonical compiler,” not the same implementation getting twenty times faster in a day. This must be stated clearly when citing.
- One disclosure where the control itself was wrong: the two Q1 controls for K3 previously used a retained Triton Canonical program that overflows when negating −128—the boundary input yields −2048, when the correct result should be 0. After the shared compiler that fixes exact I8 reduction regenerated it, artifacts from both frontends pass that input; the historical timing is retained in
triton_historical_us, and the correction note is written intriton_control_note. - The technical routes still need to be kept distinct: Weft takes the CPU route of early-importing the native TileLang AST 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 ofRuyiAI-Stack/tilelang-riscv; the three must not be conflated.
IV. Ecosystem and Adopters
TileOPs
- Cadence: Across five periods, 33, 18, 4, 0, and 0 commits were merged, making it the most active repository in the entire organization, but the main branch saw zero commits for two consecutive periods (head stuck at
9bf4bf56, 11-10 09:59), and the only write within the window was astatsbranch refresh at 23:17 (chore(stats): refresh manifest stats from 9bf4bf56...). The five performance PRs recorded in the previous period (#2894through#2898) did not appear in the in-review update list during the 11-13 window, indicating they were neither merged nor further advanced. - Capability distribution of the 33 commits in the 11-09 period: On the elementwise and reduction side,
#2869(float32 reciprocal with two 16-byte vectors per thread; of 16 million rows across 16 memory placements, 12 were originally behind torch.compile with a worst ratio of 0.9917, and after the change none lag),#2867(comparison row broadcast split into two rounds with block count halved, 16-bitnechanged to a single__hneu, fixinglogical_and/orreading NaN as true; the author marked it below the target bar),#2851,#2862; on the normalization and attention side,#2866(8192-wide RMSNorm long launch changed to two reads, four rows of 256 width per warp),#2846,#2848/#2854/#2843; on the GEMM and quantization side,#2863,#2864,#2865,#2853,#2844,#2842; on the storage and pooling side,#2868(adjacent lanes swap 16-byte int64 positions so each 16-byte store fills a 32-byte sector—vgg-block originally issued 900444 sectors for 501760 writes and was stalled by drain at warp exit). - Concentrated completion of SM89 support surface (11-09 period):
#2855enables the regular FP8 variable-length GQA kernel (the warp specialization implementation remains limited to SM90),#2840prefill sinks,#2845dense decode sinks,#2841aligned FP8 grouped GEMM; this batch means SM89 (RTX 6000 Ada, L40S and the like) is no longer an architecture that can “only run the CUDA reference implementation.” The author of#2855proactively states that “candidates always run before Torch, and execution-order effects were not controlled.” - Boundary fixes in the 18 commits of the 11-10 period:
#2878K=0 empty contraction reportsOpNotAvailableErroron SM89 (test_gemm.py -m "smoke or sm89"goes from 63 passed 6 failed to 66 passed 4 failed);#2876variable-length GQA fused RoPE hangs when the rotation half-width is padded to a value not divisible by the thread count (a newsplit_readsseparates “read plus rotate” and “write back” into two rounds);#2880float32 GLA backward requires 114688 bytes at chunk 64 while SM89 only has 101376, changed to pack the dv state term after the dA product down to 98304, with gradients bitwise identical to the default order. - The way performance figures are written is itself changing:
#2886(Softmax keeps the entire row in registers, one exponential per element) and#2871(exponential replaced with rawMUFU.EX2plus one row reciprocal, with a newexp_shifted_scaledensuring terms that would round to non-zero float32 remain normal numbers) report executed instructions down 23% at 512 columns and 21% at 2048 columns, bf16 at 512 columns 16.67 microseconds versus quack 16.35, and 2048 columns 62.99 versus 62.21;#2884(RMSNorm 64-thread block requests the SM’s 32-block residency cap from the compiler, otherwise ptxas gives 40 registers, H200 only retains 24 blocks, and short-row warp activity drops to 64%) together with#2870and#2885all point to the second-order parameter set of “how many registers per thread, how many blocks per SM, how many dispatches per block.” - Shared-memory closed loop of the four commits in the 11-11 period:
#2891forward at chunk 128 anddim_v64 requires a walk lower bound of 114688 (dim_v128 requires 131072) while SM89 only gives 101376, changed to fall back to the recurrence needing only 98560 when it does not fit;#2890backward w/u adds areuseswitch, where S, du, and Aw take over the buffers of dh, v_new, and Au after their last read, compressing 114944 to 98560, and additionally computes_parallel_live_bytesso thatrefusalalso makes SM90 reject calls that the three main-branch commits let through but that fail at runtime;#2889movescheck_held_launchbefore the per-architecture cross-marking, so on SM89 four test files go from 40 passed 9 failed 3 skipped to 35 passed 3 failed 14 skipped, and all 46 items pass on a single SM90. Performance numbers are recorded faithfully:#2891is 9% to 41% slower than FLAchunk_delta_ruleon SM89, and#2890backward is 1.530 milliseconds versus FLA 1.554 milliseconds (ratio 0.98). - Performance PRs in the 11-12 and 11-13 periods:
#2895changes the W4A16 tile split in two along K into a pair of CTAs in a cluster, receiving FP32 partials in the peer’s shared memory viast.async, reducingchunked-prefill-tilefrom 37.3 microseconds to 34.7;#2896layers a hand-written warp-specialized main loop on top, reducing it to 31.6 microseconds, 1.19 times the torch dequantized matmul;#2894changes DeepSeek-V3 prefill’sstage_scalesfrom 1 back to 0, with up-projection 269.3 → 262.0 microseconds and down-projection 142.2 → 137.2 microseconds;#2898marks the softmax two-CTA clusterbar met;#2897marks the long-K NT GEMM selectorbelow bar, quantifying that “cycle counts cannot be used as a criterion under the power wall”—the 176-wide tile uses 2.7% fewer cycles than DeepGEMM but moves 21% more operation bytes into the SM, and thus runs at a lower clock. - The nightly queue has solidified into a dashboard: three CI changes in the 11-09 period turn benchmark results into an executable queue—
#2849writes the performance gap queue into the wiki’s Perf Tracker,#2852syncs it into a GitHub project board, and#2856automatically moves items to Done once an operator catches up; at 11-11 19:52 the repository wiki’s Perf Tracker page was created (corresponding to one nightly sync). Thesnapshotsbranch at 11-13 04:33 lands onbench: 9bf4bf56... (run 43264767627), with the benchmark commit identical to the current mainline head, so that night tested the same unchanged code, and differences between two consecutive nights’ snapshots can only come from environmental fluctuation.
TileFoundry
- Beyond the five merges in the 11-09 period,
#290(isl relations remain exact under constant product divisors, avoiding scheduling analysis misjudgments due to integer approximation) and#288(counting “already-completed partial reductions” into the cost, correcting the previous bias of pricing them as incomplete) fix the analyzer’s own approximations. The in-review#292aims to support cross-CTA Partial reductions and split-K scheduling, dovetailing with the placement semantics defined by#293—#293has already 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 form of “a tensor split by CTA, each writing a partial sum, and finally combining.” - After the source-code specification reorganization in the 11-10 period,
zhen8838opened four in a row within one minute at 02:42 (#314through#317), all marked “found by reading code, not reproduced with a program”: the JIT cache key prints only the entry function and topology, with called functions appearing only by name, so two Modules with identical entry text but different callees get the same key; whentopology_levelis not given, it defaults to the first valid topology without asking which levels the target declaresmax_physical_unitsfor; loops with depth less than 4 advance the cursor by one loop-body duration rather than the entire loop span, so work after the loop does not wait for the loop to finish;passes.md§7.1 states that on-chip memory callee return values must be copied into the caller’s buffer or fail withLoweringError, butconvert_hir_to_tirhas no corresponding rejection implementation. - Two fixes in the 11-12 period are in review:
#323fixes the parser dropping slices starting at zero (when re-labelingthreads[0:16]withMeshatlayout=(16,), if the selection starts at offset 0, the parser drops the slice when constructing the new mesh, and after lowering nocontainsguard is emitted, so every thread in that layer runs this region; whereas writingwith threads[0:1] as writer:directly preserves the slice and behaves correctly);#324fixes dimension arithmetic in codegen (when loop bounds come from mesh coordinates it reportsno codegen handler for DimMul on cuda.Target; the shared scalar renderer handles comparisons but not dimension arithmetic, and the three CUDA emitters each maintain their own+ - * / %table).
topa (autonomous kernel engineering pipeline)
- 15 merges over two days in the 11-09 period solidify “one round” into commands:
topa huntfetches 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 statusshows each round’s agent and PR, andtopa review prepare <PR>performs review against a checklist. Supporting changes include limiting the PR body to the Summary carried by the squash commit, attaching the operator’s prior attempt records to queue rounds, reviewing the entire diff against the checklist before opening a PR,round stop --rmto clean up the round’s worktree, and CI running unit tests and ruff on Python 3.11 and 3.13. This is not a compiler change, but it explains why TileOPs’s perf PRs could reach 33 in two days.
Adopters and Third Parties
- Both official adopter repos had no pushes all week:
deepseek-ai/TileKernelsstuck at 10-14 (2072★),QwenLM/FlashQLAstuck at 10-30 (755★). - Third-party repos:
Index4NekoMeteor/Tilelang-AttnResimplements Kimi K3’s Attention Residual architecture with TileLang (forward only; test environment Windows 11 / 9800X3D / RTX 5080);kevin3113/models-tilelanghosts model fragment examples of DeepSeek V4.1 Flash, where L0’s “Attention output + FFN preprocessing” is chained from 7ops-tilelangoperators;Atlas-Ascend-Fork/ops-tilelang(created 10-20, 88 public repos) describes itself as “the CANN community’s TileLang high-performance operator repository for Ascend NPU,” with one push each at 11-09 19:57 and 11-12 14:23; the account name containsForkand is neither a tile-ai organization nor an official Ascend organization, so its official status cannot be determined from the repository itself and is recorded only as a lead. - Hexagon NPU backend (third-party fork):
xwhzz/tilelang-hexagonadds a TileLang backend for Qualcomm’s Hexagon NPU, developed and validated on the v79 cDSP of a Snapdragon 8 Elite phone, compilingT.copy,T.gemm,T.Parallel,T.reduce_*,T.Pipelinedinto HVX vector code, HMX matrix multiply, VTCM buffers, and DMA streams, executed on the NPU via FastRPC. LFM2-1.2B runs end-to-end on the NPU at batch 1, with prefill, decode, and sampling all written in TileLang; projections are 4-bit GPTQ (one fp16 scale per 64 columns), and the tied embedding and head are 8-bit; on-device WikiText-2 perplexity is 16.68 versus a bf16 reference of 16.29. Performance: decode 128 tokens at 80.1 token/s (bound 83 token/s, 96% of bound), prefill 1024 tokens in 176.2 milliseconds i.e. 5813 token/s (bound 129 milliseconds, 73% of bound), taken as the median of five runs ofbench.py, with timing including FastRPC call overhead and excluding tokenization and PC round-trips. The README also states that this backend and model were “rewritten by an AI coding agent under the maintainer’s guidance,” with 75 PRs filed between 2026-10-26 and 10-29, accounting for more than 80% of the current code, tests, and documentation.
V. Community, Tutorials, and Events
- Documentation sites:
tilelang.github.iohad sixUpdate docscommits from 11-06 to 11-08 (corresponding to this round’s kernel naming, multi-source caching, and CuTeDSL examples in the main repo); a check on the release date showed its most recent push was 11-11;TileOPs.github.iohad one push each on 11-09 and 11-10. Thesnapshotsbranch ofTileOPs-nightlyis written automatically every night, which looks very similar to real code commits in the “pushes within the window” statistic, so it needs to be distinguished by branch. - News side: zero hits across all five issues of the week: Google News Chinese and English queries (including groups such as “TileLang”, “TileLang OR tile-ai”, and “TileOPs or TileRT”) all returned zero within the window after filtering by publication time; the latest hits remain the 10-28 Caifuhao article and the 10-21 SemiAnalysis piece; Hacker News hits are all same-name noise (Emma real-time video persona, Vibe Slopped Cryptography, Senior Pipeline Crisis, LimbaKit, Monoprint, VoxaDub, tag classification tools, project ranking sites, etc.); on arXiv,
all:"TileLang"returns 10 results in total, and the latest is still the 11-03 disaggregated inference KV compression paper v2. Consecutive days of zero hits on the news side are normal for TileLang: project dissemination mainly goes through GitHub and developer communities, not news indexes. - 11-06 Moore Threads completed Day-0 adaptation for Kimi-K3.1 Pro: Based on MTT S5000 and MUSA, it completed adaptation for 3 trillion parameters and 1 million context, connecting the model loading, service startup, and text/image inference pipelines; the official description states that SGLang-MUSA is used to carry the upstream implementation, and the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler are used to complete migration of key computation paths; the index selection of sparse attention DSA, KV cache layout, and Prefill/Decode paths are also within the adaptation scope. Public reports mention that the validation configuration was 6 servers, each with 8 S5000 cards, 48 cards in total; the announcement itself did not disclose throughput or concurrency figures.
- The rumor from the same day needs to be flagged: The English article from Foro3D on 11-06 claimed that DeepSeek plans to train an approximately 8-trillion-parameter V5 from scratch on Huawei Ascend 950DT (Atlas SuperPoD), and claimed that the software stack uses TileLang, DeepGEMM-Ascend, and FlashMLA to raise MFU to 45%, with a procurement scale of 160,000 chips. DeepSeek has not officially announced a V5 plan; the latest publicly verifiable information is still the V4.1 series (V4.1-Flash and V4.1-Flash-1017 from 10-17), and third-party tracking sites also mark V5 as unreleased, so it can only be recorded as a rumor.
- Adjacent technology development (11-10): The Meta engineering blog published “How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale”, which is about Triton Low-Level Extensions—adding explicit warp-group and cluster-level orchestration on top of Triton’s block-based programming model, allowing developers to use execution-side capabilities such as warp specialization, asynchronous movement, and tensor memory without giving up Triton’s development speed; the article gives three production-running use cases (recommendation system training, recommendation system inference, and large-scale LLM pretraining) and a chart of the growth in the number of Meta’s internal kernels over 10 months (Triton kernel count grew about 3.8x, CUDA about 2.0x). It is another solution on the same track as TileLang: TLX extends downward from Triton with explicit orchestration, while TileLang extends upward from a tile-level DSL with automatic scheduling and multiple backends. This item is not a TileLang project development and is recorded only for comparison.
VI. Trend Observations
First, the cleanup period after a major overhaul lasts longer than the overhaul itself, and the cost lands on “programs that were previously legal are no longer legal.” After #3512 replaced PTO’s vectorization and type system together, four consecutive windows were spent patching its regression surface: execution backend default switch (11-11), integer atomics (11-11, 11-12), explicit SIMD parallel loops and scalar cast facts (11-12), and VMI descriptors (11-12). What these four categories share is not “new features computing wrong results” but “old code no longer compiles or produces different results.” For compiler projects, the debugging cost of such regressions far exceeds that of new-feature bugs—the error location often lands on a scalar or address unrelated to the fault, and the stack trace gives no hint of who installed the wrong assumption first.
Second, Ascend’s bottleneck has shifted from “can it compile” to “do declarations match implementations,” and review bandwidth is the hard ceiling on this front. After the 14 static defects in October, two waves of audits this week extended the gap from the parameter layer down to the T.simd instruction layer, with almost all findings taking the form of “declared as supported but the implementation takes another path” or “should have been rejected but silently passes”; meanwhile, more than ten corresponding fix PRs are under review simultaneously. The tilelang-ascend default branch had zero merges across four windows while the review queue kept thickening, and all 3060 daily tests pass—tests cover legal paths, while the ability to reject illegal paths depends on audits. The real acceptance criterion should shift from “pass rate” to “generative alignment between declarations and implementations”: generate from HCU_MCPU_SET the list of architectures the backend must cover, and generate from the intrinsic registry the dtype/bit-width support matrix.
Third, the backend matrix has shifted from “one player advancing alone” to four fronts starting simultaneously, with investment depth falling into four categories. This period’s eight main-repo commits are all Ascend, but in the same window ROCm opened seven commits as a series, Hygon produced ten commits on one branch in a day, Sunrise is running its release pipeline (ci/sunrise-0.1.15-1.2.1-r6 created and deleted in 80 seconds), and Hexagon pushed TileLang onto mobile DSP in the form of a personal fork. The four forms (upstream built-in backend, vendor adaptation repo, standalone distribution, third-party fork) correspond to four depths of investment, and this differentiation itself is a usable indicator of each vendor’s degree of dependence: vendor repos track versions, distributions track mainline, forks track their own hardware.
Fourth, TileFoundry’s pricing gap keeps being pushed back by the operator side, which is more dangerous than any single operator slowing down. #318 states outright that the same behavior was reported in four closed issues, all marked not planned, and still reproduces in the current build; #319 reports five at once with a 962.1 vs. 1213.6 GFLOP deviation (26%); the bf16 mantissa convention differs by a full bit between check --help and the hardware facts. The analyzer is the prerequisite for TileLang’s “price first, then choose schedule” approach, and a 26% pricing error will cause automatic selection to systematically choose wrong on boundary grids. Such gaps won’t slow down any single operator, but they will make automation quietly make wrong decisions.
Fifth, “engineering for agents” is beginning to appear at scale. The mlir-ascend harness branch gained +7632 lines in a day, consisting of a skill library, simulation trace workflows, and a Roofline calculator, serving “people or agents who write operators” rather than the compiler itself; in the same week, the Hexagon fork’s README states outright that the backend and models were rewritten by AI agents under maintainer guidance, with 75 PRs in four days accounting for over 80% of the current code. The two point in the same direction: the abstraction layer of DSLs like TileLang happens to land at the granularity where “agents can write it correctly,” so the ecosystem is starting to build toolchains for agents rather than just documentation for humans. This one is worth tracking weekly.
Sixth, the four RISC-V sources have been silent for six consecutive windows, forming a stark contrast with overall activity, but the way this week quantified the gap changed. The TVM route, MLIR route, community index page, and Sophgo TPU—all four upstreams remain at their original references; beyond the four sources, the CPU line, after switching conventions on 11-11, concluded that 553/553 fall within 1.03x of Triton, yet the same batch of test cases the day before was “41 cases faster than reference, SG2044 median 0.0580x.” Neither set of numbers is fabricated; the difference lies in whether the measured object is “author-written programs” or “shared compiler artifacts.” The real observation point is: whether this “same compiler, two frontends” approach can feed back into upstream, and whether the RVV/IME path will be made to work first by projects outside the four sources. The stagnation is not because no one is working on it, but because the official four sources have yet to form a loop of “automatically discover gaps, automatically verify, automatically queue.”
Appendix: Sources and Verification Notes
Source Verification Table
| Source | Verification Result |
|---|---|
| tile-ai org, repo by repo (30 repos) | 168-hour check from publish date: 12 pushed within the window (tilelang, TileOPs, TileFoundry, tilelang-ascend, tilelang-hygon, tilelang-mlir-ascend, tilelang-musa, tilelang-sunrise, TileOPs-nightly, tilelang.github.io, TileOPs.github.io, topa) |
| Main repo tile-ai/tilelang | Merges across five issues: 8, 1, 3, 3, 5; the five in issues 11-13 alone are +1903/−617, all Ascend; new issues across the five issues: 50, 6, 10, 35, 18 respectively |
| TileOPs | Merges across five issues: 33, 18, 4, 0, 0; one refresh of the stats branch in issues 11-13; mainline head 9bf4bf56 (11-10 09:59) |
| TileFoundry | Five merges in issue 11-09; #312 1491-file reorganization in issue 11-10; #323/#324 under review in issue 11-12; no new version |
| topa | 15 merges over two days in issue 11-09; no issues, no releases |
| tilelang-ascend | Default branch ascendc_pto has zero merges across four windows since 11-06; review queue keeps growing; daily tests of 3032 / 3060 items all passing recorded in each issue, with one failure in 11-12 reported twice (two alerts for the same run 43124767679) |
| tilelang-hygon | distributed branch merged #36 (+2022); tilelang-tileops-fix branch ten merges in one day; #37 under review; #27 has long carried one development line |
| tilelang-mlir-ascend | harness branch merged #217, #219 and three more (+7632/−24); master merged #222 CI gate consolidation; five TL dialect operators under review |
| tilelang-metax | Push in issue 11-09 was an upstream sync (tip is #3760), no own commits; most recent push 11-06 |
| tilelang-musa | Merged #11 (docs +24/−24) and #12 (README +10/−49); no code commits in the window; latest tag v0.1.14+musa.1 (09-11) |
| tilelang-sunrise | main stalled at 10-23; ci/sunrise-0.1.15-1.2.1-r6 deleted 80 seconds after creation |
| Release pages (tilelang / TileOPs / TileFoundry / musa) | No new tags all week: v0.1.15 (09-30), v0.1.0 (10-27), v0.0.4rc1 (10-25), v0.1.15+musa.1 (09-30); the 0.1.16 version-number PR #3887 has been pending since 11-09 03:07 |
| Adopters TileKernels / FlashQLA | No pushes all week (10-14 / 10-30) |
| TileLang×RISC-V four-source sweep | No change for six consecutive windows; index-page PR API rate-limited in issue 11-13, degraded to reference comparison only |
| Third-party TileLang–Weft / weft-bench | 17 and 23 commits in the 11-10 window; weft-parity released 11-11, 553/553 within 1.03× of Triton; the −128 overflow in the old K3 Q1 comparison has been disclosed and corrected |
| Third-party Hexagon NPU | xwhzz/tilelang-hexagon pushed on 11-10 and 11-12, LFM2-1.2B runs end to end, decode 80.1 token/s, prefill 176.2 ms |
| Third-party ops-tilelang | Atlas-Ascend-Fork/ops-tilelang pushed once each on 11-09 and 11-12; official status unverified |
| Google News (multiple Chinese and English queries) / Hacker News / arXiv | Zero hits across all five issues; HN hits were all same-name noise; latest relevant arXiv item remains 11-03 (outside window) |
| Adjacent technology developments | Meta TLX engineering blog published 11-10, a same-track comparison, not TileLang project activity |
Full Source List
- [1] tilelang PTO target default execution backend switched to tvm_ffi (#3516) — https://github.com/tile-ai/tilelang/pull/3516
- [2] tilelang AutoSchedule three necessary conditions tighten the cycle lower bound (#3776) — https://github.com/tile-ai/tilelang/pull/3776
- [3] tilelang integer atomics adapted to signless carrier (#3794) — https://github.com/tile-ai/tilelang/pull/3794
- [4] tilelang cross-core sync mode=0 forces batch scheduling (#3614) — https://github.com/tile-ai/tilelang/pull/3614
- [5] tilelang VMI descriptor and logical coverage validation (#3894) — https://github.com/tile-ai/tilelang/pull/3894
- [6] tilelang scalar cast fact protection (#3793) — https://github.com/tile-ai/tilelang/pull/3793
- [7] tilelang preserve explicit SIMD parallel loops (#3796) — https://github.com/tile-ai/tilelang/pull/3796
- [8] tilelang integer atomics sign-compatibility wrapper (#3792) — https://github.com/tile-ai/tilelang/pull/3792
- [9] tilelang CUDA L2 prefetch and six-level eviction priority (#3490) — https://github.com/tile-ai/tilelang/pull/3490
- [10] tilelang NaN propagation symmetric clamping (#3605) — https://github.com/tile-ai/tilelang/pull/3605
- [11] tilelang CI trigger condition consolidation and ROCm watchdog (#3911) — https://github.com/tile-ai/tilelang/pull/3911
- [12] tilelang 0.1.16 version number (#3887) — https://github.com/tile-ai/tilelang/pull/3887
- [13] tilelang release page — https://github.com/tile-ai/tilelang/releases
- [14] tilelang CuTeDSL grid sync moved to per-launch workspace (#3868) — https://github.com/tile-ai/tilelang/pull/3868
- [15] tilelang CuTeDSL signed integer truncation (#3551) — https://github.com/tile-ai/tilelang/pull/3551
- [16] tilelang CuTeDSL example gaps filled (#3685) — https://github.com/tile-ai/tilelang/pull/3685
- [17] tilelang CUDA kernel naming (#3816) — https://github.com/tile-ai/tilelang/pull/3816
- [18] tilelang multi-source cache directory (#3807) — https://github.com/tile-ai/tilelang/pull/3807
- [19] tilelang-ascend remove Z3 scheduling global lock (#3812) — https://github.com/tile-ai/tilelang/pull/3812
- [20] tilelang-ascend access_ptr region fix (#3805) — https://github.com/tile-ai/tilelang/pull/3805
- [21] tilelang-ascend tensor shift amount and Chunk O series (#1949, #1802, #1866, #1869, #1953) — https://github.com/tile-ai/tilelang-ascend/pull/1866
- [22] tilelang-ascend daily tests 3060 items all passing (#2003) — https://github.com/tile-ai/tilelang-ascend/issues/2003
- [23] tilelang-ascend cached kernel adapter path fix (#2002) — https://github.com/tile-ai/tilelang-ascend/pull/2002
- [24] tilelang-ascend duplicate alerts for the same run (#2006, #2010) — https://github.com/tile-ai/tilelang-ascend/issues/2010
- [25] tilelang Ascend multi-region duplicate physical segments (#3901) — https://github.com/tile-ai/tilelang/issues/3901
- [26] tilelang Ascend parameter validation and semantic mismatch issue cluster (#3920, #3928, #3941, #3947) — https://github.com/tile-ai/tilelang/issues/3920
- [27] tilelang Ascend T.simd predicate and vector instruction issues (#3992, #3993, #3976) — https://github.com/tile-ai/tilelang/issues/3992
- [28] tilelang validation-type fix PR cluster (#3973, #3972, #3968) — https://github.com/tile-ai/tilelang/pull/3973
- [29] tilelang CUDA cp.async two-byte, layout inference cost, Pipelined order (#3988, #3987, #3996) — https://github.com/tile-ai/tilelang/issues/3987
- [30] tilelang conditional proof workload upper bound (#3997) — https://github.com/tile-ai/tilelang/pull/3997
- [31] tilelang Metal SIMD-group destination out-of-bounds (#3907) and fix (#3912) — https://github.com/tile-ai/tilelang/pull/3912
- [32] tilelang Ascend host-mapped buffer parameters (#3909) — https://github.com/tile-ai/tilelang/pull/3909
- [33] tilelang ROCm series, seven PRs (#3954 to #3960) — https://github.com/tile-ai/tilelang/pull/3960
- [34] tile-ai/tvm TIRx ceildiv and negative scale (#93, #94) — https://github.com/tile-ai/tvm/pull/94
- [35] tilelang-hygon distributed trio (#36) — https://github.com/tile-ai/tilelang-hygon/pull/36
- [36] tilelang-hygon ds_read_vector Alt2 (#37) — https://github.com/tile-ai/tilelang-hygon/pull/37
- [37] tilelang-hygon gfx928 declared supported but missing branch (#35) — https://github.com/tile-ai/tilelang-hygon/issues/35
- [38] tilelang-hygon long-running PR #27 and tileops-fix branch — https://github.com/tile-ai/tilelang-hygon/pull/27
- [39] tilelang-mlir-ascend harness three PRs (#220, #221, #226) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/221
- [40] tilelang-mlir-ascend CI gate consolidation (#222) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/222
- [41] tilelang-mlir-ascend TL dialect operators (#228, #229, #231, #232, #233) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/233
- [42] tilelang-mlir-ascend harness knowledge base hardening (#217, #219, #225) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/219
- [43] tilelang-musa docs advanced to v0.1.12+musa.5 (#11) — https://github.com/tile-ai/tilelang-musa/pull/11
- [44] tilelang-musa README install section consolidated (#12) — https://github.com/tile-ai/tilelang-musa/pull/12
- [45] tilelang-sunrise release pipeline branch — https://github.com/tile-ai/tilelang-sunrise
- [46] TileOPs reciprocal dual vector (#2869) — https://github.com/tile-ai/TileOPs/pull/2869
- [47] TileOPs max pool whole-sector storage (#2868) — https://github.com/tile-ai/TileOPs/pull/2868
- [48] TileOPs compare row broadcast, two rounds (#2867) — https://github.com/tile-ai/TileOPs/pull/2867
- [49] TileOPs SM89 support surface (#2855, #2840, #2845, #2841) — https://github.com/tile-ai/TileOPs/pull/2855
- [50] TileOPs nightly performance queue written to wiki and dashboard (#2849, #2852, #2856) — https://github.com/tile-ai/TileOPs/pull/2849
- [51] TileOPs BMM ping-pong bandwidth (#2882) — https://github.com/tile-ai/TileOPs/pull/2882
- [52] TileOPs K=0 empty contraction and variable-length GQA hang (#2878, #2876) — https://github.com/tile-ai/TileOPs/pull/2878
- [53] TileOPs GLA backward shared memory (#2880) — https://github.com/tile-ai/TileOPs/pull/2880
- [54] TileOPs Softmax instruction count (#2886, #2871) — https://github.com/tile-ai/TileOPs/pull/2886
- [55] TileOPs DeltaNet forward split and backward buffer reuse (#2891, #2890) — https://github.com/tile-ai/TileOPs/pull/2891
- [56] TileOPs tests split by cluster and shared memory (#2889) — https://github.com/tile-ai/TileOPs/pull/2889
- [57] TileOPs W4A16 two stacked PRs (#2895, #2896) — https://github.com/tile-ai/TileOPs/pull/2896
- [58] TileOPs DeepSeek-V3 FP8 grouped GEMM A scale via TMA (#2894) — https://github.com/tile-ai/TileOPs/pull/2894
- [59] TileOPs softmax two-CTA cluster and long-K NT GEMM (#2898, #2897) — https://github.com/tile-ai/TileOPs/pull/2898
- [60] TileOPs-nightly nightly snapshots — https://github.com/tile-ai/TileOPs-nightly
- [61] TileFoundry placement value semantics (#293) — https://github.com/tile-ai/TileFoundry/pull/293
- [62] TileFoundry old-data order before overwrite (#301) — https://github.com/tile-ai/TileFoundry/pull/301
- [63] TileFoundry TMA sync plan and cross-CTA Partial (#291, #292) — https://github.com/tile-ai/TileFoundry/pull/292
- [64] TileFoundry source-spec reorganization (#312) — https://github.com/tile-ai/TileFoundry/pull/312
- [65] TileFoundry analyzer gaps (#314 to #319) — https://github.com/tile-ai/TileFoundry/issues/319
- [66] TileFoundry Mesh selection slicing and dimension arithmetic (#323, #324) — https://github.com/tile-ai/TileFoundry/pull/323
- [67] topa repo and workflow consolidation — https://github.com/tile-ai/topa
- [68] deepseek-ai/TileKernels — https://github.com/deepseek-ai/TileKernels
- [69] QwenLM/FlashQLA — https://github.com/QwenLM/FlashQLA
- [70] Tilelang-AttnRes (Kimi K3 attention residual) — https://github.com/Index4NekoMeteor/Tilelang-AttnRes
- [71] models-tilelang (DeepSeek V4.1 Flash snippet) — https://github.com/kevin3113/models-tilelang
- [72] tilelang-hexagon (Hexagon NPU backend) — https://github.com/xwhzz/tilelang-hexagon
- [73] Atlas-Ascend-Fork/ops-tilelang — https://github.com/Atlas-Ascend-Fork/ops-tilelang
- [74] TileLang–Weft (weft-parity) — https://github.com/Tianchen-ai/tilelang-weft
- [75] TileLang–Weft benchmark repo results handbook — https://github.com/Tianchen-ai/tilelang-weft-bench/blob/main/docs/results.zh-CN.md
- [76] Moore Threads Kimi-K3.1 Pro Day-0 adaptation (11-06) — https://news.qq.com/rain/a/20261106A0FH3900
- [77] Foro3D: DeepSeek V5 paraphrase (11-06) — https://foro3d.com/en/2026/noviembre/deepseek-prepara-su-v5-de-8-billones-de-parametros-sobre-silicio-huawe.html
- [78] Meta TLX engineering blog (11-10) — https://engineering.fb.com/2026/11/10/open-source/tlx-a-hardware-native-gpu-compiler-for-production-ml
Verification Notes
- This issue covers five working days, November 9 to 13, 2026, across five daily issues. Issue 11-09 is a 72-hour window spanning the weekend (11-06 07:00 to 11-09 07:00); the other four are contiguous 24-hour windows. Issue 11-09 does not overlap with the previous weekly report (through 11-06 07:00).
- Merge counts, review counts, and new issue counts are sourced from the GitHub search ground truth recorded in each daily issue (
is:merged,is:open,created), cross-checked against per-repo main-branch commits. Issues 11-09 through 11-12 are org-level counts; issue 11-13 is main-repo counts (that issue also has three commits on mlir-ascend’sharnessbranch, not counted). - Merges to non-default branches are noted branch by branch in the body: hygon
#36lands ondistributed, ten on hygon’stilelang-tileops-fixbranch, mlir-ascend#217/#219/#220/#221/#226onharness, mlir-ascend#222onmaster, and one each on TileOPsstatsand TileOPs-nightlysnapshots— these repos’mainhad zero commits in the corresponding windows, so only branch-level queries surface them. - Performance and memory figures are taken from the tables and
Result:lines in each PR body, with no secondary conversion; cases that fall below bar and cases slower than the comparison implementation (e.g.,#2891is 9% to 41% slower than FLA on SM89) are kept as stated in the original. - The two sets of figures for the third-party Weft line are listed separately with the reason for the switch; the old table (41 cases faster than reference, SG2044 median 0.0580×) is not mixed with the new table (553/553, median 1.29 to 1.66); the overflow in the old Triton artifacts for the two K3 Q1 comparisons when negating −128 is recorded per the benchmark repo’s disclosure.
- The news side had zero hits across all five issues, so no external coverage is available to cross-validate this week’s project progress; the Moore Threads Day-0 item comes from a paraphrase of an official announcement and includes a specific migration path, while the Foro3D item is a second-hand paraphrase and DeepSeek has not officially announced V5 plans, so the report flags it as a rumor.
- Collection-side limitations: GitHub anonymous API quota was exhausted early in several windows this week, so PR bodies and merge details were partly filled in via the search API, branch-level atom feeds, and diff endpoints; for a few PRs the file counts and line counts are API return values and the diffs were not reviewed line by line; the RISC-V four-source sweep in issue 11-13 was degraded to reference comparison only due to index-page PR API rate limiting.
- This report contains no local paths, proxy addresses, or intranet information; all links are publicly accessible addresses.