TileLang Daily Intelligence Report (2026-11-13)
Research window: 2026-11-12 07:00 ~ 2026-11-13 07:00 (Beijing time, 24 hours; the previous issue’s cutoff was 11-12 07:00).
Sources: Repository-by-repository check of the GitHub tile-ai organization (30 repositories, 7 pushed within the window) + organization-level merged and under-review ground-truth search (8 merged, 71 under-review updates, 18 newly opened issues within the window) + commit-by-commit check of the main branch (tilelang 5 commits, mlir-ascend and TileOPs both 0) + non-default branch check (mlir-ascend harness 3 commits, hygon tilelang-tileops-fix 10 commits, TileOPs stats 1 commit, TileOPs-nightly snapshots 1 commit) + release page check (tilelang / TileOPs / mlir-ascend, no new versions within the window) + third-party backend repository check (Hexagon NPU fork, CANN community operator repository) + Google News Chinese and English queries + Hacker News + arXiv. The dedicated RISC-V four-source inspection showed no changes in 24 hours.
This Issue’s Index
-
Today’s Focus: Five merges into the main repo are all Ascend, four of them wrapping up #3512; ROCm opens seven in a series within an hour; Ascend 950 audit changes hands and continues, with eight new issues all pointing at the low-level SIMD surface
- I. Core Project Progress
- 1.1 Ascend: Cross-core sync mode=0 forces batch scheduling, all-core barrier no longer deadlocks competing for cores with multi-stream (11-12)
- 1.2 Ascend PTO: VMI descriptors and logical coverage checks closed out together, single commit +925/−550 is the largest this issue (11-12)
- 1.3 Ascend PTO: Scalar cast facts protected, wrapping casts no longer make IntSet evaluation throw (11-12)
- 1.4 Ascend PTO: Explicit SIMD parallel loops no longer rejected by lane layout inference (11-12)
- 1.5 Ascend PTO: Integer atomics switched to signless compatibility wrapper, signed inputs now compile (11-12)
- 1.6 Mainline summary: All five merges are Ascend, +1903/−617, version still at v0.1.15 (11-12)
- II. Multi-Backend Adaptation
- 2.1 mlir-ascend: harness branch merges three in one day, +7632/−24, all on non-default branches (11-12)
- 2.2 mlir-ascend: TL dialect adds five operators in one day, reshape / sort / print / pow / xor under review (11-12)
- 2.3 Hygon:
tilelang-tileops-fixbranch sees ten commits in a single day, hipRAND random numbers and cache_warmup enter the branch (11-12) - 2.4 Hygon: PR #27 still under review, pushed content is the update to that PR branch (11-12)
- 2.5 Sunrise: Release pipeline branch
ci/sunrise-0.1.15-1.2.1-r6deleted 80 seconds after creation (11-12) - 2.6 MetaX and Moore Threads: No commits in window (most recent 11-06 / 11-10)
- 2.7 ROCm: Seven opened in a series within an hour, 22 files +624/−19 (11-12)
- II·V. RISC-V Backend and Community (Special Observation)
- 2.8 Four-source inspection: Sixth consecutive window with no changes, PR API rate-limited so only references compared (11-13)
- III. Ecosystem and Adopters
- 3.1 TileOPs: Mainline sees zero commits for second consecutive window, stats branch refreshes manifest (11-12)
- 3.2 TileOPs-nightly: Nightly snapshot still at 9bf4bf56, run 43264767627 tests unchanged code (11-13)
- 3.3 TileFoundry: Two fixes under review, Mesh selection slicing and dimension arithmetic (11-12)
- 3.4 Third-party backends: Hexagon NPU fork pushes TileLang onto phone DSP, LFM2-1.2B runs end-to-end (11-12)
- 3.5 Ascend TileLang operator repo surfaces in CANN community, official status unverified (11-12)
- IV. Community, Tutorials and Events
- 4.1 Ascend 950 audit second wave: Eight issues concentrated on predicates and the low-level SIMD surface (11-12)
- 4.2 Corresponding fixes already queued for the audit: Over ten validation-type PRs advancing in parallel (11-12)
- 4.3 Three new CUDA-side issues: cp.async two-byte, layout inference exponential time, Pipelined explicit order (11-12)
- 4.4 News side: Zero hits in Chinese and English query windows, arXiv latest still at 11-03 (11-12)
- 4.5 Daily regression: Same run opens one alert at each of two time points (11-12)
- V. Trend Observations
Today’s Highlights
The main thread of this window is cleanup. The five merges in the main repo (+1903/−617, 22 file changes) added no new operators and no new backends; four of them are wrapping up the tail of the PTO vectorization overhaul from the #3512 round: VMI descriptors were fetched incorrectly on typed-view gather/scatter and single-element loads, so #3894 fixes 6 files in one go, +925/−550; the analytic traversal bound scalar definitions too early, and IntSet threw an exception when evaluated on a wrapping cast, so #3793 adds a guard; lane-layout inference treated the vector block address of explicit SIMD as a scalar lane access, so the old form no longer compiled, and #3796 preserves the old path; codegen_pto was changed to emit si32/si64 while PTO atomics only accept signless, so #3792 performs a local reinterpretation at the call boundary. What these four share is that “after the overhaul, previously legal programs are no longer legal” — an inevitable cost of a large change, but four consecutive windows from 11-10 to 11-13 have all been paying this bill, which suggests the regression surface of the PTO line is wider than estimated.
The fifth merge, #3614, is of a different nature: it is not a bug fix but deadlock prevention. Ascend kernels have two scheduling modes: mode=0 is “launch the operator as soon as cores are available,” and mode=1 is “wait until all core resources the operator needs are in place, then launch as a whole.” If an operator that performs all-core synchronization is launched in normal mode, it may get only some of the cores and run ahead to the all-core barrier to wait for the rest, while the rest are occupied by another stream — resources form a cycle, and the device hangs outright. This change identifies “cross-core synchronization with constant 0” on the codegen side, automatically sets __schedmode__(1), and resets the flag per function to prevent leakage into subsequent entry points; at the same time, it narrows the integer-form mode in the cross-core set/wait interfaces to 0/1/2/4, leaving non-integers unvalidated to preserve symbolic quantities. The regression file is 78 lines and covers seven cases: all-AIC, all-AIV, the AIC+AIV combined protocol, three local modes, and flag reset.
The activity on the ROCm side is more noteworthy than Ascend’s. Less than an hour after the window opened, the same account filed seven PRs in a row, 22 files +624/−19, all based on 72d986e1, forming a complete set of HIP-side patches: compile flags previously went only into the cache key and not into hipcc (#3954); 8/16-lane broadcasts of 8-bit integer vectors emitted make_uint2/make_uint4, which hipcc rejects (#3955); T.gemm on small tiles reported useless messages like “Divide by zero” (#3956); T.annotate_min_blocks_per_sm was dropped on HIP, and on gfx90a a 256×128×32 fp16 GEMM was locked to one block per CU due to 143 VGPRs plus 128 AGPRs (#3957); occupancy and scratch usage showed only raw numbers without conversion (#3958); a 64-thread reduction on wave64 allocated an LDS block for nothing (#3959); in HIP, uint2/uint4 are structs, so fp16 fragments that live across loops were pushed back to the stack by SROA, consuming 80 bytes of scratch per lane (#3960). ROCm had long been recorded in this daily report as “an occasional one-off”; appearing as a series within a single day is the most obvious structural change this issue.
The third item is on the harness branch of mlir-ascend: three merges totaling +7632/−24 and 64 file changes, none of which land on main or master. The content is not compiler functionality but a “knowledge base written for people (or agents) who want to write operators”: #220 adds the Ascend950 simulation trace workflow (msprof collection, viewing in MindStudio Insight, supplementing GET/RLS events with the dump from the same run, exporting npusim SQLite for timeline analysis; all 38 existing tests pass on WSL); #221 is an installable version of the operator design guide (phase decomposition, SIMD/SIMT selection, storage budgeting, tile pipelining, plus a Cube MFU and conditional Roofline calculator, and it saves snapshots of nine upstream TileLang Skill documents for offline reading); #226 adds BMM and w4a8 Gemm quantization examples, and a conv example went in with the same batch. The value of this line is not compile capability but turning “how to write operators correctly on this hardware and compute MFU properly” into something distributable.
The fourth item is a change of hands on the quality side. Last issue, two accounts, cn-sys and JINITAIMEI121, piled up 30 Ascend issues in a day; this issue, NebulaMao filed eight in a row within a dozen or so minutes after 18:39, all running on Atlas 950 / Ascend950PR, version 0.1.15, pointing at a lower layer: after exporting any_sync/all_sync, source generation leaves an unresolved tl.any_sync; a legal int64 cross-warp reduce_sum emits an undefined HwReduce<SumOp,long>; update_mask/pintlv/pdintlv do not validate predicate width and can emit a b64 helper that does not exist at all; T.simd.pge silently degrades to pset_bXX; T.copy’s l2_cache_ctrl accepts out-of-range integers and does an int32 wrapping cast; T.set_atomic(dtype="int32") is written as documented but emits a nonexistent *_int32 CANN name; vld2 accepts mismatched DINTLV widths and 64-bit dtypes; T.ascend_pipe_barrier("PIPE_V") is silently dropped by codegen. The common shape of all eight is still one sentence: places that should report an error do not report one.
I. Core Project Progress
1.1 Ascend: Cross-core sync mode=0 forces batch scheduling, all-core barrier no longer deadlocks on core contention with multiple streams (#3614)
This change touches 5 files, +102/−1, of which 78 lines are a new regression test for the schedmode flag; the actual compiler change is only about twenty lines: in Ascend codegen, after identifying “cross-core sync with constant 0,” it sets __schedmode__(1), and makes this flag a per-function-reset member to avoid leaking into subsequent entry points.
To see why this is necessary, return to the scheduling semantics. Ascend kernel mode=0 is normal mode—it launches as soon as the partial core resources required by the operator are available, and operators from different streams can overlap; mode=1 is batch mode—it launches as a whole only after all core resources required by the operator are in place. Operators performing all-core sync (full AIC sync, full AIV sync, and the combined protocol of “AIV sync with local AIC → full AIC sync → sync back to local AIV,” which internally contains a mode=0 phase) may, when launched in normal mode, obtain only some cores and run ahead to wait at the all-core barrier, while the cores it waits for are occupied by another stream and will not be released—resources form a cycle, and the device deadlocks. Batch mode eliminates this window.
The interface was also narrowed along the way: in the cross-core set/wait interface, the integer form of mode_id accepts only 0, 1, 2, 4, and other integers are rejected outright; non-integers are not validated, to preserve the possibility of passing symbolic quantities. The docstring also adds a factual note—the modes listed in the same API documentation are not consistent across the three targets: 910B/910C support 0/1/2, and only 950 adds 4. The regression covers seven cases: full AIC, full AIV, and the AIC+AIV combination should each enable batch mode; mode=1/2/4 should remain default; and the flag does not persist between functions.
Source: https://github.com/tile-ai/tilelang/pull/3614
1.2 Ascend PTO: VMI descriptors and logical-overlap validation closed out together, a single +925/−550, the largest of this period (#3894)
#3894 is the largest change in this window (6 files, +925/−550), putting two independent matters in one branch. The first is fixing VMI memory descriptors, involving two forms: gather/scatter on typed views, and single-element loads—both take a base address and stride when constructing the descriptor that are inconsistent with the actual access. The second is removing the redundant logical-overlap rejection in the VMI frontend.
The premise of the second matter is worth stating clearly: removing the validation is not a relaxation, but because the shared-memory detector fix (#3805) has already landed, and blocking again in the frontend is duplicate interception that also falsely rejects forms such as cross-row, transpose, and dynamic views that should be allowed through. In other words, the same semantics were judged once at each of two layers; once the earlier one was fixed, the latter became a burden.
The −550 lines of deletions mainly come from consolidating these two places. The implication for downstream is that on the PTO target, the two styles—gather/scatter on typed views and single-element loads—go from “compiles but produces wrong results” to “either correct or caught by the detector.”
Source: https://github.com/tile-ai/tilelang/pull/3894
1.3 Ascend PTO: Scalar cast facts protected, wrapping casts no longer make IntSet evaluation throw (#3793)
The reproducer is tiny: u = uint32(i - 1), with i in [0, 63]. u has nothing to do with any VMI address, yet LegalizePTOVMIMemoryAccess fails on it. The root cause is the analyzer-aware traversal introduced in #3512—it eagerly binds scalar definitions into the analysis context, and when IntSet evaluates cast endpoints, if another sub-analyzer has already installed an underestimated bound, the wrapping cast throws.
The fix is to skip unsafe analysis during lowering (2 files, +571/−13). This kind of “analyzers contaminating each other’s assumptions” problem is the most time-consuming in a compiler after a major overhaul: the error is reported on a scalar unrelated to the fault, and the stack does not show who installed the wrong bound first.
Source: https://github.com/tile-ai/tilelang/pull/3793
1.4 Ascend PTO: Explicit SIMD parallel loops no longer rejected by lane-layout inference (#3796)
The PTO lane-layout inference introduced in #3512 rejects existing T.Parallel nests whenever explicit SIMD instructions are written inside. The reason is that inference treats the addresses of these loops as scalar lane accesses, whereas they actually iterate over vector blocks.
A typical example is given in the PR: a T.Parallel(2, 32) nest whose loop body uses S.vld(temp[h, block * 64]) and the corresponding S.vsts(...), processing one 64-element FP32 vector per iteration. This style was legal and commonly used before inference appeared, and after inference landed it uniformly failed to compile. The change touches 5 files, +194/−17, and the approach is to preserve the legacy explicit-SIMD parallel semantics and let inference apply only to the forms it should govern.
Source: https://github.com/tile-ai/tilelang/pull/3796
1.5 Ascend PTO: Integer atomics switched to signless compatibility wrappers, signed inputs now compile (#3792)
After #3512, codegen_pto emits pto.si32/pto.si64 for TileLang integer values, while PTO’s atomic interface accepts only signless i32/i64, so T.atomic_add, T.atomic_max, and T.atomic_min on integer buffers all failed to compile (previously mentioned in the same batch of items in the last issue; landed on main at 11:27 in this window).
The fix performs local adaptation only at the generated call boundary (4 files, +111/−36): pointers and operands are reinterpreted as signless at the same width, the pointer’s address space is preserved, and the return value is restored to the logical type; min/max retain their original signed or unsigned comparison semantics. This way the logical type system established by #3512 need not be overturned, and the adaptation is confined to the atomic layer.
Source: https://github.com/tile-ai/tilelang/pull/3792
1.6 Mainline summary: all five merges are Ascend, +1903/−617, version still at v0.1.15 (11-12)
Breaking down the five merges to the main repo within the window by direction: #3614, #3894, #3793, #3796, #3792, all landing on Ascend, with zero merges on the CUDA, ROCm, and CPU sides. Total +1903/−617 across 22 file changes, of which four are regression cleanup for #3512 and one is scheduling-mode hardening.
All non-Ascend backends are piled up in the review queue: within the main-repo window there are 50 updates under review, including on the CUDA side SM90 shared transpose lowering to ldmatrix/stmatrix (#3809), pipelined GEMM operands kept alive until completion wait (#3923), sync before TMEM auto-release (#3975), and fallback to scalar MMA for incompatible shared layouts (#3998), all in the queue; the ROCm side is the seven items in 2.7. There is no change on the version side: the latest on the main repo’s releases page is still v0.1.15 from 09-30, and the tag sequence is unchanged; TileOPs is still v0.1.0 from 10-27. These 1,903 lines in this issue will not enter downstream distribution packages until at least the next tag.
Source: https://github.com/tile-ai/tilelang/releases , https://github.com/tile-ai/tilelang/pull/3809 , https://github.com/tile-ai/tilelang/pull/3923
II. Multi-Backend Adaptation
2.1 mlir-ascend: Three merges into the harness branch in one day, +7632/−24, all on non-default branches (11-12)
None of the three merges landed on main or master; they landed on the harness branch, so they are only visible when querying by branch: #220 (14 files, +1014/−12) is the Ascend950 simulation trace skill, which strings together msprof collection, MindStudio Insight inspection, dumping GET/RLS events in the same run, and exporting npusim SQLite for timeline analysis into one executable flow; all 38 existing tests pass on WSL. #221 (45 files, +5178/−12) is the installable version of the operator design guide, covering stage decomposition, SIMD/SIMT selection, storage budgeting, and tile pipelining, with a Cube MFU and conditional Roofline calculator attached so analysis and device-test nodes can report “how much work Cube did during this period”; it also snapshots nine upstream TileLang Skill documents separately for offline reading. #226 (5 files, +1440/−0) adds BMM and w4a8 Gemm quantization examples, and a conv example (e51c1774f1) landed in the same batch.
The three merges total +7632/−24 across 64 file changes, the largest single-line landing in this window. In the previous issue these two (#220, #221) were still proposals; this issue they have entered the branch. This also explains why the repo’s main/master had zero commits that day—the work surface is entirely on the harness side.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/221 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/220 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/226
2.2 mlir-ascend: TL dialect adds five operators in one day; reshape / sort / print / pow / xor under review (11-12)
The TL dialect was filled out intensively within the window, with five proposals opened in succession from 17:49 to 05:45 the next day (Beijing time): #228 tl.xor, #229 tl.pow, #231 tl.print, #232 tl.sort, #233 tl.reshape. All are built on the TL_BinaryOp<...> base class introduced in #203, following the template of #218 (and) and #225 (or), so each individual change is thin.
Each has its own division of labor: tl.reshape is a pure view operator with no destination operand, lowering directly to memref.reinterpret_cast in TL→HIVM without passing through any HIVM hardware operator, aligned with the existing NPUIR ReshapeCodegen path; tl.sort lowers to hivm.hir.vsort, with conversion tests for both ascending and descending order and an expert_mode path; tl.print lowers to hivm.DebugOp, handling both scalar and tensor print forms, with the verifier only checking that the element type is integer or float; tl.pow and tl.xor follow the usual three stages of elementwise binary operators (dialect definition, verifyElementwiseBinary, ConvertBinary to the corresponding HIVM operator).
The significance of filling out the dialect lies in coverage: the MLIR route must be able to accept 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.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/233 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/232 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/231
2.3 Hygon: Ten commits in one day on the tilelang-tileops-fix branch; hipRAND random numbers and cache_warmup enter the branch (11-12)
The Hygon repo’s main had zero commits, but the tilelang-tileops-fix branch landed ten commits between 15:41 and 21:14 (Beijing time), aimed at “filling out the operator surface TileOPs needs”: feat(hcu): support rng operations with hipRAND adds random numbers; feat(hcu): support cache_warmup intrinsic adds the cache warmup intrinsic; feat(hcu): rebase buffer loads; [Fix] unroll the epilogue of a pipeline with a runtime trip count handles epilogue unrolling under a runtime trip count; [Fix] fix 8-bit integer vector broadcast code generation; [Fix] fix cos and sin intrinsic dispatch; [feat] fix shuffle lowering and enable full-wave shuffle reduction; plus three compile-time fixes for __builtin_bit_cast size mismatch, fp64 truncation, and hexp call ambiguity.
The 8-bit integer vector broadcast commit shares its topic with the main repo’s ROCm series #3955, indicating that these two HIP-side lines hit the same class of code-generation gap in the same week. The Hygon repo now has 15 branches—six in the feat/hcu-* series, one distributed, plus this tilelang-tileops-fix—so the work surface already covers three layers: instructions, communication, and operator compatibility.
Source: https://github.com/tile-ai/tilelang-hygon/tree/tilelang-tileops-fix
2.4 Hygon: PR #27 still under review; the pushed content is the update to that PR’s branch (11-12)
In the org-level per-repo check, the Hygon repo showed “pushed within the window,” landing as a push to the tilelang-tileops-fix branch (21:15), corresponding to the update time of #27 (Fix intrinsic lowering, precision, and integer vector codegen) at 21:15:47. That PR has been open since 10-23 and is still unmerged, but the branch keeps growing: the ten commits in this window are its content.
In other words, the Hygon side currently takes the form of “one long-lived PR carrying one development line”—the results on the branch already cover random numbers, cache warmup, shuffle reduction, and several compile-time fixes, but they only enter the mainline once this one lands as a whole.
Source: https://github.com/tile-ai/tilelang-hygon/pull/27
2.5 Sunrise: Release pipeline branch ci/sunrise-0.1.15-1.2.1-r6 created and deleted 80 seconds later (11-12)
The Sunrise backend distribution repo had no code changes in this window: main is still at fb7d3d62 from 10-23. The only action in the window was the creation and deletion of a CI branch—ci/sunrise-0.1.15-1.2.1-r6 created at 18:17:24 and deleted at 18:18:44, an 80-second interval.
The information in the branch name lies in the version combination: 0.1.15 corresponds to the upstream baseline, 1.2.1 is the second component bumped from the 1.2.0 in the 10-19 candidate branch name, and r6 means the 6th revision. This belongs to the same release sequence as the 0.1.14+sunrise.1.1.0 recorded in September, indicating that the backend is still tracking the mainline and issuing revisions at its own pace—this issue just stopped at a build rather than a commit.
Source: https://github.com/tile-ai/tilelang-sunrise
2.6 MetaX and Moore Threads: No commits in the window (most recent 11-06 / 11-10)
tilelang-metax last pushed on 11-06 (a merge PR #162 on main merging upstream main into dev, opened at 01:37 with no review), and tilelang-musa last pushed on 11-10. Neither repo had substantive commits in the window, nor new versions: metax has no release, and musa’s latest tag is still v0.1.14+musa.1 (09-11).
Source: https://github.com/tile-ai/tilelang-metax , https://github.com/tile-ai/tilelang-musa
2.7 ROCm: Seven PRs opened as a series within an hour, 22 files +624/−19 (11-12)
The seven were opened in succession within one minute, from 07:56 to 07:57 (Beijing time), by the same author, all based on 72d986e1, and together form a set of HIP-side patches:
#3954(3 files, +65/−2) passescompile_flagsandtl.device_compile_flagsinto hipcc. Previously these parameters only entered the kernel cache key and did not affect actual compilation; the only thing that could change device code generation was the process-levelHIPCC_COMPILE_FLAGS_APPEND.#3955(2 files, +70/−9) fixes broadcasting of 8-bit integer vectors: with 4 lanes,ICHECKhits a non-constant; with 8 and 16 lanes, it emitsmake_uint2/make_uint4with one argument per byte, which hipcc rejects outright.#3956(3 files, +102/−1) adds readable errors for small-tileT.gemm. 16×32 with 256 threads walks into the MFMA macro generator and reports “Divide by zero,” or hits “Layout infer conflict” when the output of one gemm is fed directly as the A operand of the next; neither case previously had usable information.#3957(6 files, +102/−5) changes the “minimum blocks per SM” annotation on HIP from “discarded” to “emit a launch bound based on waves per EU.” The motivation includes measurements: on gfx90a, a 256×128×32 fp16 GEMM compiles to 143 VGPRs plus 128 AGPRs with occupancy 1, so it can only fit one block per CU no matter how much LDS remains.#3958(2 files, +73/−0) adds occupancy reporting tohip_resource_info. hipcc already reportsOccupancyandScratchSize, but nobody converts them into blocks per CU—which is what actually determines whether a kernel can use all the LDS it requested for two blocks.#3959(2 files, +58/−2) fixes ROCm’s reduction workspace determination.AllReduceNeedsWorkspacereturns true when there are more than 32 reduction threads, whiletl::AllReduceonly uses a workspace when crossing a wavefront, so on wave64 a 64-thread reduction allocates a block of LDS it never touches; and because the workspace counts as a shared-memory access in the IR, ThreadSync wraps every such reduction in a barrier.#3960(4 files, +154/−0) keeps fp16 fragments actually in registers. In HIP,uint2/uint4are structs, so vector copies into or out of local arrays degrade to memcpy, and after SROA the array still lives on the stack: on clang 18 / ROCm 6.2.4, an fp16 fragment that resides across a loop and holds 2 rows of K and V per thread consumes 80 bytes of scratch per lane.
The seven total 22 files +624/−19, covering six layers: compile parameters, code generation, launch bounds, occupancy reporting, reduction workspace, and register residency. This is the first time a systematic set has appeared since this daily began tracking ROCm.
Source: https://github.com/tile-ai/tilelang/pull/3960 , https://github.com/tile-ai/tilelang/pull/3957 , https://github.com/tile-ai/tilelang/pull/3959
II·V. RISC-V Backend and Community (Special Focus)
2.8 Four-source inspection: Sixth consecutive window with no changes; PR API rate-limited, references compared only (11-13)
Comparing references one by one across the four established sources, nothing changed within the 24-hour window:
kubuds/tilelang-riscv-kubuds(TVM backend route): the twokubuds-prefixed branches remain atkubuds-riscv-karenfc62939 andkubuds-riscv-mahiru33e0c89;mainis only upstream sync;RuyiAI-Stack/tilelang-riscv(MLIR rewrite route,target="riscv"):mainremains at 57c548c;TongmingLAIC/tilelang-riscv(community index page):mainremains at 4499c2a, with no additions or removals in the table;TongmingLAIC/tilelang-tpu(Sophgo TPU):mainremains at 303759f,feature/sg2260e-rv-supportremains at 122bc59, and the references of the other branches are likewise unchanged.
This is the sixth consecutive window with no changes (archived on 11-07, 11-09, 11-10, 11-11, 11-12, 11-13). This issue’s script hit anonymous API rate limiting while querying the index page’s PRs, so it degraded to comparing references only and logged a note—under the established convention, empty output is recorded as “no changes” and is not treated as a collection failure.
Source: GitHub four-source reference-by-reference comparison (no changes within the window); no single landing link.
III. Ecosystem and Adopters
3.1 TileOPs: Second consecutive window with zero commits on main; stats branch refreshes manifest (11-12)
TileOPs main logged zero commits for a second consecutive window, with the main branch head still at 9bf4bf56 (11-10 09:59 Beijing time). The only write during the window was a stats branch refresh at 23:17, with the commit message chore(stats): refresh manifest stats from 9bf4bf56..., i.e. a recomputation of the manifest statistics for the current mainline.
The five performance PRs recorded last issue (#2894 through #2898) did not appear in the in-review update list during this window, indicating they were neither merged nor advanced — TileOPs is currently in a “waiting for review” rather than “in development” state.
Source: https://github.com/tile-ai/TileOPs
3.2 TileOPs-nightly: Nightly snapshot still lands on 9bf4bf56; run 43264767627 tests unchanged code (11-13)
The snapshots branch received a bench: 9bf4bf56... (run 43264767627) commit at 04:33 (Beijing time). The snapshot’s base commit matches TileOPs’ current mainline head, so this night’s benchmark results were measured against the same unchanged code.
This matters when reading the numbers: differences between two consecutive nightly snapshots can only come from environmental fluctuation and must not be treated as performance changes. Last issue’s push record for this repo came from a wiki Perf Tracker page refresh; this issue’s is a genuine benchmark snapshot commit — the two look very similar in “pushes during the window” statistics and must be distinguished by branch.
Source: https://github.com/tile-ai/TileOPs-nightly
3.3 TileFoundry: Two fixes in review — Mesh selection slicing and dimension arithmetic (11-12)
TileFoundry opened two PRs at 21:56 (Beijing time), both of the “looks right when written, miscompiles when built” variety:
#323 fixes the parser dropping slices that start at zero. When relabeling a threads[0:16] range via Mesh with layout=(16,), if the selection starts at offset 0, the parser drops the slice when constructing the new mesh, and after lowering no contains guard is emitted, so every thread in that layer runs the region; whereas the direct form with threads[0:1] as writer: preserves the slice and behaves correctly — only the relabeling path is broken.
#324 fixes dimension arithmetic in codegen. When loop bounds come from mesh coordinates (e.g. range(cta.i * 32, cta.i * 32 + 32)), CUDA codegen reports no codegen handler for DimMul on cuda.Target: the shared scalar renderer is responsible for writing loop bounds, conditions, and mbarrier phases — it handles comparisons but not dimension arithmetic — while the three CUDA emitters each maintain their own + - * / % table.
Source: https://github.com/tile-ai/TileFoundry/pull/323 , https://github.com/tile-ai/TileFoundry/pull/324
3.4 Third-party backend: Hexagon NPU fork pushes TileLang onto phone DSP, LFM2-1.2B runs end-to-end (11-12)
The out-of-organization xwhzz/tilelang-hexagon had one push during the window (landing on the fix/hexagon-dma-load-only branch, latest commit perf(hexagon): inline DMA across IEEE arithmetic scopes, 16:32 Beijing time); its main branch hexagon-backend last commit is from 11-11. This fork deserves a separate mention: it adds a Qualcomm Hexagon NPU backend to TileLang, developed and tested on the v79 cDSP of a Snapdragon 8 Elite phone, compiling T.copy, T.gemm, T.Parallel, T.reduce_*, and T.Pipelined into HVX vector code, HMX matrix multiplication, VTCM buffers, and DMA streams, executed on the NPU via FastRPC; the repo currently has 12 branches, plus Docker environment setup, SA8650 Linux board setup, and operator example documentation.
Results per the README: 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), tied embedding and head are 8-bit; on-device WikiText-2 perplexity is 16.68 versus 16.29 for bf16. Two performance figures: decode of 128 tokens at 80.1 token/s against a bound of 83 token/s (685 MB of weights per token, at a measured DSP DMA peak of 57 GB/s), reaching 96% of the bound; prefill of 1024 tokens at 176.2 ms (5813 token/s) against a bound of 129 ms (the time HMX alone needs for the matrix multiplication), reaching 73% of the bound. The numbers are the median of five bench.py runs; timing includes FastRPC call overhead but excludes tokenization and the PC round-trip driving the phone.
The most noteworthy line is in the README: this backend and the LFM2 model were “rewritten by an AI coding agent under maintainer guidance,” with 75 PRs filed between 2026-10-26 and 10-29, accounting for over 80% of the backend’s current code, tests, and documentation. Read alongside the harness thread in 2.1, the implications go beyond a single backend — see “V. Trend Observations.”
Source: https://github.com/xwhzz/tilelang-hexagon
3.5 Ascend TileLang operator repo from the CANN community surfaces; official status unverified (11-12)
Atlas-Ascend-Fork/ops-tilelang had a push during the window (14:23 Beijing time). The repo description states its positioning: “a CANN community TileLang high-performance operator repository for Ascend NPU,” uniformly managing kernel implementations, automated test cases, standard workloads, and performance baselines, built into reusable, verifiable operator assets. The repo was created on 10-20, default branch master, sparse activity.
It is recorded only as a lead because: the account name contains Fork, it is neither a tile-ai organization nor an official Ascend organization, and whether it has official backing cannot be determined from the repo itself. If the repo continues to update and cross-references with tilelang-ascend appear, it can be taken as a signal that “Ascend operator assets are beginning to have independent community-side repositories.”
Source: https://github.com/Atlas-Ascend-Fork/ops-tilelang
IV. Community, Tutorials, and Events
4.1 Second wave of Ascend 950 audit: eight issues concentrated on predicates and low-level SIMD surface (11-12)
During the window, the organization opened 18 new issues: 15 in the main repo, 2 in tilelang-ascend, and 1 in mlir-ascend. Of the 15 in the main repo, 8 were filed by NebulaMao in clusters by theme within a dozen or so minutes after 18:39, all running on Atlas 950 / Ascend950PR (npu-smi 25.7.rc1.10), version 0.1.15.
Taking the eight one by one: after any_sync/all_sync are exported, source generation leaves unresolved tl.any_sync/tl.all_sync calls; a legal int64 cross-warp reduce_sum emits an undefined HwReduce<SumOp,long>; T.simd.update_mask/pintlv/pdintlv do not validate predicate width and can emit nonexistent b64 helpers; T.simd.pge silently degrades to pset_bXX (via asc_create_mask_bXX) instead of pge_bXX; T.copy’s l2_cache_ctrl accepts out-of-range integers and casts them with int32 wraparound; T.set_atomic(..., dtype="int32") written as documented emits a nonexistent *_int32 CANN helper name; T.simd.vld2 accepts mismatched DINTLV widths and 64-bit dtypes, silently emitting ignored or illegal deinterleaving loads; T.ascend_pipe_barrier("PIPE_V") is silently dropped by AscendC codegen.
Compared with the 30 issues in the previous installment, this wave lands lower down: the previous round was mostly about kernel parameters and high-level semantics (block count, num_stages, dtype tables, pad_value routing), whereas this round is almost entirely on the factory functions for predicates and vector instructions at the T.simd.* layer. Together, the two rounds sweep the Ascend frontend’s “silent failure surface” all the way from parameter validation down to instruction encoding.
Sources: https://github.com/tile-ai/tilelang/issues/3992 , https://github.com/tile-ai/tilelang/issues/3993 , https://github.com/tile-ai/tilelang/issues/3976
4.2 Corresponding fixes for the audit already queued: a dozen-plus validation PRs advancing in parallel (11-12)
The audit is not one-way complaining. Among the 50 main-repo PRs under review and updated the same day, a large batch targets exactly these gaps, and they are almost all “reject illegal input” rather than “add new features”: #3973 changes hf32’s nearest_zero in the CCE backend from NEAREST_AWAY back to NEAREST_ZERO; #3972 aligns the GEMM input dtype table bidirectionally with the two codegen backends (float8_e5m2 was being wrongly rejected on the Python side); #3971 accepts the .dyn spelling of shared.l0c; #3969 rejects pad_value for non-UB destinations; #3968 advances the mask count for ND-to-NZ scatter so the loop can terminate; #3967 and #3970 switch to directly rejecting bit reductions and batch reductions instead of emitting types absent from the template; #3966 rejects PK4_B32 stores on 16-bit vectors; #3965 skips signal latency when running msprof on a non-main thread; #3963 switches to runtime tail masks for partial passes of ND-to-NZ; #3863 rejects FP4 lane counts whose packed size is illegal; #3862 folds the resolved NPU architecture into the kernel cache key.
These PRs share a highly consistent shape: first demonstrate that “this path makes the compiler emit something nonexistent or silently compute wrong,” then add a rejection at the frontend or a correct path during lowering. For a compiler that must be deployed across multiple domestic accelerators, this kind of change is worth more than new operators.
Sources: https://github.com/tile-ai/tilelang/pull/3973 , https://github.com/tile-ai/tilelang/pull/3972 , https://github.com/tile-ai/tilelang/pull/3968
4.3 Three new CUDA-side issues: cp.async two-byte, exponential layout-inference time, Pipelined explicit order (11-12)
The three new CUDA-side issues are each representative. #3988: fp16 T.gemm with M=1 and odd K goes through pipelined cp.async and still emits illegal 2-byte transfers (IsValidCPAsyncTransferBytes fails to catch it); reproducible on both 0.1.15 and main (348e807), while the regression form fixed by #2365 at the time passes—indicating there are still uncovered shapes in the same family. #3987: the time taken by tl.LayoutInference and tl.LowerTileOp grows exponentially with reused subexpressions in T.Parallel; reproduction only needs lowering under sm_120a, no kernel launch required. The corresponding fix #3997 is already queued (bounding the work of condition proofs; the same regression goes from 34.89 seconds—exceeding the 10-second compile limit—to passing). #3996: when T.Pipelined has explicit order/stage and the copy goes through a fragment intermediary, it reports “Fragment buffer access with non-zero index [1]”, yet that index in the error is not in the user program—the error message points at the wrong place.
CUDA fixes under review the same day also include #3998 (fall back to scalar MMA when the shared layout does not match the ldmatrix precondition, fixing #2642) and #3975 (perform a full-CTA barrier before automatically releasing single-CTA TMEM; warp 0 previously did not wait for other warps to finish using it).
Sources: https://github.com/tile-ai/tilelang/issues/3987 , https://github.com/tile-ai/tilelang/issues/3988 , https://github.com/tile-ai/tilelang/pull/3997
4.4 News side: zero hits in Chinese and English query windows; latest arXiv entry still stuck at 11-03 (11-12)
Google News Chinese and English queries returned zero hits in the window; bare queries retrieve 23 to 25 items, but the latest stops at 10-28, and the content is still follow-up reprints of the late-September batch of reports on DeepSeek open-sourcing Ascend-version TileLang components. Hacker News hits are all same-name noise (tag classification tools, project ranking sites), unrelated to TileLang. On arXiv, all:"TileLang" returns 10 entries, the latest being a 11-03 paper on decoupled-inference KV compression, also outside the window.
Zero hits on the news side for several consecutive days is normal for TileLang: project dissemination runs mainly through GitHub and developer communities, not the news indexes, and the reporting subject continues to be repository activity.
Sources: Google News Chinese and English queries (via proxy), Hacker News, arXiv; no relevant entries in the window.
4.5 Daily regression: the same run opens two alerts at two points in time (11-12)
The tilelang-ascend daily test automatically opened one issue each at 10:23 and 15:48 (#2006, #2010), both pointing to the same workflow run 43124767679, both with status “test encountered an error.” The same run being alerted twice indicates alert deduplication is not keyed by run number.
Compared with the previous installment’s 11-11 07:58 run where all 3060 items passed, this installment’s failure has no readable failure list yet, so it can only be recorded as “the Ascend repo’s CI had one failure that day and reported it twice.” Against a backdrop of zero merges for several consecutive windows, this is worth following up: if the failure is concentrated on the mainline Ascend path, it will directly block the next batch of merges after those 4 regression cleanups.
Sources: https://github.com/tile-ai/tilelang-ascend/issues/2010 , https://github.com/tile-ai/tilelang-ascend/issues/2006
V. Trend Observations
First, the cleanup period after a major overhaul is longer than the overhaul itself, and the cost falls on “previously legal programs no longer being legal.” #3512 replaced PTO’s vectorization and type system together, and the four consecutive windows since have all been 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), VMI descriptors (11-12). What these four categories of issues have in common is not “new features computing wrong results” but “old code no longer compiling or producing changed results.” For compiler projects, the debugging cost of such regressions far exceeds that of new-feature bugs—the error location is often on a scalar or address unrelated to the fault, and the stack trace gives no indication of who first installed the wrong assumption.
Second, the backend matrix has shifted from “one player advancing alone” to “multiple players starting simultaneously,” and the forms have differentiated into four categories. This issue’s five commits in the main repo are all on Ascend, but in the same window ROCm opened seven commits in one series, Hygon produced ten commits on one branch in a single day, Sunrise is running its release pipeline, and Hexagon is pushing TileLang to mobile DSP in the form of a personal fork. The four forms (upstream built-in backend, vendor adaptation repo, independent 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 the mainline, forks track their own hardware.
Third, “engineering for agents” is beginning to appear at scale. The harness branch of mlir-ascend gained +7,632 lines in one day, consisting of a skill library, a simulated trace workflow, and a Roofline calculator, serving “people or agents who write operators” rather than the compiler itself; in the same week, the README of the Hexagon fork explicitly states that the backend and models were rewritten by AI agents under the maintainer’s guidance, with 75 PRs in four days accounting for over 80% of the current code. The two developments point in the same direction: the abstraction layer of DSLs like TileLang happens to fall at the granularity where “agents can write it correctly,” so the ecosystem is beginning to lay out toolchains for agents rather than just documentation for humans. This item is worth tracking weekly.
Fourth, the four RISC-V sources have been silent for six consecutive windows, forming a stark contrast with overall activity. The TVM route, the MLIR route, the community index page, and the Sophgo TPU—all four upstreams remain at their original references, while in the same period the main repo merged five commits in a day, ROCm opened seven at once, and third-party forks are advancing a new hardware backend. Beyond the four sources, the CPU line (Weft) has not been updated since reaching an interim conclusion on 11-11. The judgment is unchanged: TileLang work on the RISC-V side remains at the stage of “routes exist, commits are few,” and the truly trackable touchpoint is still whether those two mutually independent implementation routes continue to advance.
Limitations note: This issue’s GitHub anonymous API quota (60 requests/hour) was exhausted early in collection, after which two channels that do not consume quota were used to fill in—PR added/deleted lines and file counts were tallied via the diff endpoint, and commits on non-default branches were retrieved via branch-level atom feeds; therefore the file counts and line counts for the four merges in 1.2 through 1.5 are measured values from the diff endpoint, while the descriptions in the body take the PR summaries returned by the search interface and were not verified line by line against the diff. The in-window pushes of the Hexagon fork were located to fix/hexagon-dma-load-only by comparing the atom feeds of seven branches one by one; the other branches had no new commits during the same period. The official status of the third-party operator repo cannot be verified from the repository itself and has been annotated in the body. On the news side there were only zero hits, with no external reports available for cross-verification.
Appendix: Sources and Verification Notes
| Source | Verification Result |
|---|---|
| tile-ai org, repo by repo (30 repos) | 7 pushed within the window: tilelang, TileOPs, TileOPs-nightly, tilelang-hygon, tilelang-mlir-ascend, tilelang-sunrise, TileOPs.github.io |
| Main repo tile-ai/tilelang | 5 merged (#3614, #3894, #3793, #3796, #3792), +1903/−617, all Ascend; 15 new issues |
| Non-default branches | mlir-ascend harness merged 3 (+7632/−24); hygon tilelang-tileops-fix 10; TileOPs stats 1; TileOPs-nightly snapshots 1; main in all three repos had zero commits |
| Review queue | 71 org-level PRs updated within the window: main repo 50, tilelang-ascend 12, mlir-ascend 5, TileFoundry 2, metax 1, hygon 1 |
| Releases page | tilelang latest v0.1.15 (09-30), TileOPs v0.1.0 (10-27), mlir-ascend v0.1.2.020 (06-11); no new tags within the window |
| Daily tests | tilelang-ascend run 43124767679 raised alerts twice, at 10:23 and 15:48 |
| TileLang×RISC-V four-source sweep | No change (sixth consecutive window); index-page PR API rate-limited, downgraded to reference comparison only |
| Third-party backends | Hexagon NPU fork pushed to fix/hexagon-dma-load-only within the window; CANN community ops-tilelang repo had pushes, official status unverified |
| News side | Zero hits within the window across Google News (EN/ZH), Hacker News, and arXiv |
Full Source List
- [1] tilelang #3614 (batched scheduling for cross-core sync mode=0) — https://github.com/tile-ai/tilelang/pull/3614
- [2] tilelang #3894 (VMI descriptor and logical coverage validation) — https://github.com/tile-ai/tilelang/pull/3894
- [3] tilelang #3793 (scalar cast fact preservation) — https://github.com/tile-ai/tilelang/pull/3793
- [4] tilelang #3796 (preserve explicit SIMD parallel loops) — https://github.com/tile-ai/tilelang/pull/3796
- [5] tilelang #3792 (PTO integer atomic sign adaptation) — https://github.com/tile-ai/tilelang/pull/3792
- [6] tilelang releases page — https://github.com/tile-ai/tilelang/releases
- [7] tilelang #3809 (SM90 shared transpose) — https://github.com/tile-ai/tilelang/pull/3809
- [8] tilelang #3923 (pipelined GEMM operand lifetime) — https://github.com/tile-ai/tilelang/pull/3923
- [9] tilelang #3975 (sync before TMEM release) — https://github.com/tile-ai/tilelang/pull/3975
- [10] tilelang #3998 (fall back to scalar MMA on incompatible shared layout) — https://github.com/tile-ai/tilelang/pull/3998
- [11] tilelang #3954 (pass compile args to hipcc) — https://github.com/tile-ai/tilelang/pull/3954
- [12] tilelang #3955 (8-bit integer vector broadcast) — https://github.com/tile-ai/tilelang/pull/3955
- [13] tilelang #3956 (error on small-tile gemm) — https://github.com/tile-ai/tilelang/pull/3956
- [14] tilelang #3957 (launch bound on waves per EU) — https://github.com/tile-ai/tilelang/pull/3957
- [15] tilelang #3958 (occupancy report) — https://github.com/tile-ai/tilelang/pull/3958
- [16] tilelang #3959 (wave64 reduction workspace) — https://github.com/tile-ai/tilelang/pull/3959
- [17] tilelang #3960 (fp16 fragment register residency) — https://github.com/tile-ai/tilelang/pull/3960
- [18] tilelang #3997 (conditional proof of work upper bound) — https://github.com/tile-ai/tilelang/pull/3997
- [19] tilelang issue #3987 (exponential layout inference time) — https://github.com/tile-ai/tilelang/issues/3987
- [20] tilelang issue #3988 (two-byte cp.async transfer) — https://github.com/tile-ai/tilelang/issues/3988
- [21] tilelang issue #3996 (Pipelined explicit order error) — https://github.com/tile-ai/tilelang/issues/3996
- [22] tilelang issue #3992 (T.simd.pge degradation) — https://github.com/tile-ai/tilelang/issues/3992
- [23] tilelang issue #3993 (predicate width not validated) — https://github.com/tile-ai/tilelang/issues/3993
- [24] tilelang issue #3976 (PIPE_V barrier dropped) — https://github.com/tile-ai/tilelang/issues/3976
- [25] tilelang #3973 (hf32 rounding mode mapping) — https://github.com/tile-ai/tilelang/pull/3973
- [26] tilelang #3972 (GEMM dtype table alignment) — https://github.com/tile-ai/tilelang/pull/3972
- [27] tilelang #3968 (ND-to-NZ mask count) — https://github.com/tile-ai/tilelang/pull/3968
- [28] tilelang-mlir-ascend #220 (Ascend950 simulation trace) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/220
- [29] tilelang-mlir-ascend #221 (operator design skill guide) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/221
- [30] tilelang-mlir-ascend #226 (Gemm quantization example) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/226
- [31] tilelang-mlir-ascend #233 (tl.reshape) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/233
- [32] tilelang-mlir-ascend #232 (tl.sort) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/232
- [33] tilelang-mlir-ascend #231 (tl.print) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/231
- [34] tilelang-hygon #27 (intrinsic and precision fixes) — https://github.com/tile-ai/tilelang-hygon/pull/27
- [35] tilelang-hygon
tilelang-tileops-fixbranch — https://github.com/tile-ai/tilelang-hygon/tree/tilelang-tileops-fix - [36] tilelang-sunrise (Sunrise backend distribution) — https://github.com/tile-ai/tilelang-sunrise
- [37] tilelang-metax — https://github.com/tile-ai/tilelang-metax
- [38] tilelang-musa — https://github.com/tile-ai/tilelang-musa
- [39] TileOPs — https://github.com/tile-ai/TileOPs
- [40] TileOPs-nightly — https://github.com/tile-ai/TileOPs-nightly
- [41] TileFoundry #323 (Mesh selection slice) — https://github.com/tile-ai/TileFoundry/pull/323
- [42] TileFoundry #324 (dimension arithmetic codegen) — https://github.com/tile-ai/TileFoundry/pull/324
- [43] tilelang-hexagon (Hexagon NPU backend fork) — https://github.com/xwhzz/tilelang-hexagon
- [44] ops-tilelang (CANN community operator repo, unverified) — https://github.com/Atlas-Ascend-Fork/ops-tilelang
- [45] tilelang-ascend daily test #2010 — https://github.com/tile-ai/tilelang-ascend/issues/2010
- [46] tilelang-ascend daily test #2006 — https://github.com/tile-ai/tilelang-ascend/issues/2006