TileLang Daily Intelligence Report (2026-11-11)
Research window: 2026-11-10 07:00 ~ 2026-11-11 07:00 (Beijing time, 24 hours; the previous issue’s cutoff was 11-10 07:00).
Sources: Repository-by-repository inspection of the GitHub tile-ai organization (30 repositories, 7 pushed within the window, 3 of which were automated pushes from the docs site, nightly snapshots, and gh-pages) + ground-truth search of merged and in-review items (9 merged, 42 in-review updates, 10 newly opened issues within the window) + commit-by-commit inspection of the main branch (tilelang 3, TileOPs 4, tilelang-musa 1) + release page inspection (tilelang / TileOPs / TileFoundry / tilelang-musa, no new versions within the window; the 0.1.16 version-number PR is still pending) + two adopter repositories + commit-by-commit inspection of third-party repositories (40 commits in the TileLang–Weft series, Hexagon backend repository) + three Google News queries in Chinese and English (via proxy) + Hacker News + arXiv + adjacent technology developments (Meta TLX release). The dedicated 24-hour inspection of four RISC-V sources showed no changes.
This Issue’s Index
-
Today’s Focus: Merges fell from 21 to 9; the main repo shifts to CUDA instruction-level primitives; all four TileOPs PRs concern DeltaNet’s shared-memory budget; the 0.1.16 version-number PR remains pending
- I. Core Project Progress
- 1.1 Mainline: CUDA adds L2 prefetch instructions and six eviction-priority levels; non-default priorities require SM80 or above (11-10)
- 1.2 Mainline: Symmetric clamp for NaN propagation reduced to a single min.NaN.xorsign.abs (11-10)
- 1.3 Mainline: CI no longer restarts the full pipeline on PR label changes; ROCm watchdog raised above the workflow timeout (11-11)
- 1.4 Mainline: The 0.1.16 version-number PR remains pending; no new tag in the window (11-10)
- II. Multi-Backend Adaptation
- 2.1 Ascend: Host-mapped buffer parameters under review; 529 lines added across 14 files (11-10)
- 2.2 Ascend: All 3060 daily tests pass; adapter-path fix for cached kernels under review (11-10)
- 2.3 MetaX: README installation section condensed to a single set of commands; duplicate steps and version-check commands removed (11-10)
- 2.4 Hygon: Distributed three-in-one fused operator under review; Block-Push and Block-Pull communication paths (11-10)
- 2.5 Metal: SIMD-group destination out-of-bounds write reproduced on M4; fix premised on a provable footprint (11-10)
- 2.6 mlir-ascend: Harness knowledge-base hardening merged into the harness branch; Ascend 950 AscendC becomes the sole PR gate (11-10)
- II·V. RISC-V Backend and Community (Special Watch)
- 2.7 Four-source inspection: No changes for a fourth consecutive window; both the TVM and MLIR lines remain at their original references (11-11)
- 2.8 New observation: TileLang–Weft’s CPU/RVV line saw 40 commits in one day; all 553 benchmarks pass numerical checks (11-10)
- III. Ecosystem and Adopters
- 3.1 TileOPs: W4A8 calls for 33 to 40 tokens switched to the 40-token decode tile; off-table grids 4.3% to 11.3% faster (11-10)
- 3.2 TileOPs: DeltaNet forward path split by device shared-memory budget; SM89 regains the chunk 128 path (11-10)
- 3.3 TileOPs: Backward w/u programs reuse buffers; 114944 bytes reduced to 98560 bytes (11-10)
- 3.4 TileOPs: Tests split by cluster and shared-memory facts; SM89 failures drop from 9 to 3 (11-10)
- 3.5 TileOPs: Paged GQA split count bounded by key tiles in the window; under review (11-10)
- 3.6 TileFoundry: Six analyzer and spec-gap issues in one day, two filed cross-repo by TileOPs authors (11-10)
- 3.7 Third party: Hexagon NPU backend grows from a stub into a branch that runs on-device models; LFM2-1.2B runs end-to-end on the NPU (11-10)
- IV. Community, Tutorials, and Events
- 4.1 News side: Zero hits across three Chinese and English query sets in the window; no related entries on HN or arXiv either (11-10)
- 4.2 Adjacent developments: Meta publishes TLX blog post; Triton’s low-level extension already runs on three production workloads (11-10)
- 4.3 New issue: PTO reports duplicate physical segments for multiple Cube/Vector regions; the daily test report states all 3060 tests pass (11-10)
- V. Trend Observations
Today’s Highlights
This window has only 9 merges, less than half of the previous issue’s 21, but the structure has completely changed. The main repo contributed 3, all themed around CUDA instruction-level primitives: #3490 builds L1 and L2 prefetch instructions plus six levels of eviction priority into the global load template, #3605 lowers “symmetric clamping of NaN propagation” directly into a single min.NaN.xorsign.abs, and #3911 eliminates the full CI restart triggered by PR label changes. None of these three is about scoring points for a particular operator; instead, they give kernel authors usable low-level primitives and a cheaper feedback loop.
The second item is that all 4 TileOPs merges are in DeltaNet, and all revolve around the same physical constraint: SM89 has only 101376 bytes of shared memory. #2891 makes the forward pass, at chunk 128 and head dim 128, first check whether the output pass of the state walk (dim_v 64 requires 114688, dim_v 128 requires 131072) fits; if not, it falls back to the recurrence that needs only 98560 bytes. #2890 adds a reuse switch to the backward w/u program, letting S, du, and Aw take over the buffers of dh, v_new, and Au after their last read at equal head dims, compressing 114944 back down to 98560. #2889 honestly marks the LayerNorm split cases that only hold under SM90 clusters and the two linear attention cases that require 196608 bytes, reducing the failure count on SM89 from 9 to 3. Together, the three form a complete “budget—branching—test criteria” loop.
The third item is that the version is still pending. #3887 (version number 0.1.16) has had no updates since 11-09 03:07 and remains unmerged; the latest tag is still v0.1.15 from 09-30. What is really worth noting is a passing mention in #3911: the ROCm runner’s watchdog was raised above the workflow timeout, whereas previously jobs on the AMD leg would be killed at 40 minutes—this explains why PR feedback on the AMD side is always one beat slower than on NVIDIA.
The fourth item is outside the four sources. The four upstreams of the RISC-V special inspection still show zero changes, but the third-party Tianchen-ai/tilelang-weft (an upstream fork adding the Weft CPU programming model, with RISC-V and x86 sharing one set of physical passes) made 17 commits in the window, and its benchmark repo tilelang-weft-bench made another 23, covering SG2044, K1, K3 A100, and x86 for a total of 553 operator cases and 543 external comparisons. Numerical checks pass 553/553, but only 41 cases are “faster than the reference implementation,” with a median native speedup of 0.0580x on SG2044—this is the most honest quantification of the RISC-V CPU-side gap available so far.
Sources: https://github.com/tile-ai/tilelang/pull/3490 , https://github.com/tile-ai/TileOPs/pull/2891 , https://github.com/tile-ai/TileOPs/pull/2890 , https://github.com/tile-ai/tilelang/pull/3911 , https://github.com/Tianchen-ai/tilelang-weft
I. Core Project Progress
1.1 Mainline: CUDA adds L2 prefetch instructions and six eviction priority levels; non-default priorities require SM80+ (#3490)
#3490 (6 files, +185/−1) fills in three previously missing prefetch builtins: L1 prefetch, uniform L1 prefetch, and L2 prefetch, while making “eviction priority” a first-class parameter EvictPriority, with six levels: default, evict normal, evict first, evict last, no change, and no allocate. The global load template is reworked accordingly: the previously hardcoded fixed-width load wrappers are replaced with a templated API that can specify non-coherent loads, L1 eviction priority, and the byte count for L2 prefetch.
The behavioral boundaries are clearly specified: default-priority L2 prefetch uses ordinary global or local instructions, while the remaining priorities map to evict_normal and evict_last and require compute capability 8.0 or above; on devices below that threshold these combinations are invalid. For kernel authors, the significance is that previously, to get behavior like prefetch.global.L2::evict_last, you had to write inline assembly yourself; now you can declare it in TileLang source and the compiler selects the path based on the target architecture.
Source: https://github.com/tile-ai/tilelang/pull/3490
1.2 Mainline: Symmetric clamp with NaN propagation lowered to a single min.NaN.xorsign.abs (#3605)
#3605 (7 files, +203/−1) adds T.clamp_symmetric(x, limit), which on CUDA SM86 and above clamps FP16, BF16, and FP32 operands uniformly to [-limit, limit], preserving signed zero and subnormals; if any operand is NaN, it returns a canonical NaN, without preserving the original NaN’s sign or payload. Non-positive numeric limits are outside the API contract, and the existing T.clamp is unaffected.
In implementation it lowers directly to min.NaN.xorsign.abs.*, one clamp instruction per scalar or per pair of FP16/BF16, with vector inputs going through the existing vectorizable operator registration and CUDA packed emission paths, introducing no runtime checks. The validation criteria are strict: both the language and codegen test suites pass 86 with 2 skipped; SASS inspection under CUDA 13.1 for SM86 and SM100 confirms that FMNMX.NAN.XORSIGN, HMNMX2.NAN.XORSIGN, and HMNMX2.BF16_V2.NAN.XORSIGN all appear as expected in both scalar and packed forms. The value of this kind of change lies in numerical stability: in quantized inference, losing NaN during clamping is a common source of silent errors; once the semantics are delegated to a single hardware instruction, NaN is no longer quietly turned into a boundary value.
Source: https://github.com/tile-ai/tilelang/pull/3605
1.3 Mainline: CI no longer restarts the full pipeline on PR label changes; ROCm watchdog raised above the workflow timeout (#3911)
#3911 (1 file, +2/−2) only changes trigger conditions, but the problem description is worth recording. The main CI previously triggered on both labeled and unlabeled events, and CodeRabbit applies multiple labels to a new PR at once, so the same PR produced multiple identical full CI runs within seconds; the shared concurrency group then canceled these runs one by one, keeping only the last.
The evidence is in #3908: four labels landed with the same timestamp, producing four CI runs, with the original run and the first three replacements all canceled, while no CI job reads PR labels, so these restarts provided no coverage. The fix removes these two event types from the triggers, keeping opened, synchronize, reopened, ready_for_review, and manual triggers, with the concurrency policy continuing to handle cancellation of stale head updates. The same change also raises the ROCm runner’s watchdog threshold above the workflow timeout, so AMD jobs are no longer forcibly killed at 40 minutes.
Source: https://github.com/tile-ai/tilelang/pull/3911
1.4 Mainline: The 0.1.16 version bump PR remains pending, with no new tags in the window (11-10)
#3887 ([Release] Bump version to 0.1.16) was last updated at 11-09 03:07, with no progress in the window, and remains unmerged. None of the four repos’ release pages show a new version in the window: tilelang is still v0.1.15 from 09-30 (tag sequence v0.1.15 through v0.1.11), TileOPs is v0.1.0 from 10-27, TileFoundry is v0.0.4rc1 from 10-25, and tilelang-musa is v0.1.15+musa.1 from 09-30.
Another version-related observation is on the structural side: #3490 and #3605, two CUDA primitive changes, went directly into the main branch, while #3903 (process-parallel compilation) and #3897 (CPU-side cumsum/cummax) are still under review, indicating that the feature surface of 0.1.16 is still thickening rather than closing out. Once the version number lands, downstream will need to check both the main repo tag and the backend repos’ musa.* and ascendc_pto branch lines.
Source: https://github.com/tile-ai/tilelang/pull/3887
II. Multi-Backend Adaptation
2.1 Ascend: host-mapped buffer parameters under review, 14 files with 529 new lines (#3909)
#3909 extends the explicit host-mapped buffer contract established on the CUDA side in #3802 to Ascend: previously Ascend kernels uniformly rejected CPU tensor arguments, even when their storage was visible on the NPU, and callers could only work around this by disabling tensor validation. The change declares buffers through the same T.annotate_host_mapped_buffers(...) API; both the Cython and TVM-FFI bindings validate the mapping via aclrtHostGetDevicePointer and pass down the device-visible address, while retaining the three checks on dtype, shape, strides, and execution device; the C++ binding keeps its original signature and selects the runtime query by existing device type. On the Cython side, the default remains the CUDA resolver, with an optional resolver provided only for Ascend.
This belongs to the same line as #3892 (torch.compile support), #3893 (direct dequantization of L0C accumulators), and #3891 (L2 flush API), all under review in the previous window: what Ascend is currently filling in is interface and semantic consistency, not an operator list.
Source: https://github.com/tile-ai/tilelang/pull/3909
2.2 Ascend: all 3060 daily tests pass, adapter path fix for cached kernels under review (11-10)
On the tilelang-ascend side, the two new records in this window both tell the same story—the regression surface is expanding. The daily test report in #2001 records a run at 11-10 11:38: all 3060 items passed, a 100% pass rate. #2002 ([Bugfix] Restore adapter library path when loading cached kernels) is under review and fixes the issue where the adapter library path is not restored when loading kernels from cache—the kind of problem that only appears on the cache-hit path, where “the same kernel runs correctly the first time but cannot find the library when loaded from cache the second time.”
Compared with previous issues: the latest merges on the repo’s default branch ascendc_pto are still the seven from 11-06, with no new merges in either the 11-09 or 11-10 windows, while the review queue keeps growing (this issue adds #2000 bitwise PTO uint8 wrapping, #1998 automatic synchronization of copy test kernels, and other updates). Test scale has reached 3060 items while merges have stalled, indicating that the bottleneck is indeed in review rather than development.
Source: https://github.com/tile-ai/tilelang-ascend/issues/2001 , https://github.com/tile-ai/tilelang-ascend/pull/2002
2.3 MetaX: README installation section consolidated into a single set of commands, duplicate steps and version-check command removed (#12)
#12 (1 file, +10/−49) is a net-deletion documentation change. The problem is specific: the repeated single-package installation commands in the README led users to install only TileLang-MUSA while missing TVM FFI or environment configuration. The fix keeps only one set of commands in the installation section, installing TileLang-MUSA v0.1.12+musa.5 and TVM FFI together, with links to the MUSA installation instructions before the commands and to the documentation overview after them; the Latest News and TileOPs sections now both point to the installation docs, and the duplicate commands, the separate source-install steps, and the version-check command are removed. The GEMM example is untouched word for word.
The caveat still holds: v0.1.12+musa.5 in the docs is the version number of the SDK-bundled channel, while the latest tag on the open-source release page is v0.1.15+musa.1 from 09-30; the two are not interchangeable.
Source: https://github.com/tile-ai/tilelang-musa/pull/12
2.4 Hygon: distributed trio of fused operators under review, with Block-Push and Block-Pull communication paths (#36)
#36 (6 files, +2022/−0) was updated within a single day, adding three compute-communication fused operators at once: a Block-Push version of GEMM + AllReduce, a Block-Push version of GEMM + ReduceScatter, and a Block-Pull version of AllGather + GEMM, with performance data and implementation notes written in the README. This is the first sizable distributed-direction submission in the Hygon repo—previous work there focused on the single-card side (the scaled cast #33 and the HCU reduce lane swap layout derivation #34 from 11-06).
The three operators correspond to two ways of initiating communication: in Push, the compute side actively pushes results to the peer; in Pull, the compute side pulls the peer’s allgather results. Which path to choose depends on the communication domain size and the overlap window; the README provides comparative data, suitable for topology-based selection. The PR is still open and has not been merged.
Source: https://github.com/tile-ai/tilelang-hygon/pull/36
2.5 Metal: SIMD-group destination out-of-bounds write reproduced on M4, fix premised on provable footprint (11-10)
#3907 reports a silent out-of-bounds write: when copying an explicit 8x8 metal.simdgroup matrix to an 8x1 output, the compiler emits a full simdgroup_store with a row stride of 1; the logical output is correct, but it overwrites the seven elements after the output. The reproduction environment is documented in full—Apple M4, tvm_ffi, compilation cache disabled, and extra backing storage used as a sentinel to detect the out-of-bounds write. The upstream baseline is 72d986e1.
The fix PR #3912 (2 files, +116/−1) does not add a runtime check; instead, before emitting each SIMD-group store, it requires the arithmetic analyzer to prove that the full 8x8 footprint falls both within the destination buffer and within the requested destination region; if it cannot be proven, a readable ValueError is raised during lowering, while full tiles whose last element lands exactly on the boundary are still supported. In the same window, #3906 ([Metal] Resolve the auto execution backend to tvm_ffi) is also under review, part of the same backend’s execution-path convergence. This is the third time Metal has appeared in this window this month, shifting from adding operators to fixing correctness boundaries—usually a sign that people have started running non-toy shapes on Apple chips for real.
Source: https://github.com/tile-ai/tilelang/issues/3907 , https://github.com/tile-ai/tilelang/pull/3912 , https://github.com/tile-ai/tilelang/pull/3906
2.6 mlir-ascend: harness knowledge-base hardening merged into the harness branch, Ascend 950 AscendC narrowed to the sole PR gate (11-10)
#217 (4 files, +140/−16) was merged into the harness branch rather than main, which needs to be stated explicitly—the repo’s default branch is main, and there were zero commits on main in this window. The change makes knowledge accumulated during exploration reliably persist: evidence snapshots are saved alongside hot knowledge records, preventing later updates to test reports from invalidating historical records; the same approval ID is prohibited from approving different knowledge entries; cooled knowledge is stored outside the temporary workspace; when corrupted files appear in the knowledge directory, the system prompts and allows continued viewing of other valid knowledge, but if it cannot confirm whether an approval ID is duplicated, it pauses generation of new cooled knowledge; the way “not applicable” fields are filled in is unified, and the naming of pending-review knowledge versus code candidates is distinguished. Validation is 25 CPU tests passing.
On the review side, #222 ([CI][Ascend] Make Ascend 950 AscendC the only PR gate) is worth noting: narrowing the PR gate to the single Ascend 950 AscendC line means the repo will no longer maintain parallel gates for older Ascend generations. If merged, Ascend 950 will become the sole admission baseline for mlir-ascend.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/217 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/222
II-V. RISC-V Backend and Community (Special Focus)
2.7 Four-Source Inspection: Fourth Consecutive Window with No Changes, Both TVM and MLIR Lines Remain at Prior References (11-11)
Comparing references one by one across the four established sources, all were unchanged within the 24-hour window:
kubuds/tilelang-riscv-kubuds(TVM backend route): the twokubuds-branches remain atkubuds-riscv-karenfc62939 andkubuds-riscv-mahiru33e0c89;mainis only an upstream sync;RuyiAI-Stack/tilelang-riscv(MLIR rewrite route,target="riscv"):mainremains at 57c548c;TongmingLAIC/tilelang-riscv(community index page):mainremains at 4499c2a; no projects added or removed from the table;TongmingLAIC/tilelang-tpu(SOPHGO TPU):mainremains at 303759f,feature/sg2260e-rv-supportremains at 122bc59, and other branch references are unchanged.
This is now the fourth consecutive window with no changes (archived on 11-07, 11-09, 11-10, 11-11). Per the established convention, empty script output is recorded as “no changes” and is not treated as a collection failure.
Source: GitHub four-source reference-by-reference comparison (no changes within the window); no single landing link.
2.8 New Observation: TileLang–Weft CPU/RVV Line Sees 40 Commits in One Day, All 553 Benchmark Cases Pass Numerical Checks (11-10)
Beyond the four sources, a more active RISC-V line has emerged, and this issue marks the first time it has enough factual substance to report. Tianchen-ai/tilelang-weft is a fork of upstream tilelang v0.1.15 that adds the Weft CPU programming model: on top of the native TileLang AST it introduces five categories of author information—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 uses a shared set of physical passes to cover both RISC-V and x86 simultaneously, so that the same source construct enters either the RVV/IME or the x86 SIMD/AMX lowering path. Within the window the repo saw 17 commits, and the benchmark repo tilelang-weft-bench added 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.
The benchmark results are worth quoting verbatim: 150 operator IDs form 553 platform cases (SG2044 133, K1 137, K3 A100 136, x86 147), of which 543 have a comparable external reference implementation. Numerical checks pass 553/553, but only 41 cases are “faster than the reference implementation”; the median native speedup is 0.0580× on SG2044, 0.1143× on K1, and 0.1074× on K3 A100, with the x86 side not listed in that table. The value of this data line is not in the scores but in the fact that it quantifies, for the first time at single-operator granularity, “how far TileLang-written RISC-V CPU operators are from hand-written native libraries”—and it is fully reproducible (the benchmark repo ships its own constraint files and CSV export).
Two technical routes must be distinguished: Weft follows a CPU route that imports the native TileLang AST early into its own Canonical IR, which is not the same implementation as the TVM backend continued by the kubuds- prefixed branches or the MLIR Linalg/SCF rewrite of RuyiAI-Stack/tilelang-riscv; the three must not be conflated.
Source: https://github.com/Tianchen-ai/tilelang-weft , https://github.com/Tianchen-ai/tilelang-weft-bench
III. Ecosystem and Adopters
3.1 TileOPs: W4A8 calls at 33 to 40 tokens rerouted to the 40-token decode tile, 4.3% to 11.3% faster on out-of-band grids (#2888)
#2888 (2 files, +69/−47) fixes a misalignment in a rule-based decision. On an H200 with 132 SMs, once the number of 128-row tiles for a W4A8 call at 33 to 40 tokens exceeds the SM count, the call falls into the “133 to 225 tiles, K = 7168 / 7680 / 8192” band and keeps going down the prefill path, when such calls should actually use the 40-token decode tile.
The root cause is that after one wave, the completion order of the two paths inverts: the prefill kernel splits tiles with stream-K, while the decode kernel’s consecutive terms that run earlier finish first. The original band covered only part of the grid and also wrongly admitted a set of evenly divisible grids—on those grids prefill actually finishes first. The fix narrows the band to keep only the four cases that genuinely belong to prefill (N not a multiple of 128, K not a multiple of 512, the 512−K term divisible on the SM, 64-row weight blocks a power of two and at least 512 KiB apart). On out-of-band grids it is 4.3% to 11.3% faster than main.
The same PR also adds a previously uncovered test: the decode kernel’s cross-CTA ready flag missing a reset. The Result line is marked below bar, i.e. it does not meet that item’s performance target bar, but what is corrected is the decision rule itself—”wrong path changed to right path” rather than “tuning for score.”
Source: https://github.com/tile-ai/TileOPs/pull/2888
3.2 TileOPs: DeltaNet forward dispatches by device shared-memory budget, SM89 regains the chunk 128 path (#2891)
#2891 (2 files, +78/−27) picks up the pit left by #2645: that change sent all calls with dim_k less than twice the chunk length into state walk, and walk’s output pass at chunk 128, dim_v 64 needs at least 114688 bytes of shared memory (131072 at dim_v 128), while SM89 only provides 101376, so those calls are simply unavailable on SM89; the recurrence retained in the kernel needs only 98560 bytes.
The fix makes the kernel take walk only when walk’s lower bound actually fits, otherwise recurrence, and the refusal path reads the smaller of the two bounds. Calls already served on main keep their existing programs across the SM80, SM89, and SM90 budgets.
Performance is recorded honestly: on SM89 main rejects these calls, so the baseline can only be FLA’s chunk_delta_rule forward (whose own chunk is 64); at chunk 128, dim_k 128 this change is 9% to 41% slower than FLA (TileOPs/FLA ratios across six cases: 1.41, 1.21, 1.28, 1.09, 1.28, 1.09); on SM90 main already serves them, and against main the three cases all have a ratio of 1.00, no regression. On the test side, test_deltanet_fwd_serves_chunk_128_in_sm89_shared_memory is added, checking the dispatch decision against SM89’s device facts, so it also runs on SM90 CI.
Source: https://github.com/tile-ai/TileOPs/pull/2891
3.3 TileOPs: backward w/u program reuses buffers, 114944 bytes squeezed to 98560 bytes (#2890)
#2890 (3 files, +155/−20) is the other half paired with the previous item. Since #2625, the backward w/u program must assemble du, dk_corr, and dw on the fly from the walked-out dh, and TileLang 0.1.12 places dh and S, v_new and du in separate buffers despite their lifetimes not overlapping, so the program occupies 114944 bytes, exceeding SM89’s 101376.
The fix adds a reuse switch to compute_w_u_bwd_tl, enabled only under equal head dims: S, du, and Aw take over the buffers of dh, v_new, and Au after their respective last reads, bringing the program down to 98560 bytes; the kernel sets it only when “the bound without the switch exceeds device shared memory,” and refusal likewise reads the bound with reuse. Two other spots are fixed along the way: walk now picks value tiles only among widths that “fit in a single stage” (previously on a batch 8, 16 head, head dim 128/128 16-bit case it picked a tile needing 131072 bytes, failing at launch on SM89); and _parallel_live_bytes now accounts for the c×c attention tile each chunk’s gradient program still holds while assembling partial du.
This last item cascades into a string of conclusions: with the accounting added, refusal makes SM90 also reject three calls that main lets through but that fail to run (16-bit chunk 64, head dim 256/256 needs 237568 bytes to launch on SM90), and identifies that 16-bit chunk 128, head dim 32/256 and float32 chunk 64, 32/256 fail to launch on main, while 16-bit chunk 128, 16/256 does not even compile. Backward benchmark (bf16, batch 4, 4096 tokens, 8 heads, chunk 64, head dim 128/128): 1.530 ms vs. FLA 1.554 ms, ratio 0.98.
Source: https://github.com/tile-ai/TileOPs/pull/2890
3.4 TileOPs: tests dispatch by cluster and shared-memory facts, SM89 failures drop from 9 to 3 (#2889)
#2889 (3 files, +26/−2) is the test-side companion to these two fixes. The LayerNorm fused-add case asserting “rows still normalized per row when far from zero” has a split branch that builds FusedAddLayerNormHeldKernel, and that kernel holds these rows only when the cluster can split each 12288-wide float32 row; only SM90 has clusters. On SM89 the held kernel’s rules send such rows to the fragment kernel, so the four cases added by #2811 and #2846 still fail after the numerical check passes. The fix calls check_held_launch first and skips the split cases on devices without clusters; the five 8192-wide split cases are likewise skipped on SM89.
It also explicitly marks two cases that only fit on SM90 as sm90: the bf16 GLA backward added by #2637 (chunk 128, head dim 64/64) and the float32 DeltaNet forward added by #2645 (chunk 128, 64/128), each needing 196608 bytes of shared memory. Regression data: on a single SM89, the four test files go from 40 passed, 9 failed, 3 skipped to 35 passed, 3 failed, 14 skipped; the remaining three are chunk 32 GLA backward, which need a kernel change to pass. On a single SM90 all 46 items pass in both versions, showing the skip markers did not take away SM90 coverage.
Source: https://github.com/tile-ai/TileOPs/pull/2889
3.5 TileOPs: paged GQA split count bounded by in-window key tiles, under review (#2892)
#2892 (2 files, +28/−11) is under review and changes the upper bound of GQAPagedFwdKernel._splits_for. Previously the split count was computed to cover the SM count before the window narrowed the key range, so a three-tile window would open 16 splits at batch 1; without RoPE it would keep growing toward “resident,” doubling the cost on the combine side—combine must read each split’s partial per row tile on a single CTA.
The new rule first caps the cp.async split count at the number of in-window key tiles; without RoPE, beyond the SM-covering count the upper bound is isqrt(key_tiles * block_N / block_M), the point where combine’s partial row count is comparable to one chunk’s key row count. The PR comments include the calibration data requested by #2496, covering regret statistics over 379 configurations. #2893, newly opened in the same window, is the follow-up list for this family and remains open.
Source: https://github.com/tile-ai/TileOPs/pull/2892 , https://github.com/tile-ai/TileOPs/issues/2893
3.6 TileFoundry: six analyzer and spec-gap issues in one day, two filed cross-repo by a TileOPs author (11-10)
TileFoundry opened 6 new issues in the window, from two sources.
The first group is a direct product of the source-spec reorganization in #312; zhen8838 opened four in a single minute at 02:42, all marked “found by reading code, not reproduced with a program”: #314’s JIT cache key prints only the entry function and topology, with callees appearing only by name, so two Modules with identical entry text but different callees get the same cache key; #315 defaults to the first valid topology when the caller gives no topology_level, instead of asking the target which levels declare max_physical_units; #316 in loops with depth less than 4 advances the cursor by one loop-body duration rather than the whole loop span, so work after the loop does not wait for the loop to finish; #317’s passes.md §7.1 states that on-chip memory callee return values must be copied into the caller’s buffer, otherwise it must fail with LoweringError, but convert_hir_to_tir has no corresponding rejection.
The second group was filed by TileOPs author lcy-seso at 12:04 and 12:29, both pressing the analyzer with real operators: #318 uses an elementwise clamp on H200 for placement pricing, pointing out that once the CTA grid exceeds one wave, the split input’s footprint is multiplied by the wave count, and noting the same behavior was reported in #298, #303, #307, #311, all closed as not planned, and still reproduces on the current build; #319 schedules a 4096×7168×16384 resident SM90 NT GEMM and yields five findings at once, the hardest being that check --help says 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 what it should be.
The shape of this issue group is worth noting: the analyzer’s mispricing is exposed not by unit tests but by putting real operators on it, and the filer comes from the operator-library side.
Source: https://github.com/tile-ai/TileFoundry/issues/314 , https://github.com/tile-ai/TileFoundry/issues/319 , https://github.com/tile-ai/TileFoundry/issues/318
3.7 Third party: Hexagon NPU backend grows from a shell into a branch that runs on-device models, LFM2-1.2B runs end-to-end on the NPU (11-10)
In the 11-09 window, xwhzz/tilelang-hexagon was still “README unchanged from TileLang, no verifiable backend changes”; after the 22:07 push in this window the content is filled in. The fork adds a TileLang backend for Qualcomm’s Hexagon NPU, developed and validated on the v79 cDSP of a Snapdragon 8 Elite phone: tile ops such as T.copy, T.gemm, T.Parallel, T.reduce_*, T.Pipelined compile to HVX vector code, HMX matrix multiply, VTCM buffers, and DMA streams, executed on the NPU via FastRPC.
The validation target is LFM2-1.2B: batch 1 runs end-to-end, with prefill, decode, and sampling all written in TileLang and running on the NPU; projection layers 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 vs. a bf16 reference of 16.29. The significance for TileLang’s boundary is clear: previous “non-NVIDIA, non-Ascend” backends were mostly CPUs or domestic accelerator cards, and Hexagon is the first branch to put a complete decode loop on a phone DSP, and it gives an accuracy comparison rather than throughput alone.
Source: https://github.com/xwhzz/tilelang-hexagon
IV. Community, Tutorials, and Events
4.1 News side: zero hits across three Chinese and English query windows; HN and arXiv also show no relevant entries (11-10)
Google News Chinese “TileLang” returned 28 items, “TileLang or tile language operator compilation” 10 items, and “BAAI operator or open-source operator library” 37 items; English “TileLang” returned 23 items and “TileOPs or TileRT” 10 items. After filtering by publication time, all were zero within the window; the latest hits remain the 10-28 Caifuhao article and the 10-21 SemiAnalysis piece. Hacker News hits are all same-name noise (Senior Pipeline Crisis, LimbaKit, Monoprint, VoxaDub, etc.), unrelated to TileLang. On arXiv, all:"TileLang" latest is still the 11-03 KV compression paper v2, outside the window.
Consecutive days of zero hits on the news side are normal for TileLang: project propagation runs mainly through GitHub and developer communities, not news indexes, so the reporting subject continues to follow repository activity.
Sources: Five Google News queries in Chinese and English (via proxy), Hacker News, arXiv; no relevant entries within the window.
4.2 Adjacent developments: Meta publishes TLX blog post; Triton’s low-level extensions already run on three production workloads (11-10)
On the same day, the Meta engineering blog published “How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale,” covering Triton Low-Level Extensions: adding explicit warp-group and cluster-level orchestration on top of Triton’s block-based programming model, letting developers use execution-side hardware capabilities such as warp specialization, asynchronous transfers, and tensor memory without giving up Triton’s development speed. The post lists three destinations already running in production—recommendation system training, recommendation system inference, and large-scale LLM pretraining—and provides a chart of kernel count growth inside Meta over 10 months: Triton kernel count grew about 3.8x, CUDA about 2.0x.
It belongs in the TileLang daily because it marks another solution on the same track: TileLang adds automatic scheduling and multiple backends upward from a tile-level DSL, while TLX adds explicit orchestration downward from Triton. The two give opposite answers to “who writes kernels”—TLX lets ML engineers keep using Triton and drop down locally, while TileLang has authors explicitly describe tiles and pipelines and hand them to the compiler. By comparison, what TileLang has added this year is mostly backends and operator correctness; the orchestration-layer abstraction still rests on the T.Pipelined and T.annotate_* interface set.
Sources: https://engineering.fb.com/2026/11/10/open-source/tlx-a-hardware-native-gpu-compiler-for-production-ml , https://arxiv.org/abs/2605.10905
4.3 New issues: PTO multiple Cube/Vector regions report duplicate physical sections; the rest center on the analyzer and docs (11-10)
Ten new issues were opened within the window. Besides the six TileFoundry issues already covered in 3.6, there are four more: #3901 ([BUG][PTO] Multiple explicit Cube/Vector regions fail with duplicate physical section error) reports a duplicate physical section error when multiple explicit Cube/Vector regions are declared in the same kernel—a structural limitation during the Ascend PTO lowering phase, with no fix PR yet; #2893 is a follow-up checklist for paged GQA; #2001 is the daily test report for tilelang-ascend (all 3060 items passed). The Metal one, #3907, was already covered in 2.5 along with its fix PR.
One more pattern is worth noting: both #3901 and #3907 are cases where “two individually legal regions in the same kernel only fail when combined.” Such problems never appear in single-region tests and only surface when real fused operators are assembled. Nearly all correctness issues in recent editions have this shape.
Sources: https://github.com/tile-ai/tilelang/issues/3901 , https://github.com/tile-ai/tilelang/issues/3907
V. Trend Observations
First, the main repo’s output focus is shifting from operators to primitives. #3490’s prefetch and eviction priorities, #3605’s NaN clamping, the under-review #3897 CPU scan operator, and #3903’s process-parallel compilation share a common trait: “give authors a new tool” rather than “make some operator a few points faster.” The acceptance criteria for such changes differ too: #3605 uses the presence or absence of three instructions in SASS as evidence, and #3490 uses a compute-capability threshold as its boundary—both can be rechecked independently of specific shapes. For downstream users, this means more usable patterns after 0.1.16, but each new primitive has its own architecture threshold, so cross-generation ports must be checked one by one.
Second, the shared-memory budget has become an explicit constraint on the linear-attention line. SM89’s 101376 bytes and the demand-side figures of 114688, 131072, and 196608 determine how large the chunk is, whether to use walk or recurrence, and on which architecture tests run. The solutions in #2890 and #2891 follow the same idea: rather than forcing down demand, let dispatch choose a path based on device facts and use the same bound in refusal—preferring to refuse early over crashing at launch. Once this pattern runs smoothly on DeltaNet, the GLA family will most likely copy it—the three chunk-32 failures left by #2889 are exactly the next block.
Third, the testing approach is shifting from “cross by architecture” to “judge by device facts.” #2889 did not delete or permanently skip the SM89 cases; instead it first checks check_held_launch and then decides to skip. The two new tests in #2891 and #2890 also branch decisions based on SM89 device facts, so they still run on SM90 CI. This approach is more robust than skipif markers based on device compute capability, because it verifies the decision logic itself rather than the hardware.
Fourth, TileFoundry’s mispricing problem keeps being pushed back by the operator side. #318 states outright that the same behavior was reported in four closed issues, all marked not planned; #319 gives five at once with a specific 962.1 vs. 1213.6 GFLOP deviation. The analyzer is the premise of TileLang’s “price first, then choose scheduling” route; a 26% pricing deviation will make automatic selection decide wrongly. Such gaps won’t slow any single operator, but they will make automation systematically choose wrong on boundary grids.
Fifth, divergence on the RISC-V side is widening rather than converging. The four sources (TVM route, MLIR route, SOPHGO TPU, community index) are uniformly stalled, while the third-party Weft line does 40 entries a day and has already quantified the gap down to individual operators (SG2044 median 0.0580x). Taken together, the two facts show that the stagnation is not for lack of people working on it, but because the official four sources have not formed a loop of “automatically discover gaps, automatically verify, automatically queue”—the Weft line only took off after building its own benchmark engine. If this 553-case benchmark can be cited by the official four sources as an acceptance standard, progress on the RISC-V side will for the first time have a comparable denominator.
Appendix: Sources and Verification Notes
Source Verification Table
| Source | Verification Result |
|---|---|
| tile-ai organization (30 repos) | 7 pushes within the window: tilelang, TileOPs, tilelang-musa, tilelang-mlir-ascend, plus 3 automated pushes (docs site, gh-pages, nightly snapshot) |
| tilelang main repo | 3 merges to main (#3911, #3605, #3490); 38 open updates; 2 new issues (#3901, #3907); no new tags |
| TileOPs | 4 merges to main (#2888, #2891, #2890, #2889); 2 open updates; 1 new issue (#2893) |
| TileOPs-nightly | 1 automated push to the snapshots branch (11-09 21:56); no commits on the default branch |
| TileFoundry | No merges, no pushes within the window (latest 11-09 11:33); 6 new issues (#314 through #319, all open) |
| tilelang-ascend | No merges on the default branch ascendc_pto within the window (latest 11-06); 9 open updates; daily tests 3060 items all passing (11-10 11:38) |
| tilelang-hygon | No merges within the window (latest 11-06 #33, #34); distributed fused operator #36 under review (+2022 lines) |
| tilelang-metax | No pushes within the window (latest 11-06 07:42) |
| tilelang-musa | 1 documentation change merged (#12); no new release on the releases page (latest v0.1.15+musa.1, 09-30) |
| tilelang-mlir-ascend | Zero commits on main within the window; #217 merged into the harness branch; #222 and others under review |
| Release pages (tilelang / TileOPs / TileFoundry / musa) | No new versions within the window: 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 remains open, no updates since 11-09 |
| Adopters TileKernels / FlashQLA | No pushes within the window (10-14 / 10-30) |
| Third-party TileLang–Weft / weft-bench | 17 and 23 commits respectively within the window; all 553 benchmark cases passed numerical checks, 41 cases faster than the reference implementation; latest release v0.1.15-weft.2 (11-05) |
| Third-party tilelang-hexagon | Pushed 11-10 22:07; README expanded with Hexagon NPU backend description and LFM2-1.2B end-to-end results |
| TileLang×RISC-V four-source inspection | No change (all four upstream references identical to the previous snapshot, 24-hour window; fourth consecutive window) |
| Google News (five groups, Chinese and English) / Hacker News / arXiv | Zero hits on the news side within the window; all HN hits were same-name noise; latest relevant arXiv item is 11-03 (outside window) |
| Adjacent technology developments | Meta TLX blog post published 11-10; a same-track comparison, not a TileLang project development |
Full Source List
- [1] tilelang CUDA L2 prefetch and eviction priority (#3490) — https://github.com/tile-ai/tilelang/pull/3490
- [2] tilelang NaN propagation symmetric clamping (#3605) — https://github.com/tile-ai/tilelang/pull/3605
- [3] tilelang CI trigger condition convergence and ROCm watchdog (#3911) — https://github.com/tile-ai/tilelang/pull/3911
- [4] tilelang 0.1.16 version number (#3887) — https://github.com/tile-ai/tilelang/pull/3887
- [5] tilelang Ascend host-mapped buffer parameters (#3909) — https://github.com/tile-ai/tilelang/pull/3909
- [6] tilelang JIT process-parallel compilation (#3903) — https://github.com/tile-ai/tilelang/pull/3903
- [7] tilelang CPU-side cumsum/cummax sinking (#3897) — https://github.com/tile-ai/tilelang/pull/3897
- [8] tilelang Metal automatic execution backend resolution (#3906) — https://github.com/tile-ai/tilelang/pull/3906
- [9] tilelang Metal SIMD-group destination out-of-bounds issue (#3907) — https://github.com/tile-ai/tilelang/issues/3907
- [10] tilelang Metal SIMD-group out-of-bounds write fix (#3912) — https://github.com/tile-ai/tilelang/pull/3912
- [11] tilelang PTO multi-Cube/Vector region duplicate physical segment issue (#3901) — https://github.com/tile-ai/tilelang/issues/3901
- [12] tilelang-ascend daily test report (#2001) — https://github.com/tile-ai/tilelang-ascend/issues/2001
- [13] tilelang-ascend cached kernel adapter path fix (#2002) — https://github.com/tile-ai/tilelang-ascend/pull/2002
- [14] tilelang-ascend PTO bitwise operation uint8 wrapping fix (#2000) — https://github.com/tile-ai/tilelang-ascend/pull/2000
- [15] tilelang-musa README installation section convergence (#12) — https://github.com/tile-ai/tilelang-musa/pull/12
- [16] tilelang-hygon distributed fused operator trio (#36) — https://github.com/tile-ai/tilelang-hygon/pull/36
- [17] tilelang-mlir-ascend harness knowledge base hardening (#217) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/217
- [18] tilelang-mlir-ascend Ascend 950 AscendC single-PR gate (#222) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/222
- [19] TileOPs W4A8 40-token decode tile rule (#2888) — https://github.com/tile-ai/TileOPs/pull/2888
- [20] TileOPs DeltaNet forward SM89 branching (#2891) — https://github.com/tile-ai/TileOPs/pull/2891
- [21] TileOPs DeltaNet backward w/u buffer reuse (#2890) — https://github.com/tile-ai/TileOPs/pull/2890
- [22] TileOPs tests split by cluster and shared memory (#2889) — https://github.com/tile-ai/TileOPs/pull/2889
- [23] TileOPs paged GQA split upper bound (#2892) — https://github.com/tile-ai/TileOPs/pull/2892
- [24] TileOPs paged GQA follow-up checklist issue (#2893) — https://github.com/tile-ai/TileOPs/issues/2893
- [25] TileFoundry JIT cache key missing callee function (#314) — https://github.com/tile-ai/TileFoundry/issues/314
- [26] TileFoundry topology_level default value (#315) — https://github.com/tile-ai/TileFoundry/issues/315
- [27] TileFoundry loop span costing (#316) — https://github.com/tile-ai/TileFoundry/issues/316
- [28] TileFoundry passes specification and sinking inconsistency (#317) — https://github.com/tile-ai/TileFoundry/issues/317
- [29] TileFoundry CTA grid per-wave costing (#318) — https://github.com/tile-ai/TileFoundry/issues/318
- [30] TileFoundry persistent NT GEMM five findings (#319) — https://github.com/tile-ai/TileFoundry/issues/319
- [31] Third-party TileLang–Weft (Weft CPU programming model fork) — https://github.com/Tianchen-ai/tilelang-weft
- [32] Third-party TileLang–Weft Bench (553-case CPU operator benchmark) — https://github.com/Tianchen-ai/tilelang-weft-bench
- [33] Third-party tilelang-hexagon (Hexagon NPU backend) — https://github.com/xwhzz/tilelang-hexagon
- [34] Meta TLX engineering blog (Triton low-level extension) — https://engineering.fb.com/2026/11/10/open-source/tlx-a-hardware-native-gpu-compiler-for-production-ml
- [35] Meta TLX paper (arXiv 2605.10905) — https://arxiv.org/abs/2605.10905
Verification Notes
This issue’s window is 24 hours (Wednesday). Merge counts, open counts, and new issue counts are taken from GitHub search ground truth (three query types: is:merged, is:open, created, all sorted by merged/updated/created time descending), and cross-checked against per-repo main-branch commits: 9 merges = tilelang 3 + TileOPs 4 + tilelang-musa 1 + tilelang-mlir-ascend 1, where the mlir-ascend merge went into the harness branch (that repo’s default branch is main, which had zero commits in this window), as noted in the body. Performance and shared-memory figures are taken directly from the tables and Result: lines in each PR body, with no secondary conversion; the Result lines or comparison baselines for #2888, #2891, and #2890 are recorded as in the original text, including cases that fell short of targets and cases slower than FLA. The six TileFoundry issues come from two filers: #314 through #317 were derived from code reading after specification reorganization without programmatic reproduction, while #318 and #319 include real operators and build numbers; the body distinguishes them accordingly. The RISC-V four-source inspection script output was empty and is recorded as “no change” per the established convention; the Weft line in 2.8 is a third-party observation newly discovered in this issue, not one of the four sources, and its source is explicitly labeled as per-commit history and README tables from third-party repos. The five Google News queries on the news side went through a proxy, and returned items filtered by publication time yielded zero within the window; Meta TLX is an adjacent same-track development and is not counted as TileLang project progress.