TileLang Daily Intelligence Report (2026-11-16)
Research window: 2026-11-13 07:00 ~ 2026-11-16 07:00 (Beijing time, 72 hours, covering Friday through Monday; the previous issue’s cutoff was 11-13 07:00).
Sources: Repository-by-repository checks of the GitHub tile-ai organization (30 repos, 8 pushed within the window) + organization-level merged and in-review ground-truth search (30 merged within the window; 102 in-review updates, of which 59 were newly opened within the window; 406 newly opened issues) + commit-by-commit checks of the main branch (TileOPs 15, tilelang-ascend 9, TileFoundry 3, tilelang 1, hygon 1, mlir-ascend 0) + non-default branch checks (TileOPs stats 1, TileOPs-nightly snapshots 2) + release page checks (no new tags in the six repos within the window) + Google News Chinese and English queries + Hacker News + arXiv. The special RISC-V four-source inspection showed changes in this issue.
This Issue’s Index
-
Today’s Focus: TileOPs makes fifteen commits in three days, shifting focus from the compiler to the operator library; a performance merge the same day triggers a hang that is fixed that same day; within three hours the organization receives 172 Ascend audit issues
- I. Core Project Progress
- 1.1 Main repo sees only one documentation merge in 72 hours, version still at v0.1.15 (11-13)
- 1.2 TileOPs: four performance merges on 11-13, one line each for MoE, GQA, softmax, and W4A16 (11-13)
- 1.3 TileOPs: eleven commits in a single day on 11-15, W4A16 decode tiles another 10–20% faster (11-15)
- 1.4 TileOPs: ping-pong trio, MoE shared experts beat FlashInfer-CUTLASS in every repetition (11-15)
- 1.5 TileOPs: #2901 fixes the hang introduced by #2896, barrier landed in the warpgroup branch (11-15)
- 1.6 TileOPs: #2918 lets the compiled graph be reused across processes, one pyupgrade pass changes 1926 sites (11-15)
- 1.7 TileOPs: manifest adds four spec-only operators, from Embedding to RoPE cosine-sine (11-15)
- 1.8 TileOPs: #2951 splits adaptive pooling by lane, three pooling variants on main fail to compile on narrow-long shapes (11-15)
- 1.9 TileFoundry: #325 unifies all callee naming with a single FunctionRef (11-13)
- 1.10 TileFoundry: dimension arithmetic folded into the scalar renderer, Mesh selection slicing no longer lost (11-13)
- II. Multi-Backend Adaptation (Ascend / Hygon / MetaX / Moore Threads)
- 2.1 Ascend: DSA sparse attention trio merged the same day, forward, backward, and topk reduction all in place (11-13)
- 2.2 Ascend: KDA chunk trio merged the same day, state recurrence kept entirely in UB (11-13)
- 2.3 Ascend: catlass upgraded to v2.2.0, old-version fp32 GEMM silent NaN fixed (11-13)
- 2.4 Ascend: bitwise four-in-one, uint8 PTO wrapper tail-byte out-of-bounds corrupts sentinel (11-13)
- 2.5 mlir-ascend: 22 TLIR dialect PRs under review, unary operators and cross-pipe synchronization added together (11-13 ~ 11-15)
- 2.6 mlir-ascend: inconsistent RTTI link configuration becomes the common blocker for this batch of PRs (11-15)
- 2.7 Hygon: #37 merges ds_read_vector Alt2 and LDS layout inference, #38 under review (11-13)
- 2.8 MetaX and Moore Threads: no commits in window (most recent 11-06 / 11-10)
- II·V. RISC-V Backend and Community (Special Observation)
- 2.9 kubuds TVM route adds hikari branch, five commits deliver K3 port validation and RVV decode grid (11-13 ~ 11-15)
- 2.10 Community index page PR #1 writes the kubuds fork into the README (11-14)
- 2.11 RuyiAI’s MLIR route and Sophgo TPU two lines’ references unchanged (11-16)
- III. Ecosystem and Adopters
- 3.1 TileOPs-nightly: two snapshot rounds correspond to two runs, latest round tests #2955 (11-15)
- 3.2 TileFoundry under review: NCCL communication semantics and Qwen3-MoE training example (11-13)
- 3.3 TileFoundry under review: int8 matrix multiply priced at zero compute time on H200 (11-15)
- 3.4 Adopters: TileKernels and FlashQLA no pushes in window (11-16)
- IV. Community, Tutorials, and Events
- 4.1 406 new issues in three days, two accounts contribute 314 of them (11-15)
- 4.2 cn-sys: 172 opened in a row within 1 hour 46 minutes, numbering continuous from #4208 to #4381 (11-15)
- 4.3 ilovexsir: 142 issues paired with five batch fix PRs, frontend validation systematically filled in (11-15)
- 4.4 Review queue 102 updates, 59 newly opened in window, main repo accounts for 51 (11-16)
- 4.5 News side: Chinese and English queries and arXiv zero hits in window (11-16)
- V. Trend Observations
Today’s Highlights
The center of gravity shifted this window. In the previous issue, the main repo merged five commits in a single day, all landing on Ascend; in this issue, the main repo saw only one documentation commit over 72 hours (an LSP repo link), while TileOPs merged fifteen commits in three days and TileFoundry three, plus nine in the Ascend adaptation repo—nearly all code volume moved to operator libraries and schedulers. TileOPs merged eleven commits in a single day on 11-15, among them expanding the decode-side W4A16 tile from a single consumer warpgroup to two, speeding up seven decode rows by 10–20% and stretching its lead over marlin to more than 1.20x. This is not scattered patching but a coordinated three-layer change across “operator + selector + cost model”: first the selector is made to pick the dual-warpgroup tile on more shapes, then the tile gains metadata loading sliced along K, and finally a nightly 32-row regression confirms no row got slower.
The same day produced the most notable causal chain of this issue. #2896 on 11-13 replaced the 128-row W4A16 tile with a hand-written warp-specialized main loop, cutting chunked-prefill on H200 from 37.3 microseconds to 31.6 microseconds; about 46 hours later, #2901 was fixing it—the out_shared staging of sliced_pair was written inside the conditional branch i // wg_rows == bz, and TileLang sometimes places the barrier guarding this write, which covers all 256 consumer threads, inside that same branch; the other warpgroup can never reach that barrier, the two warpgroups stall on different instances of the same barrier, and the kernel hangs. Worse, the barrier placement is not stable between two compilations of the same cache key, and CI’s persistent kernel cache keeps replaying the bad build, so gpu-smoke on main simply hung. The fix itself is plain—have both warpgroups do the staging, each storing only its own half—but it shows one thing: the correctness of automatic warp specialization no longer depends only on pass order but also on which branch the barrier is inserted into, and that location is opaque to the compilation process.
The second item is on the issue side, and its scale is more striking than the code side. In 72 hours the tile-ai org opened 406 new issues, of which a single account, cn-sys, opened 172 in 1 hour 46 minutes on the evening of 11-15 (numbered #4208–#4381 consecutively, with not one interleaved from another source), and ilovexsir opened 142 in three batches over three days; together the two account for 314, or 77% of the total. cn-sys’s batch ranges from concrete numbers in examples and READMEs (flash-attention claims forward error below 0.5%, while all five checkers in the same directory use 5%; deepgemm’s benchmark uses 8192 as cache_size, but that parameter is in MB, equivalent to allocating 8 GiB to flush a 128 MiB L2) all the way into the compiler proper: auto-schedule’s IR structure visitor misses fields, and PTO codegen’s pointer element type falls back to a hardcoded bfloat16. ilovexsir’s batch is more systematic: the 142 issues correspond to five batch PRs (#4201–#4205, opened in the same minute), each fixing a group of entry points that “should have errored but didn’t”—for example, T.copy(transpose='False') silently enables transpose because the string ‘False’ is truthy, and _make_ascend_target() uses dict(target_dict or {}), dropping all falsy config keys. A pipeline is forming between audit and fix, at the cost of a review queue backlogged to 102.
The third item is that operator migration on the Ascend side has entered the stage of “landing in batches by algorithm family.” On 11-13 alone, tilelang-ascend merged nine commits: the forward pipelined version, backward, and topk reduction trio for DSA sparse attention; the chunk_delta trio for KDA; plus the bitwise four-in-one, the alloc API docs, and a catlass submodule upgrade. These are no longer single-point fixes but organized by model: DeepSeek V3.2’s sparse MLA fuses DQK into a single 512+64=576 GEMM, the Kimi-family KDA absorbs the gated decay into a host-precomputed M matrix, and the backward operator cuts 79395 microseconds to 27388 microseconds. Meanwhile catlass went from v1.0.0-6 to v2.2.0, incidentally fixing a previously fully silent defect: the old CopyL1ToL0A/L0B had no float specialization, so fp32 GEMM output NaN or negative infinity without erroring.
The RISC-V side breaks its silence for the first time in this issue. kubuds’s TVM line opened a new kubuds-riscv-hikari branch alongside mahiru and karen; its five commits include a K3 port and validation report, unlocking the rv64gc build (the USE_LLVM=ON linking problem and an eight-thread build), grid partitioning for decode 4 tiles and prefill 8 tiles, and provenance records for the public XSMTVDotII/vmadot spec; at the same time, the community index page received its first PR, adding the kubuds fork’s project description to the README. The MLIR line (RuyiAI-Stack/tilelang-riscv) and references to the Sophgo TPU repo remain untouched, and the two lines still need to be viewed separately.
I. Core Project Progress
1.1 Main repo sees only one documentation merge in 72 hours, version still stuck at v0.1.15 (11-13)
tile-ai/tilelang’s main had only one commit within the 72-hour window: 210611dc at 11-13 16:20 (UTC), which updated the TileLang LSP repository link in the docs to a new address (#4052). No new operators, no backend changes, no release—the latest tag is still v0.1.15 from 09-30, and the version number has sat in place for six weeks.
Compared with the previous issue’s five Ascend merges in one day totaling +1903/−617, the main repo was nearly static this window. But the stillness was only on main: 51 PRs were updated in the window and 27 newly opened, among which the five batch PRs #4201 through #4205 and the cast lifetime fixes in #4138 and #4207 are substantive changes that simply haven’t crossed the review threshold—this issue’s main-repo output took the form of “audit + open PRs,” not “merges.”
Source: https://github.com/tile-ai/tilelang/pull/4052
1.2 TileOPs: four performance merges on 11-13, one line each for MoE, GQA, softmax, W4A16 (11-13)
On the first day of the window, TileOPs merged four in a row, covering four different workload types. The MoE grouped GEMM FP8 forward changed A’s scale from a producer staging buffer to a TMA box (#2894), letting the 128-row tiles of ds-v3-ep8-prefill-up and -down drop stage_scales from 1 to 0, going from 269.3 µs to 262.0 and from 142.2 to 137.2 on H200 respectively, with the ratio to DeepGEMM rising to 1.096 and 1.157. The W4A16 128-row tile was changed to a hand-written warp-specialized main loop that decodes the next K block while the previous WGMMA is still running (#2896), taking chunked-prefill from 37.3 µs to 31.6. The soft reduction put the 16-bit softmax row on a cluster of two 256-thread CTAs instead of four 128-thread CTAs, completing the “shift by scaled maximum” for the fp16 term with a single FFMA (#2898). The split upper bound for paged GQA is now jointly constrained by the number of key tiles in the window and the combine’s read volume (#2892), incidentally closing the long-pending calibration request #2496.
Source: https://github.com/tile-ai/TileOPs/pull/2894, https://github.com/tile-ai/TileOPs/pull/2896, https://github.com/tile-ai/TileOPs/pull/2898, https://github.com/tile-ai/TileOPs/pull/2892
1.3 TileOPs: eleven merges in a single day on 11-15, W4A16 decode tiles another 10–20% faster (11-15)
Eleven merges in one day on 11-15, of which #2955 was the day’s closer: the W4A16 selector moved the “at most 32 rows” deferred-scaling pick from one 128-thread consumer warpgroup to two, distributing by weight row count and splitting by K slice to preserve CTA count, while stream-K still keeps a single warpgroup. The rationale is that decode tiles are instruction-issue bound while the cost model favors a single warpgroup—measured on that path, issue slot occupancy was 49.5% and DRAM only 54%. As a result, seven decode rows are 10–20% faster than main, leading marlin by at least 1.20x, with prefill-mlp-up unchanged and still 0.7% behind torch. The same day, #2952 merged the target-side kernel’s two tables into one, copying before reading to avoid “dictionary changed size during iteration” under concurrency, and #2904 fixed a call with “no device, no named target, no process default” throwing _Unresolved object is not callable.
Source: https://github.com/tile-ai/TileOPs/pull/2955, https://github.com/tile-ai/TileOPs/pull/2952, https://github.com/tile-ai/TileOPs/pull/2904
1.4 TileOPs: ping-pong hat-trick, MoE shared expert beats FlashInfer-CUTLASS in every repetition (11-15)
Three progressive merges on the same theme. #2919 first had the grouped GEMM selector’s own 64x256 single-warpgroup compact tile run dual-warpgroup ping-pong, so each tile’s epilogue falls under the other group’s mainloop, taking the down-projection GEMM of ds-v3-grouped-t2048 from 1818 µs to 1707. #2953 extended the same technique to the selector’s own 64-row compact tiles, covering fused gate/up and short grouped tiles, and explicitly noted “short grouped tiles still keep one warpgroup because their epilogue is too short to hide anything”; the result is that FusedMoESharedExpert’s ds-v3-grouped-t2048 beats FlashInfer-CUTLASS in every repetition (median 5674 vs 5744 µs), with no program in the 32-row six-operator set getting slower. #2915 then staged the unfused ping-pong grouped GEMM output in 64-column TMA boxes, gaining a fifth ring level (an extra 40 KiB tier for 64x256 tiles).
Source: https://github.com/tile-ai/TileOPs/pull/2919, https://github.com/tile-ai/TileOPs/pull/2953, https://github.com/tile-ai/TileOPs/pull/2915
1.5 TileOPs: #2901 fixes the hang introduced by #2896, barrier landed inside a warpgroup branch (11-15)
This is the most worthwhile read of the issue. After #2896 merged, gpu-smoke on main hung: sliced_pair stages out_shared under the condition i // wg_rows == bz, and after layout lowering only the consumer warpgroup that “keeps that half tile” executes the staging loop; TileLang sometimes inserts the barrier guarding this write—covering all 256 consumer threads—inside this branch, which the other warpgroup never reaches, so the two groups stop on different instances of the same barrier and the kernel waits forever. The explicit barrier added earlier in #2899 was not enough to prevent it—the compiler can still insert barriers inside the branch on its own. The fix moves staging out of the condition: both consumer warpgroups write their own half. What’s truly striking is the second half: the barrier placement varies between two compilations of the same cache key, and the persisted CI kernel cache replays the bad build indefinitely, so one unstable instrumentation became a stable failure.
Source: https://github.com/tile-ai/TileOPs/pull/2901, https://github.com/tile-ai/TileOPs/pull/2899, https://github.com/tile-ai/TileOPs/pull/2896
1.6 TileOPs: #2918 lets compiled graphs be reused across processes, one pyupgrade touches 1926 sites (11-15)
#2918 changes the compilation-boundary key from “one uuid per process” to “operator class + parameter digest + construction-time tensor layout + the instance’s index within the class,” so later-started processes can directly reuse Inductor’s FX graph cache and AOT autograd cache; the cost is that parameters must be re-traced once assigned new values, and in-place modification is rejected outright during tracing. The same PR brings pyright --verifytypes to 100% for the tileops.backend, authoring, and manifest packages (previously 100%/50%/21%) and applies ruff’s UP rules repo-wide—1926 sites across 307 files. Such commits produce no performance numbers, but they determine nightly and CI wait times, and whether newcomers can read the code via type hints.
Source: https://github.com/tile-ai/TileOPs/pull/2918
1.7 TileOPs: manifest adds four spec-only operators, from Embedding to RoPE cos/sin (11-15)
#2920 adds “spec only, no implementation” manifest entries for the four operators missing from the five pilot models: EmbeddingFwdOp (including integer tables, for DeepSeek-V4-Flash’s hash routing), MaskedScatterFwdOp, RMSNormGatedFwdOp (with a sigmoid-gated activation parameter), and RoPECosSinFwdOp (angles per token, allowing x to be a strided view to support tail-column rotation). It also adds the tileops.indexing family, adds a torch reference implementation for each to tests/test_spec_reference.py, and prices embedding, masked scatter, and RoPE by rows read (two roofline functions plus one inline bytes, each with a hand-written oracle). The significance of this approach: nail down the interface and pricing first; the implementation can be filled in by later contributors.
Source: https://github.com/tile-ai/TileOPs/pull/2920
1.8 TileOPs: #2951 splits adaptive pooling by lane, three pooling ops on main fail to compile on narrow-long shapes (11-15)
AdaptiveAvgPool2dLanesKernel splits each bin across a power-of-two number of lanes (at most one warp), with each lane reading its own tap via an unconditional vector load with clamp, then folding with shfl_xor. The driving root cause: the original implementation ran one thread per output over the full bin, giving the unbatched ResNet head only 8 CTAs and DRAM occupancy below 3%. It also fixes a harder problem—adaptive_bin did not widen to int64 when the product overflowed int32, so all three adaptive pooling operators failed to build on main for the shape (1, 1, 150001, 16) -> (100000, 1).
Source: https://github.com/tile-ai/TileOPs/pull/2951
1.9 TileFoundry: #325 unifies all callee naming with a single FunctionRef (11-13)
The IR previously had five ways of naming callees—bound Function/PrimFunction objects, slot-bound instance calls, reserved calls, launch targets, variant tags—and each consumer resolved them its own way, with the evaluator even picking definitions from “the loaded reading” rather than module scope. #325 collapses them into a static leaf FunctionRef(path, name): the path is relative to the caller, a None segment denotes a runtime index passed as a leading operand, and core.resolve(scope, ref) becomes the sole resolver; the resolver also defines “a name declared later in @module” as an error (DeclarationOrderError). Associated cleanup includes computing StoragePlan.scratch before emission, removing relaunch, and having the host entry declare and hand off scratch.
Source: https://github.com/tile-ai/TileFoundry/pull/325
1.10 TileFoundry: dimension arithmetic folded into the scalar renderer, Mesh selection slices no longer dropped (11-13)
#324 folds DimAdd, DimSub, DimMul, DimFloorDiv, DimMod, DimMin, DimMax into render_dim_call in codegen/scalar_expr.py; previously three CUDA emitters each maintained their own + - * / % table, so writing “loop bounds from mesh coordinates” (range(cta.i * 32, cta.i * 32 + 32)) on CUDA directly reported “no codegen handler for DimMul.” #323 fixes another class of silent error: relabeling selections like Mesh(threads[0:16], layout=(16,)) lost the slice at offset 0, so lowering did not emit the contains guard and every thread in that layer ran the entire region; now relabeling selections are slices at any starting point, while whole-layer selections remain unsliced meshes.
Source: https://github.com/tile-ai/TileFoundry/pull/324, https://github.com/tile-ai/TileFoundry/pull/323
II. Multi-Backend Adaptation (Ascend / Hygon / MetaX / Moore Threads)
2.1 Ascend: DSA Sparse Attention Trio Merged the Same Day — Forward, Backward, and topk Reduction All in Place (11-13)
tilelang-ascend merged three DSA (DeepSeek Sparse Attention) operators in two slots on the first day of the window, at 03:25 and 09:49. The forward pipelined version #1870 targets V3.2’s sparse MLA prefill: each query block selects KV rows by top-K, fuses DQK into a single 512+64=576 K dimension for one GEMM (no longer three-stage wgmma), uses a persistent grid of 20 cores, and combineCV automatically switches between Cube/Vector. The backward #1928 uses a two-pass structure, with dKV written out via fp32 GM atomic scatter (under the benchmark shape, each row is contributed by roughly 448 tokens on average), dropping the cumulative time from 79,395 to 27,388 microseconds, a 65.5% reduction. #1884 adds the topk probability-mass reduction, with Expert mode, T.Scope("C") dual-streamed with two AIVs’ T.Scope("V"), and manual cross-core flags.
Sources: https://github.com/tile-ai/tilelang-ascend/pull/1870, https://github.com/tile-ai/tilelang-ascend/pull/1928, https://github.com/tile-ai/tilelang-ascend/pull/1884
2.2 Ascend: KDA Chunk Trio Merged the Same Day — State Recurrence Kept Entirely in UB (11-13)
Three KDA (Kimi Delta Attention) commits were merged the same day at 03:26 and 08:19. #1895 is the dhu path of chunk_delta_bwd: each (b, h) block’s fp32 state [DK, block_DV] stays in UB throughout the backward chunk recurrence without returning to GM, GEMMs use T.mma, and each GEMM gets its own named L0 scratch to cut false dependencies at the name granularity. #1896 is chunk_delta_h_fwd: it refactors the h_new = M@h + K^T@U M-form into a host-side precomputed M = diag(exp2(gk_last)) - K^T@W, fully absorbing the gating so the kernel has zero transcendental functions. #1799 is Chunk Delta H itself, which explicitly abandons fp32 GEMM in favor of bf16 — the rationale being that gate decay keeps the recurrence eigenvalues below 1, so a 1-ULP bf16 error does not amplify over 512 recurrence steps.
Sources: https://github.com/tile-ai/tilelang-ascend/pull/1895, https://github.com/tile-ai/tilelang-ascend/pull/1896, https://github.com/tile-ai/tilelang-ascend/pull/1799
2.3 Ascend: catlass Upgraded to v2.2.0, Silent NaN in Old fp32 GEMM Fixed (11-13)
#1989 upgrades 3rdparty/catlass from v1.0.0-6 (67fd282) to the official tag v2.2.0 (6c0a3242), and makes six TLA API adaptations to src/tl_templates/ascend/common.h (39+/48-), with all helper external signatures unchanged and zero changes to codegen or the Python layer. The primary motivation for the upgrade is a bug fix: the old CopyL1ToL0A/L0B had no float specialization, so fp32 inputs to T.gemm_v0 would output NaN or negative infinity without raising any error; after the upgrade the same kernel is bit-identical. Other benefits include absorbing 17 months of community fixes (zN fractal padding holes, cross-core flag deadlocks, interval DataCopy for small-M GM→L1), and opening four follow-up channels: HF32, GEMM+Bias fusion, quantized side-output, and FP8. The author reports full regression across four independent environments with zero regressions introduced by the upgrade.
Source: https://github.com/tile-ai/tilelang-ascend/pull/1989
2.4 Ascend: bitwise Four-in-One — uint8 PTO wrapper’s Tail-Byte Overrun Corrupts the Sentinel (11-13)
#2000 combines four PRs — bitwise_and/or/not/lshift — into one (due to conflicts sharing common.h and elementwise.py). The core issue is the PTO uint8 tand/tor wrappers: the initial fix used ceil(ColValid / 2) to handle “odd columns’ tail byte not being processed,” but review pointed out that on a 256-byte buffer, buf[0:255] would process an out-of-bounds uint16 pair, clobbering c[255] into a[255] | b[255] — a controlled test showed the sentinel 0xA5 becoming 0x03; the final approach uses floor(ColValid/2) safe pairs plus scalar handling of the trailing valid byte, with PtoSetWaitFlag providing bidirectional cross-pipe synchronization before and after. The same commit also adds a frontend binary_op size assertion on src1: previously dst=64, src0=64, src1=32 would compile and run, with src1 read as 64 elements — an out-of-bounds read of 32.
Source: https://github.com/tile-ai/tilelang-ascend/pull/2000
2.5 mlir-ascend: 22 TLIR Dialect PRs Under Review — Unary Ops and Cross-Pipe Sync Added Together (11-13 ~ 11-15)
tilelang-mlir-ascend’s main had zero commits in the window, but 22 PRs were opened, all advanced by five contributors following the same “six-touchpoint” template: TLOps.td defines the operator, TLOps.cpp writes the verifier, codegen_tlir.cc/.h does TIR→TLIR, ConvertTLToHIVM.cpp does TLIR→HIVM, two FileCheck tests, and CMakeLists.txt registration. The operators added span four categories: unary scalar functions (tl.not, abs, sqrt, rsqrt, rec, exp, ln, cos), tensor operations (tl.flip, tl.pad, tl.gather, tl.transpose, vbitcast), reductions (reduce_sum/max/min/reduce), and synchronization primitives (tl.set_flag + tl.wait_flag are both needed to form a complete producer/consumer pair). tl.flip is the first TL operator with attributes, so usePropertiesForAttributes = 0 was set at the dialect level to bypass TableGen’s Properties storage. On the memory side there are also TL_L0A/L0B/L0C/L1 address spaces and explicit mapping of local/local.var.
Sources: https://github.com/tile-ai/tilelang-mlir-ascend/pull/236, https://github.com/tile-ai/tilelang-mlir-ascend/pull/247, https://github.com/tile-ai/tilelang-mlir-ascend/pull/248, https://github.com/tile-ai/tilelang-mlir-ascend/pull/259
2.6 mlir-ascend: RTTI Link Configuration Mismatch Becomes the Common Blocker for This Batch of PRs (11-15)
Most PR descriptions add the same note: tilelangir-opt fails to link, reporting undefined reference to typeinfo for mlir::ConversionPattern, Dialect and Pass, because the project requires -frtti while the local AscendNPU-IR is built with LLVM_ENABLE_RTTI=OFF. The authors uniformly state they did not change any RTTI-related switch in their PR, with one noting “a colleague on the same project doesn’t hit this, so it may be specific to my local environment.” This does not affect code review, but it is slowing validation of the entire dialect-completion batch — several of the 22 PRs can only be proven by ctest’s dialect-level tests and cannot run the full HIVM lowering.
Sources: https://github.com/tile-ai/tilelang-mlir-ascend/pull/252, https://github.com/tile-ai/tilelang-mlir-ascend/pull/253, https://github.com/tile-ai/tilelang-mlir-ascend/pull/250
2.7 Hygon: #37 Merged ds_read_vector Alt2 and LDS Layout Inference, #38 Under Review (11-13)
tilelang-hygon’s main branch merged #37 at 12:03 on 11-13, covering the Alt2 form of ds_read_vector on HCU and the inferred LDS layout — relatively low-level instruction selection and shared-memory layout work within the HIP-family backends. Another PR, #38, was opened at 09:47 on 11-13 and is still under review; it covers skipping the local panel copy in fused AllGather-GEMM: no longer copying the panel owned by the local rank into the receive buffer, while narrowing FlattenBuffer’s index simplification to guarantee communication offsets are equivalent to the source expression, with benchmarks and correctness checks updated.
Sources: https://github.com/tile-ai/tilelang-hygon/pull/37, https://github.com/tile-ai/tilelang-hygon/pull/38
2.8 MetaX and Moore Threads: No Commits in the Window (Most Recent 11-06 / 11-10)
tilelang-metax’s most recent push was 11-06, and tilelang-musa’s was 11-10; neither had commits, new PRs, or new issues within the 72-hour window. Given the usual cadence (MetaX and Moore Threads typically follow upstream on a weekly basis), a gap in this window is within the normal range, with no sign of stalling.
Sources: https://github.com/tile-ai/tilelang-metax, https://github.com/tile-ai/tilelang-musa
II-V. RISC-V Backend and Community (Special Focus)
2.9 kubuds TVM route adds hikari branch; five commits deliver K3 port validation and RVV decode grid (11-13 ~ 11-15)
kubuds/tilelang-riscv-kubuds opened a new kubuds-riscv-hikari branch within the window (previously only mahiru and karen existed), and landed five commits forming a “compiles, runs, has data” chain: on 11-13, three commits first pinpoint public XSMTVDotII / vmadot spec leads, provide a commit-by-commit log of 54 commits (what was done, why, measured data), then change the RISC-V LLM grid to 4 tiles for decode, 8 tiles for prefill, single-stream execution; on 11-14, one commit unlocks the rv64gc build (the USE_LLVM=ON linking issue and eight-thread build); on 11-15, one commit submits the K3 port and validation report, re-verifying three existing conclusions. This continues the TVM backend route and is a separate implementation from the MLIR rewrite line; the two must not be conflated.
Sources: https://github.com/kubuds/tilelang-riscv-kubuds/commit/bf019bbe24e477d77aa06111a69ef816d347884d, https://github.com/kubuds/tilelang-riscv-kubuds/commit/c3111b8b316f81938a42f7829f74101d122e9f9d, https://github.com/kubuds/tilelang-riscv-kubuds/commit/8401be05da264db5ded98a3b119ff590396bbfc2
2.10 Community index page PR #1 adds kubuds fork to README (11-14)
TongmingLAIC/tilelang-riscv (the community entry page for TileLang for RISC-V) received its first PR since the page was created: #1, opened by lazyparser on 11-14, titled “Add tilelang-riscv-kubuds project details to README”, i.e., adding the kubuds fork’s project description to the “Community Projects” table. The index page’s own copy change is not a technical advance, but this PR marks the first formal registration of the TVM route branch on the community entry page; previously the table listed only the MLIR route and the Sophgo TPU subproject.
Source: https://github.com/TongmingLAIC/tilelang-riscv/pull/1
2.11 RuyiAI’s MLIR route and Sophgo TPU lines unchanged in references (11-16)
RuyiAI-Stack/tilelang-riscv’s main remains at 57c548cc (MLIR rewrite route, target="riscv", TIR→MLIR Linalg/SCF/MemRef/Vector→LLVM IR), and TongmingLAIC/tilelang-tpu’s 18 branch references (including feature/sg2260e-rv-support, CGO2027-submission) are identical one by one to the previous issue, with no new commits, no new branches, no new PRs. Of the four sources, three showed activity and two were silent; this issue only the kubuds side is advancing.
Sources: https://github.com/RuyiAI-Stack/tilelang-riscv, https://github.com/TongmingLAIC/tilelang-tpu
III. Ecosystem and Adopters
3.1 TileOPs-nightly: Two snapshot rounds correspond to two runs; the latest round tests #2955 (11-15)
The snapshots branch received two commits within the window: 49330f29 at 11-15 02:33 corresponds to run 43654108754, testing commit abe2086e (i.e., main after #2892 was merged), and a0566322 at 11-15 22:07 corresponds to run 43738645759, testing commit e4444654 (i.e., #2955). The stats branch’s 52e51373 at 16:38 refreshed the manifest statistics in sync. This chain is the source of several performance conclusions in this issue, and also the basis for the five “widening the lead” issues from #2946 to #2954—all of which cite ratios below 1% in run 43654108754.
Source: https://github.com/tile-ai/TileOPs-nightly/commit/a0566322, https://github.com/tile-ai/TileOPs/issues/2954
3.2 TileFoundry under review: NCCL communication semantics and Qwen3-MoE training example (11-13)
#327 is TileFoundry’s extension from “scheduler” toward “training compiler,” still under review: it adds typed Collective and Exchange communication semantics, a host-side stream and event protocol, communication provider selection, and native NCCL/CUDA fallback; it also introduces byte-region lifetime analysis, completion-aware workspace planning, and JIT identity including callees and physical bindings. On the example side, it provides a Qwen3-MoE training configuration with PP/CP/EP and DMuon FSDP/HSDP, numerical references, an operation timeline, and communication calibration.
Source: https://github.com/tile-ai/TileFoundry/pull/327
3.3 TileFoundry under review: int8 matrix multiplication priced at zero compute time on H200 (11-15)
#334 reports a roofline gap: the target description only includes throughput for bf16, f16, f32, and fp8e4m3, with no i8, and instruction.compute_dtypes does not include i8 either; as a result, analyze --roofline gets the flops count for i8 right (206.1 billion) but reports compute_ns: 0 and judges it “memory-bound”—whereas the H200’s dense INT8 tensor throughput is 1979 TOPS, on par with FP8. The second issue is that T.cuda.sm90.Wgmma does not accept s8 operands, while SM90’s wgmma.mma_async itself has an .s8 form. Both fall under “hardware capability not written into the target description.”
Source: https://github.com/tile-ai/TileFoundry/issues/334
3.4 Adopters: TileKernels and FlashQLA had no pushes within the window (11-16)
deepseek-ai/TileKernels (2074 stars) last pushed on 10-14, and QwenLM/FlashQLA (758 stars) on 10-30; neither had any commits, releases, or new issues within the 72-hour window. Upstream TileOPs has been making intensive changes to the manifest and operator signatures over these three days (#2920, #2944, #2958); adopters typically lag one to two version cycles before following up, so the next window is worth watching for alignment commits.
Source: https://github.com/deepseek-ai/TileKernels, https://github.com/QwenLM/FlashQLA
IV. Community, Tutorials, and Events
4.1 406 new issues in three days, two accounts contributed 314 of them (11-15)
Within a 72-hour window, the tile-ai organization opened 406 new issues, distributed by repo as follows: tilelang 351, TileOPs 44, TileFoundry 8, tilelang-ascend 3; by day: 50 on 11-13, 76 on 11-14, 280 on 11-15. By account: cn-sys 172, ilovexsir 142, lcy-seso 54, oisdoaiu 16, and the remaining seven combined for 22. 313 contain the word “Ascend,” accounting for 77%, in the same direction as last week’s “second wave of auditing,” just an order of magnitude larger.
Source: https://github.com/tile-ai/tilelang/issues
4.2 cn-sys: 172 issues opened in 1 hour 46 minutes, numbered consecutively from #4208 to #4381 (11-15)
All 172 of cn-sys’s issues fell between 18:41 and 20:27 (Beijing time) on the evening of 11-15, with consecutive numbers and no gaps, all tagged with the same environment: Ascend 950PR, CANN 9.0.0-beta.2, Linux 6.6.0-13, Python 3.12, source build main @ 210611d. The content falls into three categories. First, documentation-implementation drift: flash-attention’s README states “forward relative error below 0.5%,” while all five checkers in the same directory use 5%; deepgemm’s benchmark passes cache_size=8192, a parameter whose unit is MB, equivalent to allocating 8 GiB to flush a 128 MiB L2. Second, examples and tests: flash-attention’s correctness test directly reuses the full-size inputs of the performance test, three flash-attention tests lack the requires_ascend gate, and the GEMM example wastes 64–128 KiB of UB in four of seven cases. Third, the compiler itself: auto-schedule’s ControlNode::GetPipeMask does not count its own task, SubstituteVar rewrites loop bounds but misses the loop variable and For annotation, and the memory detector’s SBlockRealize accessor is empty.
Source: https://github.com/tile-ai/tilelang/issues/4381, https://github.com/tile-ai/tilelang/issues/4331, https://github.com/tile-ai/tilelang/issues/4328
4.3 ilovexsir: 142 issues paired with five batch fix PRs, frontend validation systematically filled in (11-15)
ilovexsir’s 142 issues came in three batches: 41 on 11-13, 51 on 11-14, 51 on 11-15, with a highly concentrated theme—”should have errored but didn’t.” These issues land as fixes: at 07:03 on 11-15, five batch PRs #4201 through #4205 were opened in the same minute, covering runtime and target construction, language frontend annotations and sync guards, copy/GEMM hints and padding values, SIMD’s mode/part/dist defaults and destination dtype, and VMI’s pmode/saturate/lane contract and scheduler inputs. Each PR lists the issue numbers it closes and includes an independently reproducible pure-Python script (no NPU or TVM build required). For example: T.copy(transpose='False') silently enables GM→L1 dn2nz transpose because the string ‘False’ is truthy; _make_ascend_target() uses dict(target_dict or {}), dropping all falsy config keys; the isinstance(int) guard in annotate_buffer_versions does not exclude bool.
Source: https://github.com/tile-ai/tilelang/pull/4201, https://github.com/tile-ai/tilelang/pull/4202, https://github.com/tile-ai/tilelang/pull/4203
4.4 102 updates in the review queue, 59 opened within the window, 51 in the main repo (11-16)
Within the window, 102 PRs under review had updates, of which 59 were newly opened: tilelang 27, tilelang-mlir-ascend 22, TileOPs 4, TileFoundry 3, and TileOPs.github.io, tilelang-ascend, and tilelang-hygon 1 each. By author: Areeshh 8, ghostrider0470 and HadiaSaqib 7 each, and six others 5 each. 30 merged against 102 under review—this ratio shows the current bottleneck is not writing code but reviewing it—especially the 22 isomorphic PRs in mlir-ascend, which are essentially a batch that needs the same person familiar with TLIR to go through them one by one.
Source: https://github.com/tile-ai/tilelang/pulls
4.5 News side: zero hits for Chinese and English queries and arXiv within the window (11-16)
Three Google News Chinese queries (TileLang, TileLang RISC-V, tile-ai TileOPs) returned the latest hits still being old pieces from 10-28 and 09-30; the latest English query hit was 10-14, all outside the window; Hacker News yielded only one profiler tool post unrelated to TileLang; the latest arXiv paper containing TileLang stops at 11-03 (KV compression and sharded serving direction), with no new papers in the window. The external TileLang narrative remains stuck at the late-September round of “DeepSeek open-sources Ascend toolbox,” and there are no new reports this period for cross-validation.
Source: https://news.google.com/search?q=TileLang
V. Trend Observations
First, the main repo is quiet while the operator libraries are active—the shift in focus has now persisted for two consecutive windows. In the previous issue, the main repo had five Ascend commits in a single day and TileOPs had zero commits for consecutive days; this issue reverses that: the main repo had one documentation commit in 72 hours, while TileOPs had fifteen commits over three days and TileFoundry had three. Taken together, TileLang’s development focus is moving from “the compiler proper” to “operator libraries + schedulers”: changes to the compiler proper are concentrated on fixing regressions (the cleanup phase of the PTO line), while new capabilities increasingly appear in the form of operator-library tiles, selectors, and cost models. The evidence is not just the number of merges but also the direction of dependencies—TileOPs’ performance conclusions cite TileFoundry’s roofline and scheduling, yet the TileLang main repo gave them nothing new this window.
Second, the correctness of automatic warp specialization has become a first-class risk. The chain from #2896 to #2901 shows that when a performance change delegates synchronization responsibility to compiler-inserted instrumentation, the placement of that instrumentation becomes part of correctness—and it is opaque to upper layers: compiling twice with the same cache key can place a barrier inside a branch or outside it. Combined with CI’s persistent kernel cache, an occasional bad instrumentation gets frozen into a stable failure. There is currently no structural fix, only local workarounds such as “move writes out of conditional branches”; if more tiles adopt ping-pong in the future, similar hangs will likely recur.
Third, auditing has shifted from “finding problems” to “mass-producing problems on a pipeline”. In three days, 406 issues were filed, 314 of them by two accounts: cn-sys’s 172 were written within 1 hour and 46 minutes, and ilovexsir’s 142 map directly to five batch-fix PRs. This pattern does compress the cycle from “defect discovery to fix” (issues and PRs coexist on the same day), but it shifts the pressure to the review side: 102 PRs under review, 30 merged, of which 22 isomorphic mlir-ascend PRs can only be reviewed by a very small number of people. The bottleneck in the next phase will be review bandwidth, not discovery speed.
Fourth, Ascend operator migration is entering “batch landing by model family”. The three DSA pieces correspond to DeepSeek V3.2, and the three KDA pieces correspond to the Kimi family of linear attention; nine commits in a single day all point to complete pipelines for specific models rather than scattered operators. This is good news for downstream—the cost of model migration shifts from “assembling operators one by one” to “porting by family in batches”—but it also means the quality weight of the underlying template library rises. The old catlass defect, where an fp32 GEMM output NaN without reporting any error at all, is a case in point: it hides in a third-party submodule and cannot be caught by operator-level tests.
Fifth, the four RISC-V sources show substantive progress for the first time, but only one of them is moving. kubuds’ TVM route opened a new hikari branch and provided K3 porting validation and RVV grid data; the community index page included this fork for the first time. The MLIR route (RuyiAI) and the Sophgo TPU repo references are unchanged. The division of labor between the two routes is now clear: the TVM route is doing “getting it running, with measured data,” while the MLIR route remains at “design exists, no commits.” The judgment is unchanged—this line is still a small-scale effort, worth tracking window by window, but not suitable for measuring progress week by week.
Limitations: The conclusions in this issue come from the GitHub search API and repository atomic streams, without code-level verification; the performance figures in PRs are taken from the benchmark tables the authors provided in their PR descriptions (H200, CUDA 13.2, PyTorch 2.13) and were not re-tested locally. Issue statistics rely on the search API’s total_count and five pages of paginated results (406 matches the item-by-item tally), and title classification is only a rough keyword-based split. The default branch of tilelang-ascend is ascendc_pto rather than main; the first query fetching commits from main returned 404, and it was re-queried after correction. Neither mlir-ascend’s harness nor hygon’s tilelang-tileops-fix branch had any commits during the window. There were zero news hits, so no external reports are available for cross-validation.
Appendix: Sources and Verification Notes
| Source | Verification Result |
|---|---|
| tile-ai org, repo by repo (30 repos) | 8 pushed within the window: TileOPs-nightly, TileOPs, TileOPs.github.io, tilelang-mlir-ascend, tilelang-hygon, tilelang, tilelang-ascend, TileFoundry |
| Main repo tile-ai/tilelang | 1 merge (#4052 docs), a single commit on main; 351 new issues, 51 PRs under review |
| tilelang-ascend | 9 merges into default branch ascendc_pto, all concentrated on 11-13; 3 new issues, 18 PRs under review |
| TileOPs | 15 merges into main (four on 11-13, eleven on 11-15); 44 new issues, 4 under review |
| TileFoundry | 3 merges into main; 8 new issues, 3 under review |
| tilelang-mlir-ascend | Zero commits on main; 22 PRs opened within the window, all under review |
| tilelang-hygon | 1 merge into main (#37), #38 under review |
| Non-default branches | TileOPs stats 1, nightly snapshots 2; mlir harness and hygon tileops-fix both zero |
| Releases page | No new tags in any of the six repos within the window; tilelang latest v0.1.15 (09-30), TileOPs v0.1.0 (10-27), TileFoundry v0.0.3 (10-21), tilelang-ascend v0.1.2.000 (09-09), mlir-ascend v0.1.2.020 (06-11) |
| Review queue | 102 updated within the window, 59 opened within the window (main repo 27, mlir-ascend 22, TileOPs 4, TileFoundry 3, 3 others) |
| Issues | 406 new: cn-sys 172, ilovexsir 142, lcy-seso 54, oisdoaiu 16, 22 others |
| Adopters | TileKernels (last 10-14), FlashQLA (last 10-30), no pushes within the window |
| TileLang×RISC-V four-source sweep | Changes found: kubuds added kubuds-riscv-hikari branch with 5 commits and community index page PR #1; RuyiAI and Sophgo TPU references unchanged |
| News side | Zero hits across Google News (Chinese and English), Hacker News, and arXiv within the window (arXiv latest 11-03) |
Full Source List
- [1] tilelang #4052 (LSP repo link update) — https://github.com/tile-ai/tilelang/pull/4052
- [2] TileOPs #2894 (MoE FP8 A scale via TMA) — https://github.com/tile-ai/TileOPs/pull/2894
- [3] TileOPs #2896 (W4A16 cross-K-tile single WGMMA in flight) — https://github.com/tile-ai/TileOPs/pull/2896
- [4] TileOPs #2898 (16-bit softmax row dual-CTA cluster) — https://github.com/tile-ai/TileOPs/pull/2898
- [5] TileOPs #2892 (paged GQA split upper bound) — https://github.com/tile-ai/TileOPs/pull/2892
- [6] TileOPs #2955 (W4A16 decode dual consumer warpgroup) — https://github.com/tile-ai/TileOPs/pull/2955
- [7] TileOPs #2952 (target kernel single table and concurrency-safe reads) — https://github.com/tile-ai/TileOPs/pull/2952
- [8] TileOPs #2904 (target binding returns builder) — https://github.com/tile-ai/TileOPs/pull/2904
- [9] TileOPs #2919 (selector 64x256 tile switched to ping-pong) — https://github.com/tile-ai/TileOPs/pull/2919
- [10] TileOPs #2953 (selector 64-row compact tile ping-pong) — https://github.com/tile-ai/TileOPs/pull/2953
- [11] TileOPs #2915 (grouped GEMM output in five TMA box rings) — https://github.com/tile-ai/TileOPs/pull/2915
- [12] TileOPs #2901 (both warpgroups stage, barrier moved out of branch) — https://github.com/tile-ai/TileOPs/pull/2901
- [13] TileOPs #2899 (W4A16 paired launch merged into no-reduction branch) — https://github.com/tile-ai/TileOPs/pull/2899
- [14] TileOPs #2918 (compiled graph cross-process reuse and pyupgrade) — https://github.com/tile-ai/TileOPs/pull/2918
- [15] TileOPs #2920 (four spec-only manifest entries) — https://github.com/tile-ai/TileOPs/pull/2920
- [16] TileOPs #2951 (adaptive pooling split by lane) — https://github.com/tile-ai/TileOPs/pull/2951
- [17] TileOPs #2954 (GQA lead over FA3) — https://github.com/tile-ai/TileOPs/issues/2954
- [18] TileOPs #2957 (nvbench cold measurement protocol) — https://github.com/tile-ai/TileOPs/issues/2957
- [19] TileOPs-nightly snapshot a0566322 — https://github.com/tile-ai/TileOPs-nightly/commit/a0566322
- [20] TileFoundry #325 (FunctionRef unified callee) — https://github.com/tile-ai/TileFoundry/pull/325
- [21] TileFoundry #324 (dimension arithmetic into scalar renderer) — https://github.com/tile-ai/TileFoundry/pull/324
- [22] TileFoundry #323 (renumbering selection preserves slices) — https://github.com/tile-ai/TileFoundry/pull/323
- [23] TileFoundry #327 (NCCL communication semantics and Qwen3-MoE training) — https://github.com/tile-ai/TileFoundry/pull/327
- [24] TileFoundry #334 (int8 tensor core pricing) — https://github.com/tile-ai/TileFoundry/issues/334
- [25] tilelang-ascend #1870 (Sparse MLA forward pipeline) — https://github.com/tile-ai/tilelang-ascend/pull/1870
- [26] tilelang-ascend #1928 (sparse_mla_bwd) — https://github.com/tile-ai/tilelang-ascend/pull/1928
- [27] tilelang-ascend #1884 (topk probability mass reduction) — https://github.com/tile-ai/tilelang-ascend/pull/1884
- [28] tilelang-ascend #1895 (KDA chunk_delta_bwd) — https://github.com/tile-ai/tilelang-ascend/pull/1895
- [29] tilelang-ascend #1896 (KDA chunk_delta_h_fwd) — https://github.com/tile-ai/tilelang-ascend/pull/1896
- [30] tilelang-ascend #1799 (Chunk Delta H) — https://github.com/tile-ai/tilelang-ascend/pull/1799
- [31] tilelang-ascend #1989 (catlass v2.2.0 upgrade) — https://github.com/tile-ai/tilelang-ascend/pull/1989
- [32] tilelang-ascend #2000 (bitwise four-operator merge) — https://github.com/tile-ai/tilelang-ascend/pull/2000
- [33] tilelang-ascend #1642 (alloc API docs) — https://github.com/tile-ai/tilelang-ascend/pull/1642
- [34] tilelang-mlir-ascend #236 (tl.flip) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/236
- [35] tilelang-mlir-ascend #247 (tl.set_flag) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/247
- [36] tilelang-mlir-ascend #248 (tl.wait_flag) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/248
- [37] tilelang-mlir-ascend #259 (tl.transpose) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/259
- [38] tilelang-mlir-ascend #252 (L0A/L0B/L0C/L1 address spaces) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/252
- [39] tilelang-mlir-ascend #253 (reduce four-operator) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/253
- [40] tilelang-mlir-ascend #250 (tl.ln) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/250
- [41] tilelang-hygon #37 (ds_read_vector Alt2) — https://github.com/tile-ai/tilelang-hygon/pull/37
- [42] tilelang-hygon #38 (AllGather-GEMM skips local panel copy) — https://github.com/tile-ai/tilelang-hygon/pull/38
- [43] kubuds/tilelang-riscv-kubuds hikari branch commit — https://github.com/kubuds/tilelang-riscv-kubuds/commit/bf019bbe24e477d77aa06111a69ef816d347884d
- [44] kubuds/tilelang-riscv-kubuds rv64gc build fix — https://github.com/kubuds/tilelang-riscv-kubuds/commit/c3111b8b316f81938a42f7829f74101d122e9f9d
- [45] kubuds/tilelang-riscv-kubuds decode/prefill grid — https://github.com/kubuds/tilelang-riscv-kubuds/commit/8401be05da264db5ded98a3b119ff590396bbfc2
- [46] TongmingLAIC/tilelang-riscv #1 (index page lists kubuds fork) — https://github.com/TongmingLAIC/tilelang-riscv/pull/1
- [47] RuyiAI-Stack/tilelang-riscv repo — https://github.com/RuyiAI-Stack/tilelang-riscv
- [48] TongmingLAIC/tilelang-tpu repo — https://github.com/TongmingLAIC/tilelang-tpu
- [49] tilelang main repo issue #4381 (GEMM example wastes UB) — https://github.com/tile-ai/tilelang/issues/4381
- [50] tilelang issue #4331 (auto target order inconsistent with docs) — https://github.com/tile-ai/tilelang/issues/4331
- [51] tilelang issue #4328 (GetPipeMask misses its own task) — https://github.com/tile-ai/tilelang/issues/4328
- [52] tilelang #4201 (runtime and target parameter validation) — https://github.com/tile-ai/tilelang/pull/4201
- [53] tilelang #4202 (frontend annotations and sync guards) — https://github.com/tile-ai/tilelang/pull/4202
- [54] tilelang #4203 (copy/GEMM hints and padding values) — https://github.com/tile-ai/tilelang/pull/4203
- [55] tilelang PRs under review — https://github.com/tile-ai/tilelang/pulls
- [56] deepseek-ai/TileKernels repo — https://github.com/deepseek-ai/TileKernels
- [57] QwenLM/FlashQLA repo — https://github.com/QwenLM/FlashQLA
- [58] Release pages of each repo — https://github.com/tile-ai/tilelang/releases