Research window: 2026-10-13 07:00 ~ 2026-10-14 07:00 (Beijing time, 24 hours).

Sources: GitHub (full review of pushes across 29 repos in the tile-ai organization; 10 repos had pushes within the window: tilelang, TileOPs, TileFoundry, tilelang-ascend, tilelang-hygon, tilelang-metax, tilelang-mlir-ascend, TileSight, TileOPs-nightly, TileOPs.github.io; within the window the organization merged 28, opened 42, and updated 79 under review; the main repo merged 4, TileOPs merged 18, the two Ascend repos merged 4, TileFoundry merged 2; non-default branches checked individually [ascendc_pto / dev / gcy_merge_main, etc.]; 1 new nightly pipeline snapshot, 1 deployment of the TileOPs docs site) + release page review + community repo review (PTO-ISA SuperNPUBench, easy-tilelang, SwHwCodesign, tilelang-debugger) + multiple Google News queries in Chinese and English (via proxy) + Hacker News + arXiv.

In This Issue

  • Today’s Highlights: MoE dual-operator landing and the foundry performance batch — TileOPs merges 18 PRs in a single day, nightly baseline +22 items (10-13 ~ 10-14)
  • I. Core Project Progress
    • 1.1 Main repo: Two ROCm PRs — DeepSeek mHC pre/post kernels and safe-gated KDA decode example (10-13 ~ 10-14)
    • 1.2 Main repo: Two Ascend compiler-side PRs — SimdVF multi-buffer T.fill (#3455) and bisheng temp file cleanup (#3461) (10-13)
    • 1.3 Main repo under review: SM120 mxf8f6f4 full family, Metal batch, and Ascend review cluster (10-13)
  • II. Multi-Backend Adaptation (Ascend / MetaX / Hygon / Moore Threads)
    • 2.1 Ascend: Documentation fact-check campaign lands 3 PRs, 11 under review; chunk and sparse MLA migration batch opened (10-13)
    • 2.2 MLIR Ascend: NPUIR explicit memory and pipeline control merged (#190); Cube memory allocation API under review (#201) (10-13)
    • 2.3 Hygon: Register pipeline PR updated, new branch and v0.1.15 upgrade line advancing in parallel (10-13)
    • 2.4 MetaX: main catches up to upstream v0.1.15 node; Moore Threads silent (10-13)
  • III. Ecosystem and Adopters
    • 3.1 TileOPs: MoE dual-operator landing — masked FP8 grouped GEMM and DeepSeek grouped top-k (10-13)
    • 3.2 TileOPs: SM89 four-PR closeout — GLA / DeltaNet backward and KDA prefill, plus GQA fix PR (10-13)
    • 3.3 TileOPs: foundry performance batch — 8 PRs merged, 8 newly opened — pooling / normalization / reduction rollout (10-13 ~ 10-14)
    • 3.4 TileFoundry: Qwen3.8-27B prefill example merged — first example to beat SGLang at full length (10-13)
    • 3.5 Nightly pipeline: New snapshot a79a21e4 — 1064 correctness items (+22), 1966 benchmark items (+19), zero failures (10-14)
    • 3.6 Adopters and docs: TileKernels / FlashQLA / TileRT silent; TileOPs docs site deployed (10-13)
  • IV. Community, Tutorials, and Events
    • 4.1 PTO-ISA: SuperNPUBench releases ops-20261013.2 and CUBE direct-connect commits (10-13)
    • 4.2 Community accounts: easy-tilelang tutorial repo updated; SwHwCodesign public repo silent (10-13)
    • 4.3 Media and academia: Google News / HN / arXiv zero new items in window (10-13)
  • V. Trend Observations
    • 5.1 MoE: From paired splitting to paired landing
    • 5.2 foundry performance batch enters daily-update cadence
    • 5.3 ROCm: The third closing line for DeepSeek new-architecture kernels
    • 5.4 Review queue continues to thicken, baseline growth during release silence
  • Appendix: Sources and Verification Notes
  • Full Source List

Today’s Highlights: MoE Dual Operators Land and Foundry Performance Batch — TileOPs Merges 18 PRs in a Single Day, Nightly Baseline +22 Items (10-13 ~ 10-14)

Dates: 2026-10-13, 2026-10-14　Sources: masked FP8 grouped GEMM (#2521) / DeepSeek grouped top-k (#2535) / DeepSeek mHC kernel (#3156) / Qwen3.8-27B prefill example (#229) / nightly snapshot (a79a21e4)

Within the 24-hour window, the organization merged 28 PRs (4 in the main repo, 18 in TileOPs, 3 in the Ascend adaptation repo, 1 in MLIR Ascend, 2 in TileFoundry) and opened 42 new ones; 10 repositories received pushes. The main threads of this window are two things: the dual landing of MoE operators in TileOPs and the merge wave of the foundry performance batch.

  • MoE dual operators land: masked FP8 grouped GEMM (#2521) and DeepSeek grouped top-k routing (#2535) merged on the same day — the former aligns with deep_gemm masked semantics and ships with 14 rows of workloads registered, the latter aligns with vLLM grouped_topk semantics and fixes the SM scheduling issue at T=512 (see 3.1).
  • TileOPs merges 18 PRs in a single day: besides the MoE dual operators, four SM89 backward/prefill PRs (#2546 through #2549), 8 foundry performance PRs, and a GQA fix PR (#2550) all closed on the same day (see 3.2, 3.3).
  • Two ROCm PRs in the main repo: DeepSeek mHC (manifold-constrained hyper-connections) pre/post-processing kernels (#3156) and a safe-gated KDA decode example (#3249) merged — the ROCm line closes out PRs for the third consecutive day (see 1.1).
  • TileFoundry milestone: the Qwen3.8-27B prefill example (#229) merged, the first example in that repo to outperform SGLang across all lengths of 1024 / 4096 / 8192 (see 3.4).
  • Nightly baseline keeps climbing: new snapshot a79a21e4 — 1064 correctness items (+22), 1966 benchmark items (+19), zero failures (see 3.5).

I. Core Project Progress

1.1 Main Repo: Two ROCm PRs — DeepSeek mHC Pre/Post Kernels and safe-gated KDA Decode Example (10-13 ~ 10-14)

Date: 2026-10-13, 2026-10-14　Source: DeepSeek mHC (#3156) / KDA Decode Example (#3249)

  • DeepSeek mHC Pre/Post Kernels (#3156, 2 files, +83/-6, merged early 10-14): mHC preprocessing on CUDA partitions work along 32-thread boundaries within a 96-thread block and uses TF32 GEMM inputs; on CDNA this splits the native 64-lane wavefront, and TF32 is not an MFMA input type. The HIP path now runs the split roles on full target-size wavefronts and converts FP32 mixed weights to BF16 in shared memory for MFMA computation; the CUDA side keeps the original 96-thread / TF32 layout unchanged, with added gfx942 pre/post kernel correctness coverage (32-token preprocessing, 8-token postprocessing, and other cases).
  • safe-gated KDA Decode (#3249, +360/-0): Adds a packed decode kernel — consumes post-convolution packed QKV and raw gating inputs (a, b, A_log, dt_bias), an FP32 state pool, and per-request state indices; performs Q/K L2 normalization and delta-rule recurrent updates, rewrites the specified state slot in place, and outputs [batch, 1, value_heads, value_dim]; a negative state index outputs zeros without modifying the state pool. The implementation is parameterized by dimension, with no model-name/architecture dispatch, and the documentation provides a specialized configuration for GLM-5.3-Flash TP8.

1.2 Main Repo: Two Ascend Compilation-Side PRs — SimdVF Multi-Buffer T.fill (#3455) and bisheng Temp File Cleanup (#3461) (10-13)

Date: 2026-10-13　Source: SimdVF T.fill (#3455) / bisheng Cleanup (#3461)

  • #3455 (+247/-31): Ascend SimdVF’s LowerFill previously supported only whole-buffer fills (each dimension must start at 0 and span the full buffer shape); in multi-buffer pipelines, MaterializeMultiBuffer adds a version dimension to buffers (e.g., a [64] logical buffer becomes a [2, 64] physical buffer), and a T.fill inside the loop only needs to fill the current version slot [v:v+1, 0:64], which the whole-buffer check rejected. This PR changes the target to the actual fill region, computes the write index for each VReg data block based on the region start and extent, checks constant extent, contiguity, and VReg alignment, and supports compact layouts and explicit strides; it adds 3 NPU tests (1D / 2D / two-stage pipelined multi-buffer).
  • #3461 (+2/-1): The bisheng wrapper did not pass keep_for_debug when creating temp directories, causing compiled AscendC sources to always be deleted — whereas nvcc, hipcc, and PTO wrappers all pass keep_for_debug=not env.should_cleanup_temp_files(). This PR aligns the behavior so that TILELANG_CLEANUP_TEMP_FILES=0 also takes effect on the Ascend backend.

1.3 Main Repo Under Review: SM120 mxf8f6f4 Full Family, Metal Batch, and Ascend Review Cluster (10-13)

Date: 2026-10-13　Source: SM120 mxf8f6f4 Family (#3099) / SM120 MXFP8/MXFP4 (#3495) / Metal Batch (#3194 to #3210) / Ascend Independent AIV (#3485) / Windows MSVC (#3510)

  • Two Parallel SM120 Block-Scaled MMA Lines: The kind::mxf8f6f4 full f8f6f4 operand family, sub-byte TMA producer, and official examples (#3099) and SM120 MXFP8/MXFP4 (#3495) are both under review with updates — continuing to expand the Blackwell workstation matrix after the 10-10 SM121 bring-up.
  • Metal Batch (Ten PRs): Segment-normalized accumulators, simdgroup GEMM fragments, scalar warp reductions, BF16 vector types, short constant loop unrolling, torch backend caching and parameter binding, etc. — the Metal backend has entered a continuous polishing phase.
  • Ascend Review Cluster: Independent AIV execution (#3485), access-region overlap analysis supporting repeat/stride (#3506), PTO execution routed through PTO codegen (#3316) and DeepGEMM’s PTO codegen (#3330), scalar sqrt outside VF scope (#3466), SimtVF remat_idx / vector gather pack (#3492), profiling host-device correlation (#3321), etc.
  • Infrastructure and Fixes: CI adds Windows MSVC (#3510), TMA-only producer runs warp 0 only (#3503), bundled TVM discovery and DSL declarations (#3508), test infrastructure caching IR traversal (#3502), three fill / reduce boundary PRs (#3499 to #3501), codegen shared C-family emission (#3289). Open PRs in the main repo total 161 (168 in the previous issue).

II. Multi-Backend Adaptation (Ascend / MetaX / Hygon / Moore Threads)

2.1 Ascend: Documentation Fact-Check Campaign Lands 3, 11 Under Review; Chunk and Sparse MLA Migration Batch Opened (10-13)

Date: 2026-10-13　Source: round semantics (#1612) / bitwise_rshift (#1676) / bitwise_xor (#1681) / Chunk O Forward (#1866) / sparse_mla_topk_reducesum (#1884)

  • Documentation fact-check campaign: Item-by-item verification of T.tile.* APIs (docstring rewrite + on-device testing + API docs) landed 3 PRs in the window — among them #1612 is not just documentation: it changes PTO’s round lowering from CAST_ROUND to CAST_RINT, fixing .5 midpoint rounding away from zero and aligning with ties-to-even semantics (including 17 on-device test cases and codegen assertions); #1676 / #1681 cover bitwise_rshift / bitwise_xor. Another 11 similar verification PRs are under review with updates (axpy, clamp, clear, fill, bitwise_and/or/not/lshift, exp/ln family, relu family, sin/cos, etc.).
  • Operator migration batch opened: wch0810’s chunk family migration — Chunk O Forward (#1866), chunk_o_bwd (#1802), Chunk Delta H (#1799), Chunk Scaled Dot KKT (#1803); three sparse MLA PRs — Sparse MLA SeeSaw (#1869), Sparse MLA Forward Pipelined (#1870), sparse_mla_topk_reducesum (#1884); plus an MHA example (#1879).
  • Quality and regression: Three PRs extending the accuracy standard — applying the accuracy standard to existing examples (#1882), accuracy harness and runner fixes (#1883), moving low-performance operators into examples_experiment (#1881); gemm_v0 fixes incorrect W4A4 int4xint4 GEMM results (#1731). The Ascend repo has 229 open PRs (228 in the previous issue).

2.2 MLIR Ascend: NPUIR Explicit Memory and Pipeline Control Merged (#190); Cube Memory Allocation API Under Review (#201) (10-13)

Date: 2026-10-13　Source: NPUIR control (#190) / Cube memory API (#201)

  • #190 merged (12 files, +1092/-102): Adds an explicit memory and pipeline control surface to TileLang — support for dual-row GM→UB copies with signed pitch and contiguous GM/L1 ND copies; physical buffer views, typed/reshaped views, and mutable scalar storage; exposes GEMM L0 ping-pong events and MMAD / Fixpipe unit flags (including explicit synchronization requirements); supports T.vbrc regions with scalar vector operands and runtime dimensions; adds capability-guarded atomic-mode builtins and is compatible with both versions of the Fixpipe builder API; fixes Expert signed floor div/mod, unsigned arithmetic, integer conversions, and lazy conditional lowering.
  • #201 under review: Explicit Cube memory allocation API (HhpPCc007). The MLIR Ascend repo has 32 open PRs.

2.3 Hygon: Register Pipeline PR Updated, New Branch and v0.1.15 Upgrade Line Advance in Parallel (10-13)

Date: 2026-10-13　Source: Register pipeline (#18) / Block-scaled GEMM layout branch / v0.1.15 upgrade branch

  • Register pipeline (#18): The registered branch gcy_merge_main continues to be updated — new commits in the window include “fix complex wait count” and restoring test_example_gemm_fp8.py; this PR introduces a register pipeline for the Hygon backend (coordinated with the shared-memory pipeline) and warp-diverge handling when exceeding 4 waves, and remains under review.
  • New branch created: feat/hcu-gemm-blockscaled-async-copy-layout (created and pushed on the morning of 10-13) — pointing toward block-scaled GEMM and async copy layout.
  • Upgrade line: The upgrade/upstream-v0.1.15 branch was pushed twice within the window, with the Hygon backend continuously aligning with upstream v0.1.15. The Hygon repo has 1 open PR.

2.4 MetaX: main Catches Up to Upstream v0.1.15 Node; Moore Threads Silent (10-13)

Date: 2026-10-13　Source: MetaX repo / main branch / Moore Threads repo

  • MetaX: The main branch was pushed on the morning of 10-13, with the branch head at the upstream v0.1.15 release node (#3309) — the mirror line has caught up to the upstream release; the dev branch (main adaptation work line) was last pushed on 09-24, with no new adaptation commits in the window. The MetaX repo has 16 open PRs.
  • Moore Threads: No activity since 09-30 (the most recent release branch remains silent); the Sunrise repo also saw no activity in the window.

III. Ecosystem and Adopters

3.1 TileOPs: Two MoE operators land — masked FP8 grouped GEMM and DeepSeek grouped top-k (10-13)

Date: 2026-10-13　Source: masked FP8 grouped GEMM (#2521) / DeepSeek grouped top-k (#2535) / MoE MMA refactor (#2551)

  • masked FP8 grouped GEMM (#2521, 20 files, +1788/-37, closes #2388): MoEGroupedGemmFP8FwdOp lands from spec — FP8 e4m3 grouped GEMM, 1x128 / 128x128 block scales, bf16 output, semantics aligned with deep_gemm’s masked variant; the manifest narrows to the masked layout and is marked implemented, with a 14-row workload: E=8 two rows plus EP=8 decode twelve rows (DeepSeek-V3, Qwen3-235B-A22B, Kimi-K2, batch 64 / 256). A new moe.masked_routed_metadata generator simulates expert imbalance (DeepSeek-V3 b64 empty 4/32, Kimi-K2 b64 empty 13/48); the sm90 kernel is a persistent grid of a TMA producer warp plus two WGMMA consumer warpgroups, tile walking reads row counts on device, empty tiles occupy no CTA, and each weight tile is read from HBM only once.
  • DeepSeek grouped top-k (#2535, 15 files, +1193/-69, closes #2268): FusedTopKFwdOp fills in
    num_expert_group, topk_group, and routed_scaling_factor (vLLM grouped_topk semantics), with FusedMoEFwdOp / shared-expert paths synced; fixes the scheduling issue where the default TOKENS_PER_BLOCK at T=512 lands 32 blocks on 32/132 SMs; documentation makes the fast-math boundary explicit.
  • MoE cleanup refactor (#2551, +20/-30): the compact row-wise layout is now served directly by MoEGroupedGemmMMAKernel (whose program already binary-searches for the group tail by sorted row id), removing the dedicated TightRow kernel and dispatch key — layout coverage extends to SM80 / SM86.

3.2 TileOPs: Four SM89 closures — GLA / DeltaNet backward and KDA prefill, plus a GQA fix (10-13)

Date: 2026-10-13　Source: GLA backward (#2546) / DeltaNet backward (#2547) / KDA prefill (#2548) / compact row-wise GEMM (#2549) / GQA fix (#2550)

  • GLA backward (#2546, +36/-27): hoists the inter-block key gradient computation ahead of the query gradient so the slimmed program finishes using dh before loading h — default shared memory for chunk 32, K=V=128 drops from 114688 to 81920 bytes, falling within SM89’s 101376 limit; adds FP16/BF16 cases.
  • DeltaNet backward (#2547, +166/-113): the three-stage program for chunk 64, K=V=128 fits within 90112 / 98304 / 98816 bytes; an explicit state buffer view cuts the 147968-byte w/u allocation reported by reviewers to 98816; shortens operand lifetimes, reusing state storage for keys and the inverse matrix after their last consumer.
  • KDA fused prefill (#2548, +186/-63): enables KDA fused prefill on SM89 — the compact shared-memory schedule for 128-dim inputs cuts the 256-thread allocation from 182528 to 101120 bytes; the fused path is selected only when the sequence/head block count reaches the SM count; adds four public FP16/BF16 numerical cases and replaces the outdated 2048-sequence rejection assertion with a numerical check.
  • Compact row-wise grouped GEMM (#2549, +45/-11): SM89 supports compact grouped GEMM that recovers segment tails by sorted expert id, reusing the existing MMA program (including empty experts and partial tiles).
  • GQA fix (#2550, 18 files, +341/-173): dense GQA decode’s split log-sum-exp switches to FP32 accumulation — bf16 steps by 0.5 over the 64 to 128 range, and split mixing deviates by up to 0.23 at large scores; autotune now times the program actually launched by forward (including the fused RoPE path), correcting the candidate selection criterion.

3.3 TileOPs: foundry performance batch — 8 merged, 8 newly opened — pooling / normalization / reduction spread out (10-13 ~ 10-14)

Date: 2026-10-13, 2026-10-14　Source: FP8 tiles (#2533) / short tight groups (#2536) / W4A16 (#2545) / wide dense tiles (#2552) / coop2 store (#2566) / foundry under review (#2570 to #2579)

  • GEMM (four): #2533 rearranges FP8 GEMM tiles and moves the 1x128 B scale to the wave kernel (block128 wave kernel renamed gemm_fp8_block128_wave); #2536 selects half-width tiles on compact unfused grouped calls to fill the SMs (3 of 69 descriptors change selection, all faster); #2545 W4A16 jointly scores tiles and K slices (qwen3-32b-prefill-down 78.69 µs vs 88.83, 1.129x); #2552 wide dense tiles issue 176/192 width in a single WGMMA via a new helper, and a new ping-pong kernel reselects for small M (adds the vit-l16-in21k-head workload).
  • Elementwise and store: #2557 shortens GELU’s erf chain by 5 instructions (llama-8b-ffn-decode 1.280 µs vs 1.440, 1.125x); #2565 uses fast division for SiluAndMul’s 16-bit gate (qwen3-8b-swiglu-b32 1.074x, torch-compile ratio 1.015x, flipped positive from 0.945x); #2566 stores row-by-row per warp in scenarios where TMA cannot address the output because N is not a multiple of 8 (vit-l16-in21k-head drops from 294.7 µs to 23.2 µs, 12.7x, faster than cublaslt-best’s 7.5x).
  • RoPE (#2567): adds 16-byte L2 evict-last loads, 2D neox RoPE rotates one vector per thread, and 1D interleaved single-table reads are regularized (neox-2d-2k 1.022x torch-compile, up from 0.983x).
  • Eight newly opened (10-13 evening ~ 10-14 early morning, all under review): softmax row-resident registers with evict-last (#2570), CTA inter-warp scan group tile counts (#2571), MaxPool1dIndices (#2572), AvgPool3d (#2573), RMSNorm split rows by width (#2576), SiluAndMul evict-last (#2577), any/all elementwise tests (#2578), MaxPool3dIndices (#2579) — the pooling / normalization / reduction battlefront spreads out.

3.4 TileFoundry: Qwen3.8-27B prefill example merged — first example to beat SGLang at all lengths (10-13)

Date: 2026-10-13　Source: Qwen3.8-27B prefill (#229) / scaffold merge (#227)

  • #229 (89 files, +38069): adds examples/qwen3_8_27b_prefill-tilelang/ — text prefill for a 64-layer gated-delta-net / attention hybrid, faster than SGLang at all three lengths 1024 / 4096 / 8192; this workload exceeds the H200 inflection point by 36x (at 8192 scoring accounts for only 3.1% of FLOPs, the rest is GEMM), making it FLOP-bound with no traffic to save, and the example uses analyze --roofline as a decision input; the SGLang baseline is given manually (its two columns differ by 9.1% at 8192, and the example must justify which column it takes). Previous examples were all decode and reached only a fraction of SGLang; this is the first example to surpass it at all lengths.
  • #227 (47 files, +1356/-368): tilefoundry scaffold SOURCE OUT_DIR merges — generates an editable standalone CMake project per Target codegen (parameter views, specialized dispatch, launcher, host registration, and placeholder CUDA bodies), and splits codegen artifacts and linking into six module files.

3.5 Nightly pipeline: new snapshot a79a21e4 — 1064 correctness items (+22), 1966 benchmark items (+19), zero failures (10-14)

Date: 2026-10-14　Source: nightly snapshot (a79a21e4) / snapshot metadata

  • New snapshot a79a21e4 (run 38443260386, 10-14 04:18): 1064 correctness items (62 suites, +22 vs. the previous snapshot, 0 failures 0 errors 1 skipped), 1966 benchmark items (53 suites, +19, 0 failures); environment H200, CUDA 13.2, driver 595.71.05, image cu132-torch2.13-tl-0.1.12, with tileops 0.0.1.dev1314+ga79a21e40 in the snapshot (+18 dev number vs. the previous snapshot), tilelang 0.1.12, torch 2.13.0+cu132 / triton 3.7.1 / vllm 0.27.1.
  • The correctness increment keeps pace with the three merge lines in the window — MoE, SM89, and foundry — continuing the rolling rhythm of “one merge, one notch up the chassis.”

3.6 Adopters and docs: TileKernels / FlashQLA / TileRT silent; TileOPs docs site deployed (10-13)

Date: 2026-10-13　Source: TileOPs docs site / TileKernels / FlashQLA

  • Adopters: TileKernels (DeepSeek) and FlashQLA (Qwen) both had no pushes in the window (silent since 09-30); TileRT had no activity in the window.
  • Docs site: the TileOPs site had 1 gh-pages deployment (10-13 morning, MkDocs 1.6.1); the main docs site had no new builds in the window.

IV. Community, Tutorials, and Events

4.1 PTO-ISA: SuperNPUBench Releases ops-20261013.2 and CUBE Direct-Connect Commits (10-13)

Date: 2026-10-13　Source: SuperNPUBench / ops-20261013.2 release

  • New release ops-20261013.2 (including solution/quant/dynamic_mx_quant); commits within the window: matmul_test replaces GM round-trip with direct-connect CUBE TCVT (#234), topk portable gfsim fixture and RES_CHECK ELF release (#233 / #236 merged), rms_norm tree_32k kernel archived (#216), dynamic_mx_quant integrated into one-level build.

4.2 Community Accounts: easy-tilelang Tutorial Repo Updated; SwHwCodesign Public Repos Quiet (10-13)

Date: 2026-10-13　Source: easy-tilelang tutorial repo / SwHwCodesign / tilelang-debugger

  • easy-tilelang/tilelang-tutorials (third-party Chinese tutorial repo) pushed “adapt tilelang new ascend” within the window — the tutorial includes entries such as lower-trace compilation observation for Ascend A5, adapting to the new Ascend backend.
  • SwHwCodesign account now has 2 public repos (AI-Compiler, AI-Infra), none with pushes within the window; the placeholder-repo wave recorded on 10-12 did not continue. tilelang-debugger has no updates (since 10-10).

4.3 Media and Academia: Zero New Additions Across Google News / HN / arXiv Within the Window (10-13)

Date: 2026-10-13　Source: Google News / Hacker News search / arXiv search

  • Multiple query sets in Chinese and English (including Ascend / MetaX / Hygon / Moore Threads combinations) returned zero new hits within the window; historical hits (aggregated articles and analyses from 10-07 to 10-10) have all been covered in previous issues.
  • Hacker News: zero entries over 96 hours; arXiv: no new preprints (the latest relevant item remains the 07-24 TileSight analytical model paper).

V. Trend Observations

5.1 MoE: From Pairwise Splitting to Pairwise Landing

  • The MoE surface picked up two more within the window: masked FP8 grouped GEMM (aligned with deep_gemm masked semantics) and DeepSeek grouped top-k (aligned with vLLM grouped_topk) merged on the same day, plus compact row-wise layout unified into the MMA kernel — TileOPs’ MoE coverage has shifted from “filling in operators one by one” to “closing out by model recipe,” with DeepSeek-V3 / Qwen3-235B / Kimi-K2 workloads registered in batches. On the review side, the same theme (gpt-oss swigluoai expert, permute-align, expert search skip) continues to thicken; MoE is currently the highest-weight direction for TileOPs.

5.2 Foundry Performance Batch Enters Daily-Update Cadence

  • 8 foundry performance merges plus 8 new opens in a single day: after GEMM (tile reordering, joint selection, single WGMMA), pooling, normalization, reduction, and RoPE are all rolling out, each with comparison benchmarks (torch-compile / cuBLASLt / quack). The daily-update cadence of a single author shows that TileOps’ performance engineering has shifted from “attacking key operators” to “daily leaderboard grinding across all operators,” with remaining laggards (MoE four-row, some 16-bit division rows) recorded item by item.

5.3 ROCm: The Third Closing Line for DeepSeek New-Architecture Kernels

  • Following NSA forward/decode and flash-decoding tiles (10-12 to 10-13), this window picked up mHC pre/post kernels and KDA decode examples — the ROCm line has had merges for three consecutive days, with themes highly concentrated on DeepSeek-family new architectures (NSA, mHC) and GLM-5.3 inference. The approach of clearing technical debt (CDNA wavefront size, MFMA input types, LDS/VGPR budgets) item by item indicates that ROCm support is shifting from “getting it running” to “syncing with new architecture releases.”

5.4 Review Queue Continues to Thicken, Base Growth During Release Silence

  • Single window: 28 merges, 42 new opens, 79 review updates: main repo open PRs at 161 (previous period 168), TileOPs at 26 (previous period 19), Ascend repo at 229 — main repo slimming down, TileOPs thickening. Official releases remain silent (main repo latest v0.1.15, 09-30; TileFoundry v0.0.2), but the nightly base continues to climb with correctness +22 and benchmarks +19: the three lines of merges, evaluations, and examples are accumulating static preparation for the next release window.

Appendix: Sources and Verification Notes

Source Verification Result
tile-ai org (29 repos) 10 repos had pushes in the window: tilelang (4 merged, 29 open updates, 9 new), TileOPs (18 merged, 22 open updates, 25 new), TileFoundry (2 merged, 1 new), tilelang-ascend (3 merged, 25 open updates, 5 new), tilelang-hygon (pushes to three branches), tilelang-metax (main synced), tilelang-mlir-ascend (1 merged, 1 open, 1 new), TileSight (dev branch created and branch pushes), TileOPs-nightly (1 snapshot), TileOPs.github.io (1 deploy); org total for the window: 28 merged, 42 new, 79 open updates
Main repo tilelang 4 merged (#3156, #3249, #3455, #3461); 161 open PRs (168 last issue); no new releases in the window (latest v0.1.15, 09-30)
Ascend repos (ascend / mlir-ascend) ascend: 3 merged (#1612, #1676, #1681); 229 open PRs (228 last issue). mlir-ascend: 1 merged (#190); 32 open PRs. Default branch is ascendc_pto (not main); branch-level verification performed
Hygon / MetaX / Moore Threads Hygon: pushes to three branches (gcy_merge_main, feat/hcu-gemm-blockscaled-async-copy-layout [new branch], upgrade/upstream-v0.1.15) plus 1 branch creation; #18 open. MetaX: main pushed to upstream v0.1.15 node (#3309); no new commits on dev. Moore Threads (after 09-30) and Sunrise (after 10-12) had no activity
TileOPs 18 merged (#2521, #2533, #2535, #2536, #2542, #2545, #2546 to #2552, #2555, #2557, #2565 to #2567); 26 open PRs (19 last issue)
TileOPs-nightly New snapshot a79a21e4 (10-14 04:18): 1064 correctness items (62 suites, +22), 1966 benchmark items (53 suites, +19), zero failures and zero errors
TileFoundry 2 merged (#229, #227); 1 open PR; no new releases (latest v0.0.2, 09-10)
TileSight dev branch created 10-13 with branch-level pushes (repo pushed through 10-14 00:16); public branches include main and tilelang (the latter contains TileLang frontend commits); no release content
Adopters and community TileKernels, FlashQLA had no pushes in the window (both 09-30); TileRT no activity; tilelang-debugger no updates (after 10-10); easy-tilelang tutorial repo 1 update; SwHwCodesign 2 public repos, no pushes in the window
PTO-ISA org SuperNPUBench: released ops-20261013.2; 8 commits in the window (including 2 merges)—CUBE TCVT direct connect, topk fixture and ELF, rms_norm archiving, dmxq build integration
Google News / Hacker News / arXiv Google News zero new hits in the window; HN zero items over 96 hours; arXiv no new preprints

Full Source List