TileLang Daily Intelligence Report (2026-11-19)
Research window: 2026-11-18 07:00 ~ 2026-11-19 07:00 (Beijing time, 24 hours; previous issue cutoff was 11-18 07:00).
Sources: Repository-by-repository check of the GitHub tile-ai organization; 11 of 30 repositories had pushes within the window. Organization-level merged and in-review ground-truth search: 38 merged, 43 in-review updates, 20 newly opened issues. Commit-by-commit main-branch check: TileOPs 23, tilelang 3, TileFoundry 2, topa 3, MLIR Ascend 3, harness 1, Sunrise 1, Hygon 1. Release check: tilelang remains at v0.1.15; Sunrise released 0.1.15+sunrise.1.2.1. Also includes Google News Chinese and English queries, Hacker News, arXiv, and the four-source TileLang×RISC-V sweep.
Issue Index
- Today’s Highlights: TileFoundry lands Metal target and automatic differentiation on the same day (11-18)
-
Today’s Highlights: TileOPs merges 23 commits in one day; FP8 quantization chain completed, two performance debts closed same day (11-18)
- I. Core Project Progress
- 1.1 Main repo CPU backend adds serial fallback for two logical reductions: any_of and all_of (11-18)
- 1.2 topk regression test sets resource limit for Z3, eliminating timeouts on CUDA CI (11-18)
- 1.3 Ascend autotuning tests remove cross-NPU probing to avoid contaminating subsequent cases (11-18)
- 1.4 CUDA Tile IR backend RFC resubmitted: only converts the parts Tile IR can express (11-18)
- II. Multi-Backend Adaptation (Ascend / Hygon / Moore Threads / MetaX / Sunrise)
- 2.1 MLIR Ascend mlir-dialect branch merges tl.shl, tl.shr, tl.cast in one day (11-18)
- 2.2 MLIR Ascend in review: tl.vrelu, tl.abs, tl.or, tl.xor and RoPE kernel (11-18)
- 2.3 Hygon single commit aggregates twelve fixes, covering intrinsics, precision, and integer vectors (11-18)
- 2.4 Sunrise releases 0.1.15+sunrise.1.2.1; MetaX preparing v0.1.12+musa.5 (11-18)
- II·V. RISC-V Backend and Community (Special Focus: TVM route / MLIR route / SOPHGO TPU / Community Index Page)
- 2.5.1 Four-source sweep shows no changes; citations for both routes consistent with previous issue (11-19)
- III. Ecosystem and Adopters (TileOPs target list / TileKernels / topa / Documentation site)
- 3.1 Two mHC operators retired as public API breaking change; three HyperConnection operators take over in review (11-18)
- 3.2 Seven performance target issues turn “distance to fastest library” into an acceptance checklist (11-18)
- 3.3 TileKernels comparative timing proposal: test image still pinned to tilelang 0.1.12 (11-18)
- 3.4 topa runs a full review round with one command; GPU exclusively held only during timing (11-18)
- 3.5 Two documentation site updates; adopter repositories show no commits within window (11-18)
- IV. Community, Tutorials, and Events
- 4.1 News side quiet: Chinese and English searches, Hacker News, and arXiv all show no in-window coverage (11-19)
- 4.2 Organization activity: 38 merges, 43 in-review updates in a single day; main repo at 9,001 stars (11-18)
-
V. Trend Observations
- Appendix: Materials and Verification Notes
- Source Verification Table
- Complete Source List
Today’s Highlights: TileFoundry Lands Metal Target and Automatic Differentiation on the Same Day
In a single day on 11-18, TileFoundry merged two platform-level changes. #354 adds an Apple Metal target, and #355 adds automatic differentiation. The former extends the compilation pipeline to Apple GPUs; the latter lets the framework generate backpropagation itself. Neither produces runnable kernels: Metal only goes as far as TIR, and automatic differentiation only generates HIR modules.
On the Metal side, metal.Target("apple.m2_pro.gpu") is added. The topology is threadgroup to thread. Hardware parameters are written in TOML with public sources cited. The capability table directly adopts 27 native MSL function names.
Under T.metal there are nine instruction operators in total. The matrix group is ldmatrix, stmatrix, matmul, and mma. The reduction and shuffle group is reduce, prefix, shuffle, shuffle_fill, and ballot. A single call covers G SIMD groups; quad groups are expressed with the same set of operators on a mesh of width 4.
One file per operator. typeinfer, access, and evaluator rules are written in the same file, with layouts using named constants. MSL does not specify the lane mapping for matrices; the author measured on an M2 Pro and wrote the specification accordingly. HIR is matched via hir_matches. MatMul on fragment operands lowers to matmul, and Reshard between memory and fragment lowers to ldmatrix and stmatrix. This change supersedes the earlier #294, which could not be rebased due to target interface restructuring.
On the automatic differentiation side, differentiate(module, wrt, recompute) is added. It returns two modules: <Name>Forward and <Name>Backward. The forward pass returns the result and a saved tuple. The backward pass receives the result gradient and saved, and returns gradients one by one according to wrt.
Backward rules are written per operator in their respective files and registered via the newly added interfaces/autodiff. Coverage includes all operator classes of the built-in Qwen model, plus ReLU and Log. It supports sibling and child-slot calls, tuple parameters, MeshRegion, and loops with statically countable iteration counts.
Loops only stack values that change per iteration. Values used only by induction variables, captures, and constants are recomputed in the backward pass, so layer loops do not copy weights. Gradient layouts follow DTensor.
Source: https://github.com/tile-ai/TileFoundry/pull/354.
Today’s Focus: TileOPs lands 23 merges in one day, FP8 quantization chain completed, two performance debts closed same day
On 11-18, the TileOPs main branch took 23 merges, running from 10:17 to 21:50 Beijing time. The main thread was FP8: per-token grouped quantization, short-row GEMM streaming, paged decode rewrite, and WGMMA primitivization — four commits entering main on the same day.
FP8QuantPerTokenGroupFwdOp introduces UE8M0 scales (#3026). Output is bit-for-bit identical to DeepSeek-V4-Flash’s act_quant(x, 128, "ue8m0"), with scales stored as float8_e8m0fnu. A new inplace pseudo-quantization means the strided kv[..., :448] in the model no longer needs to be copied first.
The UE8M0 path uses bfloat16x2 for reduction and scaling, and drops the clamp before the saturating e4m3 conversion. The reference implementation is vLLM’s per_token_group_fp8_quant; the numerics for DeepSeek and TileKernels are given in comments.
The FP8 paged GQA decode kernel was rewritten wholesale (#2997). Each CTA takes one request’s packed rows, one KV head, and a segment of keys. Each FP8 key/value tile is dequantized only once, and a TMA warp handles page copies. The ceiling comes from shared memory: on H200, an 8-row tile can serve 44,287 requests, and a 64-row tile 33,023.
Short-row FP8 now goes through 64-row block streaming (#3010). One producer warpgroup moves b through a deep TMA ring, and one consumer warpgroup accumulates the FP8 WGMMA partial sums for each 128-wide K step into float32.
The main branch previously padded 8 or 16 rows up to 64- or 128-row tiles. That left 160 tiles landing on 132 CTAs, or falling back to split-K. Measured throughput was only 31% to 52% of DRAM peak.
Two cleanups landed in the same batch. #3030 switches the four k32 WGMMA per step in the FP8 stream to TileLang primitives. dense_wgmma_helper.h is deleted accordingly, and the TrailingFence template parameter used only by that kernel is removed with it. #3029 fixes shared memory accounting: the scale buffer was previously summed in raw bytes without accounting for 1024-byte alignment. The shape m=16, n=163840, k=3584 therefore failed at launch.
Two performance debts were closed the same day. #3023 rewrites TopKSelect: calls with fewer than twice the SM count of rows go to a register kernel. It reads one row from global memory at a time, and a 4096-bucket histogram determines the threshold; long rows go through radix, loading four keys per 16 bytes. Three decode rows went from roughly 2x behind vLLM to 1.22 to 1.34x ahead, closing #3016.
#3022 makes the MLA cache copy each row as 16-byte vectors, with up to 1024 threads. The main branch left the copy to layout inference, turning the 448-wide latent into 7 scalar 2-byte accesses per thread. This was the direct cause of DeepSeek-V4-Flash decode lagging vLLM.
Source: https://github.com/tile-ai/TileOPs/pull/3026.
I. Core Project Progress
1.1 Main repo CPU backend adds serial lowering for two logical reductions: any_of and all_of
#4389 adds dedicated lowering for tl.any_of and tl.all_of to the CPU backend. Previously these two operators could not go through the regular CPU pipeline. The implementation adds a new LowerCPUAnyAll pass, inserted before LowerTileOp. The lowering is a serial scan in row-major order with short-circuiting: any_of stops on the first true value, all_of stops on the first false value.
Empty reductions take the standard identity elements: false for any_of, true for all_of. Numeric elements are converted to booleans by nonzero-ness, and NaN in floating point and bfloat16 is treated as true. Reductions in while conditions are reinitialized and rescanned on each condition evaluation, preserving the original semantics. The pass is also exposed to the CPU Python transformation interface.
Source: https://github.com/tile-ai/tilelang/pull/4389.
1.2 topk regression test sets a Z3 resource limit to eliminate timeouts on CUDA CI
#4407 addresses the timeout of test_topk_candidate_overflow on CUDA CI. The cause was costly Z3 queries issued during the simplification phase. The fix sets arith.z3_rlimit to 100 when the test loads the kernel, keeping the other existing JIT options. The change applies only to this regression test.
The author generated TIR locally for the four topk specializations and confirmed that Simplify completes with the limit set. The author also notes: there is no CUDA code generation on the local machine, and no remote GPU available. Full compile time and GPU correctness were not verified.
Source: https://github.com/tile-ai/tilelang/pull/4407.
1.3 Ascend autotuning test removes cross-NPU probe to avoid contaminating subsequent cases
#4401 deletes the other_npu branch in the set_autotune_inputs test. That branch performed a cross-NPU copy to determine whether the runtime supports inter-device input alignment. On some Torch-NPU and CANN combinations, this copy aborts the NPU context.
Catching RuntimeError and skipping the case does not restore the context. Subsequent unrelated cases in the same pytest process then fail as well, so the failure manifestation depends on execution order. After removal, three coverage cases remain: inputs captured on the current NPU stay on the current NPU, inputs captured on CPU are moved to the benchmark device, and passed tensors must match the benchmark device.
Source: https://github.com/tile-ai/tilelang/pull/4401.
1.4 CUDA Tile IR backend RFC resubmitted: only converts what Tile IR can express
#4402 and #4403 submit the RFC for the CUDA Tile IR backend, in Chinese and English respectively, under docs/compiler_internals/tileir/. The goal is to generate CUDA Tile IR from TileLang programs and hand it to tileiras for compilation and scheduling. Layout, pipelining, warp specialization, and instruction selection are all decided by tileiras, optimized per hardware generation.
The backend only converts semantics that Tile IR can express correctly. Semantics that cannot be converted produce a compile error and the program is rejected; full compatibility with the existing CUDA backend is not a goal. The benefits cited by the author: good performance from autotuning alone, without modifying kernels or hand-writing architecture-specific schedules.
This version is a redo of the earlier #3247. Reviewers noted that the old implementation had too many try-catch blocks, lowered case by case for each form seen, built a separate IR, and wrote passes in roughly 18,500 lines of Python. The new RFC redesigns around these four points. The PR is currently a draft, with discussion in #4403.
Source: https://github.com/tile-ai/tilelang/issues/4403.
II. Multi-Backend Adaptation (Ascend / Hygon / Moore Threads / MetaX / Sunrise)
2.1 MLIR Ascend mlir-dialect branch merges tl.shl, tl.shr, tl.cast in one day
The MLIR Ascend repo’s merges land on the mlir-dialect branch, not main. The branch head advanced from 06:23 to 08:47 (UTC) on 11-18, merging three TLIR operators in one day.
#211 adds tl.shl, lowered to hivm.hir.vshl, reusing the existing NpuirShl helper. #212 adds tl.shr, supporting only i16, i32, i64, with round as a boolean attribute defaulting to true. That PR honestly documents one known mismatch: rhs is modeled as a tensor operand, while HIVM’s vshr has a ScalarOnlyHWTrait<1> constraint. The root cause is that the constructors of NpuirShr and npuir_shr do not accept scalars, left for follow-up.
#214 adds tl.cast, with round_mode restricted to six values: round, rint, floor, ceil, trunc, odd. That PR also changes a dialect-level setting: TL_Dialect adds usePropertiesForAttributes = 0. The reason is that tl.cast is the first operator in this dialect with a non-operand attribute, and going through ODS’s default Properties generation would produce a broken CastOp. All three PRs include FileCheck tests, taken from real captured output.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/214.
2.2 MLIR Ascend under review: tl.vrelu, tl.abs, tl.or, tl.xor and RoPE kernel
Four more TLIR operators are under review. tl.vrelu is in #261, tl.abs in #238, the latter lowered to hivm.hir.vabs. tl.or is in #225, tl.xor in #228. All four follow the same pipeline of “dialect operator + TIR lowering + HIVM conversion.”
#263 takes a different path, directly adding a rotary position embedding kernel rather than filling in individual operators. Separately, #260 replaces the ascend-mfu skill in the harness with ascend-roofline. The new skill uses a Cube/Vector model to give peak, counts, roofline, MFU, BWU and hardware references. Its FMA probe calibrates to 26.3 TFLOPS on this device and build, 88% of the nominal 30 TFLOPS.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/263.
2.3 Hygon single PR bundles twelve fixes, covering intrinsics, precision and integer vectors
The Hygon repo has only one merge in the window (#27), but its contents cover twelve areas. On the intrinsic side, it fixes hexp call ambiguity and cos/sin dispatch. On the precision side, it fixes fp64 truncation failures and __builtin_bit_cast size mismatches. On the vector codegen side, it fixes shuffle lowering and enables full-wave shuffle reduction.
New capabilities include: rng operations backed by hipRAND, buffer load rebasing, and a cache_warmup intrinsic. There are also several cleanups: 8-bit integer vector broadcasting, pipelined epilogue unrolling for runtime iteration counts, shuffle mask guards, and an ordering fix in FP4 codegen. On the CI side, it pins the NumPy ABI. The commit is signed by two Hygon engineers.
Source: https://github.com/tile-ai/tilelang-hygon/pull/27.
2.4 Sunrise releases 0.1.15+sunrise.1.2.1; MetaX preparing v0.1.12+musa.5
The Sunrise repo merges #11, bumping the version to 0.1.15+sunrise.1.2.1. That PR syncs three upstream sources — TileLang, TileOPs and TileKernels — and updates the TANG and PTPU integrations. The tag appeared the same day, with the previous version being 0.1.15+sunrise.1.2.0.
The MetaX repo has no commits on main in the window. Activity is concentrated in open #13 and the release branch: the release/v0.1.12-musa.5-source branch exists, and #13 is for releasing the v0.1.12+musa.5 source and MUSA documentation. The Moore Threads main repo likewise has no new commits.
Source: https://github.com/tile-ai/tilelang-sunrise/pull/11.
II-V. RISC-V Backend and Community (Special Focus: TVM Route / MLIR Route / SOPHGO TPU / Community Index Page)
2.5.1 Four-source sweep shows no change; references for both routes match last issue
The 11-19 sweep covers four sources. First, the kubuds- prefixed branches of kubuds/tilelang-riscv-kubuds. Second, main of RuyiAI-Stack/tilelang-riscv. Third, the TongmingLAIC/tilelang-riscv index page. Fourth, main and new branches of TongmingLAIC/tilelang-tpu.
References for all four sources match the previous issue’s records, with no new commits, no new PRs, and no additions or removals in the index page’s project table. The conclusion covers only the four sources above. Neither the TVM route (kubuds- branches) nor the MLIR route (RuyiAI-Stack) saw new activity this week, so no further analysis is provided in this issue.
Source: no new commits.
III. Ecosystem and Adopters (TileOPs target list / TileKernels / topa / docs site)
3.1 Two mHC operators retired as a public API break, three HyperConnection operators take over, under review
#3021 removes MHCPreFwdOp and MHCPostFwdOp, along with their kernels, manifest entries, workloads, benchmarks, and tests, following the plan in #2796. DeepSeek-V4-Flash and GLM-5.3-Flash switch to HyperConnection’s pre, post, and head operators. The mhc.expansion manifest primitive is also removed, since only the deleted operators used it.
This change is registered in scripts/ci/api_compat_allow.txt as a reviewed public API break. That file records reviewed interface removals, and the removal decision must be confirmed during review.
The three replacement proposals are under review: #3032 adds pre, #3035 adds post, #3039 adds head. Separately, #3036 adds sqrt-softplus and hash routing for DeepSeek-V4-Flash to FusedTopKFwdOp. #3038 adds attention sink and index-only selection to DSADenseFwdOp. This group of changes points to the same thing: rearranging the sequence-modeling and attention interfaces around DeepSeek-V4’s actual structure.
Source: https://github.com/tile-ai/TileOPs/pull/3021.
3.2 Seven performance-target issues turn “gap to the fastest library” into an acceptance checklist
Seven new performance-target issues were opened in the window, each with hardware, shape, and target. #3018 tracks the gap to FlashAttention 3 on GQA variable-length vision rows with head dim 72. #3019 tracks the gap to vLLM on MLA cache writes. The corresponding fixes are #3023 and #3022, with #3022 closing #3019 the same day.
#3037 gives the full DSADense comparison table: H200 at 1500 MHz, bf16 dense peak around 811 TFLOPS. CSA rows reach about 58% of peak because each pair’s softmax and the next pair’s key gathering sit between two matrix multiplies. HCA’s 4k rows pay the cost of 64 key blocks, counting 160 keys as 192.
#3031 tracks the gap to TileKernels for HyperConnection pre on decode and medium sequences. #3034 tracks the decode rows of GatedPoolCompress, targeting SOTA on that shape.
Source: https://github.com/tile-ai/TileOPs/issues/3037.
3.3 TileKernels comparison-timing proposal: the benchmark image is still pinned to tilelang 0.1.12
#3028 proposes adding DeepSeek’s TileKernels to the benchmark comparison so that DeepSeek-V4 rows are timed directly against DeepSeek’s own kernels. The obstacle is the version: TileKernels requires tilelang no lower than 0.1.15, while the benchmark image is pinned to 0.1.12. No benchmark can currently import it.
The proposal gives three mappings. FP4QuantPerGroupFwdOp and FP8QuantPerTokenGroupFwdOp map to per_token_cast. SiLUAndMulFwdOp maps to quant.swiglu_forward. The acceptance condition is that these benchmarks actually time TileKernels in CI and that the nightly run records them as a comparison. This explains why, in the descriptions of the two quantization proposals filed the same day, TileKernels’ numbers can only be given in comments.
Source: https://github.com/tile-ai/TileOPs/issues/3028.
3.4 topa runs a full review round with one command; the GPU is exclusive only during timing
topa merged #41 and #43, turning review into repeatable rounds. #41 makes every PR change trigger a new attempt: a new head gets a full review, everything else only a recheck. Results are posted only when the PR is unchanged and still open. On the GPU side, a per-card-locked topa-bench is added, which holds a card exclusively only during timing and yields it to other rounds the rest of the time.
#43 compresses the three-step flow into a single topa review <PR>. Review comments are fixed at three single-sentence points: problem, reproduction, fix. The review directory for the same PR keeps the worktree and posted.json, and a second review fetches only when the local copy is behind GitHub. The no-change wait for a review round is relaxed from 60 minutes to 10 hours.
Beyond these two, topa opened 10 new issues in the window (#35 to #46). They are all actionable work items with fix(...) titles, focused on roofline pricing. They include computing makespan by CTA wave, pricing scatter by number of written rows, and counting resident CTAs into the grid. These are the next batch of tasks the pipeline has queued for itself.
Source: https://github.com/tile-ai/topa/pull/41.
3.5 Two docs-site updates; no commits in the two adopter repos during the window
tilelang.github.io had two commits in the window, both titled Update docs, at 11-18 17:18 and 17:24 (UTC). TileOPs.github.io merged #88, adding the index family and three new operators to the operator library and removing the two mHC operators, in sync with the interface removal in 3.1.
The two adopter repos were quiet. deepseek-ai/TileKernels last pushed on 10-14, QwenLM/FlashQLA on 10-30. The TileOPs-nightly snapshot branch had a push at 11-17 18:42 (UTC), before the start of this window.
Source: https://github.com/tile-ai/TileOPs.github.io/pull/88.
IV. Community, Tutorials, and Events
4.1 News side is empty: no in-window coverage from Chinese/English search, Hacker News, or arXiv
Six Google News queries covered tilelang, "TileLang" GPU, tile-ai, and three Chinese term sets. The only hit containing tilelang was a Korean-media piece discussing China’s AI ecosystem, whose body has nothing to do with TileLang. The rest were noise from generic terms like tile and AI. A time-sorted Hacker News search returned no in-window entries.
The most recent arXiv paper on the TileLang theme is TileSight from 11-03, with nothing new in the window. There are no news-side entries this issue; the main body of the daily comes from the repository side.
Source: no new coverage.
4.2 Organization activity: 38 merges and 43 updated open PRs in a single day; main repo at 9,001 stars
The org-level merge search returns 38, consistent with the per-repo main-branch check. Open PRs with updates number 43, and 20 new issues were opened. By distribution, TileOPs accounts for 23 merges, MLIR Ascend for 3, topa for 3, and 1 lands on the harness branch.
The main repo tilelang has 9,001 stars, 9 more than the previous issue. Releases are unchanged: tilelang is still at v0.1.15 (released 09-30), and TileOPs is still at v0.1.0 (released 10-27).
Source: https://github.com/tile-ai/tilelang.
V. Trend Observations
- Platform boundaries are expanding: a Metal target and automatic differentiation landed on the same day, extending the compiler from inference kernels toward training and Apple GPUs. Both stop at the IR layer, and the order is clear: interfaces first, backends and runtimes later.
- FP8 is entering delivery: quantization, short-row streaming, paged decode, and WGMMA primitivization all landed the same day, each benchmarked against vLLM, FlashInfer, or DeepSeek’s own implementation. Performance acceptance remains separate from the merge bar; both #3010 and #3005 self-assessments are still below the acceptance line.
- Performance debt is now managed issue by issue: seven new target issues spell out hardware, shape, and target multiple. Two were fixed and closed the same day they were opened; the list is driving delivery, not just recording debt.
- Ascend shows a two-track speed gap: the main repo this issue only wraps up tests and CI, while the MLIR route added three TLIR operators in one day. The former is converging; the latter is still in the early stage of paving one operator at a time.
- Backend distributions are starting to fork: Sunrise already has 0.1.15+sunrise.1.2.1, MetaX is preparing 0.1.12+musa.5, and Hygon is following up with merge-style fixes. With three vendors releasing separately, version alignment and the compatibility matrix will become a new maintenance cost.
Appendix: Sources and Verification Notes
Source Verification Table
| Source | Verification Result |
|---|---|
| Per-repo check of the GitHub tile-ai org | 30 repos, 11 pushed within the window; net additions to main branches: TileOPs 23, tilelang 3, TileFoundry 2, topa 3, Sunrise 1, Hygon 1; MLIR Ascend 3 merged into the mlir-dialect branch, 1 merged into the harness branch, whose main has been untouched since 11-04 |
| Org-level merged-PR ground-truth query | 38 merged: by repo, tilelang 3, TileOPs 23, topa 3, MLIR Ascend 3, harness 1, TileFoundry 2, Sunrise 1, Hygon 1, TileOPs docs site 1 — consistent with the per-repo check |
| Org-level open-PR query | 43 open and updated, concentrated in TileOPs performance and new-operator proposals, MLIR Ascend TLIR operators, and Ascend main-repo fixes |
| New issues | 20: topa 10, TileOPs 7, tilelang 1 (Tile IR RFC), tilelang-ascend 2 (daily test failure alerts) |
| Releases and tags | tilelang still v0.1.15 (09-30); Sunrise adds 0.1.15+sunrise.1.2.1; TileOPs still v0.1.0 (10-27) |
| News side | Six Google News queries (Chinese and English) returned no related articles; no Hacker News entries within the window; the latest related arXiv paper is 11-03, with no new ones in the window |
| Adopter repos | deepseek-ai/TileKernels last pushed 10-14, QwenLM/FlashQLA 10-30 — both outside the window |
| TileLang×RISC-V four-source sweep | No change: the TVM-route kubuds- prefixed branch, the MLIR-route main, the community index page, and SOPHGO TPU all match the previous issue |
Full Source List
- [1] TileFoundry #354 Metal target — https://github.com/tile-ai/TileFoundry/pull/354
- [2] TileFoundry #355 automatic differentiation — https://github.com/tile-ai/TileFoundry/pull/355
- [3] TileOPs #3026 FP8 per-token grouped quantization UE8M0 — https://github.com/tile-ai/TileOPs/pull/3026
- [4] TileOPs #2997 FP8 paged GQA decode rewrite — https://github.com/tile-ai/TileOPs/pull/2997
- [5] TileOPs #3010 FP8 short-row weight streaming — https://github.com/tile-ai/TileOPs/pull/3010
- [6] TileOPs #3030 FP8 WGMMA primitivization — https://github.com/tile-ai/TileOPs/pull/3030
- [7] TileOPs #3029 shared-memory scale buffer accounting — https://github.com/tile-ai/TileOPs/pull/3029
- [8] TileOPs #3023 TopKSelect register kernel — https://github.com/tile-ai/TileOPs/pull/3023
- [9] TileOPs #3022 MLA cache write vectorization — https://github.com/tile-ai/TileOPs/pull/3022
- [10] TileOPs #3024 DeepSeek-V4 expert post-processing — https://github.com/tile-ai/TileOPs/pull/3024
- [11] TileOPs #3001 MXFP4 expert weights for MoE grouped GEMM — https://github.com/tile-ai/TileOPs/pull/3001
- [12] TileOPs #3004 NSA variable-length attention with 256 head dim — https://github.com/tile-ai/TileOPs/pull/3004
- [13] TileOPs #2994 EOS-truncated n-gram hashing — https://github.com/tile-ai/TileOPs/pull/2994
- [14] TileOPs #3021 remove the two operators before and after mHC — https://github.com/tile-ai/TileOPs/pull/3021
- [15] TileOPs #3006 speculative decoding state rollback — https://github.com/tile-ai/TileOPs/pull/3006
- [16] TileOPs #3002 KDA state slots and value-major state — https://github.com/tile-ai/TileOPs/pull/3002
- [17] TileOPs #3003 window overflow fix — https://github.com/tile-ai/TileOPs/pull/3003
- [18] TileOPs #3025 cleanup of merges since 11-16 — https://github.com/tile-ai/TileOPs/pull/3025
- [19] tilelang #4389 CPU any_of / all_of serial fallback — https://github.com/tile-ai/tilelang/pull/4389
- [20] tilelang #4407 cap Z3 resources in topk regression test — https://github.com/tile-ai/tilelang/pull/4407
- [21] tilelang #4401 remove cross-NPU probing from Ascend autotuning tests — https://github.com/tile-ai/tilelang/pull/4401
- [22] tilelang #4403 CUDA Tile IR backend RFC discussion — https://github.com/tile-ai/tilelang/issues/4403
- [23] tilelang #4402 CUDA Tile IR backend RFC document — https://github.com/tile-ai/tilelang/pull/4402
- [24] tilelang #3804 Metal static local fill and copy — https://github.com/tile-ai/tilelang/pull/3804
- [25] MLIR Ascend #211 tl.shl — https://github.com/tile-ai/tilelang-mlir-ascend/pull/211
- [26] MLIR Ascend #212 tl.shr — https://github.com/tile-ai/tilelang-mlir-ascend/pull/212
- [27] MLIR Ascend #214 tl.cast — https://github.com/tile-ai/tilelang-mlir-ascend/pull/214
- [28] MLIR Ascend #260 roofline skill — https://github.com/tile-ai/tilelang-mlir-ascend/pull/260
- [29] MLIR Ascend #263 RoPE kernel — https://github.com/tile-ai/tilelang-mlir-ascend/pull/263
- [30] Hygon #27 intrinsic, precision, and integer vector fixes — https://github.com/tile-ai/tilelang-hygon/pull/27
- [31] Sunrise #11 release 0.1.15+sunrise.1.2.1 — https://github.com/tile-ai/tilelang-sunrise/pull/11
- [32] MetaX #13 release v0.1.12+musa.5 source — https://github.com/tile-ai/tilelang-musa/pull/13
- [33] topa #41 per-round review and GPU-exclusive timing — https://github.com/tile-ai/topa/pull/41
- [34] topa #43 complete a review round with one command — https://github.com/tile-ai/topa/pull/43
- [35] TileOPs #3028 TileKernels comparison-timing proposal — https://github.com/tile-ai/TileOPs/issues/3028
- [36] TileOPs #3037 DSADense gap to tensor-core peak — https://github.com/tile-ai/TileOPs/issues/3037
- [37] TileOPs #3018 GQA variable-length head dim 72 gap — https://github.com/tile-ai/TileOPs/issues/3018
- [38] TileOPs #3034 GatedPoolCompress decode rows — https://github.com/tile-ai/TileOPs/issues/3034
- [39] TileOPs #3031 HyperConnection pre gap — https://github.com/tile-ai/TileOPs/issues/3031
- [40] TileOPs docs site #88 index families and three new operators — https://github.com/tile-ai/TileOPs.github.io/pull/88
Verification Notes
Scope: the window is 11-18 07:00 to 11-19 07:00 Beijing time. Repository commits are counted by committer time and judged as new only after comparison with the previous issue’s cutoff. MLIR Ascend merges land on the mlir-dialect branch; main has been untouched since 11-04, and the body states the branch explicitly rather than calling it the main branch.
Collection hit unauthenticated GitHub API rate limits late in the window. A few branch details were re-checked via Atom feeds and git ls-remote; conclusions were unaffected. The RISC-V four-source sweep returned empty output, treated as no change rather than a collection failure.