TileLang Daily Intelligence Report (2026-11-17)
Research window: 2026-11-16 07:00 ~ 2026-11-17 07:00 (Beijing time, 24 hours; previous issue cutoff was 11-16 07:00).
Sources: GitHub tile-ai organization per-repo check (30 repos, 10 pushed within the window) + organization-level merged and in-review ground-truth search (30 merged, 40 newly opened PRs, 19 newly opened issues within the window) + per-commit main-branch check (TileOPs 18, tilelang 4, topa 4, TileFoundry 2, site docs 2) + release page check (tilelang still at v0.1.15, no new tags within the window) + Google News Chinese and English queries + Hacker News + TileLang×RISC-V four-source targeted inspection.
This Issue’s Index
- Today’s Focus: Public disclosure of upstream Ascend repo CI abuse and contributor credential theft
-
Today’s Focus: TileOPs merges 18 PRs in a single day; operator library shifts toward pilot-model interface alignment
- I. Core Project Progress
- 1.1 Four bugfixes merged to main repo: JIT device restoration, NVRTC single-parameter, cython dynamic shared memory, ROCm HIP runtime reuse (11-16)
- 1.2 Manifest upgraded to declarative parameter types; pilot-model interfaces completed in one pass (11-16)
- 1.3 Three new operators: Gated RMSNorm and MaskedScatter land; Embedding merged then reverted same day (11-16)
- 1.4 Three GEMM quantization tracks: INT8 W8A8 switches to 192-wide ping-pong; W4A8/W4A16 stream-K each wrap up (11-16)
- 1.5 Attention and linear attention: paged GQA calibration follow-up, DeltaNet launch config, GLA backward shared memory (11-16)
- 1.6 Tests and benchmarks: GDN scan trimmed to one line per scheduling region; fp8 cases regress to their respective targets (11-16)
- II. Multi-Backend Adaptation (Ascend / MetaX / Hygon / Moore Threads / Sunrise)
- 2.1 Ascend: One msprof cache-refresh fix in main-repo Ascend backend under review (11-16)
- 2.2 MLIR Ascend: Roofline skill enters Harness; tl.vrelu enters TLIR and lowers to HIVM (11-16)
- 2.3 Hygon and Sunrise: No Hygon main-branch commits in window; Sunrise adds a 0.1.15+sunrise.1.2.1 release PR (11-16)
- II·V. RISC-V Backend and Community (Special Watch: TVM Track / MLIR Track / SOPHGO TPU / Community Index Page)
- 2.5.1 MLIR track: RuyiAI-Stack/tilelang-riscv switches default to buddy-opt; adds GEMM and BOSC examples (11-16)
- 2.5.2 TVM track, Sophgo TPU, and community index page unchanged this issue (11-17)
- III. Ecosystem and Adopters (TileFoundry / topa / Site / TileKernels / FlashQLA)
- 3.1 TileFoundry validator refactored into a single ctx traverser; shared-memory write-conflict analysis added in sync (11-16)
- 3.2 TileFoundry opens 11 new issues in a single day, concentrated on cost model and scheduler gap-filling (11-16)
- 3.3 topa introduces new-op rounds and state machine; issues auto-routed to TileFoundry by title type (11-16)
- 3.4 Site docs follow manifest changes; no commits in either adopter repo during the window (11-16)
- IV. Community, Tutorials, and Events
- 4.1 News side quiet: No topical coverage in Google News (Chinese/English) or Hacker News (11-17)
- 4.2 Organization activity: 40 new PRs and 19 new issues in a single day; 39 under review (11-16)
-
V. Trend Observations
- Appendix: Sources and Verification Notes
- Source Verification Table
- Full Source List
Today’s Focus: Public Disclosure of Upstream Ascend Repo CI Abuse and Contributor Credential Theft
At 14:16 on 11-16 (Beijing time), contributor liggest filed issue #2016 in tile-ai/tilelang-ascend, publicly stating that their account had been impersonated on 11-15 and laying out a complete attack chain: starting at 7 a.m. on 11-15, an account named miramar-labs (itself suspected of being compromised) submitted PR #2014, which added a malicious script that collects server information when CI runs; the repo’s CI runs on self-hosted servers, and after being compromised the attacker exfiltrated a large amount of information and login credentials from the server, including a GitHub token belonging to liggest. Between 9:39 and 11:40 a.m. on 11-15, the attacker used that token to impersonate liggest: posting 18 consecutive comments under #2014, apparently using the public PR as a pickup window; filing two PRs with mining scripts in tile-ai/TileOPs (#2902, #2903, quickly closed by maintainers); and filing a similar PR #5086 in SemiAnalysisAI/InferenceX, attempting to drive the other party’s benchmark server to execute, which was discovered and closed by the author personally in the early hours of 11-16. The author states that the two mining scripts were not executed due to GitHub’s automated checks, and no execution records were found.
On the response side: the account itself had two-factor authentication and was not directly taken over; the author revoked old authorizations for GitHub CLI and VSCode, stating the token is now invalid and the account has shown no anomalies since 11-15; the malicious fork is temporarily retained for forensics; the compromised CI server awaits security handling; and #2014 also needs to be dealt with, otherwise there remains a risk of re-triggering the malicious script. The author’s recommendation is: change PRs from new contributors to require maintainer approval before triggering CI workflows that use self-hosted servers.
On the verification side: a GitHub search shows that miramar-labs submitted same-style PRs to four repositories between 23:18Z on 11-14 and 02:04Z on 11-15—tile-ai/tilelang #4148, tile-ai/tilelang-ascend #2014, tile-ai/tilelang-mlir-ascend #251, and tile-ai/TileOPs #2900—all nominally about “cache warm-up” or “CI capability probing”; all four PRs have been closed, none merged, and #2014 was closed at 11:29 on 11-16. After the issue was closed, the author added another comment stating that the earlier description “also seems to contain some exaggeration; in any case, the account has been restored,” so the above scope of damage remains based on the author’s own account and has not been independently verified by a third party. There are three direct implications for this community: the trust boundary of self-hosted CI has been penetrated by external PRs, the lifecycle of personal tokens must be brought into audits, and same-named PRs will appear simultaneously across multiple repositories.
Sources: https://github.com/tile-ai/tilelang-ascend/issues/2016 , https://github.com/tile-ai/tilelang-ascend/pull/2014 , https://github.com/tile-ai/TileOPs/pull/2902
Today’s Focus: TileOPs Merges 18 PRs in a Single Day, Operator Library Shifts Toward Pilot Model Interface Alignment
During the window, 18 commits landed on the TileOPs main branch (corresponding to 18 PRs), the highest single-day density this quarter, and the direction was highly consistent: first complete the manifest’s interface description capabilities (#2958 introduces call-time type and default, list[Tensor] and scalar-or-tensor union types; #2967 registers the seven operators that the pilot model will actually call as spec-only, and adds the parameters and inputs used by the pilot model to fifteen implemented operators), then fill in operators one after another with [Feat][foundry] PRs. Two new operators were merged that day (gated RMSNorm, MaskedScatter), while Embedding was merged and then reverted the same day, indicating that new operators must go through TileFoundry rounds rather than being merged directly. #2958 also fixed two semantic errors along the way: the RoPE family previously rotated only positions 0..S-1, making the rotation result for decode position p incorrect; and FP8QuantPerTokenGroupFwdOp was inconsistent with its reference implementation.
Sources: https://github.com/tile-ai/TileOPs/pull/2958 , https://github.com/tile-ai/TileOPs/pull/2967
I. Core Project Progress
1.1 Four bugfixes merged into the main repo: JIT device restoration, NVRTC single-argument, cython dynamic shared memory, ROCm HIP runtime reuse
Four PRs were merged into the main repo during the window, all runtime defects rather than new features. #3872 fixes the tvm_ffi host stub only setting the current device without restoring it: calling a cuda:1 tensor while on cuda:0 produces correct results but leaves the current device on cuda:1, so subsequent allocations with device="cuda" silently drift; the fix wraps the call in torch.cuda.device(...) at the adapter layer and restores on exit. #3880 fixes single-argument kernels failing to launch on the NVRTC backend: after joining parameter names and types with ", ", a single argument does not produce a tuple, and cuLaunchKernelEx throws TypeError outright; changed to a tuple literal with a trailing comma. #3882 fixes the cython backend’s dynamic shared memory limit being set only once in init() — the attribute is per-device, so kernels needing more than 48 KB fail with cudaErrorInvalidValue on other devices; changed to generating a _set_dynamic_smem() helper per kernel and recording devices already raised. #3899 fixes the HIP stub failing to find PyTorch’s bundled libamdhip64 under ROCm: PyTorch’s ROCm wheel loads it with RTLD_LOCAL, so the stub first probes already-mapped runtimes with RTLD_NOLOAD, avoiding a second runtime from dlopen that leaves module handles crossing runtimes and the first kernel launch failing.
Source: https://github.com/tile-ai/tilelang/pull/3872 , https://github.com/tile-ai/tilelang/pull/3882 , https://github.com/tile-ai/tilelang/pull/3899
1.2 Manifest upgraded to declarative parameter types; pilot model interfaces completed in one pass
#2958 extends the manifest from “only describing fixed tensors and scalars” to declaring three signature classes: tensor sequences (torch.cat), scalar-or-tensor alternatives (flash_attn’s seqlen_offsets, torch.clamp(min)), and required-but-may-be-None (mamba_ssm’s bias); call-site inputs are unified into declarations carrying type and default; the three Scalar/Tensor operator pairs previously split out for “scalar or tensor” are merged accordingly. #2967 builds on this to fill in the interface ground truth for pilot models: seven newly registered spec-only entries (including torch reference implementations), fifteen implemented entries gain the parameters and inputs actually passed by pilot models, kernels reject non-default values with OpNotAvailableError, entry status remains implemented, and benchmark rows encountering that error are recorded as “unavailable” rather than failed.
Source: https://github.com/tile-ai/TileOPs/pull/2958 , https://github.com/tile-ai/TileOPs/pull/2967
1.3 Three new operators: gated RMSNorm and MaskedScatter land, Embedding reverted same day
#2984 adds RMSNormGatedFwdOp: RMSNorm grouped along the last axis, gated by silu(z) or sigmoid(z) before or after normalization, computed entirely in float32 with a single rounding; one TileLang kernel covers all call shapes (8 lanes, each with two 16-byte accesses forming a 128-wide row, 16 rows forming a 128-thread CTA; wider groups double lanes to 1024 and sum across warps via shared memory), register cap set at 64 per thread, with the gating exponential and reciprocal each using one instruction. Benchmarks align with vLLM’s RMSNormGated, mamba_ssm’s rmsnorm_fn, fla’s rms_norm_gated, Hugging Face’s Qwen3_5RMSNormGated, plus torch reference and torch.compile, meeting the acceptance bar. #2982 adds MaskedScatterFwdOp (out-of-place masked_scatter, supporting fp16/bf16/fp32 with input and mask broadcasting): each CTA takes a complete contiguous mask segment (row count bounded at 32 KB), first counts true segments before its own segment then scans itself, reading each 16-byte vector only once (true segments read source, false segments read input), and performs a chunked count first when the segment exceeds four blocks; meets the bar against torch eager and torch.compile, while vLLM’s in-place multimodal merge is listed as incomparable due to differing semantics.
There was also a same-day reversal: after #2977 merged EmbeddingFwdOp, #2983 reverted it entirely the same day on the grounds that the PR was merged by mistake and new operators should go through a separate TileFoundry round; after the revert the entry returns to spec-only, with kernel, operator, benchmarks, tests, and workloads all removed. #2975 remains under review as the formal proposal for the same operator (gathering weight rows by token index).
Source: https://github.com/tile-ai/TileOPs/pull/2984 , https://github.com/tile-ai/TileOPs/pull/2982 , https://github.com/tile-ai/TileOPs/pull/2983
1.4 Three quantized GEMM lines: INT8 W8A8 moves to 192-wide ping-pong, W4A8/W4A16 stream-K each wrapped up
#2959 switches INT8 ping-pong GEMM to 128×192 tiles when “the busiest CTA’s column count does not exceed the 128×128 tile”, with the grid tail collapsed into 32-column-aligned storage boxes per the bf16 GEMM narrow-tile rule (160 wide on every prefill row in this period’s manifest), issuing one WGMMA of the corresponding width per width; at equal busiest-CTA workload, 192-wide tiles move fewer TMA and WGMMA shared-memory bytes per MAC and cut CTA consumer switches by a third, reducing the busiest CTA’s column count in llama3-8b prefill-qkv’s final wave from 1536 to 1504. #2960 changes W4A8 prefill’s stream-K tail: when all other partial tiles have landed, the late-arriving contributor 0 no longer writes its partial sum, and instead the contributor that completes the tile’s count finishes it, replacing each thread’s fence with a single acquire-release atomic per CTA (placed after the named barrier), with no CTA waiting on another; only programs containing stream-K tiles take the new path, and all other compilation results match main. #2964 fixes a misalignment in W4A16 stream-K: the kernel originally guarded the “pad unused slots and do ceil division” logic with cta == first_cta, but TileLang’s simplification dropped it entirely after the CTA’s second segment and folded it after the first, leaving unwritten slots in tiles shared by two CTAs and causing reductions to read dirty memory, with max error ranging from 2.4 to infinity; changed to building the table once on the host on the device that constructs the program, locating slots by “which CTA within the tile”.
Quantized GEMM coverage is filled in alongside: #2969 adds 42 W4A16 rows, 13 W4A8 rows, and 11 FP8 rows, each corresponding to a real projection (TP1 or TP8, decode batch) in Llama-3 8B/70B, Qwen3 8B/32B, or DeepSeek-V3, where W4A16 benchmark rows previously covered only 3 of 45 decode scheduling regions. Another PR, #2990 (tightening W4A16 decode tile pairing conditions), is under review, prompted by issue #2966 recording that the same batch of changes was slower than before on some 8-row and 16-row shapes.
Source: https://github.com/tile-ai/TileOPs/pull/2959 , https://github.com/tile-ai/TileOPs/pull/2964 , https://github.com/tile-ai/TileOPs/pull/2969
1.5 Attention and linear attention: paged GQA calibration follow-up, DeltaNet launch config, GLA backward shared memory
#2956 continues paged GQA calibration: the split reduction now stores the weight table with split as the primary order, one CTA per row block, and reads the request mapping generated by the scan (previously writing a [block_M, num_split] table in parallel by row, serializing at 16- and 32-way splits); GQAPagedFwdKernel.default_config uses a 128-row query tile only when packed rows actually fill it (48-row pages make the key tile narrower than one WGMMA, and decode packing gets padded to 128 rows instead); in the left-window case, RoPE preprocessing is also prepared over the in-window range. #2978 lets the WS kernel take over the case where “the window covers the entire page table and hides no keys”: previously the WS kernel rejected all windows without RoPE, so window4096/long-b1-q2048-kv2048 fell to the cp.async kernel at 2.17× the time of FA3; the PR self-assesses as still below bar — an improvement rather than a pass.
On the linear attention side, #2979 fixes DeltaNetChunkFwdOp’s launch configuration at chunk_size=16 (at most two warps for prepare and output, only one for odd V/8, and the recursive column’s warp count must divide the 8-column tile count), and #2980 changes the chunk launch cap to be derived from each GEMM’s warp tiles, fixing a batch of shapes at chunk 32/64/128 that hit “TileLang warp partition failed” (previously special-cased only for chunk 16), and making non-conforming chunk_size/dim_k/dim_v rejected at build time rather than failing randomly. #2965 brings GLA backward’s shared memory lifetime back within the 99 KiB budget on SM89 (recomputing gating factors, reloading state gradients), restoring four previously over-limit dtype/shape combinations (FP32 32×64×192, 32×32×256 and FP16/BF16 32×128×128), with three groups still rejected. #2976 marks two DeltaNet timing cases as SM90 to avoid first building an SM90-only decode warp kernel on SM89.
Source: https://github.com/tile-ai/TileOPs/pull/2956 , https://github.com/tile-ai/TileOPs/pull/2978 , https://github.com/tile-ai/TileOPs/pull/2980
1.6 Tests and benchmarks: GDN scan trimmed to one row per scheduling region, fp8 cases returned to their respective targets
#2981 trims GDNFwdOp’s workload from 151 rows / 283 cases to “one row per scheduling region”: the rule keeps the first row of each bf16 group (kernel identity, shape range, H200/SM89 scheduling region), keeps fp16 only on the first row it appears in each scheduling region, and keeps one row each for q35-gva-decode-b1 and q35-gva-decode-b8. #2991 fixes tileops_test’s device detection: _Reference._device() previously read .device from scalars, and after #2958 kernels receive call-site scalars positionally, causing 61 failures in the nightly off-cuda-suite on a0b7aeb6; it also marks three fp8 cases back to in_tree_kernels, since they check in-tree kernels’ rejection behavior while tileops_test targets always go through reference implementations.
Source: https://github.com/tile-ai/TileOPs/pull/2981 , https://github.com/tile-ai/TileOPs/pull/2991
II. Multi-backend Adaptation (Ascend / MetaX / Hygon / Moore Threads / Sunrise)
2.1 Ascend: one msprof cache-flush fix under review in the main repo’s Ascend backend
The main repo opens #4385, fixing cache-flush annotation handling for the Ascend backend’s msprof; the same window also has three documentation PRs under review — #4386, #4387, #4388 (aligning ROCm source-build prerequisites, documenting HIP stub defaults on Windows, and correcting the manual GQA throughput ratio). The backend repo tilelang-ascend has no new commits on main during the window (latest remains 11-13), but the repo logged daily test failures on 11-16 and carries this period’s most closely watched security disclosure (see Today’s Focus 1).
Source: https://github.com/tile-ai/tilelang/pull/4385
2.2 MLIR Ascend: Roofline skill enters Harness, tl.vrelu enters TLIR and lowers to HIVM
tilelang-mlir-ascend has no commits on main during the window, but opens two PRs: #260 adds a Roofline skill to the Harness, and #261 adds tl.vrelu to TLIR with the TL-to-HIVM lowering and tests. Both lines point at the “operator expression and performance modeling on the MLIR side” layer, parallel to the main repo’s manifest operator-description work.
Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/260 , https://github.com/tile-ai/tilelang-mlir-ascend/pull/261
2.3 Hygon and Sunrise: no in-window commits on Hygon main, Sunrise adds a 0.1.15+sunrise.1.2.1 release PR
tilelang-hygon’s pushed_at falls within the window, but the latest commit on main is still 11-13; in-window pushes come from non-default branch and tag activity, with no new code entering main; tilelang-metax (11-06) and tilelang-musa (11-10) saw no activity during the window. tilelang-sunrise submitted two release PRs on 11-16 (#10, #11) targeting version 0.1.15+sunrise.1.2.1, corresponding to candidate branch ci/sunrise-0.1.15-1.2.1-r6; the latest commit on main remains at 10-23, and neither PR is merged.
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 MLIR route: RuyiAI-Stack/tilelang-riscv switches to buddy-opt by default, adds GEMM and BOSC examples
RuyiAI-Stack/tilelang-riscv (the MLIR-rewritten RISC-V backend, target="riscv") advances main from 57c548cc to 2204c7ec, merging PR #6 (22 files, +835/−284, author laz174). Four substantive changes: buddy-opt enabled by default, with Buddy and LLVM toolchain roots configured separately while retaining an explicit TILELANG_RISCV_OPT override; the RISC-V compile cache key incorporates the optimizer choice (switching optimizers no longer hits the same cache); new host-side FP16 GEMM example and standalone BOSC INT8 GEMM firmware example (examples/riscv/gemm/, examples/riscv/bosc/, including launch.c and build scripts); build, target, and test docs switched to portable paths with usage notes added.
2.5.2 TVM route, SOPHGO TPU, and community index page unchanged this period
On the TVM route, the kubuds-riscv-mahiru, kubuds-riscv-karen, and kubuds-riscv-hikari branches of kubuds/tilelang-riscv-kubuds show no reference changes during the window; on the SOPHGO TPU side, TongmingLAIC/tilelang-tpu’s main and branches such as feature/sg2260e-rv-support have no new commits; the community index page TongmingLAIC/tilelang-riscv has no new PRs and an unchanged project table. Only the MLIR side saw substantive progress this period.
Source: https://github.com/RuyiAI-Stack/tilelang-riscv/pull/6
III. Ecosystem and Adopters (TileFoundry / topa / Site / TileKernels / FlashQLA)
3.1 TileFoundry validator refactored into a single ctx walker, with shared-memory write-conflict analysis added alongside
#336 consolidates the TIR validator’s previously separate walkers (walk_stmt, Binders, iter_all_stmts, the standalone interfaces/verify visitor, and the mesh_scope stack) into StmtExprVisitor: it can carry an optional ctx, visits expressions within a statement in boundary order before visiting sub-statements, and analysis no longer indexes Call mappings by structural hash. The PR also flags a warning: the FSDP training example has been unable to build since #325 and #293, and no test caught it. #326 adds shared-memory write-conflict analysis: cuda/analysis/threads.py pairs each shared-memory write with itself across threads, and only allows it if some layout can uniquely assign each byte to the originating thread, or if the relation proves two threads will not touch the same byte; previously “every thread issues the same action against an entire shared-memory tile” was accepted, and that is exactly the kind of write the lowering stage generates for fills and copies.
Source: https://github.com/tile-ai/TileFoundry/pull/336 , https://github.com/tile-ai/TileFoundry/pull/326
3.2 TileFoundry opens 11 new issues in a single day, concentrated on cost-model and scheduler gaps
#337 through #347 comprise 11 new issues, with themes heavily concentrated on “price/evaluate” — split inputs of multi-wave grids priced only once, row-priced index copies and gathers, scheduling and checking of masked row gather, DimVar slice start points, per-thread copy_async, split registers for f16 register MMA and scalar scaling, and flash-attention epilogue for zero-padded boundary tiles, among others. #344 explains why these gaps must be closed first: rotary position embedding with a caller-supplied cos/sin table needs to read table rows directly into registers per token, and without an indexed gmem-to-register copy, 5 of 14 workloads cannot be scheduled. These issues are upstream prerequisites for whether new TileOPs operators can land on schedule.
Source: https://github.com/tile-ai/TileFoundry/issues/344
3.3 topa introduces new-op rounds and a state machine, with issues auto-routed to TileFoundry by title type
Four PRs merged in the topa window. #24 gives spec-only manifest entries a formal new-op round entry point (topa round start <Op> [--issue N]) and turns round progress into an explicit state machine, rejecting at admission issues already assigned to someone else, PRs where someone else has already nominated that operator or closed the issue, and active rounds on the same operator or issue; #25 routes by TileOPs title type: [Perf] goes to perf rounds, [Feat] nominating a single spec-only operator goes to new-op, [Feat] nominating only already-implemented operators goes to the newly added extend type, and the rest go to general, and it specifies that each round times only other operator rows whose generated program differs from the main branch, marking identical programs as unchanged with a program_diff record attached; #23 fixes the round container being unable to read the checklist inside the installed package (by mounting the topa package read-only at the host path); #22 fixes idle GPUs being permanently excluded in exclusive mode — a GPU counts as unavailable only when nvidia-smi lists a compute process. One more, #26, is under review, allowing a STOPPED round to be manually restored to a state.
Source: https://github.com/tile-ai/topa/pull/24 , https://github.com/tile-ai/topa/pull/25 , https://github.com/tile-ai/topa/pull/22
3.4 Site docs follow manifest changes; no commits in either adopter repo during the window
TileOPs.github.io merged two PRs: #86 follows #2958 in describing call-time type/default, member(), list[Tensor], and random_ints, and switches the API page to the three operators ClampFwdOp, LerpFwdOp, and MaskedFillFwdOp (no longer Scalar/Tensor pairs), with the RoPE page adopting vLLM’s (positions, query, key) signature; #87 adds that validate_inputs now also validates every in_range boundary, that int members accept float, and a review note on breaking changes in 0.x minor versions. On the adopter side: neither deepseek-ai/TileKernels nor QwenLM/FlashQLA had commits during the window (latest commits on 10-14 and 10-30, respectively).
Source: https://github.com/tile-ai/TileOPs.github.io/pull/86 , https://github.com/tile-ai/TileOPs.github.io/pull/87
IV. Community, Tutorials, and Events
4.1 News-side gap: no topical coverage in Google News (Chinese or English) or Hacker News
During the window, neither the Chinese nor English Google News queries returned TileLang-related coverage (the English endpoint returned a rate-limit page this time, and the Chinese endpoint returned an empty result set); Hacker News surfaced two off-topic projects (the Tyrus language and the Daventry.io game engine), which were filtered out; the arXiv query returned nothing this time and was not included as material. The news-side gap does not affect this issue’s content, which is still composed mainly of the GitHub side.
4.2 Organization activity: 40 new PRs and 19 new issues in a single day, 39 PRs under review
Organization-level search shows 40 PRs opened during the window (TileOPs 25, topa 5, tilelang 4, tilelang-mlir-ascend 2, tilelang-sunrise 2, TileFoundry 1, site docs 1), 19 issues opened (TileFoundry 11, TileOPs 6, tilelang-ascend 2), and 39 PRs still under review at the end of the window. New operator proposals cluster around indexing (Concat and Embedding under review, MaskedScatter merged), RoPE (RoPECosSin), and normalization (RMSNorm with zero-centered weights, FusedAddRMSNorm), corresponding to the pilot model operator list TileOPs is currently rolling out.
Source: https://github.com/tile-ai/TileOPs/pulls , https://github.com/tile-ai/TileFoundry/issues
V. Trend Observations
- The operator library has become the main battleground: of the 30 merges in this issue, TileOPs accounts for 18, while the main repo has only 4 runtime bugfixes; the community’s focus has shifted from the compiler itself to operator implementation and interface ground truth.
- Interfaces precede implementation: manifest parameter types and pilot model interfaces were completed on the same day, after which new operators land one round at a time; the mistaken merge of Embedding and its same-day revert show the process is tightening, not loosening.
- Process is being tooled: topa’s new-op rounds, state machine, and issue routing, together with TileFoundry’s cost-model gap-closing, form a closed loop of “operator demand → scheduling feasibility → implementation and acceptance,” and this layer has been the fastest-growing part over the past two weeks.
- Performance fixes are starting to backfire: W4A16 paired tiles regress on some small shapes (issue #2966), and quantized GEMM optimization has entered a stage where “shapes must be tiered”; a single tile strategy can hardly yield broad gains anymore.
- Security boundaries need shoring up: incidents where the self-hosted CI was penetrated by an external PR and a personal token was stolen suggest that CI trigger permissions for new-contributor PRs should require maintainer approval by default — a point that applies equally to the entire RISC-V and AI compiler community.
- The RISC-V side is steady: this issue has only one merge on the MLIR route (buddy-opt made default and examples completed), with no activity on the TVM route or the Sophgo TPU side, which is normal alternating progress.
Appendix: Sources and Verification Notes
Source Verification Table
| Source | Verification Result |
|---|---|
| tile-ai org repo-by-repo check (30 repos) | 10 repos pushed within the window; all main branches checked commit by commit |
| Org-level merge ground truth (search issues) | 30 merges within the window, consistent with the 30 checked commit by commit against main branches |
| tilelang main repo | 4 merges, all bugfixes; version still v0.1.15, no new tag |
| TileOPs | 18 merges, 6 new issues opened, multiple PRs under review |
| TileFoundry | 2 merges, 11 new issues opened |
| topa | 4 merges, 1 under review |
| tilelang-ascend | No commits on main branch; security disclosure issue #2016 closed |
| tilelang-mlir-ascend | No commits on main branch, 2 new PRs opened |
| tilelang-hygon / metax / musa | No commits on main branches within the window |
| tilelang-sunrise | 2 release PRs under review, not merged |
| Adopters TileKernels / FlashQLA | No commits within the window |
| TileLang×RISC-V four-source sweep | Changes found: RuyiAI-Stack/tilelang-riscv merged 1; the other three sources unchanged |
| Google News (Chinese and English) | No topical coverage (English endpoint returned a rate-limit page) |
| Hacker News | Both hits irrelevant, excluded |
| arXiv | Query returned nothing, not included as material |
Full Source List
- [1] tilelang-ascend issue #2016 (CI abuse and credential theft disclosure) — https://github.com/tile-ai/tilelang-ascend/issues/2016
- [2] tilelang-ascend PR #2014 (malicious script, closed without merge) — https://github.com/tile-ai/tilelang-ascend/pull/2014
- [3] TileOPs PR #2902 (mining script, closed) — https://github.com/tile-ai/TileOPs/pull/2902
- [4] tilelang PR #3872 (tvm_ffi device restore) — https://github.com/tile-ai/tilelang/pull/3872
- [5] tilelang PR #3880 (NVRTC single argument) — https://github.com/tile-ai/tilelang/pull/3880
- [6] tilelang PR #3882 (cython dynamic shared memory) — https://github.com/tile-ai/tilelang/pull/3882
- [7] tilelang PR #3899 (ROCm HIP runtime reuse) — https://github.com/tile-ai/tilelang/pull/3899
- [8] TileOPs PR #2958 (manifest parameter types) — https://github.com/tile-ai/TileOPs/pull/2958
- [9] TileOPs PR #2967 (pilot model interface) — https://github.com/tile-ai/TileOPs/pull/2967
- [10] TileOPs PR #2984 (RMSNormGatedFwdOp) — https://github.com/tile-ai/TileOPs/pull/2984
- [11] TileOPs PR #2982 (MaskedScatterFwdOp) — https://github.com/tile-ai/TileOPs/pull/2982
- [12] TileOPs PR #2977 / #2983 (Embedding merge and revert) — https://github.com/tile-ai/TileOPs/pull/2983
- [13] TileOPs PR #2959 (INT8 W8A8 192-wide tile) — https://github.com/tile-ai/TileOPs/pull/2959
- [14] TileOPs PR #2960 (W4A8 stream-K finalization) — https://github.com/tile-ai/TileOPs/pull/2960
- [15] TileOPs PR #2964 (W4A16 slot zero-fill) — https://github.com/tile-ai/TileOPs/pull/2964
- [16] TileOPs PR #2969 (quantized GEMM benchmark rows) — https://github.com/tile-ai/TileOPs/pull/2969
- [17] TileOPs PR #2956 / #2978 (paged GQA calibration and WS window) — https://github.com/tile-ai/TileOPs/pull/2978
- [18] TileOPs PR #2979 / #2980 (DeltaNet launch configuration) — https://github.com/tile-ai/TileOPs/pull/2980
- [19] TileOPs PR #2965 (GLA backward shared memory) — https://github.com/tile-ai/TileOPs/pull/2965
- [20] TileOPs PR #2981 (GDN workload trimming) — https://github.com/tile-ai/TileOPs/pull/2981
- [21] TileOPs PR #2991 (test device detection and fp8 cases) — https://github.com/tile-ai/TileOPs/pull/2991
- [22] tilelang PR #4385 (Ascend msprof cache refresh) — https://github.com/tile-ai/tilelang/pull/4385
- [23] tilelang-mlir-ascend PR #260 / #261 (Roofline skill, tl.vrelu) — https://github.com/tile-ai/tilelang-mlir-ascend/pull/261
- [24] tilelang-sunrise PR #11 (0.1.15+sunrise.1.2.1 release) — https://github.com/tile-ai/tilelang-sunrise/pull/11
- [25] RuyiAI-Stack/tilelang-riscv PR #6 (buddy-opt default and examples) — https://github.com/RuyiAI-Stack/tilelang-riscv/pull/6
- [26] TileFoundry PR #336 (validator refactor) — https://github.com/tile-ai/TileFoundry/pull/336
- [27] TileFoundry PR #326 (shared memory write conflict analysis) — https://github.com/tile-ai/TileFoundry/pull/326
- [28] TileFoundry issue #344 (indexed gmem-to-register copy) — https://github.com/tile-ai/TileFoundry/issues/344
- [29] topa PR #24 / #25 (new-op rounds and routing) — https://github.com/tile-ai/topa/pull/25
- [30] topa PR #22 / #23 (GPU selection and container mounts) — https://github.com/tile-ai/topa/pull/22
- [31] TileOPs.github.io PR #86 / #87 (manifest documentation follow-up) — https://github.com/tile-ai/TileOPs.github.io/pull/87
- [32] TileOPs issue #2966 (W4A16 small-shape fallback) — https://github.com/tile-ai/TileOPs/issues/2966
- [33] GitHub tile-ai org repository list and push times — https://github.com/orgs/tile-ai/repositories
- [34] TileLang×RISC-V four-source sweep archive (this issue) — https://github.com/RuyiAI-Stack/tilelang-riscv