Issues

• [Perf][Attention] Two FP8 paged GQA rows are behind FlashAttention-3

  perf
  tile-ai/TileOPs#2649 · Filter by author lcy-seso opened on Oct 19, 2026

• [Perf][GEMM] Investigate W4A8 down prefill epilogue overlap

  perf
  tile-ai/TileOPs#2621 · Filter by author michaelwithu opened on Oct 18, 2026

• [Perf][GEMM] Investigate long-K W4A8 latency step at M=33

  perf
  tile-ai/TileOPs#2620 · Filter by author michaelwithu opened on Oct 18, 2026

• [TEST][Backend] guard off-cuda test coverage against over-marking cuda_only

  follow-upmulti-backendtest
  tile-ai/TileOPs#2611 · Filter by author lcy-seso opened on Oct 17, 2026

• [TEST][Backend] run the accelerator path on a registered privateuse1 test backend in ci

  follow-upmulti-backendtest
  tile-ai/TileOPs#2610 · Filter by author lcy-seso opened on Oct 17, 2026

• [Backend] Open gaps for a non-CUDA hardware backend after #2164, #2165, #2166

  enhancementmulti-backend
  tile-ai/TileOPs#2595 · Filter by author lcy-seso opened on Oct 16, 2026
  2

• [Bench][MoE] Autotune the FlashInfer fused-MoE baseline on every activation

  benchfollow-up
  tile-ai/TileOPs#2562 · Filter by author michaelwithu opened on Oct 13, 2026
  

• [Perf][Attention] model paged GQA dispatch after attention sinks

  perf
  tile-ai/TileOPs#2505 · Filter by author superAngGao opened on Oct 10, 2026
  

• [Perf][Attention] GQAPagedFwdOp: H200 regressions left by the stale-V fix in #2480

  perf
  tile-ai/TileOPs#2500 · Filter by author lcy-seso opened on Oct 10, 2026
  1

• [Perf][Attention] Define GQA decode/prefill dispatch boundaries and integrate mixed-request execution

  perf
  tile-ai/TileOPs#2496 · Filter by author superAngGao opened on Oct 9, 2026
  1

  

• [Test] Restore the dispatch tests for paths only non-SM90 architectures select

  test
  tile-ai/TileOPs#2449 · Filter by author lcy-seso opened on Oct 5, 2026
  3

• [Kernel] KDAFwdOp: compute the decay from the raw gate in the in-tree kernels

  feature
  tile-ai/TileOPs#2442 · Filter by author lcy-seso opened on Oct 5, 2026
  1

  

• [Perf][Attention] Two varlen RoPE rows ship behind the FA3 stack; row-axis packing was rejected on measurement

  perf
  tile-ai/TileOPs#2382 · Filter by author lcy-seso opened on Oct 2, 2026
  1

  

• [Perf][Ops] An eager call spends 122 us on the host around a 2.2 us kernel

  perf
  tile-ai/TileOPs#2371 · Filter by author lcy-seso opened on Oct 2, 2026
  1

• [BugFix] Mixed-architecture hosts can run a kernel compiled for another GPU

  fix
  tile-ai/TileOPs#2302 · Filter by author lcy-seso opened on Sep 28, 2026

• [Kernel] GemmFp8FwdOp: close the 1D2D block-scale gap to DeepGEMM

  featurefoundry
  tile-ai/TileOPs#2282 · Filter by author lcy-seso opened on Sep 27, 2026
  

• [Tracker] GQAPagedFwdOp: leave spec-only

  feature
  tile-ai/TileOPs#2275 · Filter by author lcy-seso opened on Sep 27, 2026

• [Tracker] GQAVarlenFwdOp: leave spec-only

  feature
  tile-ai/TileOPs#2274 · Filter by author lcy-seso opened on Sep 27, 2026

• [Tracker] TileOPs operator release plan

  featurenever-stale
  tile-ai/TileOPs#2271 · Filter by author lcy-seso opened on Sep 27, 2026
  

• [Kernel] GQADenseFwdOp / GQAVarlenFwdOp / GQAPagedFwdOp: attention sinks

  featurefoundry
  tile-ai/TileOPs#2269 · Filter by author lcy-seso opened on Sep 27, 2026
  1

  

• [META][GEMM] The tuning path: what we have measured does not generalize

  chorenever-stale
  tile-ai/TileOPs#2133 · Filter by author lcy-seso opened on Sep 15, 2026

• [META][GEMM] Scheduling paths and coverage inside the existing GEMM ops

  chorenever-stale
  tile-ai/TileOPs#2130 · Filter by author lcy-seso opened on Sep 15, 2026

• Support TileLang-MUSA distribution metadata in external MUSA backend installs

  designenhancementnever-stale
  tile-ai/TileOPs#2124 · Filter by author Rankf opened on Sep 11, 2026
  2