Issues

• [Bug] LayoutInference fails when a same-dtype fragment copy sits between two GEMMs with different layout requirements (float32)

  bug
  tile-ai/tilelang#3669 · Filter by author maxli0182 opened on Oct 26, 2026

• [BUG] T.reduce_* batch=N fails to lower on sm_100/sm_120 when the fragment layout gives each thread fewer outputs

  bug
  tile-ai/tilelang#3667 · Filter by author lucifer1004 opened on Oct 26, 2026

• [BUG][CUDA] Warp-specialized pipeline under an if fails in ptxas (setmaxnreg) when a statement follows the if

  bug
  tile-ai/tilelang#3664 · Filter by author lucifer1004 opened on Oct 25, 2026

• [BUG][CUDA] Warp specialization moves prelude statements from outside an outer T.serial loop into it (wrong results)

  bug
  tile-ai/tilelang#3663 · Filter by author lucifer1004 opened on Oct 25, 2026

• [Bug] Concurrent autotune cache repair can remove a published result

  bugpython
  tile-ai/tilelang#3659 · Filter by author yampolskiyy opened on Oct 25, 2026

• [BUG][ice-on-valid-code] ThreadSync rejects a 448-thread partial barrier as not warp-uniform under an & condition instead of emitting it

  bug
  tile-ai/tilelang#3653 · Filter by author Hughshine opened on Oct 25, 2026

• [BUG] CUDA misaligned address with align=4 input

  arithbug
  tile-ai/tilelang#3650 · Filter by author sepcnt opened on Oct 25, 2026

• [BUG] cython/nvrtc/cutedsl adapters allocate out_idx outputs on the current stream but launch on the caller's stream

  bugpython
  tile-ai/tilelang#3648 · Filter by author lucifer1004 opened on Oct 25, 2026
  1

• [BUG] ReduceOp lowering fails on sm_120 for gdn wy_fast_bwd_split (source index depends on a thread-owned reduce segment)

  bug
  tile-ai/tilelang#3647 · Filter by author lucifer1004 opened on Oct 25, 2026
  1

• [BUG][Examples] attention_sink, gqa_bwd and block_causal_attention_varlen exceed sm_120 per-block shared memory

  bug
  tile-ai/tilelang#3646 · Filter by author lucifer1004 opened on Oct 25, 2026
  3

• [BUG][CUDA] sm_100: TMEM is deallocated without a CTA barrier; with warp specialization the producer warp frees the accumulator while consumer tcgen05.mma / tcgen05.ld still use it

  bug
  tile-ai/tilelang#3645 · Filter by author jeffchen006 opened on Oct 24, 2026

• [BUG] Intermittent incorrect results with automatic warp specialization and FP8-to-BF16 shared-memory conversion

  bug
  tile-ai/tilelang#3641 · Filter by author qiaomuf opened on Oct 24, 2026

• [BUG][Fuzzer][missed-optimization] fp8->fp16 T.copy with a runtime row offset runs 3.2x slower: bounds checked twice instead of once

  bug
  tile-ai/tilelang#3635 · Filter by author Hughshine opened on Oct 24, 2026

• [BUG] VerifyBufferInit warnings have no source location for reads outside tile ops

  bugUX
  tile-ai/tilelang#3619 · Filter by author RyanL2 opened on Oct 22, 2026

• [BUG] [ROCm] T.exp, T.exp10, T.sin and T.cos lower to fast intrinsics on HIP

  bug
  tile-ai/tilelang#3618 · Filter by author jzluo opened on Oct 22, 2026

• [BUG] [Ascend] Interleaved strided UB->GM stores from AIV cores inside T.PerCoreTask() fail to compile: "does not support synchronization between AIV0 and AIV1"

  ascendbug
  tile-ai/tilelang#3616 · Filter by author liggest opened on Oct 22, 2026
  3

• [Question] Why is .add_attr_option<int64_t>("max_vector_bits", refl::DefaultValue(64)) set to 64 instead of 128?

  ascendquestion
  tile-ai/tilelang#3615 · Filter by author yuqiha opened on Oct 22, 2026
  1

• [BUG][Fuzzer][wrong-code] T.Pipelined silently races on a shared buffer accessed through T.reshape/T.view instead of pipelining it

  bug
  tile-ai/tilelang#3598 · Filter by author Hughshine opened on Oct 21, 2026
  1

• Proposal: Add TsingMicro TX81 backend as a downstream distribution (tilelang-tsingmicro)

  enhancement
  tile-ai/tilelang#3590 · Filter by author tsingmicro-public-e opened on Oct 21, 2026

• [BUG][Metal] MetalKernelAdapter binds buffers positionally, silently corrupting arguments when codegen reorders them

  bugmetalpython
  tile-ai/tilelang#3587 · Filter by author lucifer-holly opened on Oct 21, 2026

• [BUG] flash_decoding examples keep the split-KV log-sum-exp (glse) in fp16: wrong results when scores are large

  bugpython
  tile-ai/tilelang#3585 · Filter by author memxlife opened on Oct 20, 2026

• [Feature Request] Add reduce_any/reduce_all and unify logical reductions

  bugenhancement
  tile-ai/tilelang#3584 · Filter by author LeiWang1999 opened on Oct 20, 2026
  

• [BUG][Fuzzer][wrong-code] Signed x >> n with an unsigned n at least as wide as x silently zero-fills instead of sign-extending

  arithbug
  tile-ai/tilelang#3574 · Filter by author Hughshine opened on Oct 19, 2026
  2

• Wrong results: explicit linear shared-memory layout on GEMM operands when copies lower to TMA (sm_121a)

  bug
  tile-ai/tilelang#3571 · Filter by author memxlife opened on Oct 19, 2026

• [BUG] set_autotune_inputs moves captured tensors to CPU on non-CUDA hosts, breaking autotuning

  ascendbugpython
  tile-ai/tilelang#3545 · Filter by author liggest opened on Oct 16, 2026