Issues

• [Fuzzer][Candidates] Blackwell / sm_100 (tcgen05) source-only bug-candidate list — unconfirmed, awaiting silicon

  tile-ai/tilelang#2644 · Filter by author Hughshine opened on Jul 12, 2026
  

• [BUG][Fuzzer][ice-on-valid-code] T.clear of a shared buffer + warp specialization on sm_90 → fill partitioned over 384 threads → non-bijective layout ICE

  tile-ai/tilelang#2643 · Filter by author Hughshine opened on Jul 12, 2026
  1

• [BUG][Fuzzer][ice-on-valid-code] MMA T.gemm reuses an incompatible T.annotate_layout on a shared operand → launch misaligned address; sm_90 rejects it at compile

  tile-ai/tilelang#2642 · Filter by author Hughshine opened on Jul 12, 2026
  1

• [BUG][Fuzzer][wrong-code] Default (non-fast-math) path silently folds float x * 0.0 → 0.0 / x - x → 0, returning a finite value where IEEE-754 gives NaN

  tile-ai/tilelang#2638 · Filter by author Hughshine opened on Jul 12, 2026

• [BUG][Fuzzer][ice-on-valid-code] T.atomic_add(use_tma=True) crashes at launch when shared box > swizzle span

  tile-ai/tilelang#2630 · Filter by author Hughshine opened on Jul 12, 2026
  1

• [BUG][Fuzzer][ice-on-valid-code] 256-bit vectorized global store emits tl::store_global_256, defined only on sm_100 — nvcc rejects it on sm_89/sm_90

  tile-ai/tilelang#2629 · Filter by author Hughshine opened on Jul 12, 2026
  1

• [BUG][Fuzzer][wrong-code] T.rng_init() default seq is x-dim-only, silently giving y/z-differing threads one RNG stream

  tile-ai/tilelang#2625 · Filter by author Hughshine opened on Jul 12, 2026
  2

• [BUG][Fuzzer][wrong-code] T.finalize_reducer(reducer, batch>1) silently reduces only the first batch elements

  tile-ai/tilelang#2623 · Filter by author Hughshine opened on Jul 12, 2026
  1

• [BUG][Fuzzer][wrong-code] DecoupleTypeCast silently clobbers vector lanes the source left unwritten when a cast-store sits under a nested-if (or if/else) guard

  bug
  tile-ai/tilelang#2616 · Filter by author Hughshine opened on Jul 11, 2026

• [BUG][Fuzzer][wrong-code] Un-pipelined warp-specialized GEMM sharing one shared operand between both warp groups silently races on Hopper — missing loop-back __syncthreads() (pipelined num_stages>=2 is unaffected)

  bug
  tile-ai/tilelang#2613 · Filter by author Hughshine opened on Jul 11, 2026
  1

• [BUG][Fuzzer][wrong-code] tl.ThreadSync silently drops an in-loop __syncthreads() inside a threadIdx-dependent if, racing to a wrong result

  bug
  tile-ai/tilelang#2612 · Filter by author Hughshine opened on Jul 11, 2026

• [BUG][Fuzzer][ice-on-valid-code] Scalar T.cast for two dtype pairs that IsCudaVectorizableCast declares supported (float8_e8m0fnu -> bfloat16, float64 -> float4_e2m1fn) leaks an nvcc error while the vectorized form compiles

  bug
  tile-ai/tilelang#2610 · Filter by author Hughshine opened on Jul 11, 2026

• [BUG][Fuzzer][wrong-code] A cross-dtype T.view of a shared buffer that carries a swizzle (or any non-identity) layout silently reads/writes the wrong storage slots

  bug
  tile-ai/tilelang#2609 · Filter by author Hughshine opened on Jul 11, 2026
  1

• [BUG][Fuzzer][wrong-code] make_cutlass_metadata_layout (sm8x) silently emits a non-bijective, out-of-range E-metadata layout when M is not a multiple of the interleave group

  bug
  tile-ai/tilelang#2606 · Filter by author Hughshine opened on Jul 10, 2026
  2

• [BUG][Fuzzer][wrong-code] Sparse T.gemm_sp with a K not a multiple of the sparse MMA K-atom silently drops the K-tail instead of computing it or rejecting

  bug
  tile-ai/tilelang#2605 · Filter by author Hughshine opened on Jul 10, 2026
  1

• [BUG][Fuzzer][ice-on-valid-code] T.tma_gather4/T.tma_scatter4 with a constant row index segfaults the compiler instead of compiling or rejecting

  bug
  tile-ai/tilelang#2601 · Filter by author Hughshine opened on Jul 10, 2026
  1

• [BUG][Fuzzer][wrong-code] T.warp_reduce_* silently miscomputes on a partial warp and leaks an nvcc error on a float bitop (no dtype/participant guard)

  bug
  tile-ai/tilelang#2599 · Filter by author Hughshine opened on Jul 10, 2026
  1

• [BUG][Fuzzer][ice-on-valid-code] a param written only via a bare .data extern arg is marked const, leaking an nvcc error instead of compiling

  bug
  tile-ai/tilelang#2598 · Filter by author Hughshine opened on Jul 10, 2026

• [BUG][Fuzzer][wrong-code] T.annotate_restrict_buffers silently emits __restrict__ on an aliasing *_handle buffer, giving a wrong result

  bug
  tile-ai/tilelang#2596 · Filter by author Hughshine opened on Jul 10, 2026

• [BUG][Fuzzer][wrong-code] software pipelining silently reorders a read past an atomic_add to the same buffer, giving a stale result

  bug
  tile-ai/tilelang#2595 · Filter by author Hughshine opened on Jul 10, 2026
  2

• [BUG][Fuzzer][wrong-code] IfStmtBinding silently reorders a scalar read past an atomic write to the same buffer, returning the post-atomic value

  bug
  tile-ai/tilelang#2594 · Filter by author Hughshine opened on Jul 10, 2026

• [BUG][Fuzzer][wrong-code] WGMMA T.gemm with block_N ≡ 8 (mod 16) silently miscomputes the output tile on sm_90 (no error)

  bug
  tile-ai/tilelang#2593 · Filter by author Hughshine opened on Jul 10, 2026
  3

• [BUG][Fuzzer][ice-on-valid-code] 1-D bulk TMA T.copy with a non-16B-aligned base crashes CUDA_ERROR_MISALIGNED_ADDRESS instead of copying

  bug
  tile-ai/tilelang#2584 · Filter by author Hughshine opened on Jul 9, 2026

• [BUG][Fuzzer][ice-on-valid-code] uint64 constant folding crashes for an Add/Mul result in [2^63, 2^64) instead of producing the constant

  bug
  tile-ai/tilelang#2582 · Filter by author Hughshine opened on Jul 9, 2026

• [BUG][Fuzzer][wrong-code] T.copy of a column-padded shared tile silently packs the wrong pitch (interleaved rows) not the declared stride

  bug
  tile-ai/tilelang#2581 · Filter by author Hughshine opened on Jul 9, 2026