Repository Issue Activity (beta)

#2536 [BUG][Fuzzer] `T.cumsum`/`T.cummax` over a row-offset 2-D shared sub-region silently scans the wrong rows instead of the requested ones

#2530 [BUG][Fuzzer] `T.reduce_*` with a documented keep-dim `(M, 1)` destination aborts compilation with an internal `ICHECK` (`vars.size() == InputDim()` 2 vs. 1) instead of compiling

#2529 [BUG][Fuzzer] A partial-tile 1-D shared→global `T.copy` crashes the compiler with `IndexError` in `InferBulkLayout` instead of lowering or rejecting cleanly

#2528 [BUG][Fuzzer] Pipelined `T.copy` into a 1-D `block_K=32` fp16 shared buffer crashes with `CUDA error: misaligned address` instead of running

#2527 [BUG][Fuzzer] Partial-region shared→global `T.copy` with a non-16B-aligned innermost offset crashes with `cudaErrorIllegalInstruction` instead of copying

#2526 [BUG][Fuzzer] `T.copy` over a `T.make_tensor` view emits an unconditional 16-byte `int4` load/store that crashes with `CUDA error: misaligned address` instead of honoring (or rejecting on) the view's runtime base alignment

#2525 [BUG][Fuzzer] 1-D bulk TMA store of a shared→global `T.copy` whose byte size isn't a multiple of 16 silently drops the unaligned tail instead of copying the whole buffer

#2524 [BUG][Fuzzer] `T.reduce_sum`/`T.reduce_max` with `batch > 1` silently returns a wrong reduction instead of the correct one when the reduce dimension spans fewer threads than the block

#2523 [BUG][Fuzzer] `T.cumsum`/`T.cummax` over a non-contiguous 2-D shared sub-region silently returns wrong numbers instead of a correct prefix scan

#2522 [BUG][Fuzzer] WGMMA `T.gemm` with `FullRow` and >1 warpgroup along M (e.g. `block_M=256`) silently swaps 64-row output sub-blocks instead of storing each row to its correct place

#2516 [Bug] target="auto" detects arch from hardcoded device 0 — wrong-arch kernels on heterogeneous multi-GPU hosts (CUDA_ERROR_NO_BINARY_FOR_GPU)

#2394 [BUG][Fuzzer] LayoutInference crashes with an internal "Divide by zero" instead of compiling — column-broadcast into a 2-D `T.Parallel` when the tile width isn't a power of two

#2483 [BUG][Fuzzer] `T.Cast` of an integer to `bfloat16` silently truncates toward zero instead of round-to-nearest-even — int→`float16` and `float32`→bf16 round correctly

#2510 [BUG][Fuzzer] `M.ieee_*` is only correct for `float32`: `float64` fails to compile, `float16`/`bfloat16` silently miscompute

#2478 [BUG][Fuzzer] `T.atomic_addx4` on a matched `float16`/`bfloat16` destination silently returns wrong numbers (and crashes at unaligned offsets) instead of doing a 4-element atomic add — no fp16/bf16 `AtomicAddx4` specialization

#2506 [BUG][Fuzzer] `T.copy_cluster` emits a `cp.async.bulk` with a non-16B transfer size (PTX-undefined behavior; hangs the kernel on H100) instead of rounding or rejecting it

#2482 [BUG][Fuzzer] `_tir_u32_to_int_to_float` silently decodes signed sub-word ints as unsigned (zero-extends instead of sign-extending) — every negative value decodes positive

#2501 [Question] [RFC ] Is P2P already enough?

#2504 [BUG][Fuzzer] Public TMEM/TMA DSL builtins have no arch gate — a too-low target fails with a confusing `namespace "tl" has no member` nvcc error instead of a clean rejection

#2503 [BUG][Fuzzer] Stochastic-rounding `f32`→fp4/fp8 cast is gated `sm_80+` but needs `sm_100a` — ptxas aborts instead of a clean rejection

#2479 [BUG][Fuzzer] `decode_i2s_to_i4s` silently decodes signed int2 as unsigned (sign-correction branch is an empty `if constexpr`) — every negative int2 loses its sign

#2477 [BUG][Fuzzer] `decode_i4s_to_i8s` silently decodes signed int4 as `nibble − 7` instead of two's-complement, disagreeing with its own scalar sibling

#2497 [BUG] `T.Pipelined` auto-`cp.async` lowering emits local-to-local `cp_async_gs_conditional`, causing runtime `CUDA_ERROR_INVALID_ADDRESS_SPACE` on TileLang 0.1.9

#2489 [BUG][Fuzzer] Vectorized broadcast-store of a negative `int32` across a 256-bit vector silently writes the sign-extension (not the value) to half the lanes on Blackwell (sm_100)

#2481 [BUG][Fuzzer] `_tir_f32x2_to_bf16x2_to_u32` silently turns inf/NaN into ±0.0/±inf instead of preserving it — RNE bias add has no `exp == 0xFF` guard