Repository Issue Activity (beta)

#3071 Silent all-zeros miscompilation of mamba-ssm's Mamba-3 MIMO kernel on GB10 (sm_121) — depends on JIT compile order, reproduces on 0.1.8 and 0.1.13

#2623 [BUG][Fuzzer][wrong-code] `T.finalize_reducer(reducer, batch>1)` silently reduces only the first `batch` elements

#2569 [BUG][Fuzzer][ice-on-valid-code] `T.shfl_*` on `float16`/`bfloat16` fails to compile instead of shuffling the value

#3062 [BUG] export_library fails on disk-cache hit even with execution_backend="tvm_ffi"

#1729 [Feature Request] Improve layout when small fragments cause excessive thread replication

#3040 [BUG][Fuzzer][ice-on-valid-code] A width-changing `T.view` of a fragment aborts with an internal layout-inference `ICHECK` instead of a clear diagnostic

#2913 [Feature Request] Add TileLang support to Compiler Explorer (Godbolt)

#2688 [BUG][Fuzzer][wrong-code] Vectorized `T.atomic_add` silently downgrades the requested `memory_order` to relaxed

#2985 [BUG][Fuzzer][ice-on-valid-code] `T.max`/`T.min`/`T.clamp` on `float8_e4m3`/`float8_e5m2` fail to compile instead of computing the value

#3042 [BUG] 屏障误删导致的共享内存的读写冲突

#3039 [BUG][Fuzzer][ice-on-valid-code] Reading a `T.Pipelined` producer shared buffer after the loop crashes the compiler instead of compiling

#2989 [BUG][Fuzzer][ice-on-valid-code] `T.__ldg` on `float16`/`bfloat16` fails to compile instead of loading the value

#3038 [BUG][Fuzzer][wrong-code] Auto-vectorized `float32` `T.atomic_add` into shared emits a global-only vector builtin and crashes at launch (sm_90)

#3037 [BUG][Fuzzer][ice-on-valid-code] 8-bit-operand sparse `T.gemm_sp` with `K > block_K` crashes `ThreadSync` (`Cannot match type int32 vs handle`) instead of compiling

#3035 [BUG][Fuzzer][ice-on-invalid-code] `T.gemm` `k_pack` outside `{1, 2}` aborts with an internal `ICHECK` assert deep in a compiler pass instead of a clean frontend error

#3034 [BUG][Fuzzer][ice-on-invalid-code] `T.rng_rand`/`T.rng_rand_float` with no preceding `T.rng_init` emit `curand…(&)` and crash nvcc, instead of a clean "init required" error

#3033 [BUG][Fuzzer][ice-on-invalid-code] `T.finalize_reducer` on a reducer never reduced in a `T.Parallel` loop aborts with an internal `bad_optional_access` in `LayoutInference`

#3032 [BUG][Fuzzer][ice-on-invalid-code] `T.gemm` on a narrow dtype (`int8`/`fp8`/`bfloat16`) with any sub-atom `block_K` crashes nvcc instead of cleanly rejecting

#3031 [Enhancement] Pipeline planner `LOG(FATAL)`s on a legally reused shared buffer (region-overlap check has no liveness); add liveness or at least a clean diagnostic

#3030 [BUG][Fuzzer][diagnostic] T.alloc_barrier(0) is accepted through the whole compile pipeline and emits `Barrier::init(0)` with no diagnostic

#3029 [BUG][Fuzzer][ice-on-valid-code] Binding `T.rng_init()`'s documented state handle to a variable emits `void state = ;` and nvcc rejects the kernel

#3027 [BUG][Fuzzer][ice-on-valid-code] LoopVectorize's internal index-scale product overflows int32 and aborts compile for a large-stride affine access whose every real index fits int32

#3026 [BUG][Fuzzer][ice-on-valid-code] `T.loop_break()` inside `T.unroll(explicit=True)` emits bare `break;` outside any loop, so nvcc rejects the kernel instead of it compiling

#3025 [BUG][Fuzzer][wrong-code] `T.ldg*`/`T.stg*`/`T.lds*` on a `T.StridedTensor` address the wrong element — offset uses buffer shape, not the declared strides

#3024 [BUG][Fuzzer][wrong-code] `T.clamp` silently drops `NaN` and returns the `min` bound instead of propagating `NaN`