Repository Issue Activity (beta)

#11406 `tl.softmax` produces incorrect, non-row-wise results

#11407 [Pipeliner] predicateOp drops the predicate on NotSpeculatable arithmetic

#11328 [ClusterBarrierInsertion] Multicast TMA into a memdesc view misses its cluster barrier

#11408 Withdrawn

#11373 [SM80/A100] 2x perf regression 3.6.0 -> 3.7.1 at previously-optimal num_warps=4 (per-config optimum inverts); 13+ min compiles on large BLOCK sizes

#11393 OP ttg.local_atomic_scatter_rmw fails in ClusterBarrierAnalysis: LLVM ERROR: scratch buffer operations should not have any shared memory dependencies for

#11405 [TritonGPU] ttg.warp_specialize capturing a tensor aborts with report_fatal_error and no diagnostic

#11404 [Membar] A relaxed ttng.cluster_barrier is treated as a full cluster sync point

#11402 `tl.condition` never terminates when the initial while condition is false

#11352 `tl.cat` fails with misleading errors when operands differ along `dim`, despite its assertion allowing it

#11037 [AMD/gfx950] tl.atomic_add inside a while loop hangs

#11320 fp8 matmul on sm_120 runs at half the 8-bit rate

#10907 [v.3.8.0] Release Tracker

#10229 Triton NVIDIA backend: 'ttng.tensormap_create' op pipeliner doesn't know how to predicate this op (CUDA 13.0, sm_90 + sm_100)

#11378 [AMD] scf.for with runtime trip count miscompiled on gfx1151

#11146 Loop body is generated 2^depth times when lowering nested loops

#10486 Probabilistic illegal memory access from a kernel with multiple tl.dot on H100 (sm_90)

#11366 [NVIDIA] Hopper wgmma IMA on Triton 3.7.1: Invalid __shared__ read (H20 / sm90)

#10987 Triton generates pathologically large PTX for chained fp32 division-by-broadcast feeding a reduction

#11327 AMD compilation crash in TritonAMDGPUPrepareIfCombining

#10331 Docs: bring-up recipe for Triton on NVIDIA Blackwell sm_121 (GB10 / DGX Spark / consumer Blackwell)

#11344 `desc.scatter()` emits `tile::scatter4` on consumer Blackwell (sm_12x) and fails in ptxas instead of erroring at compile time

#9703 [Perf] TMA descriptor load followed by reduce regresses in 3.6.0 compared to 3.5.1

#11290 [tritongpu] Deduplicate the three hoistConvert*() driver functions in RemoveLayoutConversions

#11326 [Membar] No barrier between a caller's pending shared access and a callee's first one