Repository Issue Activity (beta)

#5817 xbyak_aarch64: Valgrind reports uninitialized stack reads in CpuInfoLinux::getLineInFile

#6035 1x1 convolution backward_weights writes past the rtus workspace (nhwc, IC < ic_block)

#5973 CPU x64: how should FP16 GEMM request accumulation_mode::f16 (AVX10.1 / native f16 acc)?

#5942 Convolution segfaults in a corner case

#5170 Proposal to add a real RISC-V hardware CI validation platform

#5912 matmul: grouped GEMM returns wrong results with binary_mul post-op on bf16 + acb weights (GPU)

#5933 [oneDNN] Fused SDPA GPU page-fault / device lost when head_dim=64 with a rank-4 attn_mask spanning more than one tile on both axes (Sq>=17, Sk>=129); rank-2 mask is safe

#5908 matmul: grouped GEMM silently produces wrong results when weight-scale memory is not canonical dense abc

#5510 oneDNN Graph API SDPA return wrong result on Arc770, Arc 140T in some cases

#5276 AArch64: CI benchdnn tests are not a subset of NIGHTLY test set

#5808 gpu: sdpa: xe_hpg head_size=128 config is 1.6-1.8x slower than a 16-subgroup alternative on Arc A770

#5730 gpu, jit:ir convolution backward_data fails at primitive creation (internal error, src/gpu/intel/conv/jit.cpp:161) for degenerate-kernel + strided geometry (kw=1, sw=1, sh>1) on Arc Battlemage

#5758 gpu: sdpa: split-barrier microkernel intermittently returns non-finite rows on Arc B580

#4381 [Graph API] Is compiled_partition thread-safe for concurrent calls on same instance?

#1667 Bad perf for matmul of ba tensors

#5689 Illegal instruction on EMR + Windows

#5393 Data race on shared brgemm ukernel AMX tile config (palette_) under concurrent set_hw_context()

#5785 gpu: sdpa: micro_sdpa quantized (INT8-key) systolic kernel produces wrong results on Intel Arc A770

#5459 Grouped Matmul (MoE): roadmap for stabilization, optimized CPU kernels, and fused gated activations

#5717 [CPU][AMX] brg_matmul bf16 kernel leaks NaN/garbage data into unrelated output rows when K dimension is not a multiple of 32

#5732 benchdnn conv can abort (failed assert) in some modes for non-ref impls

#5499 Document how to build and configure xla headers for EIGEN_ASYNC threadpool tests

#5313 AArch64: get_per_core_cache_size()==0 corrupts reorder output and crashes int8 concat

#5578 Auto-fetching and building ZenDNN from within oneDNN's build

#5564 Discussion on ZenDNN inner product implementation