Issues

• [Bug]: [GPU] Micro-kernel generators force 256-GRF mode even when oneDNN microkernel requires 512 GRF on Xe3p

  bugsupport_request
  openvinotoolkit/openvino#39394 · Filter by author bocui23 opened on Nov 2, 2026

• [Performance]: ARM64 CPU single-token SDPA repeats each K/V row's work for every query head of a GQA group, so with the default u8 KV cache GQA decode is no faster than MHA

  openvinotoolkit/openvino#39388 · Filter by author clemlesne opened on Nov 1, 2026
  1

• [Bug]: [PT FE] torch.export / dynamo=True gives container inputs the names of later forward parameters

  openvinotoolkit/openvino#39386 · Filter by author Somay-kousis opened on Nov 1, 2026

• [Bug]: CPU does not fuse the rank-5 SDPA that PyTorch enable_gqa=True converts to, and keeps full T×T score tensors (about 10x the peak memory of rank 4 at T = 8192, 44x with a stateful KV cache)

  openvinotoolkit/openvino#39385 · Filter by author clemlesne opened on Nov 1, 2026
  2

• [Bug]: [ARM CPU] Default f16 precision: ReduceSum accumulates in f16 and RoPE angle tables are computed in f16, both wrong on long sequences

  openvinotoolkit/openvino#39383 · Filter by author clemlesne opened on Nov 1, 2026
  2

• [Bug]: [NPU] Wrong results for RoPE pattern x*cos + rotate_half(x)*sin (Split→Negative→Concat) on NPU 3720

  openvinotoolkit/openvino#39379 · Filter by author ganiushin opened on Oct 31, 2026
  1

• [Bug] [NPU] LFM3-1.7B output breaks once the KV cache is longer than the 256 token sliding window

  openvinotoolkit/openvino#39370 · Filter by author Vincenz099 opened on Oct 31, 2026

• [CI] Rerun Workflow with Known Errors fails with HTTP 404 on runs without jobs

  openvinotoolkit/openvino#39367 · Filter by author blairducrayoppat opened on Oct 30, 2026

• [Docs]: README Discord server invite says expired

  openvinotoolkit/openvino#39364 · Filter by author RohanOnKeys opened on Oct 30, 2026

• [Bug]: Python benchmark_app -iop misparses tensor names containing multiple colons

  openvinotoolkit/openvino#39349 · Filter by author maindevhoon opened on Oct 30, 2026

• [Bug]:[GPU] IncreasePositionIdsPrecision does not cover YaRN RoPE: rotary angles stay fp16, wrong outputs at long positions

  bugsupport_request
  openvinotoolkit/openvino#39340 · Filter by author egeor opened on Oct 29, 2026
  1

• [Bug]: [CPU] GatedDeltaNet: infer() after CompiledModel::release_memory() with unchanged input shapes fails with "Memory has not been allocated"

  good first issue
  openvinotoolkit/openvino#39339 · Filter by author clemlesne opened on Oct 29, 2026
  4

  

• [Bug]: [CPU][aarch64] i32 eltwise operations silently lose precision above 2^24

  bugsupport_request
  openvinotoolkit/openvino#39337 · Filter by author PrashantTakale369 opened on Oct 29, 2026
  2

• [aw] Failed jobs: CI Failure Doctor

  agentic-workflows
  openvinotoolkit/openvino#39334 · Filter by author github-actions opened on Oct 29, 2026

• [aw] CI Failure Doctor: failed

  agentic-workflows
  openvinotoolkit/openvino#39333 · Filter by author github-actions opened on Oct 29, 2026

• [Feature Request]: Support installing OpenVINO without openvino-telemetry (make it an optional dependency)

  openvinotoolkit/openvino#39331 · Filter by author clemlesne opened on Oct 29, 2026
  1

  

• [Bug]: [CPU] CompiledModel::release_memory() keeps the Loop DynamicBuffer, the oneDNN scratchpad and StridedSlice index buffers sized by an earlier large input

  good first issue
  openvinotoolkit/openvino#39327 · Filter by author clemlesne opened on Oct 29, 2026
  4

  

• [Performance]: CPU Slice/StridedSlice with a dynamic input rebuilds its executor on every inference; a 1-D slice stores 16 bytes of indices per output element

  good first issue
  openvinotoolkit/openvino#39325 · Filter by author clemlesne opened on Oct 29, 2026
  3

  

• [Performance]: ARM64 CPU without 256-bit SVE: grouped-query SDPA (PyTorch enable_gqa) runs about 2.6x slower because its rank-5 MatMuls fall back to oneDNN ref_gemm (reported as jit_gemm_f32)

  openvinotoolkit/openvino#39324 · Filter by author clemlesne opened on Oct 29, 2026
  1

• [Performance]: ARM64 CPU: INT8 dynamic-quantized FullyConnected runs the 1x4 dotprod GEMV for every row count on cores without i8mm

  openvinotoolkit/openvino#39319 · Filter by author clemlesne opened on Oct 29, 2026
  1

• [GPU] Custom layer reading both outputs of a two-output SimpleGPU custom layer: name collision (BFYX) or wrong output port (ANY)

  openvinotoolkit/openvino#39300 · Filter by author TheSanchezian opened on Oct 29, 2026
  1

• [GPU] Two custom layers whose kernels are in the same .cl file fail: "Number of compiled kernels is different than kernel batch size"

  openvinotoolkit/openvino#39299 · Filter by author TheSanchezian opened on Oct 28, 2026
  1

• [GPU] Static rank-3 output of a multi-output custom layer gets a 4D layout; Assign fails with layout mismatch

  openvinotoolkit/openvino#39298 · Filter by author TheSanchezian opened on Oct 28, 2026
  1

• [GPU] Identical SimpleGPU custom layers are compiled once per layer (friendly name is embedded in the kernel source)

  openvinotoolkit/openvino#39297 · Filter by author TheSanchezian opened on Oct 28, 2026
  2

• [GPU] Access violation when a model with a SimpleGPU custom layer is loaded from CACHE_DIR

  openvinotoolkit/openvino#39296 · Filter by author TheSanchezian opened on Oct 28, 2026
  2