Issues

• [NPU compiler] Group-wise scales cost one weight-table row per output channel and group: 1 bit per weight per graph

  openvinotoolkit/npu_compiler#379 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] MatMul with a constant matrix on the LEFT of a batched input is unrolled into one small matmul per batch

  openvinotoolkit/npu_compiler#378 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] NPU 3720: weightless compilation of INT4 models is blocked by ForbidFourBitOutputs

  openvinotoolkit/npu_compiler#377 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] Weightless compilation (weights separation) fails or keeps the weights for NF4, INT4 and 2-bit models

  openvinotoolkit/npu_compiler#376 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] Strategy search and vertical fusion recompute the same results for every weight group (2-bit / grouped models)

  openvinotoolkit/npu_compiler#375 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] ContentAttr::fold() re-folds a whole concatenated weight for every DPU tile (O(tiles x weight))

  openvinotoolkit/npu_compiler#374 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU compiler] Multi-cluster strategy search serialises on the Windows heap lock (TensorDistributionMap)

  openvinotoolkit/npu_compiler#373 · Filter by author ryugyosoft-max opened on Oct 20, 2026

• [NPU37XX] LLM decode: weight DMA reaches only ~34 GB/s in the real decoder graph vs 48 GB/s in a weights-only graph (~18 ms/token lost on Qwen3-4B)

  openvinotoolkit/npu_compiler#364 · Filter by author ilyamen opened on Oct 19, 2026
  1

• u2 (2-bit) constant weights: MatMul is correct for one activation row, wrong for M ≥ 2 (NPU40XX, NPU50XX)

  openvinotoolkit/npu_compiler#355 · Filter by author ryugyosoft-max opened on Oct 11, 2026
  1

• [Bug] Unbounded ReduceMean followed by Reshape segfaults in NGraphImporter

  openvinotoolkit/npu_compiler#352 · Filter by author archledger opened on Sep 19, 2026
  1

• [Build] Windows/MSVC: CMakePresets builds fail with C2855 - shared PCH built without its consumers' conformance flags

  openvinotoolkit/npu_compiler#344 · Filter by author blairducrayoppat opened on Aug 28, 2026
  1

• [Performance] Avoid repeated sorting in AsyncDepsInfo dependency queries

  openvinotoolkit/npu_compiler#342 · Filter by author Mac-Huang opened on Aug 27, 2026

• [Intel NPU][Driver Compiler] Q/DQ Concat with two C=8 Conv producers zeros output channels 8-15

  openvinotoolkit/npu_compiler#335 · Filter by author Howard9112 opened on Aug 14, 2026

• Invalid preset: "NPU3720/Silicon"

  openvinotoolkit/npu_compiler#159 · Filter by author jcubic opened on Sep 2, 2025
  1

• Build documentation missing

  openvinotoolkit/npu_compiler#155 · Filter by author jcubic opened on Sep 1, 2025

• Cannot compile the project on Fedora Linux

  openvinotoolkit/npu_compiler#144 · Filter by author JohnNiang opened on Aug 15, 2025

• Very long compile times for moderately sized matmuls

  openvinotoolkit/npu_compiler#143 · Filter by author GlowingScrewdriver opened on Aug 13, 2025

• Do we have plan to support triton

  openvinotoolkit/npu_compiler#116 · Filter by author lcwyylcwyy opened on Jul 10, 2025

• Multi-Tenant Behavior and Resource Sharing on Intel NPU4

  openvinotoolkit/npu_compiler#105 · Filter by author Kepontry opened on Jun 30, 2025

• Mysterious build failure on naming style check.

  openvinotoolkit/npu_compiler#102 · Filter by author TommyJerryMairo opened on Jun 23, 2025
  1

• Unable to Profile DMATask on NPU40XX

  bughelp wanted
  Bug

  openvinotoolkit/npu_compiler#43 · Filter by author ColorsWind opened on Jan 6, 2025
  10