Issues

• vkML on NVIDIA RTX 4060 Laptop GPU (Windows)

  help wanted
  jeel00dev/vkml#37 · Filter by author Jayt257 opened on Aug 2, 2026

• sign returns -0.0 for +0.0 and NaN on Mesa 25.2.8, +0.0 on 26.1.5

  jeel00dev/vkml#36 · Filter by author jeel00dev opened on Jul 31, 2026

• test_invariants' RADV register model is pinned to one Mesa version, and requires_radv does not guard that

  jeel00dev/vkml#35 · Filter by author jeel00dev opened on Jul 31, 2026

• load_state_dict dropped the device, so load_module could not restore into a GPU model

  jeel00dev/vkml#34 · Filter by author jeel00dev opened on Jul 31, 2026

• A submission costs ~105us against ~9us for a dispatch, so 85-89% of a training step is submission overhead, not work

  help wanted
  jeel00dev/vkml#33 · Filter by author jeel00dev opened on Jul 31, 2026

• The optimiser is 62.7% of an MNIST MLP step: 12 submissions for 4 parameters, and it is why the GPU loses to a 6-thread CPU

  help wanted
  jeel00dev/vkml#32 · Filter by author jeel00dev opened on Jul 31, 2026
  1

• CPU backend is ~116x slower than PyTorch on the same machine and has no threading at all; a 10-epoch CIFAR-100 run is not completable

  help wanted
  jeel00dev/vkml#31 · Filter by author jeel00dev opened on Jul 31, 2026