Issues

• Smaller items worth cleaning up (tracking)

  P3
  CanReader/FastNN#30 · Filter by author CanReader opened on Aug 25, 2026

• build.rs picks CUDA paths from the host target, not the build target

  area:buildbugP3
  CanReader/FastNN#29 · Filter by author CanReader opened on Aug 25, 2026

• StepDecay and OneCycle accept values that divide by zero or underflow

  area:optimbuggood first issueP3
  CanReader/FastNN#28 · Filter by author CanReader opened on Aug 25, 2026

• Zero length sequences panic with a message about cat

  area:nnbuggood first issueP3
  CanReader/FastNN#27 · Filter by author CanReader opened on Aug 25, 2026

• permute_nd reads out of bounds for tensors of rank > 8

  area:cudabugP1
  CanReader/FastNN#26 · Filter by author CanReader opened on Aug 25, 2026

• kernels.cu is about 40% dead code and ffi.rs no longer mirrors it

  area:cudaenhancementP3
  CanReader/FastNN#25 · Filter by author CanReader opened on Aug 25, 2026

• CUDA errors are only checked at launch, so faults surface somewhere unrelated

  area:cudabugP1
  CanReader/FastNN#24 · Filter by author CanReader opened on Aug 25, 2026

• Pooling has no CUDA path, every pool round trips through the host

  area:cudaenhancementP2
  CanReader/FastNN#23 · Filter by author CanReader opened on Aug 25, 2026

• Reduction backward passes force a device sync via grad.item()

  area:autogradenhancementP2
  CanReader/FastNN#22 · Filter by author CanReader opened on Aug 25, 2026

• sum_axis and min cudaMalloc/cudaFree on every call, which syncs the device

  area:cudaenhancementP2
  CanReader/FastNN#21 · Filter by author CanReader opened on Aug 25, 2026

• Dropout builds its mask on the host and uploads it every forward

  area:nnenhancementP2
  CanReader/FastNN#20 · Filter by author CanReader opened on Aug 25, 2026

• PositionalEncoding re-uploads its whole table on every forward

  area:nnenhancementP2
  CanReader/FastNN#19 · Filter by author CanReader opened on Aug 25, 2026

• CPU broadcast ops are single threaded and clone both operands

  area:tensorenhancementP2
  CanReader/FastNN#18 · Filter by author CanReader opened on Aug 25, 2026

• The CUDA block cache is unbounded across size classes

  area:cudaenhancementP2
  CanReader/FastNN#17 · Filter by author CanReader opened on Aug 25, 2026

• The causal mask cache grows without bound

  area:nnbugP1
  CanReader/FastNN#16 · Filter by author CanReader opened on Aug 25, 2026

• A partial MNIST download gets cached forever

  area:databugP1
  CanReader/FastNN#15 · Filter by author CanReader opened on Aug 25, 2026

• A corrupt checkpoint aborts the process instead of returning an error

  area:serializebugP0
  CanReader/FastNN#14 · Filter by author CanReader opened on Aug 25, 2026

• Checkpoint writes are not atomic, a crash mid save destroys the previous one

  area:serializebugP0
  CanReader/FastNN#13 · Filter by author CanReader opened on Aug 25, 2026

• Centered RMSprop can take the square root of a negative number

  area:optimbugP1
  CanReader/FastNN#12 · Filter by author CanReader opened on Aug 25, 2026

• BatchNorm2d tracks the biased variance in running_var, CPU and CUDA disagree

  area:nnbugP1
  CanReader/FastNN#11 · Filter by author CanReader opened on Aug 25, 2026

• cuBLAS is put into TF32 mode, which quietly halves matmul precision

  area:cudabugP1
  CanReader/FastNN#10 · Filter by author CanReader opened on Aug 25, 2026

• fastnn_cuda_argmax is wrong for most inputs (non power of two block, single block)

  area:cudabugP1
  CanReader/FastNN#9 · Filter by author CanReader opened on Aug 25, 2026

• kernel_sum warp reduction is undefined behaviour on Volta and newer

  area:cudabugP1
  CanReader/FastNN#8 · Filter by author CanReader opened on Aug 25, 2026

• Device::cuda(n) silently gives you device 0 for every n

  area:cudabugP1
  CanReader/FastNN#7 · Filter by author CanReader opened on Aug 25, 2026

• sum_axis_keep and mean_axis_keep add an extra rank on 1-D input

  area:tensorbuggood first issueP1
  CanReader/FastNN#6 · Filter by author CanReader opened on Aug 25, 2026