Issues

• trn3: decode takes 36.2 ms vs 30.3 ms on trn2

  module:runtime-corequestion
  aws-neuron/torch-neuronx#1034 · Filter by author dacorvo opened on Nov 3, 2026
  1

• Compiled all_to_all_single with uneven split sizes fails on 2-D inputs and returns wrong rows on 1-D inputs

  bugmodule:torch-mlir
  aws-neuron/torch-neuronx#1033 · Filter by author dacorvo opened on Nov 3, 2026
  2

  

• Eager SDPA with dropout_p > 0: backward ignores the dropout mask (wrong grads)

  bugmodule:aten-ops
  aws-neuron/torch-neuronx#1031 · Filter by author xuan7zhang opened on Nov 2, 2026
  1

  

• Use the NKI flash attention kernel when SDPA gets an attention mask

  enhancementmodule:dynamo
  aws-neuron/torch-neuronx#1029 · Filter by author dacorvo opened on Nov 2, 2026
  1

  

• torch_mlir/extras/meta.py is missing, so every MLIR type lookup retries a failing import

  bugmodule:torch-mlir
  aws-neuron/torch-neuronx#1028 · Filter by author dacorvo opened on Nov 2, 2026
  1

  

• NEFF cache key of a compiled graph depends on where its Python files are installed

  bugmodule:dynamo
  aws-neuron/torch-neuronx#1027 · Filter by author dacorvo opened on Nov 2, 2026
  1

  

• cross_entropy fails on RowwiseParallel output: "Tensor not bound: lazy alloc kAlloc may not have executed"

  bugmodule:aten-ops
  aws-neuron/torch-neuronx#1026 · Filter by author xuan7zhang opened on Nov 1, 2026
  2

  

• device_map="auto" silently loads the model on the CPU instead of Neuron devices

  bug
  aws-neuron/torch-neuronx#1024 · Filter by author GauravM1781 opened on Oct 24, 2026
  1

  

• Tensor.to("neuron:1") on a tensor on neuron:0 keeps it on neuron:0 and raises no error

  bug
  aws-neuron/torch-neuronx#1021 · Filter by author GauravM1781 opened on Oct 22, 2026
  1