model: /Users/nayte/qwen-ane-release/models/qwen25-1.5b-ane-ctx512-lut4/qwen25_monolithic_full_lut4.mlmodelc  function: prefill
total ops: 7108
  none   4735  (66.6%)
  ANE   2360  (33.2%)
  CPU     13  (0.2%)

=== assigned-compute ops only (ANE+CPU+GPU = 2373) ===
  ANE  2360  (99.5% of compute)
  CPU    13  (0.5% of compute)
  GPU     0  (0.0% of compute)

=== top op types in 'none' ===
  const 4409
  ios18.constexpr_lut_to_dense 213
  ios18.concat 112
  ios18.add 1

=== top op types in 'ANE' ===
  ios18.mul 394
  ios18.slice_by_index 281
  ios18.transpose 271
  ios18.reshape 226
  ios18.conv 212
  ios18.add 196
  ios18.expand_dims 113
  ios18.concat 113

=== top op types in 'CPU' ===
  select 2
  ios18.greater_equal 2
  ios18.add 2
  ios18.minimum 1
  ios18.maximum 1
  ios18.sub 1
  ios18.equal 1
  ios18.gather 1
KV-machinery placement  (model: /Users/nayte/qwen-ane-release/models/qwen25-1.5b-ane-ctx512-lut4/qwen25_monolithic_full_lut4.mlmodelc  fn:prefill)
  slice_update ANE=56
  read_state ANE=56  CPU=1
  write_state ANE=56
  slice_by_index ANE=281
  gather ANE=2  CPU=1