[
  {
    "phase": 57,
    "mode": "original",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104922,
    "receive_fragments": 118035,
    "maximum_endpoint_fragments": 154,
    "transmitted_bytes": 3359520,
    "received_bytes": 6716448,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "source",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 11666,
    "receive_fragments": 24779,
    "maximum_endpoint_fragments": 26,
    "transmitted_bytes": 3359520,
    "received_bytes": 6716448,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2560,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 104922,
    "affine_copy_loops": 11666,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "destination",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104922,
    "receive_fragments": 118035,
    "maximum_endpoint_fragments": 154,
    "transmitted_bytes": 3359520,
    "received_bytes": 6716448,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 1961,
    "affine_copy_loops": 1610,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 57,
    "mode": "both",
    "label": "op Some(22) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 11666,
    "receive_fragments": 24779,
    "maximum_endpoint_fragments": 26,
    "transmitted_bytes": 3359520,
    "received_bytes": 6716448,
    "maximum_outgoing_bytes": 2560,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 2560,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 7168,
    "source_copy_max_bytes": 2560,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 106883,
    "affine_copy_loops": 13276,
    "copy_cycles_floor": 896,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([4736]) RowMajor",
    "transfers": 5710,
    "receive_fragments": 5710,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 39992320,
    "received_bytes": 39992320,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 37888,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([4736]) RowMajor",
    "transfers": 5710,
    "receive_fragments": 5710,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 39992320,
    "received_bytes": 39992320,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 37888,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 28416,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1450,
    "affine_copy_loops": 1450,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([4736]) RowMajor",
    "transfers": 5710,
    "receive_fragments": 5710,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 39992320,
    "received_bytes": 39992320,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 37888,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 37888,
    "combined_staging_max_bytes": 37888,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 37888,
    "local_copy_runs": 2830,
    "affine_copy_loops": 2830,
    "copy_cycles_floor": 4736,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 17,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([4736]) RowMajor",
    "transfers": 5710,
    "receive_fragments": 5710,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 39992320,
    "received_bytes": 39992320,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 37888,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 37888,
    "combined_staging_max_bytes": 66304,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 37888,
    "local_copy_runs": 4280,
    "affine_copy_loops": 4280,
    "copy_cycles_floor": 8288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 72]) F16 Amp(Left) -> TensorShape([32, 1, 80]) Amp(Left)",
    "transfers": 128,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 3584,
    "received_bytes": 3584,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 112,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 72]) F16 Amp(Left) -> TensorShape([32, 1, 80]) Amp(Left)",
    "transfers": 128,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 3584,
    "received_bytes": 3584,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 112,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 32,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 128,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 72]) F16 Amp(Left) -> TensorShape([32, 1, 80]) Amp(Left)",
    "transfers": 128,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 3584,
    "received_bytes": 3584,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 112,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 112,
    "combined_staging_max_bytes": 112,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 112,
    "local_copy_runs": 32,
    "affine_copy_loops": 32,
    "copy_cycles_floor": 14,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 74,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 72]) F16 Amp(Left) -> TensorShape([32, 1, 80]) Amp(Left)",
    "transfers": 128,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 3584,
    "received_bytes": 3584,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 112,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 112,
    "combined_staging_max_bytes": 112,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 112,
    "local_copy_runs": 160,
    "affine_copy_loops": 160,
    "copy_cycles_floor": 18,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "original",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 729, 4304]) F16 RowMajor -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 6275232,
    "received_bytes": 6275232,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 8608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "source",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 729, 4304]) F16 RowMajor -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 6275232,
    "received_bytes": 6275232,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 8608,
    "source_staging_max_bytes": 8608,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8608,
    "source_copy_max_bytes": 8608,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 1076,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "destination",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 729, 4304]) F16 RowMajor -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 6275232,
    "received_bytes": 6275232,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 8608,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 8608,
    "combined_staging_max_bytes": 8608,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 8608,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 1076,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 53,
    "mode": "both",
    "label": "op Some(19) LayoutRearrangement: TensorShape([2, 729, 4304]) F16 RowMajor -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 6275232,
    "received_bytes": 6275232,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 8608,
    "source_staging_max_bytes": 8608,
    "destination_staging_max_bytes": 8608,
    "combined_staging_max_bytes": 8608,
    "source_copy_max_bytes": 8608,
    "destination_copy_max_bytes": 8608,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 2152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "original",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1679616,
    "received_bytes": 1679616,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 0,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 0,
    "affine_copy_loops": 0,
    "copy_cycles_floor": 0,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "source",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1679616,
    "received_bytes": 1679616,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "destination",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1679616,
    "received_bytes": 1679616,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 58,
    "mode": "both",
    "label": "op Some(22) LayoutRearrangement: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 1679616,
    "received_bytes": 1679616,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 2304,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  }
]