[
  {
    "phase": 50,
    "mode": "original",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 62482,
    "receive_fragments": 2846614,
    "maximum_endpoint_fragments": 2109,
    "transmitted_bytes": 11588864,
    "received_bytes": 135315968,
    "maximum_outgoing_bytes": 36608,
    "maximum_incoming_bytes": 93696,
    "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": 50,
    "mode": "source",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 62320,
    "receive_fragments": 2845318,
    "maximum_endpoint_fragments": 2108,
    "transmitted_bytes": 11588864,
    "received_bytes": 135315968,
    "maximum_outgoing_bytes": 36608,
    "maximum_incoming_bytes": 93696,
    "source_staging_max_bytes": 36608,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 36608,
    "source_copy_max_bytes": 36608,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2472,
    "affine_copy_loops": 2471,
    "copy_cycles_floor": 4576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "destination",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 5441,
    "receive_fragments": 122486,
    "maximum_endpoint_fragments": 113,
    "transmitted_bytes": 11588864,
    "received_bytes": 135315968,
    "maximum_outgoing_bytes": 36608,
    "maximum_incoming_bytes": 93696,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 93696,
    "combined_staging_max_bytes": 93696,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 93696,
    "local_copy_runs": 2843664,
    "affine_copy_loops": 122021,
    "copy_cycles_floor": 11712,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 50,
    "mode": "both",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 5441,
    "receive_fragments": 122486,
    "maximum_endpoint_fragments": 113,
    "transmitted_bytes": 11588864,
    "received_bytes": 135315968,
    "maximum_outgoing_bytes": 36608,
    "maximum_incoming_bytes": 93696,
    "source_staging_max_bytes": 36608,
    "destination_staging_max_bytes": 93696,
    "combined_staging_max_bytes": 130304,
    "source_copy_max_bytes": 36608,
    "destination_copy_max_bytes": 93696,
    "local_copy_runs": 2846136,
    "affine_copy_loops": 124492,
    "copy_cycles_floor": 16288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "original",
    "label": "op Some(21) OperatorInputs: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 303042,
    "receive_fragments": 305250,
    "maximum_endpoint_fragments": 1913,
    "transmitted_bytes": 10039296,
    "received_bytes": 40562688,
    "maximum_outgoing_bytes": 31232,
    "maximum_incoming_bytes": 27648,
    "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": 55,
    "mode": "source",
    "label": "op Some(21) OperatorInputs: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 4142,
    "receive_fragments": 5246,
    "maximum_endpoint_fragments": 162,
    "transmitted_bytes": 10039296,
    "received_bytes": 40562688,
    "maximum_outgoing_bytes": 31232,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 31232,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 31232,
    "source_copy_max_bytes": 31232,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2704,
    "affine_copy_loops": 1696,
    "copy_cycles_floor": 3904,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "destination",
    "label": "op Some(21) OperatorInputs: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 4142,
    "receive_fragments": 5246,
    "maximum_endpoint_fragments": 162,
    "transmitted_bytes": 10039296,
    "received_bytes": 40562688,
    "maximum_outgoing_bytes": 31232,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 27648,
    "combined_staging_max_bytes": 27648,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 27648,
    "local_copy_runs": 2617,
    "affine_copy_loops": 1609,
    "copy_cycles_floor": 3456,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 55,
    "mode": "both",
    "label": "op Some(21) OperatorInputs: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 192, column_block: 16 })",
    "transfers": 3062,
    "receive_fragments": 4166,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 10039296,
    "received_bytes": 40562688,
    "maximum_outgoing_bytes": 31232,
    "maximum_incoming_bytes": 27648,
    "source_staging_max_bytes": 31232,
    "destination_staging_max_bytes": 27648,
    "combined_staging_max_bytes": 58880,
    "source_copy_max_bytes": 31232,
    "destination_copy_max_bytes": 27648,
    "local_copy_runs": 5321,
    "affine_copy_loops": 3305,
    "copy_cycles_floor": 7360,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 122,
    "mode": "original",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 1471,
    "receive_fragments": 1471,
    "maximum_endpoint_fragments": 1471,
    "transmitted_bytes": 17200,
    "received_bytes": 17200,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 17200,
    "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": 122,
    "mode": "source",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 1471,
    "receive_fragments": 1471,
    "maximum_endpoint_fragments": 1471,
    "transmitted_bytes": 17200,
    "received_bytes": 17200,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 17200,
    "source_staging_max_bytes": 16,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 16,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1471,
    "affine_copy_loops": 1471,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 122,
    "mode": "destination",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 1471,
    "receive_fragments": 1471,
    "maximum_endpoint_fragments": 1471,
    "transmitted_bytes": 17200,
    "received_bytes": 17200,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 17200,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 17200,
    "combined_staging_max_bytes": 17200,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 17200,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 2150,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 122,
    "mode": "both",
    "label": "op Some(41) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 1471,
    "receive_fragments": 1471,
    "maximum_endpoint_fragments": 1471,
    "transmitted_bytes": 17200,
    "received_bytes": 17200,
    "maximum_outgoing_bytes": 16,
    "maximum_incoming_bytes": 17200,
    "source_staging_max_bytes": 16,
    "destination_staging_max_bytes": 17200,
    "combined_staging_max_bytes": 17200,
    "source_copy_max_bytes": 16,
    "destination_copy_max_bytes": 17200,
    "local_copy_runs": 1472,
    "affine_copy_loops": 1472,
    "copy_cycles_floor": 2152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "original",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54502,
    "receive_fragments": 1893886,
    "maximum_endpoint_fragments": 1420,
    "transmitted_bytes": 9637632,
    "received_bytes": 100274688,
    "maximum_outgoing_bytes": 33024,
    "maximum_incoming_bytes": 70272,
    "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": 16,
    "mode": "source",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54358,
    "receive_fragments": 1892590,
    "maximum_endpoint_fragments": 1419,
    "transmitted_bytes": 9637632,
    "received_bytes": 100274688,
    "maximum_outgoing_bytes": 33024,
    "maximum_incoming_bytes": 70272,
    "source_staging_max_bytes": 33024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 33024,
    "source_copy_max_bytes": 33024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2300,
    "affine_copy_loops": 2300,
    "copy_cycles_floor": 4128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "destination",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 4786,
    "receive_fragments": 107998,
    "maximum_endpoint_fragments": 93,
    "transmitted_bytes": 9637632,
    "received_bytes": 100274688,
    "maximum_outgoing_bytes": 33024,
    "maximum_incoming_bytes": 70272,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 70272,
    "combined_staging_max_bytes": 70272,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 70272,
    "local_copy_runs": 1890013,
    "affine_copy_loops": 106503,
    "copy_cycles_floor": 8784,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 16,
    "mode": "both",
    "label": "op Some(4) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 4786,
    "receive_fragments": 107998,
    "maximum_endpoint_fragments": 93,
    "transmitted_bytes": 9637632,
    "received_bytes": 100274688,
    "maximum_outgoing_bytes": 33024,
    "maximum_incoming_bytes": 70272,
    "source_staging_max_bytes": 33024,
    "destination_staging_max_bytes": 70272,
    "combined_staging_max_bytes": 103296,
    "source_copy_max_bytes": 33024,
    "destination_copy_max_bytes": 70272,
    "local_copy_runs": 1892313,
    "affine_copy_loops": 108803,
    "copy_cycles_floor": 12912,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "original",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 729, 4304]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 4304]) Amp(Left)",
    "transfers": 196830,
    "receive_fragments": 1574640,
    "maximum_endpoint_fragments": 1374,
    "transmitted_bytes": 6275232,
    "received_bytes": 50201856,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 35328,
    "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": 54,
    "mode": "source",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 729, 4304]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 4304]) Amp(Left)",
    "transfers": 196830,
    "receive_fragments": 1574640,
    "maximum_endpoint_fragments": 1374,
    "transmitted_bytes": 6275232,
    "received_bytes": 50201856,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 35328,
    "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": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 1076,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "destination",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 729, 4304]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 4304]) Amp(Left)",
    "transfers": 17496,
    "receive_fragments": 139968,
    "maximum_endpoint_fragments": 208,
    "transmitted_bytes": 6275232,
    "received_bytes": 50201856,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 35328,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 35328,
    "combined_staging_max_bytes": 35328,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 35328,
    "local_copy_runs": 1574640,
    "affine_copy_loops": 139968,
    "copy_cycles_floor": 4416,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 54,
    "mode": "both",
    "label": "op Some(21) LayoutRearrangement: TensorShape([2, 729, 4304]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 4304]) Amp(Left)",
    "transfers": 16767,
    "receive_fragments": 134136,
    "maximum_endpoint_fragments": 115,
    "transmitted_bytes": 6275232,
    "received_bytes": 50201856,
    "maximum_outgoing_bytes": 8608,
    "maximum_incoming_bytes": 35328,
    "source_staging_max_bytes": 8608,
    "destination_staging_max_bytes": 35328,
    "combined_staging_max_bytes": 43936,
    "source_copy_max_bytes": 8608,
    "destination_copy_max_bytes": 35328,
    "local_copy_runs": 1576098,
    "affine_copy_loops": 141426,
    "copy_cycles_floor": 5492,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "original",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([25344]) RowMajor",
    "transfers": 586368,
    "receive_fragments": 586368,
    "maximum_endpoint_fragments": 1206,
    "transmitted_bytes": 74032128,
    "received_bytes": 74032128,
    "maximum_outgoing_bytes": 51840,
    "maximum_incoming_bytes": 52992,
    "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": 56,
    "mode": "source",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([25344]) RowMajor",
    "transfers": 32576,
    "receive_fragments": 32576,
    "maximum_endpoint_fragments": 67,
    "transmitted_bytes": 74032128,
    "received_bytes": 74032128,
    "maximum_outgoing_bytes": 51840,
    "maximum_incoming_bytes": 52992,
    "source_staging_max_bytes": 51840,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 51840,
    "source_copy_max_bytes": 51840,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 586368,
    "affine_copy_loops": 32576,
    "copy_cycles_floor": 6480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "destination",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([25344]) RowMajor",
    "transfers": 586368,
    "receive_fragments": 586368,
    "maximum_endpoint_fragments": 1206,
    "transmitted_bytes": 74032128,
    "received_bytes": 74032128,
    "maximum_outgoing_bytes": 51840,
    "maximum_incoming_bytes": 52992,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 52992,
    "combined_staging_max_bytes": 52992,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 52992,
    "local_copy_runs": 1488,
    "affine_copy_loops": 1488,
    "copy_cycles_floor": 6624,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 56,
    "mode": "both",
    "label": "op Some(21) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([25344]) RowMajor",
    "transfers": 32576,
    "receive_fragments": 32576,
    "maximum_endpoint_fragments": 67,
    "transmitted_bytes": 74032128,
    "received_bytes": 74032128,
    "maximum_outgoing_bytes": 51840,
    "maximum_incoming_bytes": 52992,
    "source_staging_max_bytes": 51840,
    "destination_staging_max_bytes": 52992,
    "combined_staging_max_bytes": 104832,
    "source_copy_max_bytes": 51840,
    "destination_copy_max_bytes": 52992,
    "local_copy_runs": 587856,
    "affine_copy_loops": 34064,
    "copy_cycles_floor": 13104,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 139808,
    "receive_fragments": 141247,
    "maximum_endpoint_fragments": 1013,
    "transmitted_bytes": 6713856,
    "received_bytes": 6921072,
    "maximum_outgoing_bytes": 13648,
    "maximum_incoming_bytes": 32656,
    "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": 70,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 139808,
    "receive_fragments": 141247,
    "maximum_endpoint_fragments": 1013,
    "transmitted_bytes": 6713856,
    "received_bytes": 6921072,
    "maximum_outgoing_bytes": 13648,
    "maximum_incoming_bytes": 32656,
    "source_staging_max_bytes": 13648,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13648,
    "source_copy_max_bytes": 13648,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 127129,
    "affine_copy_loops": 24271,
    "copy_cycles_floor": 1706,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 139808,
    "receive_fragments": 141247,
    "maximum_endpoint_fragments": 1013,
    "transmitted_bytes": 6713856,
    "received_bytes": 6921072,
    "maximum_outgoing_bytes": 13648,
    "maximum_incoming_bytes": 32656,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 32656,
    "combined_staging_max_bytes": 32656,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 32656,
    "local_copy_runs": 56415,
    "affine_copy_loops": 22335,
    "copy_cycles_floor": 4082,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 70,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 129600,
    "receive_fragments": 131039,
    "maximum_endpoint_fragments": 919,
    "transmitted_bytes": 6713856,
    "received_bytes": 6921072,
    "maximum_outgoing_bytes": 13648,
    "maximum_incoming_bytes": 32656,
    "source_staging_max_bytes": 13648,
    "destination_staging_max_bytes": 32656,
    "combined_staging_max_bytes": 46016,
    "source_copy_max_bytes": 13648,
    "destination_copy_max_bytes": 32656,
    "local_copy_runs": 183544,
    "affine_copy_loops": 46606,
    "copy_cycles_floor": 5788,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54501,
    "receive_fragments": 1264029,
    "maximum_endpoint_fragments": 976,
    "transmitted_bytes": 6982912,
    "received_bytes": 66847744,
    "maximum_outgoing_bytes": 22784,
    "maximum_incoming_bytes": 46848,
    "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": 61,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54357,
    "receive_fragments": 1262733,
    "maximum_endpoint_fragments": 975,
    "transmitted_bytes": 6982912,
    "received_bytes": 66847744,
    "maximum_outgoing_bytes": 22784,
    "maximum_incoming_bytes": 46848,
    "source_staging_max_bytes": 22784,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 22784,
    "source_copy_max_bytes": 22784,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2320,
    "affine_copy_loops": 2320,
    "copy_cycles_floor": 2848,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 6243,
    "receive_fragments": 107997,
    "maximum_endpoint_fragments": 95,
    "transmitted_bytes": 6982912,
    "received_bytes": 66847744,
    "maximum_outgoing_bytes": 22784,
    "maximum_incoming_bytes": 46848,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 46848,
    "combined_staging_max_bytes": 46848,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 46848,
    "local_copy_runs": 1261238,
    "affine_copy_loops": 106506,
    "copy_cycles_floor": 5856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 61,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 6243,
    "receive_fragments": 107997,
    "maximum_endpoint_fragments": 95,
    "transmitted_bytes": 6982912,
    "received_bytes": 66847744,
    "maximum_outgoing_bytes": 22784,
    "maximum_incoming_bytes": 46848,
    "source_staging_max_bytes": 22784,
    "destination_staging_max_bytes": 46848,
    "combined_staging_max_bytes": 69632,
    "source_copy_max_bytes": 22784,
    "destination_copy_max_bytes": 46848,
    "local_copy_runs": 1263558,
    "affine_copy_loops": 108826,
    "copy_cycles_floor": 8704,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "original",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 391962,
    "receive_fragments": 441500,
    "maximum_endpoint_fragments": 583,
    "transmitted_bytes": 12550304,
    "received_bytes": 25092160,
    "maximum_outgoing_bytes": 8960,
    "maximum_incoming_bytes": 17216,
    "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": 52,
    "mode": "source",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 78711,
    "receive_fragments": 128249,
    "maximum_endpoint_fragments": 144,
    "transmitted_bytes": 12550304,
    "received_bytes": 25092160,
    "maximum_outgoing_bytes": 8960,
    "maximum_incoming_bytes": 17216,
    "source_staging_max_bytes": 8960,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8960,
    "source_copy_max_bytes": 8960,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 391962,
    "affine_copy_loops": 78711,
    "copy_cycles_floor": 1120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "destination",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 391962,
    "receive_fragments": 441500,
    "maximum_endpoint_fragments": 583,
    "transmitted_bytes": 12550304,
    "received_bytes": 25092160,
    "maximum_outgoing_bytes": 8960,
    "maximum_incoming_bytes": 17216,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 17216,
    "combined_staging_max_bytes": 17216,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 17216,
    "local_copy_runs": 5356,
    "affine_copy_loops": 5071,
    "copy_cycles_floor": 2152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 52,
    "mode": "both",
    "label": "op Some(19) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([2, 729, 4304]) RowMajor",
    "transfers": 78711,
    "receive_fragments": 128249,
    "maximum_endpoint_fragments": 144,
    "transmitted_bytes": 12550304,
    "received_bytes": 25092160,
    "maximum_outgoing_bytes": 8960,
    "maximum_incoming_bytes": 17216,
    "source_staging_max_bytes": 8960,
    "destination_staging_max_bytes": 17216,
    "combined_staging_max_bytes": 26176,
    "source_copy_max_bytes": 8960,
    "destination_copy_max_bytes": 17216,
    "local_copy_runs": 397318,
    "affine_copy_loops": 83782,
    "copy_cycles_floor": 3272,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "original",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "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": 69,
    "mode": "source",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 575,
    "affine_copy_loops": 575,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "destination",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 575,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 69,
    "mode": "both",
    "label": "op None OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "original",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "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": 80,
    "mode": "source",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 575,
    "affine_copy_loops": 575,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "destination",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 575,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 80,
    "mode": "both",
    "label": "op Some(38) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 127,
    "mode": "original",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "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": 127,
    "mode": "source",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 575,
    "affine_copy_loops": 575,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 127,
    "mode": "destination",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 575,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 127,
    "mode": "both",
    "label": "op Some(44) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 575,
    "receive_fragments": 575,
    "maximum_endpoint_fragments": 575,
    "transmitted_bytes": 4600,
    "received_bytes": 4600,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 4600,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 4600,
    "combined_staging_max_bytes": 4600,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 4600,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "original",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 314748,
    "receive_fragments": 354087,
    "maximum_endpoint_fragments": 539,
    "transmitted_bytes": 10077984,
    "received_bytes": 20148768,
    "maximum_outgoing_bytes": 9472,
    "maximum_incoming_bytes": 13824,
    "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": 18,
    "mode": "source",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 209848,
    "receive_fragments": 249187,
    "maximum_endpoint_fragments": 319,
    "transmitted_bytes": 10077984,
    "received_bytes": 20148768,
    "maximum_outgoing_bytes": 9472,
    "maximum_incoming_bytes": 13824,
    "source_staging_max_bytes": 9472,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 9472,
    "source_copy_max_bytes": 9472,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 210673,
    "affine_copy_loops": 105753,
    "copy_cycles_floor": 1184,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "destination",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 314748,
    "receive_fragments": 354087,
    "maximum_endpoint_fragments": 539,
    "transmitted_bytes": 10077984,
    "received_bytes": 20148768,
    "maximum_outgoing_bytes": 9472,
    "maximum_incoming_bytes": 13824,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 13824,
    "combined_staging_max_bytes": 13824,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 13824,
    "local_copy_runs": 9648,
    "affine_copy_loops": 9415,
    "copy_cycles_floor": 1728,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 18,
    "mode": "both",
    "label": "op Some(5) OperatorInputs: TensorShape([2, 729, 3456]) F16 Amp(Left) -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 209848,
    "receive_fragments": 249187,
    "maximum_endpoint_fragments": 319,
    "transmitted_bytes": 10077984,
    "received_bytes": 20148768,
    "maximum_outgoing_bytes": 9472,
    "maximum_incoming_bytes": 13824,
    "source_staging_max_bytes": 9472,
    "destination_staging_max_bytes": 13824,
    "combined_staging_max_bytes": 23296,
    "source_copy_max_bytes": 9472,
    "destination_copy_max_bytes": 13824,
    "local_copy_runs": 220321,
    "affine_copy_loops": 115168,
    "copy_cycles_floor": 2912,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54359,
    "receive_fragments": 632879,
    "maximum_endpoint_fragments": 531,
    "transmitted_bytes": 4332800,
    "received_bytes": 33425408,
    "maximum_outgoing_bytes": 12544,
    "maximum_incoming_bytes": 23424,
    "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": 34,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 54359,
    "receive_fragments": 632879,
    "maximum_endpoint_fragments": 531,
    "transmitted_bytes": 4332800,
    "received_bytes": 33425408,
    "maximum_outgoing_bytes": 12544,
    "maximum_incoming_bytes": 23424,
    "source_staging_max_bytes": 12544,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 12544,
    "source_copy_max_bytes": 12544,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2340,
    "affine_copy_loops": 2340,
    "copy_cycles_floor": 1568,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 10619,
    "receive_fragments": 107999,
    "maximum_endpoint_fragments": 101,
    "transmitted_bytes": 4332800,
    "received_bytes": 33425408,
    "maximum_outgoing_bytes": 12544,
    "maximum_incoming_bytes": 23424,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 23424,
    "combined_staging_max_bytes": 23424,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 23424,
    "local_copy_runs": 631394,
    "affine_copy_loops": 106514,
    "copy_cycles_floor": 2928,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 34,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 1152]) Amp(Left)",
    "transfers": 10619,
    "receive_fragments": 107999,
    "maximum_endpoint_fragments": 101,
    "transmitted_bytes": 4332800,
    "received_bytes": 33425408,
    "maximum_outgoing_bytes": 12544,
    "maximum_incoming_bytes": 23424,
    "source_staging_max_bytes": 12544,
    "destination_staging_max_bytes": 23424,
    "combined_staging_max_bytes": 35968,
    "source_copy_max_bytes": 12544,
    "destination_copy_max_bytes": 23424,
    "local_copy_runs": 633734,
    "affine_copy_loops": 108854,
    "copy_cycles_floor": 4496,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "original",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 729, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 588]) Amp(Left)",
    "transfers": 27702,
    "receive_fragments": 664848,
    "maximum_endpoint_fragments": 526,
    "transmitted_bytes": 857304,
    "received_bytes": 20575296,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 15616,
    "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": 0,
    "mode": "source",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 729, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 588]) Amp(Left)",
    "transfers": 27702,
    "receive_fragments": 664848,
    "maximum_endpoint_fragments": 526,
    "transmitted_bytes": 857304,
    "received_bytes": 20575296,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 15616,
    "source_staging_max_bytes": 1176,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1176,
    "source_copy_max_bytes": 1176,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 147,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "destination",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 729, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 588]) Amp(Left)",
    "transfers": 8019,
    "receive_fragments": 192456,
    "maximum_endpoint_fragments": 255,
    "transmitted_bytes": 857304,
    "received_bytes": 20575296,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 15616,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 15616,
    "combined_staging_max_bytes": 15616,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 15616,
    "local_copy_runs": 664848,
    "affine_copy_loops": 174960,
    "copy_cycles_floor": 1952,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 0,
    "mode": "both",
    "label": "op Some(0) LayoutRearrangement: TensorShape([2, 729, 588]) F8F143 { scale_exponent: -4 } Amp(Left) -> TensorShape([2, 729, 588]) Amp(Left)",
    "transfers": 7290,
    "receive_fragments": 174960,
    "maximum_endpoint_fragments": 132,
    "transmitted_bytes": 857304,
    "received_bytes": 20575296,
    "maximum_outgoing_bytes": 1176,
    "maximum_incoming_bytes": 15616,
    "source_staging_max_bytes": 1176,
    "destination_staging_max_bytes": 15616,
    "combined_staging_max_bytes": 16792,
    "source_copy_max_bytes": 1176,
    "destination_copy_max_bytes": 15616,
    "local_copy_runs": 666306,
    "affine_copy_loops": 176418,
    "copy_cycles_floor": 2099,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 124,
    "mode": "original",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 224, column_block: 16 })",
    "transfers": 301599,
    "receive_fragments": 301599,
    "maximum_endpoint_fragments": 437,
    "transmitted_bytes": 4953344,
    "received_bytes": 4953344,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "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": 124,
    "mode": "source",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 224, column_block: 16 })",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 4953344,
    "received_bytes": 4953344,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2504,
    "affine_copy_loops": 1510,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 124,
    "mode": "destination",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 224, column_block: 16 })",
    "transfers": 2790,
    "receive_fragments": 2790,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 4953344,
    "received_bytes": 4953344,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3584,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3584,
    "local_copy_runs": 2503,
    "affine_copy_loops": 1509,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 124,
    "mode": "both",
    "label": "op Some(43) LayoutRearrangement: TensorShape([4304, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([4304, 1152]) BlockMajor(Matrix { row_block: 224, column_block: 16 })",
    "transfers": 1725,
    "receive_fragments": 1725,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 4953344,
    "received_bytes": 4953344,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 3584,
    "combined_staging_max_bytes": 7168,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 3584,
    "local_copy_runs": 5007,
    "affine_copy_loops": 3019,
    "copy_cycles_floor": 896,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 163088,
    "receive_fragments": 163088,
    "maximum_endpoint_fragments": 368,
    "transmitted_bytes": 10063104,
    "received_bytes": 10063104,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 10496,
    "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": 20,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 163088,
    "receive_fragments": 163088,
    "maximum_endpoint_fragments": 368,
    "transmitted_bytes": 10063104,
    "received_bytes": 10063104,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 10496,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13824,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 143465,
    "affine_copy_loops": 73577,
    "copy_cycles_floor": 1728,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 163088,
    "receive_fragments": 163088,
    "maximum_endpoint_fragments": 368,
    "transmitted_bytes": 10063104,
    "received_bytes": 10063104,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 10496,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 10496,
    "combined_staging_max_bytes": 10496,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 10496,
    "local_copy_runs": 163088,
    "affine_copy_loops": 74224,
    "copy_cycles_floor": 1312,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 20,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([32, 729, 80]) RowMajor",
    "transfers": 90144,
    "receive_fragments": 90144,
    "maximum_endpoint_fragments": 208,
    "transmitted_bytes": 10063104,
    "received_bytes": 10063104,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 10496,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 10496,
    "combined_staging_max_bytes": 24320,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 10496,
    "local_copy_runs": 306553,
    "affine_copy_loops": 147801,
    "copy_cycles_floor": 3040,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "original",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 209839,
    "receive_fragments": 236065,
    "maximum_endpoint_fragments": 310,
    "transmitted_bytes": 6718880,
    "received_bytes": 13432736,
    "maximum_outgoing_bytes": 4864,
    "maximum_incoming_bytes": 9216,
    "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": 63,
    "mode": "source",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 209839,
    "receive_fragments": 236065,
    "maximum_endpoint_fragments": 310,
    "transmitted_bytes": 6718880,
    "received_bytes": 13432736,
    "maximum_outgoing_bytes": 4864,
    "maximum_incoming_bytes": 9216,
    "source_staging_max_bytes": 4864,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4864,
    "source_copy_max_bytes": 4864,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1731,
    "affine_copy_loops": 1639,
    "copy_cycles_floor": 608,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "destination",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 209839,
    "receive_fragments": 236065,
    "maximum_endpoint_fragments": 310,
    "transmitted_bytes": 6718880,
    "received_bytes": 13432736,
    "maximum_outgoing_bytes": 4864,
    "maximum_incoming_bytes": 9216,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 9216,
    "combined_staging_max_bytes": 9216,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 9216,
    "local_copy_runs": 7056,
    "affine_copy_loops": 6823,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 63,
    "mode": "both",
    "label": "op Some(27) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 209839,
    "receive_fragments": 236065,
    "maximum_endpoint_fragments": 310,
    "transmitted_bytes": 6718880,
    "received_bytes": 13432736,
    "maximum_outgoing_bytes": 4864,
    "maximum_incoming_bytes": 9216,
    "source_staging_max_bytes": 4864,
    "destination_staging_max_bytes": 9216,
    "combined_staging_max_bytes": 14080,
    "source_copy_max_bytes": 4864,
    "destination_copy_max_bytes": 9216,
    "local_copy_runs": 8787,
    "affine_copy_loops": 8462,
    "copy_cycles_floor": 1760,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 42800,
    "receive_fragments": 43760,
    "maximum_endpoint_fragments": 225,
    "transmitted_bytes": 1418240,
    "received_bytes": 4367360,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "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": 1,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 2686,
    "receive_fragments": 3646,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 1418240,
    "received_bytes": 4367360,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2686,
    "affine_copy_loops": 1606,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 2686,
    "receive_fragments": 3646,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 1418240,
    "received_bytes": 4367360,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 3072,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 2569,
    "affine_copy_loops": 1489,
    "copy_cycles_floor": 384,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 1,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([588, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([588, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1606,
    "receive_fragments": 2566,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 1418240,
    "received_bytes": 4367360,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3072,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 3072,
    "combined_staging_max_bytes": 6656,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 3072,
    "local_copy_runs": 5255,
    "affine_copy_loops": 3095,
    "copy_cycles_floor": 832,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "original",
    "label": "op Some(1) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104931,
    "receive_fragments": 118044,
    "maximum_endpoint_fragments": 165,
    "transmitted_bytes": 3359808,
    "received_bytes": 6716736,
    "maximum_outgoing_bytes": 2688,
    "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": 11,
    "mode": "source",
    "label": "op Some(1) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 34983,
    "receive_fragments": 48096,
    "maximum_endpoint_fragments": 61,
    "transmitted_bytes": 3359808,
    "received_bytes": 6716736,
    "maximum_outgoing_bytes": 2688,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 2688,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 2688,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 104931,
    "affine_copy_loops": 34983,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "destination",
    "label": "op Some(1) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104931,
    "receive_fragments": 118044,
    "maximum_endpoint_fragments": 165,
    "transmitted_bytes": 3359808,
    "received_bytes": 6716736,
    "maximum_outgoing_bytes": 2688,
    "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": 2288,
    "affine_copy_loops": 1906,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 11,
    "mode": "both",
    "label": "op Some(1) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 34983,
    "receive_fragments": 48096,
    "maximum_endpoint_fragments": 61,
    "transmitted_bytes": 3359808,
    "received_bytes": 6716736,
    "maximum_outgoing_bytes": 2688,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 2688,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 7296,
    "source_copy_max_bytes": 2688,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 107219,
    "affine_copy_loops": 36889,
    "copy_cycles_floor": 912,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "original",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104926,
    "receive_fragments": 118039,
    "maximum_endpoint_fragments": 157,
    "transmitted_bytes": 3359648,
    "received_bytes": 6716576,
    "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": 46,
    "mode": "source",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104926,
    "receive_fragments": 118039,
    "maximum_endpoint_fragments": 157,
    "transmitted_bytes": 3359648,
    "received_bytes": 6716576,
    "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": 1578,
    "affine_copy_loops": 1523,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "destination",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104926,
    "receive_fragments": 118039,
    "maximum_endpoint_fragments": 157,
    "transmitted_bytes": 3359648,
    "received_bytes": 6716576,
    "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": 3168,
    "affine_copy_loops": 2827,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 46,
    "mode": "both",
    "label": "op Some(15) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 104926,
    "receive_fragments": 118039,
    "maximum_endpoint_fragments": 157,
    "transmitted_bytes": 3359648,
    "received_bytes": 6716576,
    "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": 4746,
    "affine_copy_loops": 4350,
    "copy_cycles_floor": 896,
    "replay_address_ranges_valid": true
  },
  {
    "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": 51,
    "mode": "original",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([8960]) RowMajor",
    "transfers": 41964,
    "receive_fragments": 41964,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 25824000,
    "received_bytes": 25824000,
    "maximum_outgoing_bytes": 20480,
    "maximum_incoming_bytes": 20480,
    "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": 51,
    "mode": "source",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([8960]) RowMajor",
    "transfers": 4212,
    "receive_fragments": 4212,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 25824000,
    "received_bytes": 25824000,
    "maximum_outgoing_bytes": 20480,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 20480,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 20480,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 41964,
    "affine_copy_loops": 4212,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "destination",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([8960]) RowMajor",
    "transfers": 41964,
    "receive_fragments": 41964,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 25824000,
    "received_bytes": 25824000,
    "maximum_outgoing_bytes": 20480,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 2322,
    "affine_copy_loops": 2322,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 51,
    "mode": "both",
    "label": "op Some(18) OperatorInputs: TensorShape([2, 729, 4304]) F16 Amp(Left) -> TensorShape([8960]) RowMajor",
    "transfers": 4212,
    "receive_fragments": 4212,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 25824000,
    "received_bytes": 25824000,
    "maximum_outgoing_bytes": 20480,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 20480,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 40320,
    "source_copy_max_bytes": 20480,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 44286,
    "affine_copy_loops": 6534,
    "copy_cycles_floor": 5120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 729, 72]) Amp(TransposedRight)",
    "transfers": 23040,
    "receive_fragments": 23040,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 3317760,
    "received_bytes": 3317760,
    "maximum_outgoing_bytes": 9216,
    "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": 71,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 729, 72]) Amp(TransposedRight)",
    "transfers": 23040,
    "receive_fragments": 23040,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 3317760,
    "received_bytes": 3317760,
    "maximum_outgoing_bytes": 9216,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 9216,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 9216,
    "source_copy_max_bytes": 9216,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 23040,
    "affine_copy_loops": 2880,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 729, 72]) Amp(TransposedRight)",
    "transfers": 23040,
    "receive_fragments": 23040,
    "maximum_endpoint_fragments": 80,
    "transmitted_bytes": 3317760,
    "received_bytes": 3317760,
    "maximum_outgoing_bytes": 9216,
    "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": 23040,
    "affine_copy_loops": 2880,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 71,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 729, 72]) Amp(TransposedRight)",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 3317760,
    "received_bytes": 3317760,
    "maximum_outgoing_bytes": 9216,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 9216,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 11520,
    "source_copy_max_bytes": 9216,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 46080,
    "affine_copy_loops": 5760,
    "copy_cycles_floor": 1440,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([384]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 22656,
    "maximum_incoming_bytes": 2688,
    "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": 10,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([384]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 22656,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 22656,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 22656,
    "source_copy_max_bytes": 22656,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 2832,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([384]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 22656,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 10,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([384]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 22656,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 22656,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 24960,
    "source_copy_max_bytes": 22656,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 3168,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "original",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 72,
    "receive_fragments": 2878,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 4608,
    "received_bytes": 184192,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 128,
    "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": 84,
    "mode": "source",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 36,
    "receive_fragments": 1439,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 4608,
    "received_bytes": 184192,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 128,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 72,
    "affine_copy_loops": 36,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "destination",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 72,
    "receive_fragments": 2878,
    "maximum_endpoint_fragments": 72,
    "transmitted_bytes": 4608,
    "received_bytes": 184192,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 128,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 128,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 128,
    "local_copy_runs": 1439,
    "affine_copy_loops": 1439,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 84,
    "mode": "both",
    "label": "op Some(40) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 36,
    "receive_fragments": 1439,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 4608,
    "received_bytes": 184192,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 128,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 128,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 128,
    "local_copy_runs": 1511,
    "affine_copy_loops": 1475,
    "copy_cycles_floor": 592,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9792,
    "receive_fragments": 9792,
    "maximum_endpoint_fragments": 66,
    "transmitted_bytes": 3456000,
    "received_bytes": 3456000,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 3840,
    "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": 2,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1632,
    "receive_fragments": 1632,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 3456000,
    "received_bytes": 3456000,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 21120,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 21120,
    "source_copy_max_bytes": 21120,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9792,
    "affine_copy_loops": 1632,
    "copy_cycles_floor": 2640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9792,
    "receive_fragments": 9792,
    "maximum_endpoint_fragments": 66,
    "transmitted_bytes": 3456000,
    "received_bytes": 3456000,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 3840,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 1464,
    "affine_copy_loops": 1464,
    "copy_cycles_floor": 480,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 2,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1632,
    "receive_fragments": 1632,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 3456000,
    "received_bytes": 3456000,
    "maximum_outgoing_bytes": 21120,
    "maximum_incoming_bytes": 3840,
    "source_staging_max_bytes": 21120,
    "destination_staging_max_bytes": 3840,
    "combined_staging_max_bytes": 23424,
    "source_copy_max_bytes": 21120,
    "destination_copy_max_bytes": 3840,
    "local_copy_runs": 11256,
    "affine_copy_loops": 3096,
    "copy_cycles_floor": 3120,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "original",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 143,
    "receive_fragments": 223,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 9144,
    "received_bytes": 13752,
    "maximum_outgoing_bytes": 4536,
    "maximum_incoming_bytes": 216,
    "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": 81,
    "mode": "source",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 143,
    "receive_fragments": 223,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 9144,
    "received_bytes": 13752,
    "maximum_outgoing_bytes": 4536,
    "maximum_incoming_bytes": 216,
    "source_staging_max_bytes": 4536,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4536,
    "source_copy_max_bytes": 4536,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 19,
    "affine_copy_loops": 19,
    "copy_cycles_floor": 567,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "destination",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 143,
    "receive_fragments": 223,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 9144,
    "received_bytes": 13752,
    "maximum_outgoing_bytes": 4536,
    "maximum_incoming_bytes": 216,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 216,
    "combined_staging_max_bytes": 216,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 216,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 27,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 81,
    "mode": "both",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 143,
    "receive_fragments": 223,
    "maximum_endpoint_fragments": 65,
    "transmitted_bytes": 9144,
    "received_bytes": 13752,
    "maximum_outgoing_bytes": 4536,
    "maximum_incoming_bytes": 216,
    "source_staging_max_bytes": 4536,
    "destination_staging_max_bytes": 216,
    "combined_staging_max_bytes": 4680,
    "source_copy_max_bytes": 4536,
    "destination_copy_max_bytes": 216,
    "local_copy_runs": 83,
    "affine_copy_loops": 83,
    "copy_cycles_floor": 594,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "original",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4536,
    "received_bytes": 4536,
    "maximum_outgoing_bytes": 72,
    "maximum_incoming_bytes": 4536,
    "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": 83,
    "mode": "source",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4536,
    "received_bytes": 4536,
    "maximum_outgoing_bytes": 72,
    "maximum_incoming_bytes": 4536,
    "source_staging_max_bytes": 72,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 72,
    "source_copy_max_bytes": 72,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 63,
    "affine_copy_loops": 63,
    "copy_cycles_floor": 9,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "destination",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4536,
    "received_bytes": 4536,
    "maximum_outgoing_bytes": 72,
    "maximum_incoming_bytes": 4536,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4536,
    "combined_staging_max_bytes": 4536,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4536,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 567,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 83,
    "mode": "both",
    "label": "op Some(39) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 63,
    "receive_fragments": 63,
    "maximum_endpoint_fragments": 63,
    "transmitted_bytes": 4536,
    "received_bytes": 4536,
    "maximum_outgoing_bytes": 72,
    "maximum_incoming_bytes": 4536,
    "source_staging_max_bytes": 72,
    "destination_staging_max_bytes": 4536,
    "combined_staging_max_bytes": 4536,
    "source_copy_max_bytes": 72,
    "destination_copy_max_bytes": 4536,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 60,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 20352,
    "maximum_incoming_bytes": 2688,
    "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": 9,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 20352,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 20352,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 20352,
    "source_copy_max_bytes": 20352,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 2544,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 60,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 20352,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 9,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 20352,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 20352,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 22656,
    "source_copy_max_bytes": 20352,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2880,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 121,
    "mode": "original",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 2207,
    "receive_fragments": 2942,
    "maximum_endpoint_fragments": 58,
    "transmitted_bytes": 25808,
    "received_bytes": 34408,
    "maximum_outgoing_bytes": 448,
    "maximum_incoming_bytes": 32,
    "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": 121,
    "mode": "source",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 2207,
    "receive_fragments": 2942,
    "maximum_endpoint_fragments": 58,
    "transmitted_bytes": 25808,
    "received_bytes": 34408,
    "maximum_outgoing_bytes": 448,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 448,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 448,
    "source_copy_max_bytes": 448,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 572,
    "affine_copy_loops": 115,
    "copy_cycles_floor": 56,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 121,
    "mode": "destination",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 2207,
    "receive_fragments": 2942,
    "maximum_endpoint_fragments": 58,
    "transmitted_bytes": 25808,
    "received_bytes": 34408,
    "maximum_outgoing_bytes": 448,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 32,
    "combined_staging_max_bytes": 32,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 32,
    "local_copy_runs": 1472,
    "affine_copy_loops": 1472,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 121,
    "mode": "both",
    "label": "op Some(41) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 2207,
    "receive_fragments": 2942,
    "maximum_endpoint_fragments": 58,
    "transmitted_bytes": 25808,
    "received_bytes": 34408,
    "maximum_outgoing_bytes": 448,
    "maximum_incoming_bytes": 32,
    "source_staging_max_bytes": 448,
    "destination_staging_max_bytes": 32,
    "combined_staging_max_bytes": 480,
    "source_copy_max_bytes": 448,
    "destination_copy_max_bytes": 32,
    "local_copy_runs": 2044,
    "affine_copy_loops": 1587,
    "copy_cycles_floor": 60,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2688,
    "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": 3,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 18432,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 18432,
    "source_copy_max_bytes": 18432,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 2304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 3,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18432,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 18432,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 20736,
    "source_copy_max_bytes": 18432,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2640,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18048,
    "maximum_incoming_bytes": 2688,
    "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": 8,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18048,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 18048,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 18048,
    "source_copy_max_bytes": 18048,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 2256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 54,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18048,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 8,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 9,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 18048,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 18048,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 20352,
    "source_copy_max_bytes": 18048,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2592,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2688,
    "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": 4,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 16128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 16128,
    "source_copy_max_bytes": 16128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 2016,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 4,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 16128,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 16128,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 18432,
    "source_copy_max_bytes": 16128,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2352,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 15744,
    "maximum_incoming_bytes": 2688,
    "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": 7,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 15744,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 15744,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 15744,
    "source_copy_max_bytes": 15744,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 1968,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 15744,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 7,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 15744,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 15744,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 18048,
    "source_copy_max_bytes": 15744,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "original",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 729, 72]) F16 RowMajor -> TensorShape([32, 729, 72]) RowMajor",
    "transfers": 23312,
    "receive_fragments": 23312,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3356928,
    "received_bytes": 3356928,
    "maximum_outgoing_bytes": 2304,
    "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": 33,
    "mode": "source",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 729, 72]) F16 RowMajor -> TensorShape([32, 729, 72]) RowMajor",
    "transfers": 23312,
    "receive_fragments": 23312,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3356928,
    "received_bytes": 3356928,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4608,
    "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": 23312,
    "affine_copy_loops": 1486,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "destination",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 729, 72]) F16 RowMajor -> TensorShape([32, 729, 72]) RowMajor",
    "transfers": 23312,
    "receive_fragments": 23312,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3356928,
    "received_bytes": 3356928,
    "maximum_outgoing_bytes": 2304,
    "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": 744,
    "affine_copy_loops": 744,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 33,
    "mode": "both",
    "label": "op Some(12) LayoutRearrangement: TensorShape([32, 729, 72]) F16 RowMajor -> TensorShape([32, 729, 72]) RowMajor",
    "transfers": 23312,
    "receive_fragments": 23312,
    "maximum_endpoint_fragments": 48,
    "transmitted_bytes": 3356928,
    "received_bytes": 3356928,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 24056,
    "affine_copy_loops": 2230,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 729]) F16 Amp(Left) -> TensorShape([32, 1, 729]) Amp(Left)",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 46080,
    "received_bytes": 46080,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1440,
    "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": 72,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 729]) F16 Amp(Left) -> TensorShape([32, 1, 729]) Amp(Left)",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 46080,
    "received_bytes": 46080,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1440,
    "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": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 4,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 729]) F16 Amp(Left) -> TensorShape([32, 1, 729]) Amp(Left)",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 46080,
    "received_bytes": 46080,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1440,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1440,
    "combined_staging_max_bytes": 1440,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1440,
    "local_copy_runs": 32,
    "affine_copy_loops": 32,
    "copy_cycles_floor": 180,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 72,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 729]) F16 Amp(Left) -> TensorShape([32, 1, 729]) Amp(Left)",
    "transfers": 1440,
    "receive_fragments": 1440,
    "maximum_endpoint_fragments": 45,
    "transmitted_bytes": 46080,
    "received_bytes": 46080,
    "maximum_outgoing_bytes": 32,
    "maximum_incoming_bytes": 1440,
    "source_staging_max_bytes": 32,
    "destination_staging_max_bytes": 1440,
    "combined_staging_max_bytes": 1440,
    "source_copy_max_bytes": 32,
    "destination_copy_max_bytes": 1440,
    "local_copy_runs": 1472,
    "affine_copy_loops": 1472,
    "copy_cycles_floor": 184,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2688,
    "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": 5,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13824,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 1728,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 5,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1344]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13824,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 13824,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 16128,
    "source_copy_max_bytes": 13824,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2064,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "original",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13440,
    "maximum_incoming_bytes": 2688,
    "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": 6,
    "mode": "source",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13440,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 13440,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 13440,
    "source_copy_max_bytes": 13440,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 9216,
    "affine_copy_loops": 1536,
    "copy_cycles_floor": 1680,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "destination",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 9216,
    "receive_fragments": 9216,
    "maximum_endpoint_fragments": 42,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13440,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 2688,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 1536,
    "affine_copy_loops": 1512,
    "copy_cycles_floor": 336,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 6,
    "mode": "both",
    "label": "op Some(0) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1536,
    "receive_fragments": 1536,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 13440,
    "maximum_incoming_bytes": 2688,
    "source_staging_max_bytes": 13440,
    "destination_staging_max_bytes": 2688,
    "combined_staging_max_bytes": 15744,
    "source_copy_max_bytes": 13440,
    "destination_copy_max_bytes": 2688,
    "local_copy_runs": 10752,
    "affine_copy_loops": 3048,
    "copy_cycles_floor": 2016,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 6672,
    "receive_fragments": 6672,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 6451968,
    "received_bytes": 6451968,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 4864,
    "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": 35,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 3336,
    "receive_fragments": 3336,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 6451968,
    "received_bytes": 6451968,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 4864,
    "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": 6672,
    "affine_copy_loops": 656,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 6672,
    "receive_fragments": 6672,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 6451968,
    "received_bytes": 6451968,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 4864,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4864,
    "combined_staging_max_bytes": 4864,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4864,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 608,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 35,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 3336,
    "receive_fragments": 3336,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 6451968,
    "received_bytes": 6451968,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 4864,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 4864,
    "combined_staging_max_bytes": 33280,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 4864,
    "local_copy_runs": 8112,
    "affine_copy_loops": 2096,
    "copy_cycles_floor": 4160,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 40,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 388,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 40,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2116,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 41,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 378,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 41,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2106,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 42,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 358,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 42,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2086,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 43,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 338,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 43,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2066,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 44,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 318,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 44,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2046,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 45,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 298,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 45,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2026,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 123,
    "mode": "original",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 40,
    "receive_fragments": 2878,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 17216,
    "received_bytes": 1238656,
    "maximum_outgoing_bytes": 17216,
    "maximum_incoming_bytes": 896,
    "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": 123,
    "mode": "source",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 20,
    "receive_fragments": 1439,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 17216,
    "received_bytes": 1238656,
    "maximum_outgoing_bytes": 17216,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 17216,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 17216,
    "source_copy_max_bytes": 17216,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 20,
    "copy_cycles_floor": 2152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 123,
    "mode": "destination",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 40,
    "receive_fragments": 2878,
    "maximum_endpoint_fragments": 40,
    "transmitted_bytes": 17216,
    "received_bytes": 1238656,
    "maximum_outgoing_bytes": 17216,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 1439,
    "affine_copy_loops": 1439,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 123,
    "mode": "both",
    "label": "op Some(43) LayoutRearrangement: TensorShape([2, 1, 4304]) F16 RowMajor -> TensorShape([2, 1, 4304]) RowMajor",
    "transfers": 20,
    "receive_fragments": 1439,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 17216,
    "received_bytes": 1238656,
    "maximum_outgoing_bytes": 17216,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 17216,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 17216,
    "source_copy_max_bytes": 17216,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 1479,
    "affine_copy_loops": 1459,
    "copy_cycles_floor": 2264,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 36,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 458,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 36,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2186,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 37,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 438,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 37,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2166,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 38,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 418,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 38,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2146,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "original",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 39,
    "mode": "source",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "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": 3456,
    "affine_copy_loops": 398,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "destination",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 3456,
    "receive_fragments": 3456,
    "maximum_endpoint_fragments": 38,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1728,
    "affine_copy_loops": 1728,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 39,
    "mode": "both",
    "label": "op Some(14) OperatorInputs: TensorShape([2, 729, 1152]) F16 Amp(Left) -> TensorShape([1152]) RowMajor",
    "transfers": 1728,
    "receive_fragments": 1728,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 3363840,
    "received_bytes": 3363840,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 30848,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 5184,
    "affine_copy_loops": 2126,
    "copy_cycles_floor": 3856,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "original",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 36,
    "receive_fragments": 2590,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 256,
    "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": 76,
    "mode": "source",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 18,
    "receive_fragments": 1295,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 36,
    "affine_copy_loops": 18,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "destination",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 36,
    "receive_fragments": 2590,
    "maximum_endpoint_fragments": 36,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1295,
    "affine_copy_loops": 1295,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 76,
    "mode": "both",
    "label": "op Some(37) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 RowMajor -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 18,
    "receive_fragments": 1295,
    "maximum_endpoint_fragments": 18,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 1331,
    "affine_copy_loops": 1313,
    "copy_cycles_floor": 608,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "original",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "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": 82,
    "mode": "source",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 1,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "destination",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 64,
    "affine_copy_loops": 64,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 82,
    "mode": "both",
    "label": "op Some(39) OperatorInputs: TensorShape([2, 1, 32, 2]) F32 RowMajor -> TensorShape([2, 1, 32, 2]) RowMajor",
    "transfers": 64,
    "receive_fragments": 2048,
    "maximum_endpoint_fragments": 33,
    "transmitted_bytes": 512,
    "received_bytes": 16384,
    "maximum_outgoing_bytes": 8,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 8,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 264,
    "source_copy_max_bytes": 8,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 128,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 33,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "original",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "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": 68,
    "mode": "source",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 153,
    "affine_copy_loops": 82,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "destination",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 68,
    "mode": "both",
    "label": "op Some(29) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 729,
    "affine_copy_loops": 658,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "original",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "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": 79,
    "mode": "source",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 153,
    "affine_copy_loops": 82,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "destination",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 79,
    "mode": "both",
    "label": "op Some(38) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 729,
    "affine_copy_loops": 658,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 126,
    "mode": "original",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "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": 126,
    "mode": "source",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 153,
    "affine_copy_loops": 82,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 126,
    "mode": "destination",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 16,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 576,
    "affine_copy_loops": 576,
    "copy_cycles_floor": 2,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 126,
    "mode": "both",
    "label": "op Some(44) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 863,
    "receive_fragments": 1151,
    "maximum_endpoint_fragments": 32,
    "transmitted_bytes": 6904,
    "received_bytes": 9208,
    "maximum_outgoing_bytes": 256,
    "maximum_incoming_bytes": 16,
    "source_staging_max_bytes": 256,
    "destination_staging_max_bytes": 16,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 256,
    "destination_copy_max_bytes": 16,
    "local_copy_runs": 729,
    "affine_copy_loops": 658,
    "copy_cycles_floor": 34,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "original",
    "label": "op Some(35) LayoutRearrangement: TensorShape([32, 1, 72]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 31,
    "receive_fragments": 31,
    "maximum_endpoint_fragments": 31,
    "transmitted_bytes": 4464,
    "received_bytes": 4464,
    "maximum_outgoing_bytes": 144,
    "maximum_incoming_bytes": 4464,
    "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": 75,
    "mode": "source",
    "label": "op Some(35) LayoutRearrangement: TensorShape([32, 1, 72]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 31,
    "receive_fragments": 31,
    "maximum_endpoint_fragments": 31,
    "transmitted_bytes": 4464,
    "received_bytes": 4464,
    "maximum_outgoing_bytes": 144,
    "maximum_incoming_bytes": 4464,
    "source_staging_max_bytes": 144,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 144,
    "source_copy_max_bytes": 144,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 31,
    "affine_copy_loops": 31,
    "copy_cycles_floor": 18,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "destination",
    "label": "op Some(35) LayoutRearrangement: TensorShape([32, 1, 72]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 31,
    "receive_fragments": 31,
    "maximum_endpoint_fragments": 31,
    "transmitted_bytes": 4464,
    "received_bytes": 4464,
    "maximum_outgoing_bytes": 144,
    "maximum_incoming_bytes": 4464,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 4464,
    "combined_staging_max_bytes": 4464,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 4464,
    "local_copy_runs": 1,
    "affine_copy_loops": 1,
    "copy_cycles_floor": 558,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 75,
    "mode": "both",
    "label": "op Some(35) LayoutRearrangement: TensorShape([32, 1, 72]) F16 RowMajor -> TensorShape([32, 1, 72]) RowMajor",
    "transfers": 31,
    "receive_fragments": 31,
    "maximum_endpoint_fragments": 31,
    "transmitted_bytes": 4464,
    "received_bytes": 4464,
    "maximum_outgoing_bytes": 144,
    "maximum_incoming_bytes": 4464,
    "source_staging_max_bytes": 144,
    "destination_staging_max_bytes": 4464,
    "combined_staging_max_bytes": 4464,
    "source_copy_max_bytes": 144,
    "destination_copy_max_bytes": 4464,
    "local_copy_runs": 32,
    "affine_copy_loops": 32,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "original",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 14,
    "mode": "source",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 747,
    "affine_copy_loops": 738,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "destination",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 21,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 26244,
    "affine_copy_loops": 13851,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 14,
    "mode": "both",
    "label": "op Some(3) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 738,
    "receive_fragments": 13851,
    "maximum_endpoint_fragments": 11,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 7424,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 26991,
    "affine_copy_loops": 14589,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 21,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 21,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 22,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 22,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 23,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 23,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 24,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 24,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 25,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 25,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 26,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 26,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 27,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 27,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 28,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 28,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 29,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 29,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 30,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 30,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 31,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 160,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 160,
    "receive_fragments": 7360,
    "maximum_endpoint_fragments": 7,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 31,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 64, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22240,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "original",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 25, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "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": 32,
    "mode": "source",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 25, 80]) Amp(TransposedRight)",
    "transfers": 480,
    "receive_fragments": 22080,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 8192,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 192,
    "affine_copy_loops": 96,
    "copy_cycles_floor": 1024,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "destination",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 25, 80]) Amp(TransposedRight)",
    "transfers": 192,
    "receive_fragments": 8832,
    "maximum_endpoint_fragments": 8,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 20480,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22080,
    "affine_copy_loops": 8832,
    "copy_cycles_floor": 2560,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 32,
    "mode": "both",
    "label": "op Some(12) OperatorInputs: TensorShape([32, 729, 80]) F16 Amp(TransposedRight) -> TensorShape([32, 25, 80]) Amp(TransposedRight)",
    "transfers": 96,
    "receive_fragments": 4416,
    "maximum_endpoint_fragments": 4,
    "transmitted_bytes": 655360,
    "received_bytes": 30146560,
    "maximum_outgoing_bytes": 8192,
    "maximum_incoming_bytes": 20480,
    "source_staging_max_bytes": 8192,
    "destination_staging_max_bytes": 20480,
    "combined_staging_max_bytes": 28672,
    "source_copy_max_bytes": 8192,
    "destination_copy_max_bytes": 20480,
    "local_copy_runs": 22272,
    "affine_copy_loops": 8928,
    "copy_cycles_floor": 3584,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "original",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 48,
    "mode": "source",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 747,
    "affine_copy_loops": 747,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "destination",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 48,
    "mode": "both",
    "label": "op Some(17) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 7168,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 2205,
    "affine_copy_loops": 2205,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "original",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 59,
    "mode": "source",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "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": 747,
    "affine_copy_loops": 747,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "destination",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 59,
    "mode": "both",
    "label": "op Some(25) OperatorInputs: TensorShape([2, 729, 1152]) F16 RowMajor -> TensorShape([2, 729, 1152]) RowMajor",
    "transfers": 747,
    "receive_fragments": 26973,
    "maximum_endpoint_fragments": 20,
    "transmitted_bytes": 1684224,
    "received_bytes": 8398080,
    "maximum_outgoing_bytes": 2304,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 2304,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 7168,
    "source_copy_max_bytes": 2304,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 2205,
    "affine_copy_loops": 2205,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 125,
    "mode": "original",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 1368,
    "receive_fragments": 1368,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 175104,
    "received_bytes": 175104,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2432,
    "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": 125,
    "mode": "source",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 1368,
    "receive_fragments": 1368,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 175104,
    "received_bytes": 175104,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1368,
    "affine_copy_loops": 1368,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 125,
    "mode": "destination",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 1368,
    "receive_fragments": 1368,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 175104,
    "received_bytes": 175104,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 72,
    "affine_copy_loops": 72,
    "copy_cycles_floor": 304,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 125,
    "mode": "both",
    "label": "op Some(43) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1216]) RowMajor",
    "transfers": 1368,
    "receive_fragments": 1368,
    "maximum_endpoint_fragments": 19,
    "transmitted_bytes": 175104,
    "received_bytes": 175104,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2432,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 2432,
    "combined_staging_max_bytes": 2432,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 2432,
    "local_copy_runs": 1440,
    "affine_copy_loops": 1440,
    "copy_cycles_floor": 320,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "original",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "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": 67,
    "mode": "source",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1224,
    "affine_copy_loops": 1224,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "destination",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2176,
    "combined_staging_max_bytes": 2176,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2176,
    "local_copy_runs": 72,
    "affine_copy_loops": 72,
    "copy_cycles_floor": 272,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 67,
    "mode": "both",
    "label": "op Some(28) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 2176,
    "combined_staging_max_bytes": 2176,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 2176,
    "local_copy_runs": 1296,
    "affine_copy_loops": 1296,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "original",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "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": 78,
    "mode": "source",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 128,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1224,
    "affine_copy_loops": 1224,
    "copy_cycles_floor": 16,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "destination",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 2176,
    "combined_staging_max_bytes": 2176,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 2176,
    "local_copy_runs": 72,
    "affine_copy_loops": 72,
    "copy_cycles_floor": 272,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 78,
    "mode": "both",
    "label": "op Some(37) OperatorInputs: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([1088]) RowMajor",
    "transfers": 1224,
    "receive_fragments": 1224,
    "maximum_endpoint_fragments": 17,
    "transmitted_bytes": 156672,
    "received_bytes": 156672,
    "maximum_outgoing_bytes": 128,
    "maximum_incoming_bytes": 2176,
    "source_staging_max_bytes": 128,
    "destination_staging_max_bytes": 2176,
    "combined_staging_max_bytes": 2176,
    "source_copy_max_bytes": 128,
    "destination_copy_max_bytes": 2176,
    "local_copy_runs": 1296,
    "affine_copy_loops": 1296,
    "copy_cycles_floor": 288,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "original",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2368]) RowMajor",
    "transfers": 8580,
    "receive_fragments": 8580,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 40085760,
    "received_bytes": 40085760,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 28416,
    "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": 62,
    "mode": "source",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2368]) RowMajor",
    "transfers": 8580,
    "receive_fragments": 8580,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 40085760,
    "received_bytes": 40085760,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 28416,
    "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": 1480,
    "affine_copy_loops": 1480,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "destination",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2368]) RowMajor",
    "transfers": 8580,
    "receive_fragments": 8580,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 40085760,
    "received_bytes": 40085760,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 28416,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 28416,
    "combined_staging_max_bytes": 28416,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 28416,
    "local_copy_runs": 2820,
    "affine_copy_loops": 2820,
    "copy_cycles_floor": 3552,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 62,
    "mode": "both",
    "label": "op Some(26) OperatorInputs: TensorShape([2, 729, 2304]) F16 Amp(Left) -> TensorShape([2368]) RowMajor",
    "transfers": 8580,
    "receive_fragments": 8580,
    "maximum_endpoint_fragments": 12,
    "transmitted_bytes": 40085760,
    "received_bytes": 40085760,
    "maximum_outgoing_bytes": 28416,
    "maximum_incoming_bytes": 28416,
    "source_staging_max_bytes": 28416,
    "destination_staging_max_bytes": 28416,
    "combined_staging_max_bytes": 56832,
    "source_copy_max_bytes": 28416,
    "destination_copy_max_bytes": 28416,
    "local_copy_runs": 4300,
    "affine_copy_loops": 4300,
    "copy_cycles_floor": 7104,
    "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": 65,
    "mode": "original",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 144,
    "receive_fragments": 10360,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "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": 65,
    "mode": "source",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 144,
    "receive_fragments": 10360,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 64,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 144,
    "affine_copy_loops": 72,
    "copy_cycles_floor": 8,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "destination",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 144,
    "receive_fragments": 10360,
    "maximum_endpoint_fragments": 10,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 256,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 10360,
    "affine_copy_loops": 5180,
    "copy_cycles_floor": 32,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 65,
    "mode": "both",
    "label": "op Some(28) LayoutRearrangement: TensorShape([2, 1, 1152]) F16 Amp(Left) -> TensorShape([2, 1, 1152]) RowMajor",
    "transfers": 72,
    "receive_fragments": 5180,
    "maximum_endpoint_fragments": 5,
    "transmitted_bytes": 4608,
    "received_bytes": 331520,
    "maximum_outgoing_bytes": 64,
    "maximum_incoming_bytes": 256,
    "source_staging_max_bytes": 64,
    "destination_staging_max_bytes": 256,
    "combined_staging_max_bytes": 320,
    "source_copy_max_bytes": 64,
    "destination_copy_max_bytes": 256,
    "local_copy_runs": 10504,
    "affine_copy_loops": 5252,
    "copy_cycles_floor": 40,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "original",
    "label": "op Some(2) LayoutRearrangement: TensorShape([1, 729, 1152]) F16 RowMajor -> TensorShape([1, 729, 1152]) RowMajor",
    "transfers": 2193,
    "receive_fragments": 2193,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1678472,
    "received_bytes": 1678472,
    "maximum_outgoing_bytes": 1144,
    "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": 13,
    "mode": "source",
    "label": "op Some(2) LayoutRearrangement: TensorShape([1, 729, 1152]) F16 RowMajor -> TensorShape([1, 729, 1152]) RowMajor",
    "transfers": 2193,
    "receive_fragments": 2193,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1678472,
    "received_bytes": 1678472,
    "maximum_outgoing_bytes": 1144,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 1144,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1144,
    "source_copy_max_bytes": 1144,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 2128,
    "affine_copy_loops": 2128,
    "copy_cycles_floor": 143,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 13,
    "mode": "destination",
    "label": "op Some(2) LayoutRearrangement: TensorShape([1, 729, 1152]) F16 RowMajor -> TensorShape([1, 729, 1152]) RowMajor",
    "transfers": 2193,
    "receive_fragments": 2193,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1678472,
    "received_bytes": 1678472,
    "maximum_outgoing_bytes": 1144,
    "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": 13,
    "mode": "both",
    "label": "op Some(2) LayoutRearrangement: TensorShape([1, 729, 1152]) F16 RowMajor -> TensorShape([1, 729, 1152]) RowMajor",
    "transfers": 2193,
    "receive_fragments": 2193,
    "maximum_endpoint_fragments": 6,
    "transmitted_bytes": 1678472,
    "received_bytes": 1678472,
    "maximum_outgoing_bytes": 1144,
    "maximum_incoming_bytes": 2304,
    "source_staging_max_bytes": 1144,
    "destination_staging_max_bytes": 2304,
    "combined_staging_max_bytes": 3448,
    "source_copy_max_bytes": 1144,
    "destination_copy_max_bytes": 2304,
    "local_copy_runs": 2857,
    "affine_copy_loops": 2857,
    "copy_cycles_floor": 431,
    "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": 66,
    "mode": "original",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "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": 66,
    "mode": "source",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1294,
    "affine_copy_loops": 1294,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "destination",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 1294,
    "affine_copy_loops": 1294,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 66,
    "mode": "both",
    "label": "op Some(28) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 2588,
    "affine_copy_loops": 2588,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "original",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "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": 77,
    "mode": "source",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1294,
    "affine_copy_loops": 1294,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "destination",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 1024,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 1294,
    "affine_copy_loops": 1294,
    "copy_cycles_floor": 128,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 77,
    "mode": "both",
    "label": "op Some(37) LayoutRearrangement: TensorShape([1152, 1152]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 1152]) BlockMajor(Matrix { row_block: 64, column_block: 16 })",
    "transfers": 1294,
    "receive_fragments": 1294,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 1325056,
    "received_bytes": 1325056,
    "maximum_outgoing_bytes": 1024,
    "maximum_incoming_bytes": 1024,
    "source_staging_max_bytes": 1024,
    "destination_staging_max_bytes": 1024,
    "combined_staging_max_bytes": 2048,
    "source_copy_max_bytes": 1024,
    "destination_copy_max_bytes": 1024,
    "local_copy_runs": 2588,
    "affine_copy_loops": 2588,
    "copy_cycles_floor": 256,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "original",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1152, 4304]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 4304]) BlockMajor(Matrix { row_block: 32, column_block: 16 })",
    "transfers": 1438,
    "receive_fragments": 1438,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 4951552,
    "received_bytes": 4951552,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "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": 85,
    "mode": "source",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1152, 4304]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 4304]) BlockMajor(Matrix { row_block: 32, column_block: 16 })",
    "transfers": 1438,
    "receive_fragments": 1438,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 4951552,
    "received_bytes": 4951552,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 1438,
    "affine_copy_loops": 1438,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "destination",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1152, 4304]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 4304]) BlockMajor(Matrix { row_block: 32, column_block: 16 })",
    "transfers": 1438,
    "receive_fragments": 1438,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 4951552,
    "received_bytes": 4951552,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 3584,
    "combined_staging_max_bytes": 3584,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 3584,
    "local_copy_runs": 1438,
    "affine_copy_loops": 1438,
    "copy_cycles_floor": 448,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 85,
    "mode": "both",
    "label": "op Some(40) LayoutRearrangement: TensorShape([1152, 4304]) F8F143 { scale_exponent: -4 } Amp(TransposedLeft) -> TensorShape([1152, 4304]) BlockMajor(Matrix { row_block: 32, column_block: 16 })",
    "transfers": 1438,
    "receive_fragments": 1438,
    "maximum_endpoint_fragments": 2,
    "transmitted_bytes": 4951552,
    "received_bytes": 4951552,
    "maximum_outgoing_bytes": 3584,
    "maximum_incoming_bytes": 3584,
    "source_staging_max_bytes": 3584,
    "destination_staging_max_bytes": 3584,
    "combined_staging_max_bytes": 7168,
    "source_copy_max_bytes": 3584,
    "destination_copy_max_bytes": 3584,
    "local_copy_runs": 2876,
    "affine_copy_loops": 2876,
    "copy_cycles_floor": 896,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 12,
    "mode": "original",
    "label": "op Some(1) 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": 12,
    "mode": "source",
    "label": "op Some(1) 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": 12,
    "mode": "destination",
    "label": "op Some(1) 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": 12,
    "mode": "both",
    "label": "op Some(1) 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
  },
  {
    "phase": 15,
    "mode": "original",
    "label": "op Some(3) 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": 15,
    "mode": "source",
    "label": "op Some(3) 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": 15,
    "mode": "destination",
    "label": "op Some(3) 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": 15,
    "mode": "both",
    "label": "op Some(3) 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
  },
  {
    "phase": 19,
    "mode": "original",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 729, 3456]) F16 RowMajor -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 5038848,
    "received_bytes": 5038848,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 6912,
    "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": 19,
    "mode": "source",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 729, 3456]) F16 RowMajor -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 5038848,
    "received_bytes": 5038848,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 6912,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 6912,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "destination",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 729, 3456]) F16 RowMajor -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 5038848,
    "received_bytes": 5038848,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 864,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 19,
    "mode": "both",
    "label": "op Some(5) LayoutRearrangement: TensorShape([2, 729, 3456]) F16 RowMajor -> TensorShape([2, 729, 3456]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 5038848,
    "received_bytes": 5038848,
    "maximum_outgoing_bytes": 6912,
    "maximum_incoming_bytes": 6912,
    "source_staging_max_bytes": 6912,
    "destination_staging_max_bytes": 6912,
    "combined_staging_max_bytes": 6912,
    "source_copy_max_bytes": 6912,
    "destination_copy_max_bytes": 6912,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 1728,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 47,
    "mode": "original",
    "label": "op Some(15) 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": 47,
    "mode": "source",
    "label": "op Some(15) 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": 47,
    "mode": "destination",
    "label": "op Some(15) 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": 47,
    "mode": "both",
    "label": "op Some(15) 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
  },
  {
    "phase": 49,
    "mode": "original",
    "label": "op Some(17) 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": 49,
    "mode": "source",
    "label": "op Some(17) 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": 49,
    "mode": "destination",
    "label": "op Some(17) 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": 49,
    "mode": "both",
    "label": "op Some(17) 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
  },
  {
    "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
  },
  {
    "phase": 60,
    "mode": "original",
    "label": "op Some(25) 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": 60,
    "mode": "source",
    "label": "op Some(25) 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": 60,
    "mode": "destination",
    "label": "op Some(25) 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": 60,
    "mode": "both",
    "label": "op Some(25) 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
  },
  {
    "phase": 64,
    "mode": "original",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 729, 2304]) F16 RowMajor -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3359232,
    "received_bytes": 3359232,
    "maximum_outgoing_bytes": 4608,
    "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": 64,
    "mode": "source",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 729, 2304]) F16 RowMajor -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3359232,
    "received_bytes": 3359232,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "destination",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 729, 2304]) F16 RowMajor -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3359232,
    "received_bytes": 3359232,
    "maximum_outgoing_bytes": 4608,
    "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": 729,
    "affine_copy_loops": 729,
    "copy_cycles_floor": 576,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 64,
    "mode": "both",
    "label": "op Some(27) LayoutRearrangement: TensorShape([2, 729, 2304]) F16 RowMajor -> TensorShape([2, 729, 2304]) RowMajor",
    "transfers": 729,
    "receive_fragments": 729,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 3359232,
    "received_bytes": 3359232,
    "maximum_outgoing_bytes": 4608,
    "maximum_incoming_bytes": 4608,
    "source_staging_max_bytes": 4608,
    "destination_staging_max_bytes": 4608,
    "combined_staging_max_bytes": 4608,
    "source_copy_max_bytes": 4608,
    "destination_copy_max_bytes": 4608,
    "local_copy_runs": 1458,
    "affine_copy_loops": 1458,
    "copy_cycles_floor": 1152,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "original",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 784]) F16 Amp(Left) -> TensorShape([32, 1, 736]) Amp(Left)",
    "transfers": 32,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 47104,
    "received_bytes": 188416,
    "maximum_outgoing_bytes": 1472,
    "maximum_incoming_bytes": 1472,
    "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": 73,
    "mode": "source",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 784]) F16 Amp(Left) -> TensorShape([32, 1, 736]) Amp(Left)",
    "transfers": 32,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 47104,
    "received_bytes": 188416,
    "maximum_outgoing_bytes": 1472,
    "maximum_incoming_bytes": 1472,
    "source_staging_max_bytes": 1472,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 1472,
    "source_copy_max_bytes": 1472,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 32,
    "affine_copy_loops": 32,
    "copy_cycles_floor": 184,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "destination",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 784]) F16 Amp(Left) -> TensorShape([32, 1, 736]) Amp(Left)",
    "transfers": 32,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 47104,
    "received_bytes": 188416,
    "maximum_outgoing_bytes": 1472,
    "maximum_incoming_bytes": 1472,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 1472,
    "combined_staging_max_bytes": 1472,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 1472,
    "local_copy_runs": 128,
    "affine_copy_loops": 128,
    "copy_cycles_floor": 184,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 73,
    "mode": "both",
    "label": "op Some(35) OperatorInputs: TensorShape([32, 1, 784]) F16 Amp(Left) -> TensorShape([32, 1, 736]) Amp(Left)",
    "transfers": 32,
    "receive_fragments": 128,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 47104,
    "received_bytes": 188416,
    "maximum_outgoing_bytes": 1472,
    "maximum_incoming_bytes": 1472,
    "source_staging_max_bytes": 1472,
    "destination_staging_max_bytes": 1472,
    "combined_staging_max_bytes": 1472,
    "source_copy_max_bytes": 1472,
    "destination_copy_max_bytes": 1472,
    "local_copy_runs": 160,
    "affine_copy_loops": 160,
    "copy_cycles_floor": 368,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 86,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 86,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 87,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 87,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 88,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 88,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 89,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 89,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 90,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 90,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 91,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 91,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 92,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 92,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 93,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 93,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 94,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 94,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 95,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 95,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 96,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 96,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 97,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 97,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 98,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 98,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 99,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 99,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 100,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 100,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 101,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 101,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 102,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 102,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 103,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 103,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 104,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 104,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 105,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 105,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 106,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 106,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 107,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 107,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 108,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 108,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 109,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 109,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 110,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 110,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 111,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 111,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 111,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 111,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 112,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 112,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 112,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 112,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 113,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 113,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 113,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 113,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 114,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 114,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 114,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 114,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 115,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 115,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 115,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 115,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 116,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 116,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 116,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 116,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 117,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 117,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 117,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 117,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 118,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 118,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 118,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 118,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 119,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 119,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 119,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 119,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 120,
    "mode": "original",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "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": 120,
    "mode": "source",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 0,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 0,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 120,
    "mode": "destination",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 0,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 0,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 40,
    "affine_copy_loops": 40,
    "copy_cycles_floor": 112,
    "replay_address_ranges_valid": true
  },
  {
    "phase": 120,
    "mode": "both",
    "label": "op Some(40) OperatorInputs: TensorShape([2, 1, 4304]) F16 Amp(Left) -> TensorShape([448]) RowMajor",
    "transfers": 40,
    "receive_fragments": 40,
    "maximum_endpoint_fragments": 1,
    "transmitted_bytes": 34432,
    "received_bytes": 34432,
    "maximum_outgoing_bytes": 896,
    "maximum_incoming_bytes": 896,
    "source_staging_max_bytes": 896,
    "destination_staging_max_bytes": 896,
    "combined_staging_max_bytes": 896,
    "source_copy_max_bytes": 896,
    "destination_copy_max_bytes": 896,
    "local_copy_runs": 80,
    "affine_copy_loops": 80,
    "copy_cycles_floor": 224,
    "replay_address_ranges_valid": true
  }
]