inference_3pc_bert_wan.yml

$SUITE_VARS$:

  _python_pre: "export PYTHONPATH={{ exp_code_dir }}/utils && . {{ exp_code_dir }}/.venv/bin/activate"
  _python_path: "[% my_run._python_pre %] && {{ exp_code_dir }}/.venv/bin/python"
  cmd_mpspdz: "[% my_run._python_path %] -m  python_utils.scripts.experiment_runner --player-number <PLAYER-ID>"

  commit_output: True
  convert_ring_bits: 31
  convert_n_threads: 36
  convert_chunk_size: 100000

  $INCLUDE_VARS$: dataset.yml

  network_type: lan

  consistency_args:
    abs_path_to_code_dir: "{{ exp_consistency_dir }}"
    hosts_file: "{{ exp_consistency_hosts_file }}"
    pc: kzg
    pp_args: "[% my_run.dataset_info[my_run.mpc.script_args.dataset].model_input_size %]"
    gen_pp: True
    type: pc
    use_split: True

  mpc:
    player_0_hostname: "[% exp_host_lst | json_query('[?host_type==`large_memory`].private_dns_name') | default(['<UNDEFINED-DNS>'], true) | first %]"
    abs_path_to_code_dir: "{{ exp_code_dir }}"
    player_count: 3
    protocol_setup: $FACTOR$

    compiler_args: $FACTOR$ #'100000000'] # budget was '1000000'1million # also look at B 5 -> cannot be used together with R
    domain: $FACTOR$
    custom_prime: $FACTOR$

    script_name: "inference"

    script_args:
      n_samples: 1
      n_input_parties: 3

      consistency_check: "[% my_run.consistency_args.type %]"
      sha3_approx_factor: "[% my_run.dataset_info[my_run.mpc.script_args.dataset].sha3_approx_factor | default(1) %]"

      debug: False  # for a final benchmarking run, can also use `debug: False`, this should then only release no secret info
      emulate: False
      dataset:
        $FACTOR$: [ glue-qnli ]

      # in a network with latency, a high batch size for ml.py is key to a good performance (60k raised insufficient memory error)
      batch_size: 1 #1024 #256 # was 128 before   # 60000 -> insufficient memory
      n_threads: 36 #128 #64 # TODO before was 32 -> see if this helps (128 raises too many files runtime error)

      audit_trigger_idx: 0 # select single audit trigger (from all)


    stage: # "compile" in run and then in next run "run"
      $FACTOR$: [ compile, run ] # ensure that this is last factor

wan:
  n_repetitions: 1
  host_types:
    large_memory:
      n: 3
      check_status: True
      init_roles:
        - setup-base
        - setup-consistency
        - setup-spdz-secrets
        - setup-network-delay
      $CMD$:
        - main: "[% my_run.cmd_mpspdz | replace('<PLAYER-ID>', 0) %]"
        - main: "[% my_run.cmd_mpspdz | replace('<PLAYER-ID>', 1) %]"
        - main: "[% my_run.cmd_mpspdz | replace('<PLAYER-ID>', 2) %]"

  base_experiment:

    consistency_args:
      type:
        $FACTOR$: [ pc, sha3s, cerebro ]

    network_type: wan
    mpc_type: $FACTOR$
#    mpc:
#      program_args:
#        b: 1000

  factor_levels:

    - mpc_type: sh
      mpc:
        protocol_setup: "semi_honest_3"
        domain: ring_split
        compiler_args: [ '-R', '64', '-C', '--budget', "[% my_run._runtime_info.budget_wan %]", '-Z', '3', '-D' ]
        custom_prime: null

    - mpc_type: mal
      mpc:
        protocol_setup: "sy-rep-ring-party"
        domain: ring_split
        compiler_args: [ '-R', '64', '-C', '--budget', "[% my_run._runtime_info.budget_wan %]", '-Z', '3', '-D' ]
        custom_prime: null

$ETL$: {}