Function model_provider

codegeex/megatron/tools/pretrain_codegeex.py:20–86

Build the model.

(pre_process=True, post_process=True)

def model_provider(pre_process=True, post_process=True):
    """Build the model."""

    print_rank_0("building GPT model ...")
    see_memory_usage(f"Before Building Model", force=True)

    args = get_args()
    with deepspeed.zero.Init(
        data_parallel_group=mpu.get_data_parallel_group(),
        remote_device=None if args.remote_device == "none" else args.remote_device,
        config_dict_or_path=args.deepspeed_config,
        enabled=args.zero_stage == 3,
        mpu=mpu,
    ):
        if args.deepspeed and not args.no_pipeline_parallel:
            model = CodeGeeXModelPipe(num_tokentypes=0, parallel_output=True)
            # This is a hack to give us a reference to get_batch_pipe from within training.py
            # We need to call model.set_batch_fn after deepspeed.initialize
            model._megatron_batch_fn = get_batch_pipe

            # Predompute the attention mask and store it in args. This avoids having to
            # pipeline it as an activation during training. The mask is constant, and thus
            # we can reuse it.
            attention_mask = torch.tril(
                torch.ones(
                    (1, args.seq_length, args.seq_length),
                    device=torch.cuda.current_device(),
                )
            ).view(1, 1, args.seq_length, args.seq_length)

            # Convert attention mask to binary:
            attention_mask = attention_mask < 0.5
            if args.fp16:
                attention_mask = attention_mask.half()
            elif args.bf16:
                attention_mask = attention_mask.bfloat16()

            # Attention mask must be bool.
            args.attn_mask = attention_mask.to(torch.bool)

        else:
            model = CodeGeeXModel(
                num_tokentypes=0,
                parallel_output=True,
            )
            
            if args.load_state is not None:
                timers = get_timers()
                print_rank_0("Loading warmstarting model states ...")
                timers("load-model-states").start()
                mp_rank = mpu.get_tensor_model_parallel_rank()
                if os.path.isdir(args.load_state):
                    model_path = os.path.join(
                        args.load_state, "mp_rank_{:02d}_model_states.pt".format(mp_rank)
                    )
                else:
                    model_path = args.load_state
                print_rank_0(f"Loading model from {model_path} ...")