Support large model export using multi-gpu (#17990)

### Description

This PR is to implemente a exporter which works for large language
models(LLM).
It works for models like Llama2-70b or gpt-175.

The main idea is to utilize multiple-GPU and dispatch differnet layers
to different GPU, in short, it symply implemented auto pipeline
parallelism.

For example : to export Llama2-70b, you need 8x V100-32GB or 4x A100-80G
or More GPU memories.

It would expect to export decoder-only models. For encoder-decoder
arch-like models, we didn't test it yet.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

---------

Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>

onnxruntime/python/tools/transformers/large_model_exporter.py

@@ -0,0 +1,385 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation.  All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+
+"""
+Export LLM to onnx
+"""
+import argparse
+import inspect
+import math
+import os
+import tempfile
+from pathlib import Path
+from typing import Optional
+
+import onnx
+import torch
+import transformers
+from torch import nn
+
+
+def disable_huggingface_init():
+    """do not init model twice as it slow initialization"""
+
+    torch.nn.init.kaiming_uniform_ = lambda x, *args, **kwargs: x
+    torch.nn.init.uniform_ = lambda x, *args, **kwargs: x
+    torch.nn.init.normal_ = lambda x, *args, **kwargs: x
+    torch.nn.init.constant_ = lambda x, *args, **kwargs: x
+    torch.nn.init.xavier_uniform_ = lambda x, *args, **kwargs: x
+    torch.nn.init.xavier_normal_ = lambda x, *args, **kwargs: x
+    torch.nn.init.kaiming_normal_ = lambda x, *args, **kwargs: x
+    torch.nn.init.orthogonal_ = lambda x, *args, **kwargs: x
+
+
+def get_model_parameter_size(model: nn.Module):
+    """to calculate how much memory this model needs"""
+    param_size = 0
+    param_sum = 0
+    for param in model.parameters():
+        param_size += param.nelement() * param.element_size()
+        param_sum += param.nelement()
+    buffer_size = 0
+    buffer_sum = 0
+    for buffer in model.buffers():
+        buffer_size += buffer.nelement() * buffer.element_size()
+        buffer_sum += buffer.nelement()
+    all_size = (param_size + buffer_size) / 1024 / 1024
+    return all_size
+
+
+def initialize_model_and_sample_inputs(hf_model: str, cache_dir: Optional[str], tokenizer=None):
+    """
+    get the pretrained torch model from hugginface,
+    and sample model-inputs
+    """
+
+    disable_huggingface_init()
+
+    model = transformers.AutoModelForCausalLM.from_pretrained(  # type: ignore
+        hf_model, torch_dtype=torch.float16, cache_dir=cache_dir, trust_remote_code=True
+    )
+    if tokenizer is None:
+        tokenizer = hf_model
+    tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer)  # type: ignore
+
+    sample_inputs = tuple(tokenizer("Hello, my dog is cute", return_tensors="pt").values())
+    return model, sample_inputs
+
+
+def auto_pipeline_parallel(model: nn.Module, gpulist: list, sample_inputs: tuple):
+    """Make the model executable across multiple GPUs."""
+
+    def input_gpu_device_hook(mod, inputs, kwargs):
+        modifyed_inputs = []
+        first_dev = None
+        for layer_input in inputs:
+            if type(layer_input) is not torch.Tensor:
+                modifyed_inputs.append(layer_input)
+            elif hasattr(mod, "weight"):
+                modifyed_inputs.append(layer_input.to(mod.weight.device))
+            elif hasattr(mod, "parameters"):
+                device = next(mod.parameters(), layer_input).device
+                modifyed_inputs.append(layer_input.to(device))
+            elif hasattr(next(mod.children(), None), "weight"):
+                modifyed_inputs.append(layer_input.to(next(mod.children()).weight.device))
+            elif first_dev is not None and layer_input.device != first_dev:
+                modifyed_inputs.append(layer_input.to(first_dev))
+            else:
+                modifyed_inputs.append(layer_input)
+            if first_dev is None:
+                first_dev = modifyed_inputs[0].device
+        for key, value in kwargs.items():
+            if type(value) is torch.Tensor:
+                kwargs[key] = value.to(first_dev)
+
+        return (tuple(modifyed_inputs), kwargs)
+
+    def move_layer_to_device_rurc(mod, dev):
+        mod.to(dev)
+        for layer in mod.named_children():
+            move_layer_to_device_rurc(layer[1], dev)
+
+    model = model.half()
+    all_hooks = []
+    all_hooks.append(model.register_forward_pre_hook(input_gpu_device_hook, with_kwargs=True))
+    pre_fix = next(iter(model.named_children()))[0]
+    for top_name, top_module in model.named_children():
+        for name, module in top_module.named_children():
+            all_hooks.append(module.register_forward_pre_hook(input_gpu_device_hook, with_kwargs=True))
+            if type(module) in [torch.nn.ModuleList]:
+                num_layers_on_each_gpu = math.floor(len(module) / len(gpulist))
+                for idx, attn_layer in enumerate(module):
+                    all_hooks.append(attn_layer.register_forward_pre_hook(input_gpu_device_hook, with_kwargs=True))
+
+                    to_dev = gpulist[min(idx // num_layers_on_each_gpu, len(gpulist))]
+                    attn_layer.to(to_dev)
+                    move_layer_to_device_rurc(attn_layer, to_dev)
+                    print(f"move {pre_fix}.{name}.{idx} to {to_dev}")
+            else:
+                module.to(gpulist[0])
+                print(f"move {pre_fix}.{name} to {gpulist[0]}")
+        if len(list(top_module.named_children())) == 0:
+            top_module.to(gpulist[0])
+            print(f"move {top_name} to {gpulist[0]}")
+
+    with torch.no_grad():
+        model(sample_inputs[0], attention_mask=sample_inputs[1])
+    return model
+
+
+def retrieve_onnx_inputs(model: nn.Module, sample_inputs: tuple, with_past: bool):
+    """
+    auto retrieve onnx inputs from torch model as we can't enumlate all possibilities
+    for all models
+    """
+    user_inputs = []
+
+    def hook_for_inputs(_, inputs, kwargs):
+        user_inputs.append((inputs, kwargs))
+        return user_inputs[0]
+
+    hook_handle = model.register_forward_pre_hook(hook_for_inputs, with_kwargs=True)
+
+    forward_params = inspect.signature(model.forward).parameters
+    input_keys = list(forward_params.keys())
+    default_values = [forward_params.get(key).default for key in input_keys]
+    out = model(sample_inputs[0], attention_mask=sample_inputs[1])
+    hook_handle.remove()
+    user_inputs = user_inputs[0]
+    onnx_inputs = default_values
+    for idx, _val in enumerate(user_inputs[0]):
+        onnx_inputs[idx] = user_inputs[0][idx]
+    for key, value in user_inputs[1].items():
+        idx = input_keys.index(key)
+        onnx_inputs[idx] = value
+    for idx, (key, value) in enumerate(zip(input_keys, onnx_inputs)):
+        if type(value) is torch.Tensor:
+            value.to(model.device)
+        # Didn't touch past_key_value now, please change it if you want
+        if "use_cache" in key:
+            onnx_inputs[idx] = with_past
+
+    return input_keys, onnx_inputs, out.past_key_values
+
+
+def move_to_approprate_device(model: nn.Module, sample_inputs_tp: tuple) -> nn.Module:
+    """
+    According to the model size, we will upload it to
+    CPU if has no GPU or enough GPU memory,
+    Single GPU if has only one GPU in local or model size is enough to fit one GPU
+    Multiple GPU if there is more than one gpu in local and model is too large
+    """
+    total_mem_per_cpu = torch.cuda.get_device_properties(0).total_memory / 1024 / 1024
+
+    print(f"Model_Size = {get_model_parameter_size(model)/1024} GB")
+    print(f"total_mem_per_cpu = {total_mem_per_cpu/1024} GB")
+    if get_model_parameter_size(model) > total_mem_per_cpu * 0.45:
+        device_collection = [torch.device(i) for i in range(torch.cuda.device_count())]
+        if len(device_collection) > 1:
+            print(
+                f"{len(device_collection)} GPUs are used to export onnx, \
+                   Please set CUDA_VISIBLE_DEVICES to use specific GPU group"
+            )
+            model = auto_pipeline_parallel(model, device_collection, sample_inputs_tp)
+        else:
+            print("!!!! convert model to float and export onnx using CPU")
+            model = model.cpu().float()
+    else:
+        print("Export model on a single GPU")
+        model = model.cuda().half()
+    return model
+
+
+def adapt_inputs_to_device(sample_inputs: tuple, device: torch.device) -> tuple:
+    """move inputs to device"""
+    sample_inputs_ = []
+    for sample_int in sample_inputs:
+        if isinstance(sample_int, torch.Tensor):
+            sample_inputs_.append(sample_int.to(device))
+        else:
+            sample_inputs_.append(sample_int)
+    return tuple(sample_inputs_)
+
+
+def fetch_onnx_inputs_outputs_name(
+    model: nn.Module,
+    onnx_inputs: list,
+    torch_input_names: tuple,
+    past_key_values: tuple,
+    with_past: bool,
+    input_with_past: bool,
+):
+    """fetch onnx inputs and outputs name"""
+    num_of_past_key = 0
+    kv_cache_axis = {0: "batch_size"}
+    # try get num_of_past_key and shape of past_key_value
+    if past_key_values is not None:
+        num_of_past_key = len(past_key_values)
+        seq_index = (torch.tensor(past_key_values[0][0].shape) == onnx_inputs[0].shape[-1]).nonzero().view(-1)
+        assert seq_index.numel() == 1
+        kv_cache_axis = {0: "batch_size", seq_index.item(): "seq_len"}
+
+    if not num_of_past_key:
+        num_of_past_key = model.config.num_hidden_layers
+
+    onnx_inp_names = ("input_ids", "attention_mask")
+    onnx_out_names = ("logits",)
+    onnx_dynamic_axes = {
+        "input_ids": {0: "batch_size", 1: "seq_len"},
+        "attention_mask": {0: "batch_size", 1: "seq_len"},
+    }
+    if input_with_past:
+        for i in range(num_of_past_key):
+            onnx_inp_names += (f"present_key.{i}",)
+            onnx_inp_names += (f"present_values.{i}",)
+
+            onnx_dynamic_axes[onnx_inp_names[-1]] = kv_cache_axis
+            onnx_dynamic_axes[onnx_inp_names[-2]] = kv_cache_axis
+
+    if with_past or input_with_past:
+        for i in range(num_of_past_key):
+            onnx_out_names += (f"past_key.{i}",)
+            onnx_out_names += (f"past_values.{i}",)
+            onnx_dynamic_axes[onnx_out_names[-1]] = kv_cache_axis
+            onnx_dynamic_axes[onnx_out_names[-2]] = kv_cache_axis
+
+    for idx, name in enumerate(torch_input_names):
+        if input_with_past:
+            if name == "past_key_values":
+                onnx_inputs[idx] = past_key_values
+            elif name == "attention_mask":
+                attn_mask = onnx_inputs[idx]
+                onnx_inputs[idx] = torch.cat(
+                    (attn_mask, torch.ones((attn_mask.shape[0], 1), device=attn_mask.device)), dim=1
+                )
+            elif name == "input_ids":
+                input_ids = onnx_inputs[idx]
+                onnx_inputs[idx] = input_ids[:, -1:]
+
+    return onnx_inp_names, onnx_out_names, onnx_dynamic_axes
+
+
+def do_export_internal(model: nn.Module, onnx_io_tuple: tuple, onnx_inputs: tuple, onnx_path: Path, opset: int):
+    """do export with torch.onnx.export"""
+    onnx_model_name = onnx_path.name
+    onnx_inp_names, onnx_out_names, onnx_dynamic_axes = onnx_io_tuple
+    # two step to export onnx
+    # 1. export onnx with lots of pieces of weights
+    # 2. save all weights to external data
+    with tempfile.TemporaryDirectory() as tmpdirname:
+        tmp_onnx = os.path.join(tmpdirname, "tmp.onnx")
+
+        torch.onnx.export(
+            model=model,
+            args=tuple(onnx_inputs),
+            f=tmp_onnx,
+            verbose=False,
+            opset_version=opset,
+            input_names=onnx_inp_names,
+            output_names=onnx_out_names,
+            dynamic_axes=onnx_dynamic_axes,
+        )
+
+        onnx_path.unlink(missing_ok=True)
+        (onnx_path.parent / f"{onnx_model_name}_ext.data").unlink(missing_ok=True)
+
+        onnx_model = onnx.load(str(tmp_onnx))
+        onnx.save_model(
+            onnx_model,
+            str(onnx_path),
+            save_as_external_data=(len(os.listdir(tmpdirname)) > 1),
+            all_tensors_to_one_file=True,
+            location=f"{onnx_model_name}_ext.data",
+            size_threshold=1024,
+            convert_attribute=False,
+        )
+
+
+@torch.no_grad()
+def export_onnx(hf_model: str, cache_dir: Optional[str], onnx_path_str: str, with_past: bool, opset: int):
+    """
+    do export
+    model: torch model
+    onnx_path: where the onnx model saved to
+    sample_inputs_tp: inputs for torch model
+    """
+    model, sample_inputs_tp = initialize_model_and_sample_inputs(hf_model, cache_dir)
+
+    model = move_to_approprate_device(model, sample_inputs_tp)
+
+    sample_inputs = adapt_inputs_to_device(sample_inputs_tp, next(model.parameters()).device)
+
+    # input_keys would be usesful if the model has some special inputs
+    input_keys, onnx_inputs, past_key_value = retrieve_onnx_inputs(model, sample_inputs, with_past)
+
+    onnx_io_tuple = fetch_onnx_inputs_outputs_name(model, onnx_inputs, input_keys, past_key_value, with_past, False)
+
+    onnx_model_name = "model.onnx"
+    onnx_path: Path = Path(onnx_path_str).absolute()
+    if onnx_path.suffix != ".onnx":
+        onnx_path = onnx_path / onnx_model_name
+
+    do_export_internal(model, onnx_io_tuple, onnx_inputs, onnx_path, opset)
+    if not with_past:
+        return
+
+    onnx_io_tuple = fetch_onnx_inputs_outputs_name(model, onnx_inputs, input_keys, past_key_value, with_past, True)
+
+    onnx_model_name = "model_with_past.onnx"
+    onnx_path = onnx_path.parent / onnx_model_name
+
+    do_export_internal(model, onnx_io_tuple, onnx_inputs, onnx_path, opset)
+
+
+def parse_arguments():
+    """arguments parsing."""
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "-m",
+        "--model",
+        required=True,
+        type=str,
+        default=["meta-llama/Llama-2-70b-hf"],
+        help="Pre-trained models in huggingface model hub",
+    )
+    parser.add_argument(
+        "-s",
+        "--saved_path",
+        required=False,
+        type=str,
+        default="./onnx_models/",
+        help="where the onnx model will be saved",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        required=False,
+        type=str,
+        default=None,
+        help=("cache directy of huggingface, by setting this to avoid useless downloading if you have one"),
+    )
+    parser.add_argument(
+        "--with_past",
+        action="store_true",
+        default=False,
+        help=("The tool will export onnx without past-key-value by default"),
+    )
+    parser.add_argument(
+        "--opset",
+        required=False,
+        type=int,
+        default=17,
+        help=(
+            "the opset to save onnx model, \
+              try to increase it if this opset doens't have new features you want"
+        ),
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_arguments()
+
+    export_onnx(args.model, args.cache_dir, args.saved_path, args.with_past, args.opset)