onnx/builder.py

import argparse
import onnx
import os
import requests
import shutil
import subprocess
import sys
import torch

from onnxruntime_genai.models.builder import create_model
from PIL import Image
from transformers import AutoConfig, AutoProcessor, AutoModelForCausalLM


def build_vision(args):
    # Single image:
    prompt = f"{user_prompt}<|image_1|>\nWhat is shown in this image?{prompt_suffix}{assistant_prompt}"
    url = "https://www.ilankelman.org/stopsigns/australia.jpg"
    image = Image.open(requests.get(url, stream=True).raw)
    inputs = processor(prompt, image, return_tensors="pt").to(args.execution_provider)
    inputs["pixel_values"] = inputs["pixel_values"].to(args.precision)

    # TorchScript export
    dummy_inputs = (
        inputs["pixel_values"],   # input_embeds: Optional[torch.FloatTensor] = None,
        inputs["image_sizes"],    # image_sizes: Optional[torch.FloatTensor] = None,
    )
    dynamic_axes = {
        "pixel_values": {0: "num_images", 1: "max_num_crops", 3: "height", 4: "width"},
        "image_sizes": {0: "num_images"},
        "visual_features": {0: "batch_size", 1: "num_img_tokens"},
    }
    filename = "phi-3-v-128k-instruct-vision.onnx"

    temp_folder_1 = os.path.join(args.output, "vision_init_export")
    os.makedirs(temp_folder_1, exist_ok=True)

    fpath_1 = os.path.join(temp_folder_1, filename)
    torch.onnx.export(
        model.model.vision_embed_tokens,
        args=dummy_inputs,
        f=fpath_1,
        export_params=True,
        input_names=["pixel_values", "image_sizes"],
        output_names=["visual_features"],
        dynamic_axes=dynamic_axes,
        opset_version=14,
        do_constant_folding=True,
    )

    onnx.checker.check_model(fpath_1)
    onnx.shape_inference.infer_shapes_path(fpath_1)
    onnx_model = onnx.load_model(fpath_1, load_external_data=True)

    temp_folder_2 = os.path.join(args.output, "vision_after_export")
    os.makedirs(temp_folder_2, exist_ok=True)

    fpath_2 = os.path.join(temp_folder_2, filename)
    onnx.save_model(
        onnx_model,
        fpath_2,
        save_as_external_data=True,
        all_tensors_to_one_file=True,
        location=f"{filename}.data",
        size_threshold=0,
        convert_attribute=False,
    )
    shutil.rmtree(temp_folder_1)

    # ORT transformer optimizer
    temp_folder_3 = os.path.join(args.output, "vision_after_opt")
    fpath_3 = os.path.join(temp_folder_3, filename)
    subprocess.run(
        [
            f"{sys.executable}", "-m", "onnxruntime.transformers.optimizer",
            "--input", fpath_2,
            "--output", fpath_3,
            "--model_type", "clip",
            "--num_heads", str(16),
            "--hidden_size", str(1024),
            "--use_external_data_format",
            "--opt_level", str(0),
        ]
    )
    shutil.rmtree(temp_folder_2)

    # ORT 4-bits quantizer
    fpath_4 = os.path.join(args.output, filename)
    cmd = [
        f"{sys.executable}", "-m", "onnxruntime.quantization.matmul_4bits_quantizer",
        "--input_model", fpath_3,
        "--output_model", fpath_4,
        "--block_size", str(32),
    ]
    if args.precision == torch.float32: cmd.extend(["--accuracy_level", str(4)])
    subprocess.run(cmd)
    shutil.rmtree(temp_folder_3)

def build_text_embedding(args):
    #########################################
    # Functions/variables from model builder
    #########################################
    from onnx import helper, numpy_helper, TensorProto, external_data_helper, save_model
    import numpy as np
    
    # User inputs
    io_dtype = TensorProto.FLOAT16 if args.precision == torch.float16 else TensorProto.FLOAT
    os.makedirs(args.cache_dir, exist_ok=True)

    # Map TensorProto dtypes
    to_torch_dtype = {
        TensorProto.FLOAT16: torch.float16,
        TensorProto.FLOAT: torch.float32,
    }
    to_numpy_dtype = {
        TensorProto.FLOAT16: np.float16,
        TensorProto.FLOAT: np.float32,
    }

    def make_external_tensor(np_data, name, **kwargs):
        tensor = numpy_helper.from_array(np_data)
        tensor.name = name

        filename = f"{name}.bin"
        external_data_helper.set_external_data(tensor, location=filename)
        with open(os.path.join(args.cache_dir, filename), "wb") as f:
            f.write(tensor.raw_data)
        tensor.ClearField("raw_data")
        tensor.data_location = TensorProto.EXTERNAL

        return tensor

    # Make model
    global model
    embedding = model.model.embed_tokens.weight.to(to_torch_dtype[io_dtype]).detach().cpu().numpy()
    weight_name = "model.embed_tokens.weight"
    embed_weight = make_external_tensor(embedding.astype(to_numpy_dtype[io_dtype]), weight_name)
    model = helper.make_model(
        opset_imports=[helper.make_operatorsetid('', 14), helper.make_operatorsetid('com.microsoft', 1)],
        ir_version=7,
        producer_name="onnxruntime-genai",
        producer_version="0.0.0",
        graph=helper.make_graph(
            name="main_graph",
            inputs=[helper.make_tensor_value_info("input_ids", TensorProto.INT64, shape=["batch_size", "sequence_length"])],
            outputs=[helper.make_tensor_value_info("inputs_embeds", io_dtype, shape=["batch_size", "sequence_length", config.hidden_size])],
            initializer=[embed_weight],
            value_info=[],
            nodes=[helper.make_node('Gather', inputs=[weight_name, 'input_ids'], outputs=['inputs_embeds'], name="/model/embed_tokens/Gather")],
        )
    )

    external_data_helper.load_external_data_for_model(model, args.cache_dir)

    # Delete external data files on disk before re-saving
    for path in os.listdir(args.cache_dir):
        if path.endswith(".bin"):
            os.remove(os.path.join(args.cache_dir, path))

    # Delete temporary cache dir if empty
    if len(os.listdir(args.cache_dir)) == 0:
        os.rmdir(args.cache_dir)

    # Save ONNX model with only one external data file and delete any existing duplicate copies
    filename = "phi-3-v-128k-instruct-text-embedding.onnx"
    output_path = os.path.join(args.output, filename)    
    save_model(
        model,
        output_path,
        save_as_external_data=True,
        all_tensors_to_one_file=True,
        location=f"{filename}.data",
        size_threshold=0,
        convert_attribute=False,
    )

def build_text(args):
    # Create ONNX model
    model_name = None
    precision = "int4"
    extra_options = {
        "exclude_embeds": "true",
        "filename": "phi-3-v-128k-instruct-text.onnx",
    }
    if args.precision == torch.float32: extra_options["int4_accuracy_level"] = 4
    create_model(model_name, args.input, args.output, precision, args.execution_provider, args.cache_dir, **extra_options)

def get_args():
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "-i",
        "--input",
        required=True,
        help="Path to folder on disk containing the Hugging Face config, model, tokenizer, etc.",
    )

    parser.add_argument(
        "-o",
        "--output",
        required=True,
        help="Path to folder to store ONNX model and additional files (e.g. GenAI config, external data files, etc.)",
    )

    parser.add_argument(
        "-p",
        "--precision",
        required=True,
        choices=["fp16", "fp32"],
        help="Precision to export PyTorch components with",
    )

    parser.add_argument(
        "-e",
        "--execution_provider",
        required=True,
        choices=["cpu", "cuda"],
        help="Device to export Phi-3 vision components with",
    )

    parser.add_argument(
        "-c",
        "--cache_dir",
        required=False,
        default=os.path.join('.', 'cache_dir'),
        help="Cache directory for Hugging Face files and temporary ONNX external data files",
    )

    args = parser.parse_args()
    args.precision = torch.float16 if args.precision == "fp16" else torch.float32
    return args

if __name__ == "__main__":
    user_prompt = '<|user|>\n'
    assistant_prompt = '<|assistant|>\n'
    prompt_suffix = "<|end|>\n"

    args = get_args()
    config = AutoConfig.from_pretrained(args.input, trust_remote_code=True)
    processor = AutoProcessor.from_pretrained(args.input, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(args.input, trust_remote_code=True, torch_dtype=args.precision).to(args.execution_provider)

    # Build model components
    build_vision(args)
    build_text_embedding(args)
    build_text(args)