Llava optimum ONNX inference

llava_onnx_inference.py

import os
from typing import List, Optional, Tuple

import onnxruntime as onnxrt
import requests
import torch
from PIL import Image
from transformers import AutoConfig, AutoProcessor, GenerationConfig, PreTrainedModel
from transformers.generation import GenerationMixin
from transformers.modeling_outputs import BaseModelOutput, CausalLMOutputWithPast

from optimum.utils import NormalizedConfigManager


os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"


device = torch.device("cpu")


model_name = "llava-1.5-7b-hf/"

processor = AutoProcessor.from_pretrained(model_name)
config = AutoConfig.from_pretrained(model_name)

prompt = "<image>\nUSER: What's the content of the image?\nASSISTANT:"
url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(text=prompt, images=image, return_tensors="pt")


class ORTModel(torch.nn.Module):
    def __init__(self, path, config):
        super().__init__()
        self._device = device
        self.config = config
        self.session = onnxrt.InferenceSession(path, providers=["CPUExecutionProvider"])

        self.input_names = {input_key.name: idx for idx, input_key in enumerate(self.session.get_inputs())}
        self.output_names = {output_key.name: idx for idx, output_key in enumerate(self.session.get_outputs())}


class ORTEncoder(ORTModel):
    def forward(
        self,
        input_ids: torch.FloatTensor,
        pixel_values: torch.FloatTensor,
        attention_mask: torch.LongTensor,
        **kwargs,
    ) -> BaseModelOutput:
        onnx_inputs = {
            "input_ids": input_ids.cpu().detach().numpy(),
            "pixel_values": pixel_values.cpu().detach().numpy(),
            "attention_mask": attention_mask.cpu().detach().numpy(),
        }

        # Run inference
        outputs = self.session.run(None, onnx_inputs)

        for i, output in enumerate(outputs):
            outputs[i] = torch.from_numpy(output).to(self._device)

        return (
            outputs[self.output_names["inputs_embeds"]],
            outputs[self.output_names["decoder_attention_mask"]],
            outputs[self.output_names["position_ids"]],
        )


class ORTDecoderProcessor(ORTModel):
    def forward(
        self,
        input_ids: torch.FloatTensor,
        attention_mask: torch.LongTensor,
        past_key_value: torch.FloatTensor,
        **kwargs,
    ) -> BaseModelOutput:
        onnx_inputs = {
            "input_ids": input_ids.cpu().detach().numpy(),
            "attention_mask": attention_mask.cpu().detach().numpy(),
            "past_key_values.0.key": past_key_value.cpu().detach().numpy(),
        }

        # Run inference
        outputs = self.session.run(None, onnx_inputs)

        for i, output in enumerate(outputs):
            outputs[i] = torch.from_numpy(output).to(self._device)

        return (
            outputs[self.output_names["inputs_embeds"]],
            outputs[self.output_names["decoder_attention_mask"]],
            outputs[self.output_names["position_ids"]],
        )


class ORTDecoder(ORTModel):
    def __init__(self, path, config):
        super().__init__(path, config)

        self.normalized_config = NormalizedConfigManager.get_normalized_config_class(config.text_config.model_type)(
            config.text_config
        )
        self.generation_config = GenerationConfig.from_model_config(config)

        self.key_value_input_names = [key for key in self.input_names if (".key" in key) or (".value" in key)]
        self.key_value_output_names = [key for key in self.output_names if (".key" in key) or (".value" in key)]

        self.num_pkv = 2

    def prepare_pkv(self, batch_size: int):
        if self.config.text_config.model_type in {"mistral", "llama"}:
            num_attention_heads = self.normalized_config.num_key_value_heads
        else:
            num_attention_heads = self.normalized_config.num_attention_heads

        embed_size_per_head = self.normalized_config.hidden_size // self.normalized_config.num_attention_heads

        shape = (batch_size, num_attention_heads, 0, embed_size_per_head)
        key_or_value = torch.zeros(shape, dtype=torch.float32)

        past_key_values = tuple(key_or_value for _ in range(len(self.key_value_input_names)))

        return past_key_values

    def forward(
        self,
        attention_mask: torch.LongTensor,
        position_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        past_key_values: Tuple[Tuple[torch.FloatTensor]] = None,
    ) -> CausalLMOutputWithPast:
        onnx_inputs = {
            "attention_mask": attention_mask.cpu().detach().numpy(),
            "position_ids": position_ids.cpu().detach().numpy(),
            "inputs_embeds": inputs_embeds.cpu().detach().numpy(),
        }

        if past_key_values is None:
            past_key_values = self.prepare_pkv(inputs_embeds.shape[0])
        else:
            past_key_values = tuple(
                past_key_value for pkv_per_layer in past_key_values for past_key_value in pkv_per_layer
            )

        for input_name, past_key_value in zip(self.key_value_input_names, past_key_values):
            onnx_inputs[input_name] = past_key_value.cpu().detach().numpy()

        # Run inference
        outputs = self.session.run(None, onnx_inputs)

        logits = torch.from_numpy(outputs[self.output_names["logits"]])

        past_key_values = tuple(
            torch.from_numpy(outputs[self.output_names[key]]) for key in self.key_value_output_names
        )

        past_key_values = tuple(
            past_key_values[i : i + self.num_pkv] for i in range(0, len(past_key_values), self.num_pkv)
        )

        return CausalLMOutputWithPast(logits=logits, past_key_values=past_key_values)


class ORTModelForLLava(PreTrainedModel, GenerationMixin):
    def __init__(self, *args, **kwargs):
        config = AutoConfig.from_pretrained(model_name)
        super().__init__(config)

        self.config = config
        self._device = device

        self.vision_tower = ORTEncoder(model_name + "encoder_model.onnx", config)
        self.language_model = ORTDecoder(model_name + "decoder_model.onnx", config)
        self.decoder_input_processor = ORTDecoderProcessor(model_name + "decoder_input_processor_model.onnx", config)

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        **kwargs,
    ) -> CausalLMOutputWithPast:
        if past_key_values is None:
            inputs_embeds, attention_mask, position_ids = self.vision_tower(
                input_ids=input_ids,
                pixel_values=pixel_values,
                attention_mask=attention_mask,
            )
        else:
            inputs_embeds, attention_mask, position_ids = self.decoder_input_processor(
                input_ids=input_ids,
                attention_mask=attention_mask,
                past_key_value=past_key_values[0][0][:, :, :, 0],
            )

        # Decode
        decoder_outputs = self.language_model(
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
        )

        return decoder_outputs

    def can_generate(self):
        return True

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, inputs_embeds=None, pixel_values=None, attention_mask=None, **kwargs
    ):
        if past_key_values is not None:
            cache_length = past_length = past_key_values[0][0].shape[2]

            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            elif self.config.image_token_index in input_ids:
                input_ids = input_ids[:, input_ids.shape[1] - 1 :]
            if cache_length < past_length and attention_mask is not None:
                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]

        model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "pixel_values": pixel_values,
            }
        )
        return model_inputs

    @property
    def device(self) -> torch.device:
        return self._device

    @device.setter
    def device(self, value: torch.device):
        self._device = value

    def to(self, device):
        self.device = device
        return self


model = ORTModelForLLava()

generated_ids = model.generate(**inputs, max_length=30)
out = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

print(out)

thanks

hello, thanks for your share! @mht-sharma ,Are the onnx files of vision_tower and language_model saved separately through llava?