{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "\n\n# Torch-TensorRT Distributed Inference\n\nThis interactive script is intended as a sample of distributed inference using data\nparallelism using Accelerate\nlibrary with the Torch-TensorRT workflow on GPT2 model.\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Imports and Model Definition\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "import torch\nimport torch_tensorrt\nfrom accelerate import PartialState\nfrom transformers import AutoTokenizer, GPT2LMHeadModel\n\ntokenizer = AutoTokenizer.from_pretrained(\"gpt2\")\n\n# Set input prompts for different devices\nprompt1 = \"GPT2 is a model developed by.\"\nprompt2 = \"Llama is a model developed by \"\n\ninput_id1 = tokenizer(prompt1, return_tensors=\"pt\").input_ids\ninput_id2 = tokenizer(prompt2, return_tensors=\"pt\").input_ids\n\ndistributed_state = PartialState()\n\n# Import GPT2 model and load to distributed devices\nmodel = GPT2LMHeadModel.from_pretrained(\"gpt2\").eval().to(distributed_state.device)\n\n\n# Instantiate model with Torch-TensorRT backend\nmodel.forward = torch.compile(\n    model.forward,\n    backend=\"torch_tensorrt\",\n    options={\n        \"truncate_long_and_double\": True,\n        \"enabled_precisions\": {torch.float16},\n    },\n    dynamic=False,\n)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "## Inference\n\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "# Assume there are 2 processes (2 devices)\nwith distributed_state.split_between_processes([input_id1, input_id2]) as prompt:\n    cur_input = torch.clone(prompt[0]).to(distributed_state.device)\n\n    gen_tokens = model.generate(\n        cur_input,\n        do_sample=True,\n        temperature=0.9,\n        max_length=100,\n    )\n    gen_text = tokenizer.batch_decode(gen_tokens)[0]"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.11.14"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}