Enable Langchain on Intel GPU

docs/docs/integrations/llms/ipex_llm.ipynb

@@ -4,11 +4,11 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# IPEX-LLM\n",
+    "# IPEX-LLM on Intel CPU\n",
     "\n",
     "> [IPEX-LLM](https://github.com/intel-analytics/ipex-llm/) is a PyTorch library for running LLM on Intel CPU and GPU (e.g., local PC with iGPU, discrete GPU such as Arc, Flex and Max) with very low latency. \n",
     "\n",
-    "This example goes over how to use LangChain to interact with `ipex-llm` for text generation. \n"
+    "This example goes over how to use LangChain to interact with `ipex-llm` for text generation on Intel CPU.\n"
    ]
   },
   {

docs/docs/integrations/llms/ipex_llm_gpu.ipynb

@@ -0,0 +1,258 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# IPEX-LLM on Intel GPU\n",
+    "\n",
+    "> [IPEX-LLM](https://github.com/intel-analytics/ipex-llm) is a PyTorch library for running LLM on Intel CPU and GPU (e.g., local PC with iGPU, discrete GPU such as Arc, Flex and Max) with very low latency.\n",
+    "\n",
+    "This example goes over how to use LangChain to interact with `ipex-llm` for text generation on Intel GPU. \n",
+    "\n",
+    "> **Note**\n",
+    ">\n",
+    "> It is recommended that only Windows users with Intel Arc A-Series GPU (except for Intel Arc A300-Series or Pro A60) run this Jupyter notebook directly. For other cases (e.g. Linux users, Intel iGPU, etc.), it is recommended to run the code with Python scripts in terminal for best experiences.\n",
+    "\n",
+    "## Install Prerequisites\n",
+    "To benefit from IPEX-LLM on Intel GPUs, there are several prerequisite steps for tools installation and environment preparation.\n",
+    "\n",
+    "If you are a Windows user, visit the [Install IPEX-LLM on Windows with Intel GPU Guide](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/install_windows_gpu.html), and follow [Install Prerequisites](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/install_windows_gpu.html#install-prerequisites) to update GPU driver (optional) and install Conda.\n",
+    "\n",
+    "If you are a Linux user, visit the [Install IPEX-LLM on Linux with Intel GPU](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/install_linux_gpu.html), and follow [**Install Prerequisites**](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/install_linux_gpu.html#install-prerequisites) to install GPU driver, Intel® oneAPI Base Toolkit 2024.0, and Conda.\n",
+    "\n",
+    "## Setup\n",
+    "\n",
+    "After the prerequisites installation, you should have created a conda environment with all prerequisites installed. **Start the jupyter service in this conda environment**:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "plaintext"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "%pip install -qU langchain langchain-community"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Install IEPX-LLM for running LLMs locally on Intel GPU."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "plaintext"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "%pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "> **Note**\n",
+    ">\n",
+    "> You can also use `https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/` as the extra-indel-url.\n",
+    "\n",
+    "## Runtime Configuration\n",
+    "\n",
+    "For optimal performance, it is recommended to set several environment variables based on your device:\n",
+    "\n",
+    "### For Windows Users with Intel Core Ultra integrated GPU"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "plaintext"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "\n",
+    "os.environ[\"SYCL_CACHE_PERSISTENT\"] = \"1\"\n",
+    "os.environ[\"BIGDL_LLM_XMX_DISABLED\"] = \"1\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### For Windows Users with Intel Arc A-Series GPU"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "plaintext"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "\n",
+    "os.environ[\"SYCL_CACHE_PERSISTENT\"] = \"1\""
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "> **Note**\n",
+    ">\n",
+    "> For the first time that each model runs on Intel iGPU/Intel Arc A300-Series or Pro A60, it may take several minutes to compile.\n",
+    ">\n",
+    "> For other GPU type, please refer to [here](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Overview/install_gpu.html#runtime-configuration) for Windows users, and  [here](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Overview/install_gpu.html#id5) for Linux users.\n",
+    "\n",
+    "\n",
+    "## Basic Usage\n",
+    "\n",
+    "Setting `device` to `\"xpu\"` in `model_kwargs` when initializing `IpexLLM` will put the LLM model on Intel GPU and benefit from IPEX-LLM optimizations:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import warnings\n",
+    "\n",
+    "from langchain.chains import LLMChain\n",
+    "from langchain_community.llms import IpexLLM\n",
+    "from langchain_core.prompts import PromptTemplate\n",
+    "\n",
+    "warnings.filterwarnings(\"ignore\", category=UserWarning, message=\".*padding_mask.*\")\n",
+    "template = \"USER: {question}\\nASSISTANT:\"\n",
+    "prompt = PromptTemplate(template=template, input_variables=[\"question\"])\n",
+    "\n",
+    "llm = IpexLLM.from_model_id(\n",
+    "    model_id=\"lmsys/vicuna-7b-v1.5\",\n",
+    "    model_kwargs={\n",
+    "        \"temperature\": 0,\n",
+    "        \"max_length\": 64,\n",
+    "        \"trust_remote_code\": True,\n",
+    "        \"device\": \"xpu\",\n",
+    "    },\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Use it in Chains"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "llm_chain = prompt | llm\n",
+    "\n",
+    "question = \"What is AI?\"\n",
+    "output = llm_chain.invoke(question)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Save/Load Low-bit Model\n",
+    "Alternatively, you might save the low-bit model to disk once and use `from_model_id_low_bit` instead of `from_model_id` to reload it for later use - even across different machines. It is space-efficient, as the low-bit model demands significantly less disk space than the original model. And `from_model_id_low_bit` is also more efficient than `from_model_id` in terms of speed and memory usage, as it skips the model conversion step. You can similarly set `device` to `xpu` in `model_kwargs` in order to load the LLM model to Intel GPU. "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "To save the low-bit model, use `save_low_bit` as follows."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "saved_lowbit_model_path = \"./vicuna-7b-1.5-low-bit\"  # path to save low-bit model\n",
+    "llm.model.save_low_bit(saved_lowbit_model_path)\n",
+    "del llm"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Load the model from saved lowbit model path as follows. \n",
+    "> Note that the saved path for the low-bit model only includes the model itself but not the tokenizers. If you wish to have everything in one place, you will need to manually download or copy the tokenizer files from the original model's directory to the location where the low-bit model is saved."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "llm_lowbit = IpexLLM.from_model_id_low_bit(\n",
+    "    model_id=saved_lowbit_model_path,\n",
+    "    tokenizer_id=\"lmsys/vicuna-7b-v1.5\",\n",
+    "    # tokenizer_name=saved_lowbit_model_path,  # copy the tokenizers to saved path if you want to use it this way\n",
+    "    model_kwargs={\n",
+    "        \"temperature\": 0,\n",
+    "        \"max_length\": 64,\n",
+    "        \"trust_remote_code\": True,\n",
+    "        \"device\": \"xpu\",\n",
+    "    },\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Use the loaded model in Chains:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "llm_chain = prompt | llm_lowbit\n",
+    "\n",
+    "\n",
+    "question = \"What is AI?\"\n",
+    "output = llm_chain.invoke(question)"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

libs/community/langchain_community/llms/ipex_llm.py

@@ -142,6 +142,16 @@ def _load_model(
         kwargs = kwargs or {}
 
         _tokenizer_id = tokenizer_id or model_id
+        # Set "cpu" as default device
+        if "device" not in _model_kwargs:
+            _model_kwargs["device"] = "cpu"
+
+        if _model_kwargs["device"] not in ["cpu", "xpu"]:
+            raise ValueError(
+                "IpexLLMBgeEmbeddings currently only supports device to be "
+                f"'cpu' or 'xpu', but you have: {_model_kwargs['device']}."
+            )
+        device = _model_kwargs.pop("device")
 
         try:
             tokenizer = AutoTokenizer.from_pretrained(_tokenizer_id, **_model_kwargs)
@@ -189,6 +199,8 @@ def _load_model(
                 model_kwargs=_model_kwargs,
             )
 
+        model.to(device)
+
         return cls(
             model_id=model_id,
             model=model,
@@ -238,6 +250,7 @@ def _call(
             from transformers import TextStreamer
 
             input_ids = self.tokenizer.encode(prompt, return_tensors="pt")
+            input_ids = input_ids.to(self.model.device)
             streamer = TextStreamer(
                 self.tokenizer, skip_prompt=True, skip_special_tokens=True
             )
@@ -264,6 +277,7 @@ def _call(
             return text
         else:
             input_ids = self.tokenizer.encode(prompt, return_tensors="pt")
+            input_ids = input_ids.to(self.model.device)
             if stop is not None:
                 from transformers.generation.stopping_criteria import (
                     StoppingCriteriaList,

libs/community/tests/integration_tests/llms/test_ipex_llm.py

@@ -1,4 +1,5 @@
 """Test IPEX LLM"""
+
 import os
 from typing import Any
 
@@ -12,12 +13,18 @@
     not model_ids_to_test, reason="TEST_IPEXLLM_MODEL_IDS environment variable not set."
 )
 model_ids_to_test = [model_id.strip() for model_id in model_ids_to_test.split(",")]  # type: ignore
+device = os.getenv("TEST_IPEXLLM_BGE_EMBEDDING_MODEL_DEVICE") or "cpu"
 
 
 def load_model(model_id: str) -> Any:
     llm = IpexLLM.from_model_id(
         model_id=model_id,
-        model_kwargs={"temperature": 0, "max_length": 16, "trust_remote_code": True},
+        model_kwargs={
+            "temperature": 0,
+            "max_length": 16,
+            "trust_remote_code": True,
+            "device": device,
+        },
     )
     return llm