Add LoRA adapters for the SLM Engine (microsoft#1481)

This PR addresses the following:

- Modified APIs for creating the engine and running generation, more KPI
reporting
 - New APIs supporting LoRA adapters 
    - See slm_engine.h to learn about the new APIs
 - Build script updates for building using Qualcomm QNN dependencies.
 - More unit tests - especially as examples of the LoRA APIs

---------

Co-authored-by: Avijit <[email protected]>

Author: avijit-chakroborty

examples/slm_engine/README.md

@@ -374,7 +374,7 @@ target_link_libraries(inference_server slm_engine ort ort_genai)
 
 int main(int argc, char **argv) {
 
-    auto slm_engine = microsoft::slm_engine::SLMEngine::CreateEngine(
+    auto slm_engine = microsoft::slm_engine::SLMEngine::Create(
         "path to ONNX Model Directory", "phi", true);
 
     if (!slm_engine) {

examples/slm_engine/build_scripts/build_android.sh

@@ -22,14 +22,40 @@ set -u
 # Build the docker image 
 docker build -t slm-engine-builder -f Dockerfile .
 
-# Run the docker to build dependencies
-docker run --rm -v \
-    `pwd`/../../../:`pwd`/../../../  \
-    -u $(id -u):$(id -g) -w `pwd`  \
-    slm-engine-builder python3 build_deps.py \
+# Define base build_deps command
+BUILD_DEPS_CMD="python3 build_deps.py \
     --build_ort_from_source \
     --android_sdk_path /opt/android-sdk/ \
-    --android_ndk_path /opt/android-sdk/ndk/27.2.12479018/ 
+    --android_ndk_path /opt/android-sdk/ndk/27.2.12479018/"
+
+# Docker volume mount options
+VOLUME_MOUNTS="-v `pwd`/../../../:`pwd`/../../../"
+
+# Check if USE_ORT_VERSION is defined
+if [ ! -z "${USE_ORT_VERSION:-}" ]; then
+    BUILD_DEPS_CMD="$BUILD_DEPS_CMD --ort_version_to_use $USE_ORT_VERSION"
+    echo "Using ONNX Runtime version: $USE_ORT_VERSION"
+fi
+
+# Check if QNN_SDK_HOME is defined
+if [ ! -z "${QNN_SDK_HOME:-}" ]; then
+    # Create Docker mount point for QNN SDK
+    QNN_SDK_DOCKER_PATH="/opt/qnn_sdk"
+    
+    # Add mount for QNN SDK
+    VOLUME_MOUNTS="$VOLUME_MOUNTS -v $QNN_SDK_HOME:$QNN_SDK_DOCKER_PATH"
+    
+    # Use the Docker path in build command
+    BUILD_DEPS_CMD="$BUILD_DEPS_CMD --qnn_sdk_path $QNN_SDK_DOCKER_PATH"
+    
+    echo "QNN SDK path detected, building with QNN support"
+    echo "Mounting $QNN_SDK_HOME to $QNN_SDK_DOCKER_PATH in container"
+fi
+
+# Run the docker to build dependencies
+docker run --rm $VOLUME_MOUNTS \
+    -u $(id -u):$(id -g) -w `pwd` \
+    slm-engine-builder $BUILD_DEPS_CMD
 
 # Next build the slm_engine
 docker run --rm -v \

examples/slm_engine/build_scripts/build_deps.py

@@ -228,7 +228,11 @@ def build_ort(args, build_dir, artifacts_dir):
             ]
         )
         if args.qnn_sdk_path:
-            cmd_args.extend(["--use_qnn", "--qnn_home", args.qnn_sdk_path])
+            cmd_args.extend(
+                ["--use_qnn", "static_lib", "--qnn_home", args.qnn_sdk_path]
+            )
+
+    cmd_args.extend(["--cmake_extra_defines", "onnxruntime_BUILD_UNIT_TESTS=OFF"])
 
     # now build the ORT library
     print(f"{MAGENTA}Building ONNX Runtime{CLEAR}")
@@ -587,9 +591,9 @@ def main():
     ort_home = None
     if args.build_ort_from_source:
         if args.ort_version_to_use is None:
-            # If not Windows then use 1.20.1
+            # If not Windows then use 1.22.0
             if platform.system() != "Windows":
-                args.ort_version_to_use = "v1.20.1"
+                args.ort_version_to_use = "v1.22.0"
             else:
                 args.ort_version_to_use = "main"
         ort_home = build_ort(args, dep_src_dir, artifacts_dir)
@@ -600,7 +604,7 @@ def main():
             # The ORT binaries are available as they were downloaded during the GenAI build
             # This is the supported version for most platforms
             if args.ort_version_to_use is None:
-                ORT_VERSION = "1.20.1"
+                ORT_VERSION = "1.22.0"
             else:
                 ORT_VERSION = args.ort_version_to_use
             # Copy the ORT artifacts to the artifacts directory.

examples/slm_engine/build_scripts/docker_shell.sh

@@ -0,0 +1,59 @@
+#!/bin/sh
+
+# This script builds the slm_engine for Android using docker.
+# It uses the Dockerfile in the current directory to build a docker image
+# that contains all the necessary dependencies for building the slm_engine.
+# The script then runs the docker image to build the slm_engine.
+# The script assumes that the Dockerfile is in the same directory as this script.
+# The script also assumes that the android-sdk and android-ndk are installed
+# in the /opt/android-sdk directory.
+# 
+
+# Check the architecture
+if [ "$(uname -m)" != "x86_64" ]; then
+    echo "This script is intended to run on x86_64 architecture only."
+    exit 1
+fi
+
+set -e
+set -x
+set -u
+
+# Build the docker image 
+docker build -t slm-engine-builder -f Dockerfile .
+
+# Define base build_deps command
+BUILD_DEPS_CMD="python3 build_deps.py \
+    --build_ort_from_source \
+    --android_sdk_path /opt/android-sdk/ \
+    --android_ndk_path /opt/android-sdk/ndk/27.2.12479018/"
+
+# Docker volume mount options
+VOLUME_MOUNTS="-v `pwd`/../../../:`pwd`/../../../"
+
+# Check if USE_ORT_VERSION is defined
+if [ ! -z "${USE_ORT_VERSION:-}" ]; then
+    BUILD_DEPS_CMD="$BUILD_DEPS_CMD --ort_version_to_use $USE_ORT_VERSION"
+    echo "Using ONNX Runtime version: $USE_ORT_VERSION"
+fi
+
+# Check if QNN_SDK_HOME is defined
+if [ ! -z "${QNN_SDK_HOME:-}" ]; then
+    # Create Docker mount point for QNN SDK
+    QNN_SDK_DOCKER_PATH="/opt/qnn_sdk"
+    
+    # Add mount for QNN SDK
+    VOLUME_MOUNTS="$VOLUME_MOUNTS -v $QNN_SDK_HOME:$QNN_SDK_DOCKER_PATH"
+    
+    # Use the Docker path in build command
+    BUILD_DEPS_CMD="$BUILD_DEPS_CMD --qnn_sdk_path $QNN_SDK_DOCKER_PATH"
+    
+    echo "QNN SDK path detected, building with QNN support"
+    echo "Mounting $QNN_SDK_HOME to $QNN_SDK_DOCKER_PATH in container"
+fi
+
+# Run the docker to build dependencies
+docker run --rm $VOLUME_MOUNTS \
+    -u $(id -u):$(id -g) -w `pwd` \
+    -w $HOME \
+    -it slm-engine-builder bash

examples/slm_engine/src/VERSION.txt

@@ -1 +1 @@
-1.0.0
+2.0.0

examples/slm_engine/src/cpp/gtest_main.cpp

@@ -1,10 +1,48 @@
+
 #include "gtest/gtest.h"
+#include <argparse/argparse.hpp>
+
 #include "httplib.h"
 #include "ort_genai.h"
 
+using namespace std;
+
+extern const char* MODEL_FILE_PATH;
+extern const char* ADAPTER_ROOT_DIR;
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
 
+  argparse::ArgumentParser program("slm_engine_test", "1.0",
+                                   argparse ::default_arguments::none);
+  string model_path;
+  program.add_argument("-m", "--model_path")
+      .help("Path to the model file")
+      .store_into(model_path);
+
+  string adapter_root_path;
+  program.add_argument("-m", "--adapter_root_path")
+      .help("Path to the LoRA adapter root directory")
+      .store_into(adapter_root_path);
+
+  try {
+    program.parse_args(argc, argv);
+  } catch (const std::exception& err) {
+    std::cerr << err.what() << std::endl;
+    std::cerr << program;
+    std::exit(-1);
+  }
+
+  if (!model_path.empty()) {
+    cout << "Setting Model path: " << model_path << endl;
+    MODEL_FILE_PATH = model_path.c_str();
+  }
+
+  if (!adapter_root_path.empty()) {
+    cout << "Setting Adapter path: " << adapter_root_path << endl;
+    ADAPTER_ROOT_DIR = adapter_root_path.c_str();
+  }
+
   auto status = RUN_ALL_TESTS();
 
   OgaShutdown();

examples/slm_engine/src/cpp/input_decoder.cpp

@@ -16,6 +16,7 @@ namespace slm_engine {
 // clang-format off
 // OpenAI API example
 // {
+//     "model": "name-of-the-adapter (optional)", 
 //     "messages": [
 // 		{
 // 			"role": "system",
@@ -59,6 +60,9 @@ class OpenAIInputDecoder : public InputDecoder {
           return false;
         }
       }
+      if (json_msg.contains("model")) {
+        decoded_params.LoRAAdapterName = json_msg["model"].get<string>();
+      }
       if (json_msg.contains("temperature")) {
         decoded_params.Temperature =
             json_msg["temperature"].get<float_t>();

examples/slm_engine/src/cpp/input_decoder.h

@@ -42,6 +42,9 @@ class InputDecoder {
     std::vector<std::pair<Role, std::string>> Messages;
     // The user prompt is the last message in the sequence
     std::string UserPrompt;
+    // The LoRAAdapterName is sent by the client as "model" in the
+    // OpenAI API. In our implementation, this is the name of the adapter that will be used
+    std::string LoRAAdapterName;
     uint32_t MaxGeneratedTokens;
     std::vector<std::string> StopTokens;
     float Temperature;