add CI job for phi-3-mini

.ci/scripts/test_phi_3_mini.sh

@@ -0,0 +1,116 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -exu
+
+BUILD_TYPE=${1:-Debug}
+BUILD_DIR=${3:-cmake-out}
+MODEL_DIR=examples/models/phi-3-mini
+
+echo "Building with BUILD_TYPE: $BUILD_TYPE, BUILD_DIR: $BUILD_DIR"
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+    PYTHON_EXECUTABLE=python3
+fi
+
+# Number of processes for a parallel build
+NPROC=8
+if hash nproc &> /dev/null; then NPROC=$(nproc); fi
+
+cmake_install_executorch_libraries() {
+  cmake -DPYTHON_EXECUTABLE=python \
+      -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
+      -DEXECUTORCH_ENABLE_LOGGING=1 \
+      -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+      -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+      -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+      -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+      -DEXECUTORCH_BUILD_XNNPACK=ON \
+      -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+      -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+      -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+      -B${BUILD_DIR} .
+
+  cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${BUILD_TYPE}
+}
+
+cmake_build_phi_3_mini() {
+  cmake -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
+      -DCMAKE_INSTALL_PREFIX=${BUILD_DIR} \
+      -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+      -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+      -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+      -DEXECUTORCH_BUILD_XNNPACK=ON \
+      -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+      -B${BUILD_DIR}/${MODEL_DIR} \
+      ${MODEL_DIR}
+
+  cmake --build ${BUILD_DIR}/${MODEL_DIR} -j${NPROC} --config ${BUILD_TYPE}
+}
+
+# Download and convert tokenizer.model
+prepare_tokenizer() {
+  echo "Downloading and converting tokenizer.model"
+  wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/tokenizer.model?download=true"
+  $PYTHON_EXECUTABLE -m executorch.extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+}
+
+# Export phi-3-mini model to pte
+export_phi_3_mini () {
+  echo "Exporting phi-3-mini. This will take a few minutes"
+  $PYTHON_EXECUTABLE -m executorch.examples.models.phi-3-mini.export_phi-3-mini -c "4k" -s 128 -o phi-3-mini.pte
+}
+
+run_and_verify() {
+    NOW=$(date +"%H:%M:%S")
+    echo "Starting to run phi-3-mini runner at ${NOW}"
+    if [[ ! -f "phi-3-mini.pte" ]]; then
+        echo "Export failed. Abort"
+        exit 1
+    fi
+    if [[ ! -f "tokenizer.bin" ]]; then
+        echo "tokenizer.bin is missing."
+        exit 1
+    fi
+
+    ${BUILD_DIR}/${MODEL_DIR}/phi_3_mini_runner \
+    --model_path=phi-3-mini.pte \
+    --tokenizer_path=tokenizer.bin \
+    --seq_len=128 \
+    --temperature=0 \
+    --prompt="<|system|>
+You are a helpful assistant.<|end|>
+<|user|>
+What is the capital of France?<|end|>
+<|assistant|>" > result.txt
+
+    # verify result.txt
+    RESULT=$(cat result.txt)
+    EXPECTED_RESULT="The capital of France is Paris."
+    if [[ "${RESULT}" == *"${EXPECTED_RESULT}"* ]]; then
+        echo "Expected result prefix: ${EXPECTED_RESULT}"
+        echo "Actual result: ${RESULT}"
+        echo "Success"
+        exit 0
+    else
+        echo "Expected result prefix: ${EXPECTED_RESULT}"
+        echo "Actual result: ${RESULT}"
+        echo "Failure; results not the same"
+        exit 1
+    fi
+}
+
+# Step 1. Build ExecuTorch and phi-3-mini runner
+cmake_install_executorch_libraries
+cmake_build_phi_3_mini
+
+# Step 2. Export the tokenizer and model
+prepare_tokenizer
+export_phi_3_mini
+
+# Step 3. Run and verify result
+run_and_verify

.github/workflows/pull.yml

@@ -414,3 +414,30 @@ jobs:
         PYTHON_EXECUTABLE=python bash examples/models/llama2/install_requirements.sh
         # Test llama2
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}"
+
+  test-phi-3-mini-runner-linux:
+    name: test-phi-3-mini-runner-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.24xlarge
+      docker-image: executorch-ubuntu-22.04-clang12
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
+
+        # install pybind
+        bash install_requirements.sh --pybind xnnpack
+
+        # install phi-3-mini requirements
+        bash examples/models/phi-3-mini/install_requirements.sh
+
+        # run e2e (export, tokenizer and runner)
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_phi_3_mini.sh

examples/models/phi-3-mini/install_requirements.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -x
+
+pip install transformers==4.44.2
+
+pip install sentencepiece
+
+pip list

examples/models/phi-3-mini/runner.cpp

@@ -54,17 +54,8 @@ void Runner::generate(const std::string& prompt, std::size_t max_seq_len) {
   ET_CHECK_MSG(
       encode_res.error() == Error::Ok, "Failed to encode %s", prompt.c_str());
   auto input_tokens = encode_res.get();
-
-  std::cout << "Prefilling tokens ..." << std::endl;
-  for (auto token : input_tokens) {
-    std::cout << token << " ";
-  }
-  std::cout << std::endl;
-  std::cout.flush();
   auto prev_token = input_tokens.back();
   auto current_token = prefill(input_tokens);
-
-  std::cout << "Generating tokens ..." << std::endl;
   std::cout << tokenizer_->decode(prev_token, current_token).get();
   std::cout.flush();