bring back runner-et test (#266)

* bring back runner-et test

* mac

* return state to before for sanity check

* change to using Util functions for x-platform support.  Add mac/linux runners

* mac

Author: metascroy

.github/workflows/pull.yml

@@ -679,3 +679,54 @@ jobs:
 
           echo "tests complete"
           echo "******************************************"
+  runner-et:
+    strategy:
+      matrix:
+        runner: [macos-14-xlarge]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout repo
+        uses: actions/checkout@v2
+      - name: Setup Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: 3.11
+      - name: Print machine info
+        run: |
+          uname -a
+          if [ $(uname -s) == Darwin ]; then
+            sysctl machdep.cpu.brand_string
+            sysctl machdep.cpu.core_count
+          fi
+      - name: Install requirements
+        run: |
+          echo "Intalling pip packages"
+          pip install -r requirements.txt
+
+          export TORCHCHAT_ROOT=${PWD}
+          export ENABLE_ET_PYBIND=false
+          ./scripts/install_et.sh $ENABLE_ET_PYBIND
+          cmake -S ./runner-et -B et-build/cmake-out -G Ninja
+          cmake --build ./et-build/cmake-out
+      - name: Download checkpoints
+        run: |
+          mkdir -p checkpoints/stories15M
+          pushd checkpoints/stories15M
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories15M.pt
+          wget https://github.com/karpathy/llama2.c/raw/master/tokenizer.model
+          wget https://github.com/karpathy/llama2.c/raw/master/tokenizer.bin
+          popd
+      - name: Run inference
+        run: |
+          export MODEL_DIR=${PWD}/checkpoints/stories15M
+          export PROMPT="Once upon a time in a land far away"
+
+          python torchchat.py generate --checkpoint-path ${MODEL_DIR}/stories15M.pt --temperature 0 --prompt "${PROMPT}"  > ${PWD}/output_eager
+          cat ${PWD}/output_eager
+
+          python torchchat.py export --checkpoint-path ${MODEL_DIR}/stories15M.pt --output-pte-path ${PWD}/stories15M.pte
+
+          ./et-build/cmake-out/runner_et ${PWD}/stories15M.pte -z ${MODEL_DIR}/tokenizer.bin -i "${PROMPT}" > ${PWD}/output_et
+          cat ${PWD}/output_et
+
+          echo "Tests complete."

requirements.txt

@@ -21,3 +21,9 @@ numpy
 gguf
 lm-eval
 blobfile
+
+# Build tools
+wheel
+cmake
+ninja
+zstd

runner-et/CMakeLists.txt

@@ -4,6 +4,7 @@ set(CMAKE_CXX_STANDARD 17)
 project(Torchchat)
 
 include(CMakePrintHelpers)
+include(Utils.cmake)
 set(TORCHCHAT_ROOT $ENV{TORCHCHAT_ROOT})
 cmake_print_variables(TORCHCHAT_ROOT)
 
@@ -24,13 +25,15 @@ target_link_libraries(
         portable_kernels
         cpublas
         eigen_blas
+        # The libraries below need to be whole-archived linked
         optimized_native_cpu_ops_lib
+        xnnpack_backend
+        XNNPACK
+        pthreadpool
+        cpuinfo
 )
-target_link_options(
-  optimized_native_cpu_ops_lib INTERFACE "SHELL:LINKER:--whole-archive \
-  $<TARGET_FILE:optimized_native_cpu_ops_lib> \
-  LINKER:--no-whole-archive")
-target_link_libraries(runner_et PRIVATE "$<LINK_LIBRARY:WHOLE_ARCHIVE,xnnpack_backend">)
-target_link_libraries(runner_et PRIVATE "$<LINK_LIBRARY:WHOLE_ARCHIVE,XNNPACK">)
-target_link_libraries(runner_et PRIVATE "$<LINK_LIBRARY:WHOLE_ARCHIVE,pthreadpool">)
-target_link_libraries(runner_et PRIVATE "$<LINK_LIBRARY:WHOLE_ARCHIVE,cpuinfo">)
+target_link_options_shared_lib(optimized_native_cpu_ops_lib)
+target_link_options_shared_lib(xnnpack_backend)
+target_link_options_shared_lib(XNNPACK)
+target_link_options_shared_lib(pthreadpool)
+target_link_options_shared_lib(cpuinfo)

runner-et/Utils.cmake

@@ -0,0 +1,33 @@
+# 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.
+
+# This is the funtion to use -Wl, --whole-archive to link static library NB:
+# target_link_options is broken for this case, it only append the interface link
+# options of the first library.
+function(kernel_link_options target_name)
+  # target_link_options(${target_name} INTERFACE
+  # "$<LINK_LIBRARY:WHOLE_ARCHIVE,target_name>")
+  target_link_options(
+    ${target_name} INTERFACE "SHELL:LINKER:--whole-archive \
+    $<TARGET_FILE:${target_name}> \
+    LINKER:--no-whole-archive")
+endfunction()
+
+# Same as kernel_link_options but it's for MacOS linker
+function(macos_kernel_link_options target_name)
+  target_link_options(${target_name} INTERFACE
+                      "SHELL:LINKER:-force_load,$<TARGET_FILE:${target_name}>")
+endfunction()
+
+# Ensure that the load-time constructor functions run. By default, the linker
+# would remove them since there are no other references to them.
+function(target_link_options_shared_lib target_name)
+  if(APPLE)
+    macos_kernel_link_options(${target_name})
+  else()
+    kernel_link_options(${target_name})
+  endif()
+endfunction()