Add quantized op support to llama runner (#3062)

Summary: Pull Request resolved: #3062

Reviewed By: lucylq, mikekgfb

Differential Revision: D56197863

fbshipit-source-id: c564a99d10be70fb69e554687bd506d8ff13268e

Author: larryliu0820

.ci/scripts/test_llama.sh

@@ -12,7 +12,7 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 MODEL_NAME=$1 # stories110M.pt
 BUILD_TOOL=$2 # buck2 or cmake
 DTYPE=$3 # fp16 or fp32
-MODE=${4:-"xnnpack"} # portable or xnnpack
+MODE=${4:-"xnnpack+custom"} # portable or xnnpack+custom or xnnpack+custom+qe
 if [[ $# -lt 4 ]]; then # Assuming 4 mandatory args
     echo "Expecting atleast 4 positional arguments"
     echo "Usage: [...]"
@@ -37,7 +37,7 @@ if [[ -z "${MODE:-}" ]]; then
   exit 1
 fi
 
-if [[ "${MODE}" =~ xnnpack.* ]]; then
+if [[ "${MODE}" =~ .*xnnpack.* ]]; then
   XNNPACK=ON
 else
   XNNPACK=OFF
@@ -49,6 +49,12 @@ else
   CUSTOM=OFF
 fi
 
+if [[ "${MODE}" =~ .*qe.* ]]; then
+  QE=ON
+else
+  QE=OFF
+fi
+
 if [[ -z "${BUCK:-}" ]]; then
   BUCK=buck2
 fi
@@ -84,7 +90,6 @@ cmake_build_llama_runner() {
         -DEXECUTORCH_BUILD_CUSTOM="$CUSTOM" \
         -DEXECUTORCH_BUILD_OPTIMIZED=ON \
         -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
-        -DEXECUTORCH_BUILD_OPTIMIZED=ON \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out/${dir} \
         ${dir}
@@ -126,9 +131,15 @@ fi
 # Export model.
 EXPORTED_MODEL_NAME="${EXPORTED_MODEL_NAME}.pte"
 echo "Exporting ${EXPORTED_MODEL_NAME}"
-EXPORT_ARGS="-c stories110M.pt -p ${PARAMS} -d ${DTYPE} -n ${EXPORTED_MODEL_NAME}"
-if [[ "${MODE}" == "xnnpack+kv+custom" ]]; then
-  EXPORT_ARGS="${EXPORT_ARGS} -kv --use_sdpa_with_kv_cache -X -qmode 8da4w -G 128"
+EXPORT_ARGS="-c stories110M.pt -p ${PARAMS} -d ${DTYPE} -n ${EXPORTED_MODEL_NAME} -kv"
+if [[ "${XNNPACK}" == "ON" ]]; then
+  EXPORT_ARGS="${EXPORT_ARGS} -X -qmode 8da4w -G 128"
+fi
+if [[ "${CUSTOM}" == "ON" ]]; then
+  EXPORT_ARGS="${EXPORT_ARGS} --use_sdpa_with_kv_cache"
+fi
+if [[ "${QE}" == "ON" ]]; then
+  EXPORT_ARGS="${EXPORT_ARGS} --embedding-quantize 8,1024"
 fi
 # Add dynamically linked library location
 $PYTHON_EXECUTABLE -m examples.models.llama2.export_llama ${EXPORT_ARGS}

.ci/scripts/test_quantized_aot_lib.sh

@@ -24,7 +24,7 @@ build_cmake_quantized_aot_lib() {
     && retry cmake -DBUCK2=buck2 \
       -DCMAKE_BUILD_TYPE=Release \
       -DCMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH" \
-      -DEXECUTORCH_BUILD_QUANTIZED=ON \
+      -DEXECUTORCH_BUILD_QUANTIZED_OPS_AOT=ON \
       -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" ..)
 
   cmake --build ${CMAKE_OUTPUT_DIR} -j4

.github/workflows/pull.yml

@@ -90,7 +90,7 @@ jobs:
       matrix:
         dtype: [fp32]
         build-tool: [buck2, cmake]
-        mode: [portable, xnnpack+kv+custom]
+        mode: [portable, xnnpack+custom, xnnpack+custom+qe]
       fail-fast: false
     with:
       runner: linux.2xlarge

CMakeLists.txt

@@ -164,8 +164,6 @@ option(EXECUTORCH_BUILD_QNN "Build the Qualcomm backend" OFF)
 
 option(EXECUTORCH_BUILD_OPTIMIZED "Build the optimized kernels" OFF)
 
-option(EXECUTORCH_BUILD_QUANTIZED "Build the quantized kernels" OFF)
-
 option(EXECUTORCH_BUILD_SDK "Build the ExecuTorch SDK")
 
 option(EXECUTORCH_BUILD_SIZE_TEST "Build the size test" OFF)
@@ -413,9 +411,7 @@ if(EXECUTORCH_BUILD_OPTIMIZED)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/optimized)
 endif()
 
-if(EXECUTORCH_BUILD_QUANTIZED)
-  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantized)
-endif()
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantized)
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/configurations)
 
@@ -445,19 +441,14 @@ cmake_dependent_option(
   EXECUTORCH_BUILD_HOST_TARGETS OFF)
 if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
   # Baseline libraries that executor_runner will link against.
-  set(_executor_runner_libs executorch gflags)
+  set(_executor_runner_libs executorch gflags quantized_ops_lib)
 
   if(EXECUTORCH_BUILD_OPTIMIZED)
     list(APPEND _executor_runner_libs optimized_native_cpu_ops_lib)
   else()
     list(APPEND _executor_runner_libs portable_ops_lib)
   endif()
 
-  # Generate lib to register quantized ops
-  if(EXECUTORCH_BUILD_QUANTIZED)
-    list(APPEND _executor_runner_libs quantized_ops_lib)
-  endif()
-
   add_executable(executor_runner ${_executor_runner__srcs})
   if(CMAKE_BUILD_TYPE STREQUAL "Release" AND NOT APPLE)
     target_link_options(executor_runner PRIVATE "LINKER:--gc-sections")

build/Utils.cmake

@@ -74,8 +74,6 @@ function(executorch_print_configuration_summary)
     STATUS "  EXECUTORCH_BUILD_QNN                   : ${EXECUTORCH_BUILD_QNN}")
   message(STATUS "  EXECUTORCH_BUILD_OPTIMIZED             : "
                  "${EXECUTORCH_BUILD_OPTIMIZED}")
-  message(STATUS "  EXECUTORCH_BUILD_QUANTIZED             : "
-                 "${EXECUTORCH_BUILD_QUANTIZED}")
   message(
     STATUS "  EXECUTORCH_BUILD_SDK                   : ${EXECUTORCH_BUILD_SDK}")
   message(

build/build_apple_frameworks.sh

@@ -22,7 +22,7 @@ CUSTOM=OFF
 MPS=OFF
 OPTIMIZED=OFF
 PORTABLE=OFF
-QUANTIZED=OFF
+QUANTIZED=ON
 XNNPACK=OFF
 HEADERS_PATH="include"
 EXECUTORCH_FRAMEWORK="executorch:libexecutorch.a,libexecutorch_no_prim_ops.a,libextension_apple.a,libextension_data_loader.a,libextension_module.a:$HEADERS_PATH"
@@ -51,7 +51,6 @@ usage() {
   echo "  --mps                Include this flag to build the Metal Performance Shaders backend."
   echo "  --optimized          Include this flag to build the Optimized backend."
   echo "  --portable           Include this flag to build the Portable backend."
-  echo "  --quantized          Include this flag to build the Quantized backend."
   echo "  --xnnpack            Include this flag to build the XNNPACK backend."
   echo
   echo "Example:"
@@ -74,7 +73,6 @@ for arg in "$@"; do
       --mps) MPS=ON ;;
       --optimized) OPTIMIZED=ON ;;
       --portable) PORTABLE=ON ;;
-      --quantized) QUANTIZED=ON ;;
       --xnnpack) XNNPACK=ON ;;
       *)
       if [[ -z "$SOURCE_ROOT_DIR" ]]; then
@@ -137,7 +135,6 @@ cmake_build() {
         -DEXECUTORCH_BUILD_CUSTOM=$CUSTOM \
         -DEXECUTORCH_BUILD_MPS=$MPS \
         -DEXECUTORCH_BUILD_OPTIMIZED=$OPTIMIZED \
-        -DEXECUTORCH_BUILD_QUANTIZED=$QUANTIZED \
         -DEXECUTORCH_BUILD_XNNPACK=$XNNPACK \
         ${platform_flag:+-DIOS_PLATFORM=$platform_flag}
     cmake --build . --config $MODE
@@ -181,7 +178,7 @@ append_framework_flag "$CUSTOM" "$CUSTOM_FRAMEWORK"
 append_framework_flag "$MPS" "$MPS_FRAMEWORK"
 append_framework_flag "$OPTIMIZED" "$OPTIMIZED_FRAMEWORK"
 append_framework_flag "$PORTABLE" "$PORTABLE_FRAMEWORK"
-append_framework_flag "$QUANTIZED" "$QUANTIZED_FRAMEWORK"
+append_framework_flag "ON" "$QUANTIZED_FRAMEWORK"
 append_framework_flag "$XNNPACK" "$XNNPACK_FRAMEWORK"
 
 "$SOURCE_ROOT_DIR"/build/create_frameworks.sh "${FRAMEWORK_FLAGS[@]}"

build/executorch-config.cmake

@@ -38,7 +38,7 @@ set(lib_list
     etdump bundled_program extension_data_loader ${FLATCCRT_LIB} mpsdelegate
     qnn_executorch_backend portable_ops_lib extension_module xnnpack_backend
     XNNPACK cpuinfo pthreadpool vulkan_backend optimized_kernels cpublas eigen_blas
-    optimized_ops_lib optimized_native_cpu_ops_lib
+    optimized_ops_lib optimized_native_cpu_ops_lib quantized_kernels quantized_ops_lib
 )
 foreach(lib ${lib_list})
     # Name of the variable which stores result of the find_library search

examples/models/llama2/CMakeLists.txt

@@ -44,6 +44,7 @@ set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
 set(TORCH_ROOT ${EXECUTORCH_ROOT}/third-party/pytorch)
 
 include(${EXECUTORCH_ROOT}/build/Utils.cmake)
+include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
 
 if(NOT PYTHON_EXECUTABLE)
   resolve_python_executable()
@@ -91,6 +92,7 @@ add_subdirectory(runner)
 if(EXECUTORCH_USE_TIKTOKEN)
   # find RE2 for tokenizer
   set(ABSL_ENABLE_INSTALL ON)
+  set(ABSL_PROPAGATE_CXX_STD ON)
   set(_pic_flag
     ${CMAKE_POSITION_INDEPENDENT_CODE})
   set(CMAKE_POSITION_INDEPENDENT_CODE ON)
@@ -118,6 +120,26 @@ else()
   target_link_options_shared_lib(portable_ops_lib)
 endif()
 
+# quantized ops yaml file operation
+merge_yaml(
+  FUNCTIONS_YAML ${CMAKE_CURRENT_SOURCE_DIR}/ops/quantized.yaml
+  FALLBACK_YAML ${EXECUTORCH_ROOT}/kernels/quantized/quantized.yaml
+  OUTPUT_DIR ${CMAKE_CURRENT_BINARY_DIR})
+
+gen_selected_ops("${CMAKE_CURRENT_BINARY_DIR}/merged.yaml" "" "")
+generate_bindings_for_kernels(
+    FUNCTIONS_YAML ${CMAKE_CURRENT_BINARY_DIR}/merged.yaml)
+message("Generated files ${gen_command_sources}")
+
+# quantized_merge_ops_lib: Register quantized op kernels into the runtime
+gen_operators_lib(
+  "quantized_merge_ops_lib"
+  KERNEL_LIBS quantized_kernels
+  DEPS executorch)
+target_include_directories(quantized_merge_ops_lib PUBLIC ${_common_include_directories})
+target_link_options_shared_lib(quantized_merge_ops_lib)
+list(APPEND link_libraries quantized_kernels quantized_merge_ops_lib)
+
 if(EXECUTORCH_BUILD_CUSTOM)
   target_link_options_shared_lib(custom_ops)
   list(APPEND link_libraries custom_ops)

examples/models/llama2/ops/quantized.yaml

@@ -1,10 +1,10 @@
-- func: llama_quantized::embedding_byte.out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, *, Tensor(a!) out) -> Tensor(a!)
+- func: llama_quantized::DEPRECATED_DO_NOT_USE_embedding_byte.out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels:
     - arg_meta: null
       kernel_name: torch::executor::quantized_embedding_byte_out
 
-- func: llama_quantized::embedding_byte.dtype_out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)
+- func: llama_quantized::DEPRECATED_DO_NOT_USE_embedding_byte.dtype_out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, int weight_quant_min, int weight_quant_max, Tensor indices, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
   kernels:
     - arg_meta: null

examples/models/llama2/ops/quantized_ops.py

@@ -15,22 +15,22 @@
     "llama_quantized", "DEF"
 )  # to not be confused with torch.ops.quantized.* ops.
 quantized_lib.define(
-    "embedding_byte(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
+    "DEPRECATED_DO_NOT_USE_embedding_byte(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
     "int weight_quant_min, int weight_quant_max, Tensor indices) -> Tensor",
 )
 
 quantized_lib.define(
-    "embedding_byte.out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
+    "DEPRECATED_DO_NOT_USE_embedding_byte.out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
     "int weight_quant_min, int weight_quant_max, Tensor indices, *, Tensor(a!) out) -> Tensor(a!)",
 )
 
 quantized_lib.define(
-    "embedding_byte.dtype(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
+    "DEPRECATED_DO_NOT_USE_embedding_byte.dtype(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
     "int weight_quant_min, int weight_quant_max, Tensor indices, *, ScalarType? dtype=None) -> Tensor",
 )
 
 quantized_lib.define(
-    "embedding_byte.dtype_out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
+    "DEPRECATED_DO_NOT_USE_embedding_byte.dtype_out(Tensor weight, Tensor weight_scales, Tensor? weight_zero_points, "
     "int weight_quant_min, int weight_quant_max, Tensor indices, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)",
 )
 
@@ -66,7 +66,9 @@ def embedding_byte_weight_checks(weight, weight_scales, weight_zero_points):
     ), f"Expecting weight_zero_points tensor to be None or have same number of rows as weights, but found {weight.size()} and {weight_zero_points.size()}"
 
 
-@impl(quantized_lib, "embedding_byte", "CompositeExplicitAutograd")
+@impl(
+    quantized_lib, "DEPRECATED_DO_NOT_USE_embedding_byte", "CompositeExplicitAutograd"
+)
 def embedding_byte(
     weight: torch.Tensor,
     weight_scales: torch.Tensor,
@@ -92,7 +94,7 @@ def embedding_byte(
     return torch.ops.aten.embedding.default(weight, indices)
 
 
-@impl_abstract("llama_quantized::embedding_byte.out")
+@impl_abstract("llama_quantized::DEPRECATED_DO_NOT_USE_embedding_byte.out")
 def embedding_byte_out_meta(
     weight: torch.Tensor,
     weight_scales: torch.Tensor,
@@ -112,7 +114,11 @@ def embedding_byte_out_meta(
     )
 
 
-@impl(quantized_lib, "embedding_byte.dtype", "CompositeExplicitAutograd")
+@impl(
+    quantized_lib,
+    "DEPRECATED_DO_NOT_USE_embedding_byte.dtype",
+    "CompositeExplicitAutograd",
+)
 def embedding_byte_dtype(
     weight: torch.Tensor,
     weight_scales: torch.Tensor,
@@ -140,7 +146,7 @@ def embedding_byte_dtype(
     return torch.ops.aten.embedding.default(weight, indices)
 
 
-@impl_abstract("llama_quantized::embedding_byte.dtype_out")
+@impl_abstract("llama_quantized::DEPRECATED_DO_NOT_USE_embedding_byte.dtype_out")
 def embedding_byte_dtype_out_meta(
     weight: torch.Tensor,
     weight_scales: torch.Tensor,

examples/models/llama2/quant_lib.py

@@ -105,7 +105,7 @@ def check_embedding_byte_registered():
                     'Use `python -c "import torch as _; print(_.__path__)"` to find where torch package is installed.\n'
                     "Set that as TORCH_PACKAGE_DIR.\n"
                     "Then from root executorch dir do the following:\n"
-                    "rm -rf cmake-out && mkdir cmake-out && (cd cmake-out && cmake -DBUCK2=<path-to-buck2> -DCMAKE_PREFIX_PATH=$TORCH_PACKAGE_DIR -DEXECUTORCH_BUILD_QUANTIZED=ON ..) && cmake --build . -j16\n"
+                    "rm -rf cmake-out && mkdir cmake-out && (cd cmake-out && cmake -DBUCK2=<path-to-buck2> -DCMAKE_PREFIX_PATH=$TORCH_PACKAGE_DIR -DEXECUTORCH_BUILD_QUANTIZED_OPS_AOT=ON ..) && cmake --build . -j16\n"
                     'To find the location of the lib: find cmake-out -name "libquantized_ops_aot_lib*"\n'
                     "Then specify the said library via -s <path to libquantized_ops_aot_lib.so\n"
                 )

examples/models/llama2/quantize.py

@@ -377,7 +377,7 @@ def __init__(
 
     @torch.no_grad()
     def forward(self, indices: torch.Tensor) -> torch.Tensor:
-        return torch.ops.llama_quantized.embedding_byte.dtype(
+        return torch.ops.llama_quantized.DEPRECATED_DO_NOT_USE_embedding_byte.dtype(
             self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
         )
 

examples/models/llama2/runner/targets.bzl

@@ -4,7 +4,7 @@ def _get_operator_lib(aten = False):
     if aten:
         return ["//executorch/kernels/aten:generated_lib"]
     elif runtime.is_oss:
-        return ["//executorch/kernels/portable:generated_lib", "//executorch/examples/models/llama2/custom_ops:custom_ops"]
+        return ["//executorch/kernels/portable:generated_lib", "//executorch/examples/models/llama2/custom_ops:custom_ops", "//executorch/examples/models/llama2/ops:generated_lib"]
     else:
         return ["//executorch/configurations:optimized_native_cpu_ops", "//executorch/examples/models/llama2/custom_ops:custom_ops", "//executorch/examples/models/llama2/ops:generated_lib"]
 

kernels/quantized/CMakeLists.txt

@@ -10,6 +10,9 @@
 # ~~~
 cmake_minimum_required(VERSION 3.19)
 
+option(EXECUTORCH_BUILD_QUANTIZED_OPS_AOT
+       "Build the optimized ops library for AOT export usage" OFF)
+
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
@@ -49,7 +52,7 @@ message("Generated files ${gen_command_sources}")
 # quantized_ops_aot_lib quantized_ops_lib but none of these is a common
 # dependency of the other(s). This is not allowed by the Xcode "new build
 # system".
-if(NOT CMAKE_GENERATOR STREQUAL "Xcode")
+if(NOT CMAKE_GENERATOR STREQUAL "Xcode" AND EXECUTORCH_BUILD_QUANTIZED_OPS_AOT)
   # Build a AOT library to register quantized ops into PyTorch. This is a hack.
   set(_quantized_sources
       ${_quantized_kernels__srcs}