Add torchao (#1182)

* init

* update install utils

* update

* update libs

* update torchao pin

* fix ci test

* add python et install to ci

* fix ci errors

* fixes

* fixes

* fixes

* fixes

* fixes

* fixes

* fixes

Author: metascroy

.github/workflows/pull.yml

@@ -14,6 +14,7 @@ __pycache__/
 # Build directories
 build/android/*
 et-build/*
+torchao-build/*
 runner-et/cmake-out/*
 runner-aoti/cmake-out/*
 cmake-out/

.gitignore

@@ -118,6 +118,75 @@ python3 torchchat.py export llama3 --quantize '{"embedding": {"bitwidth": 4, "gr
 python3 torchchat.py generate llama3 --pte-path llama3.pte  --prompt "Hello my name is"
 ```
 
+## Experimental TorchAO lowbit kernels
+
+### Use
+The quantization scheme a8wxdq dynamically quantizes activations to 8 bits, and quantizes the weights in a groupwise manner with a specified bitwidth and groupsize.
+It takes arguments bitwidth (2, 3, 4, 5, 6, 7), groupsize, and has_weight_zeros (true, false).
+The argument has_weight_zeros indicates whether the weights are quantized with scales only (has_weight_zeros: false) or with both scales and zeros (has_weight_zeros: true).
+Roughly speaking, {bitwidth: 4, groupsize: 256, has_weight_zeros: false} is similar to GGML's Q40 quantization scheme.
+
+You should expect high performance on ARM CPU if bitwidth is 2, 3, 4, or 5 and groupsize is divisible by 16.  With other platforms and argument choices, a slow fallback kernel will be used.  You will see warnings about this during quantization.
+
+### Setup
+To use a8wxdq, you must set up the torchao experimental kernels.  These will only work on devices with ARM CPUs, for example on Mac computers with Apple Silicon.
+
+From the torchchat root directory, run
+```
+sh torchchat/utils/scripts/build_torchao_ops.sh
+```
+
+This should take about 10 seconds to complete.  Once finished, you can use a8wxdq in torchchat.
+
+Note: if you want to use the new kernels in the AOTI and C++ runners, you must pass the flag link_torchao when running the scripts the build the runners.
+
+```
+sh torchchat/utils/scripts/build_native.sh aoti link_torchao_ops
+```
+
+```
+sh torchchat/utils/scripts/build_native.sh et link_torchao_ops
+```
+
+Note before running `sh torchchat/utils/scripts/build_native.sh et link_torchao_ops`, you must first install executorch with `sh torchchat/utils/scripts/install_et.sh` if you have not done so already.
+
+### Examples
+
+Below we show how to use the new kernels.  Except for ExecuTorch, you can specify the number of threads used by setting OMP_NUM_THREADS (as is the case with PyTorch in general).  Doing so is optional and a default number of threads will be chosen automatically if you do not specify.
+
+#### Eager mode
+```
+OMP_NUM_THREADS=6 python3 torchchat.py generate llama3.1 --device cpu --dtype float32 --quantize '{"linear:a8wxdq": {"bitwidth": 4, "groupsize": 256, "has_weight_zeros": false}}' --prompt "Once upon a time," --num-samples 5
+```
+
+#### torch.compile
+```
+OMP_NUM_THREADS=6 python3 torchchat.py generate llama3.1 --device cpu --dtype float32 --quantize '{"linear:a8wxdq": {"bitwidth": 4, "groupsize": 256, "has_weight_zeros": false}}' --compile --prompt "Once upon a time,"  --num-samples 5
+```
+
+#### AOTI
+```
+OMP_NUM_THREADS=6 python torchchat.py export llama3.1 --device cpu --dtype float32 --quantize '{"linear:a8wxdq": {"bitwidth": 4, "groupsize": 256, "has_weight_zeros": false}}' --output-dso llama3_1.so
+OMP_NUM_THREADS=6 python3 torchchat.py generate llama3.1 --dso-path llama3_1.so --prompt "Once upon a time,"  --num-samples 5
+```
+
+If you built the AOTI runner with link_torchao_ops as discussed in the setup section, you can also use the C++ runner:
+
+```
+OMP_NUM_THREADS=6 ./cmake-out/aoti_run llama3_1.so -z $HOME/.torchchat/model-cache/meta-llama/Meta-Llama-3.1-8B-Instruct/tokenizer.model -l 3 -i "Once upon a time,"
+```
+
+#### ExecuTorch
+```
+python torchchat.py export llama3.1 --device cpu --dtype float32 --quantize '{"linear:a8wxdq": {"bitwidth": 4, "groupsize": 256, "has_weight_zeros": false}}' --output-pte llama3_1.pte
+```
+
+Note: only the ExecuTorch C++ runner in torchchat when built using the instructions in the setup can run the exported *.pte file.  It will not work with the `python torchchat.py generate` command.
+
+```
+./cmake-out/et_run llama3_1.pte -z $HOME/.torchchat/model-cache/meta-llama/Meta-Llama-3.1-8B-Instruct/tokenizer.model -l 3 -i "Once upon a time,"
+```
+
 ## Quantization Profiles
 
 Four [sample profiles](https://github.com/pytorch/torchchat/tree/main/torchchat/quant_config/) are included with the torchchat distribution: `cuda.json`, `desktop.json`, `mobile.json`, `pi5.json`

docs/quantization.md

@@ -0,0 +1 @@
+63cb7a9857654784f726fec75c0dc36167094d8a

install/.pins/torchao-pin.txt

@@ -28,3 +28,7 @@ if(Torch_FOUND)
     target_link_libraries(aoti_run "${TORCH_LIBRARIES}" m)
     set_property(TARGET aoti_run PROPERTY CXX_STANDARD 17)
 endif()
+
+if (LINK_TORCHAO_OPS)
+    target_link_libraries(aoti_run "${TORCHCHAT_ROOT}/torchao-build/cmake-out/lib/liblinear_a8wxdq_ATEN${CMAKE_SHARED_LIBRARY_SUFFIX}")
+endif()

runner/aoti.cmake

@@ -116,6 +116,14 @@ if(executorch_FOUND)
     target_link_libraries(et_run PRIVATE log)
   endif()
 
+  if(LINK_TORCHAO_OPS)
+    target_link_libraries(et_run PRIVATE "$<LINK_LIBRARY:WHOLE_ARCHIVE,${TORCHCHAT_ROOT}/torchao-build/cmake-out/lib/liblinear_a8wxdq_EXECUTORCH.a>")
+    target_link_libraries(et_run PRIVATE
+      "${TORCHCHAT_ROOT}/torchao-build/cmake-out/lib/libtorchao_kernels_aarch64.a"
+       "${TORCHCHAT_ROOT}/torchao-build/cmake-out/lib/libtorchao_ops_linear_EXECUTORCH.a"
+    )
+  endif()
+
 else()
   MESSAGE(WARNING "ExecuTorch package not found")
 endif()

runner/et.cmake

@@ -96,10 +96,19 @@ def quantize_model(
                 precision = get_precision()
 
             try:
-                # Easier to ask forgiveness than permission
-                quant_handler = ao_quantizer_class_dict[quantizer](
-                    groupsize=q_kwargs["groupsize"], device=device, precision=precision
-                )
+                if quantizer == "linear:a8wxdq":
+                    quant_handler = ao_quantizer_class_dict[quantizer](
+                        device=device,
+                        precision=precision,
+                        bitwidth=q_kwargs.get("bitwidth", 4),
+                        groupsize=q_kwargs.get("groupsize", 128),
+                        has_weight_zeros=q_kwargs.get("has_weight_zeros", False),
+                    )
+                else:
+                    # Easier to ask forgiveness than permission
+                    quant_handler = ao_quantizer_class_dict[quantizer](
+                        groupsize=q_kwargs["groupsize"], device=device, precision=precision
+                    )
             except TypeError as e:
                 if "unexpected keyword argument 'device'" in str(e):
                     quant_handler = ao_quantizer_class_dict[quantizer](
@@ -861,3 +870,33 @@ def quantized_model(self) -> nn.Module:
     "linear:int4": Int4WeightOnlyQuantizer,
     "linear:a8w4dq": Int8DynActInt4WeightQuantizer,
 }
+
+try:
+    import importlib.util
+    import sys
+    import os
+    torchao_build_path = f"{os.getcwd()}/torchao-build"
+
+    # Try loading quantizer
+    torchao_experimental_quant_api_spec = importlib.util.spec_from_file_location(
+        "torchao_experimental_quant_api",
+        f"{torchao_build_path}/src/ao/torchao/experimental/quant_api.py",
+    )
+    torchao_experimental_quant_api = importlib.util.module_from_spec(torchao_experimental_quant_api_spec)
+    sys.modules["torchao_experimental_quant_api"] = torchao_experimental_quant_api
+    torchao_experimental_quant_api_spec.loader.exec_module(torchao_experimental_quant_api)
+    from torchao_experimental_quant_api import Int8DynActIntxWeightQuantizer
+    ao_quantizer_class_dict["linear:a8wxdq"] = Int8DynActIntxWeightQuantizer
+
+    # Try loading custom op
+    try:
+        import glob
+        libs = glob.glob(f"{torchao_build_path}/cmake-out/lib/liblinear_a8wxdq_ATEN.*")
+        libs = list(filter(lambda l: (l.endswith("so") or l.endswith("dylib")), libs))
+        torch.ops.load_library(libs[0])
+    except Exception as e:
+        print("Failed to torchao ops library with error: ", e)
+        print("Slow fallback kernels will be used.")
+
+except Exception as e:
+    print(f"Failed to load torchao experimental a8wxdq quantizer with error: {e}")

torchchat/utils/quantize.py

@@ -26,6 +26,7 @@ if [ $# -eq 0 ]; then
     exit 1
 fi
 
+LINK_TORCHAO_OPS=OFF
 while (( "$#" )); do
   case "$1" in
     -h|--help)
@@ -42,6 +43,11 @@ while (( "$#" )); do
       TARGET="et"
       shift
       ;;
+    link_torchao_ops)
+      echo "Linking with torchao ops..."
+      LINK_TORCHAO_OPS=ON
+      shift
+      ;;
     *)
       echo "Invalid option: $1"
       show_help
@@ -66,14 +72,28 @@ if [[ "$TARGET" == "et" ]]; then
     echo "Make sure you run install_executorch_libs"
     exit 1
   fi
+
+  if [[ "$LINK_TORCHAO_OPS" == "ON" ]]; then
+    if [ ! -d "${TORCHCHAT_ROOT}/torchao-build" ]; then
+      echo "Directory ${TORCHCHAT_ROOT}/torchao-build does not exist."
+      echo "Make sure you run clone_torchao"
+      exit 1
+    fi
+
+    source "$(dirname "${BASH_SOURCE[0]}")/install_utils.sh"
+    find_cmake_prefix_path
+    EXECUTORCH_INCLUDE_DIRS="${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/install/include;${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/src"
+    EXECUTORCH_LIBRARIES="${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/install/lib/libexecutorch_no_prim_ops.a;${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/install/lib/libextension_threadpool.a;${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/install/lib/libcpuinfo.a;${TORCHCHAT_ROOT}/${ET_BUILD_DIR}/install/lib/libpthreadpool.a"
+    install_torchao_executorch_ops
+  fi
 fi
 popd
 
 # CMake commands
 if [[ "$TARGET" == "et" ]]; then
-    cmake -S . -B ./cmake-out -DCMAKE_PREFIX_PATH=`python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'` -DET_USE_ADAPTIVE_THREADS=ON -DCMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=1" -G Ninja
+    cmake -S . -B ./cmake-out -DCMAKE_PREFIX_PATH=`python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'` -DLINK_TORCHAO_OPS="${LINK_TORCHAO_OPS}" -DET_USE_ADAPTIVE_THREADS=ON -DCMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=1" -G Ninja
 else
-    cmake -S . -B ./cmake-out -DCMAKE_PREFIX_PATH=`python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'` -DCMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=0" -G Ninja
+    cmake -S . -B ./cmake-out -DCMAKE_PREFIX_PATH=`python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'` -DLINK_TORCHAO_OPS="${LINK_TORCHAO_OPS}" -DCMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=0" -G Ninja
 fi
 cmake --build ./cmake-out --target "${TARGET}"_run
 

torchchat/utils/scripts/build_native.sh

@@ -0,0 +1,16 @@
+#!/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.
+
+
+
+source "$(dirname "${BASH_SOURCE[0]}")/install_utils.sh"
+
+pushd ${TORCHCHAT_ROOT}
+find_cmake_prefix_path
+clone_torchao
+install_torchao_aten_ops
+popd

torchchat/utils/scripts/build_torchao_ops.sh

@@ -19,10 +19,4 @@ pushd ${TORCHCHAT_ROOT}
 find_cmake_prefix_path
 clone_executorch
 install_executorch_libs $ENABLE_ET_PYBIND
-install_executorch_python_libs $ENABLE_ET_PYBIND
-# TODO: figure out the root cause of 'AttributeError: module 'evaluate'
-# has no attribute 'utils'' error from evaluate CI jobs and remove
-# `import lm_eval` from torchchat.py since it requires a specific version
-# of numpy.
-pip install numpy=='1.26.4'
 popd