Merge branch 'main' into add_mv3_testcase

Author: digantdesai

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-08434df1f2f88c9770e59246caa2ff9c6f613270
+295f2ed4d103017f7e19a7b8263ece606cd629db

.ci/docker/common/install_android.sh

@@ -70,6 +70,7 @@ install_sdk() {
   # These are the tools needed to build Android apps
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "platforms;android-34"
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "build-tools;33.0.1"
+  yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "build-tools;35.0.0"
   # And some more tools for future emulator tests
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "platform-tools"
   yes | /opt/cmdline-tools/bin/sdkmanager --sdk_root="${SDK_INSTALLATION_DIR}" --install "tools"

.ci/docker/conda-env-ci.txt

@@ -1,4 +1,4 @@
-cmake=3.22.1
+cmake=3.26.4
 ninja=1.10.2
 libuv
 llvm-openmp

.ci/scripts/test_llama_torchao_lowbit.sh

@@ -0,0 +1,85 @@
+#!/bin/bash
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# 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
+
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+export EXECUTORCH_ROOT="$(dirname "${BASH_SOURCE[0]}")/../.."
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+
+which "${PYTHON_EXECUTABLE}"
+
+# Update tokenizers submodule
+pushd $EXECUTORCH_ROOT/extension/llm/tokenizers
+echo "Update tokenizers submodule"
+git submodule update --init
+popd
+
+# Install ET with CMake
+cmake -DPYTHON_EXECUTABLE=python \
+    -DCMAKE_INSTALL_PREFIX=cmake-out \
+    -DEXECUTORCH_ENABLE_LOGGING=1 \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+    -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+    -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=OFF \
+    -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+    -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -Bcmake-out .
+cmake --build cmake-out -j16 --target install --config Release
+
+# Install llama runner with torchao
+cmake -DPYTHON_EXECUTABLE=python \
+    -DCMAKE_PREFIX_PATH=$(python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())') \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
+    -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+    -DEXECUTORCH_BUILD_XNNPACK=OFF \
+    -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
+    -DEXECUTORCH_BUILD_TORCHAO=ON \
+    -Bcmake-out/examples/models/llama \
+    examples/models/llama
+cmake --build cmake-out/examples/models/llama -j16 --config Release
+
+# Download stories llama110m artifacts
+download_stories_model_artifacts
+
+echo "Creating tokenizer.bin"
+$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+
+# Export model
+LLAMA_CHECKPOINT=stories110M.pt
+LLAMA_PARAMS=params.json
+MODEL_OUT=model.pte
+TOKENIZER=tokenizer.bin
+
+# Set low-bit quantization parameters
+QLINEAR_BITWIDTH=3 # Can be 1-8
+QLINEAR_GROUP_SIZE=128 # Must be multiple of 16
+QEMBEDDING_BITWIDTH=4 # Can be 1-8
+QEMBEDDING_GROUP_SIZE=32 # Must be multiple of 16
+
+${PYTHON_EXECUTABLE} -m examples.models.llama.export_llama \
+    --checkpoint "${LLAMA_CHECKPOINT:?}" \
+    --params "${LLAMA_PARAMS:?}" \
+    -kv \
+    --use_sdpa_with_kv_cache \
+    --output_name=${MODEL_OUT} \
+    -qmode "torchao:8da${QLINEAR_BITWIDTH}w" \
+    --group_size ${QLINEAR_GROUP_SIZE} \
+    -E "torchao:${QEMBEDDING_BITWIDTH},${QEMBEDDING_GROUP_SIZE}" \
+    --disable_dynamic_shape \
+    -d fp32
+
+# Test run
+./cmake-out/examples/models/llama/llama_main --model_path=$MODEL_OUT --tokenizer_path=$TOKENIZER --prompt="Once upon a time,"

.github/workflows/android-perf.yml

@@ -2,7 +2,7 @@ name: android-perf
 
 on:
   schedule:
-    - cron: 0 0 * * *
+    - cron: 0 0,8,16 * * *
   pull_request:
     paths:
       - .github/workflows/android-perf.yml

.github/workflows/trunk.yml

@@ -23,8 +23,8 @@ jobs:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     strategy:
       matrix:
-        # Mac runners are expensive and limited, and non reliable. 
-        # Do some basic testing for macos jobs, and rely mostly on 
+        # Mac runners are expensive and limited, and non reliable.
+        # Do some basic testing for macos jobs, and rely mostly on
         # test-models-linux-aarch64 job instead.
         model: [emformer_join, ic4, llama2, mobilebert, mv3, resnet50, vit, w2l]
         backend: [xnnpack-quantization-delegation]
@@ -288,6 +288,26 @@ jobs:
         # Test ANE llama
         ${CONDA_RUN} sh .ci/scripts/test_ane_static_llama.sh
 
+  test-llama-torchao-lowbit:
+    name: test-llama-torchao-lowbit
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    with:
+      runner: macos-m1-stable
+      python-version: '3.11'
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+        bash .ci/scripts/setup-conda.sh
+        eval "$(conda shell.bash hook)"
+
+        # Install requirements
+        ${CONDA_RUN} python install_executorch.py
+        ${CONDA_RUN} sh examples/models/llama/install_requirements.sh
+
+        # Run test
+        ${CONDA_RUN} sh .ci/scripts/test_llama_torchao_lowbit.sh
+
   test-llama-runner-linux:
     # Test Both linux x86 and linux aarch64
     name: test-llama-runner-linux

backends/arm/operators/op_avg_pool2d.py

@@ -43,12 +43,19 @@ def _build_generic_avgpool2d(
         output_zp: int,
         accumulator_type: ts.DType,
     ) -> None:
-        input_tensor = inputs[0]
 
+        input_tensor = inputs[0]
         kernel_size_list = inputs[1].special
         stride_size_list = inputs[2].special
+
         try:
             pad_size_list = inputs[3].special
+            pad_size_list = [
+                pad_size_list[0],
+                pad_size_list[0],
+                pad_size_list[1],
+                pad_size_list[1],
+            ]
         except IndexError:
             pad_size_list = [0, 0, 0, 0]
 

backends/arm/operators/op_max_pool2d.py

@@ -42,9 +42,15 @@ def define_node(
         stride = inputs[2].special
 
         try:
-            padding = [*inputs[3].special, *inputs[3].special]
+            pad_size_list = inputs[3].special
+            pad_size_list = [
+                pad_size_list[0],
+                pad_size_list[0],
+                pad_size_list[1],
+                pad_size_list[1],
+            ]
         except IndexError:
-            padding = [0, 0, 0, 0]
+            pad_size_list = [0, 0, 0, 0]
 
         accumulator_type = output.dtype
 
@@ -63,7 +69,7 @@ def define_node(
         attr.PoolAttribute(
             kernel=kernel_size,
             stride=stride,
-            pad=padding,
+            pad=pad_size_list,
             input_zp=input_zp,
             output_zp=output_zp,
             accum_dtype=accumulator_type,