Update on "Transform model to be able to use Attention Sink"

This PR adds necessary functions for transforming the model to be able to use Attention Sink.

Differential Revision: [D65571289](https://our.internmc.facebook.com/intern/diff/D65571289/)

[ghstack-poisoned]

Author: helunwencser

.ci/docker/build.sh

@@ -37,6 +37,10 @@ case "${IMAGE_NAME}" in
     ARM_SDK=yes
     CLANG_VERSION=12
     ;;
+  executorch-ubuntu-22.04-qnn-sdk)
+    QNN_SDK=yes
+    CLANG_VERSION=12
+    ;;
   executorch-ubuntu-22.04-clang12-android)
     LINTRUNNER=""
     CLANG_VERSION=12
@@ -72,6 +76,7 @@ docker build \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \
   --build-arg "ARM_SDK=${ARM_SDK:-}" \
+  --build-arg "QNN_SDK=${QNN_SDK:-}" \
   --build-arg "ANDROID_NDK_VERSION=${ANDROID_NDK_VERSION:-}" \
   -f "${OS}"/Dockerfile \
   "$@" \

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-c8a648d4dffb9f0133ff4a2ea0e660b42105d3ad
+19eff28ff3f19b50da46f5a9ff5f4d4d213806fe

.ci/docker/ci_commit_pins/torchao.txt

@@ -82,5 +82,7 @@ COPY --chown=ci-user:ci-user ./arm /opt/arm
 # Set up ARM SDK if needed
 RUN if [ -n "${ARM_SDK}" ]; then git config --global user.email "[email protected]"; git config --global user.name "OSS CI"; bash /opt/arm/setup.sh --i-agree-to-the-contained-eula /opt/arm-sdk; chown -R ci-user:ci-user /opt/arm-sdk; fi
 
+ARG QNN_SDK
+
 USER ci-user
 CMD ["bash"]

.ci/docker/ubuntu/Dockerfile

@@ -24,6 +24,8 @@
         "ic4": "linux.12xlarge",
         "resnet50": "linux.12xlarge",
         "llava": "linux.12xlarge",
+        "llama3_2_vision_encoder": "linux.12xlarge",
+        "llama3_2_text_decoder": "linux.12xlarge",
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",
@@ -88,8 +90,8 @@ def model_should_run_on_event(model: str, event: str) -> bool:
     if event == "pull_request":
         return model in ["mv3", "vit"]
     elif event == "push":
-        # 'emformer_predict' is running super slow. Only run it periodically
-        return model not in ["emformer_predict"]
+        # These are super slow. Only run it periodically
+        return model not in ["dl3", "edsr", "emformer_predict"]
     else:
         return True
 

.ci/scripts/gather_test_models.py

@@ -23,3 +23,4 @@ fi
 # of nightly. This allows CI to test against latest commits from PyTorch
 install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
+do_not_use_nightly_on_ci

.ci/scripts/setup-linux.sh

@@ -42,11 +42,12 @@ run_and_verify() {
 	-d fp32 \
 	--max_seq_length 32 \
 	--temperature 0 \
+    --show_tokens \
 	--prompt "Once upon a time," > result.txt
 
     # Verify result.txt
     RESULT=$(cat result.txt)
-    EXPECTED_RESULT="there was a little girl"
+    EXPECTED_RESULT="727, 471, 263, 2217, 7826, 4257, 365, 2354, 29889, 2296, 18012, 304, 1708, 5377, 297, 278, 6575, 845, 457, 29889, 3118, 2462, 29892, 1183, 4446, 263"
     if [[ "${RESULT}" == *"${EXPECTED_RESULT}"* ]]; then
         echo "Actual result: ${RESULT}"
         echo "Success"

.ci/scripts/test_llama_runner_eager.sh

@@ -77,7 +77,7 @@ test_model() {
     # Install requirements for export_llama
     bash examples/models/llama/install_requirements.sh
     # Test export_llama script: python3 -m examples.models.llama.export_llama
-    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
+    "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/llama/params/demo_config.json
     run_portable_executor_runner
     rm "./${MODEL_NAME}.pte"
   fi
@@ -87,6 +87,10 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
+  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" || "$MODEL_NAME" == "llama3_2_text_decoder" ]]; then
+    # Install requirements for llama vision.
+    bash examples/models/llama3_2_vision/install_requirements.sh
+  fi
   # python3 -m examples.portable.scripts.export --model_name="llama2" should works too
   "${PYTHON_EXECUTABLE}" -m examples.portable.scripts.export --model_name="${MODEL_NAME}" "${STRICT}"
   run_portable_executor_runner

.ci/scripts/test_model.sh

@@ -113,10 +113,26 @@ cmake_install_executorch_lib() {
 }
 
 download_stories_model_artifacts() {
-    # Download stories110M.pt and tokenizer from Github
+  # Download stories110M.pt and tokenizer from Github
   curl -Ls "https://huggingface.co/karpathy/tinyllamas/resolve/main/stories110M.pt" --output stories110M.pt
   curl -Ls "https://raw.githubusercontent.com/karpathy/llama2.c/master/tokenizer.model" --output tokenizer.model
   # Create params.json file
   touch params.json
   echo '{"dim": 768, "multiple_of": 32, "n_heads": 12, "n_layers": 12, "norm_eps": 1e-05, "vocab_size": 32000}' > params.json
 }
+
+do_not_use_nightly_on_ci() {
+  # An assert to make sure that we are not using PyTorch nightly on CI to prevent
+  # regression as documented in https://github.com/pytorch/executorch/pull/6564
+  TORCH_VERSION=$(pip list | grep -w 'torch ' | awk -F ' ' {'print $2'} | tr -d '\n')
+
+  # The version of PyTorch building from source looks like 2.6.0a0+gitc8a648d that
+  # includes the commit while nightly (2.6.0.dev20241019+cpu) or release (2.6.0)
+  # won't have that. Note that we couldn't check for the exact commit from the pin
+  # ci_commit_pins/pytorch.txt here because the value will be different when running
+  # this on PyTorch CI
+  if [[ "${TORCH_VERSION}" != *"+git"* ]]; then
+    echo "Unexpected torch version. Expected binary built from source, got ${TORCH_VERSION}"
+    exit 1
+  fi
+}

.ci/scripts/utils.sh

@@ -73,6 +73,8 @@ jobs:
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/llm_demo/app-debug-androidTest.apk
           curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
+          curl -o android-test-debug.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug.apk
+          curl -o android-test-debug-androidTest.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch-debug-androidTest.apk
           unzip model.zip
           mv *.pte model.pte