Update on " [ExecuTorch][BE] Split kv cache and SDPA for better code …

…sharing"

Summary:

Why?
We have coupled SDPA with kv cache for a while. Initially this was done
as we implemented sdpa_with_kv_cache custom op to reduce multiple copy
overheads from kv cache update. (This could have been done by having
separate custom kv cache update and custom sdpa op. Recent changes
enabled this.)
As a result of SDPA module owning kv cache, we get a) non-composable
implementation and b) harder to reuse model definition and components
from repos like tune. Output of this is that we have multiple definition
of the same model, llama, lying around in ET, TorchChat and Tune. This
diff and subsequent ones will try to move in the direction where custom
kv cache and custom sdpa become decoupled and composable, making it more
module-swap friendly with tune's model definition.

How.
Earlier PRs decoupled kv cache update from sdpa. So now
1. Decouple SDPA nn.Module from KV cache.
2. Standardize on KVCache and SDPA interface. That is KVCache and SDPA
   both operate on q, k, v in [B, # heads, seq_len, head_dim] formatted
   tensors.
3. 2 will introduce multiple tranposes when KVCache and SDPA are
   replaced by custom modules, but we will write graph pass to undo
   those.

Test Plan:
Existing tests.
Make sure perf doesnt regress

Differential Revision: [D67914054](https://our.internmc.facebook.com/intern/diff/D67914054)

[ghstack-poisoned]

.ci/docker/ci_commit_pins/buck2.txt

@@ -1 +1 @@
-2024-05-15
+2024-12-16

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-2ea4b56ec872424e486c4fe2d55da061067a2ed3
+0a94bb432ed75cc2d950d81b2921363218a7e459

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

@@ -1,2 +1,4 @@
 cmake=3.22.1
 ninja=1.10.2
+libuv
+pkg-config

.ci/scripts/setup-arm-baremetal-tools.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+# Copyright 2024 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# NB: This function could be used to install Arm dependencies
+# Setup arm example environment (including TOSA tools)
+git config --global user.email "github_executorch@arm.com"
+git config --global user.name "Github Executorch"
+bash examples/arm/setup.sh --i-agree-to-the-contained-eula

.ci/scripts/setup-macos.sh

@@ -131,5 +131,9 @@ if [[ -z "${GITHUB_RUNNER:-}" ]]; then
 fi
 
 print_cmake_info
-install_executorch
+install_pytorch_and_domains
+# We build PyTorch from source here instead of using nightly. This allows CI to test against
+# the pinned commit from PyTorch
+install_executorch "use-pt-pinned-commit"
 build_executorch_runner "${BUILD_TOOL}"
+do_not_use_nightly_on_ci

.ci/scripts/utils.sh

@@ -40,6 +40,42 @@ install_pip_dependencies() {
   popd || return
 }
 
+install_domains() {
+  echo "Install torchvision and torchaudio"
+  pip install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
+  pip install --no-use-pep517 --user "git+https://github.com/pytorch/vision.git@${TORCHVISION_VERSION}"
+}
+
+install_pytorch_and_domains() {
+  pushd .ci/docker || return
+  TORCH_VERSION=$(cat ci_commit_pins/pytorch.txt)
+  popd || return
+
+  git clone https://github.com/pytorch/pytorch.git
+
+  # Fetch the target commit
+  pushd pytorch || return
+  git checkout "${TORCH_VERSION}"
+  git submodule update --init --recursive
+
+  export USE_DISTRIBUTED=1
+  # Then build and install PyTorch
+  python setup.py bdist_wheel
+  pip install "$(echo dist/*.whl)"
+
+  # Grab the pinned audio and vision commits from PyTorch
+  TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
+  export TORCHAUDIO_VERSION
+  TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt)
+  export TORCHVISION_VERSION
+
+  install_domains
+
+  popd || return
+  # Print sccache stats for debugging
+  sccache --show-stats || true
+}
+
 install_flatc_from_source() {
   # NB: This function could be used to install flatbuffer from source
   pushd third-party/flatbuffers || return
@@ -59,17 +95,6 @@ install_flatc_from_source() {
   popd || return
 }
 
-install_arm() {
-  # NB: This function could be used to install Arm dependencies
-  # Setup arm example environment (including TOSA tools)
-  git config --global user.email "github_executorch@arm.com"
-  git config --global user.name "Github Executorch"
-  bash examples/arm/setup.sh --i-agree-to-the-contained-eula
-
-  # Test tosa_reference flow
-  source examples/arm/ethos-u-scratch/setup_path.sh
-}
-
 build_executorch_runner_buck2() {
   # Build executorch runtime with retry as this step is flaky on macos CI
   retry buck2 build //examples/portable/executor_runner:executor_runner

.github/workflows/apple-perf.yml

@@ -410,7 +410,7 @@ jobs:
     runs-on: linux.2xlarge
     steps:
       - name: Download the apps from GitHub
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
         with:
           # The name here needs to match the name of the upload-artifact parameter
           name: ios-apps

.github/workflows/apple.yml

@@ -53,7 +53,7 @@ jobs:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
       secrets-env: BUILD_CERTIFICATE_BASE64 EXECUTORCH_DEMO_BUILD_PROVISION_PROFILE_BASE64 KEYCHAIN_PASSWORD
-      upload-artifact: ios-apps
+      upload-artifact: ios-demo-app
       script: |
         set -eux
 
@@ -83,10 +83,10 @@ jobs:
     runs-on: linux.2xlarge
     steps:
       - name: Download the artifacts from GitHub
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
         with:
           # The name here needs to match the name of the upload-artifact parameter
-          name: ios-apps
+          name: ios-demo-app
           path: ${{ runner.temp }}/artifacts/
 
       - name: Verify the artifacts
@@ -216,7 +216,7 @@ jobs:
           role-to-assume: arn:aws:iam::308535385114:role/gha_executorch_upload-frameworks-ios
           aws-region: us-east-1
       - name: Download the artifact
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
         with:
           # NB: The name here needs to match the upload-artifact name from build-frameworks-ios job
           name: executorch-frameworks-ios
@@ -291,7 +291,7 @@ jobs:
       python-version: '3.11'
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      upload-artifact: ios-apps
+      upload-artifact: ios-benchmark-app
       secrets-env: BUILD_CERTIFICATE_BASE64 EXECUTORCH_BENCHMARK_BUILD_PROVISION_PROFILE_BASE64 KEYCHAIN_PASSWORD
       timeout: 90
       script: |

.github/workflows/pull.yml

@@ -354,13 +354,11 @@ jobs:
         EXECUTORCH_BUILD_ARM_BAREMETAL=ON \
         .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
 
-        source .ci/scripts/utils.sh
         # Install Arm dependencies
-        install_arm
-
-        # Run pytest with coverage
-        pytest -c /dev/null -v -n auto --cov=./ --cov-report=xml backends/arm/test
+        .ci/scripts/setup-arm-baremetal-tools.sh
 
+        # Run pytest without simulator
+        backends/arm/test/test_arm_baremetal.sh test_pytest
 
   test-llama-runner-qnn-linux:
     name: test-llama-runner-qnn-linux

.github/workflows/trunk.yml

@@ -146,14 +146,15 @@ jobs:
         source .ci/scripts/utils.sh
         install_executorch
 
-        install_arm
+        .ci/scripts/setup-arm-baremetal-tools.sh
 
         # Increase number of files user can monitor to bypass buck failures.
         # Hopefully this is high enough for this setup.
         sudo sysctl fs.inotify.max_user_watches=1048576 # 1024 * 1024
 
         # Test ethos-u delegate examples with run.sh
-        PYTHON_EXECUTABLE=python bash examples/arm/run.sh examples/arm/ethos-u-scratch/
+        backends/arm/test/test_arm_baremetal.sh test_run_ethosu_fvp
+
 
   test-arm-reference-delegation:
     name: test-arm-reference-delegation
@@ -172,10 +173,10 @@ jobs:
         source .ci/scripts/utils.sh
         install_executorch
 
-        install_arm
+        .ci/scripts/setup-arm-baremetal-tools.sh
 
-        # Run arm unit tests
-        pytest -c /dev/null -v -n auto --cov=./ --cov-report=xml backends/arm/test
+        # Run arm unit tests using the simulator
+        backends/arm/test/test_arm_baremetal.sh test_pytest_ethosu_fvp
 
   test-coreml-delegate:
     name: test-coreml-delegate

.gitmodules

@@ -66,7 +66,7 @@
 	url = https://github.com/pybind/pybind11.git
 [submodule "backends/cadence/fusion_g3/third-party/nnlib/nnlib-FusionG3"]
 	path = backends/cadence/fusion_g3/third-party/nnlib/nnlib-FusionG3
-	url = https://github.com/foss-xtensa/nnlib-FusionG3/
+	url = https://github.com/foss-xtensa/nnlib-FusionG3.git
 [submodule "third-party/ao"]
 	path = third-party/ao
 	url = https://github.com/pytorch/ao.git

CONTRIBUTING.md

@@ -197,7 +197,7 @@ If it's not clear how to add a test for your PR, take a look at the blame for
 the code you're modifying and find an author who has more context. Ask them
 for their help in the PR comments.
 
-TODO: Explain how to run tests locally without needing to push and wait for CI.
+The `test/run_oss_cpp_tests.sh` script will build and run C++ tests locally.
 
 ### Continuous Integration
 See https://hud.pytorch.org/hud/pytorch/executorch/main for the current state of

backends/apple/mps/mps_preprocess.py

@@ -32,6 +32,9 @@
     CompileSpec,
     PreprocessResult,
 )
+
+from executorch.exir.passes.memory_format_ops_pass import DimOrderOpsRevertPass
+from executorch.exir.program._program import _transform
 from torch.export.exported_program import ExportedProgram
 
 FORMAT = "[%(levelname)s %(asctime)s %(filename)s:%(lineno)s] %(message)s"
@@ -83,6 +86,9 @@ def preprocess(
         #    FlatBuffer graph, process the `output` nodes and add their id to
         #    the `output_ids` array in the schema.
 
+        # TODO: Remove this once we have a better support for the dim-order ops.
+        edge_program = _transform(edge_program, DimOrderOpsRevertPass())
+
         mps_graph = MPSGraph(
             version="0",
             mps_nodes=[],

backends/apple/mps/operators/constant_ops.py

@@ -79,6 +79,25 @@ def define_node(
         )
 
 
+@register_node_visitor
+class ToDimOrderEmptyVisitor(NodeVisitor):
+    target = ["dim_order_ops._empty_dim_order.default"]
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        mps_graph: MPSGraph,
+    ) -> None:
+        # We should never get here, because DimOrderOpsRevertPass replaces this with an aten.empty.memory_format op
+        # But if we do, we can't handle it ATM, so raise an exception
+        raise NotImplementedError(
+            "dim_order_ops._empty_dim_order.default is not supported yet"
+        )
+
+
 @register_node_visitor
 class FullLikeVisitor(NodeVisitor):
     target = "aten.full_like.default"

backends/apple/mps/operators/op_clone.py

@@ -33,3 +33,22 @@ def define_node(
                 )
         input_id = self.define_tensor(get_input_node(node, 0), mps_graph)
         self.tensor_to_id[node] = input_id
+
+
+@register_node_visitor
+class ToDimOrderCopyVisitor(NodeVisitor):
+    target = ["dim_order_ops._to_dim_order_copy.default"]
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        mps_graph: MPSGraph,
+    ) -> None:
+        # We should never get here, because DimOrderOpsRevertPass replaces this with an aten._to_copy op
+        # But if we do, we can't handle it ATM, so raise an exception
+        raise NotImplementedError(
+            "dim_order_ops._to_dim_order_copy.default is not supported yet"
+        )

backends/apple/mps/test/test_mps.py

@@ -1829,6 +1829,21 @@ def forward(self, x):
             Clone(), model_inputs, func_name=inspect.stack()[0].function[5:]
         )
 
+    def test_mps_backend_to_copy(self):
+        class Copy(torch.nn.Module):
+            def forward(self, x):
+                return (
+                    torch.ops.aten._to_copy.default(
+                        x + 2, memory_format=torch.contiguous_format
+                    )
+                    + x
+                )
+
+        model_inputs = (torch.randn(1, 3, 3),)
+        self.lower_and_test_with_partitioner(
+            Copy(), model_inputs, func_name=inspect.stack()[0].function[5:]
+        )
+
     def test_mps_backend_floor(self):
         class Floor(torch.nn.Module):
             def forward(self, x):

backends/apple/mps/test/test_mps_utils.py

@@ -26,10 +26,7 @@
 
 # Config for Capturing the weights, will be moved in the future
 
-# TODO(T182928844): Delegate dim order op to backend.
-_EDGE_COMPILE_CONFIG = exir.EdgeCompileConfig(
-    _check_ir_validity=False, _skip_dim_order=True
-)
+_EDGE_COMPILE_CONFIG = exir.EdgeCompileConfig(_check_ir_validity=False)
 
 
 class ansi_colors:
@@ -219,7 +216,6 @@ def lower_module_and_test_output(
             dynamic_shapes=dynamic_shapes,
             edge_compile_config=EdgeCompileConfig(
                 _check_ir_validity=False,
-                _skip_dim_order=True,  # TODO(T182928844): Delegate dim order op to backend.
             ),
         )
 
@@ -250,7 +246,6 @@ def lower_module_and_test_output(
                 export(delegated_program, sample_inputs, strict=True),
                 compile_config=exir.EdgeCompileConfig(
                     _check_ir_validity=False,
-                    _skip_dim_order=True,  # TODO(T182928844): Delegate dim order op to backend.
                 ),
             ).to_executorch(
                 config=ExecutorchBackendConfig(extract_delegate_segments=False)