2024-10-31 nightly release (f813b6a)

Author: pytorchbot

.github/workflows/_android.yml

@@ -66,6 +66,16 @@ jobs:
           # avoid permission issue
           sudo chown -R "${USER}" /opt/android
 
+      - name: Download Artifacts
+        shell: bash
+        run: |
+          set -eux
+          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
+          unzip model.zip
+          mv *.pte model.pte
+
       - name: Gradle cache
         uses: gradle/actions/setup-gradle@v3
 

.github/workflows/pull.yml

@@ -99,6 +99,8 @@ jobs:
       submodules: 'true'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 900
+      upload-artifact: android-models
+      upload-artifact-to-s3: true
       script: |
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
@@ -107,13 +109,15 @@ jobs:
         DTYPE=${{ matrix.dtype }}
         BUILD_TOOL="cmake"
         MODE=${{ matrix.mode }}
+        ARTIFACTS_DIR_NAME="artifacts-to-be-uploaded/${DTYPE}-${MODE}"
+        ARTIFACTS_DIR_NAME="${ARTIFACTS_DIR_NAME/+/-}"
 
         # Setup executorch
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
         # Install requirements for export_llama
         PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
         # Test llama2
-        PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh stories110M "${BUILD_TOOL}" "${DTYPE}" "${MODE}" "${ARTIFACTS_DIR_NAME}"
 
   test-llama-runner-linux-android:
     name: test-llama-runner-linux-android
@@ -320,6 +324,7 @@ jobs:
 
   android:
     uses: ./.github/workflows/_android.yml
+    needs: test-llama-runner-linux
 
   unittest:
     uses: ./.github/workflows/_unittest.yml

backends/apple/coreml/compiler/coreml_preprocess.py

@@ -425,12 +425,15 @@ def preprocess(
             CoreMLBackend.op_linear_quantizer_config_from_compile_specs(compile_specs)
         )
 
+        # Load the model if MODEL_TYPE is 'COMPILED_MODEL'. This step is necessary because
+        # get_compiled_model_path() requires a loaded model.
+        skip_model_load = model_type != CoreMLBackend.MODEL_TYPE.COMPILED_MODEL
         mlmodel = ct.convert(
             model=edge_program,
             source="pytorch",
             convert_to="mlprogram",
             pass_pipeline=ct.PassPipeline.DEFAULT,
-            skip_model_load=True,
+            skip_model_load=skip_model_load,
             compute_precision=model_compute_precision,
             minimum_deployment_target=minimum_deployment_target,
             compute_units=compute_units,

backends/apple/coreml/runtime/delegate/backend_delegate.h

@@ -28,7 +28,7 @@ class BackendDelegate {
         // Max models cache size in bytes.
         size_t max_models_cache_size = 10 * size_t(1024) * size_t(1024) * size_t(1024);
         // If set to `true`, delegate pre-warms the most recently used asset.
-        bool should_prewarm_asset = true;
+        bool should_prewarm_asset = false;
         // If set to `true`, delegate pre-warms the model in `init`.
         bool should_prewarm_model = true;
     };

backends/apple/coreml/runtime/delegate/com.apple.executorchcoreml_config.plist

@@ -3,7 +3,7 @@
 <plist version="1.0">
 <dict>
 	<key>shouldPrewarmAsset</key>
-	<true/>
+	<false/>
 	<key>shouldPrewarmModel</key>
 	<true/>
 	<key>maxAssetsSizeInBytes</key>

backends/apple/coreml/runtime/test/CoreMLBackendDelegateTests.mm

@@ -209,6 +209,12 @@ - (void)testStateProgramExecute {
 }
 #endif
 
+- (void)testAddMulCompiledProgramExecute {
+    NSURL *modelURL = [[self class] bundledResourceWithName:@"add_mul_compiled_coreml_all" extension:@"pte"];
+    XCTAssertNotNil(modelURL);
+    [self executeModelAtURL:modelURL nLoads:1 nExecutions:2];
+}
+
 - (void)executeMultipleModelsConcurrently:(NSArray<NSURL *> *)modelURLs
                                    nLoads:(NSUInteger)nLoads
                               nExecutions:(NSUInteger)nExecutions

backends/apple/coreml/runtime/workspace/executorchcoreml.xcodeproj/project.pbxproj

@@ -8,6 +8,7 @@
 
 /* Begin PBXBuildFile section */
 		8307EB8A2C9262060011AE6D /* state_coreml_all.pte in Resources */ = {isa = PBXBuildFile; fileRef = 8307EB892C9262060011AE6D /* state_coreml_all.pte */; };
+		838CA6872CD1965700462190 /* add_mul_compiled_coreml_all.pte in Resources */ = {isa = PBXBuildFile; fileRef = 838CA6862CD1965700462190 /* add_mul_compiled_coreml_all.pte */; };
 		83BB78A02C65DA7300274ED7 /* ETCoreMLModelDebugInfo.mm in Sources */ = {isa = PBXBuildFile; fileRef = 83BB789F2C65DA7300274ED7 /* ETCoreMLModelDebugInfo.mm */; };
 		83BB78BF2C66AAAE00274ED7 /* add_mul_coreml_all.bin in Resources */ = {isa = PBXBuildFile; fileRef = 83BB78BD2C66AAAE00274ED7 /* add_mul_coreml_all.bin */; };
 		83BB78C02C66AAAE00274ED7 /* add_mul_coreml_all.pte in Resources */ = {isa = PBXBuildFile; fileRef = 83BB78BE2C66AAAE00274ED7 /* add_mul_coreml_all.pte */; };
@@ -122,6 +123,7 @@
 
 /* Begin PBXFileReference section */
 		8307EB892C9262060011AE6D /* state_coreml_all.pte */ = {isa = PBXFileReference; lastKnownFileType = file; name = state_coreml_all.pte; path = ../test/models/state_coreml_all.pte; sourceTree = "<group>"; };
+		838CA6862CD1965700462190 /* add_mul_compiled_coreml_all.pte */ = {isa = PBXFileReference; lastKnownFileType = file; name = add_mul_compiled_coreml_all.pte; path = ../test/models/add_mul_compiled_coreml_all.pte; sourceTree = "<group>"; };
 		83BB789E2C65DA7300274ED7 /* ETCoreMLModelDebugInfo.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = ETCoreMLModelDebugInfo.h; path = ../sdk/ETCoreMLModelDebugInfo.h; sourceTree = "<group>"; };
 		83BB789F2C65DA7300274ED7 /* ETCoreMLModelDebugInfo.mm */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.objcpp; name = ETCoreMLModelDebugInfo.mm; path = ../sdk/ETCoreMLModelDebugInfo.mm; sourceTree = "<group>"; };
 		83BB78BD2C66AAAE00274ED7 /* add_mul_coreml_all.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; name = add_mul_coreml_all.bin; path = ../test/models/add_mul_coreml_all.bin; sourceTree = "<group>"; };
@@ -606,6 +608,7 @@
 				C98551992AD2542D009143F9 /* mul_coreml_all.bin */,
 				C985519C2AD2542D009143F9 /* mul_coreml_all.pte */,
 				C985519B2AD2542D009143F9 /* mv3_coreml_all.bin */,
+				838CA6862CD1965700462190 /* add_mul_compiled_coreml_all.pte */,
 				C98551982AD2542D009143F9 /* mv3_coreml_all.pte */,
 				83BB78BD2C66AAAE00274ED7 /* add_mul_coreml_all.bin */,
 				83BB78BE2C66AAAE00274ED7 /* add_mul_coreml_all.pte */,
@@ -680,6 +683,7 @@
 				C985519E2AD2542D009143F9 /* mv3_coreml_all.pte in Resources */,
 				C98551A02AD2542D009143F9 /* add_coreml_all.bin in Resources */,
 				C98551A22AD2542D009143F9 /* mul_coreml_all.pte in Resources */,
+				838CA6872CD1965700462190 /* add_mul_compiled_coreml_all.pte in Resources */,
 				8307EB8A2C9262060011AE6D /* state_coreml_all.pte in Resources */,
 				C98551A32AD2542D009143F9 /* add_coreml_all.pte in Resources */,
 			);

backends/apple/coreml/scripts/generate_test_models.sh

@@ -31,3 +31,12 @@ done
 
 echo "Executorch: Generating stateful model"
 python3 "$SCRIPT_DIR_PATH/../runtime/test/export_stateful_model.py"
+
+COMPILE_MODELS=("add_mul")
+echo "Executorch: Generating compiled model"
+for MODEL in "${COMPILE_MODELS[@]}"
+do
+  echo "Executorch: Generating compiled $MODEL model" 
+  python3 -m examples.apple.coreml.scripts.export --model_name "$MODEL" --compile
+  mv -f "$MODEL""_compiled_coreml_all.pte" "$COREML_DIR_PATH/runtime/test/models"
+done

backends/arm/_passes/arm_pass_manager.py

@@ -12,6 +12,7 @@
     AnnotateChannelsLastDimOrder,
 )
 from executorch.backends.arm._passes.cast_int64_pass import CastInt64ToInt32Pass
+from executorch.backends.arm._passes.conv1d_unsqueeze_pass import Conv1dUnsqueezePass
 from executorch.backends.arm._passes.convert_expand_copy_to_repeat import (
     ConvertExpandCopyToRepeatPass,
 )
@@ -69,6 +70,7 @@ def transform_to_backend_pipeline(
         self.add_pass(DecomposeDivPass())
         self.add_pass(InsertSqueezeAfterSumPass())
         self.add_pass(ConvertSplitToSlicePass())
+        self.add_pass(Conv1dUnsqueezePass(exported_program))
         self.add_pass(DecomposeSoftmaxesPass())
         for spec in compile_spec:
             if spec.key == "permute_memory_format":

backends/arm/_passes/arm_pass_utils.py

@@ -1,3 +1,4 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
 # Copyright 2024 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
@@ -9,11 +10,57 @@
 import torch
 import torch.fx
 
+from executorch.exir import ExportedProgram
 from executorch.exir.dialects._ops import ops as exir_ops
+
+from torch._export.utils import (
+    get_buffer,
+    get_lifted_tensor_constant,
+    get_param,
+    is_buffer,
+    is_lifted_tensor_constant,
+    is_param,
+)
 from torch._ops import OpOverload
 from torch._subclasses.fake_tensor import FakeTensor
 
 
+def is_get_attr_node(node: torch.fx.Node) -> bool:
+    """
+    Returns true if the given node is a get attr node for a tensor of the model
+    """
+    return isinstance(node, torch.fx.Node) and node.op == "get_attr"
+
+
+def is_param_node(exp_prog: ExportedProgram, node: torch.fx.Node) -> bool:
+    return (
+        is_get_attr_node(node)
+        or is_param(exp_prog, node)
+        or is_buffer(exp_prog, node)
+        or is_lifted_tensor_constant(exp_prog, node)
+    )
+
+
+def get_param_tensor(
+    exp_prog: ExportedProgram, node: torch.fx.Node
+) -> Optional[torch.Tensor]:
+    if node is None:
+        return None
+    elif is_param(exp_prog, node):
+        return get_param(exp_prog, node)
+    elif is_buffer(exp_prog, node):
+        return get_buffer(exp_prog, node)
+    elif is_lifted_tensor_constant(exp_prog, node):
+        return get_lifted_tensor_constant(exp_prog, node)
+    elif is_get_attr_node(node):
+        # This is a hack to support both lifted and unlifted graph
+        try:
+            return getattr(node.graph.owning_module, node.target)
+        except AttributeError:
+            return getattr(exp_prog.graph_module, node.target)
+    raise RuntimeError(f"unsupported param type, {node.op}.")
+
+
 def create_node(
     graph: torch.fx.Graph,
     op_target: OpOverload,