Merge branch 'main' into export-D69954671

Author: dpalmasan

.github/workflows/trunk.yml

@@ -374,7 +374,13 @@ jobs:
     secrets: inherit
     strategy:
       matrix:
-        hf_model_repo: [google/gemma-2-2b]
+        hf_model_id: [
+          google/gemma-2-2b,
+          Qwen/Qwen2.5-0.5B,
+          HuggingFaceTB/SmolLM2-135M,
+          meta-llama/Llama-3.2-1B,
+          allenai/OLMo-1B-hf
+        ]
       fail-fast: false
     with:
       secrets-env: EXECUTORCH_HF_TOKEN
@@ -389,66 +395,39 @@ jobs:
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh cmake
-
-        echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
-        rm -rf cmake-out
-        cmake \
-            -DCMAKE_INSTALL_PREFIX=cmake-out \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
-            -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
-            -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
-            -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
-            -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
-            -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
-            -DEXECUTORCH_BUILD_XNNPACK=ON \
-            -DPYTHON_EXECUTABLE=python \
-            -Bcmake-out .
-        cmake --build cmake-out -j9 --target install --config Release
-
-        echo "Build llama runner"
-        dir="examples/models/llama"
-        cmake \
-            -DCMAKE_INSTALL_PREFIX=cmake-out \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON \
-            -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
-            -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
-            -DEXECUTORCH_BUILD_XNNPACK=ON \
-            -DPYTHON_EXECUTABLE=python \
-            -Bcmake-out/${dir} \
-            ${dir}
-        cmake --build cmake-out/${dir} -j9 --config Release
         echo "::endgroup::"
 
-        echo "::group::Set up HuggingFace Dependencies"
-        if [ -z "$SECRET_EXECUTORCH_HF_TOKEN" ]; then
-          echo "::error::SECRET_EXECUTORCH_HF_TOKEN is empty. For security reason secrets won't be accessible on forked PRs. Please make sure you submit a non-forked PR."
-          exit 1
-        fi
+        echo "::group::Set up Hugging Face"
         pip install -U "huggingface_hub[cli]"
         huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
+        git clone https://github.com/huggingface/optimum-executorch
+        cd optimum-executorch
+        # There is no release yet, for CI stability, always test from the same commit on main
+        git checkout 6a7e83f3eee2976fa809335bfb78a45b1ea1cb25
+        pip install .
         pip install accelerate sentencepiece
         pip list
         echo "::endgroup::"
 
-        echo "::group::Export to ExecuTorch"
-        TOKENIZER_FILE=tokenizer.model
-        TOKENIZER_BIN_FILE=tokenizer.bin
-        ET_MODEL_NAME=et_model
-        DOWNLOADED_TOKENIZER_FILE_PATH=$(bash .ci/scripts/download_hf_hub.sh --model_id "${{ matrix.hf_model_repo }}" --files "${TOKENIZER_FILE}")
-        if [ -f "$DOWNLOADED_TOKENIZER_FILE_PATH/$TOKENIZER_FILE" ]; then
-            echo "${TOKENIZER_FILE} downloaded successfully at: $DOWNLOADED_TOKENIZER_FILE_PATH"
-            python -m extension.llm.tokenizer.tokenizer -t "$DOWNLOADED_TOKENIZER_FILE_PATH/$TOKENIZER_FILE" -o ./${TOKENIZER_BIN_FILE}
-            ls ./tokenizer.bin
-        else
-            echo "Failed to download ${TOKENIZER_FILE} from ${{ matrix.hf_model_repo }}."
-            exit 1
-        fi
-
-        python -m extension.export_util.export_hf_model -hfm=${{ matrix.hf_model_repo }} -o ${ET_MODEL_NAME}
-
-        cmake-out/examples/models/llama/llama_main --model_path=${ET_MODEL_NAME}.pte --tokenizer_path=${TOKENIZER_BIN_FILE} --prompt="My name is"
+        echo "::group::Export and Run ${{ matrix.hf_model_id }}"
+        # Pass matrix variable as environment variable
+        export MODEL_ID="${{ matrix.hf_model_id }}"
+        python -c "
+        import os
+        from optimum.executorch import ExecuTorchModelForCausalLM
+        from transformers import AutoTokenizer
+
+        model_id = os.getenv('MODEL_ID')
+        print(f'Loading model: {model_id}')
+        model = ExecuTorchModelForCausalLM.from_pretrained(model_id, recipe='xnnpack')
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        generated_text = model.text_generation(
+          tokenizer=tokenizer,
+          prompt='Simply put, the theory of relativity states that',
+          max_seq_len=64
+        )
+        print(generated_text)
+        "
         echo "::endgroup::"
 
 

backends/apple/coreml/TARGETS

@@ -76,6 +76,7 @@ runtime.cxx_python_extension(
     base_module = "",
     visibility = [
         "//executorch/examples/apple/coreml/...",
+        "@EXECUTORCH_CLIENTS",
     ],
     external_deps = [
         "pybind11",

backends/arm/operator_support/TARGETS

@@ -5,8 +5,9 @@ python_library(
     srcs = glob(["*.py"]),
     typing = True,
     deps = [
+        "//executorch/backends/arm/_passes:passes",
+        "//executorch/backends/arm:tosa_specification",
         "//executorch/backends/xnnpack/_passes:xnnpack_passes",
         "//executorch/exir:lib",
-        "//executorch/backends/arm:tosa_specification"
     ],
 )

backends/arm/test/misc/test_partition_decomposed_quantized_ops.py

@@ -60,6 +60,9 @@ def test_softplus_tosa_BI(test_data: input_t1):
     pipeline.pop_stage("check_not.exir")
     # check that all ops in exir_op except add are rejected
     pipeline.add_stage_after(
-        "partition", pipeline.tester.check, exir_op[1:], suffix="exir_post_partition"
+        "to_edge_transform_and_lower",
+        pipeline.tester.check,
+        exir_op[1:],
+        suffix="exir_post_partition",
     )
     pipeline.run()

backends/arm/test/ops/test_conv2d.py

@@ -370,7 +370,7 @@ def test_conv2d_tosa_BI(test_module):
     pipeline = TosaPipelineBI[input_t](
         test_module, test_module.get_inputs(), aten_op, exir_op
     )
-    pipeline.change_args("run_method_and_compare_outputs.0", qtol=1)
+    pipeline.change_args("run_method_and_compare_outputs", qtol=1)
     pipeline.run()
 
 

backends/arm/test/passes/test_cast_int64_pass.py

@@ -8,7 +8,7 @@
 import torch
 from executorch.backends.arm._passes.cast_int64_pass import CastInt64ToInt32Pass
 
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 input_t = Tuple[torch.Tensor]  # Input x
 
@@ -28,7 +28,7 @@ def test_int64_model_tosa_BI():
         "executorch_exir_dialects_edge__ops_dim_order_ops__to_dim_order_copy_default": 1,
         "executorch_exir_dialects_edge__ops_aten_add_Tensor": 1,
     }
-    pipeline = TestPassPipeline[input_t](
+    pipeline = PassPipeline[input_t](
         module,
         module.get_inputs(),
         tosa_version="TOSA-0.80+BI",

backends/arm/test/passes/test_fold_qdq_pass.py

@@ -9,7 +9,7 @@
 from executorch.backends.arm._passes.fold_qdq_with_annotated_qparams_pass import (
     FoldAndAnnotateQParamsPass,
 )
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 
 input_t = Tuple[torch.Tensor, torch.Tensor]  # Input x, y
@@ -32,7 +32,7 @@ def test_fold_qdq_pass_tosa_BI():
     is removed from the representation.
     """
     module = SimpleQuantizeModel()
-    pipeline = TestPassPipeline[input_t](
+    pipeline = PassPipeline[input_t](
         module,
         module.get_inputs(),
         tosa_version="TOSA-0.80+BI",

backends/arm/test/passes/test_fuse_batchnorm_pass.py

@@ -8,7 +8,7 @@
 import torch
 from executorch.backends.arm._passes.fuse_batchnorm2d_pass import FuseBatchnorm2DPass
 from executorch.backends.arm.test import common
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 input_t = Tuple[torch.Tensor]  # Input x
 
@@ -138,7 +138,7 @@ def forward(self, x):
 @common.parametrize("module", modules)
 def test_fuse_batchnorm_tosa_MI(module):
     """Test various cases where the batchnorm should and shouldn't be fused."""
-    pipeline = TestPassPipeline[input_t](
+    pipeline = PassPipeline[input_t](
         module,
         module.get_inputs(),
         tosa_version="TOSA-0.80+MI",

backends/arm/test/passes/test_insert_table_ops_pass.py

@@ -11,7 +11,7 @@
     FoldAndAnnotateQParamsPass,
 )
 from executorch.backends.arm._passes.insert_table_ops import InsertTableOpsPass
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 input_t = Tuple[torch.Tensor]  # Input x
 
@@ -27,7 +27,7 @@ def get_inputs(self) -> input_t:
 
 def test_insert_table_tosa_BI():
     module = Sigmoid()
-    pipeline = TestPassPipeline[input_t](
+    pipeline = PassPipeline[input_t](
         module,
         module.get_inputs(),
         tosa_version="TOSA-0.80+BI",

backends/arm/test/passes/test_meandim_to_averagepool2d.py

@@ -11,7 +11,7 @@
     ConvertMeanDimToAveragePoolPass,
 )
 from executorch.backends.arm.test import common
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 
 input_t = Tuple[torch.Tensor, torch.Tensor]  # Input x
@@ -65,7 +65,7 @@ def test_meandim_to_avgpool_tosa_BI(module):
     Tests the MeanDimToAveragePool2dPass which converts mean.dim to average_pool2d
     for the special case where dim is [-1, -2] and keepdim is True.
     """
-    pipeline = TestPassPipeline[input_t](
+    pipeline = PassPipeline[input_t](
         module,
         module.get_inputs(),
         tosa_version="TOSA-0.80+BI",

backends/arm/test/passes/test_unsqueeze_before_repeat_pass.py

@@ -10,7 +10,7 @@
     UnsqueezeBeforeRepeatPass,
 )
 from executorch.backends.arm.test import common
-from executorch.backends.arm.test.tester.test_pipeline import TestPassPipeline
+from executorch.backends.arm.test.tester.test_pipeline import PassPipeline
 
 input_t = Tuple[
     torch.Tensor, Dict[str, int], list[str]
@@ -47,7 +47,7 @@ def test_unsqueeze_before_repeat_tosa_MI(test_data):
     """
     module = Repeat()
     data, ops_after_pass, ops_not_after_pass = test_data
-    pipeline = TestPassPipeline(
+    pipeline = PassPipeline(
         module,
         data,
         tosa_version="TOSA-0.80+MI",