Update base for Update on "Arm backend: Add 16A8W support for view and transpose operations"

Add 16A8W quantization support for view and transpose operations in ExecutorTorch ARM backend.

This follows the pattern established for linear, mul, sigmoid, tanh, and slice operations, extending int16 support to view and transpose operations.

Changes:
- Add INT16 dtype validation support in op_transpose.py
- Add test_view_tensor_16a8w_tosa_INT test function
- Enable test_view.py in test targets configuration

The 16A8W configuration uses 16-bit activations with 8-bit weights, enabling higher precision for activations while maintaining weight efficiency.

Differential Revision: [D80511313](https://our.internmc.facebook.com/intern/diff/D80511313/)

cc digantdesai freddan80 per zingo oscarandersson8218

[ghstack-poisoned]

Author: Ninja91

.ci/scripts/test_model.sh

@@ -97,7 +97,7 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
-  if [[ "${MODEL_NAME}" == "qwen2_5" ]]; then
+  if [[ "${MODEL_NAME}" == "qwen2_5_1_5b" ]]; then
       # Install requirements for export_llama
       bash examples/models/llama/install_requirements.sh
       # Test export_llm script: python3 -m extension.llm.export.export_llm.

.github/workflows/trunk.yml

@@ -176,7 +176,7 @@ jobs:
           - model: phi_4_mini
             backend: portable
             runner: linux.arm64.m7g.4xlarge
-          - model: qwen2_5
+          - model: qwen2_5_1_5b
             backend: portable
             runner: linux.arm64.2xlarge
           - model: llama3_2_vision_encoder
@@ -823,10 +823,10 @@ jobs:
           --tsv_path ${TSV_PATH}
         echo "::endgroup::"
 
-  test-huggingface-transformers-coreml:
+  test-huggingface-transformers-macos:
     # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
     if: ${{ !github.event.pull_request.head.repo.fork }}
-    name: test-huggingface-transformers-coreml
+    name: test-huggingface-transformers-macos
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     permissions:
       id-token: write
@@ -844,10 +844,10 @@ jobs:
           # phi4-mini|xnnpack|--quantize,
           # smollm2-135m|xnnpack|--quantize,
           # smollm3-3b|xnnpack|--quantize,
+          # qwen3-1.7b|xnnpack|--quantize,
           # CoreML.
           llama3.2-1b|coreml_fp32_gpu|--quantize,
           qwen3-0.6b|coreml_fp32_gpu|--quantize,
-          qwen3-1.7b|xnnpack|--quantize,
           smollm2-135m|coreml_fp32_gpu|--quantize,
           olmo-1b|coreml_fp32_gpu|--quantize,
           bert|coreml_fp32_gpu|--quantize,

README.md

@@ -52,7 +52,7 @@ To get started you can:
 
 - Visit the [Step by Step Tutorial](https://pytorch.org/executorch/stable/getting-started.html) to get things running locally and deploy a model to a device
 - Use this [Colab Notebook](https://colab.research.google.com/drive/1qpxrXC3YdJQzly3mRg-4ayYiOjC6rue3?usp=sharing) to start playing around right away
-- Jump straight into LLM use cases by following specific instructions for popular open-source models such as [Llama](examples/models/llama/README.md), [Qwen 3](examples/models/qwen3/README.md), [Phi-4-mini](examples/models/phi_4_mini/README.md), and [Llava](examples/models/llava/README.md)
+- Jump straight into LLM use cases by following specific instructions for popular open-source models such as [Llama](examples/models/llama/README.md), [Qwen 3](examples/models/qwen3/README.md), [Phi-4-mini](examples/models/phi_4_mini/README.md), [Llava](examples/models/llava/README.md), [Voxtral](examples/models/voxtral/README.md), and [LFM2](examples/models/lfm2/README.md).
 
 ## Feedback and Engagement
 

backends/apple/coreml/TARGETS

@@ -61,16 +61,21 @@ runtime.python_library(
 )
 
 runtime.python_library(
-    name = "recipes",
-    srcs = glob([
-        "recipes/*.py",
-    ]),
+    name = "coreml_recipes",
+    srcs = [
+        "recipes/__init__.py",
+        "recipes/coreml_recipe_provider.py"
+    ],
     visibility = [
         "@EXECUTORCH_CLIENTS",
+        "//executorch/export/...",
     ],
     deps = [
         "fbsource//third-party/pypi/coremltools:coremltools",
+        ":coreml_recipe_types",
         ":backend",
+        ":partitioner",
+        ":quantizer",
         "//caffe2:torch",
         "//executorch/exir:lib",
         "//executorch/exir/backend:compile_spec_schema",
@@ -80,6 +85,20 @@ runtime.python_library(
     ],
 )
 
+runtime.python_library(
+    name = "coreml_recipe_types",
+    srcs = [
+        "recipes/coreml_recipe_types.py",
+    ],
+    visibility = [
+        "@EXECUTORCH_CLIENTS",
+        "//executorch/export/...",
+    ],
+    deps = [
+        "//executorch/export:recipe",
+    ],
+)
+
 runtime.cxx_python_extension(
     name = "executorchcoreml",
     srcs = [
@@ -124,7 +143,7 @@ runtime.python_test(
         "fbsource//third-party/pypi/pytest:pytest",
         ":partitioner",
         ":quantizer",
-        ":recipes",
+        ":coreml_recipes",
         "//caffe2:torch",
         "//pytorch/vision:torchvision",
         "fbsource//third-party/pypi/scikit-learn:scikit-learn",

backends/apple/coreml/compiler/torch_ops.py

@@ -175,7 +175,6 @@ def dequantize_affine(context, node):
         int_data.astype(quantized_np_dtype),
         zero_point,
         scale,
-        axis=-1,
         name=node.name,
     )
     context.add(output, node.name)

backends/apple/coreml/recipes/coreml_recipe_provider.py

@@ -3,6 +3,7 @@
 # Please refer to the license found in the LICENSE file in the root directory of the source tree.
 
 
+import logging
 from typing import Any, Optional, Sequence
 
 import coremltools as ct
@@ -111,8 +112,9 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
 
         unexpected = set(kwargs.keys()) - expected_keys
         if unexpected:
-            raise ValueError(
-                f"Recipe '{recipe_type.value}' received unexpected parameters: {list(unexpected)}"
+            logging.warning(
+                f"CoreML recipe '{recipe_type.value}' ignoring unexpected parameters: {list(unexpected)}. "
+                f"Expected parameters: {list(expected_keys)}"
             )
 
         self._validate_base_parameters(kwargs)
@@ -121,7 +123,13 @@ def _validate_recipe_kwargs(self, recipe_type: RecipeType, **kwargs: Any) -> Non
 
     def _get_expected_keys(self, recipe_type: RecipeType) -> set:
         """Get expected parameter keys for a recipe type"""
-        common_keys = {"minimum_deployment_target", "compute_unit"}
+        common_keys = {
+            "minimum_deployment_target",
+            "compute_unit",
+            "skip_ops_for_coreml_delegation",
+            "lower_full_graph",
+            "take_over_constant_data",
+        }
 
         if recipe_type in [
             CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_GROUP,
@@ -377,9 +385,19 @@ def _get_coreml_lowering_recipe(
         if minimum_deployment_target and minimum_deployment_target < ct.target.iOS18:
             take_over_mutable_buffer = False
 
+        # Extract additional partitioner parameters
+        skip_ops_for_coreml_delegation = kwargs.get(
+            "skip_ops_for_coreml_delegation", None
+        )
+        lower_full_graph = kwargs.get("lower_full_graph", False)
+        take_over_constant_data = kwargs.get("take_over_constant_data", True)
+
         partitioner = CoreMLPartitioner(
             compile_specs=compile_specs,
             take_over_mutable_buffer=take_over_mutable_buffer,
+            skip_ops_for_coreml_delegation=skip_ops_for_coreml_delegation,
+            lower_full_graph=lower_full_graph,
+            take_over_constant_data=take_over_constant_data,
         )
 
         edge_compile_config = EdgeCompileConfig(

backends/apple/coreml/test/test_coreml_recipes.py

@@ -185,14 +185,6 @@ def test_int4_weight_only_per_group_validation(self):
             )
         self.assertIn("must be positive", str(cm.exception))
 
-        # Test unexpected parameter
-        with self.assertRaises(ValueError) as cm:
-            self.provider.create_recipe(
-                CoreMLRecipeType.TORCHAO_INT4_WEIGHT_ONLY_PER_CHANNEL,
-                group_size=32,  # group_size not valid for per-channel
-            )
-        self.assertIn("unexpected parameters", str(cm.exception))
-
     def test_int8_weight_only_per_channel(self):
         """Test INT8 weight-only per-channel quantization"""
         model = TestHelperModules.TwoLinearModule().eval()
@@ -385,23 +377,6 @@ def forward(self, x):
         self._compare_eager_quantized_model_outputs(session, example_inputs, atol=1e-2)
         self._compare_eager_unquantized_model_outputs(session, model, example_inputs)
 
-    def test_pt2e_recipes_parameter_rejection(self):
-        """Test that PT2E recipes reject TorchAO-specific parameters"""
-        # PT2E recipes should reject TorchAO-specific parameters
-        pt2e_recipes = [
-            CoreMLRecipeType.PT2E_INT8_STATIC,
-            CoreMLRecipeType.PT2E_INT8_WEIGHT_ONLY,
-        ]
-        torchao_params = ["filter_fn", "group_size", "bits", "block_size"]
-
-        for recipe_type in pt2e_recipes:
-            for param in torchao_params:
-                with self.subTest(recipe=recipe_type.value, param=param):
-                    kwargs = {param: "dummy_value"}
-                    with self.assertRaises(ValueError) as cm:
-                        self.provider.create_recipe(recipe_type, **kwargs)
-                    self.assertIn("unexpected parameters", str(cm.exception).lower())
-
     def test_filter_fn_comprehensive(self):
         """Comprehensive test for filter_fn parameter functionality"""
 

backends/apple/coreml/test/test_torch_ops.py

@@ -27,9 +27,9 @@
 class TestTorchOps(unittest.TestCase):
     edge_compile_config = executorch.exir.EdgeCompileConfig()
 
-    def _coreml_partitioner(self):
+    def _coreml_partitioner(self, *, minimum_deployment_target=ct.target.iOS18):
         compile_specs = CoreMLBackend.generate_compile_specs(
-            minimum_deployment_target=ct.target.iOS18
+            minimum_deployment_target=minimum_deployment_target
         )
         return CoreMLPartitioner(compile_specs=compile_specs)
 
@@ -158,6 +158,33 @@ def test_dequantize_affine_c8w_embedding_b4w_linear(self):
         et_prog = delegated_program.to_executorch()
         self._compare_outputs(et_prog, model, example_inputs)
 
+    def test_dequantize_affine_c8w_embedding_c8w_linear_ios16(self):
+        model, example_inputs = self._get_test_model()
+        quantize_(
+            model,
+            IntxWeightOnlyConfig(weight_dtype=torch.int8, granularity=PerAxis(0)),
+            lambda m, fqn: isinstance(m, torch.nn.Embedding),
+        )
+        quantize_(
+            model,
+            IntxWeightOnlyConfig(weight_dtype=torch.int8, granularity=PerAxis(0)),
+        )
+        ep = torch.export.export(model, example_inputs)
+        delegated_program = executorch.exir.to_edge_transform_and_lower(
+            ep,
+            partitioner=[
+                self._coreml_partitioner(minimum_deployment_target=ct.target.iOS16)
+            ],
+        )
+        for node in delegated_program.exported_program().graph.nodes:
+            if node.op == "call_function":
+                assert node.target.__name__ in [
+                    "executorch_call_delegate",
+                    "getitem",
+                ], f"Got unexpected node target after delegation: {node.target.__name__}"
+        et_prog = delegated_program.to_executorch()
+        self._compare_outputs(et_prog, model, example_inputs)
+
     def test_dequantize_codebook_linear_per_grouped_col(self):
         model, example_inputs = self._get_test_model()
         quantize_(
@@ -298,6 +325,7 @@ def forward(self, x):
     test_runner.test_dequantize_affine_c4w_embedding()
     test_runner.test_dequantize_affine_c4w_linear()
     test_runner.test_dequantize_affine_c8w_embedding_b4w_linear()
+    test_runner.test_dequantize_affine_c8w_embedding_c8w_linear_ios16()
     test_runner.test_dequantize_codebook_linear_per_grouped_col()
     test_runner.test_dequantize_codebook_linear_per_grouped_row()
     test_runner.test_dequantize_codebook_embedding_per_grouped_col()

backends/arm/_passes/__init__.py

@@ -37,6 +37,7 @@
 from .decompose_cosine_similarity_pass import DecomposeCosineSimilarityPass  # noqa
 from .decompose_cumsum_pass import DecomposeCumsumPass  # noqa
 from .decompose_div_pass import DecomposeDivPass  # noqa
+from .decompose_div_tensor_mode import DecomposeDivTensorModePass  # noqa
 from .decompose_elu_pass import DecomposeEluPass  # noqa
 from .decompose_embedding_pass import DecomposeEmbeddingPass  # noqa  # noqa
 from .decompose_expm1_pass import DecomposeExpm1Pass  # noqa

backends/arm/_passes/arm_pass_manager.py

@@ -42,6 +42,7 @@
     DecomposeCosineSimilarityPass,
     DecomposeCumsumPass,
     DecomposeDivPass,
+    DecomposeDivTensorModePass,
     DecomposeEluPass,
     DecomposeEmbeddingPass,
     DecomposeExpm1Pass,
@@ -211,6 +212,7 @@ def _tosa_FP_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
             DecomposeMeanDimPass(exported_program.graph_module, self.tosa_spec)
         )
         self.add_pass(DecomposeNotEqualPass())
+        self.add_pass(DecomposeDivTensorModePass())
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeSoftmaxPass())
         self.add_pass(DecomposeGeluPass())
@@ -289,6 +291,7 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_pass(DecomposeNotEqualPass())
         self.add_pass(DecomposeCosineSimilarityPass())
         self.add_pass(DecomposeGluPass())
+        self.add_pass(DecomposeDivTensorModePass())
         self.add_pass(DecomposeDivPass())
         self.add_pass(DecomposeLeakyReLUPass())
         self.add_pass(DecomposeLinearVectorNormPass())