8.0b2 Release (#2308)

Author: jakesabathia2

NOTICE.txt

@@ -46,3 +46,36 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
+
+This project contains content in the files coremltools/optimize/torch/quantization/modules/conv_transpose.py and coremltools/optimize/torch/quantization/modules/conv_transpose_fused.py which are adapted from pytorch (https://github.com/pytorch/). The license for these follows:
+
+Copyright (c) 2016 Facebook, Inc (Adam Paszke)
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
+   and IDIAP Research Institute nor the names of its contributors may be
+   used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.

README.md

@@ -26,9 +26,9 @@ With coremltools, you can:
 After conversion, you can integrate the Core ML models with your app using Xcode.
 
 ## Install 8.0 Beta
-The [coremltools version 8 beta 1](https://github.com/apple/coremltools/releases/tag/8.0b1) is now out. To install, run the following command in your terminal:
+The [coremltools version 8 beta 2](https://github.com/apple/coremltools/releases/tag/8.0b2) is now out. To install, run the following command in your terminal:
 ```shell
-pip install coremltools==8.0b1
+pip install coremltools==8.0b2
 ```
 
 
@@ -46,11 +46,9 @@ pip install -U coremltools
 
 To install coremltools, see [Installing Core ML Tools](https://apple.github.io/coremltools/docs-guides/source/installing-coremltools.html). For more information, see the following:
 
-* [Release Notes](https://github.com/apple/coremltools/releases/) 
-* [Guide and examples](https://apple.github.io/coremltools/docs-guides/index.html) 
+* [Release Notes](https://github.com/apple/coremltools/releases/)
+* [Guide and examples](https://apple.github.io/coremltools/docs-guides/index.html)
 * [API Reference](https://apple.github.io/coremltools/index.html)
 * [Core ML Specification](https://apple.github.io/coremltools/mlmodel/index.html)
 * [Building from Source](BUILDING.md)
-* [Contribution Guidelines](CONTRIBUTING.md) 
-
-
+* [Contribution Guidelines](CONTRIBUTING.md)

coremltools/_deps/__init__.py

@@ -54,7 +54,7 @@ def _warn_if_above_max_supported_version(package_name, package_version, max_supp
 _HAS_SKLEARN = True
 _SKLEARN_VERSION = None
 _SKLEARN_MIN_VERSION = "0.17"
-_SKLEARN_MAX_VERSION = "1.1.2"
+_SKLEARN_MAX_VERSION = "1.5.1"
 
 
 def __get_sklearn_version(version):
@@ -197,6 +197,14 @@ def __get_sklearn_version(version):
 else:
     _HAS_SCIPY = True
 
+# ---------------------------------------------------------------------------------------
+try:
+    import transformers
+except:
+    _HAS_HF = False
+else:
+    _HAS_HF = True
+
 # General utils
 def version_ge(module, target_version):
     """

coremltools/converters/mil/backend/mil/load.py

@@ -6,7 +6,7 @@
 import os
 import warnings
 from collections import OrderedDict
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import numpy as np
 
@@ -62,12 +62,19 @@
     BlobWriter = None
 
 
-def should_use_weight_file(val):
+def should_use_weight_file(
+    val: Union[np.ndarray, np.generic],
+    specification_version: Optional[int] = _SPECIFICATION_VERSION_IOS_15,
+) -> bool:
+    # additional dtype are supported >= iOS18
+    supported_dtypes = ["float16", "float32", "uint8", "int8"]
+    if specification_version >= _SPECIFICATION_VERSION_IOS_18:
+        supported_dtypes += ["uint16", "int16", "int32", "uint32"]
     return (
         val is not None
         and isinstance(val, (np.ndarray, np.generic))
         and val.size >= 10
-        and val.dtype in ['float16', 'float32', 'uint8', 'int8']
+        and val.dtype in supported_dtypes
     )
 
 
@@ -80,13 +87,26 @@ def __init__(
         self,
         prog: Program,
         weights_dir: str,
+        specification_version: int,
     ):
         self.prog = prog
         self.weights_dir = weights_dir
+        self.specification_version = specification_version
         self.blob_writers = {}
         self.weight_id_to_file_value = {}  # mapping from weight_id to file value
         self.prog.validate(check_essential_scope=True)
 
+    @staticmethod
+    def _get_valid_kwargs(kwargs: Dict[str, Any]) -> Dict[str, Any]:
+        """
+        Get a valid kwargs to initialize a MILProtoExporter object.
+        """
+        return {
+            "prog": kwargs["prog"],
+            "weights_dir": kwargs["weights_dir"],
+            "specification_version": kwargs["specification_version"],
+        }
+
     def translate_program_attributes(self) -> Dict[str, Any]:
         """
         Get the program attributes which need to be exported to mil proto.
@@ -157,7 +177,7 @@ def get_milproto_value(self, var: Var) -> proto.MIL_pb2.Value:
         """
         Translate a pymil Var into milproto value.
         """
-        if should_use_weight_file(var.val):
+        if should_use_weight_file(var.val, self.specification_version):
             return self.create_file_value(var)
         else:
             return create_immediate_value(var)
@@ -475,9 +495,7 @@ def convert_function(self, function: Function, opset: str) -> proto.MIL_pb2.Func
             inputs=inputs, opset=opset, block_specializations={opset: block}
         )
 
-    def export(
-        self, specification_version: Optional[str] = _SPECIFICATION_VERSION_IOS_15
-    ) -> proto.MIL_pb2.Program:
+    def export(self) -> proto.MIL_pb2.Program:
         """
         Export a pymil program into mil proto with the given specification version.
         """
@@ -486,7 +504,9 @@ def export(
 
         function_protos = {}
         for func_name, func in self.prog.functions.items():
-            function_protos[func_name] = self.convert_function(func, _OPSET[specification_version])
+            function_protos[func_name] = self.convert_function(
+                func, _OPSET[self.specification_version]
+            )
 
         kwargs = {
             "version": 1,
@@ -1040,11 +1060,14 @@ def load(
         )
 
     # convert pymil program into mil proto
+    kwargs["prog"] = prog
+    kwargs["weights_dir"] = weights_dir
+    kwargs["specification_version"] = specification_version
+    exporter_kwargs = MILProtoExporter._get_valid_kwargs(kwargs)
     mil_proto_exporter = MILProtoExporter(
-        prog,
-        weights_dir,
+        **exporter_kwargs,
     )
-    mil_proto = mil_proto_exporter.export(specification_version)
+    mil_proto = mil_proto_exporter.export()
 
     # return the model provided by users
     desc = kwargs.get("model_description", None)

coremltools/converters/mil/backend/mil/test_load.py

@@ -3,7 +3,9 @@
 #  Use of this source code is governed by a BSD-3-clause license that can be
 #  found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
 
+import itertools
 import math
+import os
 import platform
 import shutil
 import tempfile
@@ -22,9 +24,68 @@
     TestConstexprLut as _TestConstexprLut,
 )
 from coremltools.converters.mil.mil.program import Symbol
+from coremltools.converters.mil.mil.types.type_mapping import string_to_nptype
 from coremltools.models.utils import _macos_version
 
 
+class TestWeightFileSerialization:
+    @staticmethod
+    @pytest.mark.parametrize(
+        "dtype, opset_version",
+        itertools.product(
+            ["fp16", "fp32", "uint8", "int8", "uint16", "int16", "int32", "uint32"],
+            [ct.target.iOS16, ct.target.iOS18],
+        ),
+    )
+    def test_weight_serialization(dtype, opset_version):
+        if dtype == "uint32":
+            # There is a pass that casts the output to CoreML supported dtype.
+            # uint32 will fail because `cast` op doesn't accept such input type.
+            pytest.skip("uint32 is not supported in `cast` op.")
+        if dtype in ["uint8", "int8", "uint16", "int16"] and opset_version == ct.target.iOS16:
+            # iOS16 doesn't support the above dtype either
+            pytest.skip("dtype not support in iOS16")
+
+        if dtype in ["fp16", "fp32", "uint8", "int8"]:
+            should_serialize_weight = True
+        else:
+            should_serialize_weight = opset_version >= ct.target.iOS18
+
+        @mb.program(input_specs=[mb.TensorSpec((1,))], opset_version=opset_version)
+        def prog(x):
+            val = np.random.rand(1000).astype(string_to_nptype(dtype))
+            return mb.const(val=val), mb.add(x=x, y=1.0)
+
+        # we don't want the const to be constant folding after casting
+        pipeline = ct.PassPipeline()
+        pipeline.set_options("common::const_elimination", {"skip_const_by_size": "-1"})
+        mlmodel = ct.convert(
+            prog,
+            minimum_deployment_target=opset_version,
+            pass_pipeline=pipeline,
+        )
+        saved_package_path = tempfile.mkdtemp(suffix=".mlpackage")
+        mlmodel.save(saved_package_path)
+
+        # check the weights are serialized as file value
+        if ct.utils._macos_version() >= (15, 0):
+            with tempfile.TemporaryDirectory() as serialize_dir:
+                os.system(f"coremlcompiler compile {saved_package_path} {serialize_dir}")
+                model_name_with_extension = os.path.basename(saved_package_path)
+                model_name_wo_extension, _ = os.path.splitext(model_name_with_extension)
+                mil_file = open(
+                    os.path.join(serialize_dir, f"{model_name_wo_extension}.mlmodelc", "model.mil")
+                )
+                mil_txt = mil_file.read()
+                if should_serialize_weight:
+                    assert f"tensor<{dtype}, [1000]>(BLOBFILE" in mil_txt
+                else:
+                    assert f"tensor<{dtype}, [1000]>(BLOBFILE" not in mil_txt
+
+        # cleanup
+        shutil.rmtree(saved_package_path)
+
+
 class TestMILFlexibleShapes:
     @mb.program(input_specs=[mb.TensorSpec(shape=[1, 3, Symbol("H"), Symbol("W")])])
     def basic_network(x):
@@ -991,9 +1052,11 @@ def prog(state, x, y):
             prog,
             convert_to="mlprogram",
             minimum_deployment_target=ct.target.iOS18,
+            skip_model_load=True,
         )
 
         mil = mlmodel.get_spec().mlProgram
+
         for function in mil.functions.values():
             for block in function.block_specializations.values():
                 ops = list(block.operations)
@@ -1005,6 +1068,50 @@ def prog(state, x, y):
                 ]
                 assert [val.type for val in ops] == expected_ops
 
+    @staticmethod
+    def test_coreml_update_state_lowering_with_prefer_state_in_downstream():
+        @mb.program(
+            input_specs=[
+                mb.StateTensorSpec((1,), dtype=types.fp16),
+                mb.TensorSpec((1,), dtype=types.fp16),
+                mb.TensorSpec((1,), dtype=types.fp16),
+                mb.TensorSpec((1,), dtype=types.fp16),
+            ],
+            opset_version=ct.target.iOS18,
+        )
+        def prog(state, x, y, z):
+            # Although seemingly not used, graph pass prefer_state_in_downstream will
+            # make its output as identiy.x
+            mb.coreml_update_state(state=state, value=x)
+            # If value only feeds into coreml_update_state,
+            # the prefer_state_in_downstream has no affects
+            mb.coreml_update_state(state=state, value=y)
+            # This is the one that really is not used
+            mb.coreml_update_state(state=state, value=z)
+            return mb.identity(x=x), mb.coreml_update_state(state=state, value=y)
+
+        mlmodel = ct.convert(
+            prog,
+            convert_to="mlprogram",
+            minimum_deployment_target=ct.target.iOS18,
+            skip_model_load=True,
+        )
+
+        mil = mlmodel.get_spec().mlProgram
+        for function in mil.functions.values():
+            for block in function.block_specializations.values():
+                ops = list(block.operations)
+                expected_ops = [
+                    "write_state",
+                    "read_state",
+                    "write_state",
+                    "write_state",
+                    "identity",
+                    "write_state",
+                    "read_state",
+                ]
+                assert [val.type for val in ops] == expected_ops
+
     @staticmethod
     @pytest.mark.skipif(ct.utils._macos_version() < (15, 0),
                     reason="State only supported on macOS 15+")

coremltools/converters/mil/converter.py

@@ -278,7 +278,7 @@ def mil_convert_to_proto(
         # If the client calls `mil_convert` directly, the `pass_pipeline` is None. To keep the
         # behaviour same as before, the quantization pass is removed in this situation.
         # TODO: rdar://106111553 ([Infra] Quantization Pass is skipped when `mil_convert` is called directly.)
-        main_pipeline = PassPipeline()
+        main_pipeline = kwargs.get("pass_pipeline", PassPipeline())
         main_pipeline.remove_passes({"common::add_fp16_cast", "common::add_int16_cast"})
     frontend_pipeline, backend_pipeline = _construct_other_pipelines(
         main_pipeline, convert_from, convert_to

coremltools/converters/mil/experimental/passes/generic_conv_scale_fusion.py

@@ -115,7 +115,7 @@ def var_constraints(pattern):
     if not is_scalar:
         conv_weight = pattern.conv.weight.val
         passed = passed and (
-            np.product(scale.shape) == Cout
+            np.prod(scale.shape) == Cout
             or (len(scale.shape) == len(conv_weight.shape) and scale.shape[1] == Cout)
             or (len(scale.shape) == len(conv_weight.shape) - 1 and scale.shape[0] == Cout)
         )