7.2 release (#2196)

Co-authored-by: yifan_shen3 <[email protected]>

coremltools/converters/mil/backend/mil/helper.py

@@ -42,16 +42,6 @@ def create_valuetype_list(length, elem_shape, dtype):
     update_listtype(v_type.listType, length, elem_shape, dtype)
     return v_type
 
-def create_valuetype_dict(key_type, value_type):
-    """
-    Return proto.MIL_pb2.ValueType with dict (dictionaryType) set
-    """
-    v_type = proto.MIL_pb2.ValueType()
-    v_type.dictionaryType.keyType.CopyFrom(types_to_proto(key_type))
-    v_type.dictionaryType.valueType.CopyFrom(types_to_proto(value_type))
-    return v_type
-
-
 def create_valuetype_tensor(shape, data_type):
     """
     Return proto.MIL_pb2.ValueType with tensor (TensorType) set.
@@ -261,40 +251,6 @@ def types_to_proto_primitive(valuetype):
         )
     return types.BUILTIN_TO_PROTO_TYPES[valuetype]
 
-
-def types_to_proto(valuetype):
-    if types.is_tensor(valuetype):
-        primitive = types_to_proto_primitive(valuetype.get_primitive())
-        return create_valuetype_tensor(valuetype.get_shape(), primitive)
-    elif types.is_tuple(valuetype):
-        v_type = proto.MIL_pb2.ValueType()
-        t_type = v_type.tupleType
-        for t in valuetype.T:
-            new_v_type = t_type.types.add()
-            new_v_type.CopyFrom(types_to_proto(t))
-        return v_type
-    elif types.is_list(valuetype):
-        elem = valuetype.T[0]
-        length = valuetype.T[1]
-        if types.is_tensor(elem):
-            dtype = types_to_proto_primitive(elem.get_primitive())
-            elem_shape = elem.get_shape()
-        elif types.is_scalar(elem):
-            dtype = types_to_proto_primitive(valuetype)
-            elem_shape = ()
-        elif types.is_str(elem):
-            dtype = types_to_proto_primitive(elem)
-            elem_shape = ()
-        else:
-            raise NotImplementedError("Only list of either tensors or scalars supported. "
-                                      "Got element of type {}".format(elem.__type_info__()))
-        return create_valuetype_list(length=length, elem_shape=elem_shape, dtype=dtype)
-    elif types.is_dict(valuetype):
-        return create_valuetype_dict(valuetype.T[0], valuetype.T[1])
-    else:
-        return create_valuetype_scalar(types_to_proto_primitive(valuetype))
-
-
 def _get_offset_by_writing_data(output_var, blob_writer):
     if output_var.val.dtype.kind == 'f' and output_var.val.dtype.itemsize == 4:
         offset = blob_writer.write_float_data(np.ascontiguousarray(output_var.val.flatten()))

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

@@ -22,7 +22,9 @@
     create_immediate_value,
     create_list_scalarvalue,
     create_scalar_value,
-    types_to_proto,
+    create_valuetype_list,
+    create_valuetype_scalar,
+    create_valuetype_tensor,
     types_to_proto_primitive,
 )
 from coremltools.converters.mil.backend.nn.load import _set_optional_inputs
@@ -158,7 +160,7 @@ def translate_const(self, op: Operation) -> proto.MIL_pb2.Operation:
             attributes={"name": create_scalar_value(op.name), "val": value},
             outputs=[
                 proto.MIL_pb2.NamedValueType(
-                    name=output_var.name, type=types_to_proto(output_var.sym_type)
+                    name=output_var.name, type=self.types_to_proto(output_var.sym_type)
                 )
             ],
         )
@@ -190,12 +192,58 @@ def translate_constexpr(self, op: Operation) -> proto.MIL_pb2.Operation:
             attributes=attributes,
             outputs=[
                 proto.MIL_pb2.NamedValueType(
-                    name=output_var.name, type=types_to_proto(output_var.sym_type)
+                    name=output_var.name, type=self.types_to_proto(output_var.sym_type)
                 )
                 for output_var in op.outputs
             ],
         )
 
+    def create_valuetype_dict(self, key_type: type, value_type: type) -> proto.MIL_pb2.ValueType:
+        """
+        Return proto.MIL_pb2.ValueType with dict (dictionaryType) set
+        """
+        v_type = proto.MIL_pb2.ValueType()
+        v_type.dictionaryType.keyType.CopyFrom(self.types_to_proto(key_type))
+        v_type.dictionaryType.valueType.CopyFrom(self.types_to_proto(value_type))
+        return v_type
+
+    def types_to_proto(self, valuetype: type) -> proto.MIL_pb2.ValueType:
+        """
+        Return proto.MIL_pb2.ValueType from PyMIL types.
+        """
+        if types.is_tensor(valuetype):
+            primitive = types_to_proto_primitive(valuetype.get_primitive())
+            return create_valuetype_tensor(valuetype.get_shape(), primitive)
+        elif types.is_tuple(valuetype):
+            v_type = proto.MIL_pb2.ValueType()
+            t_type = v_type.tupleType
+            for t in valuetype.T:
+                new_v_type = t_type.types.add()
+                new_v_type.CopyFrom(self.types_to_proto(t))
+            return v_type
+        elif types.is_list(valuetype):
+            elem = valuetype.T[0]
+            length = valuetype.T[1]
+            if types.is_tensor(elem):
+                dtype = types_to_proto_primitive(elem.get_primitive())
+                elem_shape = elem.get_shape()
+            elif types.is_scalar(elem):
+                dtype = types_to_proto_primitive(valuetype)
+                elem_shape = ()
+            elif types.is_str(elem):
+                dtype = types_to_proto_primitive(elem)
+                elem_shape = ()
+            else:
+                raise NotImplementedError(
+                    "Only list of either tensors or scalars supported. "
+                    "Got element of type {}".format(elem.__type_info__())
+                )
+            return create_valuetype_list(length=length, elem_shape=elem_shape, dtype=dtype)
+        elif types.is_dict(valuetype):
+            return self.create_valuetype_dict(valuetype.T[0], valuetype.T[1])
+        else:
+            return create_valuetype_scalar(types_to_proto_primitive(valuetype))
+
     def translate_generic_op(
         self, op: Operation, literal_params: Optional[List[str]] = None
     ) -> proto.MIL_pb2.Operation:
@@ -228,7 +276,7 @@ def translate_generic_op(
             inputs[param_name] = args
 
         outputs = [
-            proto.MIL_pb2.NamedValueType(name=v.name, type=types_to_proto(v.sym_type))
+            proto.MIL_pb2.NamedValueType(name=v.name, type=self.types_to_proto(v.sym_type))
             for v in op.outputs
         ]
         blocks = None
@@ -311,14 +359,18 @@ def feeds_to_only_constexprs(op: Operation) -> bool:
                 literal_params = ["begins", "ends", "end_masks"]
                 proto_ops.append(self.translate_generic_op(op, literal_params))
             else:
-                proto_ops.append(self.translate_generic_op(op))
+                # A single pymil op might be decomposed into multiple ops
+                ops = self.translate_generic_op(op)
+                if not isinstance(ops, list):
+                    ops = [ops]
+                proto_ops.extend(ops)
 
         inputs = []
         if not isinstance(block, Function):
             # Function is subclass of Block, but function's block has no input,
             # and hence skipping reading the block inputs.
             for var in block.inputs:
-                proto_type = types_to_proto(var.sym_type)
+                proto_type = self.types_to_proto(var.sym_type)
                 inputs.append(proto.MIL_pb2.NamedValueType(name=var.name, type=proto_type))
         output_names = [v.name for v in block.outputs]
         return proto.MIL_pb2.Block(inputs=inputs, outputs=output_names, operations=proto_ops)
@@ -331,7 +383,7 @@ def convert_function(self, function: Function, opset: str) -> proto.MIL_pb2.Func
 
         inputs = []
         for name, var in function.inputs.items():
-            proto_type = types_to_proto(var.sym_type)
+            proto_type = self.types_to_proto(var.sym_type)
             inputs.append(proto.MIL_pb2.NamedValueType(name=name, type=proto_type))
 
         return proto.MIL_pb2.Function(
@@ -467,6 +519,15 @@ def get_additional_kwargs(kwargs: Dict[str, Any]) -> Dict[str, Any]:
         """
         return {}
 
+    @staticmethod
+    def _try_convert_other_input_type(
+        input_var: Var, input_features: List[proto.Model_pb2.FeatureDescription]
+    ) -> bool:
+        """
+        Try to convert an input var with additional type.
+        """
+        return False
+
     def get_func_input(self, func: mil.Function) -> List[proto.Model_pb2.FeatureDescription]:
         """
         Utils to get function input feature description.
@@ -554,7 +615,7 @@ def get_func_input(self, func: mil.Function) -> List[proto.Model_pb2.FeatureDesc
                 input_features.append(
                     proto.Model_pb2.FeatureDescription(name=var.name, type=input_feature_type)
                 )
-            else:
+            elif not self._try_convert_other_input_type(var, input_features):
                 raise NotImplementedError(f"Unsupported input type {var.sym_type}.")
 
             if not is_input_shape_symbolic:
@@ -746,6 +807,16 @@ def get_func_output(self, func: mil.Function) -> List[proto.Model_pb2.FeatureDes
 
         return output_features
 
+    def create_model_description(
+        self,
+        input_features: List[proto.Model_pb2.FeatureDescription],
+        output_features: List[proto.Model_pb2.FeatureDescription],
+    ) -> proto.Model_pb2.ModelDescription:
+        """
+        Create model description from input and output features
+        """
+        return proto.Model_pb2.ModelDescription(input=input_features, output=output_features)
+
     def get_coreml_model(
         self,
         input: Dict[str, List[proto.Model_pb2.FeatureDescription]],
@@ -758,7 +829,7 @@ def get_coreml_model(
         # Model description
         input_features = input[self._DEFAULT_FUNCTION_NAME]
         output_features = output[self._DEFAULT_FUNCTION_NAME]
-        desc = proto.Model_pb2.ModelDescription(input=input_features, output=output_features)
+        desc = self.create_model_description(input_features, output_features)
 
         if self.classifier_config is not None:
             desc.predictedFeatureName = self.predicted_feature_name

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

@@ -12,6 +12,7 @@
     ImageType,
     RangeDim,
     Shape,
+    TensorType,
 )
 from coremltools.converters.mil.mil import types
 from coremltools.converters.mil.mil.types.symbolic import any_symbolic, any_variadic, is_symbolic
@@ -169,7 +170,7 @@ def _set_optional_inputs(proto, input_types):
     # Set default values for optional input_types
     default_map = {}
     for input_type in input_types:
-        if isinstance(input_type, ImageType):
+        if not isinstance(input_type, TensorType):
             continue
         if input_type.default_value is not None:
             default_map[input_type.name] = input_type.default_value

coremltools/converters/mil/frontend/_utils.py

@@ -512,33 +512,36 @@ def _concat_dims(dims, none_if_empty=False):
         return ab
 
 
-def _lower_scaled_dot_product_attention(q: Var, k: Var, v: Var, mask: Var, name: str) -> Var:
+def _lower_scaled_dot_product_attention(
+    q: Var, k: Var, v: Var, mask: Var, name: str, before_op: Optional[Operation] = None
+) -> Var:
     # scale the query input
     embed_size = q.shape[-1]
     if is_symbolic(embed_size):
         raise ValueError(
             "The embedding size, i.e. last dimension of the shape of query tensor"
             " cannot be symbolic, in scaled_dot_product_attention op"
         )
+
+    q, k, v = promote_input_dtypes([q, k, v])
     multiplicative_scale_factor = 1 / math.sqrt(embed_size)
-    q, k, v, multiplicative_scale_factor = promote_input_dtypes(
-        [q, k, v, multiplicative_scale_factor]
-    )
-    q = mb.mul(x=q, y=multiplicative_scale_factor)
+    if types.builtin_to_string(q.dtype) == "fp16":
+        multiplicative_scale_factor = _np.float16(multiplicative_scale_factor)
+    q = mb.mul(x=q, y=multiplicative_scale_factor, before_op=before_op)
 
     # multiply query and key input tensors
     # shape of output: (target_seq, source_seq) or (B,...,target_seq, source_seq)
-    attn_weights = mb.matmul(x=q, y=k, transpose_y=True)
+    attn_weights = mb.matmul(x=q, y=k, transpose_y=True, before_op=before_op)
 
     # add mask if applicable
     if mask is not None:
-        attn_weights = mb.add(x=attn_weights, y=mask)
+        attn_weights = mb.add(x=attn_weights, y=mask, before_op=before_op)
 
     # do softmax
-    attn_weights_normalized = mb.softmax(x=attn_weights, axis=-1)
+    attn_weights_normalized = mb.softmax(x=attn_weights, axis=-1, before_op=before_op)
 
     # multiply attn_weights and value tensor
-    res = mb.matmul(x=attn_weights_normalized, y=v, name=name)
+    res = mb.matmul(x=attn_weights_normalized, y=v, name=name, before_op=before_op)
     return res
 
 
@@ -549,7 +552,7 @@ def _construct_constexpr_affine_op(
     axis: Optional[Union[Var, int]] = None,
     name: Optional[str] = None,
     before_op: Optional[Operation] = None,
-) -> Operation:
+) -> Var:
     """Constructs the constexpr op to represent the dequantized weight from PyTorch's data."""
     # The constexpr_affine_dequantize op requires axis.
     if axis is None:

coremltools/converters/mil/frontend/torch/exir_utils.py

@@ -87,12 +87,27 @@ def extract_inputs_from_exir_program(
             val = node.meta["val"]
             assert isinstance(val, torch.Tensor), "placeholder val must be a tensor or fake tensor"
             user_inputs.append(to_coreml_tensor_type(node.name, val))
+
         elif input_spec.kind == torch.export.graph_signature.InputKind.PARAMETER:
             lifted_parameters[input_spec.arg.name] = parameters[input_spec.target]
+
         elif input_spec.kind == torch.export.graph_signature.InputKind.BUFFER:
-            lifted_buffers[input_spec.arg.name] = buffers[input_spec.target]
+            # This is a workaround on mutable buffer: Core ML does not support stateful execution,
+            # so ExecuTorch will pass mutable buffers as inputs/outputs to Core ML delegation,
+            # then in-place copy Core ML outputs into buffers
+            # On Core ML side, we do not have to do anything special with outputs,
+            # but for inputs we will need to identify ExecuTorch lifted mutable buffers
+            # as Core ML user inputs
+            if input_spec.target in exported_program.graph_signature.buffers_to_mutate.values():
+                user_inputs.append(
+                    to_coreml_tensor_type(input_spec.arg.name, buffers[input_spec.target])
+                )
+            else:
+                lifted_buffers[input_spec.arg.name] = buffers[input_spec.target]
+
         elif input_spec.kind == torch.export.graph_signature.InputKind.CONSTANT_TENSOR:
             lifted_constants[input_spec.arg.name] = exported_program.constants[input_spec.target]
+
         else:
             raise NotImplementedError(
                 "Only 4 types of inputs handled yet: user input, parameter, buffer, constant. "

coremltools/converters/mil/frontend/torch/ops.py

@@ -1678,7 +1678,7 @@ def view(context, node):
     x = inputs[0]
     shape = inputs[1]
 
-    if np.prod(shape.shape) == 0:
+    if isinstance(shape, Var) and np.prod(shape.shape) == 0:
         # Reshape to empty shape (works only for scalar) is a no op
         assert (
             np.prod(x.shape) <= 1
@@ -6694,21 +6694,15 @@ def _get_causal_attn_mask(is_causal: bool, query_var: Var, key_var: Var) -> Var:
 
 def _cast_bool_attn_mask(attn_mask: Var, query_var: Var) -> Var:
     """
-    compute float mask as:
-    mask = cast(bool_mask) + (1-cast(bool_mask)) * -30k*ones(shape(bool_mask))
+    compute float mask as (1 - cast(bool_mask)) * -30k
     """
     assert is_bool(attn_mask.dtype)
 
-    shape = mb.shape(x=attn_mask)
-    negative_inf = mb.fill(
-        shape=shape, value=_np.array([-3e4]).astype(types.nptype_from_builtin(query_var.dtype))
-    )
     mask = mb.cast(x=attn_mask, dtype=types.builtin_to_string(query_var.dtype))
     compliment_of_mask = mb.sub(
         x=_np.array([1.0]).astype(types.nptype_from_builtin(mask.dtype)), y=mask
     )
-    compliment_of_mask = mb.mul(x=negative_inf, y=compliment_of_mask)
-    return mb.add(x=mask, y=compliment_of_mask)
+    return mb.mul(x=-3e4, y=compliment_of_mask)
 
 @register_torch_op
 def scaled_dot_product_attention(context, node):