mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-07-20 19:12:24 +00:00
move all experimental ops to version 1 of ms domain (#5287)
* move all experimental ops to version 1 of ms domain * deprecate TransposeMatMul in favor of FusedMatMul * update documentation
This commit is contained in:
Ashwini Khade
GitHub
parent
2c32309e2c
commit
3f00b8db8f
5 changed files with 663 additions and 415 deletions
File diff suppressed because it is too large
Load diff
|
|
@ -412,7 +412,6 @@ and present state are optional. Present state could appear in output even when p
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(Attention)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(Attention_ver1_doc)
|
||||
.Attr("num_heads", "Number of attention heads", AttributeProto::INT)
|
||||
.Attr("unidirectional",
|
||||
|
|
@ -470,7 +469,6 @@ and present state are optional. Present state could appear in output even when p
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(QAttention)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc("Quantization of Multi-Head Self Attention.")
|
||||
.Attr("num_heads", "Number of attention heads", AttributeProto::INT)
|
||||
.Attr("unidirectional",
|
||||
|
|
@ -566,7 +564,6 @@ will be calculated.)DOC";
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(EmbedLayerNormalization)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(EmbedLayerNormalization_ver1_doc)
|
||||
.Attr("epsilon", "The epsilon value to use to avoid division by zero.", AttributeProto::FLOAT, kDefaultEmbedLayerNormEpsilon)
|
||||
.Input(0, "input_ids", "2D words IDs with shape (batch_size, sequence_length)", "T1")
|
||||
|
|
@ -628,7 +625,6 @@ GELU (Gaussian Error Linear Unit) approximation: Y=0.5*X*(1+tanh(0.797885*X+0.03
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(FastGelu)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(FastGelu_ver1_doc)
|
||||
.Input(0, "X", "input tensor", "T")
|
||||
.Input(1, "bias", "bias tensor", "T", OpSchema::Optional)
|
||||
|
|
@ -639,7 +635,6 @@ GELU (Gaussian Error Linear Unit) approximation: Y=0.5*X*(1+tanh(0.797885*X+0.03
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(SkipLayerNormalization)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc("Skip and Layer Normalization Fusion")
|
||||
.Attr("epsilon", "The epsilon value to use to avoid division by zero.", AttributeProto::FLOAT, kDefaultSkipLayerNormEpsilon)
|
||||
.Input(0, "input", "3D input tensor with shape (batch_size, sequence_length, hidden_size)", "T")
|
||||
|
|
@ -1922,6 +1917,37 @@ Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-
|
|||
FusedMatMulShapeInference(ctx);
|
||||
});
|
||||
|
||||
ONNX_CONTRIB_OPERATOR_SCHEMA(TransposeMatMul)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(2)
|
||||
.Deprecate()
|
||||
.Input(0, "A", "N-dimensional matrix A", "T")
|
||||
.Input(1, "B", "N-dimensional matrix B", "T")
|
||||
.Attr(
|
||||
"alpha",
|
||||
"Scalar multiplier for the product of the input tensors.",
|
||||
AttributeProto::FLOAT,
|
||||
1.0f)
|
||||
.Attr(
|
||||
"transA",
|
||||
"Whether A should be transposed on the last two dimensions before doing multiplication",
|
||||
AttributeProto::INT,
|
||||
static_cast<int64_t>(0))
|
||||
.Attr(
|
||||
"transB",
|
||||
"Whether B should be transposed on the last two dimensions before doing multiplication",
|
||||
AttributeProto::INT,
|
||||
static_cast<int64_t>(0))
|
||||
.Output(0, "Y", "Matrix multiply results", "T")
|
||||
.TypeConstraint(
|
||||
"T",
|
||||
{"tensor(float16)", "tensor(float)", "tensor(double)", "tensor(bfloat16)"},
|
||||
"Constrain input and output types to float tensors.")
|
||||
.SetDoc(TransposeMatMul_doc)
|
||||
.TypeAndShapeInferenceFunction([](ONNX_NAMESPACE::InferenceContext& ctx) {
|
||||
FusedMatMulShapeInference(ctx);
|
||||
});
|
||||
|
||||
ONNX_CONTRIB_OPERATOR_SCHEMA(FusedMatMul)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
|
|
@ -2843,7 +2869,6 @@ inputs by their magnitude, rather than gates inputs by their sign as in ReLUs.)D
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(Gelu)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(Gelu_ver1_doc)
|
||||
.Input(0, "X", "The input data as Tensor.", "T")
|
||||
.Output(0, "Y", "The output.", "T")
|
||||
|
|
@ -2859,7 +2884,6 @@ It's an extension of Gelu. It takes the sum of input A and bias input B as the i
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(BiasGelu)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(BiasGelu_ver1_doc)
|
||||
.Input(0, "A", "The normal input data.", "T")
|
||||
.Input(1, "B", "The bias input data that is a 1D tensor.", "T")
|
||||
|
|
@ -2885,7 +2909,6 @@ It's an extension of Gelu. It takes the sum of input A and bias input B as the i
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(Inverse)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.Input(0, "X", "Input tensor. Every matrix in the batch must be invertible.", "T")
|
||||
.Output(0, "Y", "Output tensor of the same type and shape as the input tensor.", "T")
|
||||
.TypeConstraint(
|
||||
|
|
@ -2943,7 +2966,6 @@ It's an extension of Gelu. It takes the sum of input A and bias input B as the i
|
|||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetDoc(Trilu_ver1_doc)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.Attr("upper",
|
||||
"Boolean. Indicates whether upper or lower part of matrix is retained. Default is true.",
|
||||
AttributeProto::INT,
|
||||
|
|
@ -2999,7 +3021,6 @@ It's an extension of Gelu. It takes the sum of input A and bias input B as the i
|
|||
ONNX_CONTRIB_OPERATOR_SCHEMA(BiasSoftmax)
|
||||
.SetDomain(kMSDomain)
|
||||
.SinceVersion(1)
|
||||
.SetSupportLevel(OpSchema::SupportType::EXPERIMENTAL)
|
||||
.SetDoc(
|
||||
"Y = softmax(scores + bias)) with simple broadcast on bias. "
|
||||
"Intended to specialize softmax(scores + additive_mask) commonly found in transformer models.")
|
||||
|
|
|
|||
|
|
@ -92,7 +92,7 @@ Status Environment::Initialize(std::unique_ptr<logging::LoggingManager> logging_
|
|||
#if !defined(ORT_MINIMAL_BUILD)
|
||||
// Register Microsoft domain with min/max op_set version as 1/1.
|
||||
std::call_once(schemaRegistrationOnceFlag, []() {
|
||||
ONNX_NAMESPACE::OpSchemaRegistry::DomainToVersionRange::Instance().AddDomainToVersion(onnxruntime::kMSDomain, 1, 1);
|
||||
ONNX_NAMESPACE::OpSchemaRegistry::DomainToVersionRange::Instance().AddDomainToVersion(onnxruntime::kMSDomain, 1, 2);
|
||||
ONNX_NAMESPACE::OpSchemaRegistry::DomainToVersionRange::Instance().AddDomainToVersion(onnxruntime::kMSNchwcDomain, 1, 1);
|
||||
ONNX_NAMESPACE::OpSchemaRegistry::DomainToVersionRange::Instance().AddDomainToVersion(onnxruntime::kMSFeaturizersDomain, 1, 1);
|
||||
#ifdef USE_DML
|
||||
|
|
|
|||
|
|
@ -218,6 +218,24 @@ TEST(TransposeMatMulOpTest, FloatTypeScale) {
|
|||
RunFusedMatMulTest<float>("TransposeMatMul", 1, true, true, 4.0f, true);
|
||||
}
|
||||
|
||||
TEST(TransposeMatMulOpTest, DeprecatedSchema) {
|
||||
OpTester test("TransposeMatMul", 2, onnxruntime::kMSDomain);
|
||||
|
||||
std::vector<int64_t> input0_dims{3, 4};
|
||||
std::vector<float> input0_vals{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
|
||||
|
||||
std::vector<int64_t> input1_dims{4, 3};
|
||||
std::vector<float> input1_vals{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11};
|
||||
|
||||
test.AddInput<float>("A", input0_dims, input0_vals);
|
||||
test.AddInput<float>("B", input1_dims, input1_vals);
|
||||
test.AddOutput<float>("Y", {3, 3}, {42, 48, 54, 114, 136, 158, 186, 224, 262});
|
||||
|
||||
// TransposeMatMul is deprecated from version 2 onwards of ms domain.
|
||||
test.Run(OpTester::ExpectResult::kExpectFailure);
|
||||
}
|
||||
|
||||
|
||||
} // namespace transpose_matmul
|
||||
} // namespace test
|
||||
} // namespace onnxruntime
|
||||
|
|
|
|||
|
|
@ -50,6 +50,10 @@ def format_name_with_domain(domain, schema_name): # type: (Text, Text) -> Text
|
|||
return schema_name
|
||||
|
||||
|
||||
def format_name_with_version(schema_name, version): # type: (Text, Text) -> Text
|
||||
return '{}-{}'.format(schema_name, version)
|
||||
|
||||
|
||||
def display_attr_type(v): # type: (OpSchema.AttrType) -> Text
|
||||
assert isinstance(v, OpSchema.AttrType)
|
||||
s = Text(v)
|
||||
|
|
@ -132,8 +136,9 @@ def display_schema(schema, versions): # type: (OpSchema, Sequence[OpSchema]) ->
|
|||
if len(versions) > 1:
|
||||
# TODO: link to the Changelog.md
|
||||
s += '\nOther versions of this operator: {}\n'.format(
|
||||
', '.join(display_version_link(format_name_with_domain(v.domain, v.name),
|
||||
v.since_version) for v in versions[:-1]))
|
||||
', '.join(format_name_with_version(
|
||||
format_name_with_domain(v.domain, v.name), v.since_version)
|
||||
for v in versions[:-1]))
|
||||
|
||||
# If this schema is deprecated, don't display any of the following sections
|
||||
if schema.deprecated:
|
||||
|
|
@ -310,7 +315,7 @@ def main(args): # type: (Type[Args]) -> None
|
|||
fout.write(
|
||||
"*This file is automatically generated from the\n"
|
||||
" [def files](/onnxruntime/core/graph/contrib_ops/contrib_defs.cc) via "
|
||||
"[this script](/tools/python/gen_doc.py).\n"
|
||||
"[this script](/tools/python/gen_contrib_doc.py).\n"
|
||||
" Do not modify directly and instead edit operator definitions.*\n")
|
||||
|
||||
# domain -> support level -> name -> [schema]
|
||||
|
|
|
|||
Loading…
Reference in a new issue