GraphBuilder#
yobx.xbuilder.GraphBuilder simplifies the programmatic construction
and optimization of ONNX graphs. It is the primary tool used to convert a
torch.fx.Graph into a onnx.ModelProto, but it can equally
be used standalone to build or transform any ONNX graph from scratch.
Class Hierarchy#
GraphBuilder is composed of three
cooperative base classes:
_BuilderRuntime— evaluates small constant sub-expressions (e.g. the[0, 0, -1]passed to aReshapenode) so the builder can resolve-1to the correct symbolic formula and fold constants early._ShapeRuntime— handles value-as-shape tracking needed by operators such asShape,Gather,Concat, andSlicewhen their outputs feed directly into aReshape._InferenceRuntime— walks the graph node by node, dispatching each node to the matching per-operator handler inyobx.xshape.shape_type_computeso that shapes and types are tracked for every intermediate result.
Two helper classes round out the public API:
FunctionOptions— controls whether (and how) a sub-graph is exported as a reusable ONNX local function.OptimizationOptions— selects which optimization passes run insideto_onnx.
Building a graph from scratch#
The simplest workflow is:
Construct a
GraphBuilderwith an opset version.Call
make_tensor_inputto declare each graph input.Call
make_node(or the short-handg.op.<OpType>(…)syntax) to add operators.Call
make_tensor_outputto declare each graph output.Call
to_onnxto obtain aonnx.ModelProto.
<<<
import numpy as np
import onnx
from yobx.helpers.onnx_helper import pretty_onnx
from yobx.xbuilder import GraphBuilder
TFLOAT = onnx.TensorProto.FLOAT
# 1. create builder targeting opset 18
g = GraphBuilder(18, ir_version=10)
# 2. declare inputs
g.make_tensor_input("X", TFLOAT, ("batch", "seq", 64))
g.make_tensor_input("W", TFLOAT, (64, 32))
# 3. add a MatMul node via the short-hand op accessor
result = g.op.MatMul("X", "W")
# 4. declare the output and export
g.make_tensor_output(
result, elem_type=TFLOAT, shape=("batch", "seq", 32), indexed=False
)
model = g.to_onnx()
print(f"nodes : {len(model.graph.node)}")
print(f"opset : {model.opset_import[0].version}")
print(f"output : {model.graph.output[0].name}")
print(pretty_onnx(model))
>>>
nodes : 1
opset : 18
output : _onx_matmul_X
opset: domain='' version=18
input: name='X' type=dtype('float32') shape=['batch', 'seq', 64]
input: name='W' type=dtype('float32') shape=[64, 32]
MatMul(X, W) -> _onx_matmul_X
output: name='_onx_matmul_X' type=dtype('float32') shape=['batch', 'seq', 32]
Loading an existing model#
Passing an existing onnx.ModelProto to the constructor loads it into
the builder so its nodes and initializers can be inspected, modified, or
re-optimized.
<<<
import onnx
import onnx.helper as oh
from yobx.xbuilder import GraphBuilder
TFLOAT = onnx.TensorProto.FLOAT
model = oh.make_model(
oh.make_graph(
[
oh.make_node("Add", ["X", "Y"], ["T"]),
oh.make_node("Relu", ["T"], ["Z"]),
],
"add_relu",
[
oh.make_tensor_value_info("X", TFLOAT, ["batch", 4]),
oh.make_tensor_value_info("Y", TFLOAT, ["batch", 4]),
],
[oh.make_tensor_value_info("Z", TFLOAT, ["batch", 4])],
),
opset_imports=[oh.make_opsetid("", 18)],
ir_version=10,
)
g = GraphBuilder(model)
print("input shapes:", {n: g.get_shape(n) for n in g.input_names})
print("nodes :", [node.op_type for node in g.nodes])
>>>
input shapes: {'X': ('batch', 4), 'Y': ('batch', 4)}
nodes : ['Add', 'Relu']
Initializers#
Initializers (model weights and constants) are added with
make_initializer.
The builder deduplicates small integer arrays automatically: if the same
value is added twice it returns the name of the first occurrence rather than
creating a duplicate node.
<<<
import numpy as np
import onnx
from yobx.xbuilder import GraphBuilder
TFLOAT = onnx.TensorProto.FLOAT
g = GraphBuilder(18, ir_version=10)
g.make_tensor_input("X", TFLOAT, ("batch", 64))
# Add a weight matrix as an initializer
W = np.random.randn(64, 32).astype(np.float32)
w_name = g.make_initializer("W", W, source="example")
result = g.op.MatMul("X", w_name)
g.make_tensor_output(result, elem_type=TFLOAT, shape=("batch", 32), indexed=False)
model = g.to_onnx()
print("initializer name :", list(g.initializers_dict)[0])
print("initializer shape:", list(g.initializers_dict.values())[0].shape)
>>>
initializer name : W
initializer shape: (64, 32)
Shape and type tracking#
GraphBuilder inherits the full
ShapeBuilder interface. Shapes and types
are registered for every intermediate result as nodes are added, and are used
during optimization and for populating value_info in the exported proto.
See Expected API.
Dynamic shapes#
When some input dimensions are unknown at graph-construction time, they are
represented as strings (e.g. "batch", "seq"). For graphs that are
later exported for dynamic-shape inference with torch.export, the builder
accepts a dynamic_shapes dictionary that maps input names to per-axis
dimension objects (torch.export.Dim or WrapDim).
register_dynamic_objects_from_shape
registers any string dimension names encountered in a shape so that they are
tracked as symbolic dimensions.
<<<
import onnx
from yobx.xbuilder import GraphBuilder
TFLOAT = onnx.TensorProto.FLOAT
g = GraphBuilder(18, ir_version=10)
g.make_tensor_input("X", TFLOAT, ("batch", "seq", 64))
g.make_tensor_input("Y", TFLOAT, ("batch", "seq", 64))
# symbolic dimensions are tracked automatically once shapes are set
result = g.op.Add("X", "Y")
g.make_tensor_output(
result, elem_type=TFLOAT, shape=("batch", "seq", 64), indexed=False
)
model = g.to_onnx()
out = model.graph.output[0]
dims = [
d.dim_param if d.dim_param else d.dim_value for d in out.type.tensor_type.shape.dim
]
print("output shape:", dims)
>>>
output shape: ['batch', 'seq', 64]
Optimizations#
to_onnx runs a sequence of
optimization passes by default. The set of passes is controlled by
OptimizationOptions.
Default passes (in order):
Pass |
Effect |
|---|---|
|
Remove nodes whose outputs are never consumed. |
|
Evaluate operators such as |
|
Remove |
|
Merge identical constant initializers into a single tensor, removing redundant copies. |
|
Apply user-supplied or built-in fusion patterns (e.g.
|
|
Reorder nodes to reduce peak memory by moving each |
<<<
import onnx
import onnx.helper as oh
from yobx.xbuilder import GraphBuilder, OptimizationOptions
TFLOAT = onnx.TensorProto.FLOAT
model = oh.make_model(
oh.make_graph(
[
oh.make_node("Identity", ["X"], ["X2"]),
oh.make_node("Relu", ["X2"], ["Z"]),
],
"id_relu",
[oh.make_tensor_value_info("X", TFLOAT, [None, 4])],
[oh.make_tensor_value_info("Z", TFLOAT, [None, 4])],
),
opset_imports=[oh.make_opsetid("", 18)],
ir_version=10,
)
opts = OptimizationOptions(remove_identity=True)
g = GraphBuilder(model, optimization_options=opts)
optimized = g.to_onnx()
print("nodes before:", len(model.graph.node))
print("nodes after :", len(optimized.graph.node))
>>>
nodes before: 2
nodes after : 1
Optimization report#
Passing return_optimize_report=True to
to_onnx makes the method return
a (model, stats) tuple instead of just the model. stats is a list of
dictionaries — one entry per optimization pass — that records how many nodes
were added or removed and how long each pass took.
Key |
Description |
|---|---|
|
Name of the optimization pass (e.g. |
|
Number of nodes added by this pass. |
|
Number of nodes removed by this pass. |
|
Wall-clock time spent in this pass (seconds). |
|
Iteration number (only for pattern-based passes). |
|
Sequential index of the match within the iteration (pattern passes). |
|
Number of times the pattern was matched (pattern passes). |
The list can be converted to a pandas.DataFrame for quick
exploration:
<<<
import pandas
import onnx
import onnx.helper as oh
from yobx.xbuilder import GraphBuilder, OptimizationOptions
TFLOAT = onnx.TensorProto.FLOAT
model = oh.make_model(
oh.make_graph(
[
oh.make_node("Identity", ["X"], ["X2"]),
oh.make_node("Transpose", ["X2"], ["T"], perm=[1, 0]),
oh.make_node("Transpose", ["T"], ["Z"], perm=[1, 0]),
],
"demo",
[oh.make_tensor_value_info("X", TFLOAT, [3, 4])],
[oh.make_tensor_value_info("Z", TFLOAT, [3, 4])],
),
opset_imports=[oh.make_opsetid("", 18)],
ir_version=10,
)
opts = OptimizationOptions(patterns="default")
g = GraphBuilder(model, infer_shapes_options=True, optimization_options=opts)
optimized = g.to_onnx(return_optimize_report=True)
df = pandas.DataFrame(optimized.report.stats)
# keep only rows that have numeric added/removed counts
df["added"] = df["added"].fillna(0).astype(int)
df["removed"] = df["removed"].fillna(0).astype(int)
print(df[["pattern", "added", "removed", "time_in"]].to_string(index=False))
print(f"\nnodes before: {len(model.graph.node)}")
print(f"nodes after : {len(optimized.graph.node)}")
>>>
pattern added removed time_in
dynamic_dimension_naming 0 0 1.890000e-05
check_A-dynamic_dimension_naming 0 0 9.862997e-06
check_A-opt-sub 0 0 8.260002e-06
remove_identity 1 2 3.602000e-05
check_remove_identity-0 0 0 6.582995e-06
remove_unused 0 0 2.405200e-05
check_remove_unused-1 0 0 5.993999e-06
constant_folding 0 0 9.128998e-06
apply_constant_folding_new_inits 0 0 NaN
check_constant_folding-2 0 0 5.436996e-06
remove_unused 0 0 1.118999e-05
check_remove_unused-3 0 0 5.034999e-06
patterns 0 1 4.840417e-03
check_pattern_00 0 0 1.238000e-05
match_BatchNormalizationPattern 0 0 7.495997e-06
match_BatchNormalizationTrainingPattern 0 0 3.458998e-06
match_CastPattern 0 0 2.553999e-06
match_CastCastPattern 0 0 2.647997e-06
match_ConcatGatherPattern 0 0 3.437999e-06
match_ConcatReshapePattern 0 0 2.647001e-06
match_ConvBiasNullPattern 0 0 2.774002e-06
match_ExpandPattern 0 0 3.323999e-06
match_ExpandUnsqueezeExpandPattern 0 0 2.751003e-06
match_GeluPattern 0 0 1.216002e-06
match_IdentityPattern 0 0 2.130600e-05
match_LeakyReluPattern 0 0 8.992850e-04
match_MulUnsqueezeUnsqueezePattern 0 0 4.453999e-06
match_ReshapePattern 0 0 2.756002e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 3.736001e-06
match_ShapeBasedStaticExpandPattern 0 0 2.355002e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 2.105997e-06
match_ShapeBasedIdentityPattern 0 0 7.661001e-06
match_ShapedBasedReshapePattern 0 0 2.981003e-06
match_ShapeBasedSameChildrenPattern 0 0 2.798006e-06
match_ShapeBasedShapeShapeAddPattern 0 0 2.437999e-06
match_ReshapeReshapePattern 0 0 2.713001e-06
match_SameChildrenPattern 0 0 6.253002e-06
match_SameChildrenFromInputPattern 0 0 6.032002e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 1.683041e-03
match_SqueezeAddPattern 0 0 5.612004e-06
match_SqueezeBinaryUnsqueezePattern 0 0 3.086003e-06
match_SqueezeUnsqueezePattern 0 0 4.014997e-06
match_StaticConcatReshapePattern 0 0 2.662004e-06
match_SwapExpandReshapePattern 0 0 2.253997e-06
match_SwapExpandUnsqueezePattern 0 0 2.613000e-06
match_SwapUnaryPattern 0 0 2.263800e-05
match_SwapUnsqueezeTransposePattern 0 0 9.027004e-06
match_TransposeGatherPattern 0 0 2.531000e-06
match_TransposeReshapeTransposePattern 0 0 6.237999e-06
match_TransposeTransposePattern 0 0 3.554400e-05
match_UnsqueezeOrSqueezeReshapePattern 0 0 3.419002e-06
match_UnsqueezeReshapePattern 0 0 2.569999e-06
match_UnsqueezeUnsqueezePattern 0 0 2.485001e-06
match_FunctionAttentionPattern 0 0 3.074005e-06
match_FunctionAttentionGQAPattern 0 0 3.821995e-06
insert_and_remove_nodes 0 0 7.886600e-05
apply_TransposeTransposePattern 1 2 1.442550e-04
check_pattern_A10 0 0 1.166998e-06
check_pattern_A20 0 0 9.042000e-06
remove_duplicated_shape 0 0 2.813998e-06
check_pattern_BD0 0 0 4.947004e-06
remove_identity_nodes 0 0 1.667000e-05
check_pattern_BI0 0 0 4.522000e-06
remove_unused 0 0 1.237200e-05
check_pattern_BUS0 0 0 3.869005e-06
build_graph_for_pattern 0 0 9.308002e-06
iteration_0 0 0 3.088410e-03
match_BatchNormalizationPattern 0 0 4.262001e-06
match_BatchNormalizationTrainingPattern 0 0 2.480003e-06
match_CastPattern 0 0 2.162000e-06
match_CastCastPattern 0 0 1.663000e-06
match_ConcatGatherPattern 0 0 1.868997e-06
match_ConcatReshapePattern 0 0 1.591994e-06
match_ConvBiasNullPattern 0 0 1.640001e-06
match_ExpandPattern 0 0 1.820001e-06
match_ExpandUnsqueezeExpandPattern 0 0 1.500004e-06
match_GeluPattern 0 0 9.509968e-07
match_IdentityPattern 0 0 2.514003e-06
match_LeakyReluPattern 0 0 5.296999e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.834000e-06
match_ReshapePattern 0 0 1.526998e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 1.911998e-06
match_ShapeBasedStaticExpandPattern 0 0 1.262000e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.288994e-06
match_ShapeBasedIdentityPattern 0 0 1.786997e-06
match_ShapedBasedReshapePattern 0 0 1.709006e-06
match_ShapeBasedSameChildrenPattern 0 0 1.499000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.549000e-06
match_ReshapeReshapePattern 0 0 1.538996e-06
match_SameChildrenPattern 0 0 3.221001e-06
match_SameChildrenFromInputPattern 0 0 3.775996e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 4.121001e-06
match_SqueezeAddPattern 0 0 1.433000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.233006e-06
match_SqueezeUnsqueezePattern 0 0 1.394998e-06
match_StaticConcatReshapePattern 0 0 1.398999e-06
match_SwapExpandReshapePattern 0 0 1.041997e-06
match_SwapExpandUnsqueezePattern 0 0 1.207998e-06
match_SwapUnaryPattern 0 0 1.383996e-06
match_SwapUnsqueezeTransposePattern 0 0 1.243003e-06
match_TransposeGatherPattern 0 0 1.136999e-06
match_TransposeReshapeTransposePattern 0 0 1.034998e-06
match_TransposeTransposePattern 0 0 1.275002e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.398002e-06
match_UnsqueezeReshapePattern 0 0 1.657005e-06
match_UnsqueezeUnsqueezePattern 0 0 1.215994e-06
match_FunctionAttentionPattern 0 0 1.547000e-06
match_FunctionAttentionGQAPattern 0 0 2.212000e-06
check_pattern_A20 0 0 7.545001e-06
remove_duplicated_shape 0 0 1.757006e-06
check_pattern_BD0 0 0 4.670001e-06
remove_identity_nodes 0 0 1.542100e-05
check_pattern_BI0 0 0 4.091999e-06
remove_unused 0 0 1.111900e-05
check_pattern_BUS0 0 0 3.533998e-06
build_graph_for_pattern 0 0 8.092000e-06
iteration_1 0 0 1.982500e-04
match_BatchNormalizationPattern 0 0 2.020002e-06
match_BatchNormalizationTrainingPattern 0 0 1.498003e-06
match_CastLayerNormalizationCastPattern 0 0 3.168003e-06
match_CastPattern 0 0 1.453001e-06
match_CastCastBinaryPattern 0 0 4.873000e-06
match_CastCastPattern 0 0 1.403998e-06
match_CastOpCastPattern 0 0 3.886002e-06
match_ClipClipPattern 0 0 1.756001e-06
match_ConcatEmptyPattern 0 0 2.339002e-06
match_ConcatGatherPattern 0 0 1.269000e-06
match_ConcatReshapePattern 0 0 1.175002e-06
match_ConcatTwiceUnaryPattern 0 0 2.020999e-06
match_ConstantToInitializerPattern 0 0 2.018998e-06
match_ConvBiasNullPattern 0 0 1.137996e-06
match_DropoutPattern 0 0 1.992004e-06
match_ExpandPattern 0 0 1.225002e-06
match_ExpandBroadcastPattern 0 0 1.699002e-06
match_ExpandSwapPattern 0 0 1.616005e-06
match_ExpandUnsqueezeExpandPattern 0 0 1.142995e-06
match_GathersSplitPattern 0 0 1.813998e-06
match_GeluPattern 0 0 4.800022e-07
match_IdentityPattern 0 0 1.542001e-06
match_LayerNormalizationPattern 0 0 2.096000e-06
match_LayerNormalizationScalePattern 0 0 1.651002e-06
match_LeakyReluPattern 0 0 3.305002e-06
match_MaxReluPattern 0 0 1.870001e-06
match_MulMulMulScalarPattern 0 0 1.963002e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.062996e-06
match_NotNotPattern 0 0 2.027999e-06
match_NotWherePattern 0 0 2.200002e-06
match_ReduceArgTopKPattern 0 0 2.474000e-06
match_ReduceReshapePattern 0 0 2.004999e-06
match_ReduceSumNormalizePattern 0 0 2.133995e-06
match_ReshapePattern 0 0 1.138003e-06
match_ReshapeMatMulReshapePattern 0 0 2.052999e-06
match_Reshape2Of3Pattern 0 0 2.228000e-06
match_ReshapeReshapeBinaryPattern 0 0 1.777997e-06
match_GemmTransposePattern 0 0 2.089000e-06
match_MatMulReshape2Of3Pattern 0 0 2.351997e-06
match_MulMulMatMulPattern 0 0 2.205001e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 1.303000e-06
match_ShapeBasedStaticExpandPattern 0 0 1.129003e-06
match_ShapeBasedConcatExpandPattern 0 0 2.015004e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.255001e-06
match_ShapeBasedIdentityPattern 0 0 1.162000e-06
match_ShapeBasedExpandBroadcastPattern 0 0 1.902001e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 1.938999e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 2.007000e-06
match_ShapeBasedExpandSwapPattern 0 0 2.121997e-06
match_ShapeBasedMatMulToMulPattern 0 0 2.729998e-06
match_ShapedBasedReshapePattern 0 0 1.225999e-06
match_ShapeBasedSameChildrenPattern 0 0 1.196000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.107997e-06
match_ReshapeReshapePattern 0 0 1.127002e-06
match_RotaryEmbeddingPattern 0 0 1.771004e-06
match_SameChildrenPattern 0 0 3.137000e-06
match_SameChildrenFromInputPattern 0 0 3.262998e-06
match_SequenceConstructAtPattern 0 0 2.385001e-06
match_SliceSlicePattern 0 0 2.182998e-06
match_SlicesSplitPattern 0 0 1.699998e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 4.904003e-06
match_SplitConcatPattern 0 0 2.025001e-06
match_SqueezeAddPattern 0 0 1.132001e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.122004e-06
match_SqueezeUnsqueezePattern 0 0 1.383660e-04
match_StaticConcatReshapePattern 0 0 4.524998e-06
match_Sub1MulPattern 0 0 2.444001e-06
match_SwapExpandReshapePattern 0 0 1.478002e-06
match_SwapExpandUnsqueezePattern 0 0 1.105000e-06
match_SwapRangeAddScalarPattern 0 0 1.836001e-06
match_SwapUnaryPattern 0 0 1.984001e-06
match_SwapUnsqueezeTransposePattern 0 0 1.248001e-06
match_SwitchOrderBinaryPattern 0 0 2.400004e-06
match_SwitchReshapeActivationPattern 0 0 2.311994e-06
match_TransposeEqualReshapePattern 0 0 2.139001e-06
match_TransposeGatherPattern 0 0 1.261003e-06
match_TransposeMatMulPattern 0 0 1.842003e-06
match_TransposeReshapeMatMulPattern 0 0 1.940003e-06
match_TransposeReshapeTransposePattern 0 0 1.139997e-06
match_TransposeTransposePattern 0 0 1.150001e-06
match_UnsqueezeEqualPattern 0 0 1.841996e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.150001e-06
match_UnsqueezeReshapePattern 0 0 1.315995e-06
match_UnsqueezeUnsqueezePattern 0 0 1.127002e-06
match_WhereAddPattern 0 0 2.091001e-06
match_RotaryConcatPartPattern 0 0 2.312998e-06
match_FunctionAttentionPattern 0 0 1.603999e-06
match_FunctionAttentionGQAPattern 0 0 2.338005e-06
match_FunctionCausalMaskPattern 0 0 2.062996e-06
match_FunctionCausalMaskMulAddPattern 0 0 1.665998e-06
match_FunctionCosSinCachePattern 0 0 2.006003e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 1.551001e-06
match_RMSNormalizationPattern 0 0 2.012996e-06
match_RMSNormalizationMulPattern 0 0 1.597997e-06
check_pattern_A20 0 0 1.196100e-05
remove_duplicated_shape 0 0 2.232999e-06
check_pattern_BD0 0 0 5.261005e-06
remove_identity_nodes 0 0 1.787600e-05
check_pattern_BI0 0 0 4.740003e-06
remove_unused 0 0 1.304400e-05
check_pattern_BUS0 0 0 4.280999e-06
build_graph_for_pattern 0 0 9.418000e-06
iteration_2 0 0 4.962030e-04
match_BatchNormalizationPattern 0 0 2.195004e-06
match_BatchNormalizationTrainingPattern 0 0 1.362001e-06
match_CastLayerNormalizationCastPattern 0 0 1.911998e-06
match_CastPattern 0 0 1.345004e-06
match_CastCastBinaryPattern 0 0 1.480999e-06
match_CastCastPattern 0 0 1.105000e-06
match_CastOpCastPattern 0 0 1.723005e-06
match_ClipClipPattern 0 0 1.266002e-06
match_ConcatEmptyPattern 0 0 1.838001e-06
match_ConcatGatherPattern 0 0 1.282002e-06
match_ConcatReshapePattern 0 0 1.303000e-06
match_ConcatTwiceUnaryPattern 0 0 1.449000e-06
match_ConstantToInitializerPattern 0 0 1.416003e-06
match_ConvBiasNullPattern 0 0 1.305001e-06
match_DropoutPattern 0 0 1.357002e-06
match_ExpandPattern 0 0 1.451001e-06
match_ExpandBroadcastPattern 0 0 1.560002e-06
match_ExpandSwapPattern 0 0 1.605003e-06
match_ExpandUnsqueezeExpandPattern 0 0 1.346001e-06
match_GathersSplitPattern 0 0 1.421002e-06
match_GeluPattern 0 0 4.639951e-07
match_IdentityPattern 0 0 1.504995e-06
match_LayerNormalizationPattern 0 0 1.533001e-06
match_LayerNormalizationScalePattern 0 0 1.410997e-06
match_LeakyReluPattern 0 0 4.255002e-06
match_MaxReluPattern 0 0 1.629000e-06
match_MulMulMulScalarPattern 0 0 1.501998e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.244000e-06
match_NotNotPattern 0 0 1.360997e-06
match_NotWherePattern 0 0 1.644999e-06
match_ReduceArgTopKPattern 0 0 2.305002e-06
match_ReduceReshapePattern 0 0 1.767999e-06
match_ReduceSumNormalizePattern 0 0 1.470995e-06
match_ReshapePattern 0 0 1.102999e-06
match_ReshapeMatMulReshapePattern 0 0 1.209999e-06
match_Reshape2Of3Pattern 0 0 1.330998e-06
match_ReshapeReshapeBinaryPattern 0 0 1.369997e-06
match_GemmTransposePattern 0 0 1.499000e-06
match_MatMulReshape2Of3Pattern 0 0 1.413006e-06
match_MulMulMatMulPattern 0 0 1.963999e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 1.357002e-06
match_ShapeBasedStaticExpandPattern 0 0 1.177999e-06
match_ShapeBasedConcatExpandPattern 0 0 1.452005e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.087996e-06
match_ShapeBasedIdentityPattern 0 0 1.361004e-06
match_ShapeBasedExpandBroadcastPattern 0 0 1.289001e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 1.264001e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 1.171997e-06
match_ShapeBasedExpandSwapPattern 0 0 1.437002e-06
match_ShapeBasedMatMulToMulPattern 0 0 1.355002e-06
match_ShapedBasedReshapePattern 0 0 1.261003e-06
match_ShapeBasedSameChildrenPattern 0 0 1.963002e-06
match_ShapeBasedShapeShapeAddPattern 0 0 2.151995e-06
match_ReshapeReshapePattern 0 0 1.706998e-06
match_RotaryEmbeddingPattern 0 0 2.157001e-06
match_SameChildrenPattern 0 0 4.266003e-06
match_SameChildrenFromInputPattern 0 0 4.353002e-06
match_SequenceConstructAtPattern 0 0 2.448003e-06
match_SliceSlicePattern 0 0 2.508001e-06
match_SlicesSplitPattern 0 0 2.143999e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 5.829002e-06
match_SplitConcatPattern 0 0 2.244997e-06
match_SqueezeAddPattern 0 0 1.868000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.959997e-06
match_SqueezeUnsqueezePattern 0 0 2.054003e-06
match_StaticConcatReshapePattern 0 0 2.187997e-06
match_Sub1MulPattern 0 0 1.948996e-06
match_SwapExpandReshapePattern 0 0 1.654997e-06
match_SwapExpandUnsqueezePattern 0 0 1.729000e-06
match_SwapRangeAddScalarPattern 0 0 1.966997e-06
match_SwapUnaryPattern 0 0 1.927001e-06
match_SwapUnsqueezeTransposePattern 0 0 1.913999e-06
match_SwitchOrderBinaryPattern 0 0 2.351997e-06
match_SwitchReshapeActivationPattern 0 0 2.362001e-06
match_TransposeEqualReshapePattern 0 0 2.198998e-06
match_TransposeGatherPattern 0 0 1.969005e-06
match_TransposeMatMulPattern 0 0 2.811001e-06
match_TransposeReshapeMatMulPattern 0 0 1.973000e-06
match_TransposeReshapeTransposePattern 0 0 1.679997e-06
match_TransposeTransposePattern 0 0 1.838998e-06
match_UnsqueezeEqualPattern 0 0 2.085995e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.740002e-06
match_UnsqueezeReshapePattern 0 0 1.909997e-06
match_UnsqueezeUnsqueezePattern 0 0 1.615997e-06
match_WhereAddPattern 0 0 2.049004e-06
match_RotaryConcatPartPattern 0 0 2.721004e-06
match_FunctionAttentionPattern 0 0 2.408000e-06
match_FunctionAttentionGQAPattern 0 0 3.134999e-06
match_FunctionCausalMaskPattern 0 0 2.385998e-06
match_FunctionCausalMaskMulAddPattern 0 0 2.271998e-06
match_FunctionCosSinCachePattern 0 0 2.624998e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 2.256005e-06
match_RMSNormalizationPattern 0 0 2.499000e-06
match_RMSNormalizationMulPattern 0 0 2.234003e-06
match_AttentionGQAPattern 0 0 2.565997e-06
check_pattern_A20 0 0 1.466600e-05
remove_duplicated_shape 0 0 7.069997e-06
check_pattern_BD0 0 0 9.892996e-06
remove_identity_nodes 0 0 2.449800e-05
check_pattern_BI0 0 0 6.333998e-06
remove_unused 0 0 1.708901e-05
check_pattern_BUS0 0 0 6.683003e-06
build_graph_for_pattern 0 0 1.459800e-05
iteration_3 0 0 4.115880e-04
match_BatchNormalizationPattern 0 0 4.227993e-06
match_BatchNormalizationTrainingPattern 0 0 2.293003e-06
match_CastLayerNormalizationCastPattern 0 0 3.285997e-06
match_CastPattern 0 0 2.107998e-06
match_CastCastBinaryPattern 0 0 2.563000e-06
match_CastCastPattern 0 0 2.030996e-06
match_CastOpCastPattern 0 0 2.921995e-06
match_ClipClipPattern 0 0 2.267006e-06
match_ConcatEmptyPattern 0 0 2.358996e-06
match_ConcatGatherPattern 0 0 2.097004e-06
match_ConcatReshapePattern 0 0 2.307002e-06
match_ConcatTwiceUnaryPattern 0 0 2.278000e-06
match_ConstantToInitializerPattern 0 0 2.532004e-06
match_ConvBiasNullPattern 0 0 2.417997e-06
match_DropoutPattern 0 0 2.207998e-06
match_ExpandPattern 0 0 2.230001e-06
match_ExpandBroadcastPattern 0 0 2.266002e-06
match_ExpandSwapPattern 0 0 2.480003e-06
match_ExpandUnsqueezeExpandPattern 0 0 2.100001e-06
match_GathersSplitPattern 0 0 2.626999e-06
match_GeluPattern 0 0 7.730050e-07
match_IdentityPattern 0 0 2.460001e-06
match_LayerNormalizationPattern 0 0 2.311004e-06
match_LayerNormalizationScalePattern 0 0 2.223998e-06
match_LeakyReluPattern 0 0 7.292001e-06
match_MaxReluPattern 0 0 2.499000e-06
match_MulMulMulScalarPattern 0 0 2.350003e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.963999e-06
match_NotNotPattern 0 0 6.246002e-06
match_NotWherePattern 0 0 2.319997e-06
match_ReduceArgTopKPattern 0 0 2.910005e-06
match_ReduceReshapePattern 0 0 2.945999e-06
match_ReduceSumNormalizePattern 0 0 2.233995e-06
match_ReshapePattern 0 0 2.371002e-06
match_ReshapeMatMulReshapePattern 0 0 2.096007e-06
match_Reshape2Of3Pattern 0 0 3.134999e-06
match_ReshapeReshapeBinaryPattern 0 0 2.300003e-06
match_MatMulAddPattern 0 0 3.200999e-06
match_GemmTransposePattern 0 0 2.289999e-06
match_MatMulReshape2Of3Pattern 0 0 2.722998e-06
match_MulMulMatMulPattern 0 0 2.962006e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 2.553999e-06
match_ShapeBasedStaticExpandPattern 0 0 2.216999e-06
match_ShapeBasedConcatExpandPattern 0 0 2.686997e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.951004e-06
match_ShapeBasedIdentityPattern 0 0 2.442997e-06
match_ShapeBasedExpandBroadcastPattern 0 0 2.385998e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 1.941000e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 2.278000e-06
match_ShapeBasedExpandSwapPattern 0 0 2.624001e-06
match_ShapeBasedMatMulToMulPattern 0 0 2.209999e-06
match_ShapedBasedReshapePattern 0 0 2.298999e-06
match_ShapeBasedSameChildrenPattern 0 0 2.233995e-06
match_ShapeBasedShapeShapeAddPattern 0 0 2.398003e-06
match_ReshapeReshapePattern 0 0 2.345994e-06
match_RotaryEmbeddingPattern 0 0 2.482004e-06
match_SameChildrenPattern 0 0 5.064998e-06
match_SameChildrenFromInputPattern 0 0 5.234004e-06
match_SequenceConstructAtPattern 0 0 2.662004e-06
match_SliceSlicePattern 0 0 2.114000e-06
match_SlicesSplitPattern 0 0 2.637003e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 6.703995e-06
match_SplitConcatPattern 0 0 2.804998e-06
match_SqueezeAddPattern 0 0 2.272005e-06
match_SqueezeBinaryUnsqueezePattern 0 0 2.198001e-06
match_SqueezeUnsqueezePattern 0 0 3.039000e-06
match_StaticConcatReshapePattern 0 0 1.780005e-06
match_Sub1MulPattern 0 0 1.943001e-06
match_SwapExpandReshapePattern 0 0 2.177003e-06
match_SwapExpandUnsqueezePattern 0 0 2.515997e-06
match_SwapRangeAddScalarPattern 0 0 2.355999e-06
match_SwapUnaryPattern 0 0 2.765002e-06
match_SwapUnsqueezeTransposePattern 0 0 2.646004e-06
match_SwitchOrderBinaryPattern 0 0 2.569999e-06
match_SwitchReshapeActivationPattern 0 0 2.569002e-06
match_TransposeEqualReshapePattern 0 0 2.638000e-06
match_TransposeGatherPattern 0 0 2.036999e-06
match_TransposeMatMulPattern 0 0 2.654997e-06
match_TransposeReshapeMatMulPattern 0 0 2.161003e-06
match_TransposeReshapeTransposePattern 0 0 2.110995e-06
match_TransposeTransposePattern 0 0 2.075998e-06
match_UnsqueezeEqualPattern 0 0 2.061999e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 2.472007e-06
match_UnsqueezeReshapePattern 0 0 1.930006e-06
match_UnsqueezeUnsqueezePattern 0 0 2.290995e-06
match_WhereAddPattern 0 0 2.091001e-06
match_RotaryConcatPartPattern 0 0 3.069996e-06
match_FunctionAttentionPattern 0 0 2.484005e-06
match_FunctionAttentionGQAPattern 0 0 3.699999e-06
match_FunctionCausalMaskPattern 0 0 2.799003e-06
match_FunctionCausalMaskMulAddPattern 0 0 2.026005e-06
match_FunctionCosSinCachePattern 0 0 2.492001e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 2.474000e-06
match_RMSNormalizationPattern 0 0 2.243003e-06
match_RMSNormalizationMulPattern 0 0 2.403001e-06
match_AttentionGQAPattern 0 0 2.006003e-06
check_pattern_A20 0 0 1.664200e-05
remove_duplicated_shape 0 0 3.464003e-06
check_pattern_BD0 0 0 7.737006e-06
remove_identity_nodes 0 0 2.757100e-05
check_pattern_BI0 0 0 7.579001e-06
remove_unused 0 0 1.861100e-05
check_pattern_BUS0 0 0 7.021998e-06
build_graph_for_pattern 0 0 1.399700e-05
check_patterns-4 0 0 1.192100e-05
remove_unused 0 0 1.737000e-05
check_remove_unused-5 0 0 7.354000e-06
remove_identity 0 0 1.912900e-05
check_remove_identity-6 0 0 6.077003e-06
constant_folding 0 0 1.197300e-05
apply_constant_folding_new_inits 0 0 NaN
check_constant_folding-7 0 0 5.965005e-06
remove_unused 0 0 1.261799e-05
check_remove_unused-8 0 0 5.820999e-06
remove_duplicated_initializer 0 0 2.285000e-06
check_remove_duplicated_initializer-9 0 0 5.987000e-06
remove_identity 0 0 1.674300e-05
check_remove_identity-10 0 0 5.715003e-06
remove_unused 0 0 1.257700e-05
check_remove_unused-11 0 0 5.808004e-06
order 0 0 5.281100e-05
check_orderA 0 0 9.283001e-06
check_orderL 0 0 6.076996e-06
shape_order 0 0 2.519200e-05
order 0 0 NaN
check_order-12 0 0 6.309005e-06
optimization 0 2 5.244711e-03
nodes before: 3
nodes after : 1
The report can be aggregated by pass name:
<<<
import pandas
import onnx
import onnx.helper as oh
from yobx.xbuilder import GraphBuilder, OptimizationOptions
TFLOAT = onnx.TensorProto.FLOAT
model = oh.make_model(
oh.make_graph(
[
oh.make_node("Identity", ["X"], ["X2"]),
oh.make_node("Transpose", ["X2"], ["T"], perm=[1, 0]),
oh.make_node("Transpose", ["T"], ["Z"], perm=[1, 0]),
],
"demo",
[oh.make_tensor_value_info("X", TFLOAT, [3, 4])],
[oh.make_tensor_value_info("Z", TFLOAT, [3, 4])],
),
opset_imports=[oh.make_opsetid("", 18)],
ir_version=10,
)
opts = OptimizationOptions(patterns="default")
g = GraphBuilder(model, infer_shapes_options=True, optimization_options=opts)
art = g.to_onnx(return_optimize_report=True)
df = pandas.DataFrame(art.report.stats)
for c in ["added", "removed"]:
df[c] = df[c].fillna(0).astype(int)
agg = df.groupby("pattern")[["added", "removed", "time_in"]].sum()
agg = agg[(agg["added"] > 0) | (agg["removed"] > 0)].sort_values(
"removed", ascending=False
)
print(agg.to_string())
>>>
added removed time_in
pattern
apply_TransposeTransposePattern 1 2 0.000339
optimization 0 2 0.014814
remove_identity 1 2 0.000212
patterns 0 1 0.013827
Local functions#
A sub-graph can be exported as a reusable ONNX local function (a
FunctionProto) by passing a FunctionOptions instance to
to_onnx.
<<<
import onnx
from yobx.xbuilder import GraphBuilder, FunctionOptions
TFLOAT = onnx.TensorProto.FLOAT
g = GraphBuilder(18, ir_version=10, as_function=True)
g.make_tensor_input("X", TFLOAT, ("batch", 64))
r = g.op.Relu("X")
g.make_tensor_output(r, indexed=False)
func = g.to_onnx(
function_options=FunctionOptions(
export_as_function=True,
name="MyRelu",
domain="my.domain",
),
inline=False,
)
print(type(func).__name__)
print("function name :", func.name)
print("function domain:", func.domain)
>>>
FunctionProto
function name : MyRelu
function domain: my.domain
Debugging#
GraphBuilder respects several
environment variables that help narrow down construction or optimization
problems:
Environment variable |
Effect |
|---|---|
|
Raises an exception the moment result |
|
Raises an exception the moment result |
|
Raises an exception the moment result |
|
Raises an exception the moment a node produces output |
|
Prints extra information for shape-as-value tracking (e.g. inputs
to |
|
Prints which constant is being evaluated. |
|
Prints details when nodes from a local function domain are added. |
|
Raises an exception when a shape is missing for a result that should have one. |
|
Raises an exception as soon as a null/empty shape is encountered. |
|
Prints a message every time dynamic dimension |
|
Prints a message every time a node producing |
In addition,
get_debug_msg
returns a detailed text dump of the builder’s internal state (known shapes,
types, ranks, constants, and node list) which can be printed or logged whenever
an assertion fails.
pretty_text returns a
human-readable representation of the whole graph (inputs, initializers, nodes,
outputs) and is useful for quick visual inspection:
<<<
import onnx
import onnx.helper as oh
from yobx.xbuilder import GraphBuilder
TFLOAT = onnx.TensorProto.FLOAT
model = oh.make_model(
oh.make_graph(
[
oh.make_node("Add", ["X", "Y"], ["T"]),
oh.make_node("Relu", ["T"], ["Z"]),
],
"add_relu",
[
oh.make_tensor_value_info("X", TFLOAT, ["batch", 4]),
oh.make_tensor_value_info("Y", TFLOAT, ["batch", 4]),
],
[oh.make_tensor_value_info("Z", TFLOAT, ["batch", 4])],
),
opset_imports=[oh.make_opsetid("", 18)],
ir_version=10,
)
g = GraphBuilder(model)
print(g.pretty_text())
>>>
dyn---: batch -> WrapSym(batch)
dynrev: batch -> [('batch', SymInt(batch))]
dynsrc: batch -> [{batch:('input_name', 'X'), batch:('axis', 0)}, {batch:('input_name', 'Y'), batch:('axis', 0)}, {batch:('input_name', 'Z'), batch:('axis', 0)}]
opset: : 18
input:: X |T1: batch x 4
input:: Y |T1: batch x 4
Add: X, Y -> T |T1: batch x 4
Relu: T -> Z |T1: batch x 4
output:: Z |T1: batch x 4