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 2.412600e-05
check_A-dynamic_dimension_naming 0 0 1.023900e-05
check_A-opt-sub 0 0 8.780000e-06
remove_identity 1 2 5.059400e-05
check_remove_identity-0 0 0 6.589001e-06
remove_unused 0 0 1.749100e-05
check_remove_unused-1 0 0 5.660000e-06
constant_folding 0 0 1.032400e-05
apply_constant_folding_new_inits 0 0 NaN
check_constant_folding-2 0 0 5.735000e-06
remove_unused 0 0 1.148300e-05
check_remove_unused-3 0 0 5.370000e-06
patterns 0 1 6.222457e-03
check_pattern_00 0 0 1.264300e-05
match_BatchNormalizationPattern 0 0 1.523200e-05
match_BatchNormalizationTrainingPattern 0 0 5.405999e-06
match_CastPattern 0 0 5.589000e-06
match_CastCastPattern 0 0 4.624000e-06
match_ConcatGatherPattern 0 0 4.457000e-06
match_ConcatReshapePattern 0 0 5.620000e-06
match_ConvBiasNullPattern 0 0 4.347000e-06
match_PadConvPattern 0 0 3.985000e-06
match_ExpandPattern 0 0 3.831000e-06
match_ExpandUnsqueezeExpandPattern 0 0 4.000000e-06
match_GeluPattern 0 0 1.978000e-06
match_IdentityPattern 0 0 3.631900e-05
match_LeakyReluPattern 0 0 1.542497e-03
match_MulUnsqueezeUnsqueezePattern 0 0 1.072500e-05
match_ReshapePattern 0 0 7.439000e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 5.073000e-06
match_ShapeBasedStaticExpandPattern 0 0 4.237000e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 6.577000e-06
match_ShapeBasedIdentityPattern 0 0 2.028500e-05
match_ShapedBasedReshapePattern 0 0 5.422000e-06
match_ShapeBasedSameChildrenPattern 0 0 4.451999e-06
match_ShapeBasedShapeShapeAddPattern 0 0 4.070000e-06
match_ReshapeReshapePattern 0 0 4.938000e-06
match_SameChildrenPattern 0 0 1.264900e-05
match_SameChildrenFromInputPattern 0 0 1.080400e-05
match_SoftmaxCrossEntropyLossCastPattern 0 0 2.235091e-03
match_SqueezeAddPattern 0 0 7.651000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 3.587001e-06
match_SqueezeUnsqueezePattern 0 0 4.801000e-06
match_StaticConcatReshapePattern 0 0 5.470000e-06
match_SwapExpandReshapePattern 0 0 3.797000e-06
match_SwapExpandUnsqueezePattern 0 0 3.009000e-06
match_SwapUnaryPattern 0 0 2.944500e-05
match_SwapUnsqueezeTransposePattern 0 0 1.260000e-05
match_TransposeGatherPattern 0 0 4.281001e-06
match_TransposeReshapeTransposePattern 0 0 8.899000e-06
match_TransposeTransposePattern 0 0 4.959000e-05
match_UnsqueezeOrSqueezeReshapePattern 0 0 5.414000e-06
match_UnsqueezeReshapePattern 0 0 3.812000e-06
match_UnsqueezeUnsqueezePattern 0 0 4.054000e-06
match_FunctionAttentionPattern 0 0 4.400999e-06
match_FunctionAttentionGQAPattern 0 0 4.134999e-06
insert_and_remove_nodes 0 0 1.016260e-04
apply_TransposeTransposePattern 1 2 1.846580e-04
check_pattern_A10 0 0 1.381000e-06
check_pattern_A20 0 0 1.074400e-05
remove_duplicated_shape 0 0 2.977000e-06
check_pattern_BD0 0 0 5.379000e-06
remove_identity_nodes 0 0 1.694400e-05
check_pattern_BI0 0 0 4.599000e-06
remove_unused 0 0 1.141600e-05
check_pattern_BUS0 0 0 3.901000e-06
build_graph_for_pattern 0 0 9.595000e-06
iteration_0 0 0 4.528918e-03
match_BatchNormalizationPattern 0 0 5.000999e-06
match_BatchNormalizationTrainingPattern 0 0 3.062000e-06
match_CastPattern 0 0 2.458000e-06
match_CastCastPattern 0 0 2.199000e-06
match_ConcatGatherPattern 0 0 2.639000e-06
match_ConcatReshapePattern 0 0 3.368999e-06
match_ConvBiasNullPattern 0 0 2.835000e-06
match_PadConvPattern 0 0 2.363000e-06
match_ExpandPattern 0 0 2.740000e-06
match_ExpandUnsqueezeExpandPattern 0 0 2.129999e-06
match_GeluPattern 0 0 1.169000e-06
match_IdentityPattern 0 0 4.260000e-06
match_LeakyReluPattern 0 0 6.004000e-06
match_MulUnsqueezeUnsqueezePattern 0 0 2.547999e-06
match_ReshapePattern 0 0 2.777000e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 2.578000e-06
match_ShapeBasedStaticExpandPattern 0 0 2.531000e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 2.184000e-06
match_ShapeBasedIdentityPattern 0 0 2.072000e-06
match_ShapedBasedReshapePattern 0 0 2.269000e-06
match_ShapeBasedSameChildrenPattern 0 0 2.664000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.924000e-06
match_ReshapeReshapePattern 0 0 2.939000e-06
match_SameChildrenPattern 0 0 4.332999e-06
match_SameChildrenFromInputPattern 0 0 7.586000e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 4.184000e-06
match_SqueezeAddPattern 0 0 4.073000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.918000e-06
match_SqueezeUnsqueezePattern 0 0 1.748000e-06
match_StaticConcatReshapePattern 0 0 1.728999e-06
match_SwapExpandReshapePattern 0 0 1.386000e-06
match_SwapExpandUnsqueezePattern 0 0 1.638000e-06
match_SwapUnaryPattern 0 0 1.588000e-06
match_SwapUnsqueezeTransposePattern 0 0 1.767999e-06
match_TransposeGatherPattern 0 0 1.591000e-06
match_TransposeReshapeTransposePattern 0 0 1.573000e-06
match_TransposeTransposePattern 0 0 1.525000e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.847000e-06
match_UnsqueezeReshapePattern 0 0 2.082000e-06
match_UnsqueezeUnsqueezePattern 0 0 1.746000e-06
match_FunctionAttentionPattern 0 0 2.004999e-06
match_FunctionAttentionGQAPattern 0 0 2.067000e-06
check_pattern_A20 0 0 8.506000e-06
remove_duplicated_shape 0 0 1.736999e-06
check_pattern_BD0 0 0 4.835000e-06
remove_identity_nodes 0 0 1.503400e-05
check_pattern_BI0 0 0 4.073000e-06
remove_unused 0 0 9.822000e-06
check_pattern_BUS0 0 0 3.851001e-06
build_graph_for_pattern 0 0 8.378999e-06
iteration_1 0 0 2.475500e-04
match_BatchNormalizationPattern 0 0 2.725999e-06
match_BatchNormalizationTrainingPattern 0 0 2.224000e-06
match_CastLayerNormalizationCastPattern 0 0 3.548000e-06
match_CastPattern 0 0 1.820999e-06
match_CastCastBinaryPattern 0 0 2.693000e-06
match_CastCastPattern 0 0 1.394000e-06
match_CastOpCastPattern 0 0 3.194999e-06
match_ClipClipPattern 0 0 3.150000e-06
match_ConcatEmptyPattern 0 0 2.387000e-06
match_ConcatGatherPattern 0 0 1.757000e-06
match_ConcatReshapePattern 0 0 1.952000e-06
match_ConcatTwiceUnaryPattern 0 0 3.344000e-06
match_ConstantToInitializerPattern 0 0 2.395001e-06
match_ConvBiasNullPattern 0 0 1.717000e-06
match_PadConvPattern 0 0 1.711001e-06
match_DropoutPattern 0 0 2.380000e-06
match_ExpandPattern 0 0 2.076999e-06
match_ExpandBroadcastPattern 0 0 2.710000e-06
match_ExpandSwapPattern 0 0 3.376000e-06
match_ExpandUnsqueezeExpandPattern 0 0 1.482001e-06
match_GathersSplitPattern 0 0 2.946000e-06
match_GeluPattern 0 0 7.499993e-07
match_IdentityPattern 0 0 2.133000e-06
match_LayerNormalizationPattern 0 0 3.393000e-06
match_LayerNormalizationScalePattern 0 0 2.570000e-06
match_LeakyReluPattern 0 0 4.541001e-06
match_MaxReluPattern 0 0 3.053000e-06
match_MulMulMulScalarPattern 0 0 3.347000e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.775001e-06
match_NotNotPattern 0 0 2.578000e-06
match_NotWherePattern 0 0 2.403001e-06
match_ReduceArgTopKPattern 0 0 2.852000e-06
match_ReduceReshapePattern 0 0 2.571999e-06
match_ReduceSumNormalizePattern 0 0 4.138999e-06
match_ReshapePattern 0 0 2.319000e-06
match_ReshapeMatMulReshapePattern 0 0 2.857000e-06
match_Reshape2Of3Pattern 0 0 3.012000e-06
match_ReshapeReshapeBinaryPattern 0 0 2.392000e-06
match_GemmTransposePattern 0 0 2.787000e-06
match_MatMulReshape2Of3Pattern 0 0 3.310000e-06
match_MulMulMatMulPattern 0 0 2.147000e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 1.782000e-06
match_ShapeBasedStaticExpandPattern 0 0 1.617001e-06
match_ShapeBasedConcatExpandPattern 0 0 2.436001e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.915000e-06
match_ShapeBasedIdentityPattern 0 0 1.632000e-06
match_ShapeBasedExpandBroadcastPattern 0 0 2.696999e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 2.522000e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 2.154000e-06
match_ShapeBasedExpandSwapPattern 0 0 2.822000e-06
match_ShapeBasedMatMulToMulPattern 0 0 3.151999e-06
match_ShapedBasedReshapePattern 0 0 1.724000e-06
match_ShapeBasedSameChildrenPattern 0 0 1.988000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.610000e-06
match_ReshapeReshapePattern 0 0 1.840000e-06
match_RotaryEmbeddingPattern 0 0 2.870000e-06
match_SameChildrenPattern 0 0 3.249000e-06
match_SameChildrenFromInputPattern 0 0 3.717000e-06
match_SequenceConstructAtPattern 0 0 2.304000e-06
match_SplitToSequenceSequenceAtPattern 0 0 2.450000e-06
match_SliceSlicePattern 0 0 2.813999e-06
match_SlicesSplitPattern 0 0 3.005900e-05
match_SoftmaxCrossEntropyLossCastPattern 0 0 9.768000e-06
match_SplitConcatPattern 0 0 4.363000e-06
match_SqueezeAddPattern 0 0 3.406000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 2.153000e-06
match_SqueezeUnsqueezePattern 0 0 2.004000e-06
match_StaticConcatReshapePattern 0 0 2.626000e-06
match_Sub1MulPattern 0 0 3.799000e-06
match_SwapExpandReshapePattern 0 0 1.577000e-06
match_SwapExpandUnsqueezePattern 0 0 1.378000e-06
match_SwapRangeAddScalarPattern 0 0 3.508800e-05
match_SwapUnaryPattern 0 0 3.096000e-06
match_SwapUnsqueezeTransposePattern 0 0 2.191000e-06
match_SwitchOrderBinaryPattern 0 0 3.320000e-06
match_SwitchReshapeActivationPattern 0 0 4.078000e-06
match_TransposeEqualReshapePattern 0 0 3.247000e-06
match_TransposeGatherPattern 0 0 1.622000e-06
match_TransposeMatMulPattern 0 0 2.810000e-06
match_TransposeReshapeMatMulPattern 0 0 2.633000e-06
match_TransposeReshapeTransposePattern 0 0 1.467000e-06
match_TransposeTransposePattern 0 0 1.748999e-06
match_UnsqueezeEqualPattern 0 0 2.993000e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 2.195000e-06
match_UnsqueezeReshapePattern 0 0 1.937000e-06
match_UnsqueezeUnsqueezePattern 0 0 1.628000e-06
match_WhereAddPattern 0 0 3.448001e-06
match_RotaryConcatPartPattern 0 0 3.098000e-06
match_FunctionAttentionPattern 0 0 2.340000e-06
match_FunctionAttentionGQAPattern 0 0 2.387000e-06
match_FunctionCausalMaskPattern 0 0 3.170000e-06
match_FunctionCausalMaskMulAddPattern 0 0 2.576000e-06
match_FunctionCosSinCachePattern 0 0 2.381000e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 2.728000e-06
match_RMSNormalizationPattern 0 0 2.532000e-06
match_RMSNormalizationMulPattern 0 0 2.375999e-06
check_pattern_A20 0 0 1.225300e-05
remove_duplicated_shape 0 0 1.898001e-06
check_pattern_BD0 0 0 4.955000e-06
remove_identity_nodes 0 0 1.709200e-05
check_pattern_BI0 0 0 4.768001e-06
remove_unused 0 0 1.358400e-05
check_pattern_BUS0 0 0 4.195000e-06
build_graph_for_pattern 0 0 8.780000e-06
iteration_2 0 0 5.426640e-04
match_BatchNormalizationPattern 0 0 2.848999e-06
match_BatchNormalizationTrainingPattern 0 0 2.195000e-06
match_CastLayerNormalizationCastPattern 0 0 2.543999e-06
match_CastPattern 0 0 1.945999e-06
match_CastCastBinaryPattern 0 0 1.868000e-06
match_CastCastPattern 0 0 1.686000e-06
match_CastOpCastPattern 0 0 2.369000e-06
match_ClipClipPattern 0 0 1.983000e-06
match_ConcatEmptyPattern 0 0 1.762000e-06
match_ConcatGatherPattern 0 0 2.163000e-06
match_ConcatReshapePattern 0 0 2.830000e-06
match_ConcatTwiceUnaryPattern 0 0 2.272999e-06
match_ConstantToInitializerPattern 0 0 1.796000e-06
match_ConvBiasNullPattern 0 0 1.651000e-06
match_PadConvPattern 0 0 1.558000e-06
match_DropoutPattern 0 0 1.813000e-06
match_ExpandPattern 0 0 1.665000e-06
match_ExpandBroadcastPattern 0 0 1.642000e-06
match_ExpandSwapPattern 0 0 1.959001e-06
match_ExpandUnsqueezeExpandPattern 0 0 1.411000e-06
match_GathersSplitPattern 0 0 1.846000e-06
match_GeluPattern 0 0 9.129999e-07
match_IdentityPattern 0 0 2.015000e-06
match_LayerNormalizationPattern 0 0 2.110000e-06
match_LayerNormalizationScalePattern 0 0 1.950000e-06
match_LeakyReluPattern 0 0 5.248000e-06
match_MaxReluPattern 0 0 1.988001e-06
match_MulMulMulScalarPattern 0 0 1.788000e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.583000e-06
match_NotNotPattern 0 0 1.790000e-06
match_NotWherePattern 0 0 2.149000e-06
match_ReduceArgTopKPattern 0 0 2.035000e-06
match_ReduceReshapePattern 0 0 1.971001e-06
match_ReduceSumNormalizePattern 0 0 1.810000e-06
match_ReshapePattern 0 0 1.795001e-06
match_ReshapeMatMulReshapePattern 0 0 2.215000e-06
match_Reshape2Of3Pattern 0 0 2.381001e-06
match_ReshapeReshapeBinaryPattern 0 0 1.638999e-06
match_GemmTransposePattern 0 0 1.661000e-06
match_MatMulReshape2Of3Pattern 0 0 2.177000e-06
match_MulMulMatMulPattern 0 0 2.076000e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 2.170000e-06
match_ShapeBasedStaticExpandPattern 0 0 1.604000e-06
match_ShapeBasedConcatExpandPattern 0 0 2.152000e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.844000e-06
match_ShapeBasedIdentityPattern 0 0 1.710000e-06
match_ShapeBasedExpandBroadcastPattern 0 0 1.765000e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 1.535000e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 3.997000e-06
match_ShapeBasedExpandSwapPattern 0 0 2.112000e-06
match_ShapeBasedMatMulToMulPattern 0 0 1.576001e-06
match_ShapedBasedReshapePattern 0 0 2.076000e-06
match_ShapeBasedSameChildrenPattern 0 0 1.986000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.696000e-06
match_ReshapeReshapePattern 0 0 1.767001e-06
match_RotaryEmbeddingPattern 0 0 2.025000e-06
match_SameChildrenPattern 0 0 3.808000e-06
match_SameChildrenFromInputPattern 0 0 3.842000e-06
match_SequenceConstructAtPattern 0 0 1.697000e-06
match_SplitToSequenceSequenceAtPattern 0 0 1.752000e-06
match_SliceSlicePattern 0 0 1.512000e-06
match_SlicesSplitPattern 0 0 1.769000e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 3.685999e-06
match_SplitConcatPattern 0 0 1.852000e-06
match_SqueezeAddPattern 0 0 1.796000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.450000e-06
match_SqueezeUnsqueezePattern 0 0 1.559000e-06
match_StaticConcatReshapePattern 0 0 1.599000e-06
match_Sub1MulPattern 0 0 1.499000e-06
match_SwapExpandReshapePattern 0 0 1.405000e-06
match_SwapExpandUnsqueezePattern 0 0 1.384999e-06
match_SwapRangeAddScalarPattern 0 0 1.402000e-06
match_SwapUnaryPattern 0 0 1.657000e-06
match_SwapUnsqueezeTransposePattern 0 0 1.341999e-06
match_SwitchOrderBinaryPattern 0 0 1.519999e-06
match_SwitchReshapeActivationPattern 0 0 1.687000e-06
match_TransposeEqualReshapePattern 0 0 1.626000e-06
match_TransposeGatherPattern 0 0 1.564000e-06
match_TransposeMatMulPattern 0 0 1.705000e-06
match_TransposeReshapeMatMulPattern 0 0 1.599999e-06
match_TransposeReshapeTransposePattern 0 0 1.449000e-06
match_TransposeTransposePattern 0 0 1.388000e-06
match_UnsqueezeEqualPattern 0 0 1.314001e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.777000e-06
match_UnsqueezeReshapePattern 0 0 1.530000e-06
match_UnsqueezeUnsqueezePattern 0 0 1.367999e-06
match_WhereAddPattern 0 0 1.621000e-06
match_RotaryConcatPartPattern 0 0 1.693999e-06
match_FunctionAttentionPattern 0 0 1.793000e-06
match_FunctionAttentionGQAPattern 0 0 1.779000e-06
match_FunctionCausalMaskPattern 0 0 1.843000e-06
match_FunctionCausalMaskMulAddPattern 0 0 1.701001e-06
match_FunctionCosSinCachePattern 0 0 1.661000e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 1.574999e-06
match_RMSNormalizationPattern 0 0 1.641000e-06
match_RMSNormalizationMulPattern 0 0 1.412001e-06
match_AttentionGQAPattern 0 0 2.403001e-06
check_pattern_A20 0 0 7.721001e-06
remove_duplicated_shape 0 0 1.261999e-06
check_pattern_BD0 0 0 7.464000e-06
remove_identity_nodes 0 0 1.348200e-05
check_pattern_BI0 0 0 4.292000e-06
remove_unused 0 0 9.450000e-06
check_pattern_BUS0 0 0 3.944000e-06
build_graph_for_pattern 0 0 7.560000e-06
iteration_3 0 0 3.722380e-04
match_BatchNormalizationPattern 0 0 2.631000e-06
match_BatchNormalizationTrainingPattern 0 0 1.621001e-06
match_CastLayerNormalizationCastPattern 0 0 2.113000e-06
match_CastPattern 0 0 1.529000e-06
match_CastCastBinaryPattern 0 0 1.525999e-06
match_CastCastPattern 0 0 1.678000e-06
match_CastOpCastPattern 0 0 2.587000e-06
match_ClipClipPattern 0 0 1.746000e-06
match_ConcatEmptyPattern 0 0 1.512000e-06
match_ConcatGatherPattern 0 0 1.773000e-06
match_ConcatReshapePattern 0 0 2.224000e-06
match_ConcatTwiceUnaryPattern 0 0 1.847000e-06
match_ConstantToInitializerPattern 0 0 1.708000e-06
match_ConvBiasNullPattern 0 0 1.435000e-06
match_PadConvPattern 0 0 1.658000e-06
match_DropoutPattern 0 0 1.498000e-06
match_ExpandPattern 0 0 1.310001e-06
match_ExpandBroadcastPattern 0 0 1.636000e-06
match_ExpandSwapPattern 0 0 1.585000e-06
match_ExpandUnsqueezeExpandPattern 0 0 2.112000e-06
match_GathersSplitPattern 0 0 1.486000e-06
match_GeluPattern 0 0 7.720000e-07
match_IdentityPattern 0 0 1.895000e-06
match_LayerNormalizationPattern 0 0 1.635000e-06
match_LayerNormalizationScalePattern 0 0 1.581000e-06
match_LeakyReluPattern 0 0 4.304000e-06
match_MaxReluPattern 0 0 1.869000e-06
match_MulMulMulScalarPattern 0 0 1.636000e-06
match_MulUnsqueezeUnsqueezePattern 0 0 1.465000e-06
match_NotNotPattern 0 0 1.497000e-06
match_NotWherePattern 0 0 1.462000e-06
match_ReduceArgTopKPattern 0 0 1.992000e-06
match_ReduceReshapePattern 0 0 1.668000e-06
match_ReduceSumNormalizePattern 0 0 1.532000e-06
match_ReshapePattern 0 0 1.656000e-06
match_ReshapeMatMulReshapePattern 0 0 1.702000e-06
match_Reshape2Of3Pattern 0 0 4.373001e-06
match_ReshapeReshapeBinaryPattern 0 0 1.503000e-06
match_MatMulAddPattern 0 0 3.337000e-06
match_GemmTransposePattern 0 0 1.924000e-06
match_MatMulReshape2Of3Pattern 0 0 1.896000e-06
match_MulMulMatMulPattern 0 0 1.759000e-06
match_ShapeBasedReshapeIsSqueezePattern 0 0 1.642999e-06
match_ShapeBasedStaticExpandPattern 0 0 1.535000e-06
match_ShapeBasedConcatExpandPattern 0 0 2.051000e-06
match_ShapeBasedEditDistanceReshapePattern 0 0 1.742000e-06
match_ShapeBasedIdentityPattern 0 0 1.625001e-06
match_ShapeBasedExpandBroadcastPattern 0 0 1.642999e-06
match_ShapeBasedExpandBroadcastMatMulPattern 0 0 1.850000e-06
match_ShapeBasedExpandCastWhereSwapPattern 0 0 1.754001e-06
match_ShapeBasedExpandSwapPattern 0 0 1.694000e-06
match_ShapeBasedMatMulToMulPattern 0 0 1.519001e-06
match_ShapedBasedReshapePattern 0 0 1.881000e-06
match_ShapeBasedSameChildrenPattern 0 0 1.585000e-06
match_ShapeBasedShapeShapeAddPattern 0 0 1.539000e-06
match_ReshapeReshapePattern 0 0 1.584000e-06
match_RotaryEmbeddingPattern 0 0 1.549000e-06
match_SameChildrenPattern 0 0 3.095000e-06
match_SameChildrenFromInputPattern 0 0 3.654000e-06
match_SequenceConstructAtPattern 0 0 1.603999e-06
match_SplitToSequenceSequenceAtPattern 0 0 1.581000e-06
match_SliceSlicePattern 0 0 1.484000e-06
match_SlicesSplitPattern 0 0 1.418000e-06
match_SoftmaxCrossEntropyLossCastPattern 0 0 3.781000e-06
match_SplitConcatPattern 0 0 1.503000e-06
match_SqueezeAddPattern 0 0 1.780000e-06
match_SqueezeBinaryUnsqueezePattern 0 0 1.318000e-06
match_SqueezeUnsqueezePattern 0 0 1.459000e-06
match_StaticConcatReshapePattern 0 0 1.544000e-06
match_Sub1MulPattern 0 0 1.409000e-06
match_SwapExpandReshapePattern 0 0 1.314999e-06
match_SwapExpandUnsqueezePattern 0 0 1.252000e-06
match_SwapRangeAddScalarPattern 0 0 1.218999e-06
match_SwapUnaryPattern 0 0 1.799001e-06
match_SwapUnsqueezeTransposePattern 0 0 1.455000e-06
match_SwitchOrderBinaryPattern 0 0 1.599000e-06
match_SwitchReshapeActivationPattern 0 0 1.999000e-06
match_TransposeEqualReshapePattern 0 0 1.558999e-06
match_TransposeGatherPattern 0 0 1.544000e-06
match_TransposeMatMulPattern 0 0 1.535001e-06
match_TransposeReshapeMatMulPattern 0 0 1.404001e-06
match_TransposeReshapeTransposePattern 0 0 1.437999e-06
match_TransposeTransposePattern 0 0 1.353000e-06
match_UnsqueezeEqualPattern 0 0 1.358000e-06
match_UnsqueezeOrSqueezeReshapePattern 0 0 1.462000e-06
match_UnsqueezeReshapePattern 0 0 1.466000e-06
match_UnsqueezeUnsqueezePattern 0 0 1.533999e-06
match_WhereAddPattern 0 0 1.845000e-06
match_RotaryConcatPartPattern 0 0 1.427000e-06
match_FunctionAttentionPattern 0 0 1.792000e-06
match_FunctionAttentionGQAPattern 0 0 1.744999e-06
match_FunctionCausalMaskPattern 0 0 1.661000e-06
match_FunctionCausalMaskMulAddPattern 0 0 1.894000e-06
match_FunctionCosSinCachePattern 0 0 1.713001e-06
match_FunctionHalfRotaryEmbeddingPattern 0 0 4.078000e-06
match_RMSNormalizationPattern 0 0 1.631000e-06
match_RMSNormalizationMulPattern 0 0 1.549000e-06
match_AttentionGQAPattern 0 0 1.426000e-06
check_pattern_A20 0 0 6.324000e-06
remove_duplicated_shape 0 0 1.220000e-06
check_pattern_BD0 0 0 4.336000e-06
remove_identity_nodes 0 0 1.249700e-05
check_pattern_BI0 0 0 4.166001e-06
remove_unused 0 0 8.479000e-06
check_pattern_BUS0 0 0 3.847000e-06
build_graph_for_pattern 0 0 7.280000e-06
check_patterns-4 0 0 8.571999e-06
remove_unused 0 0 9.122000e-06
check_remove_unused-5 0 0 4.173000e-06
remove_identity 0 0 1.076000e-05
check_remove_identity-6 0 0 3.915000e-06
constant_folding 0 0 2.523000e-05
apply_constant_folding_new_inits 0 0 NaN
check_constant_folding-7 0 0 9.793000e-06
remove_unused 0 0 1.090200e-05
check_remove_unused-8 0 0 4.818000e-06
remove_duplicated_initializer 0 0 2.463999e-06
check_remove_duplicated_initializer-9 0 0 4.114000e-06
remove_identity 0 0 1.384200e-05
check_remove_identity-10 0 0 4.088000e-06
remove_unused 0 0 7.635001e-06
check_remove_unused-11 0 0 5.627000e-06
order 0 0 4.743200e-05
check_orderA 0 0 5.808000e-06
check_orderL 0 0 4.021001e-06
shape_order 0 0 1.857000e-05
order 0 0 NaN
check_order-12 0 0 4.566000e-06
optimization 0 2 6.615706e-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 0.000211
optimization 0 2 0.004884
remove_identity 1 2 0.000049
patterns 0 1 0.004588
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,
)
proto = func.proto
print(type(proto).__name__)
print("function name :", proto.name)
print("function domain:", proto.domain)
>>>
FunctionProto
function name : MyRelu
function domain: my.domain
Debugging GraphBuilder with Environment Variables#
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