Operator Graph¶
Importable from syssim. The traced intermediate representation of a model step. Advanced —
needed only when inspecting or building graphs directly.
- class OperatorType¶
Enum of operator categories:
GEMM,ATTN,MATH(compute);COLLECTIVE(communication);MEMORY;BARRIER,STREAM_SYNC(sync).
- class TensorMeta(shape, dtype, device)¶
Frozen dataclass describing a tensor:
shape(tuple),dtype(str),device(str).- to_dict()¶
Serialize to
{"shape": [...], "dtype": ..., "device": ...}.
- class OperatorNode(name, op_type, config=None, predecessors=None, stream_id=0, device_id=0, inputs=None, outputs=None, estimated_time_ms=0.0)¶
A single node in the graph.
predecessorslists same-stream FIFO and cross-stream sync dependencies;inputs/outputsare lists ofTensorMeta.- to_dict()¶
Serialize all fields to a dict.
- class OperatorGraph(name='model')¶
DAG of
OperatorNodewith multi-stream critical-path analysis. Holdsoperators(dict by name) andstreams(set of stream ids);len(graph)is the operator count.- add_operator(node)¶
Add an
OperatorNodeto the graph.
- validate()¶
Validate the DAG: reference integrity and cycle detection (DFS coloring).
- topological_sort()¶
Kahn’s algorithm; cached until the graph is modified.
- to_json()¶
Serialize the graph to JSON.
- to_dot()¶
Graphviz DOT representation, color-coded by op type.
- summary()¶
Human-readable summary of op counts and total time.