.xoptim.patterns_investigation

modules

• .xoptim.patterns_investigation.element_wise
• .xoptim.patterns_investigation.llm_patterns

class experimental_experiment.xoptim.patterns_investigation.SimplifyingEasyPatternFunction(verbose: int = 0, priority: int = 0, min_opset: int = 1)

Base class to build investigation patterns. See FunctionPowTanhPattern to see how to use it. It replaces a subgraph with a local function.

add_local_functions_to_builder(g2: GraphBuilder)

Adds missing local functions ot understand the applied pattern. There is none by default but there some pattern may need to overwrite this.

post_apply_pattern(g, *nodes)

Method to overload to apply as step after the pattern was applied.

experimental_experiment.xoptim.patterns_investigation.get_investigation_patterns(verbose: int = 0) → List[PatternOptimization]

Returns a default list of patterns for investigations. They do nothing but prints information if verbose > 0.

<<<

from experimental_experiment.xoptim.patterns_api import pattern_table_doc
from experimental_experiment.xoptim.patterns_investigation import (
    get_investigation_patterns,
)

print(pattern_table_doc(get_investigation_patterns(), as_rst=True))

>>>

	name	short_name	priority	doc
0	BinaryInvestigation	BinaryInvestigation	1	Looks into
1	FunctionPackedMatMulPattern	FunctionPackedMatMul	1	Replaces multiple MatMul (X,A), (X,B) by (X, concat(A,B))…
2	FunctionPowTanhPattern	FunctionPowTanh	0	Moves the nodes in match_pattern into a local function. .. runpython
3	FunctionSplitRotaryMulPattern	FunctionSplitRotaryMul	0	Moves the nodes in match_pattern into a local function. This is only applicable when the pattern to match has more than one node. .. runpython