MAX graphs takes along time to compile, around 300-500 ms even for smaller graphs as well.
The compilation becomes a bit faster on next consecutive code execution due to caching but still the time compilation time for graphs remains above 250 ms.
Note: enable_model_ir_cache is already true in modular.cfg file.
Steps to reproduce
Code example:
fn main() raises:
var session = engine.InferenceSession()
start = now()
var graph1 = Graph(in_types=List[Type](TensorType(DType.float32, "a","m")))
var transposed = ops.transpose(graph1[0],-1,-2)
graph1.output(transposed)
graph1.verify()
var transpose = session.load(graph1)
end1 = now()
print("total compile time for graph1: ",(end1 - start)/1000000000)
var graph2 = Graph(in_types=List[Type](TensorType(DType.float32, "m", "n"), TensorType(DType.float32, "n","x")))
var out3 = graph2[0] @ graph2[1]
graph2.output(out3)
graph2.verify()
var multiplication = session.load(graph2)
end2 = now()
print("total compile time for graph2: ",(end2 - end1)/1000000000)
The output is following:
total compile time for graph1: 0.31278938899999997
total compile time for graph2: 0.451578655
System information
OS: Ubuntu 22.04.3 LTS
Mojo version: mojo 24.5.0 (e8aacb95)
max 24.5.0 (e8aacb95)
Modular version 24.5.0
Bug description
MAX graphs takes along time to compile, around 300-500 ms even for smaller graphs as well. The compilation becomes a bit faster on next consecutive code execution due to caching but still the time compilation time for graphs remains above 250 ms. Note: enable_model_ir_cache is already true in modular.cfg file.
Steps to reproduce
Code example:
The output is following:
System information