Multiple fixes to enable multi-GPUs forward execution

Some of these fix allow multi-GPU forward execution of cloned graphs.

• Correctly escape HTML special characters in GraphView::save();
• Added const version of refCastFrom();
• Make GraphView::add() from another GraphView preserve ordered inputs/outputs;
• Operators attributes were not properly cloned (should fix aidge!105 (merged)).

This now works:

aidge_model0 = aidge_onnx.load_onnx("mymodel.onnx")
aidge_model1 = aidge_model0.clone()
aidge_model0.set_backend("cuda", 0)
aidge_model1.set_backend("cuda", 1)

aidge_models = aidge_core.GraphView()
aidge_models.add(aidge_model0)
aidge_models.add(aidge_model1)

scheduler = aidge_core.ParallelScheduler(aidge_models)
scheduler.forward(True, [in_model0, in_model1])

The models aidge_model0 and aidge_model1 will be run in parallel on 2 GPUs, yet in a synchronized fashion (the same layers are executed at the same time). This should be generalizable to any number of device/backend combinations!

This also works:

aidge_model0 = aidge_onnx.load_onnx("mymodel.onnx")
aidge_model1 = aidge_model0.clone()
aidge_model0.set_backend("cuda", 0)
aidge_model1.set_backend("cuda", 1)

scheduler0 = aidge_core.SequentialScheduler(aidge_model0)
scheduler1 = aidge_core.SequentialScheduler(aidge_model1)

import concurrent.futures

with concurrent.futures.ThreadPoolExecutor(max_workers=2) as e:
    e.submit(scheduler0.forward, True, [in_model0])
    e.submit(scheduler1.forward, True, [in_model1])

Here models aidge_model0 and aidge_model1 are run in parallel fully asynchronously.

changed milestone to %aidge v0.6.0

added Fix 🔥🔥 label

assigned to @olivierbichler

changed title from Added htmlEscape to Multiple fixes

changed the description

added 2 commits

• 0fa98cce - Added const version of refCastFrom
• 9b4eae91 - Make inputs/outputs order deterministic when adding a GraphView to a GraphView

Compare with previous version

added 4 commits

• badf5d93 - 1 commit from branch dev
• 5f0a18f5 - Added htmlEscape
• 2480e05d - Added const version of refCastFrom
• 050faae9 - Make inputs/outputs order deterministic when adding a GraphView to a GraphView

Compare with previous version

added 1 commit

• 7d665ab6 - Correct inputs/outputs order

Compare with previous version

added 1 commit

• 7957949f - Fixed attributes not properly cloned

Compare with previous version

changed the description

changed title from Multiple fixes to Multiple fixes to enable multi-GPUs forward execution

added StatusReview Ready label

changed the description

changed the description

merged

mentioned in merge request aidge!105 (merged)