[PT2] MinMax #3166
[PT2] MinMax #3166
Conversation
github-actions
bot
added
NNCF PT
| from nncf.experimental.torch2.function_hook.nncf_graph.nncf_graph_builder import GraphModelWrapper | ||
|
|
||
|
|
||
| class PT2Engine(Engine): |
There was a problem hiding this comment.
Can we reuse PTEngine as it done for the TorchFX backend?
There was a problem hiding this comment.
We can , but what a reason?
I have try to keep one way dependence torch2 from torch
There was a problem hiding this comment.
I was asked to remove the PTFXEngine by @alexsu52, await the same comment from him
| def _apply_insertion_transformation( | ||
| self, model: nn.Module, transformations: List[PT2InsertionCommand] |
There was a problem hiding this comment.
What do you think about introduction of transformations as it done for torchFX? https://github.com/openvinotoolkit/nncf/blob/develop/nncf/experimental/torch/fx/model_transformer.py#L195-L208
There was a problem hiding this comment.
don't see any advantages in this approach (using function instead of commands), but it makes development and debugging much more difficult.
This discuss should be part of refactoring of ModelTransformer,
There was a problem hiding this comment.
Ok, I have to ask so be sure PT2 and FX backends are aligned. I'll approve it, but what're the difficulties with transformations? Why closures worse for the debug?
|
Are conformance tests numbers available? |
(N) - experimental tracing without F/BC |
Changes
Introduce TORCH2 backend
MinMax algorithms for torch2 backend
Add handle_torch_function for quantization function to trace it by torch_function
Related tickets
152996