- Starting with PyTorch 2.9, jit.trace is no longer maintained.
- The goal of this effort is to port to the new
torch.export path.
This will require a significant refactor of the torch_graph module to replace the current logic. It may enable better analysis of symbolic dimensions, allowing more fine-grained serialization (by resolving constant dimensions directly when known as invariant at export time).
torch.jit has been tested and still works, but a bug has already surfaced with the conjunction of Python 3.14 and PyTorch >=2.9, causing RAM usage to explode (details).
torch.exportpath.This will require a significant refactor of the
torch_graphmodule to replace the current logic. It may enable better analysis of symbolic dimensions, allowing more fine-grained serialization (by resolving constant dimensions directly when known as invariant at export time).torch.jithas been tested and still works, but a bug has already surfaced with the conjunction of Python 3.14 and PyTorch >=2.9, causing RAM usage to explode (details).