Make PyTorch models up to 40% faster! Thunder is a source to source compiler for PyTorch. It enables using different hardware executors at once; across one or thousands of GPUs.
This is a continuation of the work in #262 . The phase 2 proposal targets the re-use of compiled thunder program by supporting TensorProxy with dynamic shape and adjusting caching policy accordingly.
This is an umbrella issue to track progress. I would like to leave detailed technical discussion in separate issues and link them. This section will be gradually populated as we start picking up speed:
shape propagation modeling in thunder
[ ] Enabling NumberProxy in TensorProxy.shape (toy PR #1027)
[x] Adding shape prim (PR #1113)
[ ] Issue that blocks the #1182
[ ] shape logic in prim needs to be modeled in trace Issue: #471
I think a better design should make this a constraint, rather than opaquely treat it as a cache miss by backend. Will target this when we proceed to static constraints.
🚀 Feature
This is a continuation of the work in #262 . The phase 2 proposal targets the re-use of compiled thunder program by supporting TensorProxy with dynamic shape and adjusting caching policy accordingly.
This is an umbrella issue to track progress. I would like to leave detailed technical discussion in separate issues and link them. This section will be gradually populated as we start picking up speed:
shape propagation modeling in thunder
Static constraints and prologue checks
Handling of existing shape checks