My proposal is now implemented.
I ended up completely replacing the content of graph_plan_memory.cc with a
python implementation:
- Redirect to Python:
https://github.com/tum-ei-eda/tvm/blob/e9184d948edd58635e79c3f21355f2b83b361401/src/relay/backend/graph_plan_memory.cc#L890
- Main implementat
Hi @aca88
thanks for your interest!
For the evaluated models, we just used a single schedule as given by TVM:
https://github.com/tum-ei-eda/tvm/blob/tumeda_memplan/python/tvm/relay/memplan.py#L187
You are right that for more complex graphs, we would have to evaluate more
schedules to find t
While switching to TVMC, I noticed a "virtual_device" property on the top-level
relay module function. It was not properly propagated through my relay passes
and caused an assertion in lowering to TE, with:
Check failed: (!virtual_device->IsFullyUnconstrained()) is false
at:
```
File "
Hi Mark, thank you for clarifying.
If I'm not interested in using the virtual_device feature, is there a way to
disable it? The issue is that without the patch above, it is not possible to
use any pass that is based on the ExprMutator in Python, because the TE
Compiler complains with the abov