1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
|
> :warning: **This is an experimental feature**
# Static Runtime
The premise of this approach is that a small subset of neural networks are well represented by a
completely flattened dataflow graph.
TorchScript supports a far more feature programming paradigm,
so many models will not work out of the box.
## Assumptions
This is a list of current assumptions for use with
this feature.
- Inference only execution
- Single CPU device
After `torch.jit.freeze` and inlining/constant propagation is run on the model:
- No control flow
- No submodule invocations
- No references to `self`
- Inlined weights (i.e. no calls to `GetAttr`)
## Planned features
- Memory planning
- Operator dispatch inlining
- Operator subsitution
- Weight layout transformations (pre-packing)
- Lowering to `torch.jit.tensorexpr`
|
