OpenVINO reduce System RAM usage

This commit is contained in:
Disty0
2023-12-23 16:42:24 +03:00
parent 843340d3b3
commit 9dcc76af02
2 changed files with 3 additions and 0 deletions
+1
View File
@@ -76,6 +76,7 @@
- compatibility improvements
- **OpenVINO**, thanks @disty0
- **8 bit support for CPUs**
- Reduce System RAM usage when compiling
- add *Directory for OpenVINO cache* option to *System Paths*
- remove Intel ARC specific 1024x1024 workaround
- **HDR controls**
+2
View File
@@ -397,6 +397,8 @@ def openvino_fx(subgraph, example_inputs):
if inputs_reversed:
example_inputs.reverse()
model = make_fx(subgraph)(*example_inputs)
# Deleting unused subgraphs doesn't do anything, so we cast it down to fp8
subgraph = subgraph.to(dtype=torch.float8_e4m3fn)
for node in model.graph.nodes:
if node.target == torch.ops.aten.mul_.Tensor:
node.target = torch.ops.aten.mul.Tensor