Create optimize.py
Browse files- optimize.py +24 -0
optimize.py
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from torch.utils.cpp_extension import CUDA_HOME
|
| 3 |
+
|
| 4 |
+
def optimize_model(model):
|
| 5 |
+
"""Apply various optimizations"""
|
| 6 |
+
# Mixed precision
|
| 7 |
+
model.half()
|
| 8 |
+
|
| 9 |
+
# CUDA optimizations
|
| 10 |
+
if torch.cuda.is_available():
|
| 11 |
+
model = model.to('cuda')
|
| 12 |
+
torch.backends.cudnn.benchmark = True
|
| 13 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 14 |
+
|
| 15 |
+
# Compile with torch.compile (PyTorch 2.0+)
|
| 16 |
+
if hasattr(torch, 'compile'):
|
| 17 |
+
model = torch.compile(model, mode="reduce-overhead")
|
| 18 |
+
|
| 19 |
+
return model
|
| 20 |
+
|
| 21 |
+
def memory_optimization():
|
| 22 |
+
"""Memory optimization techniques"""
|
| 23 |
+
torch.cuda.empty_cache()
|
| 24 |
+
torch.backends.cudnn.deterministic = False
|