deftdawg/bitsandbytes-rocm
1
0
Fork 0
forked from mrq/bitsandbytes-rocm
Code Pull Requests Activity

pre-cast

Browse Source
This commit is contained in:
justheuristic 2022-09-18 01:20:34 +03:00
parent d9b8789818
commit 6a826c41a6
1 changed files with 2 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tests/test_modules.py
View File

@ -538,14 +538,11 @@ def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
assert mlp.fc1.weight.device.type == "cuda" assert mlp.fc1.weight.device.type == "cuda"
assert mlp.fc2.weight.device.type == "cuda" assert mlp.fc2.weight.device.type == "cuda"
mlp = ( mlp = MLP8bit(
MLP8bit(
32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward 32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward
) )
.to(torch.float16)
.to("cuda")
)
w1, w2 = mlp.fc1.weight.clone(), mlp.fc2.weight.clone() w1, w2 = mlp.fc1.weight.clone(), mlp.fc2.weight.clone()
mlp = mlp.cuda().half()
for i in range(100): for i in range(100):
b1 = torch.randn(16, 8, 32, device="cuda").half() b1 = torch.randn(16, 8, 32, device="cuda").half()