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

debug

Browse Source
This commit is contained in:
justheuristic 2022-09-18 01:13:58 +03:00
parent 5d65817101
commit d9b8789818
1 changed files with 8 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
tests/test_modules.py
View File

@ -545,6 +545,7 @@ def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
.to(torch.float16) .to(torch.float16)
.to("cuda") .to("cuda")
) )
w1, w2 = mlp.fc1.weight.clone(), mlp.fc2.weight.clone()
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()
@ -567,8 +568,15 @@ def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
assert o1.requires_grad assert o1.requires_grad
grad_proj = torch.randn_like(o1) grad_proj = torch.randn_like(o1)
mlp.zero_grad()
(o1 * grad_proj).sum().backward() (o1 * grad_proj).sum().backward()
grad_ref = grad_proj.flatten(2) @ w2 @ w1
assert torch.allclose(b1.grad, grad_ref)