forked from mrq/bitsandbytes-rocm
cast device
This commit is contained in:
parent
e4086a2758
commit
01b4c6a048
|
@ -568,7 +568,8 @@ def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
|
|||
mlp.zero_grad()
|
||||
(o1 * grad_proj).sum().backward()
|
||||
grad_ref = grad_proj.flatten(2) @ w2.half() @ w1.half()
|
||||
assert torch.allclose(b1.grad, grad_ref)
|
||||
scale = grad_ref.abs().mean()
|
||||
assert torch.allclose(b1.grad, grad_ref, rtol=0, atol=0.1 * scale)
|
||||
|
||||
|
||||
|
||||
|
|
Loading…
Reference in New Issue
Block a user