mrq/bitsandbytes-rocm
1
0
Fork 1
Code Issues 1 Pull Requests Packages Projects Releases Wiki Activity

cast before allclose

Browse Source
This commit is contained in:
justheuristic 2022-09-18 01:24:27 +03:00
parent 95dafc6475
commit 28a9313ddc
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tests/test_modules.py
View File

@ -541,7 +541,7 @@ def test_linear8bitlt_no_fp16_weights(threshold, memory_efficient_backward):
mlp = MLP8bit(
32, 64, threshold=threshold, has_fp16_weights=False, memory_efficient_backward=memory_efficient_backward
)
w1, w2 = mlp.fc1.weight.clone(), mlp.fc2.weight.clone() # note: we grad original weights before quantization,
w1, w2 = mlp.fc1.weight.clone().cuda(), mlp.fc2.weight.clone().cuda() # grab weights before quantization,
mlp = mlp.cuda().half() # and this line triggers quantization
for i in range(100):
@ -567,7 +567,7 @@ 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.to(grad_proj.device) @ w1.to(grad_proj.device)
grad_ref = grad_proj.flatten(2) @ w2.to() @ w1.to(grad_proj.device)
assert torch.allclose(b1.grad, grad_ref)