Update save.py

This commit is contained in:
Daniel Han-Chen 2024-03-10 18:09:58 +11:00
commit 08da057f04

View file

@ -90,14 +90,14 @@ def _merge_lora(layer, name):
W = fast_dequantize(W, quant_state)
else:
dtype = W.dtype
W = W.to(torch.float32).t()
# W = W.t()
# W = W.to(torch.float32).t()
W = W.t()
if A is not None:
# sAB = (A.t().to(torch.float32) @ (s * B.t().to(torch.float32)))
# W += sAB
W.addmm_(A.t().to(torch.float32), B.t().to(torch.float32), alpha = s)
# W.addmm_(A.t().to(W.dtype), B.t().to(W.dtype), alpha = s)
# W.addmm_(A.t().to(torch.float32), B.t().to(torch.float32), alpha = s)
W.addmm_(A.t().to(W.dtype), B.t().to(W.dtype), alpha = s)
# if not torch.isfinite(W).all():
maximum_element = torch.max(W.min().abs(), W.max())
if not torch.isfinite(maximum_element).item():