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

cast properly

Browse Source
This commit is contained in:
justheuristic 2022-09-18 00:28:03 +03:00
parent 702cc72018
commit 45dc1983e9
1 changed files with 1 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
bitsandbytes/autograd/_functions.py
View File

@ -294,7 +294,7 @@ class MatMul8bitLt(torch.autograd.Function):
(outliers * state.SCB.view(-1, 1) / 127.0) (outliers * state.SCB.view(-1, 1) / 127.0)
.t() .t()
.contiguous() .contiguous()
.half() .to(B.dtype)
) )
CA[:, state.idx.long()] = 0 CA[:, state.idx.long()] = 0
CAt[:, state.idx.long()] = 0 CAt[:, state.idx.long()] = 0
@ -321,7 +321,6 @@ class MatMul8bitLt(torch.autograd.Function):
# 4. Mixed-precision decomposition matmul # 4. Mixed-precision decomposition matmul
if coo_tensorA is not None and subA is not None: if coo_tensorA is not None and subA is not None:
assert subA.dtype == state.subB.dtype, (subA.dtype, state.subB.dtype)
output += torch.matmul(subA, state.subB) output += torch.matmul(subA, state.subB)
# 5. Save state # 5. Save state