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

delete CxB from state

Browse Source
This commit is contained in:
dbaranchuk 2022-08-24 01:33:04 +03:00
parent 876387dc0c
commit ef2936a90d
1 changed files with 4 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
bitsandbytes/nn/modules.py
View File

@ -260,11 +260,10 @@ class Linear8bitLt(nn.Linear):
out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state) out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
# if not self.state.has_fp16_weights and self.state.CB is not None: if not self.state.has_fp16_weights and self.state.CxB is not None:
# we converted 8-bit row major to turing/ampere format in the first inference pass # In this version, we convert 8-bit row major to turing/ampere format at each inference pass
# we no longer need the row-major weight # Thus, we delete CxB from the state. TODO: do not store it in the state in the first place.
# del self.state.CB del self.state.CxB
# self.weight.data = self.state.CxB
return out return out