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

un-fuse bias

Browse Source
This commit is contained in:
justheuristic 2022-09-17 23:51:28 +03:00
parent 56a074f6dc
commit e9b87112ee
1 changed files with 6 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
bitsandbytes/autograd/_functions.py
View File

@ -316,15 +316,14 @@ class MatMul8bitLt(torch.autograd.Function):
if bias is None or bias.dtype == torch.float16:
output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=bias)
output = output.to(A_dtype)
delayed_bias = None
else: # apply bias separately
output = F.mm_dequant(out32, Sout32, SCA, state.SCB, bias=None)
output = output.to(A_dtype).add_(bias)
delayed_bias = bias
# 4. Mixed-precision decomposition matmul
if coo_tensorA is not None and subA is not None:
output += torch.matmul(subA, state.subB)
output.addmm_(subA, state.subB)
# 5. Save state
ctx.state = state
@ -341,6 +340,9 @@ class MatMul8bitLt(torch.autograd.Function):
ctx.tensor_states = (None, None)
ctx.save_for_backward(None, None)
output = output.to(A_dtype)
if delayed_bias is not None:
output.add_(delayed_bias)
clone_func = torch.clone if len(output_shape) == 3 else lambda x : x
return clone_func(output.view(output_shape))