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

cast edge case

Browse Source
This commit is contained in:
justheuristic 2022-09-18 00:36:46 +03:00
parent cbfdf0b5ef
commit ab9dee062d
1 changed files with 1 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
bitsandbytes/autograd/_functions.py
View File

@ -221,9 +221,6 @@ class MatMul8bitLt(torch.autograd.Function):
# 3. Matmul # 3. Matmul
# 4. Mixed-precision decomposition matmul # 4. Mixed-precision decomposition matmul
# 5. Save state # 5. Save state
requires_gradA = A.requires_grad
requires_gradB = B.requires_grad
requires_gradBias = bias is not None and bias.requires_grad
formatB = state.formatB formatB = state.formatB
input_shape = A.shape input_shape = A.shape
if state.outlier_pool is None: if state.outlier_pool is None:
@ -330,7 +327,7 @@ class MatMul8bitLt(torch.autograd.Function):
ctx.grad_shape = input_shape ctx.grad_shape = input_shape
ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype
if requires_gradA or requires_gradB: if any(ctx.needs_input_grad[:2]):
ctx.tensors = (CAt, subA) ctx.tensors = (CAt, subA)
ctx.tensor_states = (SCAt, state.idx) ctx.tensor_states = (SCAt, state.idx)
else: else: