diff options
| author | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:36:46 +0300 |
|---|---|---|
| committer | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:36:46 +0300 |
| commit | ab9dee062d791ef343ff5f9e8c2c85dc094219ed (patch) | |
| tree | 88ce521928a5abd9ddd46a9f6de51e20deb14482 /bitsandbytes | |
| parent | cbfdf0b5efe4923ba4533c274ce83072b7e502b5 (diff) | |
cast edge case
Diffstat (limited to 'bitsandbytes')
| -rw-r--r-- | bitsandbytes/autograd/_functions.py | 5 |
1 files changed, 1 insertions, 4 deletions
diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py index d0e48b7..1d0002c 100644 --- a/bitsandbytes/autograd/_functions.py +++ b/bitsandbytes/autograd/_functions.py @@ -221,9 +221,6 @@ class MatMul8bitLt(torch.autograd.Function): # 3. Matmul # 4. Mixed-precision decomposition matmul # 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 input_shape = A.shape if state.outlier_pool is None: @@ -330,7 +327,7 @@ class MatMul8bitLt(torch.autograd.Function): 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 - if requires_gradA or requires_gradB: + if any(ctx.needs_input_grad[:2]): ctx.tensors = (CAt, subA) ctx.tensor_states = (SCAt, state.idx) else: |