diff options
| author | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:24:20 +0300 |
|---|---|---|
| committer | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:24:20 +0300 |
| commit | 14048a3c16bf3e60754b4218ec40a01e6a7f213c (patch) | |
| tree | baf4dde5cfb84d15ecc256dff5d2e135e98ed9f9 | |
| parent | 5b169f18e4894a82b8681139727d45a4dd61c4b1 (diff) | |
safer cast
| -rw-r--r-- | bitsandbytes/autograd/_functions.py | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py index 93304f9..6a225b7 100644 --- a/bitsandbytes/autograd/_functions.py +++ b/bitsandbytes/autograd/_functions.py @@ -371,7 +371,7 @@ class MatMul8bitLt(torch.autograd.Function): gradB32, SgradB32 = F.igemmlt(C32grad, CxAt, Sgrad, SAt) grad_B = F.mm_dequant(gradB32, SgradB32, SCgradt, SCAt).to(ctx.dtype_B) if state.threshold > 0.0 and subA is not None: - grad_B[:, idx].addmm_(grad_output.t(), subA) + grad_B[:, idx] += torch.mm(grad_output.t(), subA) if req_gradA: if state.CBt is not None: @@ -384,6 +384,7 @@ class MatMul8bitLt(torch.autograd.Function): grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape).to(ctx.dtype_A) elif state.CB is not None: + raise NotImplementedError("WIP") CB = state.CB.to(ctx.dtype_B) CB.mul_(state.SCB.unsqueeze(1).div_(127.0).to(CB.dtype)) grad_A = torch.matmul(grad_output, CB).view(ctx.grad_shape).to(ctx.dtype_A) |