diff options
| author | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:52:53 +0300 |
|---|---|---|
| committer | justheuristic <justheuristic@gmail.com> | 2022-09-18 00:52:53 +0300 |
| commit | 591f60395a1e9c62f291e23c91af45cc699f072c (patch) | |
| tree | 1f4ff32a1e490d9a872286cdf2d4f43eb0f1df2a /bitsandbytes/autograd | |
| parent | 579b8c782f5240d589ca65ef950054734db97ae1 (diff) | |
add memory efficient backward
Diffstat (limited to 'bitsandbytes/autograd')
| -rw-r--r-- | bitsandbytes/autograd/_functions.py | 1 |
1 files changed, 0 insertions, 1 deletions
diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py index 6674a82..daf9ba0 100644 --- a/bitsandbytes/autograd/_functions.py +++ b/bitsandbytes/autograd/_functions.py @@ -381,7 +381,6 @@ 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) |