add memory efficient backward

2 files changed, 28 insertions, 24 deletions

diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py
index 4dbf129..63e8ad5 100644
--- a/bitsandbytes/autograd/_functions.py
+++ b/bitsandbytes/autograd/_functions.py
@@ -245,11 +245,10 @@ class MatMul8bitLt(torch.autograd.Function):
                 subA = A[:, idx]
                 state.subB = B[:, idx].t().contiguous()
                 state.idx = idx
-            else:
-                if state.CxB is None:
-                    # B in in 8-bit row-major, we can transform it back to 16-bit to extract outlier dimensions
-                    # we also need to convert it to the turing/ampere format
-                    state.CxB, state.SB = F.transform(state.CB, to_order=formatB)
+            elif state.CxB is None:
+                # B in in 8-bit row-major, we can transform it back to 16-bit to extract outlier dimensions
+                # we also need to convert it to the turing/ampere format
+                state.CxB, state.SB = F.transform(state.CB, to_order=formatB)
         else:
             if not state.has_fp16_weights and state.CxB is None:
                 state.CxB, state.SB = F.transform(state.CB, to_order=formatB)
@@ -280,12 +279,6 @@ class MatMul8bitLt(torch.autograd.Function):
 
             outlier_idx = torch.unique(coo_tensorA.colidx)
             state.idx = outlier_idx
-            # state.outlier_pool.add_outliers(outlier_idx, A.shape[-1])
-            # if state.use_pool and state.outlier_pool.model_dim == A.shape[-1]:
-            #    # do not use pool for 2nd FFN layer
-            #    state.idx = state.outlier_pool.get_current_outlier_idx().to(A.device)
-            # else:
-            #    state.idx = outlier_idx
             outliers = F.extract_outliers(state.CxB, state.SB, state.idx.int())
             state.subB = (
                 (outliers * state.SCB.view(-1, 1) / 127.0)
@@ -343,12 +336,9 @@ class MatMul8bitLt(torch.autograd.Function):
         SCAt, idx = ctx.tensor_states
         formatB = ctx.formatB
         state = ctx.state
-        assert (
-            state.has_fp16_weights
-        ), "Backprop only supported for fp16 weights."
 
         if len(grad_output.shape) == 3:
-            grad_output = grad_output.view(
+            grad_output = grad_output.reshape(
                 -1, grad_output.shape[-1]
             ).contiguous()
 
@@ -365,11 +355,20 @@ class MatMul8bitLt(torch.autograd.Function):
 
         if req_gradA:
             C32grad, Sgrad = F.transform(Cgrad, "col32")
-            if state.CxBt is None:
-                state.CxBt, state.SBt = F.transform(
-                    state.CBt, to_order=formatB, transpose=True
-                )
-            gradA32, SgradA32 = F.igemmlt(C32grad, state.CxBt, Sgrad, state.SBt)
+            if state.CxBt is None and state.has_fp16_weights:
+                CBt = state.CBt
+            elif state.CxBt is None:
+                assert state.CBt is None
+                CB = state.CB.half()
+                SCB = state.SCB.unsquezee(1).half()
+                SCBt = state.SCBt.unsquezee(1).half()
+                Bt = (CB * SCB).t().contiguous()
+                CBt = (Bt / SCBt).t().to(torch.int8)
+
+            CxBt, SBt = F.transform(
+                CBt, to_order=formatB, transpose=True
+            )
+            gradA32, SgradA32 = F.igemmlt(C32grad, CxBt, Sgrad, SBt)
             grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape)
 
         if req_gradBias:
diff --git a/bitsandbytes/nn/modules.py b/bitsandbytes/nn/modules.py
index b222f54..ef7fefc 100644
--- a/bitsandbytes/nn/modules.py
+++ b/bitsandbytes/nn/modules.py
@@ -148,10 +148,12 @@ class Int8Params(torch.nn.Parameter):
         has_fp16_weights=False,
         CB=None,
         SCB=None,
+        SCBt=None,
     ):
         cls.has_fp16_weights = has_fp16_weights
         cls.CB = None
         cls.SCB = None
+        cls.SCBt = None
         if data is None:
             data = torch.empty(0)
         return torch.Tensor._make_subclass(cls, data, requires_grad)
@@ -165,10 +167,10 @@ class Int8Params(torch.nn.Parameter):
             B = self.data.contiguous().half().cuda(device)
             CB, CBt, SCB, SCBt, coo_tensorB = bnb.functional.double_quant(B)
             del CBt
-            del SCBt
             self.data = CB
             setattr(self, "CB", CB)
             setattr(self, "SCB", SCB)
+            setattr(self, "SCBt", SCBt)
 
         return self
 
@@ -210,6 +212,7 @@ class Int8Params(torch.nn.Parameter):
             )
             new_param.CB = self.CB
             new_param.SCB = self.SCB
+            new_param.SCB = self.SCBt
 
             return new_param
 
@@ -240,8 +243,10 @@ class Linear8bitLt(nn.Linear):
     def init_8bit_state(self):
         self.state.CB = self.weight.CB
         self.state.SCB = self.weight.SCB
+        self.state.SCBt = self.weight.SCBt
         self.weight.CB = None
         self.weight.SCB = None
+        self.weight.SCBt = None
 
     def forward(self, x):
         self.state.is_training = self.training
@@ -255,11 +260,11 @@ class Linear8bitLt(nn.Linear):
 
         out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
 
-        if not self.state.has_fp16_weights and self.state.CB is not None:
+        # if not self.state.has_fp16_weights and self.state.CB is not None:
             # we converted 8-bit row major to turing/ampere format in the first inference pass
             # we no longer need the row-major weight
-            del self.state.CB
-            self.weight.data = self.state.CxB
+            # del self.state.CB
+            # self.weight.data = self.state.CxB
 
         return out