add memory effcient backward option

2 files changed, 52 insertions, 10 deletions

diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py
index 226cbb5..271c690 100644
--- a/bitsandbytes/autograd/_functions.py
+++ b/bitsandbytes/autograd/_functions.py
@@ -1,5 +1,6 @@
 import operator
 import torch
+import bitsandbytes as bnb
 import bitsandbytes.functional as F
 
 from dataclasses import dataclass
@@ -187,6 +188,8 @@ class MatmulLtState:
     use_pool = False
     formatB = F.get_special_format_str()
 
+    memory_efficient_backward = False
+
     def reset_grads(self):
         self.CB = None
         self.CxB = None
@@ -283,6 +286,12 @@ 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)
@@ -332,13 +341,15 @@ class MatMul8bitLt(torch.autograd.Function):
         clone_func = torch.clone if len(output_shape) == 3 else lambda x : x
         return clone_func(output.view(output_shape))
 
+    @staticmethod
     def backward(ctx, grad_output):
         if ctx.is_empty:
             bias_grad = (None if ctx.bias is None else torch.zeros_like(ctx.bias))
             return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None
-
         req_gradA, req_gradB, req_gradBias = ctx.req_grads
-        assert not req_gradB, "TODO: support weight updates as well"
+        CAt, subA = ctx.tensors
+        SCAt, idx = ctx.tensor_states
+        formatB = ctx.formatB
         state = ctx.state
 
         # Cast grad_output to fp16
@@ -352,11 +363,31 @@ class MatMul8bitLt(torch.autograd.Function):
 
         grad_A = grad_B = grad_bias = None
 
+        Cgrad, Cgradt, SCgrad, SCgradt, coo_tensor = F.double_quant(grad_output)
+        if req_gradB:
+            CxAt, SAt = F.transform(CAt, formatB, transpose=True)
+            C32grad, Sgrad = F.transform(Cgradt, "col32", transpose=True)
+            gradB32, SgradB32 = F.igemmlt(C32grad, CxAt, Sgrad, SAt)
+            grad_B = F.mm_dequant(gradB32, SgradB32, SCgradt, SCAt)
+            if state.threshold > 0.0 and subA is not None:
+                grad_B[:, idx] += torch.matmul(grad_output.t(), subA)
+
         if req_gradA:
-            CB = state.CB.half()
-            SCB = (state.SCB.unsqueeze(1) / 127.0).half()
-            CB *= SCB
-            grad_A = torch.mm(grad_output, CB).view(ctx.grad_shape)
+            if state.CBt:
+                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)
+                grad_A = F.mm_dequant(gradA32, SgradA32, SCgrad, state.SCBt).view(ctx.grad_shape)
+            elif state.CB:
+                CB = state.CB.half()
+                SCB = (state.SCB.unsqueeze(1) / 127.0).half()
+                CB *= SCB
+                grad_A = torch.mm(grad_output, CB).view(ctx.grad_shape)
+            else:
+                raise Exception('State must contain either CBt or CB matrix')
 
         if req_gradBias:
             grad_bias = grad_output.sum(0)
@@ -367,6 +398,9 @@ class MatMul8bitLt(torch.autograd.Function):
         return grad_A, grad_B, None, grad_bias, None
 
 
+matmul = MatMul8bitLt.apply
+
+
 def matmul(
     A: tensor,
     B: tensor,
diff --git a/bitsandbytes/nn/modules.py b/bitsandbytes/nn/modules.py
index 3e32c8e..00d0c61 100644
--- a/bitsandbytes/nn/modules.py
+++ b/bitsandbytes/nn/modules.py
@@ -223,6 +223,7 @@ class Linear8bitLt(nn.Linear):
         has_fp16_weights=True,
         threshold=0.0,
         index=None,
+        memory_efficient_backward=False
     ):
         super(Linear8bitLt, self).__init__(
             input_features, output_features, bias
@@ -232,6 +233,7 @@ class Linear8bitLt(nn.Linear):
 
         self.state.threshold = threshold
         self.state.has_fp16_weights = has_fp16_weights
+        self.state.memory_efficient_backward = memory_efficient_backward
         if threshold > 0.0 and not has_fp16_weights:
             self.state.use_pool = True
 
@@ -255,10 +257,16 @@ 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.CxB is not None:
-            # In this version, we convert 8-bit row major to turing/ampere format at each inference pass
-            # Thus, we delete CxB from the state. TODO: do not store it in the state in the first place. 
-            del self.state.CxB
+        if not self.state.has_fp16_weights:
+            if not self.state.memory_efficient_backward 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
+            elif self.state.memory_efficient_backward and self.state.CxB is not None:
+                # For memory efficient backward, we convert 8-bit row major to turing/ampere format at each inference pass.
+                # Thus, we delete CxB from the state. 
+                del self.state.CxB
 
         return out