add memory effcient backward option

1 files changed, 12 insertions, 4 deletions

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