Checkpoint to reduce fp8_weight tensor saved for backwards

float8_experimental/float8_dynamic_linear.py

@@ -40,6 +40,27 @@ def backward(ctx, gradY):
         )
 
 
+def cast_weight_linear(
+    x_fp8: Float8Tensor, weight: torch.Tensor, scale: torch.Tensor, bias, emulate: bool
+) -> torch.Tensor:
+    """Cast weight to fp8_e4m3fn and do linear
+    Why a new function for something that can be inlined?
+    Because we want to call torch utils checkpoint on this function.
+    We always want to recompute the cast of the weight to fp8 since we can, trivially
+    fuse this into the transpose/contiguous of the weight during the backwards.
+
+    Args:
+        x_fp8 (Float8Tensor): input activation in fp8
+        weight (torch.Tensor): weight tensor in higher precision
+        scale (torch.Tensor): scale tensor for weight
+        bias: bias tensor in higher precision
+        emulate (bool): whether to emulate fp8 matmul logic in float32
+    """
+    w_fp8 = Float8Tensor.to_float8(weight, scale, torch.float8_e4m3fn, emulate=emulate)
+    y = torch.nn.functional.linear(x_fp8, w_fp8, bias)
+    return y
+
+
 class Float8DynamicLinear(torch.nn.Linear):
     """
     A wrapper around a `torch.nn.Linear` module which does fp8 compute. By on the fly
@@ -48,9 +69,16 @@ class Float8DynamicLinear(torch.nn.Linear):
 
     def forward(self, x):
         x_fp8 = self.cast_to_float8(x)
-        w_fp8 = self.cast_to_float8(self.weight)
-
-        y = torch.nn.functional.linear(x_fp8, w_fp8, self.bias)
+        scale = tensor_to_scale(self.weight, torch.float8_e4m3fn)
+        y = torch.utils.checkpoint.checkpoint(
+            cast_weight_linear,
+            x_fp8,
+            self.weight,
+            scale,
+            self.bias,
+            self.emulate,
+            use_reentrant=False,
+        )
 
         # Cast gradY to float8_e5m2 during backward
         y = self.cast_to_float8e5m2_bw(y)