[TBD if for land] bring back torch.autograd.Function

Summary:

This approach is more readable as we add additional scaling options.

For now, seeing how many things break in 2024-07 with
torch.autograd.Function + subclasses + compile.

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: a7abb00
Pull Request resolved: #316

Author: vkuzo

float8_experimental/float8_linear.py

@@ -71,6 +71,58 @@ def _maybe_initialize_amaxes_scales_for_float8_cast(
         scale.copy_(new_scale)
 
 
+# this code was resurrected from https://github.com/pytorch-labs/float8_experimental/pull/128/files
+@torch._dynamo.allow_in_graph
+class manual_float8_matmul(torch.autograd.Function):
+    """
+    Like torch.matmul, but with the arguments in float8 
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        x_fp8,
+        w_fp8_t,
+    ):
+        ctx.save_for_backward(x_fp8, w_fp8_t)
+        # the reshapes are needed in order to make the shapes compatible with 
+        # torch.mm
+        orig_shape = x_fp8.shape
+        x_fp8_reshaped = x_fp8.reshape(-1, orig_shape[-1])
+        res_bits = torch.mm(x_fp8_reshaped, w_fp8_t)
+        res_bits = res_bits.reshape(*orig_shape[:-1], res_bits.shape[-1])
+        return res_bits
+
+    @staticmethod
+    def backward(ctx, go_fp8):
+        x_fp8, w_fp8_t = ctx.saved_tensors
+
+        # the reshapes are needed in order to make the shapes compatible with 
+        # torch.mm
+        go_fp8_orig_shape = go_fp8.shape
+        go_fp8_reshaped = go_fp8.reshape(-1, go_fp8_orig_shape[-1])
+
+        # calculate dL/dX
+        dL_dX = torch.mm(
+            go_fp8_reshaped,
+            w_fp8_t.t(),
+        )
+        dL_dX = dL_dX.reshape(*go_fp8_orig_shape[:-1], dL_dX.shape[-1])
+
+        x_fp8_orig_shape = x_fp8.shape
+        x_fp8_reshaped = x_fp8.reshape(-1, x_fp8_orig_shape[-1])
+
+        # calculate dL/dW
+        # Note: the variant below is slightly faster on LLaMa 3 8B pretraining
+        # compared to than calculating `dL_dW_t = x_fp8_t @ go_fp8_reshaped`
+        dL_dW = torch.mm(
+            go_fp8_reshaped.t(),
+            x_fp8_reshaped,
+        )
+
+        return dL_dX, dL_dW.t()
+
+
 @torch._dynamo.allow_in_graph
 class NoopFwToFloat8E5M2Bw(torch.autograd.Function):
     """
@@ -410,7 +462,7 @@ def forward(self, input: torch.Tensor) -> torch.Tensor:
         x_fp8 = self.cast_x_to_float8(input, self.is_amax_initialized)
         w_fp8 = self.cast_w_to_float8(self.weight, self.is_amax_initialized)
 
-        y = torch.matmul(x_fp8, w_fp8.t())
+        y = manual_float8_matmul.apply(x_fp8, w_fp8.t())
 
         # Cast gradY to float8_e5m2 during backward
         y = self.cast_y_to_float8_in_bw(y)

float8_experimental/float8_tensor.py

@@ -101,7 +101,9 @@ def choose_scaled_mm_config(
             a_linear_mm_config.dL_dX == b_linear_mm_config.dL_dX
         ), f"linear_mm_config.dL_dX mismatch: {a_linear_mm_config.dL_dX} vs {b_linear_mm_config.dL_dX}"
         return a_linear_mm_config.dL_dX
-    elif a_role is GemmInputRole.DL_DY and b_role is GemmInputRole.X:
+    elif (a_role is GemmInputRole.X and b_role is GemmInputRole.DL_DY) or (
+        a_role is GemmInputRole.DL_DY and b_role is GemmInputRole.X
+    ):
         assert (
             a_linear_mm_config.dL_dW == b_linear_mm_config.dL_dW
         ), f"linear_mm_config.dL_dW mismatch: {a_linear_mm_config.dL_dW} vs {b_linear_mm_config.dL_dW}"