Fix graph breaks in tensor subclass (#131)

Summary:
For more detailed understanding of status see:
#106

But this removes all graph breaks on the main work branch

Pull Request resolved: #131

Reviewed By: albanD

Differential Revision: D50758815

Pulled By: drisspg

fbshipit-source-id: 1502601099988b1eba666306e327eb724eb14989

float8_experimental/__init__.py

@@ -1,4 +1,5 @@
-# Lets wait to define the top level interface
-# from float8_experimental.float8_tensor import Float8Tensor
+# Lets define a few top level things here
+from float8_experimental.float8_tensor import Float8Tensor
+from float8_experimental.float8_linear import Float8Linear
 
-# __all__ = ["Float8Tensor"]
+__all__ = ["Float8Tensor", "Float8Linear"]

float8_experimental/float8_linear.py

@@ -18,7 +18,7 @@
 )
 
 from float8_experimental.float8_python_api import mm_float8
-from float8_experimental.float8_tensor import Float8Tensor
+from float8_experimental.float8_tensor import Float8Tensor, to_float8
 
 from float8_experimental.float8_utils import (
     amax_history_to_scale,
@@ -174,9 +174,7 @@ class DelayedScalingRecipe:
 
 class Float8LinearMixin(object):
     def __init__(self, *args, **kwargs):
-        delayed_scaling_recipe = kwargs.pop(
-            "delayed_scaling_recipe", DelayedScalingRecipe()
-        )
+        delayed_scaling_recipe = kwargs.pop("delayed_scaling_recipe", DelayedScalingRecipe())
         super().__init__(*args, **kwargs)
 
         # TODO(future): have a unique recipe per buffer instead of one per
@@ -239,10 +237,7 @@ def cast_x_to_float8(self, x, is_amax_initialized):
             torch.float8_e4m3fn,
             is_amax_initialized,
         )
-        x_fp8 = Float8Tensor.to_float8(
-            x, self.fp8_scale_x, torch.float8_e4m3fn, self.fp8_amax_x
-        )
-
+        x_fp8 = to_float8(x, self.fp8_scale_x, torch.float8_e4m3fn, self.fp8_amax_x)
         return x_fp8
 
     def cast_w_to_float8(self, w, is_amax_initialized):
@@ -256,9 +251,7 @@ def cast_w_to_float8(self, w, is_amax_initialized):
             torch.float8_e4m3fn,
             is_amax_initialized,
         )
-        w_fp8 = Float8Tensor.to_float8(
-            w, self.fp8_scale_w, torch.float8_e4m3fn, self.fp8_amax_w
-        )
+        w_fp8 = to_float8(w, self.fp8_scale_w, torch.float8_e4m3fn, self.fp8_amax_w)
         return w_fp8
 
     def cast_y_to_float8_in_bw(self, y):
@@ -275,9 +268,7 @@ def cast_y_to_float8_in_bw(self, y):
 
     def float8_mm(self, x_fp8, w_fp8, is_amax_initialized):
         scale_fn_name = self.recipe.scale_fn_name
-        y = float8_linear.apply(
-            x_fp8, w_fp8, is_amax_initialized, scale_fn_name, self.emulate
-        )
+        y = float8_linear.apply(x_fp8, w_fp8, is_amax_initialized, scale_fn_name, self.emulate)
         return y
 
     def float8_pre_forward(self, x):
@@ -416,9 +407,7 @@ def sync_float8_amax_and_scale_history(model: torch.nn.Module) -> None:
         #
         _update_history_with_new_amax(child.fp8_amax_x, child.fp8_amax_history_x)
         _update_history_with_new_amax(child.fp8_amax_w, child.fp8_amax_history_w)
-        _update_history_with_new_amax(
-            child.fp8_amax_dL_dY, child.fp8_amax_history_dL_dY
-        )
+        _update_history_with_new_amax(child.fp8_amax_dL_dY, child.fp8_amax_history_dL_dY)
 
         #
         # 3. calculate the scales

float8_experimental/float8_tensor.py

@@ -75,6 +75,20 @@ def backward(ctx, g):
             return g, None, None, None
 
 
+def to_float8(tensor: torch.Tensor, scale: torch.Tensor, float8_dtype: torch.dtype, amax_buffer:torch.Tensor =None) -> "Float8Tensor":
+    """ Converts a higher precision tensor to float8 in a differentiable way.
+
+    Args:
+        tensor: the tensor to convert
+        scale: the scale to use to convert the tensor
+        float8_dtype: the float8 dtype to use
+        amax_buffer: a buffer to store the amax value in prior to conversion
+
+    Returns:
+        Float8Tensor: a float8 tensor
+    """
+    return ToFloat8ConstrFunc.apply(tensor, scale, float8_dtype, amax_buffer)
+
 class FromFloat8ConstrFunc(torch.autograd.Function):
     """
     A differentiable conversion from fp8
@@ -86,7 +100,7 @@ def forward(ctx, tensor):
 
     @staticmethod
     def backward(ctx, g):
-        return Float8Tensor.to_float8(g), None, None
+        return to_float8(g), None, None
 
 
 class Float8Tensor(torch.Tensor):
@@ -154,10 +168,6 @@ def __tensor_unflatten__(inner_tensors: Dict, metadata):
     def to_original_precision(self):
         return FromFloat8ConstrFunc.apply(self)
 
-    @classmethod
-    def to_float8(cls, tensor, scale, float8_dtype, amax_buffer=None):
-        return ToFloat8ConstrFunc.apply(tensor, scale, float8_dtype, amax_buffer)
-
     @classmethod
     def __torch_dispatch__(cls, func, types, args, kwargs=None):
         # 1. tracing through __torch_function__ logic is not supported yet in

test/test_base.py

@@ -15,7 +15,7 @@
 )
 from float8_experimental.float8_linear_nots import Float8LinearNoTensorSubclass
 from float8_experimental.float8_python_api import mm_float8
-from float8_experimental.float8_tensor import Float8Tensor
+from float8_experimental.float8_tensor import Float8Tensor, to_float8
 
 from float8_experimental.float8_utils import (
     amax_to_scale,
@@ -39,7 +39,7 @@ def test_preserves_dtype(self):
         for hp_dtype, lp_dtype in itertools.product(hp_dtypes, lp_dtypes):
             x1_hp = torch.randn(4, 4, dtype=hp_dtype)
             x1_s = tensor_to_scale(x1_hp, lp_dtype)
-            x2_lp = Float8Tensor.to_float8(x1_hp, x1_s, lp_dtype)
+            x2_lp = to_float8(x1_hp, x1_s, lp_dtype)
             x3_hp = x2_lp.to_original_precision()
             self.assertTrue(x3_hp.dtype == hp_dtype)
 
@@ -248,8 +248,8 @@ def test_scaled_mm_vs_emulated(self, base_dtype):
         a_scale = tensor_to_scale(a, input_dtype).float()
         b_scale = tensor_to_scale(b, input_dtype).float()
 
-        a_fp8 = Float8Tensor.to_float8(a, a_scale, input_dtype)
-        b_fp8 = Float8Tensor.to_float8(b, b_scale, input_dtype)
+        a_fp8 = to_float8(a, a_scale, input_dtype)
+        b_fp8 = to_float8(b, b_scale, input_dtype)
 
         out_scaled_mm, output_amax_scaled = mm_float8(
             a_fp8, b_fp8, output_dtype=output_dtype, emulate=False