[wip] hooks

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

Author: vkuzo

float8_experimental/config.py

@@ -16,3 +16,11 @@
 # according to their microbatching/pipeline parallel setup.
 # Note: this is currently a global flag for simplicity and dynamo performance.
 weight_cache_enabled = False
+
+#
+# Other
+#
+
+# If True, dynamic linear uses hooks for activation casting
+dynamic_use_activation_hooks = True
+# dynamic_use_activation_hooks = False

float8_experimental/dynamic_linear/dynamic_float8_linear.py

@@ -11,6 +11,7 @@
 
 from float8_experimental.float8_tensor import Float8Tensor
 from float8_experimental.float8_utils import tensor_to_scale, to_fp8_saturated
+import float8_experimental.config as config
 
 
 class NoopFwToFloat8E5M2Bw(torch.autograd.Function):
@@ -38,6 +39,22 @@ def backward(ctx, gradY):
             None,
         )
 
+def cast_x_to_float8_e4m3fn_pre_hook(module, args):
+    """
+    Hook to cast the incoming activation to `torch.float8_e4m3fn`
+    """
+    return module.cast_to_float8(args[0])
+
+def cast_dldy_to_float8_e5m2_backward_pre_hook(module, grad_output):
+    """
+    Hook to cast the incoming gradient to `torch.float8_e5m2`
+    """
+    gradY = grad_output[0]
+    gradY_scale = tensor_to_scale(gradY, torch.float8_e5m2)
+    gradY_scaled = gradY * gradY_scale
+    bits_fp8 = to_fp8_saturated(gradY_scaled, torch.float8_e5m2)
+    tensor_fp8 = Float8Tensor(bits_fp8, gradY_scale, gradY.dtype, emulate=module.emulate)
+    return (tensor_fp8,)
 
 class Float8DynamicLinear(torch.nn.Linear):
     """
@@ -48,9 +65,16 @@ class Float8DynamicLinear(torch.nn.Linear):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.add_weight_tag()
+        self.use_activation_hooks = config.dynamic_use_activation_hooks
 
     def forward(self, x):
-        x_fp8 = self.cast_to_float8(x)
+        # cast x to float8_e4m3fn
+        if self.use_activation_hooks:
+            x_fp8 = x
+        else:
+            x_fp8 = self.cast_to_float8(x)
+
+        # cast w to float8_e4m3fn
         if getattr(self, "_w_fp8", None) is not None:  # FSDP handled the cast
             w_fp8 = self._w_fp8
         else:
@@ -59,7 +83,10 @@ def forward(self, x):
         y = torch.nn.functional.linear(x_fp8, w_fp8, self.bias)
 
         # Cast gradY to float8_e5m2 during backward
-        y = self.cast_to_float8e5m2_bw(y)
+        if self.use_activation_hooks:
+            pass
+        else:
+            y = self.cast_to_float8e5m2_bw(y)
 
         return y
 
@@ -69,6 +96,7 @@ def add_weight_tag(self):
         self.weight._is_fp8_weight = True
 
     def cast_to_float8(self, inpt_tensor):
+        # TODO rename this function to clarify e4m3
         scale = tensor_to_scale(inpt_tensor, torch.float8_e4m3fn)
         return Float8Tensor.to_float8(
             inpt_tensor, scale, torch.float8_e4m3fn, emulate=self.emulate
@@ -92,4 +120,14 @@ def from_float(cls, mod, emulate: bool = False):
         new_mod.bias = mod.bias
         new_mod.emulate = emulate
         new_mod.add_weight_tag()
+
+        new_mod.use_activation_hooks = config.dynamic_use_activation_hooks
+        if new_mod.use_activation_hooks:
+            # install the hooks
+            # TODO(future): figure out why using backward pre-hooks does not 
+            # work here:
+            # 1. repro code: https://gist.github.com/vkuzo/27a3f6ca48e50ba1134b077f0dba254c
+            # 2. repro output: https://gist.github.com/vkuzo/728eae9dcc627e130829d122daa982e7
+            new_mod.register_forward_pre_hook(cast_x_to_float8_e4m3fn_pre_hook)
+            new_mod.register_full_backward_pre_hook(cast_dldy_to_float8_e5m2_backward_pre_hook)
         return new_mod