ops: sync the offload stream with the consumption of w&b

This sync is nessacary as pytorch will queue cuda async frees on the
same stream as created to tensor. In the case of async offload, this
will be on the offload stream.

Weights and biases can go out of scope in python which then
triggers the pytorch garbage collector to queue the free operation on
the offload stream possible before the compute stream has used the
weight. This causes a use after free on weight data leading to total
corruption of some workflows.

So sync the offload stream with the compute stream after the weight
has been used so the free has to wait for the weight to be used.

The cast_bias_weight is extended in a backwards compatible way with
the new behaviour opt-in on a defaulted parameter. This handles
custom node packs calling cast_bias_weight and defeatures
async-offload for them (as they do not handle the race).

The pattern is now:

cast_bias_weight(... , offloadable=True) #This might be offloaded
thing(weight, bias, ...)
uncast_bias_weight(...)

Author: rattus128

comfy/ops.py

@@ -70,8 +70,12 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
 def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
+
 @torch.compiler.disable()
-def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False):
+    # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass
+    # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This
+    # will add async-offload support to your cast and improve performance.
     if input is not None:
         if dtype is None:
             dtype = input.dtype
@@ -80,7 +84,11 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
         if device is None:
             device = input.device
 
-    offload_stream = comfy.model_management.get_offload_stream(device)
+    if offloadable:
+        offload_stream = comfy.model_management.get_offload_stream(device)
+    else:
+        offload_stream = None
+
     if offload_stream is not None:
         wf_context = offload_stream
     else:
@@ -105,7 +113,24 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
                 weight = f(weight)
 
     comfy.model_management.sync_stream(device, offload_stream)
-    return weight, bias
+    if offloadable:
+        return weight, bias, offload_stream
+    else:
+        #Legacy function signature
+        return weight, bias
+
+
+def uncast_bias_weight(s, weight, bias, offload_stream):
+    if offload_stream is None:
+        return
+    if weight is not None:
+        device = weight.device
+    else:
+        if bias is None:
+            return
+        device = bias.device
+    offload_stream.wait_stream(comfy.model_management.current_stream(device))
+
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
@@ -118,8 +143,10 @@ def reset_parameters(self):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.linear(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -133,8 +160,10 @@ def reset_parameters(self):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -148,8 +177,10 @@ def reset_parameters(self):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -172,8 +203,10 @@ def _conv_forward(self, input, weight, bias, *args, **kwargs):
                 return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return self._conv_forward(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = self._conv_forward(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -187,8 +220,10 @@ def reset_parameters(self):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -203,11 +238,14 @@ def reset_parameters(self):
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
                 bias = None
-            return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+                offload_stream = None
+            x = torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -223,11 +261,15 @@ def reset_parameters(self):
 
         def forward_comfy_cast_weights(self, input):
             if self.weight is not None:
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
             else:
                 weight = None
-            return comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
-            # return torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+                bias = None
+                offload_stream = None
+            x = comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
+            # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -246,10 +288,12 @@ def forward_comfy_cast_weights(self, input, output_size=None):
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose2d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.conv_transpose2d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -268,10 +312,12 @@ def forward_comfy_cast_weights(self, input, output_size=None):
                 input, output_size, self.stride, self.padding, self.kernel_size,
                 num_spatial_dims, self.dilation)
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.conv_transpose1d(
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.conv_transpose1d(
                 input, weight, bias, self.stride, self.padding,
                 output_padding, self.groups, self.dilation)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -289,8 +335,11 @@ def forward_comfy_cast_weights(self, input, out_dtype=None):
             output_dtype = out_dtype
             if self.weight.dtype == torch.float16 or self.weight.dtype == torch.bfloat16:
                 out_dtype = None
-            weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype)
-            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
+            weight, bias, offload_stream = cast_bias_weight(self, device=input.device, dtype=out_dtype, offloadable=True)
+            x = torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
+
 
         def forward(self, *args, **kwargs):
             run_every_op()
@@ -356,7 +405,7 @@ def fp8_linear(self, input):
     input_shape = input.shape
     input_dtype = input.dtype
     if len(input.shape) == 3:
-        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype)
+        w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True)
         w = w.t()
 
         scale_weight = self.scale_weight
@@ -379,6 +428,8 @@ def fp8_linear(self, input):
         else:
             o = torch._scaled_mm(input, w, out_dtype=input_dtype, scale_a=scale_input, scale_b=scale_weight)
 
+        uncast_bias_weight(self, w, bias, offload_stream)
+
         if isinstance(o, tuple):
             o = o[0]
 
@@ -405,8 +456,10 @@ def forward_comfy_cast_weights(self, input):
                 except Exception as e:
                     logging.info("Exception during fp8 op: {}".format(e))
 
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.linear(input, weight, bias)
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            x = torch.nn.functional.linear(input, weight, bias)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
 
 def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
     logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
@@ -434,12 +487,14 @@ def forward_comfy_cast_weights(self, input):
                     if out is not None:
                         return out
 
-                weight, bias = cast_bias_weight(self, input)
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
 
                 if weight.numel() < input.numel(): #TODO: optimize
-                    return torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
+                    x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias)
                 else:
-                    return torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                    x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias)
+                uncast_bias_weight(self, weight, bias, offload_stream)
+                return x
 
             def convert_weight(self, weight, inplace=False, **kwargs):
                 if inplace: