Lower T5 memory usage by a few hundred MB.

Author: comfyanonymous

comfy/ldm/hydit/models.py

@@ -355,7 +355,7 @@ def forward(self,
         if self.use_style_cond:
             if style is None:
                 style = torch.zeros((extra_vec.shape[0],), device=x.device, dtype=torch.int)
-            style_embedding = self.style_embedder(style)
+            style_embedding = self.style_embedder(style, out_dtype=x.dtype)
             extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
 
         # Concatenate all extra vectors

comfy/ops.py

@@ -19,17 +19,27 @@
 import torch
 import comfy.model_management
 
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False):
+    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+
 def cast_to_input(weight, input, non_blocking=False):
-    return weight.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking)
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+
+def cast_bias_weight(s, input=None, dtype=None, device=None):
+    if input is not None:
+        if dtype is None:
+            dtype = input.dtype
+        if device is None:
+            device = input.device
 
-def cast_bias_weight(s, input):
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(input.device)
+    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to_input(s.bias, input, non_blocking=non_blocking)
+        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
         if s.bias_function is not None:
             bias = s.bias_function(bias)
-    weight = cast_to_input(s.weight, input, non_blocking=non_blocking)
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
     if s.weight_function is not None:
         weight = s.weight_function(weight)
     return weight, bias
@@ -176,14 +186,19 @@ def reset_parameters(self):
             self.bias = None
             return None
 
-        def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse)
+        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)
 
         def forward(self, *args, **kwargs):
             if self.comfy_cast_weights:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
+                if "out_dtype" in kwargs:
+                    kwargs.pop("out_dtype")
                 return super().forward(*args, **kwargs)
 
     @classmethod

comfy/text_encoders/t5.py

@@ -1,6 +1,7 @@
 import torch
 import math
 from comfy.ldm.modules.attention import optimized_attention_for_device
+import comfy.ops
 
 class T5LayerNorm(torch.nn.Module):
     def __init__(self, hidden_size, eps=1e-6, dtype=None, device=None, operations=None):
@@ -11,7 +12,7 @@ def __init__(self, hidden_size, eps=1e-6, dtype=None, device=None, operations=No
     def forward(self, x):
         variance = x.pow(2).mean(-1, keepdim=True)
         x = x * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight.to(device=x.device, dtype=x.dtype) * x
+        return comfy.ops.cast_to_input(self.weight, x) * x
 
 activations = {
     "gelu_pytorch_tanh": lambda a: torch.nn.functional.gelu(a, approximate="tanh"),
@@ -82,7 +83,7 @@ def __init__(self, model_dim, inner_dim, num_heads, relative_attention_bias, dty
         if relative_attention_bias:
             self.relative_attention_num_buckets = 32
             self.relative_attention_max_distance = 128
-            self.relative_attention_bias = torch.nn.Embedding(self.relative_attention_num_buckets, self.num_heads, device=device)
+            self.relative_attention_bias = operations.Embedding(self.relative_attention_num_buckets, self.num_heads, device=device, dtype=dtype)
 
     @staticmethod
     def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
@@ -132,7 +133,7 @@ def _relative_position_bucket(relative_position, bidirectional=True, num_buckets
         relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
         return relative_buckets
 
-    def compute_bias(self, query_length, key_length, device):
+    def compute_bias(self, query_length, key_length, device, dtype):
         """Compute binned relative position bias"""
         context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
         memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
@@ -143,7 +144,7 @@ def compute_bias(self, query_length, key_length, device):
             num_buckets=self.relative_attention_num_buckets,
             max_distance=self.relative_attention_max_distance,
         )
-        values = self.relative_attention_bias(relative_position_bucket)  # shape (query_length, key_length, num_heads)
+        values = self.relative_attention_bias(relative_position_bucket, out_dtype=dtype)  # shape (query_length, key_length, num_heads)
         values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
         return values
 
@@ -152,7 +153,7 @@ def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
         k = self.k(x)
         v = self.v(x)
         if self.relative_attention_bias is not None:
-            past_bias = self.compute_bias(x.shape[1], x.shape[1], x.device)
+            past_bias = self.compute_bias(x.shape[1], x.shape[1], x.device, x.dtype)
 
         if past_bias is not None:
             if mask is not None:
@@ -225,7 +226,7 @@ def __init__(self, config_dict, dtype, device, operations):
 
         self.encoder = T5Stack(self.num_layers, model_dim, model_dim, config_dict["d_ff"], config_dict["dense_act_fn"], config_dict["is_gated_act"], config_dict["num_heads"], config_dict["model_type"] != "umt5", dtype, device, operations)
         self.dtype = dtype
-        self.shared = torch.nn.Embedding(config_dict["vocab_size"], model_dim, device=device)
+        self.shared = operations.Embedding(config_dict["vocab_size"], model_dim, device=device, dtype=dtype)
 
     def get_input_embeddings(self):
         return self.shared
@@ -234,5 +235,5 @@ def set_input_embeddings(self, embeddings):
         self.shared = embeddings
 
     def forward(self, input_ids, *args, **kwargs):
-        x = self.shared(input_ids)
+        x = self.shared(input_ids, out_dtype=kwargs.get("dtype", torch.float32))
         return self.encoder(x, *args, **kwargs)