sync with repo 28.08

comfy/ops.py

@@ -18,29 +18,42 @@
 
 import torch
 import comfy.model_management
+from comfy.cli_args import args
 
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+    if device is None or weight.device == device:
+        if not copy:
+            if dtype is None or weight.dtype == dtype:
+                return weight
+        return weight.to(dtype=dtype, copy=copy)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False):
-    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+    r = torch.empty_like(weight, dtype=dtype, device=device)
+    r.copy_(weight, non_blocking=non_blocking)
+    return r
 
-def cast_to_input(weight, input, non_blocking=False):
-    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+def cast_to_input(weight, input, non_blocking=False, copy=True):
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
-def cast_bias_weight(s, input=None, dtype=None, device=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     if input is not None:
         if dtype is None:
             dtype = input.dtype
+        if bias_dtype is None:
+            bias_dtype = dtype
         if device is None:
             device = input.device
 
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
-        if s.bias_function is not None:
+        has_function = s.bias_function is not None
+        bias = cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
+        if has_function:
             bias = s.bias_function(bias)
-    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
-    if s.weight_function is not None:
+
+    has_function = s.weight_function is not None
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
+    if has_function:
         weight = s.weight_function(weight)
     return weight, bias
 
@@ -238,3 +251,59 @@ class manual_cast(disable_weight_init):
 
     class Embedding(disable_weight_init.Embedding):
         comfy_cast_weights = True
+
+
+def fp8_linear(self, input):
+    dtype = self.weight.dtype
+    if dtype not in [torch.float8_e4m3fn]:
+        return None
+
+    if len(input.shape) == 3:
+        inn = input.reshape(-1, input.shape[2]).to(dtype)
+        non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
+        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input.dtype)
+        w = w.t()
+
+        scale_weight = self.scale_weight
+        scale_input = self.scale_input
+        if scale_weight is None:
+            scale_weight = torch.ones((1), device=input.device, dtype=torch.float32)
+            if scale_input is None:
+                scale_input = scale_weight
+        if scale_input is None:
+            scale_input = torch.ones((1), device=input.device, dtype=torch.float32)
+
+        if bias is not None:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
+
+        if isinstance(o, tuple):
+            o = o[0]
+
+        return o.reshape((-1, input.shape[1], self.weight.shape[0]))
+    return None
+
+class fp8_ops(manual_cast):
+    class Linear(manual_cast.Linear):
+        def reset_parameters(self):
+            self.scale_weight = None
+            self.scale_input = None
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            out = fp8_linear(self, input)
+            if out is not None:
+                return out
+
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+
+def pick_operations(weight_dtype, compute_dtype, load_device=None):
+    if compute_dtype is None or weight_dtype == compute_dtype:
+        return disable_weight_init
+    if args.fast:
+        if comfy.model_management.supports_fp8_compute(load_device):
+            return fp8_ops
+    return manual_cast