offload support, bug fixes, remove mixins

comfy/ops.py

@@ -77,6 +77,9 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     # will add async-offload support to your cast and improve performance.
     if input is not None:
         if dtype is None:
+            if isinstance(input, QuantizedTensor):
+                dtype = input._layout_params["orig_dtype"]
+            else:
                 dtype = input.dtype
         if bias_dtype is None:
             bias_dtype = dtype
@@ -534,18 +537,7 @@ if CUBLAS_IS_AVAILABLE:
 # ==============================================================================
 # Mixed Precision Operations
 # ==============================================================================
-from .quant_ops import QuantizedTensor
+from .quant_ops import QuantizedTensor, QUANT_ALGOS
-
-QUANT_FORMAT_MIXINS = {
-    "float8_e4m3fn": {
-        "dtype": torch.float8_e4m3fn,
-        "layout_type": "TensorCoreFP8Layout",
-        "parameters": {
-            "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
-            "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False),
-        }
-    }
-}
 
 class MixedPrecisionOps(disable_weight_init):
     _layer_quant_config = {}
@@ -596,23 +588,24 @@ class MixedPrecisionOps(disable_weight_init):
                 if quant_format is None:
                     raise ValueError(f"Unknown quantization format for layer {layer_name}")
 
-                mixin = QUANT_FORMAT_MIXINS[quant_format]
+                qconfig = QUANT_ALGOS[quant_format]
-                self.layout_type = mixin["layout_type"]
+                self.layout_type = qconfig["comfy_tensor_layout"]
 
-                scale_key = f"{prefix}weight_scale"
+                weight_scale_key = f"{prefix}weight_scale"
                 layout_params = {
-                    'scale': state_dict.pop(scale_key, None),
+                    'scale': state_dict.pop(weight_scale_key, None),
-                    'orig_dtype': MixedPrecisionOps._compute_dtype
+                    'orig_dtype': MixedPrecisionOps._compute_dtype,
+                    'block_size': qconfig.get("group_size", None),
                 }
                 if layout_params['scale'] is not None:
-                    manually_loaded_keys.append(scale_key)
+                    manually_loaded_keys.append(weight_scale_key)
 
                 self.weight = torch.nn.Parameter(
-                    QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params),
+                    QuantizedTensor(weight.to(device=device), self.layout_type, layout_params),
                     requires_grad=False
                 )
 
-                for param_name, param_value in mixin["parameters"].items():
+                for param_name in qconfig["parameters"]:
                     param_key = f"{prefix}{param_name}"
                     _v = state_dict.pop(param_key, None)
                     if _v is None:
@@ -643,7 +636,7 @@ class MixedPrecisionOps(disable_weight_init):
             if (getattr(self, 'layout_type', None) is not None and
                 getattr(self, 'input_scale', None) is not None and
                 not isinstance(input, QuantizedTensor)):
-                input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype)
+                input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype)
             return self._forward(input, self.weight, self.bias)
 
 

comfy/quant_ops.py

@@ -74,6 +74,12 @@ def _copy_layout_params(params):
             new_params[k] = v
     return new_params
 
+def _copy_layout_params_inplace(src, dst, non_blocking=False):
+    for k, v in src.items():
+        if isinstance(v, torch.Tensor):
+            dst[k].copy_(v, non_blocking=non_blocking)
+        else:
+            dst[k] = v
 
 class QuantizedLayout:
     """
@@ -318,13 +324,13 @@ def generic_to_dtype_layout(func, args, kwargs):
 def generic_copy_(func, args, kwargs):
     qt_dest = args[0]
     src = args[1]
-
+    non_blocking = args[2] if len(args) > 2 else False
     if isinstance(qt_dest, QuantizedTensor):
         if isinstance(src, QuantizedTensor):
             # Copy from another quantized tensor
-            qt_dest._qdata.copy_(src._qdata)
+            qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking)
             qt_dest._layout_type = src._layout_type
-            qt_dest._layout_params = _copy_layout_params(src._layout_params)
+            _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking)
         else:
             # Copy from regular tensor - just copy raw data
             qt_dest._qdata.copy_(src)
@@ -336,6 +342,26 @@ def generic_copy_(func, args, kwargs):
 def generic_has_compatible_shallow_copy_type(func, args, kwargs):
     return True
 
+
+@register_generic_util(torch.ops.aten.empty_like.default)
+def generic_empty_like(func, args, kwargs):
+    """Empty_like operation - creates an empty tensor with the same quantized structure."""
+    qt = args[0]
+    if isinstance(qt, QuantizedTensor):
+        # Create empty tensor with same shape and dtype as the quantized data
+        hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"])
+        new_qdata = torch.empty_like(qt._qdata, **kwargs)
+
+        # Handle device transfer for layout params
+        target_device = kwargs.get('device', new_qdata.device)
+        new_params = _move_layout_params_to_device(qt._layout_params, target_device)
+
+        # Update orig_dtype if dtype is specified
+        new_params['orig_dtype'] = hp_dtype
+
+        return QuantizedTensor(new_qdata, qt._layout_type, new_params)
+    return func(*args, **kwargs)
+
 # ==============================================================================
 # FP8 Layout + Operation Handlers
 # ==============================================================================
@@ -378,6 +404,13 @@ class TensorCoreFP8Layout(QuantizedLayout):
     def get_plain_tensors(cls, qtensor):
         return qtensor._qdata, qtensor._layout_params['scale']
 
+QUANT_ALGOS = {
+    "float8_e4m3fn": {
+        "storage_t": torch.float8_e4m3fn,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreFP8Layout",
+    },
+}
 
 LAYOUTS = {
     "TensorCoreFP8Layout": TensorCoreFP8Layout,