comfy/ops.py

@@ -449,7 +449,7 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            if len(self.weight_function) == 0 and len(self.bias_function) == 0:
+            if not self.training:
                 try:
                     out = fp8_linear(self, input)
                     if out is not None:

comfy/quant_ops.py

@@ -241,6 +241,9 @@ class QuantizedTensor(torch.Tensor):
     def storage(self):
         return self._qdata.storage()
 
+    def untyped_storage(self):
+        return self._qdata.untyped_storage()
+
 # ==============================================================================
 # Generic Utilities (Layout-Agnostic Operations)
 # ==============================================================================
@@ -252,6 +255,12 @@ def _create_transformed_qtensor(qt, transform_fn):
 
 
 def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"):
+    if target_dtype is not None and target_dtype != qt.dtype:
+        logging.warning(
+            f"QuantizedTensor: dtype conversion requested to {target_dtype}, "
+            f"but not supported for quantized tensors. Ignoring dtype."
+        )
+
     if target_layout is not None and target_layout != torch.strided:
         logging.warning(
             f"QuantizedTensor: layout change requested to {target_layout}, "
@@ -271,8 +280,6 @@ def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=
             logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}")
             new_q_data = qt._qdata.to(device=target_device)
             new_params = _move_layout_params_to_device(qt._layout_params, target_device)
-            if target_dtype is not None:
-                new_params["orig_dtype"] = target_dtype
             new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params)
             logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}")
             return new_qt
@@ -396,7 +403,7 @@ class TensorCoreFP8Layout(QuantizedLayout):
     def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False):
         orig_dtype = tensor.dtype
 
-        if isinstance(scale, str) and scale == "recalculate":
+        if scale == "recalculate":
             scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
 
         if scale is not None:

comfy/text_encoders/ovis.py

@@ -55,9 +55,12 @@ class OvisTEModel(sd1_clip.SD1ClipModel):
         return out, pooled, {}
 
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None):
     class OvisTEModel_(OvisTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None: