diff --git a/README.md b/README.md
index 8142f595bcd0..f51807ad5df1 100644
--- a/README.md
+++ b/README.md
@@ -200,7 +200,7 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended.
+Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies.
 
 Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
 
@@ -242,7 +242,7 @@ RDNA 4 (RX 9000 series):
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
 
 1. To install PyTorch xpu, use the following command:
 
@@ -252,10 +252,6 @@ This is the command to install the Pytorch xpu nightly which might have some per
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
-
-1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py
index e5d0d17c1d6e..427ea19c1179 100644
--- a/comfy/ldm/qwen_image/model.py
+++ b/comfy/ldm/qwen_image/model.py
@@ -236,10 +236,10 @@ def forward(
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_normed = self.img_norm1(hidden_states)
-        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
-        txt_normed = self.txt_norm1(encoder_hidden_states)
-        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1)
+        del img_mod1
+        txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
+        del txt_mod1
 
         img_attn_output, txt_attn_output = self.attn(
             hidden_states=img_modulated,
@@ -248,16 +248,20 @@ def forward(
             image_rotary_emb=image_rotary_emb,
             transformer_options=transformer_options,
         )
+        del img_modulated
+        del txt_modulated
 
         hidden_states = hidden_states + img_gate1 * img_attn_output
         encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+        del img_attn_output
+        del txt_attn_output
+        del img_gate1
+        del txt_gate1
 
-        img_normed2 = self.img_norm2(hidden_states)
-        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
+        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2)
         hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
 
-        txt_normed2 = self.txt_norm2(encoder_hidden_states)
-        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
 
         return encoder_hidden_states, hidden_states
diff --git a/comfy/model_management.py b/comfy/model_management.py
index d8913082adaa..a21df54b3f38 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -503,10 +503,7 @@ def model_load(self, lowvram_model_memory=0, force_patch_weights=False):
         use_more_vram = lowvram_model_memory
         if use_more_vram == 0:
             use_more_vram = 1e32
-        if use_more_vram > 0:
-            self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
-        else:
-            self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights)
+        self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
 
         real_model = self.model.model
 
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
index 68b0a9192753..cf1b0d4412bc 100644
--- a/comfy/model_patcher.py
+++ b/comfy/model_patcher.py
@@ -928,6 +928,9 @@ def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
                 extra_memory += (used - self.model.model_loaded_weight_memory)
 
             self.patch_model(load_weights=False)
+            if extra_memory < 0 and not unpatch_weights:
+                self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights)
+                return 0
             full_load = False
             if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0:
                 self.apply_hooks(self.forced_hooks, force_apply=True)