From 6b778afdc938427116cb72e683d28e42baf67082 Mon Sep 17 00:00:00 2001
From: blepping <157360029+blepping@users.noreply.github.com>
Date: Mon, 8 Sep 2025 23:07:45 -0600
Subject: [PATCH 1/3] Fix attempt to create a tensor with dtype "target" (#330)

* Fix attempt to create a tensor with dtype "target"

* Check if dequant_dtype is None before string comparison with target
---
 ops.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ops.py b/ops.py
index 41d42f5..ff9b960 100644
--- a/ops.py
+++ b/ops.py
@@ -153,7 +153,7 @@ class GGMLLayer(torch.nn.Module):
         # Take into account space required for dequantizing the largest tensor
         if self.largest_layer:
             shape = getattr(self.weight, "tensor_shape", self.weight.shape)
-            dtype = self.dequant_dtype or torch.float16
+            dtype = (self.dequant_dtype and self.dequant_dtype != "target") or torch.float16
             temp = torch.empty(*shape, device=torch.device("meta"), dtype=dtype)
             destination[prefix + "temp.weight"] = temp
 

From be2a08330d7ec232d684e50ab938870d7529471e Mon Sep 17 00:00:00 2001
From: blepping <157360029+blepping@users.noreply.github.com>
Date: Sun, 14 Sep 2025 15:10:50 -0600
Subject: [PATCH 2/3] Fix dequant_dtype handling (#334)

* Fix dequant_dtype handling

* Simplify dequant dtype selection expression
---
 ops.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ops.py b/ops.py
index ff9b960..88a352e 100644
--- a/ops.py
+++ b/ops.py
@@ -153,7 +153,7 @@ class GGMLLayer(torch.nn.Module):
         # Take into account space required for dequantizing the largest tensor
         if self.largest_layer:
             shape = getattr(self.weight, "tensor_shape", self.weight.shape)
-            dtype = (self.dequant_dtype and self.dequant_dtype != "target") or torch.float16
+            dtype = self.dequant_dtype if self.dequant_dtype and self.dequant_dtype != "target" else torch.float16
             temp = torch.empty(*shape, device=torch.device("meta"), dtype=dtype)
             destination[prefix + "temp.weight"] = temp
 

From 100c06cda840b0406f96fd7c3c57076d443f82b6 Mon Sep 17 00:00:00 2001
From: rattus <46076784+rattus128@users.noreply.github.com>
Date: Tue, 4 Nov 2025 00:57:29 +1000
Subject: [PATCH 3/3] nodes: mmap detach weights before they are pinned (#355)

Comfy core recently introduced a feature where weights may be pinned
when loading, particularly for the case of offloading.

Intercept this, and immediately detached each weight before the pinning.
This avoids a crash that at least some users are experiencing.

Use a little dict on the modules to keep track of whats already done,
and when the catch-all detacher loop comes through, use this dict
(which has already done modules removed) as the iterator basis.
---
 nodes.py | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/nodes.py b/nodes.py
index 4159142..08849df 100644
--- a/nodes.py
+++ b/nodes.py
@@ -76,8 +76,21 @@ class GGUFModelPatcher(comfy.model_patcher.ModelPatcher):
         # TODO: Find another way to not unload after patches
         return super().unpatch_model(device_to=device_to, unpatch_weights=unpatch_weights)
 
+
+    def pin_weight_to_device(self, key):
+        op_key = key.rsplit('.', 1)[0]
+        if self.named_modules_to_munmap is not None and op_key in self.named_modules_to_munmap:
+            # TODO: possible to OOM, find better way to detach
+            self.named_modules_to_munmap[op_key].to(self.load_device).to(self.offload_device)
+            del self.named_modules_to_munmap[op_key]
+        super().pin_weight_to_device(key)
+
     mmap_released = False
+
     def load(self, *args, force_patch_weights=False, **kwargs):
+        if not self.mmap_released:
+            self.named_modules_to_munmap = dict(self.model.named_modules())
+
         # always call `patch_weight_to_device` even for lowvram
         super().load(*args, force_patch_weights=True, **kwargs)
 
@@ -85,7 +98,7 @@ class GGUFModelPatcher(comfy.model_patcher.ModelPatcher):
         if not self.mmap_released:
             linked = []
             if kwargs.get("lowvram_model_memory", 0) > 0:
-                for n, m in self.model.named_modules():
+                for n, m in self.named_modules_to_munmap.items():
                     if hasattr(m, "weight"):
                         device = getattr(m.weight, "device", None)
                         if device == self.offload_device:
@@ -102,6 +115,7 @@ class GGUFModelPatcher(comfy.model_patcher.ModelPatcher):
                     # TODO: possible to OOM, find better way to detach
                     m.to(self.load_device).to(self.offload_device)
             self.mmap_released = True
+        self.named_modules_to_munmap = None
 
     def clone(self, *args, **kwargs):
         src_cls = self.__class__