diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py
index a68cb8439..37b2a8ee3 100644
--- a/comfy/ldm/modules/attention.py
+++ b/comfy/ldm/modules/attention.py
@@ -336,6 +336,10 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
     steps = 1
 
 
+    if mem_free_total <= 0:
+        # DirectML doesn't expose free VRAM — assume 4GB free as a safe fallback for 6GB cards
+        mem_free_total = 4 * (1024 ** 3)
+
     if mem_required > mem_free_total:
         steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
         # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 6b4d4b770..a14627842 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -544,7 +544,11 @@ def module_mmap_residency(module, free=False):
     for k in sd:
         t = sd[k]
         module_mem += t.nbytes
-        storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
+        try:
+            storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
+        except NotImplementedError:
+            # DirectML (AMD) tensors are opaque — no host storage to inspect; skip mmap tracking
+            continue
         if not getattr(storage, "_comfy_tensor_mmap_touched", False):
             continue
         mmap_touched_mem += t.nbytes
@@ -1328,7 +1332,12 @@ def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
                 continue
             if comfy.memory_management.read_tensor_file_slice_into(tensor, dest_view):
                 continue
-            storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
+            try:
+                storage = tensor._qdata.untyped_storage() if isinstance(tensor, comfy.quant_ops.QuantizedTensor) else tensor.untyped_storage()
+            except NotImplementedError:
+                # DirectML tensors are opaque — skip mmap marking, just copy
+                dest_view.copy_(tensor, non_blocking=non_blocking)
+                continue
             if hasattr(storage, "_comfy_tensor_mmap_touched"):
                 storage._comfy_tensor_mmap_touched = True
             dest_view.copy_(tensor, non_blocking=non_blocking)