mm: redefine free memory for Windows

As commented.

comfy/model_management.py

@@ -27,6 +27,7 @@ import weakref
 import gc
 import os
 from contextlib import nullcontext
+import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
 
@@ -584,9 +585,15 @@ WINDOWS = any(platform.win32_ver())
 
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
+    import comfy.windows
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
     if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
         EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
+    def get_free_ram():
+        return comfy.windows.get_free_ram()
+else:
+    def get_free_ram():
+        return psutil.virtual_memory().available
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -617,7 +624,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
         ram_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
             memory_to_free = memory_required - get_free_memory(device)
-            ram_to_free = ram_required - psutil.virtual_memory().available
+            ram_to_free = ram_required - get_free_ram()
 
         if current_loaded_models[i].model.is_dynamic() and for_dynamic:
             #don't actually unload dynamic models for the sake of other dynamic models
@@ -822,7 +829,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
         return torch_dev
     else:
         return cpu_dev

comfy/ops.py

@@ -131,11 +131,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             xfer_dest = cast_dest
 
     pin = None
-    if signature is not None:
-        #If we are able to increase our load level (e.g. user reduces resolution or batch number)
-        #reclaim the pin previously used for offload.
-        comfy.pinned_memory.unpin_memory(s)
-    elif not resident:
+    if signature is None and not resident:
         #prepare a new pin
         assert comfy.pinned_memory.get_pin(s) is None
         comfy.pinned_memory.pin_memory(s)

comfy/windows.py

@@ -0,0 +1,52 @@
+import ctypes
+import logging
+import psutil
+from ctypes import wintypes
+
+import comfy_aimdo.control
+
+psapi = ctypes.WinDLL("psapi")
+kernel32 = ctypes.WinDLL("kernel32")
+
+class PERFORMANCE_INFORMATION(ctypes.Structure):
+    _fields_ = [
+        ("cb", wintypes.DWORD),
+        ("CommitTotal", ctypes.c_size_t),
+        ("CommitLimit", ctypes.c_size_t),
+        ("CommitPeak", ctypes.c_size_t),
+        ("PhysicalTotal", ctypes.c_size_t),
+        ("PhysicalAvailable", ctypes.c_size_t),
+        ("SystemCache", ctypes.c_size_t),
+        ("KernelTotal", ctypes.c_size_t),
+        ("KernelPaged", ctypes.c_size_t),
+        ("KernelNonpaged", ctypes.c_size_t),
+        ("PageSize", ctypes.c_size_t),
+        ("HandleCount", wintypes.DWORD),
+        ("ProcessCount", wintypes.DWORD),
+        ("ThreadCount", wintypes.DWORD),
+    ]
+
+def get_free_ram():
+    #Windows is way too conservative and chalks recently used uncommitted model RAM
+    #as "in-use". So, calculate free RAM for the sake of general use as the greater of:
+    #
+    #1: What psutil says
+    #2: Total Memory - (Committed Memory - VRAM in use)
+    #
+    #We have to subtract VRAM in use from the comitted memory as WDDM creates a naked
+    #commit charge for all VRAM used just incase it wants to page it all out. This just
+    #isn't realistic so "overcommit" on our calculations by just subtracting it off.
+
+    pi = PERFORMANCE_INFORMATION()
+    pi.cb = ctypes.sizeof(pi)
+
+    if not psapi.GetPerformanceInfo(ctypes.byref(pi), pi.cb):
+        logging.warning("WARNING: Failed to query windows performance info. RAM usage may be sub optimal")
+        return psutil.virtual_memory().available
+
+    committed = pi.CommitTotal * pi.PageSize
+    total = pi.PhysicalTotal * pi.PageSize
+
+    return max(psutil.virtual_memory().available,
+               total - (committed - comfy_aimdo.control.get_total_vram_usage()))
+