pins: implement freeing intermediate for pinned memory

Pinning is more important than inactive intermediates and the stream
pin buffer is more important than even active intermediates.

comfy/memory_management.py

@@ -157,7 +157,7 @@ def set_ram_cache_release_state(callback, headroom):
     extra_ram_release_callback = callback
     RAM_CACHE_HEADROOM = max(0, int(headroom))
 
-def extra_ram_release(target):
+def extra_ram_release(target, free_active=False):
     if extra_ram_release_callback is None:
         return 0
-    return extra_ram_release_callback(target)
+    return extra_ram_release_callback(target, free_active=free_active)

comfy/model_management.py

@@ -1208,6 +1208,7 @@ def resize_pin_buffer(pin_buffer, size):
     if size <= old_size:
         return True
     growth = size - old_size
+    comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM, free_active=True)
     ensure_pin_budget(growth, evict_active=True)
     try:
         pin_buffer.extend(size=size, reallocate=True)
@@ -1389,6 +1390,7 @@ def pin_memory(tensor):
         return False
 
     size = tensor.nbytes
+    comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM)
     ensure_pin_budget(size)
 
     ptr = tensor.data_ptr()

comfy/pinned_memory.py

@@ -17,6 +17,7 @@ def pin_memory(module, subset="weights", size=None):
     if size is None:
         size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
     offset = hostbuf.size
+    comfy.memory_management.extra_ram_release(comfy.memory_management.RAM_CACHE_HEADROOM)
     comfy.model_management.ensure_pin_budget(size)
 
     try: