Account for the VRAM cost of weight offloading (#10733)

* mm: default to 0 for NUM_STREAMS

Dont count the compute stream as an offload stream. This makes async
offload accounting easier.

* mm: remove 128MB minimum

This is from a previous offloading system requirement. Remove it to
make behaviour of the loader and partial unloader consistent.

* mp: order the module list by offload expense

Calculate an approximate offloading temporary VRAM cost to offload a
weight and primary order the module load list by that. In the simple
case this is just the same as the module weight, but with Loras, a
weight with a lora consumes considerably more VRAM to do the Lora
application on-the-fly.

This will slightly prioritize lora weights, but is really for
proper VRAM offload accounting.

* mp: Account for the VRAM cost of weight offloading

when checking the VRAM headroom, assume that the weight needs to be
offloaded, and only load if it has space for both the load and offload
 * the number of streams.

As the weights are ordered from largest to smallest by offload cost
this is guaranteed to fit in VRAM (tm), as all weights that follow
will be smaller.

Make the partial unload aware of this system as well by saving the
budget for offload VRAM to the model state and accounting accordingly.
Its possible that partial unload increases the size of the largest
offloaded weights, and thus needs to unload a little bit more than
asked to accomodate the bigger temp buffers.

Honor the existing codes floor on model weight loading of 128MB by
having the patcher honor this separately withough regard to offloading.
Otherwise when MM specifies its 128MB minimum, MP will see the biggest
weights, and budget that 128MB to only offload buffer and load nothing
which isnt the intent of these minimums. The same clamp applies in
case of partial offload of the currently loading model.

Author: rattus128

comfy/model_management.py

@@ -689,7 +689,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             loaded_memory = loaded_model.model_loaded_memory()
             current_free_mem = get_free_memory(torch_dev) + loaded_memory
 
-            lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
+            lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
             lowvram_model_memory = lowvram_model_memory - loaded_memory
 
             if lowvram_model_memory == 0:
@@ -1012,7 +1012,7 @@ def force_channels_last():
 
 
 STREAMS = {}
-NUM_STREAMS = 1
+NUM_STREAMS = 0
 if args.async_offload:
     NUM_STREAMS = 2
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
@@ -1030,7 +1030,7 @@ def current_stream(device):
 stream_counters = {}
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
-    if NUM_STREAMS <= 1:
+    if NUM_STREAMS == 0:
         return None
 
     if device in STREAMS:

comfy/model_patcher.py

@@ -148,6 +148,15 @@ def __call__(self, weight):
         else:
             return out
 
+#The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3
+LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3
+
+def low_vram_patch_estimate_vram(model, key):
+    weight, set_func, convert_func = get_key_weight(model, key)
+    if weight is None:
+        return 0
+    return weight.numel() * torch.float32.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR
+
 def get_key_weight(model, key):
     set_func = None
     convert_func = None
@@ -269,6 +278,9 @@ def __init__(self, model, load_device, offload_device, size=0, weight_inplace_up
         if not hasattr(self.model, 'current_weight_patches_uuid'):
             self.model.current_weight_patches_uuid = None
 
+        if not hasattr(self.model, 'model_offload_buffer_memory'):
+            self.model.model_offload_buffer_memory = 0
+
     def model_size(self):
         if self.size > 0:
             return self.size
@@ -662,7 +674,16 @@ def _load_list(self):
                     skip = True # skip random weights in non leaf modules
                     break
             if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
-                loading.append((comfy.model_management.module_size(m), n, m, params))
+                module_mem = comfy.model_management.module_size(m)
+                module_offload_mem = module_mem
+                if hasattr(m, "comfy_cast_weights"):
+                    weight_key = "{}.weight".format(n)
+                    bias_key = "{}.bias".format(n)
+                    if weight_key in self.patches:
+                        module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key)
+                    if bias_key in self.patches:
+                        module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key)
+                loading.append((module_offload_mem, module_mem, n, m, params))
         return loading
 
     def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
@@ -676,20 +697,22 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
 
             load_completely = []
             offloaded = []
+            offload_buffer = 0
             loading.sort(reverse=True)
             for x in loading:
-                n = x[1]
-                m = x[2]
-                params = x[3]
-                module_mem = x[0]
+                module_offload_mem, module_mem, n, m, params = x
 
                 lowvram_weight = False
 
+                potential_offload = max(offload_buffer, module_offload_mem * (comfy.model_management.NUM_STREAMS + 1))
+                lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory
+
                 weight_key = "{}.weight".format(n)
                 bias_key = "{}.bias".format(n)
 
                 if not full_load and hasattr(m, "comfy_cast_weights"):
-                    if mem_counter + module_mem >= lowvram_model_memory:
+                    if not lowvram_fits:
+                        offload_buffer = potential_offload
                         lowvram_weight = True
                         lowvram_counter += 1
                         lowvram_mem_counter += module_mem
@@ -723,9 +746,11 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
                     if hasattr(m, "comfy_cast_weights"):
                         wipe_lowvram_weight(m)
 
-                    if full_load or mem_counter + module_mem < lowvram_model_memory:
+                    if full_load or lowvram_fits:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
+                    else:
+                        offload_buffer = potential_offload
 
                 if cast_weight and hasattr(m, "comfy_cast_weights"):
                     m.prev_comfy_cast_weights = m.comfy_cast_weights
@@ -766,7 +791,7 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
                 logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
@@ -778,6 +803,7 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
             self.model.lowvram_patch_counter += patch_counter
             self.model.device = device_to
             self.model.model_loaded_weight_memory = mem_counter
+            self.model.model_offload_buffer_memory = offload_buffer
             self.model.current_weight_patches_uuid = self.patches_uuid
 
             for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
@@ -831,6 +857,7 @@ def unpatch_model(self, device_to=None, unpatch_weights=True):
                 self.model.to(device_to)
                 self.model.device = device_to
             self.model.model_loaded_weight_memory = 0
+            self.model.model_offload_buffer_memory = 0
 
             for m in self.model.modules():
                 if hasattr(m, "comfy_patched_weights"):
@@ -849,13 +876,14 @@ def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=Fals
             patch_counter = 0
             unload_list = self._load_list()
             unload_list.sort()
+            offload_buffer = self.model.model_offload_buffer_memory
+
             for unload in unload_list:
-                if memory_to_free < memory_freed:
+                if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed:
                     break
-                module_mem = unload[0]
-                n = unload[1]
-                m = unload[2]
-                params = unload[3]
+                module_offload_mem, module_mem, n, m, params = unload
+
+                potential_offload = (comfy.model_management.NUM_STREAMS + 1) * module_offload_mem
 
                 lowvram_possible = hasattr(m, "comfy_cast_weights")
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
@@ -906,15 +934,18 @@ def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=Fals
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False
                         memory_freed += module_mem
+                        offload_buffer = max(offload_buffer, potential_offload)
                         logging.debug("freed {}".format(n))
 
                         for param in params:
                             self.pin_weight_to_device("{}.{}".format(n, param))
 
+
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
-            logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter))
+            self.model.model_offload_buffer_memory = offload_buffer
+            logging.info("Unloaded partially: {:.2f} MB freed, {:.2f} MB remains loaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(memory_freed / (1024 * 1024), self.model.model_loaded_weight_memory / (1024 * 1024), offload_buffer / (1024 * 1024), self.model.lowvram_patch_counter))
             return memory_freed
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):