Dynamic VRAM fixes - Ace 1.5 performance + a VRAM leak (#12368)

* revert threaded model loader change

This change was only needed to get around the pytorch 2.7 mempool bugs,
and should have been reverted along with #12260. This fixes a different
memory leak where pytorch gets confused about cache emptying.

* load non comfy weights

* MPDynamic: Pre-generate the tensors for vbars

Apparently this is an expensive operation that slows down things.

* bump to aimdo 1.8

New features:
watermark limit feature
logging enhancements
-O2 build on linux

comfy/model_management.py

@@ -19,7 +19,7 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+from comfy.cli_args import args, PerformanceFeature
 import threading
 import torch
 import sys
@@ -651,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -747,26 +747,6 @@ def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, m
         current_loaded_models.insert(0, loaded_model)
     return
 
-def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load):
-    with torch.inference_mode():
-        load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        soft_empty_cache()
-
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
-    #Deliberately load models outside of the Aimdo mempool so they can be retained accross
-    #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are
-    #thread local. So exploit that to escape context
-    if enables_dynamic_vram():
-        t = threading.Thread(
-            target=load_models_gpu_thread,
-            args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        )
-        t.start()
-        t.join()
-    else:
-        load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights,
-                             minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
-
 def load_model_gpu(model):
     return load_models_gpu([model])
 
@@ -1226,21 +1206,16 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         if dtype is None:
             dtype = weight._model_dtype
 
-        r = torch.empty_like(weight, dtype=dtype, device=device)
-
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
-            raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
-            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
             if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
                 weight._v_signature = signature
                 #Send it over
                 v_tensor.copy_(weight, non_blocking=non_blocking)
-            #always take a deep copy even if _v is good, as we have no reasonable point to unpin
-            #a non comfy weight
-            r.copy_(v_tensor)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-            return r
+            return v_tensor.to(dtype=dtype)
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
 
         if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
             #Offloaded casting could skip this, however it would make the quantizations

comfy/model_patcher.py

@@ -1492,7 +1492,9 @@ class ModelPatcherDynamic(ModelPatcher):
             if vbar is not None:
                 vbar.prioritize()
 
-            #We have way more tools for acceleration on comfy weight offloading, so always
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
             #prioritize the non-comfy weights (note the order reverse).
             loading = self._load_list(prio_comfy_cast_weights=True)
             loading.sort(reverse=True)
@@ -1541,6 +1543,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     if vbar is not None and not hasattr(m, "_v"):
                         m._v = vbar.alloc(v_weight_size)
+                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
                     allocated_size += v_weight_size
 
                 else:
@@ -1555,8 +1558,10 @@ class ModelPatcherDynamic(ModelPatcher):
                         weight_size = geometry.numel() * geometry.element_size()
                         if vbar is not None and not hasattr(weight, "_v"):
                             weight._v = vbar.alloc(weight_size)
+                            weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
                             weight._model_dtype = model_dtype
                         allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
 
             logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
 

comfy/ops.py

@@ -87,7 +87,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
 
     signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
     if signature is not None:
-        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+        xfer_dest = s._v_tensor
     resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
 
     if not resident:

execution.py

@@ -13,8 +13,11 @@ from contextlib import nullcontext
 
 import torch
 
+from comfy.cli_args import args
 import comfy.memory_management
 import comfy.model_management
+import comfy_aimdo.model_vbar
+
 from latent_preview import set_preview_method
 import nodes
 from comfy_execution.caching import (
@@ -527,8 +530,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                 finally:
                     if allocator is not None:
+                        if args.verbose == "DEBUG":
+                            comfy_aimdo.model_vbar.vbars_analyze()
                         comfy.model_management.reset_cast_buffers()
-                        torch.cuda.synchronize()
+                        comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data

requirements.txt

@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.1.7
+comfy-aimdo>=0.1.8
 requests
 
 #non essential dependencies: