Merge 6fd6ffd023 into de9ada6a41

* mp: fix full dynamic unloading

This was not unloading dynamic models when requesting a full unload via
the unpatch() code path.

This was ok, i your workflow was all dynamic models but fails with big
VRAM leaks if you need to fully unload something for a regular ModelPatcher

It also fices the "unload models" button.

* mm: load models outside of Aimdo Mempool

In dynamic_vram mode, escape the Aimdo mempool and load into the regular
mempool. Use a dummy thread to do it.

comfy/cli_args.py

@@ -151,6 +151,7 @@ parser.add_argument("--force-non-blocking", action="store_true", help="Force Com
 
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
+parser.add_argument("--disable-fp8-compute", action="store_true", help="Prevent ComfyUI from activating fp8 compute in Nvidia cards that support it. Can prevent some issues with some models not suitable for fp8 compute.")
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
 

comfy/model_management.py

@@ -19,7 +19,8 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature
+from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+import threading
 import torch
 import sys
 import platform
@@ -650,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -746,8 +747,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         current_loaded_models.insert(0, loaded_model)
     return
 
-def load_model_gpu(model):
-    return load_models_gpu([model])
+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 loaded_models(only_currently_used=False):
     output = []
@@ -1112,11 +1130,11 @@ def get_cast_buffer(offload_stream, device, size, ref):
             return None
         if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
             #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
-            torch.cuda.synchronize()
+            synchronize()
             del STREAM_CAST_BUFFERS[offload_stream]
             del cast_buffer
             #FIXME: This doesn't work in Aimdo because mempool cant clear cache
-            torch.cuda.empty_cache()
+            soft_empty_cache()
         with wf_context:
             cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
             STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
@@ -1132,9 +1150,7 @@ def reset_cast_buffers():
     for offload_stream in STREAM_CAST_BUFFERS:
         offload_stream.synchronize()
     STREAM_CAST_BUFFERS.clear()
-    if comfy.memory_management.aimdo_allocator is None:
-        #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-        torch.cuda.empty_cache()
+    soft_empty_cache()
 
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
@@ -1284,7 +1300,7 @@ def discard_cuda_async_error():
         a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         _ = a + b
-        torch.cuda.synchronize()
+        synchronize()
     except torch.AcceleratorError:
         #Dump it! We already know about it from the synchronous return
         pass
@@ -1639,6 +1655,8 @@ def supports_fp8_compute(device=None):
 
     if not is_nvidia():
         return False
+    if args.disable_fp8_compute:
+        return False
 
     props = torch.cuda.get_device_properties(device)
     if props.major >= 9:
@@ -1688,6 +1706,12 @@ def lora_compute_dtype(device):
     LORA_COMPUTE_DTYPES[device] = dtype
     return dtype
 
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1713,9 +1737,6 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-#TODO: might be cleaner to put this somewhere else
-import threading
-
 class InterruptProcessingException(Exception):
     pass
 

comfy/model_patcher.py

@@ -1597,7 +1597,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
         if unpatch_weights:
             self.partially_unload_ram(1e32)
-            self.partially_unload(None)
+            self.partially_unload(None, 1e32)
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
         assert not force_patch_weights #See above