reverse logic of gpu_mem_reserved

- gpu_mem_reserved now indicates the amount of VRAM that will be reserved
  for model caching (similar to max_cache_size).

invokeai/app/services/config.py

@@ -26,7 +26,7 @@ InvokeAI:
     max_cache_size: 6
     always_use_cpu: false
     free_gpu_mem: false
-    gpu_mem_reserved: 1
+    gpu_mem_reserved: 2.7
   Features:
     nsfw_checker: true
     restore: true
@@ -366,7 +366,7 @@ setting environment variables INVOKEAI_<setting>.
     free_gpu_mem        : bool = Field(default=False, description="If true, purge model from GPU after each generation.", category='Memory/Performance')
     max_loaded_models   : int = Field(default=3, gt=0, description="(DEPRECATED: use max_cache_size) Maximum number of models to keep in memory for rapid switching", category='Memory/Performance')
     max_cache_size      : float = Field(default=6.0, gt=0, description="Maximum memory amount used by model cache for rapid switching", category='Memory/Performance')
-    gpu_mem_reserved    : float = Field(default=1.75, ge=0, description="Amount of VRAM to reserve for use during generation", category='Memory/Performance')
+    gpu_mem_reserved    : float = Field(default=2.75, ge=0, description="Amount of VRAM reserved for model storage", category='Memory/Performance')
     precision           : Literal[tuple(['auto','float16','float32','autocast'])] = Field(default='float16',description='Floating point precision', category='Memory/Performance')
     sequential_guidance : bool = Field(default=False, description="Whether to calculate guidance in serial instead of in parallel, lowering memory requirements", category='Memory/Performance')
     xformers_enabled    : bool = Field(default=True, description="Enable/disable memory-efficient attention", category='Memory/Performance')

invokeai/backend/model_management/model_cache.py

@@ -37,7 +37,7 @@ from .models import BaseModelType, ModelType, SubModelType, ModelBase
 DEFAULT_MAX_CACHE_SIZE = 6.0
 
 # amount of GPU memory to hold in reserve for use by generations (GB)
-DEFAULT_GPU_MEM_RESERVED= 1.75
+DEFAULT_GPU_MEM_RESERVED= 2.75
 
 # actual size of a gig
 GIG = 1073741824
@@ -350,17 +350,18 @@ class ModelCache(object):
 
     def _offload_unlocked_models(self, size_needed: int=0):
         reserved = self.gpu_mem_reserved * GIG
+        vram_in_use = torch.cuda.memory_allocated()
+        self.logger.debug(f'{(vram_in_use/GIG):.2f}GB VRAM used for models; max allowed={(reserved/GIG):.2f}GB')
         for model_key, cache_entry in sorted(self._cached_models.items(), key=lambda x:x[1].size):
-            free_mem, used_mem = torch.cuda.mem_get_info()
-            free_mem -= reserved
-            self.logger.debug(f'Require {(size_needed/GIG):.2f}GB VRAM. Have {(free_mem/GIG):.2f}GB available ({(reserved/GIG):.2f} reserved).')
-            if free_mem > size_needed:
+            if vram_in_use <= reserved:
                 break
             if not cache_entry.locked and cache_entry.loaded:
                 self.logger.debug(f'Offloading {model_key} from {self.execution_device} into {self.storage_device}')
                 with VRAMUsage() as mem:
                     cache_entry.model.to(self.storage_device)
                 self.logger.debug(f'GPU VRAM freed: {(mem.vram_used/GIG):.2f} GB')
+                vram_in_use += mem.vram_used  # note vram_used is negative
+                self.logger.debug(f'{(vram_in_use/GIG):.2f}GB VRAM used for models; max allowed={(reserved/GIG):.2f}GB')
         
     def _local_model_hash(self, model_path: Union[str, Path]) -> str:
         sha = hashlib.sha256()

invokeai/backend/model_management/model_manager.py

@@ -340,6 +340,7 @@ class ModelManager(object):
             precision = precision,
             sequential_offload = sequential_offload,
             logger = logger,
+            gpu_mem_reserved = self.app_config.gpu_mem_reserved
         )
         self.cache_keys = dict()