The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
mp: use look-ahead actuals for stream offload VRAM calculation (#11096)
rattus6be85c792022
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.pyindex df2d8e827..3dcac3eef 100644--- a/comfy/model_patcher.py+++ b/comfy/model_patcher.py@@ -699,12 +699,12 @@ class ModelPatcher:offloaded = []offload_buffer = 0loading.sort(reverse=True)- for x in loading:+ for i, x in enumerate(loading):module_offload_mem, module_mem, n, m, params = xlowvram_weight = False- potential_offload = max(offload_buffer, module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem))+ potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]]))lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memoryweight_key = "{}.weight".format(n)@@ -876,14 +876,18 @@ class ModelPatcher:patch_counter = 0unload_list = self._load_list()unload_list.sort()+offload_buffer = self.model.model_offload_buffer_memory+ if len(unload_list) > 0:+ NS = comfy.model_management.NUM_STREAMS+ offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NSfor unload in unload_list:if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed:breakmodule_offload_mem, module_mem, n, m, params = unload- potential_offload = module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem)+ potential_offload = module_offload_mem + sum(offload_weight_factor)lowvram_possible = hasattr(m, "comfy_cast_weights")if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:@@ -935,6 +939,8 @@ class ModelPatcher:m.comfy_patched_weights = Falsememory_freed += module_memoffload_buffer = max(offload_buffer, potential_offload)+ offload_weight_factor.append(module_mem)+ offload_weight_factor.pop(0)logging.debug("freed {}".format(n))for param in params: