From e632e5de281b91dd7199636dd6d82126fbfb07d5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 9 Nov 2025 15:06:39 -0800 Subject: [PATCH 1/3] Add logging for model unloading. (#10692) --- comfy/model_patcher.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 5a31a8734..17e06a869 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -909,6 +909,7 @@ class ModelPatcher: 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)) return memory_freed def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): From dea899f22125d38a8b48147d6cce89a2b659fdeb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 9 Nov 2025 15:51:33 -0800 Subject: [PATCH 2/3] Unload weights if vram usage goes up between runs. (#10690) --- comfy/model_management.py | 11 +++++++++-- comfy/model_patcher.py | 20 +++++++++++++------- 2 files changed, 22 insertions(+), 9 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 7012df858..a4410f2ec 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -503,7 +503,11 @@ class LoadedModel: use_more_vram = lowvram_model_memory if use_more_vram == 0: use_more_vram = 1e32 - self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) + if use_more_vram > 0: + self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) + else: + self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights) + real_model = self.model.model if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None: @@ -689,7 +693,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu 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(0.1, lowvram_model_memory - loaded_memory) + lowvram_model_memory = lowvram_model_memory - loaded_memory + + if lowvram_model_memory == 0: + lowvram_model_memory = 0.1 if vram_set_state == VRAMState.NO_VRAM: lowvram_model_memory = 0.1 diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 17e06a869..68b0a9192 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -843,7 +843,7 @@ class ModelPatcher: self.object_patches_backup.clear() - def partially_unload(self, device_to, memory_to_free=0): + def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False): with self.use_ejected(): hooks_unpatched = False memory_freed = 0 @@ -887,13 +887,19 @@ class ModelPatcher: module_mem += move_weight_functions(m, device_to) if lowvram_possible: if weight_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, weight_key) - m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(weight_key) + else: + _, set_func, convert_func = get_key_weight(self.model, weight_key) + m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) + patch_counter += 1 if bias_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, bias_key) - m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(bias_key) + else: + _, set_func, convert_func = get_key_weight(self.model, bias_key) + m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) + patch_counter += 1 cast_weight = True if cast_weight: From c350009236e5d172a3050c04043ea70a301378ca Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 10 Nov 2025 13:52:11 +1000 Subject: [PATCH 3/3] ops: Put weight cast on the offload stream (#10697) This needs to be on the offload stream. This reproduced a black screen with low resolution images on a slow bus when using FP8. --- comfy/ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 733bff99d..96dffa85d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -110,9 +110,9 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of for f in s.bias_function: bias = f(bias) - weight = weight.to(dtype=dtype) - if weight_has_function: + if weight_has_function or weight.dtype != dtype: with wf_context: + weight = weight.to(dtype=dtype) for f in s.weight_function: weight = f(weight)