wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-02-16 08:22:36 +08:00
Code Issues Actions 15 Packages Projects Releases Wiki Activity

dynamic_vram: Training fixes (#12442)

Browse Source
This commit is contained in:
rattus 2026-02-13 12:29:37 -08:00 committed by GitHub
parent e03fe8b591
commit 8902907d7a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 14 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
comfy/model_patcher.py
View File

@ -1561,6 +1561,8 @@ class ModelPatcherDynamic(ModelPatcher):
allocated_size += weight_size
vbar.set_watermark_limit(allocated_size)
move_weight_functions(m, device_to)
logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
self.model.device = device_to
@ -1601,6 +1603,8 @@ class ModelPatcherDynamic(ModelPatcher):
if unpatch_weights:
self.partially_unload_ram(1e32)
self.partially_unload(None, 1e32)
for m in self.model.modules():
move_weight_functions(m, device_to)
def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
assert not force_patch_weights #See above

11
comfy_extras/nodes_train.py
View File

@ -1035,7 +1035,7 @@ class TrainLoraNode(io.ComfyNode):
io.Boolean.Input(
"offloading",
default=False,
tooltip="Depth level for gradient checkpointing.",
tooltip="Offload the Model to RAM. Requires Bypass Mode.",
),
io.Combo.Input(
"existing_lora",
@ -1124,6 +1124,15 @@ class TrainLoraNode(io.ComfyNode):
lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype)
mp.set_model_compute_dtype(dtype)
if mp.is_dynamic():
if not bypass_mode:
logging.info("Training MP is Dynamic - forcing bypass mode. Start comfy with --highvram to force weight diff mode")
bypass_mode = True
offloading = True
elif offloading:
if not bypass_mode:
logging.info("Training Offload selected - forcing bypass mode. Set bypass = True to remove this message")
# Prepare latents and compute counts
latents, num_images, multi_res = _prepare_latents_and_count(
latents, dtype, bucket_mode