jinx/ComfyUI
1
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-02 23:14:49 +08:00
Code Issues Activity

Add experimental --async-offload lowvram weight offloading. (#7820)

Browse Source 
This should speed up the lowvram mode a bit. It currently is only enabled when --async-offload is used but it will be enabled by default in the future if there are no problems.
This commit is contained in:
comfyanonymous
2025-04-26 13:11:21 -07:00
committed by GitHub
parent b685b8a4e0
commit 0dcc75ca54
3 changed files with 50 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
comfy/ops.py
View File

@@ -37,20 +37,23 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
if device is None:
device = input.device
offload_stream = comfy.model_management.get_offload_stream(device)
bias = None
non_blocking = comfy.model_management.device_supports_non_blocking(device)
if s.bias is not None:
has_function = len(s.bias_function) > 0
bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
if has_function:
for f in s.bias_function:
bias = f(bias)
has_function = len(s.weight_function) > 0
weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
if has_function:
for f in s.weight_function:
weight = f(weight)
comfy.model_management.sync_stream(device, offload_stream)
return weight, bias
class CastWeightBiasOp: