Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

Auto detect memory size ffor balaced offload

Browse Source
This commit is contained in:
Disty0
2024-08-11 19:01:49 +03:00
parent 4ba173280b
commit fb89e26b54
2 changed files with 13 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_models.py
+1 -1
View File
@@ -918,7 +918,7 @@ def load_diffuser(checkpoint_info=None, already_loaded_state_dict=None, timer=No
}
if shared.opts.diffusers_offload_mode == "balanced":
diffusers_load_config['device_map'] = "balanced"
if shared.opts.diffusers_offload_max_memory != 0:
if shared.opts.diffusers_offload_max_memory > 0:
device_index = torch.device(devices.device).index
if device_index is None:
device_index = 0
modules/shared.py
+12 -7
View File
@@ -353,20 +353,25 @@ def temp_disable_extensions():
return disabled
mem_stat = memory_stats()
if "gpu" in mem_stat:
gpu_memory = mem_stat['gpu']['total']
else:
gpu_memory = 0
if not (cmd_opts.lowvram or cmd_opts.medvram):
mem_stat = memory_stats()
if "gpu" in mem_stat:
if mem_stat['gpu']['total'] <= 4:
if gpu_memory <= 4:
cmd_opts.lowvram = True
offload_mode_default = "sequential"
log.info(f"VRAM: Detected={mem_stat['gpu']['total']} GB Optimization=lowvram")
elif mem_stat['gpu']['total'] <= 8:
log.info(f"VRAM: Detected={gpu_memory} GB Optimization=lowvram")
elif gpu_memory <= 8:
cmd_opts.medvram = True
offload_mode_default = "cpu"
log.info(f"VRAM: Detected={mem_stat['gpu']['total']} GB Optimization=medvram")
log.info(f"VRAM: Detected={gpu_memory} GB Optimization=medvram")
else:
offload_mode_default = "none"
log.info(f"VRAM: Detected={mem_stat['gpu']['total']} GB Optimization=none")
log.info(f"VRAM: Detected={gpu_memory} GB Optimization=none")
if devices.backend == "directml": # Force BMM for DirectML instead of SDP
@@ -531,7 +536,7 @@ options_templates.update(options_section(('diffusers', "Diffusers Settings"), {
"diffusers_extract_ema": OptionInfo(False, "Use model EMA weights when possible"),
"diffusers_generator_device": OptionInfo("GPU", "Generator device", gr.Radio, {"choices": ["GPU", "CPU", "Unset"]}),
"diffusers_offload_mode": OptionInfo(offload_mode_default, "Model offload mode", gr.Radio, {"choices": ['none', 'balanced', 'cpu', 'sequential']}),
"diffusers_offload_max_memory": OptionInfo(0, "Max memory for balanced offload mode in GB", gr.Slider, {"minimum": 0, "maximum": 24, "step": 0.1,}),
"diffusers_offload_max_memory": OptionInfo(gpu_memory * 0.8, "Max memory for balanced offload mode in GB", gr.Slider, {"minimum": 0, "maximum": gpu_memory, "step": 0.1,}),
"diffusers_vae_upcast": OptionInfo("default", "VAE upcasting", gr.Radio, {"choices": ['default', 'true', 'false']}),
"diffusers_vae_slicing": OptionInfo(True, "VAE slicing"),
"diffusers_vae_tiling": OptionInfo(cmd_opts.lowvram or cmd_opts.medvram, "VAE tiling"),