From ab470444591e425290db72db9ebc3127f5520449 Mon Sep 17 00:00:00 2001 From: deepdiffuser Date: Fri, 10 Mar 2023 04:29:09 -0800 Subject: [PATCH] add multi-gpu support for 4bit gptq LLaMA --- modules/models.py | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) diff --git a/modules/models.py b/modules/models.py index 3e6cea18..14443c89 100644 --- a/modules/models.py +++ b/modules/models.py @@ -110,7 +110,18 @@ def load_model(model_name): exit() model = load_quant(path_to_model, Path(f"models/{pt_model}"), 4) - model = model.to(torch.device('cuda:0')) + + if shared.args.gpu_memory: + max_memory = {} + for i in range(len(shared.args.gpu_memory)): + max_memory[i] = f"{shared.args.gpu_memory[i]}GiB" + max_memory['cpu'] = f"{shared.args.cpu_memory or '99'}GiB" + + import accelerate + device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory) + model = accelerate.dispatch_model(model, device_map=device_map) + else: + model = model.to(torch.device('cuda:0')) # Custom else: