From 4a216ded433ded315106e2989c5ff7dec1c49304 Mon Sep 17 00:00:00 2001
From: Ljzd-PRO <63289359+Ljzd-PRO@users.noreply.github.com>
Date: Thu, 13 Oct 2022 02:07:49 +0800
Subject: [PATCH] load models to VRAM when using `--lowram` param

load models to VRM instead of RAM (for machines which have bigger VRM than RAM such as free Google Colab server)
---
 modules/sd_models.py | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/modules/sd_models.py b/modules/sd_models.py
index 0a55b4c32..78a198b9c 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -134,7 +134,12 @@ def load_model_weights(model, checkpoint_info):
 
     print(f"Loading weights [{sd_model_hash}] from {checkpoint_file}")
 
-    pl_sd = torch.load(checkpoint_file, map_location="cpu")
+    if shared.cmd_opts.lowram:
+        print("Load to VRAM if GPU is available (low RAM)")
+        pl_sd = torch.load(checkpoint_file)
+    else:
+        pl_sd = torch.load(checkpoint_file, map_location="cpu")
+
     if "global_step" in pl_sd:
         print(f"Global Step: {pl_sd['global_step']}")
 
@@ -158,7 +163,13 @@ def load_model_weights(model, checkpoint_info):
 
     if os.path.exists(vae_file):
         print(f"Loading VAE weights from: {vae_file}")
-        vae_ckpt = torch.load(vae_file, map_location="cpu")
+
+        if shared.cmd_opts.lowram:
+            print("Load to VRAM if GPU is available (low RAM)")
+            vae_ckpt = torch.load(vae_file)
+        else:
+            vae_ckpt = torch.load(vae_file, map_location="cpu")
+
         vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss"}
 
         model.first_stage_model.load_state_dict(vae_dict)