From 57eb54b838faa383c10079e1bb5471b7bee6a695 Mon Sep 17 00:00:00 2001
From: Extraltodeus <extraltodeus@gmail.com>
Date: Sat, 22 Oct 2022 00:11:07 +0200
Subject: implement CUDA device selection by ID

---
 modules/devices.py | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)

(limited to 'modules/devices.py')

diff --git a/modules/devices.py b/modules/devices.py
index eb422583..8a159282 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -1,7 +1,6 @@
+import sys, os, shlex
 import contextlib
-
 import torch
-
 from modules import errors
 
 # has_mps is only available in nightly pytorch (for now), `getattr` for compatibility
@@ -9,10 +8,26 @@ has_mps = getattr(torch, 'has_mps', False)
 
 cpu = torch.device("cpu")
 
+def extract_device_id(args, name):
+    for x in range(len(args)):
+        if name in args[x]: return args[x+1]
+    return None
 
 def get_optimal_device():
     if torch.cuda.is_available():
-        return torch.device("cuda")
+        # CUDA device selection support:
+        if "shared" not in sys.modules:
+            commandline_args = os.environ.get('COMMANDLINE_ARGS', "") #re-parse the commandline arguments because using the shared.py module creates an import loop.
+            sys.argv += shlex.split(commandline_args)
+            device_id = extract_device_id(sys.argv, '--device-id')
+        else:
+            device_id = shared.cmd_opts.device_id
+            
+        if device_id is not None:
+            cuda_device = f"cuda:{device_id}"
+            return torch.device(cuda_device)
+        else:
+            return torch.device("cuda")
 
     if has_mps:
         return torch.device("mps")
-- 
cgit v1.2.3


From 50b5504401e50b6c94eba41b37fe212b2f27b792 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Sat, 22 Oct 2022 14:04:14 +0300
Subject: remove parsing command line from devices.py

---
 modules/devices.py | 14 +++++---------
 modules/lowvram.py |  9 ++++-----
 2 files changed, 9 insertions(+), 14 deletions(-)

(limited to 'modules/devices.py')

diff --git a/modules/devices.py b/modules/devices.py
index 8a159282..dc1f3cdd 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -15,14 +15,10 @@ def extract_device_id(args, name):
 
 def get_optimal_device():
     if torch.cuda.is_available():
-        # CUDA device selection support:
-        if "shared" not in sys.modules:
-            commandline_args = os.environ.get('COMMANDLINE_ARGS', "") #re-parse the commandline arguments because using the shared.py module creates an import loop.
-            sys.argv += shlex.split(commandline_args)
-            device_id = extract_device_id(sys.argv, '--device-id')
-        else:
-            device_id = shared.cmd_opts.device_id
-            
+        from modules import shared
+
+        device_id = shared.cmd_opts.device_id
+
         if device_id is not None:
             cuda_device = f"cuda:{device_id}"
             return torch.device(cuda_device)
@@ -49,7 +45,7 @@ def enable_tf32():
 
 errors.run(enable_tf32, "Enabling TF32")
 
-device = device_interrogate = device_gfpgan = device_bsrgan = device_esrgan = device_scunet = device_codeformer = get_optimal_device()
+device = device_interrogate = device_gfpgan = device_bsrgan = device_esrgan = device_scunet = device_codeformer = None
 dtype = torch.float16
 dtype_vae = torch.float16
 
diff --git a/modules/lowvram.py b/modules/lowvram.py
index 7eba1349..f327c3df 100644
--- a/modules/lowvram.py
+++ b/modules/lowvram.py
@@ -1,9 +1,8 @@
 import torch
-from modules.devices import get_optimal_device
+from modules import devices
 
 module_in_gpu = None
 cpu = torch.device("cpu")
-device = gpu = get_optimal_device()
 
 
 def send_everything_to_cpu():
@@ -33,7 +32,7 @@ def setup_for_low_vram(sd_model, use_medvram):
         if module_in_gpu is not None:
             module_in_gpu.to(cpu)
 
-        module.to(gpu)
+        module.to(devices.device)
         module_in_gpu = module
 
     # see below for register_forward_pre_hook;
@@ -51,7 +50,7 @@ def setup_for_low_vram(sd_model, use_medvram):
     # send the model to GPU. Then put modules back. the modules will be in CPU.
     stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model
     sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = None, None, None
-    sd_model.to(device)
+    sd_model.to(devices.device)
     sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = stored
 
     # register hooks for those the first two models
@@ -70,7 +69,7 @@ def setup_for_low_vram(sd_model, use_medvram):
         # so that only one of them is in GPU at a time
         stored = diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed
         diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = None, None, None, None
-        sd_model.model.to(device)
+        sd_model.model.to(devices.device)
         diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = stored
 
         # install hooks for bits of third model
-- 
cgit v1.2.3