Option to use CPU for random number generation.

Makes a given manual seed generate the same images across different
platforms, independently of the GPU architecture in use.

Fixes #9613.

modules/devices.py

@@ -92,14 +92,18 @@ def cond_cast_float(input):
 
 
 def randn(seed, shape):
+    from modules.shared import opts
+
     torch.manual_seed(seed)
-    if device.type == 'mps':
+    if opts.use_cpu_randn or device.type == 'mps':
         return torch.randn(shape, device=cpu).to(device)
     return torch.randn(shape, device=device)
 
 
 def randn_without_seed(shape):
-    if device.type == 'mps':
+    from modules.shared import opts
+
+    if opts.use_cpu_randn or device.type == 'mps':
         return torch.randn(shape, device=cpu).to(device)
     return torch.randn(shape, device=device)
 

modules/sd_samplers_common.py

@@ -60,3 +60,12 @@ def store_latent(decoded):
 
 class InterruptedException(BaseException):
     pass
+
+if opts.use_cpu_randn:
+    import torchsde._brownian.brownian_interval
+
+    def torchsde_randn(size, dtype, device, seed):
+        generator = torch.Generator(devices.cpu).manual_seed(int(seed))
+        return torch.randn(size, dtype=dtype, device=devices.cpu, generator=generator).to(device)
+
+    torchsde._brownian.brownian_interval._randn = torchsde_randn

modules/sd_samplers_kdiffusion.py

@@ -190,7 +190,7 @@ class TorchHijack:
             if noise.shape == x.shape:
                 return noise
 
-        if x.device.type == 'mps':
+        if opts.use_cpu_randn or x.device.type == 'mps':
             return torch.randn_like(x, device=devices.cpu).to(x.device)
         else:
             return torch.randn_like(x)

modules/shared.py

@@ -331,6 +331,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "comma_padding_backtrack": OptionInfo(20, "Increase coherency by padding from the last comma within n tokens when using more than 75 tokens", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1 }),
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}),
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
+    "use_cpu_randn": OptionInfo(False, "Use CPU for random number generation to make manual seeds generate the same image across platforms. This may change existing seeds."),
 }))
 
 options_templates.update(options_section(('compatibility', "Compatibility"), {