From 58c19545c83fa6925c9ce2216ee64964eb5129ce Mon Sep 17 00:00:00 2001
From: hidenorly <twitte.harold@gmail.com>
Date: Tue, 21 Nov 2023 01:13:53 +0900
Subject: Add FP32 fallback support on sd_vae_approx

This tries to execute interpolate with FP32 if it failed.

Background is that
on some environment such as Mx chip MacOS devices, we get error as follows:

```
"torch/nn/functional.py", line 3931, in interpolate
        return torch._C._nn.upsample_nearest2d(input, output_size, scale_factors)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    RuntimeError: "upsample_nearest2d_channels_last" not implemented for 'Half'
```

In this case, ```--no-half``` doesn't help to solve. Therefore this commits add the FP32 fallback execution to solve it.

Note that the submodule may require additional modifications. The following is the example modification on the other submodule.

```repositories/stable-diffusion-stability-ai/ldm/modules/diffusionmodules/openaimodel.py

class Upsample(nn.Module):
..snip..
    def forward(self, x):
        assert x.shape[1] == self.channels
        if self.dims == 3:
            x = F.interpolate(
                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
            )
        else:
            try:
                x = F.interpolate(x, scale_factor=2, mode="nearest")
            except:
                x = F.interpolate(x.to(th.float32), scale_factor=2, mode="nearest").to(x.dtype)
        if self.use_conv:
            x = self.conv(x)
        return x
..snip..
```

You can see the FP32 fallback execution as same as sd_vae_approx.py.
---
 modules/sd_vae_approx.py | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

(limited to 'modules')

diff --git a/modules/sd_vae_approx.py b/modules/sd_vae_approx.py
index 3965e223..8370493f 100644
--- a/modules/sd_vae_approx.py
+++ b/modules/sd_vae_approx.py
@@ -21,7 +21,13 @@ class VAEApprox(nn.Module):
 
     def forward(self, x):
         extra = 11
-        x = nn.functional.interpolate(x, (x.shape[2] * 2, x.shape[3] * 2))
+        try:
+            x = nn.functional.interpolate(x, (x.shape[2] * 2, x.shape[3] * 2))
+        except RuntimeError as e:
+            if "not implemented for" in str(e) and "Half" in str(e):
+                x = nn.functional.interpolate(x.to(torch.float32), (x.shape[2] * 2, x.shape[3] * 2)).to(x.dtype)
+            else:
+                print(f"An unexpected RuntimeError occurred: {str(e)}")
         x = nn.functional.pad(x, (extra, extra, extra, extra))
 
         for layer in [self.conv1, self.conv2, self.conv3, self.conv4, self.conv5, self.conv6, self.conv7, self.conv8, ]:
-- 
cgit v1.2.1