From faed465a0b1a7d19669568738c93e04907c10415 Mon Sep 17 00:00:00 2001 From: brkirch Date: Tue, 25 Oct 2022 02:01:57 -0400 Subject: [PATCH] MPS Upscalers Fix Get ESRGAN, SCUNet, and SwinIR working correctly on MPS by ensuring memory is contiguous for tensor views before sending to MPS device. --- modules/devices.py | 4 ++++ modules/esrgan_model.py | 2 +- modules/scunet_model.py | 3 +-- modules/swinir_model.py | 2 +- 4 files changed, 7 insertions(+), 4 deletions(-) diff --git a/modules/devices.py b/modules/devices.py index 033a42d5..7511e1dc 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -81,3 +81,7 @@ def autocast(disable=False): return contextlib.nullcontext() return torch.autocast("cuda") + +# MPS workaround for https://github.com/pytorch/pytorch/issues/79383 +def mps_contiguous(input_tensor, device): return input_tensor.contiguous() if device.type == 'mps' else input_tensor +def mps_contiguous_to(input_tensor, device): return mps_contiguous(input_tensor, device).to(device) diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py index a49e2258..a13cf6ac 100644 --- a/modules/esrgan_model.py +++ b/modules/esrgan_model.py @@ -190,7 +190,7 @@ def upscale_without_tiling(model, img): img = img[:, :, ::-1] img = np.ascontiguousarray(np.transpose(img, (2, 0, 1))) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(devices.device_esrgan) + img = devices.mps_contiguous_to(img.unsqueeze(0), devices.device_esrgan) with torch.no_grad(): output = model(img) output = output.squeeze().float().cpu().clamp_(0, 1).numpy() diff --git a/modules/scunet_model.py b/modules/scunet_model.py index 36a996bf..59532274 100644 --- a/modules/scunet_model.py +++ b/modules/scunet_model.py @@ -54,9 +54,8 @@ class UpscalerScuNET(modules.upscaler.Upscaler): img = img[:, :, ::-1] img = np.moveaxis(img, 2, 0) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(device) + img = devices.mps_contiguous_to(img.unsqueeze(0), device) - img = img.to(device) with torch.no_grad(): output = model(img) output = output.squeeze().float().cpu().clamp_(0, 1).numpy() diff --git a/modules/swinir_model.py b/modules/swinir_model.py index facd262d..4253b66d 100644 --- a/modules/swinir_model.py +++ b/modules/swinir_model.py @@ -111,7 +111,7 @@ def upscale( img = img[:, :, ::-1] img = np.moveaxis(img, 2, 0) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(devices.device_swinir) + img = devices.mps_contiguous_to(img.unsqueeze(0), devices.device_swinir) with torch.no_grad(), precision_scope("cuda"): _, _, h_old, w_old = img.size() h_pad = (h_old // window_size + 1) * window_size - h_old