From bdaa36c84470adbdce3e98c01a69af5e95adfb02 Mon Sep 17 00:00:00 2001 From: brkirch Date: Fri, 30 Sep 2022 23:53:25 -0400 Subject: When device is MPS, use CPU for GFPGAN instead GFPGAN will not work if the device is MPS, so default to CPU instead. --- modules/devices.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/devices.py') diff --git a/modules/devices.py b/modules/devices.py index 07bb2339..08bb26d6 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -34,7 +34,7 @@ errors.run(enable_tf32, "Enabling TF32") device = get_optimal_device() -device_codeformer = cpu if has_mps else device +device_gfpgan = device_codeformer = cpu if device.type == 'mps' else device def randn(seed, shape): -- cgit v1.2.3 From eeab7aedf532680a6ae9058ee272450bb07e41eb Mon Sep 17 00:00:00 2001 From: brkirch Date: Tue, 4 Oct 2022 04:24:35 -0400 Subject: Add --use-cpu command line option Remove MPS detection to use CPU for GFPGAN / CodeFormer and add a --use-cpu command line option. --- modules/devices.py | 5 ++--- modules/esrgan_model.py | 9 ++++----- modules/scunet_model.py | 8 ++++---- modules/shared.py | 9 +++++++-- 4 files changed, 17 insertions(+), 14 deletions(-) (limited to 'modules/devices.py') diff --git a/modules/devices.py b/modules/devices.py index 5d9c7a07..b5a0cd29 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -1,8 +1,8 @@ import torch -# has_mps is only available in nightly pytorch (for now), `getattr` for compatibility from modules import errors +# has_mps is only available in nightly pytorch (for now), `getattr` for compatibility has_mps = getattr(torch, 'has_mps', False) cpu = torch.device("cpu") @@ -32,8 +32,7 @@ def enable_tf32(): errors.run(enable_tf32, "Enabling TF32") -device = get_optimal_device() -device_gfpgan = device_codeformer = cpu if device.type == 'mps' else device +device = device_gfpgan = device_esrgan = device_scunet = device_codeformer = get_optimal_device() dtype = torch.float16 def randn(seed, shape): diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py index 4aed9283..d17e730f 100644 --- a/modules/esrgan_model.py +++ b/modules/esrgan_model.py @@ -6,8 +6,7 @@ from PIL import Image from basicsr.utils.download_util import load_file_from_url import modules.esrgam_model_arch as arch -from modules import shared, modelloader, images -from modules.devices import has_mps +from modules import shared, modelloader, images, devices from modules.paths import models_path from modules.upscaler import Upscaler, UpscalerData from modules.shared import opts @@ -97,7 +96,7 @@ class UpscalerESRGAN(Upscaler): model = self.load_model(selected_model) if model is None: return img - model.to(shared.device) + model.to(devices.device_esrgan) img = esrgan_upscale(model, img) return img @@ -112,7 +111,7 @@ class UpscalerESRGAN(Upscaler): print("Unable to load %s from %s" % (self.model_path, filename)) return None - pretrained_net = torch.load(filename, map_location='cpu' if has_mps else None) + pretrained_net = torch.load(filename, map_location='cpu' if shared.device.type == 'mps' else None) crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32) pretrained_net = fix_model_layers(crt_model, pretrained_net) @@ -127,7 +126,7 @@ def upscale_without_tiling(model, img): img = img[:, :, ::-1] img = np.moveaxis(img, 2, 0) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(shared.device) + img = img.unsqueeze(0).to(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 7987ac14..fb64b740 100644 --- a/modules/scunet_model.py +++ b/modules/scunet_model.py @@ -8,7 +8,7 @@ import torch from basicsr.utils.download_util import load_file_from_url import modules.upscaler -from modules import shared, modelloader +from modules import devices, modelloader from modules.paths import models_path from modules.scunet_model_arch import SCUNet as net @@ -51,12 +51,12 @@ class UpscalerScuNET(modules.upscaler.Upscaler): if model is None: return img - device = shared.device + device = devices.device_scunet img = np.array(img) img = img[:, :, ::-1] img = np.moveaxis(img, 2, 0) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(shared.device) + img = img.unsqueeze(0).to(device) img = img.to(device) with torch.no_grad(): @@ -69,7 +69,7 @@ class UpscalerScuNET(modules.upscaler.Upscaler): return PIL.Image.fromarray(output, 'RGB') def load_model(self, path: str): - device = shared.device + device = devices.device_scunet if "http" in path: filename = load_file_from_url(url=self.model_url, model_dir=self.model_path, file_name="%s.pth" % self.name, progress=True) diff --git a/modules/shared.py b/modules/shared.py index 2a599e9c..7899ab8d 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -12,7 +12,7 @@ import modules.interrogate import modules.memmon import modules.sd_models import modules.styles -from modules.devices import get_optimal_device +import modules.devices as devices from modules.paths import script_path, sd_path sd_model_file = os.path.join(script_path, 'model.ckpt') @@ -46,6 +46,7 @@ parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.") parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization") parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find") +parser.add_argument("--use-cpu", nargs='+',choices=['SD', 'GFPGAN', 'ESRGAN', 'SCUNet', 'CodeFormer'], help="use CPU for specified modules", default=[]) parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests") parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None) parser.add_argument("--show-negative-prompt", action='store_true', help="does not do anything", default=False) @@ -63,7 +64,11 @@ parser.add_argument("--enable-console-prompts", action='store_true', help="print cmd_opts = parser.parse_args() -device = get_optimal_device() + +devices.device, devices.device_gfpgan, devices.device_esrgan, devices.device_scunet, devices.device_codeformer = \ +(devices.cpu if x in cmd_opts.use_cpu else devices.get_optimal_device() for x in ['SD', 'GFPGAN', 'ESRGAN', 'SCUNet', 'CodeFormer']) + +device = devices.device batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram) parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram -- cgit v1.2.3 From 27ddc24fdee1fbe709054a43235ab7f9c51b3e9f Mon Sep 17 00:00:00 2001 From: brkirch Date: Tue, 4 Oct 2022 05:18:17 -0400 Subject: Add BSRGAN to --add-cpu --- modules/bsrgan_model.py | 6 +++--- modules/devices.py | 2 +- modules/shared.py | 6 +++--- 3 files changed, 7 insertions(+), 7 deletions(-) (limited to 'modules/devices.py') diff --git a/modules/bsrgan_model.py b/modules/bsrgan_model.py index e62c6657..3bd80791 100644 --- a/modules/bsrgan_model.py +++ b/modules/bsrgan_model.py @@ -8,7 +8,7 @@ import torch from basicsr.utils.download_util import load_file_from_url import modules.upscaler -from modules import shared, modelloader +from modules import devices, modelloader from modules.bsrgan_model_arch import RRDBNet from modules.paths import models_path @@ -44,13 +44,13 @@ class UpscalerBSRGAN(modules.upscaler.Upscaler): model = self.load_model(selected_file) if model is None: return img - model.to(shared.device) + model.to(devices.device_bsrgan) torch.cuda.empty_cache() img = np.array(img) img = img[:, :, ::-1] img = np.moveaxis(img, 2, 0) / 255 img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(shared.device) + img = img.unsqueeze(0).to(devices.device_bsrgan) with torch.no_grad(): output = model(img) output = output.squeeze().float().cpu().clamp_(0, 1).numpy() diff --git a/modules/devices.py b/modules/devices.py index b5a0cd29..b7899632 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -32,7 +32,7 @@ def enable_tf32(): errors.run(enable_tf32, "Enabling TF32") -device = device_gfpgan = device_esrgan = device_scunet = device_codeformer = get_optimal_device() +device = device_gfpgan = device_bsrgan = device_esrgan = device_scunet = device_codeformer = get_optimal_device() dtype = torch.float16 def randn(seed, shape): diff --git a/modules/shared.py b/modules/shared.py index 7899ab8d..95b98a06 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -46,7 +46,7 @@ parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.") parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization") parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find") -parser.add_argument("--use-cpu", nargs='+',choices=['SD', 'GFPGAN', 'ESRGAN', 'SCUNet', 'CodeFormer'], help="use CPU for specified modules", default=[]) +parser.add_argument("--use-cpu", nargs='+',choices=['SD', 'GFPGAN', 'BSRGAN', 'ESRGAN', 'SCUNet', 'CodeFormer'], help="use CPU for specified modules", default=[]) parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests") parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None) parser.add_argument("--show-negative-prompt", action='store_true', help="does not do anything", default=False) @@ -65,8 +65,8 @@ parser.add_argument("--enable-console-prompts", action='store_true', help="print cmd_opts = parser.parse_args() -devices.device, devices.device_gfpgan, devices.device_esrgan, devices.device_scunet, devices.device_codeformer = \ -(devices.cpu if x in cmd_opts.use_cpu else devices.get_optimal_device() for x in ['SD', 'GFPGAN', 'ESRGAN', 'SCUNet', 'CodeFormer']) +devices.device, devices.device_gfpgan, devices.device_bsrgan, devices.device_esrgan, devices.device_scunet, devices.device_codeformer = \ +(devices.cpu if x in cmd_opts.use_cpu else devices.get_optimal_device() for x in ['SD', 'GFPGAN', 'BSRGAN', 'ESRGAN', 'SCUNet', 'CodeFormer']) device = devices.device -- cgit v1.2.3 From 6c6ae28bf5fd1e8bc3e8f64a3430b6f29f338f77 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Tue, 4 Oct 2022 12:32:22 +0300 Subject: send all three of GFPGAN's and codeformer's models to CPU memory instead of just one for #1283 --- modules/codeformer_model.py | 12 ++++++++++-- modules/devices.py | 10 ++++++++++ modules/gfpgan_model.py | 14 ++++++++++++-- modules/processing.py | 16 +++++++++------- 4 files changed, 41 insertions(+), 11 deletions(-) (limited to 'modules/devices.py') diff --git a/modules/codeformer_model.py b/modules/codeformer_model.py index a29f3855..e6d9fa4f 100644 --- a/modules/codeformer_model.py +++ b/modules/codeformer_model.py @@ -69,10 +69,14 @@ def setup_model(dirname): self.net = net self.face_helper = face_helper - self.net.to(devices.device_codeformer) return net, face_helper + def send_model_to(self, device): + self.net.to(device) + self.face_helper.face_det.to(device) + self.face_helper.face_parse.to(device) + def restore(self, np_image, w=None): np_image = np_image[:, :, ::-1] @@ -82,6 +86,8 @@ def setup_model(dirname): if self.net is None or self.face_helper is None: return np_image + self.send_model_to(devices.device_codeformer) + self.face_helper.clean_all() self.face_helper.read_image(np_image) self.face_helper.get_face_landmarks_5(only_center_face=False, resize=640, eye_dist_threshold=5) @@ -113,8 +119,10 @@ def setup_model(dirname): if original_resolution != restored_img.shape[0:2]: restored_img = cv2.resize(restored_img, (0, 0), fx=original_resolution[1]/restored_img.shape[1], fy=original_resolution[0]/restored_img.shape[0], interpolation=cv2.INTER_LINEAR) + self.face_helper.clean_all() + if shared.opts.face_restoration_unload: - self.net.to(devices.cpu) + self.send_model_to(devices.cpu) return restored_img diff --git a/modules/devices.py b/modules/devices.py index ff82f2f6..12aab665 100644 --- a/modules/devices.py +++ b/modules/devices.py @@ -1,3 +1,5 @@ +import contextlib + import torch # has_mps is only available in nightly pytorch (for now), `getattr` for compatibility @@ -57,3 +59,11 @@ def randn_without_seed(shape): return torch.randn(shape, device=device) + +def autocast(): + from modules import shared + + if dtype == torch.float32 or shared.cmd_opts.precision == "full": + return contextlib.nullcontext() + + return torch.autocast("cuda") diff --git a/modules/gfpgan_model.py b/modules/gfpgan_model.py index dd3fbcab..5586b554 100644 --- a/modules/gfpgan_model.py +++ b/modules/gfpgan_model.py @@ -37,22 +37,32 @@ def gfpgann(): print("Unable to load gfpgan model!") return None model = gfpgan_constructor(model_path=model_file, upscale=1, arch='clean', channel_multiplier=2, bg_upsampler=None) - model.gfpgan.to(shared.device) loaded_gfpgan_model = model return model +def send_model_to(model, device): + model.gfpgan.to(device) + model.face_helper.face_det.to(device) + model.face_helper.face_parse.to(device) + + def gfpgan_fix_faces(np_image): model = gfpgann() if model is None: return np_image + + send_model_to(model, devices.device) + np_image_bgr = np_image[:, :, ::-1] cropped_faces, restored_faces, gfpgan_output_bgr = model.enhance(np_image_bgr, has_aligned=False, only_center_face=False, paste_back=True) np_image = gfpgan_output_bgr[:, :, ::-1] + model.face_helper.clean_all() + if shared.opts.face_restoration_unload: - model.gfpgan.to(devices.cpu) + send_model_to(model, devices.cpu) return np_image diff --git a/modules/processing.py b/modules/processing.py index 0a4b6198..9cbecdd8 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -1,4 +1,3 @@ -import contextlib import json import math import os @@ -330,9 +329,8 @@ def process_images(p: StableDiffusionProcessing) -> Processed: infotexts = [] output_images = [] - precision_scope = torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext - ema_scope = (contextlib.nullcontext if cmd_opts.lowvram else p.sd_model.ema_scope) - with torch.no_grad(), precision_scope("cuda"), ema_scope(): + + with torch.no_grad(): p.init(all_prompts, all_seeds, all_subseeds) if state.job_count == -1: @@ -351,8 +349,9 @@ def process_images(p: StableDiffusionProcessing) -> Processed: #uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt]) #c = p.sd_model.get_learned_conditioning(prompts) - uc = prompt_parser.get_learned_conditioning(len(prompts) * [p.negative_prompt], p.steps) - c = prompt_parser.get_learned_conditioning(prompts, p.steps) + with devices.autocast(): + uc = prompt_parser.get_learned_conditioning(len(prompts) * [p.negative_prompt], p.steps) + c = prompt_parser.get_learned_conditioning(prompts, p.steps) if len(model_hijack.comments) > 0: for comment in model_hijack.comments: @@ -361,7 +360,9 @@ def process_images(p: StableDiffusionProcessing) -> Processed: if p.n_iter > 1: shared.state.job = f"Batch {n+1} out of {p.n_iter}" - samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength) + with devices.autocast(): + samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength).to(devices.dtype) + if state.interrupted: # if we are interruped, sample returns just noise @@ -386,6 +387,7 @@ def process_images(p: StableDiffusionProcessing) -> Processed: devices.torch_gc() x_sample = modules.face_restoration.restore_faces(x_sample) + devices.torch_gc() image = Image.fromarray(x_sample) -- cgit v1.2.3