Significantly reduce thread abuse for faster model moving

This will move all major gradio calls into the main thread rather than random gradio threads.
This ensures that all torch.module.to() are performed in main thread to completely possible avoid GPU fragments.
In my test now model moving is 0.7 ~ 1.2 seconds faster, which means all 6GB/8GB VRAM users will get 0.7 ~ 1.2 seconds faster per image on SDXL.

modules/img2img.py

@@ -15,6 +15,7 @@
 import modules.processing as processing
 from modules.ui import plaintext_to_html
 import modules.scripts
+from modules_forge import main_thread
 
 
 def process_batch(p, input_dir, output_dir, inpaint_mask_dir, args, to_scale=False, scale_by=1.0, use_png_info=False, png_info_props=None, png_info_dir=None):
@@ -146,7 +147,7 @@ def process_batch(p, input_dir, output_dir, inpaint_mask_dir, args, to_scale=Fal
     return batch_results
 
 
-def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps: int, sampler_name: str, mask_blur: int, mask_alpha: float, inpainting_fill: int, n_iter: int, batch_size: int, cfg_scale: float, image_cfg_scale: float, denoising_strength: float, selected_scale_tab: int, height: int, width: int, scale_by: float, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, img2img_batch_inpaint_mask_dir: str, override_settings_texts, img2img_batch_use_png_info: bool, img2img_batch_png_info_props: list, img2img_batch_png_info_dir: str, request: gr.Request, *args):
+def img2img_function(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps: int, sampler_name: str, mask_blur: int, mask_alpha: float, inpainting_fill: int, n_iter: int, batch_size: int, cfg_scale: float, image_cfg_scale: float, denoising_strength: float, selected_scale_tab: int, height: int, width: int, scale_by: float, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, img2img_batch_inpaint_mask_dir: str, override_settings_texts, img2img_batch_use_png_info: bool, img2img_batch_png_info_props: list, img2img_batch_png_info_dir: str, request: gr.Request, *args):
     override_settings = create_override_settings_dict(override_settings_texts)
 
     is_batch = mode == 5
@@ -244,3 +245,7 @@ def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_s
         processed.images = []
 
     return processed.images + processed.extra_images, generation_info_js, plaintext_to_html(processed.info), plaintext_to_html(processed.comments, classname="comments")
+
+
+def img2img(id_task: str, mode: int, prompt: str, negative_prompt: str, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps: int, sampler_name: str, mask_blur: int, mask_alpha: float, inpainting_fill: int, n_iter: int, batch_size: int, cfg_scale: float, image_cfg_scale: float, denoising_strength: float, selected_scale_tab: int, height: int, width: int, scale_by: float, resize_mode: int, inpaint_full_res: bool, inpaint_full_res_padding: int, inpainting_mask_invert: int, img2img_batch_input_dir: str, img2img_batch_output_dir: str, img2img_batch_inpaint_mask_dir: str, override_settings_texts, img2img_batch_use_png_info: bool, img2img_batch_png_info_props: list, img2img_batch_png_info_dir: str, request: gr.Request, *args):
+    return main_thread.run_and_wait_result(img2img_function, id_task, mode, prompt, negative_prompt, prompt_styles, init_img, sketch, init_img_with_mask, inpaint_color_sketch, inpaint_color_sketch_orig, init_img_inpaint, init_mask_inpaint, steps, sampler_name, mask_blur, mask_alpha, inpainting_fill, n_iter, batch_size, cfg_scale, image_cfg_scale, denoising_strength, selected_scale_tab, height, width, scale_by, resize_mode, inpaint_full_res, inpaint_full_res_padding, inpainting_mask_invert, img2img_batch_input_dir, img2img_batch_output_dir, img2img_batch_inpaint_mask_dir, override_settings_texts, img2img_batch_use_png_info, img2img_batch_png_info_props, img2img_batch_png_info_dir, request, *args)

modules/initialize.py

@@ -149,24 +149,9 @@ def initialize_rest(*, reload_script_modules=False):
     sd_unet.list_unets()
     startup_timer.record("scripts list_unets")
 
-    def load_model():
-        """
-        Accesses shared.sd_model property to load model.
-        After it's available, if it has been loaded before this access by some extension,
-        its optimization may be None because the list of optimizaers has neet been filled
-        by that time, so we apply optimization again.
-        """
-        from modules import devices
-        devices.torch_npu_set_device()
-
-        shared.sd_model  # noqa: B018
-
-        if sd_hijack.current_optimizer is None:
-            sd_hijack.apply_optimizations()
-
-        devices.first_time_calculation()
-    if not shared.cmd_opts.skip_load_model_at_start:
-        Thread(target=load_model).start()
+    from modules_forge import main_thread
+    import modules.sd_models
+    main_thread.async_run(modules.sd_models.model_data.get_sd_model)
 
     from modules import shared_items
     shared_items.reload_hypernetworks()

modules/initialize_util.py

@@ -170,10 +170,11 @@ def sigint_handler(sig, frame):
 def configure_opts_onchange():
     from modules import shared, sd_models, sd_vae, ui_tempdir, sd_hijack
     from modules.call_queue import wrap_queued_call
+    from modules_forge import main_thread
 
-    shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: sd_models.reload_model_weights()), call=False)
-    shared.opts.onchange("sd_vae", wrap_queued_call(lambda: sd_vae.reload_vae_weights()), call=False)
-    shared.opts.onchange("sd_vae_overrides_per_model_preferences", wrap_queued_call(lambda: sd_vae.reload_vae_weights()), call=False)
+    shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: main_thread.run_and_wait_result(sd_models.reload_model_weights)), call=False)
+    shared.opts.onchange("sd_vae", wrap_queued_call(lambda: main_thread.run_and_wait_result(sd_vae.reload_vae_weights)), call=False)
+    shared.opts.onchange("sd_vae_overrides_per_model_preferences", wrap_queued_call(lambda: main_thread.run_and_wait_result(sd_vae.reload_vae_weights)), call=False)
     shared.opts.onchange("temp_dir", ui_tempdir.on_tmpdir_changed)
     shared.opts.onchange("gradio_theme", shared.reload_gradio_theme)
     shared.opts.onchange("cross_attention_optimization", wrap_queued_call(lambda: sd_hijack.model_hijack.redo_hijack(shared.sd_model)), call=False)

modules/launch_utils.py

@@ -511,6 +511,11 @@ def start():
     else:
         webui.webui()
 
+    from modules_forge import main_thread
+
+    main_thread.loop()
+    return
+
 
 def dump_sysinfo():
     from modules import sysinfo

modules/shared_state.py

@@ -2,6 +2,8 @@
 import logging
 import threading
 import time
+import traceback
+import torch
 
 from modules import errors, shared, devices
 from typing import Optional
@@ -134,6 +136,7 @@ def end(self):
 
         devices.torch_gc()
 
+    @torch.inference_mode()
     def set_current_image(self):
         """if enough sampling steps have been made after the last call to this, sets self.current_image from self.current_latent, and modifies self.id_live_preview accordingly"""
         if not shared.parallel_processing_allowed:
@@ -142,6 +145,7 @@ def set_current_image(self):
         if self.sampling_step - self.current_image_sampling_step >= shared.opts.show_progress_every_n_steps and shared.opts.live_previews_enable and shared.opts.show_progress_every_n_steps != -1:
             self.do_set_current_image()
 
+    @torch.inference_mode()
     def do_set_current_image(self):
         if self.current_latent is None:
             return
@@ -156,11 +160,14 @@ def do_set_current_image(self):
 
             self.current_image_sampling_step = self.sampling_step
 
-        except Exception:
+        except Exception as e:
+            # traceback.print_exc()
+            # print(e)
             # when switching models during genration, VAE would be on CPU, so creating an image will fail.
             # we silently ignore this error
             errors.record_exception()
 
+    @torch.inference_mode()
     def assign_current_image(self, image):
         self.current_image = image
         self.id_live_preview += 1

modules/txt2img.py

@@ -9,6 +9,7 @@
 from modules.ui import plaintext_to_html
 from PIL import Image
 import gradio as gr
+from modules_forge import main_thread
 
 
 def txt2img_create_processing(id_task: str, request: gr.Request, prompt: str, negative_prompt: str, prompt_styles, steps: int, sampler_name: str, n_iter: int, batch_size: int, cfg_scale: float, height: int, width: int, enable_hr: bool, denoising_strength: float, hr_scale: float, hr_upscaler: str, hr_second_pass_steps: int, hr_resize_x: int, hr_resize_y: int, hr_checkpoint_name: str, hr_sampler_name: str, hr_prompt: str, hr_negative_prompt, override_settings_texts, *args, force_enable_hr=False):
@@ -56,7 +57,7 @@ def txt2img_create_processing(id_task: str, request: gr.Request, prompt: str, ne
     return p
 
 
-def txt2img_upscale(id_task: str, request: gr.Request, gallery, gallery_index, generation_info, *args):
+def txt2img_upscale_function(id_task: str, request: gr.Request, gallery, gallery_index, generation_info, *args):
     assert len(gallery) > 0, 'No image to upscale'
     assert 0 <= gallery_index < len(gallery), f'Bad image index: {gallery_index}'
 
@@ -100,7 +101,7 @@ def txt2img_upscale(id_task: str, request: gr.Request, gallery, gallery_index, g
     return new_gallery, json.dumps(geninfo), plaintext_to_html(processed.info), plaintext_to_html(processed.comments, classname="comments")
 
 
-def txt2img(id_task: str, request: gr.Request, *args):
+def txt2img_function(id_task: str, request: gr.Request, *args):
     p = txt2img_create_processing(id_task, request, *args)
 
     with closing(p):
@@ -119,3 +120,11 @@ def txt2img(id_task: str, request: gr.Request, *args):
         processed.images = []
 
     return processed.images + processed.extra_images, generation_info_js, plaintext_to_html(processed.info), plaintext_to_html(processed.comments, classname="comments")
+
+
+def txt2img_upscale(id_task: str, request: gr.Request, gallery, gallery_index, generation_info, *args):
+    return main_thread.run_and_wait_result(txt2img_upscale_function, id_task, request, gallery, gallery_index, generation_info, *args)
+
+
+def txt2img(id_task: str, request: gr.Request, *args):
+    return main_thread.run_and_wait_result(txt2img_function, id_task, request, *args)

modules_forge/main_thread.py

@@ -0,0 +1,68 @@
+# This file is the main thread that handles all gradio calls for major t2i or i2i processing.
+# Other gradio calls (like those from extensions) are not influenced.
+# By using one single thread to process all major calls, model moving is significantly faster.
+
+
+import time
+import traceback
+import threading
+
+
+lock = threading.Lock()
+last_id = 0
+waiting_list = []
+finished_list = []
+
+
+class Task:
+    def __init__(self, task_id, func, args, kwargs):
+        self.task_id = task_id
+        self.func = func
+        self.args = args
+        self.kwargs = kwargs
+        self.result = None
+
+    def work(self):
+        self.result = self.func(*self.args, **self.kwargs)
+
+
+def loop():
+    global lock, last_id, waiting_list, finished_list
+    while True:
+        time.sleep(0.01)
+        if len(waiting_list) > 0:
+            with lock:
+                task = waiting_list.pop(0)
+            try:
+                task.work()
+            except Exception as e:
+                traceback.print_exc()
+                print(e)
+            with lock:
+                finished_list.append(task)
+
+
+def async_run(func, *args, **kwargs):
+    global lock, last_id, waiting_list, finished_list
+    with lock:
+        last_id += 1
+        new_task = Task(task_id=last_id, func=func, args=args, kwargs=kwargs)
+        waiting_list.append(new_task)
+    return new_task.task_id
+
+
+def run_and_wait_result(func, *args, **kwargs):
+    global lock, last_id, waiting_list, finished_list
+    current_id = async_run(func, *args, **kwargs)
+    while True:
+        time.sleep(0.01)
+        finished_task = None
+        for t in finished_list.copy():  # thread safe shallow copy without needing a lock
+            if t.task_id == current_id:
+                finished_task = t
+                break
+        if finished_task is not None:
+            with lock:
+                finished_list.remove(finished_task)
+            return finished_task.result
+

webui.py

@@ -6,8 +6,10 @@
 from modules import timer
 from modules import initialize_util
 from modules import initialize
-
+from threading import Thread
 from modules_forge.initialization import initialize_forge
+from modules_forge import main_thread
+
 
 startup_timer = timer.startup_timer
 startup_timer.record("launcher")
@@ -18,6 +20,8 @@
 
 initialize.check_versions()
 
+initialize.initialize()
+
 
 def create_api(app):
     from modules.api.api import Api
@@ -27,12 +31,10 @@ def create_api(app):
     return api
 
 
-def api_only():
+def api_only_worker():
     from fastapi import FastAPI
     from modules.shared_cmd_options import cmd_opts
 
-    initialize.initialize()
-
     app = FastAPI()
     initialize_util.setup_middleware(app)
     api = create_api(app)
@@ -49,11 +51,10 @@ def api_only():
     )
 
 
-def webui():
+def webui_worker():
     from modules.shared_cmd_options import cmd_opts
 
     launch_api = cmd_opts.api
-    initialize.initialize()
 
     from modules import shared, ui_tempdir, script_callbacks, ui, progress, ui_extra_networks
 
@@ -157,10 +158,20 @@ def webui():
         initialize.initialize_rest(reload_script_modules=True)
 
 
+def api_only():
+    Thread(target=api_only_worker, daemon=True).start()
+
+
+def webui():
+    Thread(target=webui_worker, daemon=True).start()
+
+
 if __name__ == "__main__":
     from modules.shared_cmd_options import cmd_opts
 
     if cmd_opts.nowebui:
         api_only()
     else:
         webui()
+
+    main_thread.loop()