Merge branch 'main' into req-xformers

.github/workflows/build-cloud-img.yml

@@ -21,6 +21,7 @@ env:
 
 jobs:
   docker:
+    if: github.event.pull_request.draft == false
     strategy:
       fail-fast: false
       matrix:

.github/workflows/build-container.yml

@@ -8,6 +8,7 @@ on:
 
 jobs:
   docker:
+    if: github.event.pull_request.draft == false
     strategy:
       fail-fast: false
       matrix:

.github/workflows/lint-frontend.yml

@@ -14,6 +14,7 @@ defaults:
 
 jobs:
   lint-frontend:
+    if: github.event.pull_request.draft == false
     runs-on: ubuntu-22.04
     steps:
       - name: Setup Node 18

.github/workflows/mkdocs-material.yml

@@ -7,6 +7,7 @@ on:
 
 jobs:
   mkdocs-material:
+    if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     steps:
       - name: checkout sources

.github/workflows/pyflakes.yml

@@ -9,6 +9,7 @@ on:
 jobs:
   pyflakes:
     name: runner / pyflakes
+    if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v2

ldm/generate.py

@@ -20,6 +20,7 @@ import torch
 import transformers
 from PIL import Image, ImageOps
 from diffusers.pipeline_utils import DiffusionPipeline
+from diffusers.utils.import_utils import is_xformers_available
 from omegaconf import OmegaConf
 from pytorch_lightning import seed_everything, logging
 
@@ -203,6 +204,14 @@ class Generate:
             self.precision = choose_precision(self.device)
         Globals.full_precision = self.precision=='float32'
 
+        if is_xformers_available():
+            if not Globals.disable_xformers:
+                print('>> xformers memory-efficient attention is available and enabled')
+            else:
+                print('>> xformers memory-efficient attention is available but disabled')
+        else:
+            print('>> xformers not installed')
+            
         # model caching system for fast switching
         self.model_manager = ModelManager(mconfig,self.device,self.precision,max_loaded_models=max_loaded_models)
         # don't accept invalid models

ldm/invoke/generator/diffusers_pipeline.py

@@ -597,11 +597,20 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
             output = InvokeAIStableDiffusionPipelineOutput(images=image, nsfw_content_detected=[], attention_map_saver=result_attention_maps)
             return self.check_for_safety(output, dtype=conditioning_data.dtype)
 
-    def non_noised_latents_from_image(self, init_image, *, device, dtype):
+    def non_noised_latents_from_image(self, init_image, *, device: torch.device, dtype):
         init_image = init_image.to(device=device, dtype=dtype)
         with torch.inference_mode():
+            if device.type == 'mps':
+                # workaround for torch MPS bug that has been fixed in https://github.com/kulinseth/pytorch/pull/222
+                # TODO remove this workaround once kulinseth#222 is merged to pytorch mainline
+                self.vae.to('cpu')
+                init_image = init_image.to('cpu')
             init_latent_dist = self.vae.encode(init_image).latent_dist
             init_latents = init_latent_dist.sample().to(dtype=dtype)  # FIXME: uses torch.randn. make reproducible!
+            if device.type == 'mps':
+                self.vae.to(device)
+                init_latents = init_latents.to(device)
+
         init_latents = 0.18215 * init_latents
         return init_latents