fix sdxl

core/config/api_settings.py

@@ -40,7 +40,7 @@ class APIConfig:
     data_type: Literal[
         "float32", "float16", "bfloat16", "float8_e4m3fn", "float8_e5m2"
     ] = "float16"
-    use_minimal_sdxl_pipeline: bool = True  # slower, but works better
+    use_minimal_sdxl_pipeline: bool = False  # slower, but works better
 
     # VRAM optimizations
     # whether to run both parts of CFG>1 generations in one call. Increases VRAM usage during inference,

core/inference/sdxl/pipeline.py

@@ -646,8 +646,8 @@ def do_denoise(
                 # perform guidance
                 if do_classifier_free_guidance:
                     if not split_latents_into_two:
-                        if isinstance(noise_pred, list):
-                            noise_pred = noise_pred[0]
+                        # if isinstance(noise_pred, list):  # type: ignore
+                        #     noise_pred = noise_pred[0]
                         noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)  # type: ignore
                     noise_pred = calculate_cfg(
                         j, noise_pred_text, noise_pred_uncond, guidance_scale, t, additional_pred=noise_pred_vanilla  # type: ignore

core/inference/utilities/convert_from_ckpt.py

@@ -1148,7 +1148,6 @@ def download_from_original_stable_diffusion_ckpt(
             model_type = "SDXL-Refiner"
         if image_size is None:
             image_size = 1024
-    print(model_type)
 
     # Check if we have a SDXL or SD model and initialize default pipeline
     pipeline_class = StableDiffusionPipeline  # type: ignore
@@ -1237,7 +1236,10 @@ def download_from_original_stable_diffusion_ckpt(
     )
 
     with init_empty_weights():
-        if model_type in ["SDXL", "SDXL-Refiner"] and volta_config.api.use_minimal_sdxl_pipeline:
+        if (
+            model_type in ["SDXL", "SDXL-Refiner"]
+            and volta_config.api.use_minimal_sdxl_pipeline
+        ):
             unet = SDXLUNet2D()
         else:
             unet = UNet2DConditionModel(**unet_config)

core/inference/utilities/latents.py

@@ -1,3 +1,4 @@
+from contextlib import ExitStack
 import logging
 import math
 from time import time
@@ -18,6 +19,7 @@
 from core.inference.utilities.philox import PhiloxGenerator
 
 from .random import randn
+from core.optimizations.autocast_utils import autocast
 
 logger = logging.getLogger(__name__)
 
@@ -315,8 +317,14 @@ def prepare_latents(
     else:
         if image.shape[1] != 4:
             image = pad_tensor(image, pipe.vae_scale_factor)
-            init_latent_dist = pipe.vae.encode(image.to(config.api.device, dtype=pipe.vae.dtype)).latent_dist  # type: ignore
-            init_latents = init_latent_dist.sample(generator=generator)
+            with ExitStack() as gs:
+                if pipe.vae.config["force_upcast"] or config.api.upcast_vae:
+                    gs.enter_context(autocast(dtype=torch.float32))
+                init_latent_dist = pipe.vae.encode(image.to(config.api.device, dtype=pipe.vae.dtype)).latent_dist  # type: ignore
+
+                if pipe.vae.config["force_upcast"] or config.api.upcast_vae:
+                    gs.enter_context(autocast(dtype=config.api.load_dtype))
+            init_latents = init_latent_dist.sample(generator=generator)  # type: ignore
             init_latents = 0.18215 * init_latents
             init_latents = torch.cat([init_latents] * batch_size, dim=0)  # type: ignore
         else: