add sd3 to starter models

invokeai/app/services/config/config_default.py

@@ -105,7 +105,6 @@ class InvokeAIAppConfig(BaseSettings):
         vram: Amount of VRAM reserved for model storage (GB).
         convert_cache: Maximum size of on-disk converted models cache (GB).
         lazy_offload: Keep models in VRAM until their space is needed.
-        load_sd3_encoder_3: Load the memory-intensive SD3 text_encoder_3.
         log_memory_usage: If True, a memory snapshot will be captured before and after every model cache operation, and the result will be logged (at debug level). There is a time cost to capturing the memory snapshots, so it is recommended to only enable this feature if you are actively inspecting the model cache's behaviour.
         device: Preferred execution device. `auto` will choose the device depending on the hardware platform and the installed torch capabilities.<br>Valid values: `auto`, `cpu`, `cuda`, `cuda:1`, `mps`
         precision: Floating point precision. `float16` will consume half the memory of `float32` but produce slightly lower-quality images. The `auto` setting will guess the proper precision based on your video card and operating system.<br>Valid values: `auto`, `float16`, `bfloat16`, `float32`
@@ -176,7 +175,6 @@ class InvokeAIAppConfig(BaseSettings):
     vram:                         float = Field(default=DEFAULT_VRAM_CACHE, ge=0, description="Amount of VRAM reserved for model storage (GB).")
     convert_cache:                float = Field(default=DEFAULT_CONVERT_CACHE, ge=0, description="Maximum size of on-disk converted models cache (GB).")
     lazy_offload:                  bool = Field(default=True,               description="Keep models in VRAM until their space is needed.")
-    load_sd3_encoder_3:            bool = Field(default=False,              description="Load the memory-intensive SD3 text_encoder_3.")
     log_memory_usage:              bool = Field(default=False,              description="If True, a memory snapshot will be captured before and after every model cache operation, and the result will be logged (at debug level). There is a time cost to capturing the memory snapshots, so it is recommended to only enable this feature if you are actively inspecting the model cache's behaviour.")
 
     # DEVICE

invokeai/backend/model_manager/load/model_cache/model_cache_default.py

@@ -241,9 +241,15 @@ class ModelCache(ModelCacheBase[AnyModel]):
             if vram_in_use <= reserved:
                 break
 
-            # only way to remove a quantized model from VRAM is to
+            # Special handling of the stable-diffusion-3:text_encoder_3
+            # submodel, when the user has loaded a quantized model.
+            # The only way to remove the quantized version of this model from VRAM is to
             # delete it completely - it can't be moved from device to device
+            # This also contains a workaround for quantized models that
+            # persist indefinitely in VRAM
             if cache_entry.is_quantized:
+                self._empty_quantized_state_dict(cache_entry.model)
+                cache_entry.model = None
                 self._delete_cache_entry(cache_entry)
                 vram_in_use = torch.cuda.memory_allocated() + size_required
                 continue
@@ -426,3 +432,17 @@ class ModelCache(ModelCacheBase[AnyModel]):
         del cache_entry
         gc.collect()
         TorchDevice.empty_cache()
+
+    def _empty_quantized_state_dict(self, model: AnyModel) -> None:
+        """Set all keys of a model's state dict to None.
+
+        This is a partial workaround for a poorly-understood bug in
+        transformers' support for quantized T5EncoderModels (text_encoder_3
+        of SD3). This allows most of the model to be unloaded from VRAM, but
+        still leaks 8K of VRAM each time the model is unloaded. Using the quantized
+        version of stable-diffusion-3-medium is NOT recommended.
+        """
+        assert isinstance(model, torch.nn.Module)
+        sd = model.state_dict()
+        for k in sd.keys():
+            sd[k] = None

invokeai/backend/model_manager/load/model_loaders/stable_diffusion.py

@@ -40,7 +40,7 @@ class StableDiffusionDiffusersModel(GenericDiffusersLoader):
     model_base_to_model_type = {
         BaseModelType.StableDiffusion1: "FrozenCLIPEmbedder",
         BaseModelType.StableDiffusion2: "FrozenOpenCLIPEmbedder",
-        BaseModelType.StableDiffusion3: "SD3",  # non-functional, for completeness only
+        BaseModelType.StableDiffusion3: "SD3",
         BaseModelType.StableDiffusionXL: "SDXL",
         BaseModelType.StableDiffusionXLRefiner: "SDXL-Refiner",
     }

invokeai/backend/model_manager/starter_models.py

@@ -122,6 +122,13 @@ STARTER_MODELS: list[StarterModel] = [
         type=ModelType.Main,
         dependencies=[sdxl_fp16_vae_fix],
     ),
+    StarterModel(
+        name="Stable Diffusion 3",
+        base=BaseModelType.StableDiffusion3,
+        source="stabilityai/stable-diffusion-3-medium-diffusers",
+        description="The OG Stable Diffusion 3 base model (beta).",
+        type=ModelType.Main,
+    ),
     # endregion
     # region VAE
     sdxl_fp16_vae_fix,

pyproject.toml

@@ -34,26 +34,26 @@ classifiers = [
 dependencies = [
   # Core generation dependencies, pinned for reproducible builds.
   "accelerate==0.30.1",
-  "bitsandbytes",
+  "bitsandbytes==0.43.1",
   "clip_anytorch==2.6.0",       # replacing "clip @ https://github.com/openai/CLIP/archive/eaa22acb90a5876642d0507623e859909230a52d.zip",
   "compel==2.0.2",
   "controlnet-aux==0.0.7",
-  "diffusers[torch]",
+  "diffusers[torch]==0.29.0",
   "invisible-watermark==0.2.0", # needed to install SDXL base and refiner using their repo_ids
   "mediapipe==0.10.7",          # needed for "mediapipeface" controlnet model
-  "numpy==1.26.4",              # >1.24.0 is needed to use the 'strict' argument to np.testing.assert_array_equal()
+  "numpy==1.23.5",              # >1.24.0 is needed to use the 'strict' argument to np.testing.assert_array_equal()
   "onnx==1.15.0",
   "onnxruntime==1.16.3",
   "opencv-python==4.9.0.80",
-  "pytorch-lightning==2.1.3",
+  "pytorch-lightning",
   "safetensors==0.4.3",
   "timm==0.6.13",               # needed to override timm latest in controlnet_aux, see  https://github.com/isl-org/ZoeDepth/issues/26
-  "torch",
+  "torch==2.2.2",
   "torchmetrics==0.11.4",
   "torchsde==0.2.6",
-  "torchvision",
-  "transformers",
-  "sentencepiece",
+  "torchvision==0.17.2",
+  "transformers==4.41.1",
+  "sentencepiece==0.1.99",
 
   # Core application dependencies, pinned for reproducible builds.
   "fastapi-events==0.11.0",