feat(api): enable optimizations for SD pipelines based on env vars (#155)

api/onnx_web/chain/upscale_stable_diffusion.py

@@ -5,6 +5,7 @@ import torch
 from diffusers import StableDiffusionUpscalePipeline
 from PIL import Image
 
+from ..diffusion.load import optimize_pipeline
 from ..diffusion.pipeline_onnx_stable_diffusion_upscale import (
     OnnxStableDiffusionUpscalePipeline,
 )
@@ -52,6 +53,8 @@ def load_stable_diffusion(
     if not server.show_progress:
         pipe.set_progress_bar_config(disable=True)
 
+    optimize_pipeline(server, pipe)
+
     server.cache.set("diffusion", cache_key, pipe)
     run_gc([device])
 

api/onnx_web/diffusion/load.py

@@ -17,6 +17,7 @@ from diffusers import (
     KDPM2DiscreteScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
+    StableDiffusionPipeline,
 )
 
 try:
@@ -87,6 +88,32 @@ def get_tile_latents(
     return full_latents[:, :, y:yt, x:xt]
 
 
+def optimize_pipeline(
+    server: ServerContext,
+    pipe: StableDiffusionPipeline,
+) -> None:
+    if "attention-slicing" in server.optimizations:
+        logger.debug("enabling attention slicing on SD pipeline")
+        pipe.enable_attention_slicing()
+
+    if "vae-slicing" in server.optimizations:
+        logger.debug("enabling VAE slicing on SD pipeline")
+        pipe.enable_vae_slicing()
+
+    if "sequential-cpu-offload" in server.optimizations:
+        logger.debug("enabling sequential CPU offload on SD pipeline")
+        pipe.enable_sequential_cpu_offload()
+    elif "model-cpu-offload" in server.optimizations:
+        # TODO: check for accelerate
+        logger.debug("enabling model CPU offload on SD pipeline")
+        pipe.enable_model_cpu_offload()
+
+    if "memory-efficient-attention" in server.optimizations:
+        # TODO: check for xformers
+        logger.debug("enabling memory efficient attention for SD pipeline")
+        pipe.enable_xformers_memory_efficient_attention()
+
+
 def load_pipeline(
     server: ServerContext,
     pipeline: DiffusionPipeline,
@@ -151,6 +178,8 @@ def load_pipeline(
         if not server.show_progress:
             pipe.set_progress_bar_config(disable=True)
 
+        optimize_pipeline(server, pipe)
+
         if device is not None and hasattr(pipe, "to"):
             pipe = pipe.to(device.torch_str())
 

api/onnx_web/utils.py

@@ -28,6 +28,7 @@ class ServerContext:
         cache: ModelCache = None,
         cache_path: str = None,
         show_progress: bool = True,
+        optimizations: List[str] = [],
     ) -> None:
         self.bundle_path = bundle_path
         self.model_path = model_path
@@ -42,6 +43,7 @@ class ServerContext:
         self.cache = cache or ModelCache(num_workers)
         self.cache_path = cache_path or path.join(model_path, ".cache")
         self.show_progress = show_progress
+        self.optimizations = optimizations
 
     @classmethod
     def from_environ(cls):
@@ -64,6 +66,7 @@ class ServerContext:
             image_format=environ.get("ONNX_WEB_IMAGE_FORMAT", "png"),
             cache=ModelCache(limit=cache_limit),
             show_progress=get_boolean(environ, "ONNX_WEB_SHOW_PROGRESS", True),
+            optimizations=environ.get("ONNX_WEB_OPTIMIZATIONS", "").split(","),
         )