add vae compile to sd-webui

onediff_sd_webui_extensions/README.md

@@ -11,6 +11,11 @@ Updated on DEC 26, 2023. Device: RTX 3090. Resolution: 1024x1024
 | --------------- | --------------- | ------------------ | ---------------------- |
 | 2.99it/s        | 6.40it/s        | 6.71it/s           | 224.41%                |
 
+Time to enerate a 1024x1024 image with sdxl (30 steps) on 3090
+| torch(Baseline) | TensorRT-v9.0.1 | onediff(Optimized) | Percentage improvement |
+| --------------- | --------------- | ------------------ | ---------------------- |
+| 11.03s          | 5.55            | 5.29s              | 208.51%                |
+
 ## Installation Guide
 
 It is recommended to create a Python virtual environment in advance. For example `conda create -n sd-webui python=3.10`.

onediff_sd_webui_extensions/compile_ldm.py

@@ -1,10 +1,11 @@
-import math
+import os
 import oneflow as flow
 from onediff.infer_compiler import oneflow_compile, register
 
 from ldm.modules.attention import BasicTransformerBlock, CrossAttention
 from ldm.modules.diffusionmodules.openaimodel import ResBlock, UNetModel
 from ldm.modules.diffusionmodules.util import GroupNorm32
+from modules import shared
 from sd_webui_onediff_utils import (
     CrossAttentionOflow,
     GroupNorm32Oflow,
@@ -59,3 +60,19 @@ def compile_ldm_unet(unet_model, *, use_graph=True, options={}):
         if isinstance(module, ResBlock):
             module.use_checkpoint = False
     return oneflow_compile(unet_model, use_graph=use_graph, options=options)
+
+
+class SD21CompileCtx(object):
+    """to avoid results for NaN when the model is v2-1_768-ema-pruned"""
+
+    _var_name = "ONEFLOW_ATTENTION_ALLOW_HALF_PRECISION_ACCUMULATION"
+
+    def __enter__(self):
+        self._original = os.getenv(self._var_name)
+        if shared.opts.sd_model_checkpoint.startswith("v2-1"):
+            os.environ[self._var_name] = "0"
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self._original is not None:
+            os.environ[self._var_name] = self._original
+        return False

onediff_sd_webui_extensions/compile_vae.py

@@ -0,0 +1,38 @@
+from modules import shared
+from modules.sd_vae_approx import model as get_vae_model, sd_vae_approx_models
+from onediff.infer_compiler import oneflow_compile
+
+__all__ = ["VaeCompileCtx"]
+
+
+compiled_models = {}
+
+
+class VaeCompileCtx(object):
+    def __init__(self, use_graph=True, options={}):
+        self._use_graph = use_graph
+        self._options = options
+        # https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/75336dfc84cae280036bc52a6805eb10d9ae30ba/modules/sd_vae_approx.py#L43
+        self._model_name = (
+            "vaeapprox-sdxl.pt"
+            if getattr(shared.sd_model, "is_sdxl", False)
+            else "model.pt"
+        )
+        self._original_model = get_vae_model()
+
+    def __enter__(self):
+        if self._original_model is None:
+            return
+        global compiled_models
+        model = compiled_models.get(self._model_name)
+        if model is None:
+            model = oneflow_compile(
+                self._original_model, use_graph=self._use_graph, options=self._options
+            )
+            compiled_models[self._model_name] = model
+        sd_vae_approx_models[self._model_name] = model
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self._original_model is not None:
+            sd_vae_approx_models[self._model_name] = self._original_model
+        return False

onediff_sd_webui_extensions/scripts/onediff.py

@@ -5,8 +5,9 @@
 import modules.shared as shared
 from modules.processing import process_images
 
+from compile_ldm import compile_ldm_unet, SD21CompileCtx
 from compile_sgm import compile_sgm_unet
-from compile_ldm import compile_ldm_unet
+from compile_vae import VaeCompileCtx
 
 from onediff.optimization.quant_optimizer import (
     quantize_model,
@@ -33,11 +34,7 @@ def is_compiled(ckpt_name):
 
 
 def compile_unet(
-    unet_model,
-    quantization=False,
-    *,
-    use_graph=True,
-    options={},
+    unet_model, quantization=False, *, use_graph=True, options={},
 ):
     from ldm.modules.diffusionmodules.openaimodel import UNetModel as UNetModelLDM
     from sgm.modules.diffusionmodules.openaimodel import UNetModel as UNetModelSGM
@@ -57,6 +54,22 @@ def compile_unet(
         return unet_model
 
 
+class UnetCompileCtx(object):
+    """The unet model is stored in a global variable.
+    The global variables need to be replaced with compiled_unet before process_images is run,
+    and then the original model restored so that subsequent reasoning with onediff disabled meets expectations.
+    """
+
+    def __enter__(self):
+        self._original_model = shared.sd_model.model.diffusion_model
+        global compiled_unet
+        shared.sd_model.model.diffusion_model = compiled_unet
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        shared.sd_model.model.diffusion_model = self._original_model
+        return False
+
+
 class Script(scripts.Script):
     def title(self):
         return "onediff_diffusion_model"
@@ -68,7 +81,7 @@ def ui(self, is_img2img):
         """
         if not varify_can_use_quantization():
             ret = gr.HTML(
-                    """
+                """
                     <div style="padding: 20px; border: 1px solid #e0e0e0; border-radius: 5px; background-color: #f9f9f9;">
                         <div style="font-size: 18px; font-weight: bold; margin-bottom: 15px; color: #31708f;">
                             Hints Message
@@ -88,7 +101,7 @@ def ui(self, is_img2img):
                         </p>
                     </div>
                     """
-                )
+            )
 
         else:
             ret = gr.components.Checkbox(label="Model Quantization(int8) Speed Up")
@@ -107,22 +120,11 @@ def run(self, p, quantization=False):
         )
 
         if not is_compiled(ckpt_name):
-            graph_file = generate_graph_path(
-                ckpt_name, original_diffusion_model.__class__.__name__
-            )
-            graph_file_device = shared.device
-            compile_options = {
-                "graph_file_device": graph_file_device,
-                "graph_file": graph_file,
-            }
             compiled_unet = compile_unet(
-                original_diffusion_model,
-                quantization=quantization,
-                options=compile_options,
+                original_diffusion_model, quantization=quantization,
             )
             compiled_ckpt_name = ckpt_name
 
-        shared.sd_model.model.diffusion_model = compiled_unet
-        proc = process_images(p)
-        shared.sd_model.model.diffusion_model = original_diffusion_model
+        with UnetCompileCtx(), VaeCompileCtx(), SD21CompileCtx():
+            proc = process_images(p)
         return proc

src/onediff/optimization/quant_optimizer.py

@@ -16,7 +16,7 @@
 def varify_can_use_quantization():
     if not is_quantization_enabled():
         message = get_support_message()
-        logger.error(message)
+        logger.warn(message)
         return False
     return True