add onediff_comfy_nodes/benchmarks

.github/workflows/examples.yml

@@ -282,7 +282,7 @@ jobs:
       - name: Start ComfyUI Web Service
         if: matrix.test-suite == 'comfy'
         run: |
-          docker exec -w /app/ComfyUI -d ${{ env.CONTAINER_NAME }} sh -c "python3 /app/ComfyUI/main.py --port 8188 --extra-model-paths-config /src/onediff/tests/comfyui/extra_model_paths.yaml > /app/ComfyUI/onediff_comfyui.log 2>&1"
+          docker exec -w /app/ComfyUI -d ${{ env.CONTAINER_NAME }} sh -c "python3 /app/ComfyUI/main.py --gpu-only --disable-cuda-malloc --port 8188 --extra-model-paths-config /src/onediff/tests/comfyui/extra_model_paths.yaml > /app/ComfyUI/onediff_comfyui.log 2>&1"
           sleep 30
       # print to check if comfy is launched successfully
       - run: docker exec ${{ env.CONTAINER_NAME }} ps aux
@@ -299,12 +299,18 @@ jobs:
             }
           }
 
-          run_comfy_test "workflows/sdxl-unet-speedup-graph-saver.json" 200
-          run_comfy_test "workflows/sdxl-control-lora-speedup.json" 200
-          run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/ipadapter_advanced.json" 200
-          run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/deep-cache.json" 600
-          run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/deep-cache-with-lora.json" 800
+          # run_comfy_test "workflows/sdxl-unet-speedup-graph-saver.json" 200
+          # run_comfy_test "workflows/sdxl-control-lora-speedup.json" 200
+          # run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/ipadapter_advanced.json" 200
+          # run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/deep-cache.json" 600
+          # run_comfy_test "/share_nfs/hf_models/comfyui_resources/workflows/deep-cache-with-lora.json" 800
           # run_comfy_test "workflows/text-to-video-speedup.json" 5000
+          docker exec -w /src/onediff/onediff_comfy_nodes/benchmarks ${{ env.CONTAINER_NAME }} bash scripts/install_env.sh /app/ComfyUI
+          docker exec -w /src/onediff/onediff_comfy_nodes/benchmarks ${{ env.CONTAINER_NAME }} bash scripts/run_oneflow_case.sh || {
+              echo "Test fails! print the ComfyUI logs..."
+              docker exec onediff-test cat /app/ComfyUI/onediff_comfyui.log
+              false
+            }
 
       - name: Show ComfyUI Log
         if: matrix.test-suite == 'comfy'

.gitignore

@@ -180,3 +180,4 @@ unet_graphs
 
 # onediff_sd_webui_extensions
 onediff_sd_webui_extensions/compiled_caches/
+onediff_comfy_nodes/benchmarks/results/

onediff_comfy_nodes/_config.py

@@ -1,5 +1,6 @@
 import os
 import sys
+import torch
 import folder_paths
 
 __all__ = [
@@ -8,6 +9,14 @@
     "is_disable_oneflow_backend",
 ]
 
+# https://huggingface.co/blog/sd3#performance-optimizations-for-sd3
+torch.set_float32_matmul_precision("high")
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+
+
 # https://github.com/comfyanonymous/ComfyUI/blob/master/folder_paths.py#L9
 os.environ["COMFYUI_ROOT"] = folder_paths.base_path
 _default_backend = os.environ.get("ONEDIFF_COMFY_NODES_DEFAULT_BACKEND", "oneflow")

onediff_comfy_nodes/_nodes.py

@@ -44,6 +44,7 @@ def speedup(
         model,
         inplace: bool = False,
         custom_booster: Optional[BoosterScheduler] = None,
+        booster_settings: Optional[BoosterSettings] = None,
         *args,
         **kwargs
     ) -> Tuple:
@@ -60,15 +61,15 @@ def speedup(
         Returns:
             Tuple: Tuple containing the optimized model.
         """
-        if not hasattr(self, "booster_settings"):
+        if booster_settings is None and not hasattr(self, "booster_settings"):
             self.booster_settings = BoosterSettings(tmp_cache_key=str(uuid.uuid4()))
 
         if custom_booster:
             booster = custom_booster
             booster.inplace = inplace
         else:
             booster = BoosterScheduler(BasicBoosterExecutor(), inplace=inplace)
-        booster.settings = self.booster_settings
+        booster.settings = self.booster_settings if booster_settings is None else booster_settings
         return (booster(model, *args, **kwargs),)
 
 
@@ -93,6 +94,9 @@ def INPUT_TYPES(s):
 
     RETURN_TYPES = ("VAE",)
 
+    def speedup(self, vae, inplace=False, custom_booster: BoosterScheduler = None):
+        return super().speedup(vae, inplace, custom_booster)
+
 
 class ControlnetSpeedup:
     @classmethod
@@ -193,7 +197,7 @@ def onediff_load_controlnet(self, control_net_name, custom_booster=None):
         return (controlnet,)
 
 
-class OneDiffCheckpointLoaderSimple(CheckpointLoaderSimple):
+class OneDiffCheckpointLoaderSimple(CheckpointLoaderSimple, SpeedupMixin):
     @classmethod
     def INPUT_TYPES(s):
         return {
@@ -207,40 +211,27 @@ def INPUT_TYPES(s):
     CATEGORY = "OneDiff/Loaders"
     FUNCTION = "onediff_load_checkpoint"
 
-    @staticmethod
-    def _load_checkpoint(
-        ckpt_name, vae_speedup="disable", custom_booster: BoosterScheduler = None
-    ):
-        """Loads a checkpoint, applying speedup techniques."""
-
-        ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
-        out = comfy.sd.load_checkpoint_guess_config(
-            ckpt_path,
-            output_vae=True,
-            output_clip=True,
-            embedding_directory=folder_paths.get_folder_paths("embeddings"),
-        )
-
-        # Unpack outputs
-        modelpatcher, clip, vae = out[:3]
+    def __init__(self) -> None:
+        super().__init__()
+        self.unet_booster_settings = BoosterSettings(tmp_cache_key=str(uuid.uuid4()))
+        self.vae_booster_settings = BoosterSettings(tmp_cache_key=str(uuid.uuid4()))
 
-        # Apply custom booster if provided, otherwise use a basic one
-        custom_booster = custom_booster or BoosterScheduler(BasicBoosterExecutor())
-        modelpatcher = custom_booster(modelpatcher, ckpt_name=ckpt_name)
+    @torch.inference_mode()
+    def onediff_load_checkpoint(
+        self, ckpt_name, vae_speedup="disable", custom_booster: BoosterScheduler = None,
+    ):
+        modelpatcher, clip, vae = self.load_checkpoint(ckpt_name)
+        modelpatcher = self.speedup(
+            modelpatcher, inplace=True, custom_booster=custom_booster, booster_settings=self.unet_booster_settings
+        )[0]
 
-        # Apply VAE speedup if enabled
         if vae_speedup == "enable":
-            vae = BoosterScheduler(BasicBoosterExecutor())(vae, ckpt_name=ckpt_name)
+            vae = self.speedup(vae, inplace=True, custom_booster=custom_booster,booster_settings = self.vae_booster_settings)[0]
 
         # Set weight inplace update
         modelpatcher.weight_inplace_update = True
-
-        return modelpatcher, clip, vae
-
-    @torch.inference_mode()
-    def onediff_load_checkpoint(
-        self, ckpt_name, vae_speedup="disable", custom_booster: BoosterScheduler = None,
-    ):
-        out = self._load_checkpoint(ckpt_name, vae_speedup, custom_booster)
-        # Return the loaded checkpoint (modelpatcher, clip, vae)
-        return out
+        return (
+            modelpatcher,
+            clip,
+            vae,
+        )

onediff_comfy_nodes/benchmarks/README.md

@@ -0,0 +1,17 @@
+## Overview
+This benchmarks the production environment stability by testing:
+- ✅ Memory prediction time
+- ✅ First AUTOTUNE duration
+- ✅ Startup time with warmup
+- [x] TODO Join CI
+- [ ] https://github.com/Fannovel16/comfyui_controlnet_aux 耗时测试
+- [ ] SSIM 验证准确度
+
+##  “disable auto tune”
+https://discuss.pytorch.org/t/how-to-use-auto-tune-with-pytorch/64669
+
+
+- [ ] lora 切换测试
+- [ ] load checkpoint 切换测试
+- [ ] 文档节点整理 
+- [ ] AnimateDiff 支持

onediff_comfy_nodes/benchmarks/resources/baseline/ComfyUI_IPAdapter_plus/ipadapter_advanced.json

@@ -0,0 +1,180 @@
+{
+  "3": {
+    "inputs": {
+      "seed": 1,
+      "steps": 30,
+      "cfg": 6.5,
+      "sampler_name": "ddpm",
+      "scheduler": "karras",
+      "denoise": 1,
+      "model": [
+        "14",
+        0
+      ],
+      "positive": [
+        "6",
+        0
+      ],
+      "negative": [
+        "7",
+        0
+      ],
+      "latent_image": [
+        "5",
+        0
+      ]
+    },
+    "class_type": "KSampler",
+    "_meta": {
+      "title": "KSampler"
+    }
+  },
+  "4": {
+    "inputs": {
+      "ckpt_name": "sd15/020.realisticVisionV51_v51VAE.safetensors"
+    },
+    "class_type": "CheckpointLoaderSimple",
+    "_meta": {
+      "title": "Load Checkpoint"
+    }
+  },
+  "5": {
+    "inputs": {
+      "width": 512,
+      "height": 512,
+      "batch_size": 1
+    },
+    "class_type": "EmptyLatentImage",
+    "_meta": {
+      "title": "Empty Latent Image"
+    }
+  },
+  "6": {
+    "inputs": {
+      "text": "in a peaceful spring morning a woman wearing a white shirt is sitting in a park on a bench\n\nhigh quality, detailed, diffuse light",
+      "clip": [
+        "4",
+        1
+      ]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": {
+      "title": "CLIP Text Encode (Prompt)"
+    }
+  },
+  "7": {
+    "inputs": {
+      "text": "blurry, noisy, messy, lowres, jpeg, artifacts, ill, distorted, malformed",
+      "clip": [
+        "4",
+        1
+      ]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": {
+      "title": "CLIP Text Encode (Prompt)"
+    }
+  },
+  "8": {
+    "inputs": {
+      "samples": [
+        "3",
+        0
+      ],
+      "vae": [
+        "4",
+        2
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAE Decode"
+    }
+  },
+  "9": {
+    "inputs": {
+      "filename_prefix": "IPAdapter",
+      "images": [
+        "8",
+        0
+      ]
+    },
+    "class_type": "SaveImage",
+    "_meta": {
+      "title": "Save Image"
+    }
+  },
+  "12": {
+    "inputs": {
+      "image": "input_image_vermeer.png",
+      "upload": "image"
+    },
+    "class_type": "LoadImage",
+    "_meta": {
+      "title": "Load Image"
+    }
+  },
+  "14": {
+    "inputs": {
+      "weight": 0.8,
+      "weight_type": "linear",
+      "combine_embeds": "concat",
+      "start_at": 0,
+      "end_at": 1,
+      "embeds_scaling": "V only",
+      "model": [
+        "4",
+        0
+      ],
+      "ipadapter": [
+        "15",
+        0
+      ],
+      "image": [
+        "17",
+        0
+      ],
+      "clip_vision": [
+        "16",
+        0
+      ]
+    },
+    "class_type": "IPAdapterAdvanced",
+    "_meta": {
+      "title": "IPAdapter Advanced"
+    }
+  },
+  "15": {
+    "inputs": {
+      "ipadapter_file": "ip-adapter-plus_sd15.safetensors"
+    },
+    "class_type": "IPAdapterModelLoader",
+    "_meta": {
+      "title": "IPAdapter Model Loader"
+    }
+  },
+  "16": {
+    "inputs": {
+      "clip_name": "SD1.5/pytorch_model.bin"
+    },
+    "class_type": "CLIPVisionLoader",
+    "_meta": {
+      "title": "Load CLIP Vision"
+    }
+  },
+  "17": {
+    "inputs": {
+      "interpolation": "LANCZOS",
+      "crop_position": "top",
+      "sharpening": 0.15,
+      "image": [
+        "12",
+        0
+      ]
+    },
+    "class_type": "PrepImageForClipVision",
+    "_meta": {
+      "title": "Prep Image For ClipVision"
+    }
+  }
+}