make ip_adapters work with stable-fast

only compile model the first time :-)

probe for availability of stable-fast compiler and triton at startup time

simplify config logic

docs/contributing/INVOCATIONS.md

@@ -11,7 +11,7 @@ complex functionality.
 
 InvokeAI Nodes can be found in the `invokeai/app/invocations` directory. These can be used as examples to create your own nodes.
 
-New nodes should be added to a subfolder in the `nodes` directory found at the root level of the InvokeAI installation location. Nodes added to this folder will be imported upon application startup. 
+New nodes should be added to a subfolder in `nodes` direction found at the root level of the InvokeAI installation location. Nodes added to this folder will be able to be used upon application startup. 
 
 Example `nodes`  subfolder structure:
 ```py

invokeai/app/services/config/config_default.py

@@ -271,6 +271,7 @@ class InvokeAIAppConfig(InvokeAISettings):
     attention_slice_size: Literal["auto", "balanced", "max", 1, 2, 3, 4, 5, 6, 7, 8] = Field(default="auto", description='Slice size, valid when attention_type=="sliced"', json_schema_extra=Categories.Generation)
     force_tiled_decode  : bool = Field(default=False, description="Whether to enable tiled VAE decode (reduces memory consumption with some performance penalty)", json_schema_extra=Categories.Generation)
     png_compress_level  : int = Field(default=6, description="The compress_level setting of PIL.Image.save(), used for PNG encoding. All settings are lossless. 0 = fastest, largest filesize, 9 = slowest, smallest filesize", json_schema_extra=Categories.Generation)
+    stable_fast         : bool = Field(default=True, description="Enable stable-fast performance optimizations, if the library is installed and functional", json_schema_extra=Categories.Generation)
 
     # QUEUE
     max_queue_size      : int = Field(default=10000, gt=0, description="Maximum number of items in the session queue", json_schema_extra=Categories.Queue)

invokeai/backend/ip_adapter/attention_processor.py

@@ -141,7 +141,6 @@ class IPAttnProcessor2_0(torch.nn.Module):
                 ip_hidden_states = ipa_embed
 
                 # Expected ip_hidden_state shape: (batch_size, num_ip_images, ip_seq_len, ip_image_embedding)
-
                 ip_key = ipa_weights.to_k_ip(ip_hidden_states)
                 ip_value = ipa_weights.to_v_ip(ip_hidden_states)
 

invokeai/backend/ip_adapter/ip_attention_weights.py

@@ -12,6 +12,8 @@ class IPAttentionProcessorWeights(torch.nn.Module):
         super().__init__()
         self.to_k_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
         self.to_v_ip = torch.nn.Linear(in_dim, out_dim, bias=False)
+        for param in self.parameters():
+            param.requires_grad = False
 
 
 class IPAttentionWeights(torch.nn.Module):

invokeai/backend/model_management/model_cache.py

@@ -24,12 +24,14 @@ import sys
 import time
 from contextlib import suppress
 from dataclasses import dataclass, field
+from importlib.util import find_spec
 from pathlib import Path
 from typing import Any, Dict, Optional, Type, Union, types
 
 import torch
 
 import invokeai.backend.util.logging as logger
+from invokeai.app.services.config import InvokeAIAppConfig
 from invokeai.backend.model_management.memory_snapshot import MemorySnapshot, get_pretty_snapshot_diff
 from invokeai.backend.model_management.model_load_optimizations import skip_torch_weight_init
 
@@ -39,6 +41,26 @@ from .models import BaseModelType, ModelBase, ModelType, SubModelType
 if choose_torch_device() == torch.device("mps"):
     from torch import mps
 
+SFAST_AVAILABLE = False
+TRITON_AVAILABLE = False
+XFORMERS_AVAILABLE = False
+SFAST_CONFIG = None
+
+TRITON_AVAILABLE = find_spec("triton") is not None
+XFORMERS_AVAILABLE = find_spec("xformers") is not None
+
+try:
+    from sfast.compilers.diffusion_pipeline_compiler import CompilationConfig, compile_unet, compile_vae
+
+    SFAST_CONFIG = CompilationConfig.Default()
+    SFAST_CONFIG.enable_cuda_graph = True
+    SFAST_CONFIG.enable_xformers = XFORMERS_AVAILABLE
+    SFAST_CONFIG.enable_triton = TRITON_AVAILABLE
+    SFAST_AVAILABLE = True
+except ImportError:
+    pass
+
+
 # Maximum size of the cache, in gigs
 # Default is roughly enough to hold three fp16 diffusers models in RAM simultaneously
 DEFAULT_MAX_CACHE_SIZE = 6.0
@@ -110,6 +132,7 @@ class _CacheRecord:
 class ModelCache(object):
     def __init__(
         self,
+        app_config: InvokeAIAppConfig,
         max_cache_size: float = DEFAULT_MAX_CACHE_SIZE,
         max_vram_cache_size: float = DEFAULT_MAX_VRAM_CACHE_SIZE,
         execution_device: torch.device = torch.device("cuda"),
@@ -122,6 +145,7 @@ class ModelCache(object):
         log_memory_usage: bool = False,
     ):
         """
+        :param app_config: InvokeAIAppConfig for application
         :param max_cache_size: Maximum size of the RAM cache [6.0 GB]
         :param execution_device: Torch device to load active model into [torch.device('cuda')]
         :param storage_device: Torch device to save inactive model in [torch.device('cpu')]
@@ -135,6 +159,7 @@ class ModelCache(object):
             behaviour.
         """
         self.model_infos: Dict[str, ModelBase] = {}
+        self.app_config = app_config
         # allow lazy offloading only when vram cache enabled
         self.lazy_offloading = lazy_offloading and max_vram_cache_size > 0
         self.precision: torch.dtype = precision
@@ -239,6 +264,9 @@ class ModelCache(object):
             snapshot_before = self._capture_memory_snapshot()
             with skip_torch_weight_init():
                 model = model_info.get_model(child_type=submodel, torch_dtype=self.precision)
+            if SFAST_AVAILABLE and self.app_config.stable_fast and submodel:
+                model = self._compile_model(model, submodel)
+
             snapshot_after = self._capture_memory_snapshot()
             end_load_time = time.time()
 
@@ -322,6 +350,16 @@ class ModelCache(object):
                     f"{get_pretty_snapshot_diff(snapshot_before, snapshot_after)}"
                 )
 
+    def _compile_model(self, model: Any, model_type: SubModelType) -> Any:
+        if model_type == SubModelType("unet"):
+            self.logger.info("SFast-compiling unet model")
+            return compile_unet(model, SFAST_CONFIG)
+        elif model_type == SubModelType("vae"):
+            self.logger.info("SFast-compiling vae model")
+            return compile_vae(model, SFAST_CONFIG)
+        else:
+            return model
+
     class ModelLocker(object):
         def __init__(self, cache, key, model, gpu_load, size_needed):
             """

invokeai/backend/model_management/model_manager.py

@@ -344,6 +344,7 @@ class ModelManager(object):
         self.app_config = InvokeAIAppConfig.get_config()
         self.logger = logger
         self.cache = ModelCache(
+            app_config=self.app_config,
             max_cache_size=max_cache_size,
             max_vram_cache_size=self.app_config.vram_cache_size,
             lazy_offloading=self.app_config.lazy_offload,

invokeai/frontend/web/public/locales/zh_CN.json

@@ -1119,7 +1119,10 @@
         "deletedInvalidEdge": "已删除无效的边缘 {{source}} -> {{target}}",
         "unknownInput": "未知输入：{{name}}",
         "prototypeDesc": "此调用是一个原型 (prototype)。它可能会在本项目更新期间发生破坏性更改，并且随时可能被删除。",
-        "betaDesc": "此调用尚处于测试阶段。在稳定之前，它可能会在项目更新期间发生破坏性更改。本项目计划长期支持这种调用。"
+        "betaDesc": "此调用尚处于测试阶段。在稳定之前，它可能会在项目更新期间发生破坏性更改。本项目计划长期支持这种调用。",
+        "newWorkflow": "新建工作流",
+        "newWorkflowDesc": "是否创建一个新的工作流？",
+        "newWorkflowDesc2": "当前工作流有未保存的更改。"
     },
     "controlnet": {
         "resize": "直接缩放",
@@ -1635,7 +1638,7 @@
         "openWorkflow": "打开工作流",
         "clearWorkflowSearchFilter": "清除工作流检索过滤器",
         "workflowLibrary": "工作流库",
-        "downloadWorkflow": "下载工作流",
+        "downloadWorkflow": "保存到文件",
         "noRecentWorkflows": "无最近工作流",
         "workflowSaved": "已保存工作流",
         "workflowIsOpen": "工作流已打开",
@@ -1648,8 +1651,9 @@
         "deleteWorkflow": "删除工作流",
         "workflows": "工作流",
         "noDescription": "无描述",
-        "uploadWorkflow": "上传工作流",
-        "userWorkflows": "我的工作流"
+        "uploadWorkflow": "从文件中加载",
+        "userWorkflows": "我的工作流",
+        "newWorkflowCreated": "已创建新的工作流"
     },
     "app": {
         "storeNotInitialized": "商店尚未初始化"

invokeai/frontend/web/src/app/store/store.ts

@@ -34,6 +34,7 @@ import { actionSanitizer } from './middleware/devtools/actionSanitizer';
 import { actionsDenylist } from './middleware/devtools/actionsDenylist';
 import { stateSanitizer } from './middleware/devtools/stateSanitizer';
 import { listenerMiddleware } from './middleware/listenerMiddleware';
+import { authToastMiddleware } from 'services/api/authToastMiddleware';
 
 const allReducers = {
   canvas: canvasReducer,
@@ -96,6 +97,7 @@ export const createStore = (uniqueStoreKey?: string, persist = true) =>
       })
         .concat(api.middleware)
         .concat(dynamicMiddlewares)
+        .concat(authToastMiddleware)
         .prepend(listenerMiddleware.middleware),
     enhancers: (getDefaultEnhancers) => {
       const _enhancers = getDefaultEnhancers().concat(autoBatchEnhancer());

invokeai/frontend/web/src/services/api/authToastMiddleware.ts

@@ -5,12 +5,10 @@ import { t } from 'i18next';
 import { z } from 'zod';
 
 const zRejectedForbiddenAction = z.object({
-  action: z.object({
-    payload: z.object({
-      status: z.literal(403),
-      data: z.object({
-        detail: z.string(),
-      }),
+  payload: z.object({
+    status: z.literal(403),
+    data: z.object({
+      detail: z.string(),
     }),
   }),
 });
@@ -22,8 +20,8 @@ export const authToastMiddleware: Middleware =
         const parsed = zRejectedForbiddenAction.parse(action);
         const { dispatch } = api;
         const customMessage =
-          parsed.action.payload.data.detail !== 'Forbidden'
-            ? parsed.action.payload.data.detail
+          parsed.payload.data.detail !== 'Forbidden'
+            ? parsed.payload.data.detail
             : undefined;
         dispatch(
           addToast({
@@ -32,7 +30,7 @@ export const authToastMiddleware: Middleware =
             description: customMessage,
           })
         );
-      } catch {
+      } catch (error) {
         // no-op
       }
     }

pyproject.toml

@@ -113,6 +113,7 @@ dependencies = [
 "onnx" = ["onnxruntime"]
 "onnx-cuda" = ["onnxruntime-gpu"]
 "onnx-directml" = ["onnxruntime-directml"]
+"stable-fast" = ["stable-fast"]
 
 [project.scripts]