Merge branch 'main' into rocm57

.nvmrc

@@ -0,0 +1 @@
+v22.12.0

docs/features/low-vram.md

@@ -28,11 +28,12 @@ It is possible to fine-tune the settings for best performance or if you still ge
 
 ## Details and fine-tuning
 
-Low-VRAM mode involves 3 features, each of which can be configured or fine-tuned:
+Low-VRAM mode involves 4 features, each of which can be configured or fine-tuned:
 
-- Partial model loading
-- Dynamic RAM and VRAM cache sizes
-- Working memory
+- Partial model loading (`enable_partial_loading`)
+- Dynamic RAM and VRAM cache sizes (`max_cache_ram_gb`, `max_cache_vram_gb`)
+- Working memory (`device_working_mem_gb`)
+- Keeping a RAM weight copy (`keep_ram_copy_of_weights`)
 
 Read on to learn about these features and understand how to fine-tune them for your system and use-cases.
 
@@ -67,12 +68,20 @@ As of v5.6.0, the caches are dynamically sized. The `ram` and `vram` settings ar
 But, if your GPU has enough VRAM to hold models fully, you might get a perf boost by manually setting the cache sizes in `invokeai.yaml`:
 
 ```yaml
-# Set the RAM cache size to as large as possible, leaving a few GB free for the rest of your system and Invoke.
-# For example, if your system has 32GB RAM, 28GB is a good value.
+# The default max cache RAM size is logged on InvokeAI startup. It is determined based on your system RAM / VRAM.
+# You can override the default value by setting `max_cache_ram_gb`.
+# Increasing `max_cache_ram_gb` will increase the amount of RAM used to cache inactive models, resulting in faster model
+# reloads for the cached models.
+# As an example, if your system has 32GB of RAM and no other heavy processes, setting the `max_cache_ram_gb` to 28GB
+# might be a good value to achieve aggressive model caching.
 max_cache_ram_gb: 28
-# Set the VRAM cache size to be as large as possible while leaving enough room for the working memory of the tasks you will be doing.
-# For example, on a 24GB GPU that will be running unquantized FLUX without any auxiliary models,
-# 18GB is a good value.
+# The default max cache VRAM size is adjusted dynamically based on the amount of available VRAM (taking into
+# consideration the VRAM used by other processes).
+# You can override the default value by setting `max_cache_vram_gb`. Note that this value takes precedence over the
+# `device_working_mem_gb`.
+# It is recommended to set the VRAM cache size to be as large as possible while leaving enough room for the working
+# memory of the tasks you will be doing. For example, on a 24GB GPU that will be running unquantized FLUX without any
+# auxiliary models, 18GB might be a good value.
 max_cache_vram_gb: 18
 ```
 
@@ -109,6 +118,15 @@ device_working_mem_gb: 4
 
     Once decoding completes, the model manager "reclaims" the extra VRAM allocated as working memory for future model loading operations.
 
+### Keeping a RAM weight copy
+
+Invoke has the option of keeping a RAM copy of all model weights, even when they are loaded onto the GPU. This optimization is _on_ by default, and enables faster model switching and LoRA patching. Disabling this feature will reduce the average RAM load while running Invoke (peak RAM likely won't change), at the cost of slower model switching and LoRA patching. If you have limited RAM, you can disable this optimization:
+
+```yaml
+# Set to false to reduce the average RAM usage at the cost of slower model switching and LoRA patching.
+keep_ram_copy_of_weights: false
+```
+
 ### Disabling Nvidia sysmem fallback (Windows only)
 
 On Windows, Nvidia GPUs are able to use system RAM when their VRAM fills up via **sysmem fallback**. While it sounds like a good idea on the surface, in practice it causes massive slowdowns during generation.

invokeai/app/invocations/batch.py

@@ -0,0 +1,200 @@
+from typing import Literal
+
+from pydantic import BaseModel
+
+from invokeai.app.invocations.baseinvocation import (
+    BaseInvocation,
+    BaseInvocationOutput,
+    Classification,
+    invocation,
+    invocation_output,
+)
+from invokeai.app.invocations.fields import (
+    ImageField,
+    Input,
+    InputField,
+    OutputField,
+)
+from invokeai.app.invocations.primitives import (
+    FloatOutput,
+    ImageOutput,
+    IntegerOutput,
+    StringOutput,
+)
+from invokeai.app.services.shared.invocation_context import InvocationContext
+
+BATCH_GROUP_IDS = Literal[
+    "None",
+    "Group 1",
+    "Group 2",
+    "Group 3",
+    "Group 4",
+    "Group 5",
+]
+
+
+class NotExecutableNodeError(Exception):
+    def __init__(self, message: str = "This class should never be executed or instantiated directly."):
+        super().__init__(message)
+
+    pass
+
+
+class BaseBatchInvocation(BaseInvocation):
+    batch_group_id: BATCH_GROUP_IDS = InputField(
+        default="None",
+        description="The ID of this batch node's group. If provided, all batch nodes in with the same ID will be 'zipped' before execution, and all nodes' collections must be of the same size.",
+        input=Input.Direct,
+        title="Batch Group",
+    )
+
+    def __init__(self):
+        raise NotExecutableNodeError()
+
+
+@invocation(
+    "image_batch",
+    title="Image Batch",
+    tags=["primitives", "image", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class ImageBatchInvocation(BaseBatchInvocation):
+    """Create a batched generation, where the workflow is executed once for each image in the batch."""
+
+    images: list[ImageField] = InputField(
+        default=[], min_length=1, description="The images to batch over", input=Input.Direct
+    )
+
+    def invoke(self, context: InvocationContext) -> ImageOutput:
+        raise NotExecutableNodeError()
+
+
+@invocation(
+    "string_batch",
+    title="String Batch",
+    tags=["primitives", "string", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class StringBatchInvocation(BaseBatchInvocation):
+    """Create a batched generation, where the workflow is executed once for each string in the batch."""
+
+    strings: list[str] = InputField(
+        default=[], min_length=1, description="The strings to batch over", input=Input.Direct
+    )
+
+    def invoke(self, context: InvocationContext) -> StringOutput:
+        raise NotExecutableNodeError()
+
+
+@invocation(
+    "integer_batch",
+    title="Integer Batch",
+    tags=["primitives", "integer", "number", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class IntegerBatchInvocation(BaseBatchInvocation):
+    """Create a batched generation, where the workflow is executed once for each integer in the batch."""
+
+    integers: list[int] = InputField(
+        default=[],
+        min_length=1,
+        description="The integers to batch over",
+    )
+
+    def invoke(self, context: InvocationContext) -> IntegerOutput:
+        raise NotExecutableNodeError()
+
+
+@invocation_output("integer_generator_output")
+class IntegerGeneratorOutput(BaseInvocationOutput):
+    integers: list[int] = OutputField(description="The generated integers")
+
+
+class IntegerGeneratorField(BaseModel):
+    pass
+
+
+@invocation(
+    "integer_generator",
+    title="Integer Generator",
+    tags=["primitives", "int", "number", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class IntegerGenerator(BaseInvocation):
+    """Generated a range of integers for use in a batched generation"""
+
+    generator: IntegerGeneratorField = InputField(
+        description="The integer generator.",
+        input=Input.Direct,
+        title="Generator Type",
+    )
+
+    def __init__(self):
+        raise NotExecutableNodeError()
+
+    def invoke(self, context: InvocationContext) -> IntegerGeneratorOutput:
+        raise NotExecutableNodeError()
+
+
+@invocation(
+    "float_batch",
+    title="Float Batch",
+    tags=["primitives", "float", "number", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class FloatBatchInvocation(BaseBatchInvocation):
+    """Create a batched generation, where the workflow is executed once for each float in the batch."""
+
+    floats: list[float] = InputField(
+        default=[],
+        min_length=1,
+        description="The floats to batch over",
+    )
+
+    def invoke(self, context: InvocationContext) -> FloatOutput:
+        raise NotExecutableNodeError()
+
+
+@invocation_output("float_generator_output")
+class FloatGeneratorOutput(BaseInvocationOutput):
+    """Base class for nodes that output a collection of floats"""
+
+    floats: list[float] = OutputField(description="The generated floats")
+
+
+class FloatGeneratorField(BaseModel):
+    pass
+
+
+@invocation(
+    "float_generator",
+    title="Float Generator",
+    tags=["primitives", "float", "number", "batch", "special"],
+    category="primitives",
+    version="1.0.0",
+    classification=Classification.Special,
+)
+class FloatGenerator(BaseInvocation):
+    """Generated a range of floats for use in a batched generation"""
+
+    generator: FloatGeneratorField = InputField(
+        description="The float generator.",
+        input=Input.Direct,
+        title="Generator Type",
+    )
+
+    def __init__(self):
+        raise NotExecutableNodeError()
+
+    def invoke(self, context: InvocationContext) -> FloatGeneratorOutput:
+        raise NotExecutableNodeError()

invokeai/app/invocations/denoise_latents.py

@@ -40,6 +40,7 @@
 from invokeai.app.util.controlnet_utils import prepare_control_image
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
 from invokeai.backend.model_manager import BaseModelType, ModelVariantType
+from invokeai.backend.model_manager.config import AnyModelConfig
 from invokeai.backend.model_patcher import ModelPatcher
 from invokeai.backend.patches.layer_patcher import LayerPatcher
 from invokeai.backend.patches.model_patch_raw import ModelPatchRaw
@@ -85,6 +86,7 @@ def get_scheduler(
     scheduler_info: ModelIdentifierField,
     scheduler_name: str,
     seed: int,
+    unet_config: AnyModelConfig,
 ) -> Scheduler:
     """Load a scheduler and apply some scheduler-specific overrides."""
     # TODO(ryand): Silently falling back to ddim seems like a bad idea. Look into why this was added and remove if
@@ -103,6 +105,9 @@ def get_scheduler(
         "_backup": scheduler_config,
     }
 
+    if hasattr(unet_config, "prediction_type"):
+        scheduler_config["prediction_type"] = unet_config.prediction_type
+
     # make dpmpp_sde reproducable(seed can be passed only in initializer)
     if scheduler_class is DPMSolverSDEScheduler:
         scheduler_config["noise_sampler_seed"] = seed
@@ -829,6 +834,9 @@ def _new_invoke(self, context: InvocationContext) -> LatentsOutput:
         seed, noise, latents = self.prepare_noise_and_latents(context, self.noise, self.latents)
         _, _, latent_height, latent_width = latents.shape
 
+        # get the unet's config so that we can pass the base to sd_step_callback()
+        unet_config = context.models.get_config(self.unet.unet.key)
+
         conditioning_data = self.get_conditioning_data(
             context=context,
             positive_conditioning_field=self.positive_conditioning,
@@ -848,6 +856,7 @@ def _new_invoke(self, context: InvocationContext) -> LatentsOutput:
             scheduler_info=self.unet.scheduler,
             scheduler_name=self.scheduler,
             seed=seed,
+            unet_config=unet_config,
         )
 
         timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(
@@ -859,9 +868,6 @@ def _new_invoke(self, context: InvocationContext) -> LatentsOutput:
             denoising_end=self.denoising_end,
         )
 
-        # get the unet's config so that we can pass the base to sd_step_callback()
-        unet_config = context.models.get_config(self.unet.unet.key)
-
         ### preview
         def step_callback(state: PipelineIntermediateState) -> None:
             context.util.sd_step_callback(state, unet_config.base)
@@ -1030,6 +1036,7 @@ def _lora_loader() -> Iterator[Tuple[ModelPatchRaw, float]]:
                 scheduler_info=self.unet.scheduler,
                 scheduler_name=self.scheduler,
                 seed=seed,
+                unet_config=unet_config,
             )
 
             pipeline = self.create_pipeline(unet, scheduler)

invokeai/app/invocations/flux_model_loader.py

@@ -10,6 +10,10 @@
 from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, OutputField, UIType
 from invokeai.app.invocations.model import CLIPField, ModelIdentifierField, T5EncoderField, TransformerField, VAEField
 from invokeai.app.services.shared.invocation_context import InvocationContext
+from invokeai.app.util.t5_model_identifier import (
+    preprocess_t5_encoder_model_identifier,
+    preprocess_t5_tokenizer_model_identifier,
+)
 from invokeai.backend.flux.util import max_seq_lengths
 from invokeai.backend.model_manager.config import (
     CheckpointConfigBase,
@@ -74,8 +78,8 @@ def invoke(self, context: InvocationContext) -> FluxModelLoaderOutput:
         tokenizer = self.clip_embed_model.model_copy(update={"submodel_type": SubModelType.Tokenizer})
         clip_encoder = self.clip_embed_model.model_copy(update={"submodel_type": SubModelType.TextEncoder})
 
-        tokenizer2 = self.t5_encoder_model.model_copy(update={"submodel_type": SubModelType.Tokenizer2})
-        t5_encoder = self.t5_encoder_model.model_copy(update={"submodel_type": SubModelType.TextEncoder2})
+        tokenizer2 = preprocess_t5_tokenizer_model_identifier(self.t5_encoder_model)
+        t5_encoder = preprocess_t5_encoder_model_identifier(self.t5_encoder_model)
 
         transformer_config = context.models.get_config(transformer)
         assert isinstance(transformer_config, CheckpointConfigBase)

invokeai/app/invocations/flux_text_encoder.py

@@ -2,7 +2,7 @@
 from typing import Iterator, Literal, Optional, Tuple
 
 import torch
-from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer
+from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5Tokenizer, T5TokenizerFast
 
 from invokeai.app.invocations.baseinvocation import BaseInvocation, Classification, invocation
 from invokeai.app.invocations.fields import (
@@ -76,7 +76,7 @@ def _t5_encode(self, context: InvocationContext) -> torch.Tensor:
             context.models.load(self.t5_encoder.tokenizer) as t5_tokenizer,
         ):
             assert isinstance(t5_text_encoder, T5EncoderModel)
-            assert isinstance(t5_tokenizer, T5Tokenizer)
+            assert isinstance(t5_tokenizer, (T5Tokenizer, T5TokenizerFast))
 
             t5_encoder = HFEncoder(t5_text_encoder, t5_tokenizer, False, self.t5_max_seq_len)