Initial commit

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+RobertML.png filter=lfs diff=lfs merge=lfs -text
+backup.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1 @@
+# speedmax

pyproject.toml

@@ -0,0 +1,48 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission by RobertML for the 4090 Flux contest"
+requires-python = ">=3.10,<3.13"
+version = "8"
+dependencies = [
+  "diffusers==0.31.0",
+  "transformers==4.46.2",
+  "accelerate==1.1.0",
+  "omegaconf==2.3.0",
+  "torch==2.6.0",
+  "protobuf==5.28.3",
+  "sentencepiece==0.2.0",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "gitpython>=3.1.43",
+  "hf_transfer==0.1.8",
+  "torchao==0.6.1",
+  "setuptools>=75.3.0",
+  "para-attn==0.3.15",
+  "git-lfs<=1.6"
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "manbeast3b/flux.1-schnell-full1"
+revision = "cb1b599b0d712b9aab2c4df3ad27b050a27ec146"
+
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-vae_e3m2"
+revision = "da0d2cd7815792fb40d084dbd8ed32b63f153d8d"
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/caching.py

@@ -0,0 +1,294 @@
+# caching.py
+
+import functools
+import unittest
+import contextlib
+import dataclasses
+from collections import defaultdict
+from typing import DefaultDict, Dict
+import torch
+from diffusers import DiffusionPipeline, FluxTransformer2DModel
+
+
+@dataclasses.dataclass
+class CacheContext:
+    buffers: Dict[str, torch.Tensor] = dataclasses.field(default_factory=dict)
+    incremental_name_counters: DefaultDict[str, int] = dataclasses.field(default_factory=lambda: defaultdict(int))
+
+    def get_incremental_name(self, name=None):
+        if name is None:
+            name = "default"
+        idx = self.incremental_name_counters[name]
+        self.incremental_name_counters[name] += 1
+        return f"{name}_{idx}"
+
+    def reset_incremental_names(self):
+        self.incremental_name_counters.clear()
+
+    @torch.compiler.disable
+    def get_buffer(self, name):
+        return self.buffers.get(name)
+
+    @torch.compiler.disable
+    def set_buffer(self, name, buffer):
+        self.buffers[name] = buffer
+
+    def clear_buffers(self):
+        self.buffers.clear()
+
+
+@torch.compiler.disable
+def get_buffer(name):
+    cache_context = get_current_cache_context()
+    assert cache_context is not None, "cache_context must be set before"
+    return cache_context.get_buffer(name)
+
+
+@torch.compiler.disable
+def set_buffer(name, buffer):
+    cache_context = get_current_cache_context()
+    assert cache_context is not None, "cache_context must be set before"
+    cache_context.set_buffer(name, buffer)
+
+
+_current_cache_context = None
+
+
+def create_cache_context():
+    return CacheContext()
+
+
+def get_current_cache_context():
+    return _current_cache_context
+
+
+def set_current_cache_context(cache_context=None):
+    global _current_cache_context
+    _current_cache_context = cache_context
+
+
+@contextlib.contextmanager
+def cache_context(cache_context):
+    global _current_cache_context
+    old_cache_context = _current_cache_context
+    _current_cache_context = cache_context
+    try:
+        yield
+    finally:
+        _current_cache_context = old_cache_context
+
+
+
+@torch.compiler.disable
+def apply_prev_hidden_states_residual(hidden_states, encoder_hidden_states):
+    hidden_states_residual = get_buffer("hidden_states_residual")
+    assert hidden_states_residual is not None, "hidden_states_residual must be set before"
+    hidden_states = hidden_states_residual + hidden_states
+
+    encoder_hidden_states_residual = get_buffer("encoder_hidden_states_residual")
+    assert encoder_hidden_states_residual is not None, "encoder_hidden_states_residual must be set before"
+    encoder_hidden_states = encoder_hidden_states_residual + encoder_hidden_states
+
+    hidden_states = hidden_states.contiguous()
+    encoder_hidden_states = encoder_hidden_states.contiguous()
+
+    return hidden_states, encoder_hidden_states
+
+
+def are_two_tensors_similar(t1, t2, *, threshold=0.85):
+    mean_diff = (t1 - t2).abs().mean()
+    mean_t1 = t1.abs().mean()
+    diff = mean_diff / mean_t1
+    return diff.item() < threshold
+
+@torch.compiler.disable
+def get_can_use_cache(first_hidden_states_residual, threshold, parallelized=False):
+    prev_first_hidden_states_residual = get_buffer("first_hidden_states_residual")
+    can_use_cache = prev_first_hidden_states_residual is not None and are_two_tensors_similar(
+        prev_first_hidden_states_residual,
+        first_hidden_states_residual,
+    )
+    return can_use_cache
+
+
+class CachedTransformerBlocks(torch.nn.Module):
+    def __init__(
+        self,
+        transformer_blocks,
+        single_transformer_blocks=None,
+        *,
+        transformer=None,
+        residual_diff_threshold,
+        return_hidden_states_first=True,
+    ):
+        super().__init__()
+        self.transformer = transformer
+        self.transformer_blocks = transformer_blocks
+        self.single_transformer_blocks = single_transformer_blocks
+        self.residual_diff_threshold = residual_diff_threshold
+        self.return_hidden_states_first = return_hidden_states_first
+
+    def forward(self, hidden_states, encoder_hidden_states, *args, **kwargs):
+        if self.residual_diff_threshold <= 0.0:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+                if not self.return_hidden_states_first:
+                    hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+            if self.single_transformer_blocks is not None:
+                hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+                for block in self.single_transformer_blocks:
+                    hidden_states = block(hidden_states, *args, **kwargs)
+                hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :]
+            return (
+                (hidden_states, encoder_hidden_states)
+                if self.return_hidden_states_first
+                else (encoder_hidden_states, hidden_states)
+            )
+
+        original_hidden_states = hidden_states
+        first_transformer_block = self.transformer_blocks[0]
+        hidden_states, encoder_hidden_states = first_transformer_block(
+            hidden_states, encoder_hidden_states, *args, **kwargs
+        )
+        if not self.return_hidden_states_first:
+            hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+        first_hidden_states_residual = hidden_states - original_hidden_states
+        del original_hidden_states
+
+        can_use_cache = get_can_use_cache(
+            first_hidden_states_residual,
+            threshold=self.residual_diff_threshold,
+            parallelized=self.transformer is not None and getattr(self.transformer, "_is_parallelized", False),
+        )
+
+        torch._dynamo.graph_break()
+        if can_use_cache:
+            del first_hidden_states_residual
+            hidden_states, encoder_hidden_states = apply_prev_hidden_states_residual(
+                hidden_states, encoder_hidden_states
+            )
+        else:
+            set_buffer("first_hidden_states_residual", first_hidden_states_residual)
+            del first_hidden_states_residual
+            (
+                hidden_states,
+                encoder_hidden_states,
+                hidden_states_residual,
+                encoder_hidden_states_residual,
+            ) = self.call_remaining_transformer_blocks(hidden_states, encoder_hidden_states, *args, **kwargs)
+            set_buffer("hidden_states_residual", hidden_states_residual)
+            set_buffer("encoder_hidden_states_residual", encoder_hidden_states_residual)
+        torch._dynamo.graph_break()
+
+        return (
+            (hidden_states, encoder_hidden_states)
+            if self.return_hidden_states_first
+            else (encoder_hidden_states, hidden_states)
+        )
+
+    def call_remaining_transformer_blocks(self, hidden_states, encoder_hidden_states, *args, **kwargs):
+        original_hidden_states = hidden_states
+        original_encoder_hidden_states = encoder_hidden_states
+        for block in self.transformer_blocks[1:]:
+            hidden_states, encoder_hidden_states = block(hidden_states, encoder_hidden_states, *args, **kwargs)
+            if not self.return_hidden_states_first:
+                hidden_states, encoder_hidden_states = encoder_hidden_states, hidden_states
+        if self.single_transformer_blocks is not None:
+            hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+            for block in self.single_transformer_blocks:
+                hidden_states = block(hidden_states, *args, **kwargs)
+            encoder_hidden_states, hidden_states = hidden_states.split(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+
+        # hidden_states_shape = hidden_states.shape
+        # encoder_hidden_states_shape = encoder_hidden_states.shape
+        hidden_states = hidden_states.reshape(-1).contiguous().reshape(original_hidden_states.shape)
+        encoder_hidden_states = (
+            encoder_hidden_states.reshape(-1).contiguous().reshape(original_encoder_hidden_states.shape)
+        )
+
+        # hidden_states = hidden_states.contiguous()
+        # encoder_hidden_states = encoder_hidden_states.contiguous()
+
+        hidden_states_residual = hidden_states - original_hidden_states
+        encoder_hidden_states_residual = encoder_hidden_states - original_encoder_hidden_states
+
+        hidden_states_residual = hidden_states_residual.reshape(-1).contiguous().reshape(original_hidden_states.shape)
+        encoder_hidden_states_residual = (
+            encoder_hidden_states_residual.reshape(-1).contiguous().reshape(original_encoder_hidden_states.shape)
+        )
+
+        return hidden_states, encoder_hidden_states, hidden_states_residual, encoder_hidden_states_residual
+
+
+def apply_cache_on_transformer(
+    transformer: FluxTransformer2DModel,
+    *,
+    residual_diff_threshold=0.1,
+):
+    cached_transformer_blocks = torch.nn.ModuleList(
+        [
+            CachedTransformerBlocks(
+                transformer.transformer_blocks,
+                transformer.single_transformer_blocks,
+                transformer=transformer,
+                residual_diff_threshold=residual_diff_threshold,
+                return_hidden_states_first=False,
+            )
+        ]
+    )
+    dummy_single_transformer_blocks = torch.nn.ModuleList()
+
+    original_forward = transformer.forward
+
+    @functools.wraps(original_forward)
+    def new_forward(
+        self,
+        *args,
+        **kwargs,
+    ):
+        with unittest.mock.patch.object(
+            self,
+            "transformer_blocks",
+            cached_transformer_blocks,
+        ), unittest.mock.patch.object(
+            self,
+            "single_transformer_blocks",
+            dummy_single_transformer_blocks,
+        ):
+            return original_forward(
+                *args,
+                **kwargs,
+            )
+
+    transformer.forward = new_forward.__get__(transformer)
+
+    return transformer
+
+
+def apply_cache_on_pipe(
+    pipe: DiffusionPipeline,
+    *,
+    shallow_patch: bool = False,
+    **kwargs,
+):
+    original_call = pipe.__class__.__call__
+
+    if not getattr(original_call, "_is_cached", False):
+
+        @functools.wraps(original_call)
+        def new_call(self, *args, **kwargs):
+            with cache_context(create_cache_context()):
+                return original_call(self, *args, **kwargs)
+
+        pipe.__class__.__call__ = new_call
+
+        new_call._is_cached = True
+
+    if not shallow_patch:
+        apply_cache_on_transformer(pipe.transformer, **kwargs)
+
+    pipe._is_cached = True
+
+    return pipe

src/flux_schnell_edge_inference.egg-info/PKG-INFO

@@ -0,0 +1,16 @@
+Metadata-Version: 2.2
+Name: flux-schnell-edge-inference
+Version: 8
+Summary: An edge-maxxing model submission by RobertML for the 4090 Flux contest
+Requires-Python: <3.13,>=3.10
+Requires-Dist: diffusers==0.31.0
+Requires-Dist: transformers==4.46.2
+Requires-Dist: accelerate==1.1.0
+Requires-Dist: omegaconf==2.3.0
+Requires-Dist: torch==2.6.0
+Requires-Dist: protobuf==5.28.3
+Requires-Dist: sentencepiece==0.2.0
+Requires-Dist: edge-maxxing-pipelines@ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines
+Requires-Dist: gitpython>=3.1.43
+Requires-Dist: hf_transfer==0.1.8
+Requires-Dist: torchao==0.6.1

src/flux_schnell_edge_inference.egg-info/SOURCES.txt

@@ -0,0 +1,17 @@
+README.md
+pyproject.toml
+src/main.py
+src/pipeline.py
+src/first_block_cache/__init__.py
+src/first_block_cache/utils.py
+src/first_block_cache/diffusers_adapters/__init__.py
+src/first_block_cache/diffusers_adapters/cogvideox.py
+src/first_block_cache/diffusers_adapters/flux.py
+src/first_block_cache/diffusers_adapters/hunyuan_video.py
+src/first_block_cache/diffusers_adapters/mochi.py
+src/flux_schnell_edge_inference.egg-info/PKG-INFO
+src/flux_schnell_edge_inference.egg-info/SOURCES.txt
+src/flux_schnell_edge_inference.egg-info/dependency_links.txt
+src/flux_schnell_edge_inference.egg-info/entry_points.txt
+src/flux_schnell_edge_inference.egg-info/requires.txt
+src/flux_schnell_edge_inference.egg-info/top_level.txt

src/flux_schnell_edge_inference.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

src/flux_schnell_edge_inference.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+start_inference = main:main

src/flux_schnell_edge_inference.egg-info/requires.txt

@@ -0,0 +1,11 @@
+diffusers==0.31.0
+transformers==4.46.2
+accelerate==1.1.0
+omegaconf==2.3.0
+torch==2.6.0
+protobuf==5.28.3
+sentencepiece==0.2.0
+edge-maxxing-pipelines@ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines
+gitpython>=3.1.43
+hf_transfer==0.1.8
+torchao==0.6.1

src/flux_schnell_edge_inference.egg-info/top_level.txt

@@ -0,0 +1,3 @@
+first_block_cache
+main
+pipeline

src/main.py

@@ -0,0 +1,55 @@
+import atexit
+from io import BytesIO
+from multiprocessing.connection import Listener
+from os import chmod, remove
+from os.path import abspath, exists
+from pathlib import Path
+from git import Repo
+import torch
+
+from PIL.JpegImagePlugin import JpegImageFile
+from pipelines.models import TextToImageRequest
+from pipeline import load_pipeline, infer
+SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
+
+
+def at_exit():
+    torch.cuda.empty_cache()
+
+
+def main():
+    atexit.register(at_exit)
+
+    print(f"Loading pipeline")
+    pipeline = load_pipeline()
+
+    print(f"Pipeline loaded, creating socket at '{SOCKET}'")
+
+    if exists(SOCKET):
+        remove(SOCKET)
+
+    with Listener(SOCKET) as listener:
+        chmod(SOCKET, 0o777)
+
+        print(f"Awaiting connections")
+        with listener.accept() as connection:
+            print(f"Connected")
+            generator = torch.Generator("cuda")
+            while True:
+                try:
+                    request = TextToImageRequest.model_validate_json(connection.recv_bytes().decode("utf-8"))
+                except EOFError:
+                    print(f"Inference socket exiting")
+
+                    return
+                image = infer(request, pipeline, generator.manual_seed(request.seed))
+                data = BytesIO()
+                image.save(data, format=JpegImageFile.format)
+
+                packet = data.getvalue()
+
+                connection.send_bytes(packet )
+
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,141 @@
+
+import os
+import gc
+import torch
+from PIL.Image import Image
+from dataclasses import dataclass
+from diffusers import DiffusionPipeline, AutoencoderTiny, FluxTransformer2DModel
+from transformers import T5EncoderModel
+from huggingface_hub.constants import HF_HUB_CACHE
+from torchao.quantization import quantize_, int8_weight_only, float8_weight_only
+from caching import apply_cache_on_pipe
+from pipelines.models import TextToImageRequest
+from torch import Generator
+
+# Configuration settings using a dataclass for clarity
+@dataclass
+class Config:
+    CKPT_ID: str = "black-forest-labs/FLUX.1-schnell"
+    CKPT_REVISION: str = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+    DEVICE: str = "cuda"
+    DTYPE = torch.bfloat16
+    PYTORCH_CUDA_ALLOC_CONF: str = "expandable_segments:True"
+
+def _initialize_environment():
+    """Set up PyTorch and CUDA environment variables for optimal performance."""
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.enabled = True
+    torch.backends.cudnn.benchmark = True
+    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = Config.PYTORCH_CUDA_ALLOC_CONF
+
+def _clear_gpu_memory():
+    """Free up GPU memory to prevent memory-related issues."""
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def _load_text_encoder_model():
+    """Load the text encoder model with specified configuration."""
+    return T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16",
+        revision="1b9c856aadb864af93c1dcdc226c2774fa67bc86",
+        torch_dtype=Config.DTYPE
+    ).to(memory_format=torch.channels_last)
+
+def _load_vae_model():
+    """Load the variational autoencoder (VAE) model with specified configuration."""
+    return AutoencoderTiny.from_pretrained(
+        "RobertML/FLUX.1-schnell-vae_e3m2",
+        revision="da0d2cd7815792fb40d084dbd8ed32b63f153d8d",
+        torch_dtype=Config.DTYPE
+    )
+
+def _load_transformer_model():
+    """Load the transformer model from a specific cached path."""
+    # transformer_path = os.path.join(
+    #     HF_HUB_CACHE,"models--manbeast3b--flux.1-schnell-full1/snapshots/cb1b599b0d712b9aab2c4df3ad27b050a27ec146", 
+        
+    # )
+    transformer_path = os.path.join(
+        HF_HUB_CACHE,
+        "models--manbeast3b--flux.1-schnell-full1/snapshots/cb1b599b0d712b9aab2c4df3ad27b050a27ec146", 
+        "transformer"
+    )
+    return FluxTransformer2DModel.from_pretrained(
+        transformer_path,
+        torch_dtype=Config.DTYPE,
+        use_safetensors=False
+    ).to(memory_format=torch.channels_last)
+
+def _warmup_pipeline(pipeline):
+    """Warm up the pipeline by running it with an empty prompt to initialize internal caches."""
+    for _ in range(3):
+        pipeline(prompt=" ")
+
+def load_pipeline():
+    """
+    Load and configure the diffusion pipeline for text-to-image generation.
+
+    Returns:
+        DiffusionPipeline: The configured pipeline ready for inference.
+    """
+    _clear_gpu_memory()
+
+    # Load individual components
+    text_encoder = _load_text_encoder_model()
+    vae = _load_vae_model()
+    transformer = _load_transformer_model()
+
+    # Assemble the diffusion pipeline
+    pipeline = DiffusionPipeline.from_pretrained(
+        Config.CKPT_ID,
+        vae=vae,
+        revision=Config.CKPT_REVISION,
+        transformer=transformer,
+        text_encoder_2=text_encoder,
+        torch_dtype=Config.DTYPE,
+    ).to(Config.DEVICE)
+
+    # Apply optimizations
+    apply_cache_on_pipe(pipeline)
+    pipeline.to(memory_format=torch.channels_last)
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune")
+    quantize_(pipeline.vae, int8_weight_only())
+    quantize_(pipeline.vae, float8_weight_only())
+
+    # Warm up the pipeline to ensure readiness
+    _warmup_pipeline(pipeline)
+
+    return pipeline
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: DiffusionPipeline, generator: Generator) -> Image:
+    """
+    Generate an image from a text prompt using the diffusion pipeline.
+
+    Args:
+        request (TextToImageRequest): The request containing the prompt and image parameters.
+        pipeline (DiffusionPipeline): The pre-loaded diffusion pipeline.
+        generator (Generator): The random seed generator for reproducibility.
+
+    Returns:
+        Image: The generated image in PIL format.
+    """
+    image = pipeline(
+        prompt=request.prompt,
+        generator=generator,
+        guidance_scale=0.0,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+        output_type="pil"
+    ).images[0]
+    return image
+
+# Initialize environment settings when the module is imported
+_initialize_environment()
+
+# For compatibility with other scripts, alias load_pipeline as load
+load = load_pipeline