sd-mecha 1.1.0rc3

State dict recipe merger

sd-mecha

import sd_mecha

# create the merge plan
a = sd_mecha.model("path/to/model_a.safetensors")
b = sd_mecha.model("path/to/model_b.safetensors")
recipe = sd_mecha.weighted_sum(a, b, alpha=0.5)

# merge!
sd_mecha.merge(recipe, output="path/to/model_out.safetensors")

sd-mecha is a general memory-efficient model merging library. It can merge any model:

• Diffusion models
• LLMs
• VLMs
• Aesthetic scorers
• etc.

Features

• Memory efficient model merging: merge a very large number of models in a single execution
• Textual and interpretable format for storage and execution (.mecha)
• Extensible library interface:
  • add custom models
  • add custom merge methods
• Builtin support for popular diffusion models:
  • Stable Diffusion 1.5
  • Stable Diffusion XL
  • Stable Diffusion 3
  • FLUX Schnell/Dev
• Merge LyCORIS networks together and to checkpoints
• Block-wise hyperparameters for precise control of blocks (aka MBW)

Install

pip install sd-mecha

Make sure to install the appropriate release of torch to get the best performance.

Usage

Merge models

To merge models, mecha uses recipes. A recipe is a list of instructions that describes the exact steps needed to derive a state dict from inputs.

Here's an example script that merges three models:

import sd_mecha

# create the merge plan
model_a = sd_mecha.model("path/to/model_a.safetensors")
model_b = sd_mecha.model("path/to/model_b.safetensors")
recipe = sd_mecha.weighted_sum(model_a, model_b, alpha=0.5)

# merge!
sd_mecha.merge(recipe, output="path/to/model_out.safetensors")

See the examples directory for more examples.

Get Model-Specific Information

The library uses a "model config" to designate any specific set of keys of a certain shape.

It is possible to list all available model configs through the sd_mecha.extensions.model_configs module:

from sd_mecha.extensions import model_configs

all_configs = model_configs.get_all()

print([config.identifier for config in all_configs])
# ["sd1-ldm-base", "sdxl-sgm-base", "sd3-sgm-base", ...]

A component of a model config is a subset of keys of the config that belong to the same logical group. For example, all keys starting with "first_stage_model." in Stable Diffusion models belong to the component "vae".

It is possible to query the different components of a model config:

from sd_mecha.extensions import model_configs

config = model_configs.resolve("sd1-ldm")
for component_id, component in config.components().items():
      # component.keys contains the state dict keys that the component owns
      print(f"{component_id}")

# this prints:
#   clip_l
#   vae
#   diffuser

Motivation

Keeping track of full merge recipes has always been a problem for me. I needed something that allows to store merge recipes in a readable format while also being executable. I also needed something that allows to fully merge an entire tree of models without having to save intermediate models to disk.

Typically, mergers load all models in memory before initiating the merge process. This can be very inefficient when the merge focuses on each key individually:

sd-mecha doesn't have this problem as it saves keys as soon as it can:

This allows to merge a very large number of models simultaneously on low-end hardware.