MiSS

MiSS (Matrix Shard Sharing) is a PEFT method that achieves a good balance between model performance and computational efficiency. It requires only a single trainable matrix and introduces a shard-sharing mechanism distinct from LoRA.

The abstract from the paper is:

Parameter-Efficient Fine-Tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), effectively reduce the number of trainable parameters in Large Language Models (LLMs). However, as model scales continue to grow, the demand for computational resources remains a significant challenge. Existing LoRA variants often struggle to strike an optimal balance between adaptability (model performance and convergence speed) and efficiency (computational overhead, memory usage, and initialization time). This paper introduces MiSS (Matrix Shard Sharing), a novel PEFT approach that addresses this trade-off through a simple shard-sharing mechanism. MiSS leverages the insight that a low-rank adaptation can be achieved by decomposing the weight matrix into multiple fragment matrices and utilizing a shared, trainable common fragment. This method constructs the low-rank update matrix through the replication of these shared, partitioned shards. We also propose a hardware-efficient and broadly applicable implementation for MiSS. Extensive experiments conducted on a range of tasks, alongside a systematic analysis of computational performance, demonstrate MiSS’s superiority. The results show that MiSS significantly outperforms standard LoRA and its prominent variants in both model performance metrics and computational efficiency, including initialization speed and training throughput. By effectively balancing expressive power and resource utilization, MiSS offers a compelling solution for efficiently adapting large-scale models.

When to use MiSS

MiSS is a good choice when:

• You want faster initialization and higher training throughput than advanced LoRA initialization schemes that use expensive setups (such as PiSSA, LoRA-GA, or OLoRA).
• You want a drop-in alternative to LoRA with minimal configuration changes.

If you need stronger expressiveness at the cost of some efficiency, consider the bat initialization variant (see below).

init_weights modes

MiSS supports three initialization modes via the init_weights parameter:

• True (default): Standard MiSS initialization. Best starting point for most use cases.
• "bat": Enables nonlinear updates across different shards. Produces better results than standard MiSS but uses more memory and is approximately twice as slow. Use this when performance is the priority over efficiency.
• "mini": Uses a smaller rank along the out_features dimension, controlled by mini_r. This reduces trainable parameters further. When using this mode, mini_r must be set and out_features must be divisible by mini_r.

Quick start

import torch
from peft import MissConfig, get_peft_model
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
tokenizer.pad_token_id = tokenizer.eos_token_id

# Standard MiSS
config = MissConfig(
    r=64,
    miss_dropout=0.01,
    task_type="CAUSAL_LM"
)

# BAT variant — better performance, more memory
# config = MissConfig(
#     r=64,
#     init_weights="bat",
#     task_type="CAUSAL_LM"
# )

# Mini variant — fewer trainable parameters
# config = MissConfig(
#     r=64,
#     init_weights="mini",
#     mini_r=8,
#     task_type="CAUSAL_LM"
# )

model = get_peft_model(model, config)
model.print_trainable_parameters()

For a full fine-tuning example including training and inference, see the MiSS fine-tuning example.

MissConfig

MissModel

>>> from diffusers import StableDiffusionPipeline
>>> from peft import MissModel, MissConfig

>>> config_te = MissConfig(
...     r=8,
...     target_modules=["k_proj", "q_proj", "v_proj", "out_proj", "fc1", "fc2"],
...     init_weights=True,
... )
>>> config_unet = MissConfig(
...     r=8,
...     target_modules=[
...         "proj_in",
...         "proj_out",
...         "to_k",
...         "to_q",
...         "to_v",
...         "to_out.0",
...         "ff.net.0.proj",
...         "ff.net.2",
...     ],
...     init_weights=True,
... )

>>> model = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
>>> model.text_encoder = MissModel(model.text_encoder, config_te, "default")
>>> model.unet = MissModel(model.unet, config_unet, "default")