wenmin-wu/nlp-8bit-optimizer-embedding-override

8-Bit Optimizer with Embedding Override

Overview

Optimizer states (momentum + variance in Adam) consume 2x the model size. BitsAndBytes 8-bit AdamW quantizes these states to 8-bit, cutting optimizer memory by ~75%. However, embedding layers are sensitive to quantization noise. Override embedding weights to use 32-bit optimizer states while keeping everything else at 8-bit.

Quick Start

import bitsandbytes as bnb

def set_embedding_parameters_bits(embeddings, optim_bits=32):
    for attr in ("word_embeddings", "position_embeddings", "token_type_embeddings"):
        if hasattr(embeddings, attr):
            bnb.optim.GlobalOptimManager.get_instance().register_module_override(
                getattr(embeddings, attr), "weight", {"optim_bits": optim_bits}
            )

model = AutoModel.from_pretrained("microsoft/deberta-v3-base")
trainable_params = [p for p in model.parameters() if p.requires_grad]

optimizer = bnb.optim.AdamW(trainable_params, lr=2e-5, optim_bits=8)
set_embedding_parameters_bits(model.embeddings)

Workflow

1. Install bitsandbytes (pip install bitsandbytes)
2. Replace torch.optim.AdamW with bnb.optim.AdamW(optim_bits=8)
3. Register 32-bit overrides for embedding submodules via GlobalOptimManager
4. Train normally — optimizer automatically handles quantized state updates

Key Decisions

• Which layers to override: Embeddings are critical; other layers tolerate 8-bit well
• Memory savings: ~75% optimizer memory reduction (32-bit → 8-bit states)
• Compatibility: Works with any PyTorch model, not just HuggingFace
• Combine with: Layer freezing + gradient checkpointing for maximum memory efficiency

References

• Optimization approaches for Transformers