grillycompression 0.1.0

Activation, KV-cache, and communication compression pipelines — optional grilly extension

GrillyCompression

Activation, KV-cache, and communication compression pipelines — optional grilly extension.

Features

• Block DCT Codec — 4x4 DCT + scalar quantization with configurable error bounds
• Activation Compression — 30-60% VRAM savings on intermediate tensors
• KV-Cache Compression — 3-5x compression on cached K/V pages
• Communication Compression — 19-37% gradient volume reduction for multi-GPU
• Adaptive Quality — tight bounds for embeddings, looser for FFN activations
• Error Feedback — residual compression for gradient communication

Quick Start

pip install grillycompression

from grillycompression import BlockDCTCodec, ActivationCompressor, KVCacheCompressor

# Compress activations
compressor = ActivationCompressor(quality=32, adaptive=True)
compressed = compressor.compress(activation_tensor, layer_type="activation")
restored = compressor.decompress(compressed)

# Compress KV-cache pages
kv_comp = KVCacheCompressor(quality=48)
compressed_page = kv_comp.compress_page(k_data, v_data)
k_restored, v_restored = kv_comp.decompress_page(compressed_page)

Requirements

• Python 3.12+
• grilly >= 0.4.0
• numpy

License

MIT