Open in tt-awesome →

DFloat11-TT

community
by Tomas Petros · Python ·

DFloat11 lossless LLM weight compression ported to Blackhole. Huffman-codes only the exponent field of BFloat16 weights, cutting them to about 70% of their size with bit-for-bit identical outputs, with a C++ extension built against tt-metal, a compressor and an on-device evaluation path.

dfloat11 compression huffman bfloat16 llm tt-metal cpp
blackhole