Zyphra

/

ZAYA1-8B

like 587

Follow

Zyphra 1.19k

Since the June refactor moved the original 80-layer architecture to
Zyphra/ZAYA1-8B-legacy, quants built against it (like mine) don't show up
in this repo's Quantizations list. Leaving them here in case anyone's still
running the pre-refactor checkpoint and wants a 4-bit weights+activations
option:

• Ttimms/zaya1-8b-nvfp4-w4a4-uniform — 6.02 GB, zero BF16 exemptions
• Ttimms/zaya1-8b-nvfp4-w4a4 — 9.46 GB, best measured accuracy

NVFP4 W4A4 (weights and activations quantized) on a 16 GB consumer Blackwell
card (RTX 5070 Ti), built with compressed-tensors + a from-source vLLM SM120
CUTLASS build. Measured accuracy (paired McNemar vs. the 9.46 GB build) and
throughput are documented on the model cards, including a known CUDA-graph
correctness issue on SM120 and how to avoid it (enforce_eager=True).

Full writeup and reproduction steps: https://github.com/t-timms/zaya1-godspeed