What's that?
The goal: Make a competitive 4-bit quality quant of TheDrummer/Big-Alice-28B-v1 using SOTA quant types from ik_llama.cpp, allowing the resulting gguf to fit into 16gb VRAM without KVO, accounting for system overhead.
The result: Mixed precision quantization of TheDrummer/Big-Alice-28B-v1 utilizing 4-bit trellis quants for FFN tensors that is theoretically smarter and faster than IQ4_XS.
quantized with ik_llama.cpp build: 9d07d868
incompatible with mainline llama.cpp
Layout
| Layer | Dims | Dims | Quant |
|---|---|---|---|
| token_embd | 5120 | 131072.0 | iq4_k |
| blk | 40 | ||
| attn_k | 5120 | 1024 | iq5_ks |
| attn_norm | 5120 | 1 | f32 |
| attn_q | 5120 | 4096 | iq5_ks |
| attn_v | 5120 | 1024 | iq6_k |
| attn_output | 4096 | 5120 | iq5_k |
| ffn_down | 14336 | 5120 | iq4_kt |
| ffn_gate | 5120 | 14336 | iq4_kt |
| ffn_norm | 5120 | 1 | f32 |
| ffn_up | 5120 | 14336 | iq4_kt |
| output | 5120 | 131072 | iq6_k |
| output_norm | 5120 | 1 | f32 |
Conclusion
In retrospect, it probably wasn't the best idea for 16GB VRAM - the PCI-E travel overhead from KV cache in RAM is MASSIVE for upscaled Apriel (especially for my 8-lane 5060 Ti.). This one remains an ok option at 20gb+ VRAM, or on 16gb GPUs with a less impotent PCI-E config. A smaller 3.44 bpw ffn + trellis 4.0 bpw attention quant is in the overn for strictly 16gb inference.
Regardless, this one is provided as is. WYSIWYG.
Cheers
SillyTilly - for the base model.
TheDrummer - for the finetuned model.
mradermacher - for the imatrix + the myriad of quants we all benefit from.
ikawrakow and contributors of ik_llama.cpp - I probably misused your ingenious creation.