about0/llama-v2-chinese-ymcui-alpaca-GGML-13B
1
LLaMA-v2-chinese-alpaca-13B-GGML (ymcui)
Here are the GGML converted and/or quantized models for ymcui's Chinese LLaMA-v2 Alpaca 13B.
!NOTE! The GGML filetype is outdated. Prefer GGUF format going forward.
Explanation of quantisation methods
<details> <summary>Click to see details</summary>
Methods:
- type-0 (Q40, Q50, Q8_0) - weights w are obtained from quants q using w = d * q, where d is the block scale.
- type-1 (Q41, Q51) - weights are given by w = d * q + m, where m is the block minimum
The new methods available are:
- GGMLTYPEQ2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
- GGMLTYPEQ3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
- GGMLTYPEQ4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
- GGMLTYPEQ5K - "type-1" 5-bit quantization. Same super-block structure as GGMLTYPEQ4K resulting in 5.5 bpw
- GGMLTYPEQ6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
- GGMLTYPEQ8K - "type-0" 8-bit quantization. Only used for quantizing intermediate results. The difference to the existing Q80 is that the block size is 256. All 2-6 bit dot products are implemented for this quantization type.
This is exposed via llama.cpp quantization types that define various "quantization mixes" as follows:
- LLAMAFTYPEMOSTLYQ2K - uses GGMLTYPEQ4K for the attention.vw and feedforward.w2 tensors, GGMLTYPEQ2_K for the other tensors.
- LLAMAFTYPEMOSTLYQ3KS - uses GGMLTYPEQ3K for all tensors
- LLAMAFTYPEMOSTLYQ3KM - uses GGMLTYPEQ4K for the attention.wv, attention.wo, and feedforward.w2 tensors, else GGMLTYPEQ3K
- LLAMAFTYPEMOSTLYQ3KL - uses GGMLTYPEQ5K for the attention.wv, attention.wo, and feedforward.w2 tensors, else GGMLTYPEQ3K
- LLAMAFTYPEMOSTLYQ4KS - uses GGMLTYPEQ4K for all tensors
- LLAMAFTYPEMOSTLYQ4KM - uses GGMLTYPEQ6K for half of the attention.wv and feedforward.w2 tensors, else GGMLTYPEQ4K
- LLAMAFTYPEMOSTLYQ5KS - uses GGMLTYPEQ5K for all tensors
- LLAMAFTYPEMOSTLYQ5KM - uses GGMLTYPEQ6K for half of the attention.wv and feedforward.w2 tensors, else GGMLTYPEQ5K
- LLAMAFTYPEMOSTLYQ6K- uses 6-bit quantization (GGMLTYPEQ8_K) for all tensors </details>
Provided files
Model Sources
- Repository: [https://github.com/ymcui/Chinese-LLaMA-Alpaca-2]
