leafspark's picture
Update README.md
9284f10 verified
|
raw
history blame
1.13 kB
metadata
license: unknown
library_name: transformers
pipeline_tag: text-generation

Deepseek-V2-Chat-GGUF

Quantizised from https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat

Using llama.cpp fork: https://github.com/fairydreaming/llama.cpp/tree/deepseek-v2

Warning: This will not work unless you compile llama.cpp from the repo provided!

How to use:

  • Find the relevant directory
  • Download all files
  • Run merge.py
  • Merged GGUF should appear

Quants:

  • bf16 (finished, currently splitting and uploading) [size: 439gb]
  • f32 (may require some time to upload, after q8_0) [estimated size: ~800gb]
  • q8_0 (after bf16) [estimated size: 233.27gb]
  • q4_k_m (after q8_0) [estimated size: 133.10gb]
  • q2_k (after q4_k_m) [estimated size: ~65gb]
  • q3_k_s (low priority) [estimated size: 96.05gb]

If quantize.exe supports it I will make RTN quants (edit: it doesn't, will try building from fork).

Note: the bf16 GGUF does not have some DeepSeek v2 specific parameters, will look into adding them