steampunque/Llama-3.3-70B-Instruct-Hybrid-GGUF

Llama.cpp hybrid layer quantization of Llama 3.3 70B Instruct by meta-llama

Original model: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct

The hybrid quant employs different quantization levels on a per layer basis to enable both high performance and small file size at the same time. The quants employed are all K to avoid slow CPU or older GPU processing of IQ quants. Three quants are available for the model as follows:

Q3_S_H : Smallest Q3_K based quant available

  LAYER_TYPES='[
   [0 ,"Q4_K_M"],[1 ,"Q3_K_L"],[2 ,"Q3_K_M"],[3 ,"Q3_K_S"],[4 ,"Q3_K_S"],[5 ,"Q3_K_S"],[6 ,"Q3_K_S"],[7 ,"Q3_K_S"],
   [8 ,"Q3_K_S"],[9 ,"Q3_K_S"],[10,"Q3_K_S"],[11,"Q3_K_S"],[12,"Q3_K_S"],[13,"Q3_K_S"],[14,"Q3_K_S"],[15,"Q3_K_S"],
   [16,"Q3_K_S"],[17,"Q3_K_S"],[18,"Q3_K_S"],[19,"Q3_K_S"],[20,"Q3_K_S"],[21,"Q3_K_S"],[22,"Q3_K_S"],[23,"Q3_K_S"],
   [24,"Q3_K_S"],[25,"Q3_K_S"],[26,"Q3_K_S"],[27,"Q3_K_S"],[28,"Q3_K_S"],[29,"Q3_K_S"],[30,"Q3_K_S"],[31,"Q3_K_S"],
   [32,"Q3_K_S"],[33,"Q3_K_S"],[34,"Q3_K_S"],[35,"Q3_K_S"],[36,"Q3_K_S"],[37,"Q3_K_S"],[38,"Q3_K_S"],[39,"Q3_K_S"],
   [40,"Q3_K_M"],[41,"Q3_K_S"],[42,"Q3_K_M"],[43,"Q3_K_S"],[44,"Q3_K_M"],[45,"Q3_K_S"],[46,"Q3_K_M"],[47,"Q3_K_S"],
   [48,"Q3_K_M"],[49,"Q3_K_S"],[50,"Q3_K_M"],[51,"Q3_K_S"],[52,"Q3_K_M"],[53,"Q3_K_S"],[54,"Q3_K_M"],[55,"Q3_K_S"],
   [56,"Q3_K_M"],[57,"Q3_K_S"],[58,"Q3_K_M"],[59,"Q3_K_S"],[60,"Q3_K_M"],[61,"Q3_K_S"],[62,"Q3_K_M"],[63,"Q3_K_S"],
   [64,"Q3_K_M"],[65,"Q3_K_M"],[66,"Q3_K_M"],[67,"Q3_K_M"],[68,"Q3_K_M"],[69,"Q3_K_M"],[70,"Q3_K_M"],[71,"Q3_K_M"],
   [72,"Q3_K_M"],[73,"Q3_K_M"],[74,"Q3_K_M"],[75,"Q3_K_M"],[76,"Q3_K_M"],[77,"Q3_K_L"],[78,"Q4_K_S"],[79,"Q4_K_M"]
   ]'
  FLAGS="--token-embedding-type Q4_K --output-tensor-type Q5_K --layer-types-high"

Q3_K_H : Slightly larger Q3_K based quant

   LAYER_TYPES='[
   [0 ,"Q4_K_M"],[1 ,"Q3_K_L"],[2 ,"Q3_K_M"],[3 ,"Q3_K_M"],[4 ,"Q3_K_S"],[5 ,"Q3_K_M"],[6 ,"Q3_K_S"],[7 ,"Q3_K_M"],
   [8 ,"Q3_K_S"],[9 ,"Q3_K_M"],[10,"Q3_K_S"],[11,"Q3_K_M"],[12,"Q3_K_S"],[13,"Q3_K_M"],[14,"Q3_K_S"],[15,"Q3_K_M"],
   [16,"Q3_K_M"],[17,"Q3_K_S"],[18,"Q3_K_M"],[19,"Q3_K_S"],[20,"Q3_K_M"],[21,"Q3_K_S"],[22,"Q3_K_M"],[23,"Q3_K_S"],
   [24,"Q3_K_M"],[25,"Q3_K_S"],[26,"Q3_K_M"],[27,"Q3_K_S"],[28,"Q3_K_M"],[29,"Q3_K_S"],[30,"Q3_K_M"],[31,"Q3_K_S"],
   [32,"Q3_K_M"],[33,"Q3_K_S"],[34,"Q3_K_M"],[35,"Q3_K_S"],[36,"Q3_K_M"],[37,"Q3_K_S"],[38,"Q3_K_M"],[39,"Q3_K_S"],
   [40,"Q3_K_M"],[41,"Q3_K_S"],[42,"Q3_K_M"],[43,"Q3_K_S"],[44,"Q3_K_M"],[45,"Q3_K_S"],[46,"Q3_K_M"],[47,"Q3_K_S"],
   [48,"Q3_K_M"],[49,"Q3_K_S"],[50,"Q3_K_M"],[51,"Q3_K_S"],[52,"Q3_K_M"],[53,"Q3_K_S"],[54,"Q3_K_M"],[55,"Q3_K_S"],
   [56,"Q3_K_M"],[57,"Q3_K_S"],[58,"Q3_K_M"],[59,"Q3_K_S"],[60,"Q3_K_M"],[61,"Q3_K_S"],[62,"Q3_K_M"],[63,"Q3_K_S"],
   [64,"Q3_K_M"],[65,"Q3_K_M"],[66,"Q3_K_M"],[67,"Q3_K_M"],[68,"Q3_K_M"],[69,"Q3_K_M"],[70,"Q3_K_M"],[71,"Q3_K_M"],
   [72,"Q3_K_M"],[73,"Q3_K_M"],[74,"Q3_K_M"],[75,"Q3_K_M"],[76,"Q3_K_L"],[77,"Q3_K_L"],[78,"Q4_K_S"],[79,"Q4_K_M"]
   ]'
  FLAGS="--token-embedding-type Q4_K --output-tensor-type Q5_K --layer-types-high"

Q4_K_H : Largest and best performance quant

  LAYER_TYPES='[
   [0 ,"Q4_K_M"],[1 ,"Q4_K_M"],[2 ,"Q4_K_S"],[3 ,"Q4_K_S"],[4 ,"Q3_K_M"],[5 ,"Q3_K_L"],[6 ,"Q3_K_M"],[7 ,"Q3_K_L"],
   [8 ,"Q3_K_M"],[9 ,"Q3_K_L"],[10,"Q3_K_M"],[11,"Q3_K_L"],[12,"Q3_K_M"],[13,"Q3_K_L"],[14,"Q3_K_M"],[15,"Q3_K_L"],
   [16,"Q3_K_L"],[17,"Q3_K_M"],[18,"Q3_K_L"],[19,"Q3_K_M"],[20,"Q3_K_L"],[21,"Q3_K_M"],[22,"Q3_K_L"],[23,"Q3_K_M"],
   [24,"Q3_K_L"],[25,"Q3_K_M"],[26,"Q3_K_L"],[27,"Q3_K_M"],[28,"Q3_K_L"],[29,"Q3_K_M"],[30,"Q3_K_L"],[31,"Q3_K_M"],
   [32,"Q3_K_L"],[33,"Q3_K_M"],[34,"Q3_K_L"],[35,"Q3_K_M"],[36,"Q3_K_L"],[37,"Q3_K_M"],[38,"Q3_K_L"],[39,"Q3_K_M"],
   [40,"Q3_K_L"],[41,"Q3_K_M"],[42,"Q3_K_L"],[43,"Q3_K_M"],[44,"Q3_K_L"],[45,"Q3_K_M"],[46,"Q3_K_L"],[47,"Q3_K_M"],
   [48,"Q3_K_L"],[49,"Q3_K_M"],[50,"Q3_K_L"],[51,"Q3_K_M"],[52,"Q3_K_L"],[53,"Q3_K_M"],[54,"Q3_K_L"],[55,"Q3_K_M"],
   [56,"Q3_K_L"],[57,"Q3_K_M"],[58,"Q3_K_L"],[59,"Q3_K_M"],[60,"Q3_K_L"],[61,"Q3_K_M"],[62,"Q3_K_L"],[63,"Q3_K_M"],
   [64,"Q4_K_S"],[65,"Q3_K_L"],[66,"Q4_K_S"],[67,"Q3_K_L"],[68,"Q4_K_S"],[69,"Q3_K_L"],[70,"Q4_K_S"],[71,"Q3_K_L"],
   [72,"Q4_K_S"],[73,"Q4_K_S"],[74,"Q4_K_M"],[75,"Q4_K_S"],[76,"Q4_K_M"],[77,"Q5_K_S"],[78,"Q5_K_M"],[79,"Q6_K"  ]
   ]'
   FLAGS="--token-embedding-type Q4_K --output-tensor-type Q6_K"

All three quants were optimized to maintain knowledge preservation and reasoning performance using a small set of curated test/evaluation prompts. All three quants score 100% on the eval prompts but the Q3 quants sometimes get a little goofy, giving wrong answer then correcting itself with the right one, or adding some non sequiter with the answer etc. Q4_K_H is rock solid. Note that use of Q2_K or Q2_K_S was not possible with this model since any Q2 use even at deep layers threw the model immediately into either noncoherence or large knowledge loss.

Comparison:

Quant	size	PPL	Comment
Q3_S_H	32.6e9	4.8	Q3_K dominant with Q4_K embedding
Q3_K_H	33.4e9	4.8	" "
Q3_K_M	34.3e9	4.9	Fails parts of eval prompt set
Q4_K_H	37.5e9	4.5	Best available quant
IQ4_XS	38.3e9	4.4	Q4_K embedding Q6_K output

Usage:

This model may be used together with fixie-ai ultravox-v0_5-llama-3_3-70b or ultravox-v0_6-llama-3_3-70b to enable it to process audio (.mp3 and .wav files) and text inputs and generate text outputs. The mmproj file are made available here: https://huggingface.co/steampunque/ultravox-v0_5-llama-3_3-70b-Hybrid-GGUF , https://huggingface.co/steampunque/ultravox-v0_6-llama-3_3-70b-Hybrid-GGUF More information about running multimedia may be found in the docs in the mtmd readme in the tools directory of the llama.cpp source tree https://github.com/ggml-org/llama.cpp/blob/master/tools/mtmd/README.md.

Benchmarks:

A partial set of benchmarks for the model will eventually be given here: https://huggingface.co/spaces/steampunque/benchlm

Download the file from below:

Link	Type	Size/e9 B	Notes
Llama-3.3-70B-Instruct.Q3_S_H.gguf	Q3_S_H	32.6e9 B	1.7B smaller than Q3_K_M
Llama-3.3-70B-Instruct.Q3_K_H.gguf	Q3_K_H	33.4e9 B	0.9B smaller than Q3_K_M
Llama-3.3-70B-Instruct.Q4_K_H.gguf	Q4_K_H	37.5e9 B	0.8B smaller than IQ4_XS
ultravox-v0_5-llama-3_3-70b.mmproj.gguf	mmproj	1.38e9 B	multimedia projector
ultravox-v0_6-llama-3_3-70b.mmproj.gguf	mmproj	1.38e9 B	multimedia projector

A discussion thread about the hybrid layer quant approach can be found here on the llama.cpp git repository:

https://github.com/ggml-org/llama.cpp/discussions/13040