Llamacpp imatrix Quantizations of EVA-LLaMA-3.33-70B-v0.0

Using llama.cpp release b4273 for quantization.

Original model: https://huggingface.co/EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0

All quants made using imatrix option with dataset from here

Run them in LM Studio

Prompt format

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Download a file (not the whole branch) from below:

Filename	Quant type	File Size	Split	Description
EVA-LLaMA-3.33-70B-v0.0-Q8_0.gguf	Q8_0	74.98GB	true	Extremely high quality, generally unneeded but max available quant.
EVA-LLaMA-3.33-70B-v0.0-Q6_K.gguf	Q6_K	57.89GB	true	Very high quality, near perfect, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q5_K_M.gguf	Q5_K_M	49.95GB	true	High quality, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q5_K_S.gguf	Q5_K_S	48.66GB	false	High quality, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q4_K_M.gguf	Q4_K_M	42.52GB	false	Good quality, default size for most use cases, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q4_K_S.gguf	Q4_K_S	40.35GB	false	Slightly lower quality with more space savings, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q4_0.gguf	Q4_0	40.12GB	false	Legacy format, offers online repacking for ARM CPU inference.
EVA-LLaMA-3.33-70B-v0.0-IQ4_NL.gguf	IQ4_NL	40.05GB	false	Similar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference.
EVA-LLaMA-3.33-70B-v0.0-Q4_0_8_8.gguf	Q4_0_8_8	39.97GB	false	Optimized for ARM and AVX inference. Requires 'sve' support for ARM (see details below). Don't use on Mac.
EVA-LLaMA-3.33-70B-v0.0-Q4_0_4_8.gguf	Q4_0_4_8	39.97GB	false	Optimized for ARM inference. Requires 'i8mm' support (see details below). Don't use on Mac.
EVA-LLaMA-3.33-70B-v0.0-Q4_0_4_4.gguf	Q4_0_4_4	39.97GB	false	Optimized for ARM inference. Should work well on all ARM chips, not for use with GPUs. Don't use on Mac.
EVA-LLaMA-3.33-70B-v0.0-Q3_K_XL.gguf	Q3_K_XL	38.06GB	false	Uses Q8_0 for embed and output weights. Lower quality but usable, good for low RAM availability.
EVA-LLaMA-3.33-70B-v0.0-IQ4_XS.gguf	IQ4_XS	37.90GB	false	Decent quality, smaller than Q4_K_S with similar performance, recommended.
EVA-LLaMA-3.33-70B-v0.0-Q3_K_L.gguf	Q3_K_L	37.14GB	false	Lower quality but usable, good for low RAM availability.
EVA-LLaMA-3.33-70B-v0.0-Q3_K_M.gguf	Q3_K_M	34.27GB	false	Low quality.
EVA-LLaMA-3.33-70B-v0.0-IQ3_M.gguf	IQ3_M	31.94GB	false	Medium-low quality, new method with decent performance comparable to Q3_K_M.
EVA-LLaMA-3.33-70B-v0.0-Q3_K_S.gguf	Q3_K_S	30.91GB	false	Low quality, not recommended.
EVA-LLaMA-3.33-70B-v0.0-IQ3_XXS.gguf	IQ3_XXS	27.47GB	false	Lower quality, new method with decent performance, comparable to Q3 quants.
EVA-LLaMA-3.33-70B-v0.0-Q2_K_L.gguf	Q2_K_L	27.40GB	false	Uses Q8_0 for embed and output weights. Very low quality but surprisingly usable.
EVA-LLaMA-3.33-70B-v0.0-Q2_K.gguf	Q2_K	26.38GB	false	Very low quality but surprisingly usable.
EVA-LLaMA-3.33-70B-v0.0-IQ2_M.gguf	IQ2_M	24.12GB	false	Relatively low quality, uses SOTA techniques to be surprisingly usable.
EVA-LLaMA-3.33-70B-v0.0-IQ2_S.gguf	IQ2_S	22.24GB	false	Low quality, uses SOTA techniques to be usable.
EVA-LLaMA-3.33-70B-v0.0-IQ2_XS.gguf	IQ2_XS	21.14GB	false	Low quality, uses SOTA techniques to be usable.
EVA-LLaMA-3.33-70B-v0.0-IQ2_XXS.gguf	IQ2_XXS	19.10GB	false	Very low quality, uses SOTA techniques to be usable.
EVA-LLaMA-3.33-70B-v0.0-IQ1_M.gguf	IQ1_M	16.75GB	false	Extremely low quality, not recommended.

Embed/output weights

Some of these quants (Q3_K_XL, Q4_K_L etc) are the standard quantization method with the embeddings and output weights quantized to Q8_0 instead of what they would normally default to.

Downloading using huggingface-cli

Click to view download instructions

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download bartowski/EVA-LLaMA-3.33-70B-v0.0-GGUF --include "EVA-LLaMA-3.33-70B-v0.0-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download bartowski/EVA-LLaMA-3.33-70B-v0.0-GGUF --include "EVA-LLaMA-3.33-70B-v0.0-Q8_0/*" --local-dir ./

You can either specify a new local-dir (EVA-LLaMA-3.33-70B-v0.0-Q8_0) or download them all in place (./)

Q4_0_X_X information

New: Thanks to efforts made to have online repacking of weights in this PR, you can now just use Q4_0 if your llama.cpp has been compiled for your ARM device.

Similarly, if you want to get slightly better performance, you can use IQ4_NL thanks to this PR which will also repack the weights for ARM, though only the 4_4 for now. The loading time may be slower but it will result in an overall speed incrase.

Click to view Q4_0_X_X information

These are *NOT* for Metal (Apple) or GPU (nvidia/AMD/intel) offloading, only ARM chips (and certain AVX2/AVX512 CPUs).

If you're using an ARM chip, the Q4_0_X_X quants will have a substantial speedup. Check out Q4_0_4_4 speed comparisons on the original pull request

To check which one would work best for your ARM chip, you can check AArch64 SoC features (thanks EloyOn!).

If you're using a CPU that supports AVX2 or AVX512 (typically server CPUs and AMD's latest Zen5 CPUs) and are not offloading to a GPU, the Q4_0_8_8 may offer a nice speed as well:

Click to view benchmarks on an AVX2 system (EPYC7702)

model	size	params	backend	threads	test	t/s	% (vs Q4_0)
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp512	204.03 ± 1.03	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp1024	282.92 ± 0.19	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	pp2048	259.49 ± 0.44	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg128	39.12 ± 0.27	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg256	39.31 ± 0.69	100%
qwen2 3B Q4_0	1.70 GiB	3.09 B	CPU	64	tg512	40.52 ± 0.03	100%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp512	301.02 ± 1.74	147%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp1024	287.23 ± 0.20	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	pp2048	262.77 ± 1.81	101%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg128	18.80 ± 0.99	48%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg256	24.46 ± 3.04	83%
qwen2 3B Q4_K_M	1.79 GiB	3.09 B	CPU	64	tg512	36.32 ± 3.59	90%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp512	271.71 ± 3.53	133%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp1024	279.86 ± 45.63	100%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	pp2048	320.77 ± 5.00	124%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg128	43.51 ± 0.05	111%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg256	43.35 ± 0.09	110%
qwen2 3B Q4_0_8_8	1.69 GiB	3.09 B	CPU	64	tg512	42.60 ± 0.31	105%

Q4_0_8_8 offers a nice bump to prompt processing and a small bump to text generation

Which file should I choose?

Click here for details

A great write up with charts showing various performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.

The I-quants are not compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.

Credits

Thank you kalomaze and Dampf for assistance in creating the imatrix calibration dataset.

Thank you ZeroWw for the inspiration to experiment with embed/output.

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski