πŸ“‹ Model Description

quantized_by: bartowski pipeline_tag: image-text-to-text language:
  • en
license: apache-2.0 base_model: Qwen/Qwen2-VL-2B-Instruct tags:
  • multimodal

Llamacpp imatrix Quantizations of Qwen2-VL-2B-Instruct

Using llama.cpp release b4327 for quantization.

Original model: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct

All quants made using imatrix option with dataset from here

How to run

Since this a new vision model, I'll add special instructions this one time

If you've build llama.cpp locally, you'll want to run:

./llama-qwen2vl-cli -m /models/Qwen2-VL-2B-Instruct-Q40.gguf --mmproj /models/mmproj-Qwen2-VL-2B-Instruct-f32.gguf -p 'Describe this image.' --image '/models/testimage.jpg'

And the model will output the answer. Very simple stuff, similar to other llava models, just make sure you use the correct binary!

Prompt format

<|im_start|>system
{systemprompt}<|imend|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

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

FilenameQuant typeFile SizeSplitDescription
Qwen2-VL-2B-Instruct-f16.gguff163.09GBfalseFull F16 weights.
Qwen2-VL-2B-Instruct-Q80.ggufQ801.65GBfalseExtremely high quality, generally unneeded but max available quant.
Qwen2-VL-2B-Instruct-Q6KL.ggufQ6KL1.33GBfalseUses Q80 for embed and output weights. Very high quality, near perfect, recommended.
Qwen2-VL-2B-Instruct-Q6K.ggufQ6K1.27GBfalseVery high quality, near perfect, recommended.
Qwen2-VL-2B-Instruct-Q5KL.ggufQ5KL1.18GBfalseUses Q80 for embed and output weights. High quality, recommended.
Qwen2-VL-2B-Instruct-Q5KM.ggufQ5K_M1.13GBfalseHigh quality, recommended.
Qwen2-VL-2B-Instruct-Q5KS.ggufQ5K_S1.10GBfalseHigh quality, recommended.
Qwen2-VL-2B-Instruct-Q4KL.ggufQ4KL1.04GBfalseUses Q80 for embed and output weights. Good quality, recommended.
Qwen2-VL-2B-Instruct-Q4KM.ggufQ4K_M0.99GBfalseGood quality, default size for most use cases, recommended.
Qwen2-VL-2B-Instruct-Q4KS.ggufQ4K_S0.94GBfalseSlightly lower quality with more space savings, recommended.
Qwen2-VL-2B-Instruct-Q40.ggufQ400.94GBfalseLegacy format, offers online repacking for ARM and AVX CPU inference.
Qwen2-VL-2B-Instruct-IQ4NL.ggufIQ4NL0.94GBfalseSimilar to IQ4_XS, but slightly larger. Offers online repacking for ARM CPU inference.
Qwen2-VL-2B-Instruct-Q3KXL.ggufQ3KXL0.94GBfalseUses Q80 for embed and output weights. Lower quality but usable, good for low RAM availability.
Qwen2-VL-2B-Instruct-IQ4XS.ggufIQ4XS0.90GBfalseDecent quality, smaller than Q4KS with similar performance, recommended.
Qwen2-VL-2B-Instruct-Q3KL.ggufQ3K_L0.88GBfalseLower quality but usable, good for low RAM availability.
Qwen2-VL-2B-Instruct-Q3KM.ggufQ3K_M0.82GBfalseLow quality.
Qwen2-VL-2B-Instruct-IQ3M.ggufIQ3M0.78GBfalseMedium-low quality, new method with decent performance comparable to Q3KM.
Qwen2-VL-2B-Instruct-Q3KS.ggufQ3K_S0.76GBfalseLow quality, not recommended.
Qwen2-VL-2B-Instruct-IQ3XS.ggufIQ3XS0.73GBfalseLower quality, new method with decent performance, slightly better than Q3KS.
Qwen2-VL-2B-Instruct-Q2KL.ggufQ2KL0.73GBfalseUses Q80 for embed and output weights. Very low quality but surprisingly usable.
Qwen2-VL-2B-Instruct-Q2K.ggufQ2K0.68GBfalseVery low quality but surprisingly usable.
Qwen2-VL-2B-Instruct-IQ2M.ggufIQ2M0.60GBfalseRelatively low quality, uses SOTA techniques to be surprisingly usable.

Embed/output weights

Some of these quants (Q3KXL, Q4KL 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/Qwen2-VL-2B-Instruct-GGUF --include "Qwen2-VL-2B-Instruct-Q4KM.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/Qwen2-VL-2B-Instruct-GGUF --include "Qwen2-VL-2B-Instruct-Q8_0/*" --local-dir ./

You can either specify a new local-dir (Qwen2-VL-2B-Instruct-Q8_0) or download them all in place (./)

ARM/AVX information

Previously, you would download Q4044/48/8_8, and these would have their weights interleaved in memory in order to improve performance on ARM and AVX machines by loading up more data in one pass.

Now, however, there is something called "online repacking" for weights. details in this PR. If you use Q40 and your hardware would benefit from repacking weights, it will do it automatically on the fly.

As of llama.cpp build b4282 you will not be able to run the Q40XX files and will instead need to use Q4_0.

Additionally, if you want to get slightly better quality for , you can use IQ4NL 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 Q40X_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 Q40XX quants will have a substantial speedup. Check out Q4044 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 Q408_8 may offer a nice speed as well:


Click to view benchmarks on an AVX2 system (EPYC7702)

modelsizeparamsbackendthreadstestt/s% (vs Q4_0)
qwen2 3B Q4_01.70 GiB3.09 BCPU64pp512204.03 Β± 1.03100%
qwen2 3B Q4_01.70 GiB3.09 BCPU64pp1024282.92 Β± 0.19100%
qwen2 3B Q4_01.70 GiB3.09 BCPU64pp2048259.49 Β± 0.44100%
qwen2 3B Q4_01.70 GiB3.09 BCPU64tg12839.12 Β± 0.27100%
qwen2 3B Q4_01.70 GiB3.09 BCPU64tg25639.31 Β± 0.69100%
qwen2 3B Q4_01.70 GiB3.09 BCPU64tg51240.52 Β± 0.03100%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64pp512301.02 Β± 1.74147%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64pp1024287.23 Β± 0.20101%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64pp2048262.77 Β± 1.81101%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64tg12818.80 Β± 0.9948%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64tg25624.46 Β± 3.0483%
qwen2 3B Q4KM1.79 GiB3.09 BCPU64tg51236.32 Β± 3.5990%
qwen2 3B Q408_81.69 GiB3.09 BCPU64pp512271.71 Β± 3.53133%
qwen2 3B Q408_81.69 GiB3.09 BCPU64pp1024279.86 Β± 45.63100%
qwen2 3B Q408_81.69 GiB3.09 BCPU64pp2048320.77 Β± 5.00124%
qwen2 3B Q408_81.69 GiB3.09 BCPU64tg12843.51 Β± 0.05111%
qwen2 3B Q408_81.69 GiB3.09 BCPU64tg25643.35 Β± 0.09110%
qwen2 3B Q408_81.69 GiB3.09 BCPU64tg51242.60 Β± 0.31105%
Q408_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 'QXKX', like Q5KM.

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 IQXX, like IQ3M. 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

πŸ“‚ GGUF File List

πŸ“ Filename πŸ“¦ Size ⚑ Download
Qwen2-VL-2B-Instruct-IQ2_M.gguf
LFS Q2
573.21 MB Download
Qwen2-VL-2B-Instruct-IQ3_M.gguf
LFS Q3
740.68 MB Download
Qwen2-VL-2B-Instruct-IQ3_XS.gguf
LFS Q3
697.8 MB Download
Qwen2-VL-2B-Instruct-IQ4_NL.gguf
LFS Q4
892.95 MB Download
Qwen2-VL-2B-Instruct-IQ4_XS.gguf
LFS Q4
854.24 MB Download
Qwen2-VL-2B-Instruct-Q2_K.gguf
LFS Q2
644.97 MB Download
Qwen2-VL-2B-Instruct-Q2_K_L.gguf
LFS Q2
698.88 MB Download
Qwen2-VL-2B-Instruct-Q3_K_L.gguf
LFS Q3
839.39 MB Download
Qwen2-VL-2B-Instruct-Q3_K_M.gguf
LFS Q3
786 MB Download
Qwen2-VL-2B-Instruct-Q3_K_S.gguf
LFS Q3
725.69 MB Download
Qwen2-VL-2B-Instruct-Q3_K_XL.gguf
LFS Q3
893.29 MB Download
Qwen2-VL-2B-Instruct-Q4_0.gguf
Recommended LFS Q4
894.1 MB Download
Qwen2-VL-2B-Instruct-Q4_K_L.gguf
LFS Q4
994.27 MB Download
Qwen2-VL-2B-Instruct-Q4_K_M.gguf
LFS Q4
940.37 MB Download
Qwen2-VL-2B-Instruct-Q4_K_S.gguf
LFS Q4
896.75 MB Download
Qwen2-VL-2B-Instruct-Q5_K_L.gguf
LFS Q5
1.1 GB Download
Qwen2-VL-2B-Instruct-Q5_K_M.gguf
LFS Q5
1.05 GB Download
Qwen2-VL-2B-Instruct-Q5_K_S.gguf
LFS Q5
1.02 GB Download
Qwen2-VL-2B-Instruct-Q6_K.gguf
LFS Q6
1.19 GB Download
Qwen2-VL-2B-Instruct-Q6_K_L.gguf
LFS Q6
1.24 GB Download
Qwen2-VL-2B-Instruct-Q8_0.gguf
LFS Q8
1.53 GB Download
Qwen2-VL-2B-Instruct-f16.gguf
LFS FP16
2.88 GB Download
mmproj-Qwen2-VL-2B-Instruct-f16.gguf
LFS FP16
1.24 GB Download
mmproj-Qwen2-VL-2B-Instruct-f32.gguf
LFS
2.48 GB Download