Xuan-Son Nguyen de4c07f937 clip : cap max image size 1024 for qwen vl model (#13478) 8 месяцев назад
..
android 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
CMakeLists.txt a634d75d1b mtmd : move helpers to dedicated file (#13442) 8 месяцев назад
README-quantize.md 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
README.md 053367d149 mtmd : support InternVL 2.5 and 3 (#13422) 8 месяцев назад
clip-impl.h 3eac209319 mtmd : support InternVL 3 38B and 78B mmproj (#13443) 8 месяцев назад
clip-quantize-cli.cpp 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
clip.cpp de4c07f937 clip : cap max image size 1024 for qwen vl model (#13478) 8 месяцев назад
clip.h 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
convert_image_encoder_to_gguf.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
deprecation-warning.cpp 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
glmedge-convert-image-encoder-to-gguf.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
glmedge-surgery.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
llava.cpp 27ebfcacba llama : do not crash if there is no CPU backend (#13395) 8 месяцев назад
llava.h 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
llava_surgery.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
llava_surgery_v2.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
minicpmv-convert-image-encoder-to-gguf.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
minicpmv-surgery.py 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
mtmd-cli.cpp 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
mtmd-helper.cpp a634d75d1b mtmd : move helpers to dedicated file (#13442) 8 месяцев назад
mtmd.cpp a634d75d1b mtmd : move helpers to dedicated file (#13442) 8 месяцев назад
mtmd.h a634d75d1b mtmd : move helpers to dedicated file (#13442) 8 месяцев назад
qwen2vl-test.cpp 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
requirements.txt 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
test-1.jpeg 9b61acf060 mtmd : rename llava directory to mtmd (#13311) 8 месяцев назад
tests.sh 053367d149 mtmd : support InternVL 2.5 and 3 (#13422) 8 месяцев назад

README-quantize.md

Quantizing CLIP Visual Projector

This is the tool for quantizing the CLIP visual projector model. Quantization reduces the precision of the model's weights, which can significantly decrease the model size and improve inference speed, often with minimal impact on performance.

Usage

To quantize a CLIP visual projector model, use the following command:

./bin/llama-llava-clip-quantize-cli /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf <type>

After the quantization, the visual projector can be used freely with the existing LLAVA cli (LLAVA, Qwen2VL, etc).

Arguments

  • /path/to/ggml-model-f32.gguf: The path to the input model file in FP32 or FP16 format.
  • /path/to/ggml-model-quantized.gguf: The path where the quantized model will be saved.
  • <type>: The quantization type to apply. This should be an integer corresponding to one of the quantization types defined in the enum ggml_type.

Quantization Types

The following quantization types are supported, based on the enum ggml_type definition:

  • 2 - q4_0: 4-bit quantization with a single scale value.
  • 3 - q4_1: 4-bit quantization with a separate scale value for each block.
  • 6 - q5_0: 5-bit quantization with a single scale value.
  • 7 - q5_1: 5-bit quantization with a separate scale value for each block.
  • 8 - q8_0: 8-bit quantization with a single scale value.

Example

To quantize a model using the q4_0 quantization type, you would run:

./bin/llama-llava-clip-quantize-cli /path/to/ggml-model-f32.gguf /path/to/ggml-model-quantized.gguf 2

This command will generate a quantized model at /path/to/ggml-model-quantized.gguf using the q4_0 quantization method.

Notes

  • Quantization can lead to a loss in model accuracy, depending on the chosen quantization type. It is recommended to evaluate the quantized model's performance on your specific task to ensure it meets your requirements.
  • The quantized model will typically be smaller in size and faster to run, making it more suitable for deployment in resource-constrained environments.