|
| 1 | +# Granite Vision |
| 2 | + |
| 3 | +Download the model and point your `GRANITE_MODEL` environment variable to the path. |
| 4 | + |
| 5 | +```bash |
| 6 | +$ git clone https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview |
| 7 | +$ export GRANITE_MODEL=./granite-vision-3.1-2b-preview |
| 8 | +``` |
| 9 | + |
| 10 | + |
| 11 | +### 1. Running llava surgery v2. |
| 12 | +First, we need to run the llava surgery script as shown below: |
| 13 | + |
| 14 | +`python llava_surgery_v2.py -C -m $GRANITE_MODEL` |
| 15 | + |
| 16 | +You should see two new files (`llava.clip` and `llava.projector`) written into your model's directory, as shown below. |
| 17 | + |
| 18 | +```bash |
| 19 | +$ ls $GRANITE_MODEL | grep -i llava |
| 20 | +llava.clip |
| 21 | +llava.projector |
| 22 | +``` |
| 23 | + |
| 24 | +We should see that the projector and visual encoder get split out into the llava files. Quick check to make sure they aren't empty: |
| 25 | +```python |
| 26 | +import os |
| 27 | +import torch |
| 28 | + |
| 29 | +MODEL_PATH = os.getenv("GRANITE_MODEL") |
| 30 | +if not MODEL_PATH: |
| 31 | + raise ValueError("env var GRANITE_MODEL is unset!") |
| 32 | + |
| 33 | +encoder_tensors = torch.load(os.path.join(MODEL_PATH, "llava.clip")) |
| 34 | +projector_tensors = torch.load(os.path.join(MODEL_PATH, "llava.projector")) |
| 35 | + |
| 36 | +assert len(encoder_tensors) > 0 |
| 37 | +assert len(projector_tensors) > 0 |
| 38 | +``` |
| 39 | + |
| 40 | +If you actually inspect the `.keys()` of the loaded tensors, you should see a lot of `vision_model` tensors in the `encoder_tensors`, and 5 tensors (`'multi_modal_projector.linear_1.bias'`, `'multi_modal_projector.linear_1.weight'`, `'multi_modal_projector.linear_2.bias'`, `'multi_modal_projector.linear_2.weight'`, `'image_newline'`) in the multimodal `projector_tensors`. |
| 41 | + |
| 42 | + |
| 43 | +### 2. Creating the Visual Component GGUF |
| 44 | +To create the GGUF for the visual components, we need to write a config for the visual encoder; make sure the config contains the correct `image_grid_pinpoints` |
| 45 | + |
| 46 | + |
| 47 | +Note: we refer to this file as `$VISION_CONFIG` later on. |
| 48 | +```json |
| 49 | +{ |
| 50 | + "_name_or_path": "siglip-model", |
| 51 | + "architectures": [ |
| 52 | + "SiglipVisionModel" |
| 53 | + ], |
| 54 | + "image_grid_pinpoints": [ |
| 55 | + [384,768], |
| 56 | + [384,1152], |
| 57 | + [384,1536], |
| 58 | + [384,1920], |
| 59 | + [384,2304], |
| 60 | + [384,2688], |
| 61 | + [384,3072], |
| 62 | + [384,3456], |
| 63 | + [384,3840], |
| 64 | + [768,384], |
| 65 | + [768,768], |
| 66 | + [768,1152], |
| 67 | + [768,1536], |
| 68 | + [768,1920], |
| 69 | + [1152,384], |
| 70 | + [1152,768], |
| 71 | + [1152,1152], |
| 72 | + [1536,384], |
| 73 | + [1536,768], |
| 74 | + [1920,384], |
| 75 | + [1920,768], |
| 76 | + [2304,384], |
| 77 | + [2688,384], |
| 78 | + [3072,384], |
| 79 | + [3456,384], |
| 80 | + [3840,384] |
| 81 | + ], |
| 82 | + "mm_patch_merge_type": "spatial_unpad", |
| 83 | + "hidden_size": 1152, |
| 84 | + "image_size": 384, |
| 85 | + "intermediate_size": 4304, |
| 86 | + "model_type": "siglip_vision_model", |
| 87 | + "num_attention_heads": 16, |
| 88 | + "num_hidden_layers": 27, |
| 89 | + "patch_size": 14, |
| 90 | + "layer_norm_eps": 1e-6, |
| 91 | + "hidden_act": "gelu_pytorch_tanh", |
| 92 | + "projection_dim": 0, |
| 93 | + "vision_feature_layer": [-24, -20, -12, -1] |
| 94 | +} |
| 95 | +``` |
| 96 | + |
| 97 | +Create a new directory to hold the visual components, and copy the llava.clip/projector files, as well as the vision config into it. |
| 98 | + |
| 99 | +```bash |
| 100 | +$ ENCODER_PATH=$PWD/visual_encoder |
| 101 | +$ mkdir $ENCODER_PATH |
| 102 | + |
| 103 | +$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin |
| 104 | +$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/ |
| 105 | +$ cp $VISION_CONFIG $ENCODER_PATH/config.json |
| 106 | +``` |
| 107 | + |
| 108 | +At which point you should have something like this: |
| 109 | +```bash |
| 110 | +$ ls $ENCODER_PATH |
| 111 | +config.json llava.projector pytorch_model.bin |
| 112 | +``` |
| 113 | + |
| 114 | +Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the siglip visual encoder - in the transformers model, you can find these numbers in the [preprocessor_config.json](https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview/blob/main/preprocessor_config.json). |
| 115 | +```bash |
| 116 | +$ python convert_image_encoder_to_gguf.py \ |
| 117 | + -m $ENCODER_PATH \ |
| 118 | + --llava-projector $ENCODER_PATH/llava.projector \ |
| 119 | + --output-dir $ENCODER_PATH \ |
| 120 | + --clip-model-is-vision \ |
| 121 | + --clip-model-is-siglip \ |
| 122 | + --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 |
| 123 | +``` |
| 124 | + |
| 125 | +this will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the abs path of this file as the `$VISUAL_GGUF_PATH.` |
| 126 | + |
| 127 | + |
| 128 | +### 3. Creating the LLM GGUF. |
| 129 | +The granite vision model contains a granite LLM as its language model. For now, the easiest way to get the GGUF for LLM is by loading the composite model in `transformers` and exporting the LLM so that it can be directly converted with the normal conversion path. |
| 130 | + |
| 131 | +First, set the `LLM_EXPORT_PATH` to the path to export the `transformers` LLM to. |
| 132 | +``` |
| 133 | +$ export LLM_EXPORT_PATH=$PWD/granite_vision_llm |
| 134 | +``` |
| 135 | + |
| 136 | +```python |
| 137 | +import os |
| 138 | +import transformers |
| 139 | + |
| 140 | +MODEL_PATH = os.getenv("GRANITE_MODEL") |
| 141 | +if not MODEL_PATH: |
| 142 | + raise ValueError("env var GRANITE_MODEL is unset!") |
| 143 | + |
| 144 | +LLM_EXPORT_PATH = os.getenv("LLM_EXPORT_PATH") |
| 145 | +if not MODEL_PATH: |
| 146 | + raise ValueError("env var LLM_EXPORT_PATH is unset!") |
| 147 | + |
| 148 | +tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH) |
| 149 | + |
| 150 | +# NOTE: granite vision support was added to transformers very recently (4.49); |
| 151 | +# if you get size mismatches, your version is too old. |
| 152 | +# If you are running with an older version, set `ignore_mismatched_sizes=True` |
| 153 | +# as shown below; it won't be loaded correctly, but the LLM part of the model that |
| 154 | +# we are exporting will be loaded correctly. |
| 155 | +model = transformers.AutoModelForImageTextToText.from_pretrained(MODEL_PATH, ignore_mismatched_sizes=True) |
| 156 | + |
| 157 | +tokenizer.save_pretrained(LLM_EXPORT_PATH) |
| 158 | +model.language_model.save_pretrained(LLM_EXPORT_PATH) |
| 159 | +``` |
| 160 | + |
| 161 | +Now you can convert the exported LLM to GGUF with the normal converter in the root of the llama cpp project. |
| 162 | +```bash |
| 163 | +$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm.gguf |
| 164 | +... |
| 165 | +$ python convert_hf_to_gguf.py --outfile $LLM_GGUF_PATH $LLM_EXPORT_PATH |
| 166 | +``` |
| 167 | + |
| 168 | + |
| 169 | +### 4. Running the Model in Llama cpp |
| 170 | +Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. Sample usage: |
| 171 | + |
| 172 | +Note - the test image shown below can be found [here](https://github-production-user-asset-6210df.s3.amazonaws.com/10740300/415512792-d90d5562-8844-4f34-a0a5-77f62d5a58b5.jpg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20250221%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20250221T054145Z&X-Amz-Expires=300&X-Amz-Signature=86c60be490aa49ef7d53f25d6c973580a8273904fed11ed2453d0a38240ee40a&X-Amz-SignedHeaders=host). |
| 173 | + |
| 174 | +```bash |
| 175 | +$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \ |
| 176 | + --mmproj $VISUAL_GGUF_PATH \ |
| 177 | + --image cherry_blossom.jpg \ |
| 178 | + -c 16384 \ |
| 179 | + -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat type of flowers are in this picture?\n<|assistant|>\n" \ |
| 180 | + --temp 0 |
| 181 | +``` |
| 182 | + |
| 183 | +Sample response: `The flowers in the picture are cherry blossoms, which are known for their delicate pink petals and are often associated with the beauty of spring.` |
0 commit comments