Merge branch 'upstream' into concedo_experimental

# Conflicts: # .github/workflows/build.yml # Makefile # README.md # docs/backend/SYCL.md # examples/llava/README.md # examples/llava/clip.cpp # examples/run/run.cpp # examples/server/README.md # examples/sycl/run-llama2.sh # ggml/CMakeLists.txt # ggml/src/ggml-cpu/CMakeLists.txt # ggml/src/ggml-cuda/CMakeLists.txt # ggml/src/ggml-hip/CMakeLists.txt # ggml/src/ggml-musa/CMakeLists.txt # ggml/src/ggml-sycl/CMakeLists.txt # tests/test-backend-ops.cpp
2026-05-08 18:30:50 +00:00 · 2025-02-26 18:44:45 +08:00 · 2025-02-26 18:44:45 +08:00 · 722fc2dbf1
commit 722fc2dbf1
parent 50eae1ffeb 53e4db1012
79 changed files with 9330 additions and 5299 deletions
--- a/examples/llava/README-granitevision.md
+++ b/examples/llava/README-granitevision.md
@ -0,0 +1,183 @@
+# Granite Vision
+
+Download the model and point your `GRANITE_MODEL` environment variable to the path.
+
+```bash
+$ git clone https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview
+$ export GRANITE_MODEL=./granite-vision-3.1-2b-preview
+```
+
+
+### 1. Running llava surgery v2.
+First, we need to run the llava surgery script as shown below:
+
+`python llava_surgery_v2.py -C -m $GRANITE_MODEL`
+
+You should see two new files (`llava.clip` and `llava.projector`) written into your model's directory, as shown below.
+
+```bash
+$ ls $GRANITE_MODEL | grep -i llava
+llava.clip
+llava.projector
+```
+
+We should see that the projector and visual encoder get split out into the llava files. Quick check to make sure they aren't empty:
+```python
+import os
+import torch
+
+MODEL_PATH = os.getenv("GRANITE_MODEL")
+if not MODEL_PATH:
+    raise ValueError("env var GRANITE_MODEL is unset!")
+
+encoder_tensors = torch.load(os.path.join(MODEL_PATH, "llava.clip"))
+projector_tensors = torch.load(os.path.join(MODEL_PATH, "llava.projector"))
+
+assert len(encoder_tensors) > 0
+assert len(projector_tensors) > 0
+```
+
+If you actually inspect the `.keys()` of the loaded tensors, you should see a lot of `vision_model` tensors in the `encoder_tensors`, and 5 tensors (`'multi_modal_projector.linear_1.bias'`, `'multi_modal_projector.linear_1.weight'`, `'multi_modal_projector.linear_2.bias'`, `'multi_modal_projector.linear_2.weight'`, `'image_newline'`) in the multimodal `projector_tensors`.
+
+
+### 2. Creating the Visual Component GGUF
+To create the GGUF for the visual components, we need to write a config for the visual encoder; make sure the config contains the correct `image_grid_pinpoints`
+
+
+Note: we refer to this file as `$VISION_CONFIG` later on.
+```json
+{
+    "_name_or_path": "siglip-model",
+    "architectures": [
+      "SiglipVisionModel"
+    ],
+    "image_grid_pinpoints": [
+        [384,768],
+        [384,1152],
+        [384,1536],
+        [384,1920],
+        [384,2304],
+        [384,2688],
+        [384,3072],
+        [384,3456],
+        [384,3840],
+        [768,384],
+        [768,768],
+        [768,1152],
+        [768,1536],
+        [768,1920],
+        [1152,384],
+        [1152,768],
+        [1152,1152],
+        [1536,384],
+        [1536,768],
+        [1920,384],
+        [1920,768],
+        [2304,384],
+        [2688,384],
+        [3072,384],
+        [3456,384],
+        [3840,384]
+    ],
+    "mm_patch_merge_type": "spatial_unpad",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14,
+    "layer_norm_eps": 1e-6,
+    "hidden_act": "gelu_pytorch_tanh",
+    "projection_dim": 0,
+    "vision_feature_layer": [-24, -20, -12, -1]
+}
+```
+
+Create a new directory to hold the visual components, and copy the llava.clip/projector files, as well as the vision config into it.
+
+```bash
+$ ENCODER_PATH=$PWD/visual_encoder
+$ mkdir $ENCODER_PATH
+
+$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin
+$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/
+$ cp $VISION_CONFIG $ENCODER_PATH/config.json
+```
+
+At which point you should have something like this:
+```bash
+$ ls $ENCODER_PATH
+config.json             llava.projector         pytorch_model.bin
+```
+
+Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the siglip visual encoder - in the transformers model, you can find these numbers in the [preprocessor_config.json](https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview/blob/main/preprocessor_config.json).
+```bash
+$ python convert_image_encoder_to_gguf.py \
+    -m $ENCODER_PATH \
+    --llava-projector $ENCODER_PATH/llava.projector \
+    --output-dir $ENCODER_PATH \
+    --clip-model-is-vision \
+    --clip-model-is-siglip \
+    --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+```
+
+this will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the abs path of this file as the `$VISUAL_GGUF_PATH.`
+
+
+### 3. Creating the LLM GGUF.
+The granite vision model contains a granite LLM as its language model. For now, the easiest way to get the GGUF for LLM is by loading the composite model in `transformers` and exporting the LLM so that it can be directly converted with the normal conversion path.
+
+First, set the `LLM_EXPORT_PATH` to the path to export the `transformers` LLM to.
+```
+$ export LLM_EXPORT_PATH=$PWD/granite_vision_llm
+```
+
+```python
+import os
+import transformers
+
+MODEL_PATH = os.getenv("GRANITE_MODEL")
+if not MODEL_PATH:
+    raise ValueError("env var GRANITE_MODEL is unset!")
+
+LLM_EXPORT_PATH = os.getenv("LLM_EXPORT_PATH")
+if not MODEL_PATH:
+    raise ValueError("env var LLM_EXPORT_PATH is unset!")
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH)
+
+# NOTE: granite vision support was added to transformers very recently (4.49);
+# if you get size mismatches, your version is too old.
+# If you are running with an older version, set `ignore_mismatched_sizes=True`
+# as shown below; it won't be loaded correctly, but the LLM part of the model that
+# we are exporting will be loaded correctly.
+model = transformers.AutoModelForImageTextToText.from_pretrained(MODEL_PATH, ignore_mismatched_sizes=True)
+
+tokenizer.save_pretrained(LLM_EXPORT_PATH)
+model.language_model.save_pretrained(LLM_EXPORT_PATH)
+```
+
+Now you can convert the exported LLM to GGUF with the normal converter in the root of the llama cpp project.
+```bash
+$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm.gguf
+...
+$ python convert_hf_to_gguf.py --outfile $LLM_GGUF_PATH $LLM_EXPORT_PATH
+```
+
+
+### 4. Running the Model in Llama cpp
+Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. Sample usage:
+
+Note - the test image shown below can be found [here](https://github-production-user-asset-6210df.s3.amazonaws.com/10740300/415512792-d90d5562-8844-4f34-a0a5-77f62d5a58b5.jpg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20250221%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20250221T054145Z&X-Amz-Expires=300&X-Amz-Signature=86c60be490aa49ef7d53f25d6c973580a8273904fed11ed2453d0a38240ee40a&X-Amz-SignedHeaders=host).
+
+```bash
+$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
+    --mmproj $VISUAL_GGUF_PATH \
+    --image cherry_blossom.jpg \
+    -c 16384 \
+    -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" \
+    --temp 0
+```
+
+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.`
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@ -43,6 +43,7 @@
 #include <map>
 #include <regex>
 #include <stdexcept>
+#include <unordered_set>
 #include <vector>
 #include <sstream>
 #include <cinttypes>
@ -123,6 +124,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
 #define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"

 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@ -447,8 +449,9 @@ struct clip_hparams {

    char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)

-    int32_t image_grid_pinpoints[32];
+    std::vector<int32_t> image_grid_pinpoints;
    int32_t image_crop_resolution;
+    std::unordered_set<int32_t> vision_feature_layer;
 };

 struct clip_layer {
@ -588,6 +591,7 @@ struct clip_ctx {
    struct clip_vision_model vision_model;
    projector_type proj_type = PROJECTOR_TYPE_MLP;

+    int32_t max_feature_layer;
    float image_mean[3];
    float image_std[3];
    bool use_gelu = false;
@ -654,7 +658,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
    const int hidden_size          = hparams.hidden_size;
    const int n_head               = hparams.n_head;
    const int d_head               = hidden_size / n_head;
-    int n_layer                    = hparams.n_layer;
    const float eps                = hparams.eps;
    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};

@ -755,13 +758,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
    }

+    std::vector<struct ggml_tensor *> embedding_stack;
+    const auto & vision_feature_layer = hparams.vision_feature_layer;
+
    // loop over layers
-    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
-        n_layer += 1;
-    }
-    for (int il = 0; il < n_layer - 1; il++) {
+    for (int il = 0; il < ctx->max_feature_layer; il++) {
        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states

+        // If this is an embedding feature layer, save the output.
+        // NOTE: 0 index here refers to the input to the encoder.
+        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
+            embedding_stack.push_back(embeddings);
+        }
+
        //const size_t nb_q_w = model.layers[il].q_w->nb[0];

        // layernorm1
@ -849,7 +858,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        cur = ggml_add(ctx0, embeddings, cur);

        embeddings = cur;
-
    }

    // post-layernorm
@ -860,6 +868,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
    }

+    // final layer is a vision feature layer
+    if (vision_feature_layer.find(ctx->max_feature_layer) != vision_feature_layer.end()) {
+        embedding_stack.push_back(embeddings);
+    }
+
+    // If feature layers are explicitly set, stack them (if we have multiple)
+    if (!embedding_stack.empty()) {
+        embeddings = embedding_stack[0];
+        for (size_t i = 1; i < embedding_stack.size(); i++) {
+            embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
+        }
+    }
+
    // llava projector
    if (ctx->has_llava_projector) {
        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@ -1455,14 +1476,26 @@ if(enable_gpu_clip)
            int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
            int n = gguf_get_arr_n(ctx, idx);
            const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
-            for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
-                hparams.image_grid_pinpoints[i] = pinpoints[i];
+            for (int i = 0; i < n; ++i) {
+                hparams.image_grid_pinpoints.push_back(pinpoints[i]);
            }
-            if (n < 32)
-                hparams.image_grid_pinpoints[n] = 0;
-        } catch (std::runtime_error & /*e*/) {
-            hparams.image_grid_pinpoints[0]=0;
-        }
+        } catch (std::runtime_error & /*e*/) { }
+
+        // Load the vision feature layer indices if they are explicitly provided;
+        // if multiple vision feature layers are present, the values will be concatenated
+        // to form the final visual features.
+        // NOTE: gguf conversions should standardize the values of the vision feature layer to
+        // be non-negative, since we use -1 to mark values as unset here.
+        try {
+            int idx = get_key_idx(ctx, KEY_FEATURE_LAYER);
+            int n = gguf_get_arr_n(ctx, idx);
+
+            const int32_t * vision_feature_layer = (const int32_t *)gguf_get_arr_data(ctx, idx);
+
+            for (int i = 0; i < n; ++i) {
+                hparams.vision_feature_layer.insert(vision_feature_layer[i]);
+            }
+        } catch (std::runtime_error & /*e*/) { }

        try {
            int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
@ -1488,6 +1521,9 @@ if(enable_gpu_clip)
            new_clip->image_std[i]  = std_data[i];
        }

+        // Calculate the deepest feature layer based on hparams and projector type
+        new_clip->max_feature_layer = get_deepest_feature_layer(new_clip);
+
        if (verbosity >= 2) {
            LOG_INF("\n%s: vision model hparams\n", __func__);
            LOG_INF("image_size         %d\n", hparams.image_size);
@ -1501,8 +1537,13 @@ if(enable_gpu_clip)
            LOG_INF("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
            LOG_INF("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
            LOG_INF("v_image_grid_pinpoints: ");
-            for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
-                LOG_INF("%d ", hparams.image_grid_pinpoints[i]);
+            for (const auto & pp : hparams.image_grid_pinpoints) {
+                LOG_INF("%d ", pp);
+            }
+            LOG_INF("\n");
+            LOG_INF("v_vision_feature_layer: ");
+            for (const auto & feature_layer: hparams.vision_feature_layer) {
+                LOG_INF("%d ", feature_layer);
            }
            LOG_INF("\n");
            LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
@ -1741,11 +1782,11 @@ void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
    }
 }

-static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
+void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img) {
    img->nx = nx;
    img->ny = ny;
    img->buf.resize(3 * nx * ny);
-    memcpy(img->buf.data(), data, img->buf.size());
+    memcpy(img->buf.data(), rgb_pixels, img->buf.size());
 }

 bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
@ -1755,7 +1796,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
        LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
        return false;
    }
-    build_clip_img_from_data(data, nx, ny, img);
+    clip_build_img_from_pixels(data, nx, ny, img);
    stbi_image_free(data);
    return true;
 }
@ -1846,14 +1887,14 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
        uint8_t* letterboxed_image = make_new_letterbox_img(data, nx, ny, nc, new_width, new_height);
        if(letterboxed_image!=nullptr)
        {
-            build_clip_img_from_data(letterboxed_image, new_width, new_height, img);
+            clip_build_img_from_pixels(letterboxed_image, new_width, new_height, img);
            free(letterboxed_image);
            letterboxed_image = nullptr;
        }
    }
    else
    {
-        build_clip_img_from_data(data, nx, ny, img);
+        clip_build_img_from_pixels(data, nx, ny, img);
    }
    stbi_image_free(data);
    return true;
@ -2334,10 +2375,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
            }
        }
    } else {
-        if (params.image_grid_pinpoints[0] != 0) {
+        if (!params.image_grid_pinpoints.empty()) {
            // "spatial_unpad" with "anyres" processing for llava-1.6
            std::vector<std::pair<int, int>> possible_resolutions;
-            for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            for (size_t i = 0; i < params.image_grid_pinpoints.size(); i+=2) {
                possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
            }
            std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
@ -2503,7 +2544,14 @@ const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
 }

 const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
-    return ctx->vision_model.hparams.image_grid_pinpoints;
+    if (ctx->vision_model.hparams.image_grid_pinpoints.size()) {
+        return &ctx->vision_model.hparams.image_grid_pinpoints.front();
+    }
+    return nullptr;
+}
+
+size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_grid_pinpoints.size();
 }

 int clip_n_patches(const struct clip_ctx * ctx) {
@ -3038,6 +3086,28 @@ bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
    return ctx->has_qwen2vl_merger;
 }

+// Determine the number of encoder layers to iterate over
+int get_deepest_feature_layer(const struct clip_ctx * ctx) {
+    // Get the index of the second to last layer; this is the
+    // default for models that have a llava projector
+    const auto & hparams = ctx->vision_model.hparams;
+    int n_layer = hparams.n_layer - 1;
+    int deepest_feature_layer = -1;
+
+    // Handle other projectors; incrementing here indicates that we
+    // should use the last encoder layer for the vision features.
+    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
+        n_layer += 1;
+    }
+
+    // If we set explicit vision feature layers, only go up to the deepest one
+    for (const auto & feature_layer : hparams.vision_feature_layer) {
+        if (feature_layer > deepest_feature_layer) {
+            deepest_feature_layer = feature_layer;
+        }
+    }
+    return deepest_feature_layer < 0 ? n_layer : deepest_feature_layer;
+}

 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
    clip_image_f32 clip_img;
--- a/examples/llava/clip.h
+++ b/examples/llava/clip.h
@ -55,6 +55,7 @@ CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);

 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);

 CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
@ -73,6 +74,9 @@ CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);

+/** build image from pixels decoded by other libraries instead of stb_image.h for better performance. The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes */
+CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img);
+
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);

 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
@ -89,11 +93,13 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);

 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);

+CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
+
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);

-CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);

 CLIP_API void set_clip_uses_gpu(bool usegpu);

--- a/examples/llava/convert_image_encoder_to_gguf.py
+++ b/examples/llava/convert_image_encoder_to_gguf.py
@ -6,7 +6,7 @@ import re
 import torch
 import numpy as np
 from gguf import *
-from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel, SiglipVisionModel

 TEXT = "clip.text"
 VISION = "clip.vision"
@ -37,6 +37,18 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b


 def get_tensor_name(name: str) -> str:
+    # Standardize the transformers llava next keys for
+    # image newline / mm projector with the classes in haotian-liu LLaVA
+    if name == "image_newline":
+        return "model.image_newline"
+    if name.startswith("multi_modal_projector"):
+        name = name.replace("multi_modal_projector", "mm")
+        if "linear_1" in name:
+            name = name.replace("linear_1", "0")
+        if "linear_2" in name:
+            name = name.replace("linear_2", "2")
+        return name
+
    if "projection" in name:
        return name
    if "mm_projector" in name:
@ -83,8 +95,14 @@ ap.add_argument("--vision-only", action="store_true", required=False,
                help="Save a vision-only model. It can't be used to encode texts")
 ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
-ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+
+# Selectable visual encoders that are compatible with this script
+encoder_group = ap.add_mutually_exclusive_group()
+encoder_group.add_argument("--clip-model-is-openclip", action="store_true", required=False,
                help="The clip model is from openclip (for ViT-SO400M type))")
+encoder_group.add_argument("--clip-model-is-siglip", action="store_true", required=False,
+                help="the visual encoder is Siglip.")
+
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
@ -109,7 +127,12 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir

-if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+if (
+    args.clip_model_is_vision or
+    not os.path.exists(dir_model + "/vocab.json") or
+    args.clip_model_is_openclip or
+    args.clip_model_is_siglip
+):
    vocab = None
    tokens = None
 else:
@ -137,7 +160,10 @@ ftype = 1
 if args.use_f32:
    ftype = 0

-if args.clip_model_is_vision or args.clip_model_is_openclip:
+if args.clip_model_is_siglip:
+    model = SiglipVisionModel.from_pretrained(dir_model)
+    processor = None
+elif args.clip_model_is_vision or args.clip_model_is_openclip:
    model = CLIPVisionModel.from_pretrained(dir_model)
    processor = None
 else:
@ -187,26 +213,71 @@ else:
 if has_text_encoder:
    assert t_hparams is not None
    assert tokens is not None
+    if args.clip_model_is_siglip:
+        text_projection_dim = 0
+    else:
+        text_projection_dim = t_hparams.get("projection_dim", config["projection_dim"])
    # text_model hparams
    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
-    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32("clip.text.projection_dim", text_projection_dim)
    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
    fout.add_token_list(tokens)

+
+
+def get_non_negative_vision_feature_layers(v_hparams):
+    """
+    Determine the vision feature layer(s) for the llava model, which are indices into the
+    hidden states of the visual encoder. Note that the hidden states array generally takes the
+    form:
+
+        [<emb input>, <output of enc block 0>, ... <output of enc block num_hidden_layers>]
+
+    so feature indices should be offset as n+1 to get the output of encoder block n.
+    We convert all vision feature layers to non-negative so that -1 can be used in
+    the model as an unset value. If no vision feature layer is found, we leave it unset.
+    """
+    num_hidden_layers = v_hparams["num_hidden_layers"]
+    to_non_negative = lambda layer_idx: layer_idx  if layer_idx >= 0 else num_hidden_layers + layer_idx + 1
+    feature_layers_key = None
+    # Key used for llava models in transformers
+    if "vision_feature_layer" in config:
+        feature_layers_key = "vision_feature_layer"
+    # Key used for llava models in the original format
+    elif "mm_vision_select_layer" in config:
+        feature_layers_key = "mm_vision_select_layer"
+    if feature_layers_key is not None:
+        feature_layers = config[feature_layers_key]
+        if isinstance(feature_layers, int):
+            feature_layers = [feature_layers]
+        return [to_non_negative(feature_layer) for feature_layer in feature_layers]
+
+# Determine if we have explicitly specified vision feature layers in our config
+feature_layers = get_non_negative_vision_feature_layers(v_hparams)
+
 if has_vision_encoder:
-    # vision_model hparams
+    # Siglip does not have a visual projector; set projection dim to 0
+    if args.clip_model_is_siglip:
+        visual_projection_dim = 0
+    else:
+        visual_projection_dim = v_hparams.get("projection_dim", config["projection_dim"])
+
+    # set vision_model hparams
    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
-    fout.add_uint32("clip.vision.projection_dim", v_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32("clip.vision.projection_dim", visual_projection_dim)
    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
-    block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
+    if feature_layers:
+        block_count = max(feature_layers)
+    else:
+        block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
                            #     /**
                            #      "image_grid_pinpoints": [
@ -258,7 +329,8 @@ if has_vision_encoder:
        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
    if "mm_projector_type" in v_hparams:
        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
-
+    if feature_layers:
+        fout.add_array("clip.vision.feature_layer", feature_layers)

    if processor is not None:
        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean  # pyright: ignore[reportAttributeAccessIssue]
@ -274,7 +346,13 @@ fout.add_bool("clip.use_gelu", use_gelu)


 if has_llava_projector:
-    model.vision_model.encoder.layers.pop(-1)
+    # By default, we drop the last layer for llava projector
+    # models unless we have explicitly set vision feature layers
+    if feature_layers is None:
+        model.vision_model.encoder.layers.pop(-1)
+    else:
+        model.vision_model.encoder.layers = model.vision_model.encoder.layers[:max(feature_layers)]
+
    projector = torch.load(args.llava_projector)
    for name, data in projector.items():
        name = get_tensor_name(name)
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@ -353,9 +353,10 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);

        const int32_t * image_grid = clip_image_grid(ctx_clip);
+        const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);

        std::vector<std::pair<int, int>> grid_pinpoints;
-        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+        for (size_t i = 0; i < num_gridpoints; i += 2) {
            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
        }

@ -405,7 +406,8 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }

 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    int num_max_patches = 6;
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
    if (clip_is_minicpmv(ctx_clip)) {
        num_max_patches = 10;
    }
--- a/examples/llava/llava_surgery_v2.py
+++ b/examples/llava/llava_surgery_v2.py
@ -33,6 +33,33 @@ def save_model(model, file_path, file_type):
    else:
        torch.save(model, file_path)

+# Helpers to match weight names from specific components or
+# determine if a saved shard contains that component
+def is_vision_tower(weight_name):
+    return (
+        weight_name.startswith("model.vision_tower") or
+        weight_name.startswith("vit.") or
+        weight_name.startswith("vision_tower")
+    )
+
+def is_newline(weight_name):
+    return (
+        weight_name.startswith("model.image_newline") or
+        weight_name.startswith("image_newline")
+    )
+
+def is_mm_projector(weight_name):
+    return (
+        weight_name.startswith("model.mm_projector") or
+        weight_name.startswith("vision_proj.") or
+        weight_name.startswith("multi_modal_projector")
+    )
+
+def newline_criteria(checkpoint):
+    return any(is_newline(k) for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(is_mm_projector(k) for k in checkpoint.keys())

 # Adapted function to clean vision tower from checkpoint
 def clean_vision_tower_from_checkpoint(checkpoint_path):
@ -40,7 +67,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
    # file_type = 'pytorch'
    model_path = os.path.dirname(checkpoint_path)
    print(f"Searching for vision tower tensors in {checkpoint_path}")
-    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+    clip_tensors = [k for k, v in checkpoint.items() if is_vision_tower(k)]

    if len(clip_tensors) > 0:
        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
@ -84,12 +111,6 @@ def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):

    return newline_checkpoint_path, projector_checkpoint_path

-def newline_criteria(checkpoint):
-    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
-
-def proj_criteria(checkpoint):
-    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
-

 # Command-line interface setup
 ap = argparse.ArgumentParser()
@ -123,14 +144,14 @@ first_checkpoint = None
 if newline_checkpoint_path is not None:
    print(f"Taking newline from {newline_checkpoint_path}")
    first_checkpoint, file_type = load_model(newline_checkpoint_path)
-    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if is_newline(k)]

 # Load the checkpoint
 mm_tensors = []
 last_checkpoint = None
 if projector_checkpoint_path is not None:
    last_checkpoint, file_type = load_model(projector_checkpoint_path)
-    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+    mm_tensors = [k for k, v in last_checkpoint.items() if is_mm_projector(k)]

 if len(mm_tensors) == 0:
    if last_checkpoint is not None:
@ -155,5 +176,5 @@ if len(projector) > 0:
    save_model(projector, f"{args.model}/llava.projector", 'pytorch')

 print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@ -521,8 +521,13 @@ static json oaicompat_completion_params_parse(const json & body) {
        throw std::runtime_error("Only one completion choice is allowed");
    }

+    // Handle "echo" field
+    if (json_value(body, "echo", false)) {
+        throw std::runtime_error("Only no echo is supported");
+    }
+
    // Params supported by OAI but unsupported by llama.cpp
-    static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
+    static const std::vector<std::string> unsupported_params { "best_of", "suffix" };
    for (const auto & param : unsupported_params) {
        if (body.contains(param)) {
            throw std::runtime_error("Unsupported param: " + param);
@ -598,7 +603,7 @@ static json oaicompat_completion_params_parse(
    inputs.tool_choice           = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto")));
    inputs.json_schema           = json_schema.is_null() ? "" : json_schema.dump();
    inputs.grammar               = grammar;
-    inputs.add_generation_prompt = true;
+    inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
    inputs.use_jinja             = use_jinja;
    inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
    inputs.extract_reasoning     = reasoning_format != COMMON_REASONING_FORMAT_NONE;