Merge 70a23863dc into 549279d804

.gitignore

@@ -60,6 +60,7 @@ models-mnt
 /libllama.so
 /llama-bench
 /llava-cli
+/minicpmv-cli
 /lookahead
 /lookup
 /lookup-create

Makefile

@@ -1,7 +1,7 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search  \
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli minicpmv-cli baby-llama beam-search  \
 	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
@@ -878,6 +878,12 @@ llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/cli
 	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
 
+minicpmv-cli: examples/minicpmv/minicpmv-cli.cpp examples/minicpmv/clip.h examples/minicpmv/clip.cpp examples/minicpmv/minicpmv.h examples/minicpmv/minicpmv.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/minicpmv/clip.cpp  -o $(call GET_OBJ_FILE, examples/minicpmv/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c examples/minicpmv/minicpmv.cpp -o $(call GET_OBJ_FILE, examples/minicpmv/minicpmv.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/minicpmv/clip.cpp examples/minicpmv/minicpmv.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/minicpmv/clip.cpp) $(call GET_OBJ_FILE, examples/minicpmv/minicpmv.cpp) -o $@ $(LDFLAGS)
+
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

examples/CMakeLists.txt

@@ -26,6 +26,7 @@ else()
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
+    add_subdirectory(minicpmv)
     if (LLAMA_SYCL)
         add_subdirectory(sycl)
     endif()

examples/minicpmv/CMakeLists.txt

@@ -0,0 +1,37 @@
+add_library(minicpmv OBJECT
+            minicpmv.cpp
+            minicpmv.h
+            clip.cpp
+            clip.h
+            )
+
+target_link_libraries(minicpmv PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+
+target_include_directories(minicpmv PUBLIC .)
+target_include_directories(minicpmv PUBLIC ../..)
+target_include_directories(minicpmv PUBLIC ../../common)
+
+target_compile_features(minicpmv PRIVATE cxx_std_11)
+
+add_library(minicpmv_static STATIC $<TARGET_OBJECTS:minicpmv>)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(minicpmv PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(minicpmv PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(minicpmv_shared SHARED $<TARGET_OBJECTS:minicpmv>)
+    target_link_libraries(minicpmv_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS minicpmv_shared LIBRARY)
+endif()
+
+if (NOT MSVC)
+    target_compile_options(minicpmv PRIVATE -Wno-cast-qual) # stb_image.h
+endif()
+
+if(TARGET BUILD_INFO)
+    add_dependencies(minicpmv BUILD_INFO)
+endif()
+
+set(TARGET minicpmv-cli)
+add_executable(minicpmv-cli minicpmv-cli.cpp)
+install(TARGETS minicpmv-cli RUNTIME)
+target_link_libraries(minicpmv-cli PRIVATE common minicpmv ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(minicpmv PRIVATE cxx_std_11)

examples/minicpmv/README.md

@@ -0,0 +1,19 @@
+# 所有命令在 llama.cpp 根目录执行，模型位于根目录上级目录处
+# All command should be executed under the root path of llama.cpp repo. We assume the MiniCPM-V-2 model are put in its parent folder.
+
+```bash
+make
+make minicpmv-cli
+
+python ./examples/minicpmv/minicpm-surgery.py -m ../MiniCPM-V-2
+python ./examples/minicpmv/convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2 --llava-projector ../MiniCPM-V-2/llava.projector --output-dir ../MiniCPM-V-2 --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+python ./convert-hf-to-gguf.py ../MiniCPM-V-2/MiniCPM
+./minicpmv-cli -m ../MiniCPM-V-2/MiniCPM/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2/mmproj-model-f16.gguf -c 4096 --temp 0.6 --top-p 0.8 --top-k 100 --repeat-penalty 1.0 --image ../test.jpg -p "这张图里有什么?"
+
+# or run quantize int4 version
+./quantize ../MiniCPM-V-2/MiniCPM/ggml-model-f16.gguf ../MiniCPM-V-2/MiniCPM/ggml-model-Q4_K_M.gguf Q4_K_M
+./minicpmv-cli -m ../MiniCPM-V-2/MiniCPM/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2/mmproj-model-f16.gguf -c 4096 --temp 0.6 --top-p 0.8 --top-k 100 --repeat-penalty 1.0 --image ../test.jpg -p "这张图里有什么?"
+
+# or run in interactive mode
+./minicpmv-cli -m ../MiniCPM-V-2/MiniCPM/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2/mmproj-model-f16.gguf -c 4096 --temp 0.6 --top-p 0.8 --top-k 100 --repeat-penalty 1.0 --image ../test.jpg -i
+```

examples/minicpmv/android/adb_run.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+
+model_dir="/Users/cxt/model/llm/mobileVLM/MobileVLM-1.7B_processed"
+projector_name="mmproj-model-f16.gguf"
+llama_name="ggml-model-q4_k.gguf"
+img_dir="/Users/cxt/model/llm"
+img_name="demo.jpg"
+prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:"
+# img_name="cat.jpeg"
+# prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
+
+program_dir="build_64/bin"
+binName="minicpmv-cli"
+n_threads=4
+
+
+deviceDir="/data/local/tmp"
+saveDir="output"
+if [ ! -d ${saveDir} ]; then
+    mkdir ${saveDir}
+fi
+
+
+function android_run() {
+    # # copy resource into device
+    # adb push ${model_dir}/${projector_name} ${deviceDir}/${projector_name}
+    # adb push ${model_dir}/${llama_name} ${deviceDir}/${llama_name}
+    adb push ${img_dir}/${img_name} ${deviceDir}/${img_name}
+    # copy program into device
+    adb push ${program_dir}/${binName} ${deviceDir}/${binName}
+    adb shell "chmod 0777 ${deviceDir}/${binName}"
+
+    # run
+    adb shell "echo cd ${deviceDir} ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 > ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt"
+    adb shell "cd ${deviceDir}; pwd; ${deviceDir}/${binName} \
+                                                 -m ${deviceDir}/${llama_name} \
+                                                 --mmproj ${deviceDir}/${projector_name} \
+                                                 -t ${n_threads} \
+                                                 --image ${deviceDir}/${img_name} \
+                                                 -p \"${prompt}\" \
+                                                 >> ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt 2>&1"
+    adb pull ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt ${saveDir}
+}
+
+android_run
+
+echo "android_run is Done!"

examples/minicpmv/clip.h

@@ -0,0 +1,85 @@
+#ifndef CLIP_H
+#define CLIP_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define CLIP_API __declspec(dllexport)
+#        else
+#            define CLIP_API __declspec(dllimport)
+#        endif
+#    else
+#        define CLIP_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define CLIP_API
+#endif
+
+struct clip_ctx;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct clip_ctx;
+
+struct clip_image_u8_batch {
+    struct clip_image_u8 * data;
+    size_t size;
+};
+
+struct clip_image_f32_batch {
+    struct clip_image_f32 * data;
+    size_t size;
+};
+
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+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 int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
+
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
+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);
+
+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 */
+CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
+
+/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
+
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // CLIP_H

examples/minicpmv/convert-image-encoder-to-gguf.py

@@ -0,0 +1,403 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+import timm
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if has_llava and name in ["visual_projection.weight"]:
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+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,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+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)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.48145466, 0.4578275, 0.40821073]
+default_image_std = [0.26862954, 0.26130258, 0.27577711]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# 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:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+# if args.clip_model_is_vision or args.clip_model_is_openclip:
+#     model = CLIPVisionModel.from_pretrained(dir_model)
+#     processor = None
+# else:
+#     model = CLIPModel.from_pretrained(dir_model)
+#     processor = CLIPProcessor.from_pretrained(dir_model)
+model = timm.create_model(
+    "vit_so400m_patch14_siglip_384.webli",
+    pretrained=False,
+    num_classes=0,
+    dynamic_img_size=True,
+    dynamic_img_pad=True,
+)
+processor = None
+if model.attn_pool is not None:
+    model.attn_pool = torch.nn.Identity()
+
+model.blocks = model.blocks[:-1]
+model.load_state_dict(torch.load(os.path.join(dir_model, "llava.clip")))
+
+fname_middle = None
+has_text_encoder = True
+has_vision_encoder = True
+has_llava_projector = False
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.llava_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_llava_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_llava_projector", has_llava_projector)
+fout.add_file_type(ftype)
+if args.text_only:
+    fout.add_description("text-only CLIP model")
+elif args.vision_only and not has_llava_projector:
+    fout.add_description("vision-only CLIP model")
+elif has_llava_projector:
+    fout.add_description("image encoder for LLaVA")
+    # add projector type
+    fout.add_string("clip.projector_type", "resampler")
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", 448)
+    fout.add_uint32("clip.vision.patch_size", 14)
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), 1152)
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    block_count = 26
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+
+    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
+        image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
+    else:
+        image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+        image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+use_gelu = True
+fout.add_bool("clip.use_gelu", use_gelu)
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+def _replace_name_resampler(s, v):
+    if re.match("resampler.pos_embed", s):
+        return {
+            s: v,
+            re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(2304, (448//14, 448//14))),
+        }
+    if re.match("resampler.proj", s):
+        return {
+            re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
+        }
+    if re.match("resampler.attn.in_proj_.*", s):
+        return {
+            re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
+            re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
+            re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
+        }
+    return {s: v}
+
+if has_llava_projector:
+    projector = torch.load(args.llava_projector)
+    new_state_dict = {}
+    for k, v in projector.items():
+        kvs = _replace_name_resampler(k, v)
+        for nk, nv in kvs.items():
+            new_state_dict[nk] = nv
+    projector = new_state_dict
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        data = data.squeeze().numpy()
+
+        n_dims = len(data.shape)
+        if ftype == 1:
+            if name[-7:] == ".weight" and n_dims == 2:
+                print("  Converting to float16")
+                data = data.astype(np.float16)
+                ftype_cur = 1
+            else:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+        else:
+            if data.dtype != np.float32:
+                print("  Converting to float32")
+                data = data.astype(np.float32)
+                ftype_cur = 0
+
+        fout.add_tensor(name, data)
+        print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+
+    print("Projector tensors added\n")
+
+def _replace_name(s, v):
+    if re.match("blocks.([0-9]+).attn.qkv.weight", s):
+        return {
+            re.sub("blocks.([0-9]+).attn.qkv.weight", "vision_model.encoder.layers.\\1.self_attn.q_proj.weight", s): v.chunk(3, dim=0)[0],
+            re.sub("blocks.([0-9]+).attn.qkv.weight", "vision_model.encoder.layers.\\1.self_attn.k_proj.weight", s): v.chunk(3, dim=0)[1],
+            re.sub("blocks.([0-9]+).attn.qkv.weight", "vision_model.encoder.layers.\\1.self_attn.v_proj.weight", s): v.chunk(3, dim=0)[2],
+        }
+    if re.match("blocks.([0-9]+).attn.qkv.bias", s):
+        return {
+            re.sub("blocks.([0-9]+).attn.qkv.bias", "vision_model.encoder.layers.\\1.self_attn.q_proj.bias", s): v.chunk(3, dim=0)[0],
+            re.sub("blocks.([0-9]+).attn.qkv.bias", "vision_model.encoder.layers.\\1.self_attn.k_proj.bias", s): v.chunk(3, dim=0)[1],
+            re.sub("blocks.([0-9]+).attn.qkv.bias", "vision_model.encoder.layers.\\1.self_attn.v_proj.bias", s): v.chunk(3, dim=0)[2],
+        }
+    if re.match("pos_embed", s):
+        from timm.layers import resample_abs_pos_embed
+        s = re.sub("pos_embed", "vision_model.embeddings.position_embedding", s)
+        v = resample_abs_pos_embed(v, (448//14, 448//14), num_prefix_tokens=0)
+        return {s: v}
+
+    s = re.sub("patch_embed.proj.weight", "vision_model.embeddings.patch_embedding.proj.weight", s)
+    s = re.sub("patch_embed.proj.bias", "vision_model.embeddings.patch_embedding.proj.bias", s)
+
+    # norm
+    s = re.sub("blocks.([0-9]+).norm([0-9]+).weight", "vision_model.encoder.layers.\\1.layer_norm\\2.weight", s)
+    s = re.sub("blocks.([0-9]+).norm([0-9]+).bias", "vision_model.encoder.layers.\\1.layer_norm\\2.bias", s)
+
+    s = re.sub("blocks.([0-9]+).attn.proj.weight", "vision_model.encoder.layers.\\1.self_attn.out_proj.weight", s)
+    s = re.sub("blocks.([0-9]+).attn.proj.bias", "vision_model.encoder.layers.\\1.self_attn.out_proj.bias", s)
+
+    s = re.sub("blocks.([0-9]+).mlp.fc([0-9]+).weight", "vision_model.encoder.layers.\\1.mlp.fc\\2.weight", s)
+    s = re.sub("blocks.([0-9]+).mlp.fc([0-9]+).bias", "vision_model.encoder.layers.\\1.mlp.fc\\2.bias", s)
+
+    s = re.sub("norm.weight", "vision_model.post_layernorm.weight", s)
+    s = re.sub("norm.bias", "vision_model.post_layernorm.bias", s)
+
+    return {s: v}
+
+state_dict = model.state_dict()
+new_state_dict = {}
+for k, v in state_dict.items():
+    kvs = _replace_name(k, v)
+    for nk, nv in kvs.items():
+        new_state_dict[nk] = nv
+state_dict = new_state_dict
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)

examples/minicpmv/minicpm-surgery.py

@@ -0,0 +1,58 @@
+import argparse
+import glob
+import os, json
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to LLaVA v1.5 model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float().cpu() for name in mm_tensors}
+torch.save(projector, f"{args.model}/llava.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vpm.", ""): checkpoint[name].float().cpu() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/llava.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+config = model.llm.config
+config._name_or_path = "openbmb/CPM-2B"
+config.auto_map = {
+    "AutoConfig": "configuration_minicpm.MiniCPMConfig",
+    "AutoModel": "modeling_minicpm.MiniCPMModel",
+    "AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
+    "AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
+}
+model.llm.save_pretrained(f"{args.model}/MiniCPM")
+tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+tok.save_pretrained(f"{args.model}/MiniCPM")
+os.system(f"cp {args.model}/modeling_minicpm.py {args.model}/MiniCPM/modeling_minicpm.py")
+os.system(f"cp {args.model}/tokenizer.json {args.model}/MiniCPM/tokenizer.json")
+with open(f"{args.model}/MiniCPM/tokenizer_config.json", "r") as f:
+    d = json.load(f)
+    d.pop("auto_map")
+    d["tokenizer_class"] = "LlamaTokenizer"
+    if "add_prefix_space" in d:
+        d.pop("add_prefix_space")
+with open(f"{args.model}/MiniCPM/tokenizer_config.json", "w") as f:
+    json.dump(d, f, indent=2)
+
+
+print("Done!")
+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.")

examples/minicpmv/minicpmv-cli.cpp

@@ -0,0 +1,331 @@
+#include "ggml.h"
+#include "log.h"
+#include "common.h"
+#include "clip.h"
+#include "minicpmv.h"
+#include "llama.h"
+
+#include "base64.hpp"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
+    int N = (int) tokens.size();
+    for (int i = 0; i < N; i += n_batch) {
+        int n_eval = (int) tokens.size() - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
+    std::vector<llama_token> tokens;
+    tokens.push_back(id);
+    return eval_tokens(ctx_llama, tokens, 1, n_past);
+}
+
+static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
+    std::string              str2     = str;
+    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
+    return true;
+}
+
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    static std::string ret;
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
+        ret = "</s>";
+    } else {
+        ret = llama_token_to_piece(ctx_llama, id);
+    }
+    eval_id(ctx_llama, id, n_past);
+    return ret.c_str();
+}
+
+static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
+static const char* IMG_BASE64_TAG_END = "\">";
+
+static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
+    begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
+    end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
+}
+
+static bool prompt_contains_image(const std::string& prompt) {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    return (begin != std::string::npos);
+}
+
+// replaces the base64 image tag in the prompt with `replacement`
+static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
+    size_t img_base64_str_start, img_base64_str_end;
+    find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
+    if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
+        LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        return NULL;
+    }
+
+    auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
+    auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
+    auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
+
+    auto required_bytes = base64::required_encode_size(base64_str.size());
+    auto img_bytes = std::vector<unsigned char>(required_bytes);
+    base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
+
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
+    if (!embed) {
+        LOG_TEE("%s: could not load image from base64 string.\n", __func__);
+        return NULL;
+    }
+
+    return embed;
+}
+
+static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
+    size_t begin, end;
+    find_image_tag_in_prompt(prompt, begin, end);
+    if (begin == std::string::npos || end == std::string::npos) {
+        return prompt;
+    }
+    auto pre = prompt.substr(0, begin);
+    auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
+    return pre + replacement + post;
+}
+
+struct llava_context {
+    struct clip_ctx * ctx_clip = NULL;
+    struct llama_context * ctx_llama = NULL;
+    struct llama_model * model = NULL;
+};
+
+static void show_additional_info(int /*argc*/, char ** argv) {
+    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+}
+
+static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
+
+    // load and preprocess the image
+    llava_image_embed * embed = NULL;
+    auto prompt = params->prompt;
+    if (prompt_contains_image(prompt)) {
+        if (!params->image.empty()) {
+            LOG_TEE("using base64 encoded image instead of command line image path\n");
+        }
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        if (!embed) {
+            LOG_TEE("%s: can't load image from prompt\n", __func__);
+            return NULL;
+        }
+        params->prompt = remove_image_from_prompt(prompt);
+    } else {
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
+        if (!embed) {
+            LOG_TEE("%s: is %s really an image file?\n", __func__, params->image.c_str());
+            return NULL;
+        }
+    }
+
+    return embed;
+}
+
+static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, std::vector<std::vector<struct llava_image_embed *>> image_embed_slices, gpt_params * params, int &n_past) {
+    std::string system_prompt;
+
+    system_prompt = "<用户>";
+
+    eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, true);
+    llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
+    if (!image_embed_slices.empty()) {
+        eval_string(ctx_llava->ctx_llama, std::string("</image><slice>").c_str(), params->n_batch, &n_past, false);
+        for (int i = 0; i < image_embed_slices.size(); ++i) {
+            eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
+            for (int j = 0; j < image_embed_slices[i].size(); ++j) {
+                llava_eval_image_embed(ctx_llava->ctx_llama, image_embed_slices[i][j], params->n_batch, &n_past);
+                if (j != image_embed_slices[i].size() - 1) {
+                    eval_string(ctx_llava->ctx_llama, std::string("</image><image>").c_str(), params->n_batch, &n_past, false);
+                } else {
+                    if (i != image_embed_slices.size() - 1) {
+                        eval_string(ctx_llava->ctx_llama, std::string("</image>\n").c_str(), params->n_batch, &n_past, false);
+                    } else {
+                        eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
+                    }
+                }
+            }
+        }
+        eval_string(ctx_llava->ctx_llama, std::string("</slice>\n").c_str(), params->n_batch, &n_past, false);
+
+    } else {
+        eval_string(ctx_llava->ctx_llama, std::string("</image>\n").c_str(), params->n_batch, &n_past, false);
+    }
+}
+
+static void process_user_input(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false) {
+    std::string user_prompt = prompt + "<AI>";
+    if (!is_first) user_prompt = "<用户>" + prompt;
+    const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
+
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
+
+    LOG_TEE("\n");
+
+    // generate the response
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    std::string response = "";
+    for (int i = 0; i < max_tgt_len; i++) {
+        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
+        if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
+        printf("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+        if (strstr(response.c_str(), "<用户>")) break; // minicpm-v 
+
+        fflush(stdout);
+    }
+
+    llama_sampling_free(ctx_sampling);
+    printf("\n");
+}
+
+static struct llava_context * llava_init(gpt_params * params) {
+    const char * clip_path = params->mmproj.c_str();
+
+    auto prompt = params->prompt;
+    if (prompt.empty()) {
+        prompt = "describe the image in detail.";
+    }
+
+    auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
+
+    llama_backend_init();
+    llama_numa_init(params->numa);
+
+    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+
+    llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
+    if (model == NULL) {
+        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        return NULL;
+    }
+
+    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
+
+    llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx_llama == NULL) {
+        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        return NULL;
+    }
+
+    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+
+    ctx_llava->ctx_llama = ctx_llama;
+    ctx_llava->ctx_clip = ctx_clip;
+    ctx_llava->model = model;
+    return ctx_llava;
+}
+
+static void llava_free(struct llava_context * ctx_llava) {
+    if (ctx_llava->ctx_clip) {
+        clip_free(ctx_llava->ctx_clip);
+        ctx_llava->ctx_clip = NULL;
+    }
+
+    llama_free(ctx_llava->ctx_llama);
+    llama_free_model(ctx_llava->model);
+    llama_backend_free();
+}
+
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("llava", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
+    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
+        gpt_print_usage(argc, argv, params);
+        show_additional_info(argc, argv);
+        return 1;
+    }
+
+    auto ctx_llava = llava_init(&params);
+    if (ctx_llava == NULL) {
+        LOG_TEE("%s: error: failed to init llava\n", __func__);
+        return 1;
+    }
+
+    // auto image_embed = load_image(ctx_llava, &params);
+    auto image_embed_and_slices = llava_image_embed_make_with_filename_slice(ctx_llava->ctx_clip, params.n_threads, params.image.c_str());
+    auto image_embed = image_embed_and_slices.first;
+    auto image_embed_slices = image_embed_and_slices.second;
+
+    if (!image_embed) {
+        LOG_TEE("%s: is %s really an image file?\n", __func__, params.image.c_str());
+        return 1;
+    }
+
+    // process the prompt
+    if (params.prompt.empty() && params.interactive == false) {
+        LOG_TEE("prompt should be given or interactive mode should be on");
+        return 1;
+    }
+
+    int n_past = 0;
+    process_image(ctx_llava, image_embed, image_embed_slices, &params, n_past);
+
+    if (!params.prompt.empty()) {
+        LOG_TEE("<用户>%s\n", params.prompt.c_str());
+        LOG_TEE("<AI>");
+        process_user_input(ctx_llava, &params, params.prompt, n_past, true);
+    } else {
+        while (true) {
+            LOG_TEE("<用户>");
+            std::string prompt;
+            std::getline(std::cin, prompt);
+            LOG_TEE("<AI>");
+            process_user_input(ctx_llava, &params, prompt, n_past, true);
+        }
+    }
+
+    llama_print_timings(ctx_llava->ctx_llama);
+
+    llava_image_embed_free(image_embed);
+    llava_free(ctx_llava);
+    return 0;
+}

examples/minicpmv/minicpmv.cpp

@@ -0,0 +1,556 @@
+#include "clip.h"
+#include "common.h"
+#include "llama.h"
+#include "minicpmv.h"
+#include "base64.hpp"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+#include <numeric>
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    int second;
+};
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+    struct {
+        struct ggml_tensor * newline;
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
+    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
+    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
+        if (newline_tmp->buffer == NULL) {
+            LOG_TEE("newline_tmp tensor buffer is NULL\n");
+        }
+        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
+    } else {
+        model.newline->data = newline_tmp->data;
+        if (model.newline->data == NULL) {
+            LOG_TEE("newline_tmp tensor data is NULL\n");
+        }
+    }
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
+
+
+static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+        LOG_TEE("%s: unable to preprocess image\n", __func__);
+        delete[] img_res_v.data;
+        return false;
+    }
+
+    const int64_t t_img_enc_start_us = ggml_time_us();
+
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        *n_img_pos = clip_n_patches(ctx_clip);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
+        if (!encoded) {
+            LOG_TEE("Unable to encode image\n");
+
+            return false;
+        }
+    } else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        LOG_TEE("%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);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
+
+        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
+
+        int n_img_pos_out;
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        *n_img_pos = n_img_pos_out;
+
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
+    }
+
+    LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
+    const int64_t t_img_enc_end_us = ggml_time_us();
+    float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
+
+    LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+
+    return true;
+}
+
+bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
+        // make sure that the correct mmproj was used, i.e., compare apples to apples
+    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+    auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
+    if (n_image_embd != n_llama_embd) {
+        LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        return false;
+    }
+    return true;
+}
+
+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) {
+    float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
+    if (!image_embd) {
+        LOG_TEE("Unable to allocate memory for image embeddings\n");
+        return false;
+    }
+
+    int n_img_pos;
+    if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
+        LOG_TEE("%s: cannot encode image, aborting\n", __func__);
+        free(image_embd);
+        return false;
+    }
+    *image_embd_out = image_embd;
+    *n_img_pos_out = n_img_pos;
+
+    return true;
+}
+
+bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
+    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+
+    for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
+        int n_eval = image_embed->n_image_pos - i;
+        if (n_eval > n_batch) {
+            n_eval = n_batch;
+        }
+        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
+        if (llama_decode(ctx_llama, batch)) {
+            LOG_TEE("%s : failed to eval\n", __func__);
+            return false;
+        }
+        *n_past += n_eval;
+    }
+    return true;
+}
+
+struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+    clip_image_u8 * img = clip_image_u8_init();
+    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
+        clip_image_u8_free(img);
+        LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
+        return NULL;
+    }
+
+    float* image_embed = NULL;
+    int n_image_pos = 0;
+    bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
+    if (!image_embed_result) {
+        clip_image_u8_free(img);
+        LOG_TEE("%s: coulnd't embed the image\n", __func__);
+        return NULL;
+    }
+
+    clip_image_u8_free(img);
+    auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+    result->embed = image_embed;
+    result->n_image_pos = n_image_pos;
+    return result;
+}
+
+std::vector<std::vector<clip_image_u8 *>> slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14, const bool never_split=false) {
+
+    const int original_width = img->nx;
+    const int original_height = img->ny;
+    const float log_ratio = log(1.0*original_width/original_height); //
+    const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
+    const int multiple = fmin(ceil(ratio), max_slice_nums);
+
+    std::vector<std::vector<clip_image_u8 *>> images;
+    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    if(multiple > 1){
+
+        std::vector<int> candidate_split_grids_nums;
+        for (int i : {multiple - 1, multiple, multiple + 1}) {
+            if (i == 1 || i > max_slice_nums) {
+                continue;
+            }
+            candidate_split_grids_nums.push_back(i);
+        }
+
+        std::vector<std::pair<int, int>> candidate_grids;
+
+        for (int split_grids_nums : candidate_split_grids_nums) {
+            int m = 1;
+            while (m <= split_grids_nums) {
+                if (split_grids_nums % m == 0) {
+                    candidate_grids.emplace_back(m, split_grids_nums / m);
+                }
+                ++m;
+            }
+        }
+
+        std::pair<int, int> best_grid{1, 1};
+        float min_error = std::numeric_limits<float>::infinity();
+
+        for (const auto& grid : candidate_grids) {
+            float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
+            if (error < min_error) {
+                best_grid = grid;
+                min_error = error;
+            }
+        }
+        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+        
+        // split_to_patches
+        int width = img->nx;
+        int height = img->ny;
+        int grid_x = int(width / best_grid.first);
+        int grid_y = int(height / best_grid.second);
+        for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
+            images.push_back(std::vector<clip_image_u8 *>());
+            for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
+                clip_image_u8 * patch = clip_image_u8_init();
+                patch->nx = grid_x;
+                patch->ny = grid_y;
+                patch->buf.resize(3 * patch->nx * patch->ny);
+                for (int y = patches_i; y < patches_i + grid_y; ++y) {
+                    for (int x = patches_j; x < patches_j + grid_x; ++x) {
+                        const int i = 3 * (y * img->nx + x);
+                        const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
+                        patch->buf[j]   = img->buf[i];
+                        patch->buf[j+1] = img->buf[i+1];
+                        patch->buf[j+2] = img->buf[i+2];
+                    }
+                }
+                images[images.size()-1].push_back(patch);
+            }
+        }
+
+
+    }
+
+    return images;
+}
+
+
+std::vector<std::vector<struct llava_image_embed *>> llava_image_embed_make_with_bytes_slice(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+    clip_image_u8 * img = clip_image_u8_init();
+    if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
+        clip_image_u8_free(img);
+        LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
+        return std::vector<std::vector<struct llava_image_embed *>>();
+    }
+    std::vector<std::vector<clip_image_u8 *>> imgs = slice_image(img);
+    for (size_t i = 0; i < imgs.size(); ++i){
+        for (size_t j = 0; j < imgs[i].size(); ++j) {
+            LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+        }
+    }
+    std::vector<std::vector<llava_image_embed *>> results;
+
+    for (size_t i = 0; i < imgs.size(); ++i){
+        results.push_back(std::vector<llava_image_embed *>());
+        for (size_t j = 0; j < imgs[i].size(); ++j) {
+            float* image_embed = NULL;
+            int n_image_pos = 0;
+            bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, imgs[i][j], &image_embed, &n_image_pos);
+            if (!image_embed_result) {
+                clip_image_u8_free(img);
+                LOG_TEE("%s: coulnd't embed the image\n", __func__);
+                return std::vector<std::vector<struct llava_image_embed *>>();
+            }
+
+            auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
+            result->embed = image_embed;
+            result->n_image_pos = n_image_pos;
+            results[i].push_back(result);
+        }
+    }
+    clip_image_u8_free(img);
+    return results;
+}
+
+
+static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
+    auto file = fopen(path, "rb");
+    if (file == NULL) {
+        LOG_TEE("%s: can't read file %s\n", __func__, path);
+        return false;
+    }
+
+    fseek(file, 0, SEEK_END);
+    auto fileSize = ftell(file);
+    fseek(file, 0, SEEK_SET);
+
+    auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
+    if (buffer == NULL) {
+        LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        perror("Memory allocation error");
+        fclose(file);
+        return false;
+    }
+    errno = 0;
+    size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
+    if (ferror(file)) {
+        die_fmt("read error: %s", strerror(errno));
+    }
+    if (ret != (size_t) fileSize) {
+        die("unexpectedly reached end of file");
+    }
+    fclose(file); // Close the file
+
+    *bytesOut = buffer;
+    *sizeOut = fileSize;
+    return true;
+}
+
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+    unsigned char* image_bytes;
+    long image_bytes_length;
+    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
+    if (!loaded) {
+        LOG_TEE("%s: failed to load %s\n", __func__, image_path);
+        return NULL;
+    }
+
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    free(image_bytes);
+
+    return embed;
+}
+
+std::pair<struct llava_image_embed *, std::vector<std::vector<struct llava_image_embed *>>> llava_image_embed_make_with_filename_slice(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+    unsigned char* image_bytes;
+    long image_bytes_length;
+    auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
+    if (!loaded) {
+        LOG_TEE("%s: failed to load %s\n", __func__, image_path);
+        return {NULL, std::vector<std::vector<struct llava_image_embed *>>()};
+    }
+
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    std::vector<std::vector<struct llava_image_embed *>> embeds = llava_image_embed_make_with_bytes_slice(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    free(image_bytes);
+    return {embed, embeds};
+}
+
+
+void llava_image_embed_free(struct llava_image_embed * embed) {
+    free(embed->embed);
+    free(embed);
+}

examples/minicpmv/minicpmv.h

@@ -0,0 +1,53 @@
+#ifndef LLAVA_H
+#define LLAVA_H
+
+#include "ggml.h"
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAVA_API __declspec(dllexport)
+#        else
+#            define LLAVA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAVA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAVA_API
+#endif
+
+struct clip_ctx;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct llava_image_embed {
+    float * embed;
+    int n_image_pos;
+};
+
+/** sanity check for clip <-> llava embed size match */
+LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
+
+LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+
+/** build an image embed from image file bytes */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
+/** build an image embed from a path to an image filename */
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
+/** free an embedding made with llava_image_embed_make_* */
+
+/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
+LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
+
+#ifdef __cplusplus
+}
+#endif
+
+LLAVA_API std::vector<std::vector<struct llava_image_embed *>> llava_image_embed_make_with_bytes_slice(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
+LLAVA_API std::pair<struct llava_image_embed *, std::vector<std::vector<struct llava_image_embed *>>> llava_image_embed_make_with_filename_slice(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
+
+#endif

examples/minicpmv/requirements.txt

@@ -0,0 +1,3 @@
+-r ../../requirements/requirements-convert.txt
+pillow~=10.2.0
+torch~=2.1.1

gguf-py/gguf/constants.py

@@ -645,6 +645,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     ],
     MODEL_ARCH.MINICPM: [
         MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_NORM,

llama.cpp

@@ -5120,18 +5120,16 @@ static bool llm_load_tensors(
             case LLM_ARCH_MINICPM:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                    // if output is NULL, init from the input tok embed
+                    if (model.output == NULL) {
+                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                        ml.n_created--; // artificial tensor
+                        ml.size_data += ggml_nbytes(model.output);
+                    }
 
                     // output
-                    {
+                    model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        if (model.arch != LLM_ARCH_MINICPM){
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            // if output is NULL, init from the input tok embed
-                            if (model.output == NULL) {
-                                model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
-                            }
-                        }
-                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         ggml_context * ctx_layer = ctx_for_layer(i);
@@ -10096,7 +10094,9 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
 
         // scale the input embeddings
-        inpL = ggml_scale(ctx0, inpL, scale_embd);
+        if (batch.token) {
+            inpL = ggml_scale(ctx0, inpL, scale_embd);
+        }
         cb(inpL, "inp_scaled", -1);
 
         // inp_pos - contains the positions
@@ -10212,7 +10212,7 @@ struct llm_build_context {
         cb(cur, "lmhead_scaling", -1);
 
         // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
+        cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);