update vendored llama.cpp and ggml (#11823)

* TEMPORARY: Update the llama.cpp upstream to my fork's Granite Four branch

This will be redone once my branch is merged upstream in llama.cpp

* feat: Update all patches

There are a number that are no longer needed at all:

- 0003-embeddings: Embeddings entirely overhauled on master
- 0008-ensure-KV-cache-is-fully-defragmented: KV caching entirely
    overhauled on master
- 0019-metal-add-mean-kernel-14267: Merged upstream
- 0020-CUDA-add-mean-operation-14313: Merged upstream

* feat: Sync llama.cpp and ggml

* fix: Update rsync-filter for all moved/new/removed files

* fix: Add files missing from sync

* fix: Update ggml rsync-filter for new ggml-cpu/arch subdirs

* fix: Add ggml files missing from sync

* fix: Narrow llama.cpp rsync-filter to not include mtmd main tool cpp files

* fix: Remove mtmd main cpp files

* fix: Add missing include in sampling_ext.cpp

* fix: Update llama.go to use mtmd instead of clip/llava

* fix: Add patch for mtmd_input_text

* chore: Ignore *.patched in the patch directory

* fix: Fix support for arch-specific ggml-cpu source files with new arrangement

In https://github.com/ggml-org/llama.cpp/pull/13892, all arch-specific
implementations were split out into a nested tree structure under
ggml-cpu/arch. This conflicts with standard CGO layout where all
arch-specific source files are expected to live in the same directory as
the parent go module and use suffixes based on GOOS and GOARCH. As such,
there were really two options for getting this to work:

1. Add a patch on top of the GGML sync to rearrange the files to match the
GO layout convention
2. Use CGO directives to conditionally include the nested source files in
the compilation units

This commit does (2) in order to minimize the set of changes needed on top
of the upstream file layout. To get this to work, there are two key things
needed:

1. In cpu.go, #cgo directives are added to explicitly set __${GOARCH}__ in
the preprocessor directives
2. In arch-impls.c|cpp, use an #ifdef | #elif defined | #endif chain to
explicitly include the .c|.cpp files for the given architecture from the
nested directory

* fix: Use mtmd_helper to correctly load the bitmap for the image

* fix: Apply patch for mtmd_text_input

* fix: Add missing stb to llama.cpp rsync-filter

* fix: Add sync'ed stb vendored header

* fix: Use c++17 and include vendor for go wrapper modules

* fix: Update patch 0015 for upstream implementation of uuid

* feat: Bump to the latest tip of the branch

* fix: Update patches for bump

* feat: Bump back to the cenral repo and point at the latest master

This includes granite 4 and a number of other model architectures!

* fix: Revert changes to ggml export GPU UUID patch

* fix: Add patch for GGML_VERSION and GGML_COMMIT constants

* feat: Sync all patched code

* build: Include cmake/common.cmake in ggml sync

* build: Add top-level include for GNUINstallDirs in CMakeLists.txt

This is used to populate CMAKE_INSTALL_BINDIR

* fix: Add a patch to avoid power throttling API on non-msvc windows builds

* fix: Sync patch changes for ggml-cpu.c

* feat: Bump llama.cpp to 4a4f42

This picks up support for Kimi K2 and PLaMO-2

* feat: Sync llama.cpp

* fix: Handle multi-chunk image encodings from mtmd

* fix: Re-number patches after merge with `main`

* feat: Bump to 41e78c in the makefile

* fix: Fix Solar and argsort/copy patches after bump

* fix: Remove Gemma3n CUDA Graphs patch

It was implemented upstream:
https://github.com/ggml-org/llama.cpp/pull/14741

* feat: Sync llama.cpp / ggml after latest bump

* build: Remove unnecessary CFLAGS definitions in cpu.go

* fix: Remove unnecessary additions in the rsync-filter

* fix: Remove unused vendored code for chat template parsing

* Revert "fix: Remove Gemma3n CUDA Graphs patch"

This reverts commit d724caced3ce21f08924d4b7801f94ce6638f6ea.

* fix: Update 0020 CUDA Graphs for gemma3n to keep both llama.cpp and ollama fixes

https://github.com/ollama/ollama/pull/11195#issuecomment-3137312394



* fix: Sync ggml-cuda.cu after keeping both style cuda graph fixes for gemma3n

* unwind mxfp4 patch

Prepare to bump ggml with their impl for mxfp4

* bump

* fix windows build error

* Convert tensors at load time

Repack the mxfp4 tensors as ggmls kernels expect them to be.

* convert mlp bf16 to f32

* buffer the conversion better

* reshape earlier

* openai swiglu

* add ids

* split qkv, gate_up

* fix nested alt tags

* fast attention

* remove debug messages

* fix lint

* remove redundant test

* remap values only if source/target are different

* add back i32->i32 copy

* refactor cpu quants

* clean up vendor

* update patch instructions

* clean up patches

* remove webgpu

* update mem

* also handle gpt-oss

* revert convert changes

---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>

llama/llama.cpp/tools/mtmd/mtmd-audio.h

+#pragma once
+
+#include "ggml.h"
+
+#include <cstdint>
+#include <vector>
+#include <string>
+
+#define WHISPER_ASSERT GGML_ASSERT
+
+#define WHISPER_SAMPLE_RATE 16000
+#define WHISPER_N_FFT       400
+#define WHISPER_HOP_LENGTH  160
+#define WHISPER_CHUNK_SIZE  30
+
+#define COMMON_SAMPLE_RATE 16000
+
+namespace whisper_preprocessor {
+
+struct whisper_mel {
+    int n_len;
+    int n_len_org;
+    int n_mel;
+
+    std::vector<float> data;
+};
+
+struct whisper_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+bool preprocess_audio(
+        const float * samples,
+        size_t n_samples,
+        const whisper_filters & filters,
+        std::vector<whisper_mel> & output);
+
+} // namespace whisper_preprocessor
+
+namespace whisper_precalc_filters {
+
+whisper_preprocessor::whisper_filters get_128_bins();
+
+} // namespace whisper_precalc_filters

llama/llama.cpp/tools/mtmd/mtmd-helper.cpp

+// fix problem with std::min and std::max
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include "mtmd.h"
+#include "mtmd-helper.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <cinttypes>
+#include <vector>
+
+//#define MTMD_AUDIO_DEBUG
+
+#define MINIAUDIO_IMPLEMENTATION
+#ifndef MTMD_AUDIO_DEBUG
+#   define MA_NO_ENCODING
+#endif
+#define MA_NO_DEVICE_IO
+#define MA_NO_RESOURCE_MANAGER
+#define MA_NO_NODE_GRAPH
+#define MA_NO_ENGINE
+#define MA_NO_GENERATION
+#define MA_API static
+#include "miniaudio/miniaudio.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb/stb_image.h"
+
+#define LOG_INF(...) fprintf(stdout, __VA_ARGS__)
+#define LOG_ERR(...) fprintf(stderr, __VA_ARGS__)
+
+size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
+    size_t n_tokens = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        n_tokens += mtmd_input_chunk_get_n_tokens(chunk);
+    }
+    return n_tokens;
+}
+
+llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
+    llama_pos n_pos = 0;
+    for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+        n_pos += mtmd_input_chunk_get_n_pos(chunk);
+    }
+    return n_pos;
+}
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
+    std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for image
+    void set_position_mrope_2d(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for audio
+    void set_position_mrope_1d(llama_pos pos_0, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            pos[i                     ] = pos_0 + i;
+            pos[i + batch.n_tokens    ] = pos_0 + i;
+            pos[i + batch.n_tokens * 2] = pos_0 + i;
+            pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.reserve(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                // assume n_tokens is less than or equal to batch.n_tokens
+                // batch.n_tokens is number of **total** tokens
+                // n_tokens is number of viewed token
+                size_t src_idx = i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(),
+                    pos.data() + src_idx,
+                    pos.data() + src_idx + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
+};
+
+// Helper function for decoding an image whose embeddings have already been calculated
+int32_t mtmd_helper_decode_image_chunk(
+        mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        float * encoded_embd,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        llama_pos * new_n_past) {
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+    const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_ERR("failed to decode chunk: input chunk not of image/audio type\n");
+        return -1;
+    }
+
+    const llama_model * model = llama_get_model(lctx);
+    int n_mmproj_embd = llama_model_n_embd(model);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
+
+    int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
+    int32_t i_batch = 0;
+    int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
+    decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+    if (mtmd_decode_use_mrope(ctx)) {
+        if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+            if (!image_tokens) {
+                LOG_ERR("failed to decode chunk: image tokens are null\n");
+                return -1;
+            }
+            const int nx = mtmd_image_tokens_get_nx(image_tokens);
+            const int ny = mtmd_image_tokens_get_ny(image_tokens);
+            batch_embd.set_position_mrope_2d(n_past, nx, ny, seq_id);
+        } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+            batch_embd.set_position_mrope_1d(n_past, seq_id);
+        } else {
+            GGML_ABORT("invalid chunk type for M-RoPE");
+        }
+    } else {
+        batch_embd.set_position_normal(n_past, seq_id);
+    }
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, false);
+        // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+    }
+
+    while (i_batch < n_img_batches) { // split into batches
+        int pos_offset = i_batch*n_batch;
+        int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+        llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
+
+        LOG_INF("decoding %s batch %d/%d, n_tokens_batch = %d\n", name, i_batch+1, n_img_batches, n_tokens_batch);
+
+        int64_t t1 = ggml_time_ms();
+        int32_t ret = llama_decode(lctx, batch_embd_view);
+        if (ret != 0) {
+            LOG_ERR("failed to decode %s\n", name);
+            llama_set_causal_attn(lctx, true); // restore causal attn
+            return ret;
+        }
+
+        LOG_INF("%s decoded (batch %d/%d) in %" PRId64 " ms\n", name, i_batch+1, n_img_batches, ggml_time_ms() - t1);
+
+        i_batch++;
+    }
+
+    n_past += mtmd_input_chunk_get_n_pos(chunk);
+    *new_n_past = n_past;
+
+    if (mtmd_decode_use_non_causal(ctx)) {
+        llama_set_causal_attn(lctx, true);
+    }
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
+        struct llama_context * lctx,
+        const mtmd_input_chunk * chunk,
+        llama_pos n_past,
+        llama_seq_id seq_id,
+        int32_t n_batch,
+        bool logits_last,
+        llama_pos * new_n_past) {
+    int32_t ret;
+    llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
+    auto chunk_type = mtmd_input_chunk_get_type(chunk);
+
+    if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        size_t n_tokens;
+        const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
+        // LOG_INF("decoding text chunk, n_tokens = %zu\n", n_tokens);
+        size_t i = 0;
+        while (i < n_tokens) { // split into batches
+            text_batch.n_tokens = 0; // clear the batch
+            for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
+                int32_t j = text_batch.n_tokens;
+                text_batch.token   [j]    = tokens[i];
+                text_batch.pos     [j]    = n_past++;
+                text_batch.n_seq_id[j]    = 1;
+                text_batch.seq_id  [j][0] = seq_id;
+                text_batch.logits  [j]    = false;
+
+                text_batch.n_tokens++;
+            }
+            bool is_last_token = (i == n_tokens);
+            if (logits_last && is_last_token) {
+                text_batch.logits[text_batch.n_tokens - 1] = true;
+            }
+            ret = llama_decode(lctx, text_batch);
+            if (ret != 0) {
+                LOG_ERR("failed to decode text\n");
+                llama_batch_free(text_batch);
+                return ret;
+            }
+            *new_n_past += text_batch.n_tokens;
+        }
+
+    } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE || chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
+        int64_t t0 = ggml_time_ms();
+
+        LOG_INF("encoding %s slice...\n", name);
+
+        ret = mtmd_encode_chunk(ctx, chunk);
+        if (ret != 0) {
+            LOG_ERR("failed to encode %s slice\n", name);
+            llama_batch_free(text_batch);
+            return ret;
+        }
+
+        LOG_INF("%s slice encoded in %" PRId64 " ms\n", name, ggml_time_ms() - t0);
+
+        float * embd = mtmd_get_output_embd(ctx);
+        ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past);
+        if (ret != 0) {
+            LOG_ERR("failed to decode %s\n", name);
+            llama_batch_free(text_batch);
+            return ret;
+        }
+    } else {
+        GGML_ABORT("chunk type not supported");
+    }
+
+    llama_batch_free(text_batch);
+    return 0;
+}
+
+int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
+                                struct llama_context * lctx,
+                                const mtmd_input_chunks * chunks,
+                                llama_pos n_past,
+                                llama_seq_id seq_id,
+                                int32_t n_batch,
+                                bool logits_last,
+                                llama_pos * new_n_past) {
+    size_t n_chunks = mtmd_input_chunks_size(chunks);
+    if (n_chunks == 0) {
+        LOG_ERR("no chunks to eval\n");
+        return 0;
+    }
+
+    for (size_t i = 0; i < n_chunks; i++) {
+        bool chunk_logits_last = (i == n_chunks - 1) && logits_last;
+        auto chunk = mtmd_input_chunks_get(chunks, i);
+
+        int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past);
+        if (res != 0) {
+            LOG_ERR("failed to eval chunk %zu\n", i);
+            return res;
+        }
+        *new_n_past = n_past;
+    }
+
+    return 0;
+}
+
+namespace audio_helpers {
+
+static bool is_audio_file(const char * buf, size_t len) {
+    if (len < 12) {
+        return false;
+    }
+
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    bool is_wav = memcmp(buf, "RIFF", 4) == 0 && memcmp(buf + 8, "WAVE", 4) == 0;
+    bool is_mp3 = len >= 3 && (
+        memcmp(buf, "ID3", 3) == 0 ||
+        // Check for MPEG sync word (simplified check)
+        ((unsigned char)buf[0] == 0xFF && ((unsigned char)buf[1] & 0xE0) == 0xE0)
+    );
+    bool is_flac = memcmp(buf, "fLaC", 4) == 0;
+
+    return is_wav || is_mp3 || is_flac;
+}
+
+// returns true if the buffer is a valid audio file
+static bool decode_audio_from_buf(const unsigned char * buf_in, size_t len, int target_sampler_rate, std::vector<float> & pcmf32_mono) {
+    ma_result result;
+    const int channels = 1;
+    ma_decoder_config decoder_config = ma_decoder_config_init(ma_format_f32, channels, target_sampler_rate);
+    ma_decoder decoder;
+
+    result = ma_decoder_init_memory(buf_in, len, &decoder_config, &decoder);
+    if (result != MA_SUCCESS) {
+        return false;
+    }
+
+    ma_uint64 frame_count;
+    ma_uint64 frames_read;
+    result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+    pcmf32_mono.resize(frame_count);
+    result = ma_decoder_read_pcm_frames(&decoder, pcmf32_mono.data(), frame_count, &frames_read);
+    if (result != MA_SUCCESS) {
+        ma_decoder_uninit(&decoder);
+        return false;
+    }
+
+#ifdef MTMD_AUDIO_DEBUG
+    // save audio to wav file
+    ma_encoder_config config = ma_encoder_config_init(ma_encoding_format_wav, ma_format_f32, 1, target_sampler_rate);
+    ma_encoder encoder;
+    ma_encoder_init_file("output.wav", &config, &encoder);
+    ma_encoder_write_pcm_frames(&encoder, pcmf32_mono.data(), pcmf32_mono.size(), &frames_read);
+    ma_encoder_uninit(&encoder);
+#endif
+
+    ma_decoder_uninit(&decoder);
+    return true;
+}
+
+} // namespace audio_helpers
+
+mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len) {
+    if (audio_helpers::is_audio_file((const char *)buf, len)) {
+        std::vector<float> pcmf32;
+        int bitrate = mtmd_get_audio_bitrate(ctx);
+        if (bitrate < 0) {
+            LOG_ERR("This model does not support audio input\n");
+            return nullptr;
+        }
+        if (!audio_helpers::decode_audio_from_buf(buf, len, bitrate, pcmf32)) {
+            LOG_ERR("Unable to read WAV audio file from buffer\n");
+            return nullptr;
+        }
+        return mtmd_bitmap_init_from_audio(pcmf32.size(), pcmf32.data());
+    }
+
+    // otherwise, we assume it's an image
+    mtmd_bitmap * result = nullptr;
+    {
+        int nx, ny, nc;
+        auto * data = stbi_load_from_memory(buf, len, &nx, &ny, &nc, 3);
+        if (!data) {
+            LOG_ERR("%s: failed to decode image bytes\n", __func__);
+            return nullptr;
+        }
+        result = mtmd_bitmap_init(nx, ny, data);
+        stbi_image_free(data);
+    }
+    return result;
+}
+
+mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname) {
+    std::vector<unsigned char> buf;
+    FILE * f = fopen(fname, "rb");
+    if (!f) {
+        LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
+        return nullptr;
+    }
+
+    fseek(f, 0, SEEK_END);
+    long file_size = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    buf.resize(file_size);
+
+    size_t n_read = fread(buf.data(), 1, file_size, f);
+    fclose(f);
+    if (n_read != (size_t)file_size) {
+        LOG_ERR("Failed to read entire file %s", fname);
+        return nullptr;
+    }
+
+    return mtmd_helper_bitmap_init_from_buf(ctx, buf.data(), buf.size());
+}

llama/llama.cpp/tools/mtmd/mtmd-helper.h

+#ifndef MTMD_HELPER_H
+#define MTMD_HELPER_H
+
+#include "ggml.h"
+#include "llama.h"
+#include "mtmd.h"
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// libmtmd helper functions
+//
+// Please note that these helpers are not guaranteed to be stable.
+// BREAKING CHANGES are expected.
+//
+
+// helper function to construct a mtmd_bitmap from a file
+// it calls mtmd_helper_bitmap_init_from_buf() internally
+// returns nullptr on failure
+// this function is thread-safe
+MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname);
+
+// helper function to construct a mtmd_bitmap from a buffer containing a file
+// supported formats:
+//     image: formats supported by stb_image: jpg, png, bmp, gif, etc.
+//     audio: formats supported by miniaudio: wav, mp3, flac
+// note: audio files will be auto-detected based on magic bytes
+// returns nullptr on failure
+// this function is thread-safe
+MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len);
+
+// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache
+MTMD_API size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks);
+
+// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
+// normally, n_pos is equal to n_tokens, but for M-RoPE it is different
+MTMD_API llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks);
+
+// helper function that automatically:
+// 1. run llama_decode() on text chunks
+// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
+// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
+// otherwise, returns 0 on success
+// this function is NOT thread-safe
+MTMD_API int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
+                                         struct llama_context * lctx,
+                                         const mtmd_input_chunks * chunks,
+                                         llama_pos n_past,
+                                         llama_seq_id seq_id,
+                                         int32_t n_batch,
+                                         bool logits_last,
+                                         llama_pos * new_n_past);
+
+// works like mtmd_helper_eval_chunks(), but only for a single chunk
+// this function is NOT thread-safe
+MTMD_API int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
+                                               struct llama_context * lctx,
+                                               const mtmd_input_chunk * chunk,
+                                               llama_pos n_past,
+                                               llama_seq_id seq_id,
+                                               int32_t n_batch,
+                                               bool logits_last,
+                                               llama_pos * new_n_past);
+
+// helper function to decode an image whose embeddings have already been calculated
+// this helper will handle batching and pre/post decoding setup (for ex. gemma 3 requires non-causal attention)
+// ret 0 on success, -1 on chunk not being a valid image chunk, 1 on decode failure
+MTMD_API int32_t mtmd_helper_decode_image_chunk(mtmd_context * ctx,
+                                                struct llama_context * lctx,
+                                                const mtmd_input_chunk * chunk,
+                                                float * encoded_embd,
+                                                llama_pos n_past,
+                                                llama_seq_id seq_id,
+                                                int32_t n_batch,
+                                                llama_pos * new_n_past);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+//
+// C++ wrappers
+//
+
+#endif

llama/llama.cpp/tools/mtmd/mtmd.cpp

+#include "clip.h"
+#include "clip-impl.h"
+#include "mtmd.h"
+#include "mtmd-audio.h"
+
+#include "llama.h"
+
+#include <algorithm>
+#include <cerrno>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
+#include <vector>
+
+// represents raw image data, layout is RGBRGBRGB...
+// length of data must be nx * ny * 3
+struct mtmd_bitmap {
+    uint32_t nx;
+    uint32_t ny;
+    std::vector<unsigned char> data;
+    std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking
+    bool is_audio = false; // true if the bitmap is audio
+};
+
+struct mtmd_image_tokens {
+    uint32_t nx; // number of tokens in x direction
+    uint32_t ny; // number of tokens in y direction
+    bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
+    uint32_t n_tokens() const { return nx * ny; }
+    clip_image_f32_batch batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
+
+    mtmd_image_tokens clone() {
+        return mtmd_image_tokens{
+            nx,
+            ny,
+            use_mrope_pos,
+            batch_f32.clone(),
+            id
+        };
+    }
+};
+using mtmd_image_tokens_ptr = std::unique_ptr<mtmd_image_tokens>;
+
+struct mtmd_audio_tokens {
+    uint32_t n_tokens; // number of tokens
+    clip_image_f32_batch batch_f32; // preprocessed image patches
+    std::string id; // optional user-defined ID, useful for KV cache tracking
+
+    mtmd_audio_tokens clone() {
+        return mtmd_audio_tokens{
+            n_tokens,
+            batch_f32.clone(),
+            id
+        };
+    }
+};
+using mtmd_audio_tokens_ptr = std::unique_ptr<mtmd_audio_tokens>;
+
+struct mtmd_input_chunk {
+    mtmd_input_chunk_type type;
+    std::vector<llama_token> tokens_text;
+    mtmd_image_tokens_ptr tokens_image;
+    mtmd_audio_tokens_ptr tokens_audio;
+};
+
+struct mtmd_input_chunks {
+    std::vector<mtmd_input_chunk> entries;
+};
+
+// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
+// models not having it (llava-1.6) will process embeddings without any special tokens in-between
+enum mtmd_slice_tmpl {
+    MTMD_SLICE_TMPL_NONE,
+    MTMD_SLICE_TMPL_MINICPMV_2_5,
+    MTMD_SLICE_TMPL_MINICPMV_2_6,
+    MTMD_SLICE_TMPL_LLAMA4,
+    // TODO @ngxson : add support for idefics (SmolVLM)
+};
+
+mtmd_input_text* mtmd_input_text_init(const char * text, bool add_special, bool parse_special) {
+    return new mtmd_input_text{text, add_special, parse_special};
+}
+
+void mtmd_input_text_free(mtmd_input_text* input_text) {
+    if (input_text) {
+        delete input_text;
+    }
+}
+
+const char * mtmd_default_marker() {
+    return "<__media__>";
+}
+
+mtmd_context_params mtmd_context_params_default() {
+    mtmd_context_params params;
+    params.use_gpu = true;
+    params.print_timings = true;
+    params.n_threads = 4;
+    params.verbosity = GGML_LOG_LEVEL_INFO;
+    params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
+    params.media_marker = mtmd_default_marker();
+    return params;
+}
+
+struct mtmd_context {
+    struct clip_ctx * ctx_v; // vision
+    struct clip_ctx * ctx_a; // audio
+    const struct llama_model * text_model;
+    std::vector<float> image_embd_v; // image embedding vector
+
+    bool print_timings;
+    int n_threads;
+    std::string media_marker;
+    const int n_embd_text;
+
+    // these are not token, but strings used to mark the beginning and end of image/audio embeddings
+    std::string img_beg;
+    std::string img_end;
+    std::string aud_beg;
+    std::string aud_end;
+
+    // for llava-uhd style models, we need special tokens in-between slices
+    // minicpmv calls them "slices", llama 4 calls them "tiles"
+    mtmd_slice_tmpl slice_tmpl    = MTMD_SLICE_TMPL_NONE;
+    llama_token tok_ov_img_start  = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_ov_img_end    = LLAMA_TOKEN_NULL; // overview image
+    llama_token tok_slices_start  = LLAMA_TOKEN_NULL; // start of all slices
+    llama_token tok_slices_end    = LLAMA_TOKEN_NULL; // end of all slices
+    llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice start
+    llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice end
+    llama_token tok_sli_img_mid   = LLAMA_TOKEN_NULL; // between 2 slices
+    llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
+    bool        tok_row_end_trail = false;
+    bool        ov_img_first      = false;
+
+    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
+
+    // for whisper, we pre-calculate the mel filter bank
+    whisper_preprocessor::whisper_filters w_filters;
+
+    // TODO @ngxson : add timings
+
+    mtmd_context(const char * mmproj_fname,
+                   const llama_model * text_model,
+                   const mtmd_context_params & ctx_params) :
+        text_model   (text_model),
+        print_timings(ctx_params.print_timings),
+        n_threads    (ctx_params.n_threads),
+        media_marker (ctx_params.media_marker),
+        n_embd_text  (llama_model_n_embd(text_model))
+    {
+        if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
+            throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
+        }
+
+        if (media_marker.empty()) {
+            throw std::runtime_error("media_marker must not be empty");
+        }
+
+        clip_context_params ctx_clip_params;
+        ctx_clip_params.use_gpu   = ctx_params.use_gpu;
+        ctx_clip_params.verbosity = ctx_params.verbosity;
+        auto res = clip_init(mmproj_fname, ctx_clip_params);
+        ctx_v = res.ctx_v;
+        ctx_a = res.ctx_a;
+        if (!ctx_v && !ctx_a) {
+            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
+        }
+
+        // if both vision and audio mmproj are present, we need to validate their n_embd
+        if (ctx_v && ctx_a) {
+            int n_embd_v = clip_n_mmproj_embd(ctx_v);
+            int n_embd_a = clip_n_mmproj_embd(ctx_a);
+            if (n_embd_v != n_embd_a) {
+                throw std::runtime_error(string_format(
+                    "mismatch between vision and audio mmproj (n_embd_v = %d, n_embd_a = %d)\n",
+                    n_embd_v, n_embd_a));
+            }
+        }
+
+        // since we already validate n_embd of vision and audio mmproj,
+        // we can safely assume that they are the same
+        int n_embd_clip = clip_n_mmproj_embd(ctx_v ? ctx_v : ctx_a);
+        if (n_embd_text != n_embd_clip) {
+            throw std::runtime_error(string_format(
+                "mismatch between text model (n_embd = %d) and mmproj (n_embd = %d)\n"
+                "hint: you may be using wrong mmproj\n",
+                n_embd_text, n_embd_clip));
+        }
+        if (ctx_v) {
+            init_vision();
+        }
+        if (ctx_a) {
+            init_audio();
+        }
+    }
+
+    void init_vision() {
+        GGML_ASSERT(ctx_v != nullptr);
+        use_mrope = clip_is_qwen2vl(ctx_v);
+
+        projector_type proj = clip_get_projector_type(ctx_v);
+        int minicpmv_version = clip_is_minicpmv(ctx_v);
+        if (minicpmv_version == 2) {
+            // minicpmv 2.5 format:
+            // <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_5;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_slices_start  = lookup_token("<slice>");
+            tok_slices_end    = lookup_token("</slice>");
+            tok_sli_img_start = tok_ov_img_start;
+            tok_sli_img_end   = tok_ov_img_end;
+            tok_row_end       = lookup_token("\n");
+            tok_row_end_trail = false; // no trailing end-of-row token
+            ov_img_first      = true;
+
+        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5) {
+            // minicpmv 2.6 format:
+            // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
+            slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;
+            tok_ov_img_start  = lookup_token("<image>");
+            tok_ov_img_end    = lookup_token("</image>");
+            tok_sli_img_start = lookup_token("<slice>");
+            tok_sli_img_end   = lookup_token("</slice>");
+            tok_row_end       = lookup_token("\n");
+            tok_row_end_trail = false; // no trailing end-of-row token
+            ov_img_first      = true;
+
+        } else if (minicpmv_version != 0) {
+            GGML_ASSERT(false && "unsupported minicpmv version");
+        } else if (proj == PROJECTOR_TYPE_LLAMA4) {
+            // llama 4 format:
+            // <|image_start|>
+            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
+            //     (slice) <|tile_x_separator|> (slice) <|tile_x_separator|> ... <|tile_y_separator|>
+            //     ... <|tile_y_separator|>   <-- trailing end-of-row token
+            // <|image|> (overview)           <-- overview image is last
+            // <|image_end|>
+            slice_tmpl        = MTMD_SLICE_TMPL_LLAMA4;
+            tok_ov_img_start  = lookup_token("<|image|>");
+            tok_sli_img_mid   = lookup_token("<|tile_x_separator|>");
+            tok_row_end       = lookup_token("<|tile_y_separator|>");
+            tok_row_end_trail = true; // add trailing end-of-row token
+            ov_img_first      = false; // overview image is last
+        }
+
+        // set boi/eoi
+        if (proj == PROJECTOR_TYPE_GEMMA3) {
+            // <start_of_image> ... (image embeddings) ... <end_of_image>
+            img_beg = "<start_of_image>";
+            img_end = "<end_of_image>";
+
+        } else if (proj == PROJECTOR_TYPE_IDEFICS3) {
+            // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
+            img_beg = "<fake_token_around_image><global-img>";
+            img_end = "<fake_token_around_image>";
+
+        } else if (proj == PROJECTOR_TYPE_PIXTRAL) {
+            // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
+            img_end = "[IMG_END]";
+
+        } else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL) {
+            // <|vision_start|> ... (image embeddings) ... <|vision_end|>
+            img_beg = "<|vision_start|>";
+            img_end = "<|vision_end|>";
+
+        } else if (proj == PROJECTOR_TYPE_LLAMA4) {
+            // (more details in mtmd_context constructor)
+            img_beg = "<|image_start|>";
+            img_end = "<|image_end|>";
+            LOG_WRN("%s: llama 4 vision is known to have degraded quality:\n"
+                    "    https://github.com/ggml-org/llama.cpp/pull/13282\n", __func__);
+
+        } else if (proj == PROJECTOR_TYPE_INTERNVL) {
+            // <img> ... (image embeddings) ... </img>
+            img_beg = "<img>";
+            img_end = "</img>";
+
+        }
+    }
+
+    void init_audio() {
+        GGML_ASSERT(ctx_a != nullptr);
+        projector_type proj = clip_get_projector_type(ctx_a);
+
+        if (clip_has_whisper_encoder(ctx_a)) {
+            // TODO @ngxson : check if model n_mel is 128 or 80
+            w_filters = whisper_precalc_filters::get_128_bins();
+        }
+
+        LOG_WRN("%s: audio input is in experimental stage and may have reduced quality:\n"
+                "    https://github.com/ggml-org/llama.cpp/discussions/13759\n", __func__);
+
+        if (proj == PROJECTOR_TYPE_QWEN2A) {
+            // <|audio_bos|> ... (embeddings) ... <|audio_eos|>
+            aud_beg = "<|audio_bos|>";
+            aud_end = "<|audio_eos|>";
+
+        } else if (proj == PROJECTOR_TYPE_ULTRAVOX) {
+            // [BEGIN_AUDIO] ... (embeddings) ...
+            aud_beg = "[BEGIN_AUDIO]";
+
+        }
+    }
+
+    // get clip ctx based on chunk type
+    clip_ctx * get_clip_ctx(const mtmd_input_chunk * chunk) const {
+        if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            return ctx_v;
+        } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+            return ctx_a;
+        }
+        GGML_ABORT("unknown chunk type");
+    }
+
+    projector_type proj_type_v() const {
+        return ctx_v ? clip_get_projector_type(ctx_v) : PROJECTOR_TYPE_UNKNOWN;
+    }
+
+    projector_type proj_type_a() const {
+        return ctx_a ? clip_get_projector_type(ctx_a) : PROJECTOR_TYPE_UNKNOWN;
+    }
+
+    ~mtmd_context() {
+        clip_free(ctx_a);
+        clip_free(ctx_v);
+    }
+
+private:
+    llama_token lookup_token(const std::string & token_text) {
+        const llama_vocab * vocab = llama_model_get_vocab(text_model);
+        const int n_vocab = llama_vocab_n_tokens(vocab);
+        for (int i = 0; i < n_vocab; i++) {
+            if (token_to_piece(vocab, i, true) == token_text) {
+                return i;
+            }
+        }
+        return LLAMA_TOKEN_NULL;
+    }
+
+    std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
+        std::string piece;
+        piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
+        const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+        if (n_chars < 0) {
+            piece.resize(-n_chars);
+            int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
+            GGML_ASSERT(check == -n_chars);
+        } else {
+            piece.resize(n_chars);
+        }
+        return piece;
+    }
+};
+
+mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+        const struct llama_model * text_model,
+        const struct mtmd_context_params ctx_params) {
+    try {
+        return new mtmd_context(mmproj_fname, text_model, ctx_params);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return nullptr;
+    }
+}
+
+void mtmd_free(mtmd_context * ctx) {
+    if (ctx) {
+        delete ctx;
+    }
+}
+
+struct mtmd_tokenizer {
+    mtmd_context * ctx;
+    std::vector<const mtmd_bitmap *> bitmaps;
+
+    std::string input_text;
+    bool add_special;
+    bool parse_special;
+    const llama_vocab * vocab;
+
+    mtmd_input_chunks cur;
+
+    mtmd_tokenizer(mtmd_context * ctx,
+            const mtmd_input_text * text,
+            const mtmd_bitmap ** bitmaps,
+            size_t n_bitmaps) : ctx(ctx), bitmaps(bitmaps, bitmaps + n_bitmaps) {
+        add_special   = text->add_special;
+        parse_special = text->parse_special;
+        input_text    = text->text;
+        vocab         = llama_model_get_vocab(ctx->text_model);
+
+        // for compatibility, we convert image marker to media marker
+        string_replace_all(input_text, MTMD_DEFAULT_IMAGE_MARKER, ctx->media_marker);
+    }
+
+    int32_t tokenize(mtmd_input_chunks * output) {
+        cur.entries.clear();
+        std::vector<std::string> parts = split_text(input_text, ctx->media_marker);
+        size_t i_bm = 0; // index of the current bitmap
+        for (auto & part : parts) {
+            if (part == ctx->media_marker) {
+                // this is a marker, we should add the next bitmap
+                if (i_bm >= bitmaps.size()) {
+                    LOG_ERR("%s: error: number of bitmaps (%zu) does not match number of markers (%zu)\n",
+                            __func__, bitmaps.size(), parts.size() - 1);
+                    return 1;
+                }
+                const mtmd_bitmap * bitmap = bitmaps[i_bm++];
+                int32_t res = add_media(bitmap);
+                if (res != 0) {
+                    return res;
+                }
+            } else {
+                // this is a text part, we should add it as text
+                add_text(part, parse_special);
+            }
+        }
+
+        if (add_special && llama_vocab_get_add_bos(vocab)) {
+            // if first chunk is text, we add BOS token to first text chunk
+            // otherwise, create a new text chunk with BOS token
+            if (!cur.entries.empty() && cur.entries[0].type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+                // add BOS token to the beginning of first text chunk
+                cur.entries[0].tokens_text.insert(cur.entries[0].tokens_text.begin(), llama_vocab_bos(vocab));
+            } else {
+                // create a new text chunk with BOS token at the beginning
+                mtmd_input_chunk bos_chunk{
+                    MTMD_INPUT_CHUNK_TYPE_TEXT,
+                    {llama_vocab_bos(vocab)},
+                    nullptr, // image tokens
+                    nullptr, // audio tokens
+                };
+                cur.entries.insert(cur.entries.begin(), std::move(bos_chunk));
+            }
+        }
+
+        if (add_special && llama_vocab_get_add_eos(vocab)) {
+            // if last chunk is text, we add EOS token to it
+            add_text({llama_vocab_eos(vocab)});
+        }
+
+        if (i_bm != bitmaps.size()) {
+            LOG_ERR("%s: error: number of bitmaps (%zu) does not match number of markers (%zu)\n",
+                    __func__, bitmaps.size(), parts.size() - 1);
+            return 1;
+        }
+
+        *output = std::move(cur);
+
+        return 0;
+    }
+
+    void add_text(const std::string & txt, bool parse_special) {
+        LOG_DBG("%s: %s\n", __func__, txt.c_str());
+        auto tokens = mtmd_tokenize_text_internal(vocab, txt, /* add_special */ false, parse_special);
+        add_text(tokens);
+    }
+
+    void add_text(const std::vector<llama_token> & tokens) {
+        if (tokens.empty()) {
+            return;
+        }
+        // if last entry is also a text chunk, add tokens to it instead of creating new chunk
+        if (!cur.entries.empty() && cur.entries.back().type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            cur.entries.back().tokens_text.insert(
+                                            cur.entries.back().tokens_text.end(),
+                                            tokens.begin(),
+                                            tokens.end());
+        } else {
+            mtmd_input_chunk chunk{
+                MTMD_INPUT_CHUNK_TYPE_TEXT,
+                tokens,
+                nullptr, // image tokens
+                nullptr, // audio tokens
+            };
+            cur.entries.emplace_back(std::move(chunk));
+        }
+    }
+
+    int32_t add_media(const mtmd_bitmap * bitmap) {
+        if (!bitmap->is_audio) {
+            // handle image
+
+            if (!ctx->ctx_v) {
+                LOG_ERR("%s: error: model does not support vision input\n", __func__);
+                return 2;
+            }
+
+            if (!ctx->img_beg.empty()) {
+                add_text(ctx->img_beg, true); // add image begin token
+            }
+
+            // convert mtmd_bitmap to clip_image_u8
+            clip_image_u8_ptr img_u8(clip_image_u8_init());
+            img_u8->nx = bitmap->nx;
+            img_u8->ny = bitmap->ny;
+            img_u8->buf.resize(bitmap->data.size());
+            std::memcpy(img_u8->buf.data(), bitmap->data.data(), img_u8->nx * img_u8->ny * 3);
+
+            // preprocess image
+            clip_image_f32_batch batch_f32;
+            bool ok = clip_image_preprocess(ctx->ctx_v, img_u8.get(), &batch_f32);
+            if (!ok) {
+                LOG_ERR("Unable to preprocess image\n");
+                return 2;
+            }
+
+            // handle llava-uhd style preprocessing
+            if (
+                ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5
+                || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6
+                || ctx->slice_tmpl == MTMD_SLICE_TMPL_LLAMA4
+            ) {
+                const int n_col = batch_f32.grid_x;
+                const int n_row = batch_f32.grid_y;
+                // split batch into chunks of single images
+                // NOTE: batch_f32 will be invalidated after this call
+                auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmap->id);
+                GGML_ASSERT(chunks.size() > 0);
+
+                auto ov_chunk = std::move(chunks.front());
+                chunks.erase(chunks.begin());
+
+                // add overview image (first)
+                if (ctx->ov_img_first) {
+                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_ov_img_start});
+                    }
+                    cur.entries.emplace_back(std::move(ov_chunk));
+                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_ov_img_end});
+                    }
+                }
+
+                // add slices (or tiles)
+                if (!chunks.empty()) {
+                    GGML_ASSERT((int)chunks.size() == n_row * n_col);
+                    if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_slices_start});
+                    }
+                    for (int y = 0; y < n_row; y++) {
+                        for (int x = 0; x < n_col; x++) {
+                            const bool is_last_in_row = (x == n_col - 1);
+                            if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
+                                add_text({ctx->tok_sli_img_start});
+                            }
+                            cur.entries.emplace_back(std::move(chunks[y * n_col + x]));
+                            if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
+                                add_text({ctx->tok_sli_img_end});
+                            }
+                            if (!is_last_in_row && ctx->tok_sli_img_mid != LLAMA_TOKEN_NULL) {
+                                add_text({ctx->tok_sli_img_mid});
+                            }
+                        }
+                        if ((y != n_row - 1 || ctx->tok_row_end_trail) && ctx->tok_row_end != LLAMA_TOKEN_NULL) {
+                            add_text({ctx->tok_row_end});
+                        }
+                    }
+                    if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_slices_end});
+                    }
+                }
+
+                // add overview image (last)
+                if (!ctx->ov_img_first) {
+                    if (ctx->tok_ov_img_start != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_ov_img_start});
+                    }
+                    cur.entries.emplace_back(std::move(ov_chunk));
+                    if (ctx->tok_ov_img_end != LLAMA_TOKEN_NULL) {
+                        add_text({ctx->tok_ov_img_end});
+                    }
+                }
+
+            } else {
+                size_t n_tokens = 0;
+                for (const auto & entry : batch_f32.entries) {
+                    n_tokens += clip_n_output_tokens(ctx->ctx_v, entry.get());
+                }
+
+                mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+                if (ctx->use_mrope) {
+                    // for Qwen2VL, we need this information for M-RoPE decoding positions
+                    image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
+                    image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, batch_f32.entries[0].get());
+                    image_tokens->use_mrope_pos = true;
+                } else {
+                    // other models, we only need the total number of tokens
+                    image_tokens->nx = n_tokens;
+                    image_tokens->ny = 1;
+                }
+                image_tokens->batch_f32 = std::move(batch_f32);
+                image_tokens->id = bitmap->id; // optional
+
+                LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
+                LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
+                LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
+
+                mtmd_input_chunk chunk{
+                    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                    {}, // text tokens
+                    std::move(image_tokens),
+                    nullptr, // audio tokens
+                };
+                cur.entries.emplace_back(std::move(chunk));
+            }
+
+            if (!ctx->img_end.empty()) {
+                add_text(ctx->img_end, true); // add image end token
+            }
+
+        } else {
+            // handle audio
+
+            if (!ctx->ctx_a) {
+                LOG_ERR("%s: error: model does not support audio input\n", __func__);
+                return 2;
+            }
+
+            if (bitmap->data.size() == 0) {
+                LOG_ERR("%s: error: empty audio data\n", __func__);
+                return 2;
+            }
+
+            if (!ctx->aud_beg.empty()) {
+                add_text(ctx->aud_beg, true); // add audio begin token
+            }
+
+            // preprocess audio
+            GGML_ASSERT(ctx->w_filters.n_mel); // make sure we have filter preloaded
+            std::vector<whisper_preprocessor::whisper_mel> mel_spec_chunks;
+            const float * samples = (const float *)bitmap->data.data();
+            size_t n_samples = bitmap->data.size() / sizeof(float);
+            bool ok = whisper_preprocessor::preprocess_audio(samples, n_samples, ctx->w_filters, mel_spec_chunks);
+            if (!ok) {
+                LOG_ERR("Unable to preprocess audio\n");
+                return 2;
+            }
+
+            // consider each mel_spec as a separate audio chunk
+            // TODO: maybe support batching, but this may come with memory cost
+            for (auto & mel_spec : mel_spec_chunks) {
+                clip_image_f32_ptr mel_f32(clip_image_f32_init());
+                mel_f32->nx  = mel_spec.n_len;
+                mel_f32->ny  = mel_spec.n_mel;
+                mel_f32->buf = std::move(mel_spec.data);
+                size_t n_tokens = clip_n_output_tokens(ctx->ctx_a, mel_f32.get());
+
+                clip_image_f32_batch batch_f32;
+                batch_f32.is_audio = true;
+                batch_f32.entries.push_back(std::move(mel_f32));
+
+                mtmd_audio_tokens_ptr audio_tokens(new mtmd_audio_tokens);
+                audio_tokens->n_tokens = n_tokens;
+                audio_tokens->batch_f32 = std::move(batch_f32);
+                audio_tokens->id = bitmap->id; // optional
+
+                LOG_DBG("audio_tokens->n_tokens = %d\n", audio_tokens->n_tokens);
+
+                mtmd_input_chunk chunk{
+                    MTMD_INPUT_CHUNK_TYPE_AUDIO,
+                    {}, // text tokens
+                    nullptr, // image tokens
+                    std::move(audio_tokens),
+                };
+                cur.entries.emplace_back(std::move(chunk));
+            }
+
+            if (!ctx->aud_end.empty()) {
+                add_text(ctx->aud_end, true); // add audio end token
+            }
+        }
+
+        return 0;
+    }
+
+    std::vector<mtmd_input_chunk> split_batch_to_chunk(clip_image_f32_batch && batch_f32, const std::string & id) {
+        std::vector<mtmd_input_chunk> chunks;
+
+        for (auto & entry : batch_f32.entries) {
+            mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+            image_tokens->nx = clip_n_output_tokens(ctx->ctx_v, entry.get());
+            image_tokens->ny = 1;
+            image_tokens->batch_f32.entries.push_back(std::move(entry));
+            image_tokens->id = id;
+
+            mtmd_input_chunk chunk{
+                MTMD_INPUT_CHUNK_TYPE_IMAGE,
+                {}, // text tokens
+                std::move(image_tokens),
+                nullptr, // audio tokens
+            };
+            chunks.emplace_back(std::move(chunk));
+        }
+
+        return chunks;
+    }
+
+    // for example: "a <__media__> b <__media__> c" --> "a", "<__media__>", "b", "<__media__>", "c"
+    static std::vector<std::string> split_text(const std::string & input, const std::string & delimiter) {
+        std::vector<std::string> result;
+        if (input.empty()) {
+            return result;
+        }
+        size_t start = 0;
+        size_t pos = 0;
+        while ((pos = input.find(delimiter, start)) != std::string::npos) {
+            if (pos > start) {
+                result.push_back(input.substr(start, pos - start));
+            }
+            result.push_back(delimiter);
+            start = pos + delimiter.length();
+        }
+        if (start < input.length()) {
+            result.push_back(input.substr(start));
+        }
+        return result;
+    }
+
+    // copied from common_tokenize
+    static std::vector<llama_token> mtmd_tokenize_text_internal(
+        const struct llama_vocab * vocab,
+               const std::string & text,
+                            bool   add_special,
+                            bool   parse_special) {
+        // upper limit for the number of tokens
+        int n_tokens = text.length() + 2 * add_special;
+        std::vector<llama_token> result(n_tokens);
+        n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        if (n_tokens < 0) {
+            result.resize(-n_tokens);
+            int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+            GGML_ASSERT(check == -n_tokens);
+        } else {
+            result.resize(n_tokens);
+        }
+        return result;
+    }
+};
+
+int32_t mtmd_tokenize(mtmd_context * ctx,
+            mtmd_input_chunks * output,
+            const mtmd_input_text * text,
+            const mtmd_bitmap ** bitmaps,
+            size_t n_bitmaps) {
+    mtmd_tokenizer tokenizer(ctx, text, bitmaps, n_bitmaps);
+    return tokenizer.tokenize(output);
+}
+
+int32_t mtmd_encode_chunk(mtmd_context * ctx, const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        LOG_WRN("mtmd_encode_chunk has no effect for text chunks\n");
+        return 0;
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        if (!ctx->ctx_v) {
+            LOG_ERR("%s: model does not support vision input\n", __func__);
+            return 1;
+        }
+        return mtmd_encode(ctx, chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        if (!ctx->ctx_a) {
+            LOG_ERR("%s: model does not support audio input\n", __func__);
+            return 1;
+        }
+        int n_mmproj_embd = ctx->n_embd_text;
+        ctx->image_embd_v.resize(chunk->tokens_audio->n_tokens * n_mmproj_embd);
+        bool ok = clip_image_batch_encode(
+            ctx->ctx_a,
+            ctx->n_threads,
+            &chunk->tokens_audio->batch_f32,
+            ctx->image_embd_v.data());
+        return ok ? 0 : 1;
+    }
+
+    LOG_ERR("%s: unknown chunk type %d\n", __func__, (int)chunk->type);
+    return 1;
+}
+
+int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
+    clip_ctx * ctx_clip = ctx->ctx_v;
+    if (!ctx_clip) {
+        LOG_ERR("%s: this API does not support non-vision input, please use mtmd_encode_chunk instead\n", __func__);
+        return 1;
+    }
+    int n_mmproj_embd = clip_n_mmproj_embd(ctx_clip);
+    ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
+    bool ok = false;
+
+    if (clip_is_llava(ctx_clip) || clip_is_minicpmv(ctx_clip) || clip_is_glm(ctx_clip)) {
+        // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
+        const auto & entries = image_tokens->batch_f32.entries;
+        for (size_t i = 0; i < entries.size(); i++) {
+            int n_tokens_per_image = clip_n_output_tokens(ctx_clip, entries[i].get());
+            ok = clip_image_encode(
+                ctx_clip,
+                ctx->n_threads,
+                entries[i].get(),
+                ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
+        }
+    } else {
+        ok = clip_image_batch_encode(
+            ctx_clip,
+            ctx->n_threads,
+            &image_tokens->batch_f32,
+            ctx->image_embd_v.data());
+    }
+
+    return ok ? 0 : 1;
+}
+
+float * mtmd_get_output_embd(mtmd_context * ctx) {
+    return ctx->image_embd_v.data();
+}
+
+bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
+    if (ctx->ctx_v && clip_get_projector_type(ctx->ctx_v) == PROJECTOR_TYPE_GEMMA3) {
+        return true;
+    }
+    return false;
+}
+
+bool mtmd_decode_use_mrope(mtmd_context * ctx) {
+    return ctx->use_mrope;
+}
+
+bool mtmd_support_vision(mtmd_context * ctx) {
+    return ctx->ctx_v != nullptr;
+}
+
+bool mtmd_support_audio(mtmd_context * ctx) {
+    return ctx->ctx_a != nullptr;
+}
+
+int mtmd_get_audio_bitrate(mtmd_context * ctx) {
+    if (!ctx->ctx_a) {
+        return -1;
+    }
+    // for now, we assume that all audio models have the same bitrate
+    return 16000; // 16kHz
+}
+
+//
+// public API functions
+//
+
+// mtmd_bitmap
+
+mtmd_bitmap * mtmd_bitmap_init(uint32_t nx,
+                               uint32_t ny,
+                               const unsigned char * data) {
+    mtmd_bitmap * bitmap = new mtmd_bitmap;
+    bitmap->nx = nx;
+    bitmap->ny = ny;
+    size_t data_size = (size_t)nx * ny * 3;
+    bitmap->data.resize(data_size);
+    std::memcpy(bitmap->data.data(), data, data_size);
+    return bitmap;
+}
+
+mtmd_bitmap * mtmd_bitmap_init_from_audio(size_t n_samples,
+                                          const float * data) {
+    mtmd_bitmap * bitmap = new mtmd_bitmap;
+    bitmap->nx = n_samples;
+    bitmap->ny = 1;
+    bitmap->is_audio = true;
+    size_t data_size = n_samples * sizeof(float);
+    bitmap->data.resize(data_size);
+    std::memcpy(bitmap->data.data(), data, data_size);
+    return bitmap;
+}
+
+uint32_t mtmd_bitmap_get_nx(const mtmd_bitmap * bitmap) {
+    return bitmap->nx;
+}
+
+uint32_t mtmd_bitmap_get_ny(const mtmd_bitmap * bitmap) {
+    return bitmap->ny;
+}
+
+const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap) {
+    return bitmap->data.data();
+}
+
+size_t mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap) {
+    return bitmap->data.size();
+}
+
+bool mtmd_bitmap_is_audio(const mtmd_bitmap * bitmap) {
+    return bitmap->is_audio;
+}
+
+const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap) {
+    return bitmap->id.c_str();
+}
+
+void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id) {
+    if (id) {
+        bitmap->id = std::string(id);
+    } else {
+        bitmap->id.clear();
+    }
+}
+
+void mtmd_bitmap_free(mtmd_bitmap * bitmap) {
+    if (bitmap) {
+        delete bitmap;
+    }
+}
+
+// mtmd_input_chunks
+
+mtmd_input_chunks * mtmd_input_chunks_init() {
+    return new mtmd_input_chunks;
+}
+
+size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks) {
+    return chunks->entries.size();
+}
+
+const mtmd_input_chunk * mtmd_input_chunks_get(const mtmd_input_chunks * chunks, size_t idx) {
+    if (idx >= chunks->entries.size()) {
+        return nullptr;
+    }
+    return &chunks->entries[idx];
+}
+
+void mtmd_input_chunks_free(mtmd_input_chunks * chunks) {
+    if (chunks) {
+        delete chunks;
+    }
+}
+
+// mtmd_input_chunk
+
+enum mtmd_input_chunk_type mtmd_input_chunk_get_type(const mtmd_input_chunk * chunk) {
+    return chunk->type;
+}
+
+const llama_token * mtmd_input_chunk_get_tokens_text(const mtmd_input_chunk * chunk, size_t * n_tokens_output) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        *n_tokens_output = chunk->tokens_text.size();
+        return chunk->tokens_text.data();
+    }
+    *n_tokens_output = 0;
+    return nullptr;
+}
+
+const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return chunk->tokens_image.get();
+    }
+    return nullptr;
+}
+
+size_t mtmd_input_chunk_get_n_tokens(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        return chunk->tokens_text.size();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return mtmd_image_tokens_get_n_tokens(chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->n_tokens;
+    } else {
+        GGML_ABORT("invalid chunk type");
+    }
+}
+
+llama_pos mtmd_input_chunk_get_n_pos(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+        return chunk->tokens_text.size();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return mtmd_image_tokens_get_n_pos(chunk->tokens_image.get());
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->n_tokens;
+    } else {
+        GGML_ABORT("invalid chunk type");
+    }
+}
+
+const char * mtmd_input_chunk_get_id(const mtmd_input_chunk * chunk) {
+    if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+        return chunk->tokens_image->id.c_str();
+    } else if (chunk->type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+        return chunk->tokens_audio->id.c_str();
+    }
+    return nullptr;
+}
+
+mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk) {
+    mtmd_input_chunk * copy = new mtmd_input_chunk{
+        chunk->type,
+        chunk->tokens_text,
+        nullptr,
+        nullptr,
+    };
+    if (chunk->tokens_image) {
+        // copy the image tokens
+        copy->tokens_image = mtmd_image_tokens_ptr(new mtmd_image_tokens());
+        *copy->tokens_image = chunk->tokens_image->clone();
+    }
+    if (chunk->tokens_audio) {
+        // copy the audio tokens
+        copy->tokens_audio = mtmd_audio_tokens_ptr(new mtmd_audio_tokens());
+        *copy->tokens_audio = chunk->tokens_audio->clone();
+    }
+    return copy;
+}
+
+void mtmd_input_chunk_free(mtmd_input_chunk * chunk) {
+    if (chunk) {
+        delete chunk;
+    }
+}
+
+// mtmd_image_tokens
+
+size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->n_tokens();
+}
+
+size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->nx;
+}
+
+size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->ny;
+}
+
+const char * mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
+    return image_tokens->id.c_str();
+}
+
+llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
+    if (image_tokens->use_mrope_pos) {
+        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
+    }
+    return image_tokens->n_tokens();
+}
+
+// test function
+
+mtmd_input_chunks * mtmd_test_create_input_chunks() {
+    mtmd_input_chunks * chunks = mtmd_input_chunks_init();
+    if (!chunks) {
+        return nullptr;
+    }
+
+    // create a text chunk
+    std::vector<llama_token> tokens_text = { 1, 2, 3, 4, 5 };
+    mtmd_input_chunk chunk_text{
+        MTMD_INPUT_CHUNK_TYPE_TEXT,
+        std::move(tokens_text),
+        nullptr, // image tokens
+        nullptr, // audio tokens
+    };
+    chunks->entries.emplace_back(std::move(chunk_text));
+
+    // create an image chunk
+    mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
+    image_tokens->nx = 4;
+    image_tokens->ny = 4;
+    image_tokens->batch_f32.entries.resize(16);
+    image_tokens->id = "image_1";
+    mtmd_input_chunk chunk_image{
+        MTMD_INPUT_CHUNK_TYPE_IMAGE,
+        {}, // text tokens
+        std::move(image_tokens),
+        nullptr, // audio tokens
+    };
+    chunks->entries.emplace_back(std::move(chunk_image));
+
+    return chunks;
+}

llama/llama.cpp/tools/mtmd/mtmd.go

 package mtmd
 
-// #cgo CXXFLAGS: -std=c++11
-// #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../common
+// #cgo CXXFLAGS: -std=c++17
+// #cgo CPPFLAGS: -I${SRCDIR}/../../include -I${SRCDIR}/../../common -I${SRCDIR}/../../vendor
 // #cgo CPPFLAGS: -I${SRCDIR}/../../../../ml/backend/ggml/ggml/include
 import "C"

llama/llama.cpp/tools/mtmd/mtmd.h

+#ifndef MTMD_H
+#define MTMD_H
+
+#include "ggml.h"
+#include "llama.h"
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+#include <string>
+#include <vector>
+#include <cinttypes>
+#include <memory>
+#endif
+
+/**
+ * libmtmd: A library for multimodal support in llama.cpp.
+ *
+ * WARNING: This API is experimental and subject to many BREAKING CHANGES.
+ *          Issues related to API usage may receive lower priority support.
+ *
+ * For the usage, see an example in mtmd-cli.cpp
+ */
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define MTMD_API __declspec(dllexport)
+#        else
+#            define MTMD_API __declspec(dllimport)
+#        endif
+#    else
+#        define MTMD_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define MTMD_API
+#endif
+
+// deprecated marker, use mtmd_default_marker() instead
+#define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum mtmd_input_chunk_type {
+    MTMD_INPUT_CHUNK_TYPE_TEXT,
+    MTMD_INPUT_CHUNK_TYPE_IMAGE,
+    MTMD_INPUT_CHUNK_TYPE_AUDIO,
+};
+
+// opaque types
+struct mtmd_context;
+struct mtmd_bitmap;
+struct mtmd_image_tokens;
+struct mtmd_input_chunk;
+struct mtmd_input_chunks;
+
+struct mtmd_input_text {
+    const char * text;
+    bool add_special;
+    bool parse_special;
+};
+
+//
+// C API
+//
+
+typedef struct mtmd_context      mtmd_context;
+typedef struct mtmd_bitmap       mtmd_bitmap;
+typedef struct mtmd_image_tokens mtmd_image_tokens;
+typedef struct mtmd_input_chunk  mtmd_input_chunk;
+typedef struct mtmd_input_chunks mtmd_input_chunks;
+typedef struct mtmd_input_text   mtmd_input_text;
+
+MTMD_API mtmd_input_text* mtmd_input_text_init(const char * text, bool add_special, bool parse_special);
+MTMD_API void mtmd_input_text_free(mtmd_input_text* input_text);
+
+struct mtmd_context_params {
+    bool use_gpu;
+    bool print_timings;
+    int n_threads;
+    enum ggml_log_level verbosity;
+    const char * image_marker; // deprecated, use media_marker instead
+    const char * media_marker;
+};
+
+MTMD_API const char * mtmd_default_marker(void);
+
+MTMD_API struct mtmd_context_params mtmd_context_params_default(void);
+
+// initialize the mtmd context
+// return nullptr on failure
+MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
+                                            const struct llama_model * text_model,
+                                            const struct mtmd_context_params ctx_params);
+
+MTMD_API void mtmd_free(mtmd_context * ctx);
+
+// whether we need to set non-causal mask before llama_decode
+MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
+
+// whether the current model use M-RoPE for llama_decode
+MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
+
+// whether the current model supports vision input
+MTMD_API bool mtmd_support_vision(mtmd_context * ctx);
+
+// whether the current model supports audio input
+MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
+
+// get audio bitrate in Hz, for example 16000 for Whisper
+// return -1 if audio is not supported
+MTMD_API int mtmd_get_audio_bitrate(mtmd_context * ctx);
+
+// mtmd_bitmap
+//
+// if bitmap is image:
+//     length of data must be nx * ny * 3
+//     the data is in RGBRGBRGB... format
+// if bitmap is audio:
+//     length of data must be n_samples * sizeof(float)
+//     the data is in float format (PCM F32)
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init           (uint32_t nx, uint32_t ny, const unsigned char * data);
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init_from_audio(size_t n_samples,         const float         * data);
+MTMD_API uint32_t              mtmd_bitmap_get_nx     (const mtmd_bitmap * bitmap);
+MTMD_API uint32_t              mtmd_bitmap_get_ny     (const mtmd_bitmap * bitmap);
+MTMD_API const unsigned char * mtmd_bitmap_get_data   (const mtmd_bitmap * bitmap);
+MTMD_API size_t                mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap);
+MTMD_API bool                  mtmd_bitmap_is_audio   (const mtmd_bitmap * bitmap);
+MTMD_API void                  mtmd_bitmap_free       (mtmd_bitmap * bitmap);
+// bitmap ID is optional, but useful for KV cache tracking
+// these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
+MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap);
+MTMD_API void         mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id);
+
+
+// mtmd_input_chunks
+//
+// this is simply a list of mtmd_input_chunk
+// the elements can only be populated via mtmd_tokenize()
+MTMD_API mtmd_input_chunks *      mtmd_input_chunks_init(void);
+MTMD_API size_t                   mtmd_input_chunks_size(const mtmd_input_chunks * chunks);
+MTMD_API const mtmd_input_chunk * mtmd_input_chunks_get (const mtmd_input_chunks * chunks, size_t idx);
+MTMD_API void                     mtmd_input_chunks_free(mtmd_input_chunks * chunks);
+
+// mtmd_input_chunk
+//
+// the instance will be constructed via mtmd_tokenize()
+// it will be freed along with mtmd_input_chunks
+MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type        (const mtmd_input_chunk * chunk);
+MTMD_API const llama_token *        mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output);
+MTMD_API const mtmd_image_tokens *  mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
+MTMD_API size_t                     mtmd_input_chunk_get_n_tokens    (const mtmd_input_chunk * chunk);
+// returns nullptr for ID on text chunk
+MTMD_API const char *               mtmd_input_chunk_get_id          (const mtmd_input_chunk * chunk);
+// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd_input_chunk * chunk);
+
+// in case you want to use custom logic to handle the chunk (i.e. KV cache management)
+// you can move the chunk ownership to your own code by copying it
+// remember to free the chunk when you are done with it
+MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk);
+MTMD_API void               mtmd_input_chunk_free(mtmd_input_chunk * chunk);
+
+
+// mtmd_image_tokens
+//
+// the instance will be constructed via mtmd_tokenize()
+// it will be freed along with mtmd_input_chunk
+MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
+MTMD_API size_t       mtmd_image_tokens_get_nx      (const mtmd_image_tokens * image_tokens);
+MTMD_API size_t       mtmd_image_tokens_get_ny      (const mtmd_image_tokens * image_tokens);
+MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens); // TODO: deprecate
+// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens); // TODO: deprecate
+
+// tokenize an input text prompt and a list of bitmaps (images/audio)
+// the prompt must have the input image marker (default: "<__media__>") in it
+// the default marker is defined by mtmd_default_marker()
+// the marker will be replaced with the image/audio chunk
+// for example:
+//   "here is an image: <__media__>\ndescribe it in detail."
+//   this will gives 3 chunks:
+//   1. "here is an image: <start_of_image>"
+//   2. (image/audio tokens)
+//   3. "<end_of_image>\ndescribe it in detail."
+// number of bitmaps must be equal to the number of markers in the prompt
+// this function is thread-safe (shared ctx)
+// return values:
+//   0 on success
+//   1 on number of bitmaps not matching the number of markers
+//   2 on image preprocessing error
+MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
+                               mtmd_input_chunks * output,
+                               const mtmd_input_text * text,
+                               const mtmd_bitmap ** bitmaps,
+                               size_t n_bitmaps);
+
+// returns 0 on success
+// TODO: deprecate
+MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
+                             const mtmd_image_tokens * image_tokens);
+
+// returns 0 on success
+MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
+                                   const mtmd_input_chunk * chunk);
+
+// get output embeddings from the last encode pass
+// the reading size (in bytes) is equal to:
+// llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
+MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
+
+/////////////////////////////////////////
+
+// test function, to be used in test-mtmd-c-api.c
+MTMD_API mtmd_input_chunks * mtmd_test_create_input_chunks(void);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif
+
+//
+// C++ wrappers
+//
+
+#ifdef __cplusplus
+
+namespace mtmd {
+
+struct mtmd_context_deleter {
+    void operator()(mtmd_context * val) { mtmd_free(val); }
+};
+using context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
+
+struct mtmd_bitmap_deleter {
+    void operator()(mtmd_bitmap * val) { mtmd_bitmap_free(val); }
+};
+using bitmap_ptr = std::unique_ptr<mtmd_bitmap, mtmd_bitmap_deleter>;
+
+struct mtmd_input_chunks_deleter {
+    void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
+};
+using input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
+
+struct mtmd_input_chunk_deleter {
+    void operator()(mtmd_input_chunk * val) { mtmd_input_chunk_free(val); }
+};
+using input_chunk_ptr = std::unique_ptr<mtmd_input_chunk, mtmd_input_chunk_deleter>;
+
+struct bitmap {
+    bitmap_ptr ptr;
+    bitmap() : ptr(nullptr) {}
+    bitmap(mtmd_bitmap * bitmap) : ptr(bitmap) {}
+    bitmap(bitmap && other) noexcept : ptr(std::move(other.ptr)) {}
+    bitmap(uint32_t nx, uint32_t ny, const unsigned char * data) {
+        ptr.reset(mtmd_bitmap_init(nx, ny, data));
+    }
+    ~bitmap() = default;
+    uint32_t nx() { return mtmd_bitmap_get_nx(ptr.get()); }
+    uint32_t ny() { return mtmd_bitmap_get_ny(ptr.get()); }
+    const unsigned char * data() { return mtmd_bitmap_get_data(ptr.get()); }
+    size_t n_bytes() { return mtmd_bitmap_get_n_bytes(ptr.get()); }
+    std::string id() { return mtmd_bitmap_get_id(ptr.get()); }
+    void set_id(const char * id) { mtmd_bitmap_set_id(ptr.get(), id); }
+};
+
+struct bitmaps {
+    std::vector<bitmap> entries;
+    ~bitmaps() = default;
+    // return list of pointers to mtmd_bitmap
+    // example:
+    //   auto bitmaps_c_ptr = bitmaps.c_ptr();
+    //   int32_t res = mtmd_tokenize(... bitmaps_c_ptr.data(), bitmaps_c_ptr.size());
+    std::vector<const mtmd_bitmap *> c_ptr() {
+        std::vector<const mtmd_bitmap *> res(entries.size());
+        for (size_t i = 0; i < entries.size(); i++) {
+            res[i] = entries[i].ptr.get();
+        }
+        return res;
+    }
+};
+
+struct input_chunks {
+    input_chunks_ptr ptr;
+    input_chunks() = default;
+    input_chunks(mtmd_input_chunks * chunks) : ptr(chunks) {}
+    ~input_chunks() = default;
+    size_t size() { return mtmd_input_chunks_size(ptr.get()); }
+    const mtmd_input_chunk * operator[](size_t idx) {
+        return mtmd_input_chunks_get(ptr.get(), idx);
+    }
+};
+
+} // namespace mtmd
+
+#endif
+
+#endif

llama/llama.cpp/vendor/nlohmann/json_fwd.hpp

+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_
+#define INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+#include <cstdint> // int64_t, uint64_t
+#include <map> // map
+#include <memory> // allocator
+#include <string> // string
+#include <vector> // vector
+
+// #include <nlohmann/detail/abi_macros.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013 - 2025 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// This file contains all macro definitions affecting or depending on the ABI
+
+#ifndef JSON_SKIP_LIBRARY_VERSION_CHECK
+    #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH)
+        #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 12 || NLOHMANN_JSON_VERSION_PATCH != 0
+            #warning "Already included a different version of the library!"
+        #endif
+    #endif
+#endif
+
+#define NLOHMANN_JSON_VERSION_MAJOR 3   // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_MINOR 12  // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_PATCH 0   // NOLINT(modernize-macro-to-enum)
+
+#ifndef JSON_DIAGNOSTICS
+    #define JSON_DIAGNOSTICS 0
+#endif
+
+#ifndef JSON_DIAGNOSTIC_POSITIONS
+    #define JSON_DIAGNOSTIC_POSITIONS 0
+#endif
+
+#ifndef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON 0
+#endif
+
+#if JSON_DIAGNOSTICS
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS _diag
+#else
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS
+#endif
+
+#if JSON_DIAGNOSTIC_POSITIONS
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS _dp
+#else
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS
+#endif
+
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON _ldvcmp
+#else
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION
+    #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0
+#endif
+
+// Construct the namespace ABI tags component
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c) json_abi ## a ## b ## c
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b, c) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b, c)
+
+#define NLOHMANN_JSON_ABI_TAGS                                       \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
+            NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS,                       \
+            NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON, \
+            NLOHMANN_JSON_ABI_TAG_DIAGNOSTIC_POSITIONS)
+
+// Construct the namespace version component
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \
+    _v ## major ## _ ## minor ## _ ## patch
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(major, minor, patch) \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch)
+
+#if NLOHMANN_JSON_NAMESPACE_NO_VERSION
+#define NLOHMANN_JSON_NAMESPACE_VERSION
+#else
+#define NLOHMANN_JSON_NAMESPACE_VERSION                                 \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(NLOHMANN_JSON_VERSION_MAJOR, \
+                                           NLOHMANN_JSON_VERSION_MINOR, \
+                                           NLOHMANN_JSON_VERSION_PATCH)
+#endif
+
+// Combine namespace components
+#define NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) a ## b
+#define NLOHMANN_JSON_NAMESPACE_CONCAT(a, b) \
+    NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b)
+
+#ifndef NLOHMANN_JSON_NAMESPACE
+#define NLOHMANN_JSON_NAMESPACE               \
+    nlohmann::NLOHMANN_JSON_NAMESPACE_CONCAT( \
+            NLOHMANN_JSON_ABI_TAGS,           \
+            NLOHMANN_JSON_NAMESPACE_VERSION)
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_BEGIN
+#define NLOHMANN_JSON_NAMESPACE_BEGIN                \
+    namespace nlohmann                               \
+    {                                                \
+    inline namespace NLOHMANN_JSON_NAMESPACE_CONCAT( \
+                NLOHMANN_JSON_ABI_TAGS,              \
+                NLOHMANN_JSON_NAMESPACE_VERSION)     \
+    {
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_END
+#define NLOHMANN_JSON_NAMESPACE_END                                     \
+    }  /* namespace (inline namespace) NOLINT(readability/namespace) */ \
+    }  // namespace nlohmann
+#endif
+
+
+/*!
+@brief namespace for Niels Lohmann
+@see https://github.com/nlohmann
+@since version 1.0.0
+*/
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/*!
+@brief default JSONSerializer template argument
+
+This serializer ignores the template arguments and uses ADL
+([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl))
+for serialization.
+*/
+template<typename T = void, typename SFINAE = void>
+struct adl_serializer;
+
+/// a class to store JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/
+template<template<typename U, typename V, typename... Args> class ObjectType =
+         std::map,
+         template<typename U, typename... Args> class ArrayType = std::vector,
+         class StringType = std::string, class BooleanType = bool,
+         class NumberIntegerType = std::int64_t,
+         class NumberUnsignedType = std::uint64_t,
+         class NumberFloatType = double,
+         template<typename U> class AllocatorType = std::allocator,
+         template<typename T, typename SFINAE = void> class JSONSerializer =
+         adl_serializer,
+         class BinaryType = std::vector<std::uint8_t>, // cppcheck-suppress syntaxError
+         class CustomBaseClass = void>
+class basic_json;
+
+/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+/// @sa https://json.nlohmann.me/api/json_pointer/
+template<typename RefStringType>
+class json_pointer;
+
+/*!
+@brief default specialization
+@sa https://json.nlohmann.me/api/json/
+*/
+using json = basic_json<>;
+
+/// @brief a minimal map-like container that preserves insertion order
+/// @sa https://json.nlohmann.me/api/ordered_map/
+template<class Key, class T, class IgnoredLess, class Allocator>
+struct ordered_map;
+
+/// @brief specialization that maintains the insertion order of object keys
+/// @sa https://json.nlohmann.me/api/ordered_json/
+using ordered_json = basic_json<nlohmann::ordered_map>;
+
+NLOHMANN_JSON_NAMESPACE_END
+
+#endif  // INCLUDE_NLOHMANN_JSON_FWD_HPP_

llama/llama.go

@@ -6,6 +6,7 @@ package llama
 #cgo CXXFLAGS: -std=c++17
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/include
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/common
+#cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/vendor
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/tools/mtmd
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/src
 #cgo CPPFLAGS: -I${SRCDIR}/../ml/backend/ggml/ggml/include
@@ -13,8 +14,8 @@ package llama
 #include <stdlib.h>
 #include "ggml.h"
 #include "llama.h"
-#include "clip.h"
-#include "llava.h"
+#include "mtmd.h"
+#include "mtmd-helper.h"
 #include "gguf.h"
 
 #include "sampling_ext.h"
@@ -148,27 +149,23 @@ func (c *Context) Model() *Model {
 }
 
 func (c *Context) KvCacheSeqAdd(seqId int, p0 int, p1 int, delta int) {
-	C.llama_kv_self_seq_add(c.c, C.int(seqId), C.int(p0), C.int(p1), C.int(delta))
+	C.llama_memory_seq_add(C.llama_get_memory(c.c), C.int(seqId), C.int(p0), C.int(p1), C.int(delta))
 }
 
 func (c *Context) KvCacheSeqRm(seqId int, p0 int, p1 int) bool {
-	return bool(C.llama_kv_self_seq_rm(c.c, C.int(seqId), C.int(p0), C.int(p1)))
+	return bool(C.llama_memory_seq_rm(C.llama_get_memory(c.c), C.int(seqId), C.int(p0), C.int(p1)))
 }
 
 func (c *Context) KvCacheSeqCp(srcSeqId int, dstSeqId int, p0 int, p1 int) {
-	C.llama_kv_self_seq_cp(c.c, C.int(srcSeqId), C.int(dstSeqId), C.int(p0), C.int(p1))
+	C.llama_memory_seq_cp(C.llama_get_memory(c.c), C.int(srcSeqId), C.int(dstSeqId), C.int(p0), C.int(p1))
 }
 
 func (c *Context) KvCacheClear() {
-	C.llama_kv_self_clear(c.c)
-}
-
-func (c *Context) KvCacheDefrag() {
-	C.llama_kv_self_defrag(c.c)
+	C.llama_memory_clear(C.llama_get_memory(c.c), true)
 }
 
 func (c *Context) KvCacheCanShift() bool {
-	return bool(C.llama_kv_self_can_shift(c.c))
+	return bool(C.llama_memory_can_shift(C.llama_get_memory(c.c)))
 }
 
 // Get the embeddings for a sequence id
@@ -460,52 +457,75 @@ func (m *Model) NEmbd() int {
 }
 
 // vision processing
-type ClipContext struct {
-	c *C.struct_clip_ctx
+type MtmdContext struct {
+	c *C.struct_mtmd_context
 }
 
-func NewClipContext(llamaContext *Context, modelPath string) (*ClipContext, error) {
+func NewMtmdContext(llamaContext *Context, modelPath string) (*MtmdContext, error) {
 	mp := C.CString(modelPath)
 	defer C.free(unsafe.Pointer(mp))
-	c := C.clip_model_load(mp, 1)
-	if c == nil {
-		return nil, fmt.Errorf("unable to load clip model: %v", modelPath)
-	}
+	// TODO: Support non-default params
+	cp := C.mtmd_context_params_default()
 
-	projEmbedSize := int(C.clip_n_mmproj_embd(c))
-	modelEmbedSize := llamaContext.Model().NEmbd()
-	if projEmbedSize != modelEmbedSize {
-		return nil, fmt.Errorf("projector embedding size (%d) does not match model (%d)", projEmbedSize, modelEmbedSize)
+	// NOTE: The model and projector embedding lengths are checked during init
+	c := C.mtmd_init_from_file(mp, C.llama_get_model(llamaContext.c), cp)
+	if c == nil {
+		return nil, fmt.Errorf("unable to load mmtd model: %v", modelPath)
 	}
 
-	return &ClipContext{c: c}, nil
+	return &MtmdContext{c: c}, nil
 }
 
-func (c *ClipContext) Free() {
-	C.clip_free(c.c)
+func (c *MtmdContext) Free() {
+	C.mtmd_free(c.c)
 }
 
-func (c *ClipContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32, error) {
-	l := C.llava_image_embed_make_with_bytes(c.c, C.int(llamaContext.numThreads), (*C.uchar)(unsafe.Pointer(&data[0])), C.int(len(data)))
-	if l == nil {
-		return nil, errors.New("unable to make llava embedding from image")
-	}
+func (c *MtmdContext) NewEmbed(llamaContext *Context, data []byte) ([][]float32, error) {
+	// Initialize the input chunks pointer
+	ic := C.mtmd_input_chunks_init()
+	defer C.mtmd_input_chunks_free(ic)
+
+	// Initialize an empty text prompt so we can tokenize
+	it := C.mtmd_input_text_init(C.mtmd_default_marker(), true, true)
+	defer C.mtmd_input_text_free(it)
 
-	numTokens := int(l.n_image_pos)
+	// Initialize a bitmap with the image data
+	bm := C.mtmd_helper_bitmap_init_from_buf(c.c, (*C.uchar)(unsafe.Pointer(&data[0])), C.size_t(len(data)))
+	defer C.mtmd_bitmap_free(bm)
+
+	// Tokenize the image
+	if C.int32_t(0) != C.mtmd_tokenize(c.c, ic, it, &bm, 1) {
+		return nil, errors.New("unable to tokenize mtmd embedding from image")
+	}
+	nChunks := C.mtmd_input_chunks_size(ic)
 	numEmbed := llamaContext.Model().NEmbd()
+	lastChunkSize := 0
+	for i := range int(nChunks) {
+		chunk := C.mtmd_input_chunks_get(ic, C.size_t(i))
+		numTokens := int(C.mtmd_input_chunk_get_n_tokens(chunk))
+		lastChunkSize = numTokens
+
+		// Encode the chunk
+		if C.int32_t(0) != C.mtmd_encode_chunk(c.c, chunk) {
+			return nil, errors.New("unable to encode mtmd image chunk")
+		}
+	}
 
-	s := unsafe.Slice((*float32)(l.embed), numEmbed*numTokens)
+	// Get the embeddings
+	embed := make([][]float32, lastChunkSize)
+	embd := C.mtmd_get_output_embd(c.c)
+	if nil == embd {
+		return nil, errors.New("failed to get image embedding")
+	}
 
-	embed := make([][]float32, numTokens)
+	// Extend the embedding array for each token
+	s := unsafe.Slice((*float32)(embd), numEmbed*lastChunkSize)
 	rows := make([]float32, len(s))
 	copy(rows, s)
-
-	for i := range embed {
+	for i := range lastChunkSize {
 		embed[i] = rows[i*numEmbed : (i+1)*numEmbed]
 	}
 
-	C.llava_image_embed_free(l)
-
 	return embed, nil
 }
 

llama/patches/.gitignore

+*.patched

llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch

@@ -12,19 +12,18 @@ MSVC and freed by Clang, which can cause problems.
 This moves freeing of the buffers into the backends to avoid the
 problem.
 ---
- ggml/src/ggml-backend.cpp              | 9 +++++++--
- ggml/src/ggml-cann/ggml-cann.cpp       | 2 ++
- ggml/src/ggml-cuda/ggml-cuda.cu        | 3 +++
- ggml/src/ggml-kompute/ggml-kompute.cpp | 1 +
- ggml/src/ggml-metal/ggml-metal.m       | 1 +
- ggml/src/ggml-opencl/ggml-opencl.cpp   | 1 +
- ggml/src/ggml-rpc/ggml-rpc.cpp         | 1 +
- ggml/src/ggml-sycl/ggml-sycl.cpp       | 3 +++
- ggml/src/ggml-vulkan/ggml-vulkan.cpp   | 2 ++
- 9 files changed, 21 insertions(+), 2 deletions(-)
+ ggml/src/ggml-backend.cpp            | 9 +++++++--
+ ggml/src/ggml-cann/ggml-cann.cpp     | 2 ++
+ ggml/src/ggml-cuda/ggml-cuda.cu      | 3 +++
+ ggml/src/ggml-metal/ggml-metal.m     | 1 +
+ ggml/src/ggml-opencl/ggml-opencl.cpp | 1 +
+ ggml/src/ggml-rpc/ggml-rpc.cpp       | 1 +
+ ggml/src/ggml-sycl/ggml-sycl.cpp     | 3 +++
+ ggml/src/ggml-vulkan/ggml-vulkan.cpp | 2 ++
+ 8 files changed, 20 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index b30b4cb3..0ce73a99 100644
+index 1b9d29e9..97f47abd 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -107,7 +107,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -35,7 +34,7 @@ index b30b4cb3..0ce73a99 100644
  }
  
  size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-@@ -544,6 +543,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -529,6 +528,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
  
      free(ctx->buffers);
      free(ctx);
@@ -43,7 +42,7 @@ index b30b4cb3..0ce73a99 100644
  }
  
  static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -1871,6 +1871,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -1890,6 +1890,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
  
  static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_aligned_free(buffer->context, buffer->size);
@@ -55,7 +54,7 @@ index b30b4cb3..0ce73a99 100644
  }
  
  static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-@@ -1918,7 +1923,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -1937,7 +1942,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
  };
  
  static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -65,10 +64,10 @@ index b30b4cb3..0ce73a99 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index e2617b06..242e50a7 100644
+index cf575b36..ca1addfa 100755
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
-@@ -800,6 +800,7 @@ static void ggml_backend_cann_buffer_free_buffer(
+@@ -826,6 +826,7 @@ static void ggml_backend_cann_buffer_free_buffer(
      ggml_backend_cann_buffer_context* ctx =
          (ggml_backend_cann_buffer_context*)buffer->context;
      delete ctx;
@@ -76,7 +75,7 @@ index e2617b06..242e50a7 100644
  }
  
  /**
-@@ -1472,6 +1473,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
+@@ -1572,6 +1573,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
   */
  static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
      ACL_CHECK(aclrtFreeHost(buffer->context));
@@ -85,10 +84,10 @@ index e2617b06..242e50a7 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index b4b85abc..cb0d8528 100644
+index d9110491..37ee2a6d 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -534,6 +534,7 @@ struct ggml_backend_cuda_buffer_context {
+@@ -567,6 +567,7 @@ struct ggml_backend_cuda_buffer_context {
  static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
      delete ctx;
@@ -96,7 +95,7 @@ index b4b85abc..cb0d8528 100644
  }
  
  static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -790,6 +791,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -822,6 +823,7 @@ struct ggml_backend_cuda_split_buffer_context {
  static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
      delete ctx;
@@ -104,7 +103,7 @@ index b4b85abc..cb0d8528 100644
  }
  
  static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1067,6 +1069,7 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
+@@ -1103,6 +1105,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
  
  static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      CUDA_CHECK(cudaFreeHost(buffer->context));
@@ -112,23 +111,11 @@ index b4b85abc..cb0d8528 100644
  }
  
  static void * ggml_cuda_host_malloc(size_t size) {
-diff --git a/ggml/src/ggml-kompute/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp
-index 50579227..2799a0a5 100644
---- a/ggml/src/ggml-kompute/ggml-kompute.cpp
-+++ b/ggml/src/ggml-kompute/ggml-kompute.cpp
-@@ -1911,6 +1911,7 @@ static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer
-         ggml_vk_free_memory(*memory);
-     }
-     delete memory;
-+    delete buffer;
- }
- 
- static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 576f9581..1b56f858 100644
+index cb8eff4a..7bccc7bf 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -5214,6 +5214,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -6032,6 +6032,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
      }
  
      free(ctx);
@@ -137,10 +124,10 @@ index 576f9581..1b56f858 100644
  
  static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index 05a2f4e6..392cc18d 100644
+index 8ba1e00d..8163e8dc 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -1940,6 +1940,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -2745,6 +2745,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -149,22 +136,22 @@ index 05a2f4e6..392cc18d 100644
  
  static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index 4f0abb5a..de1ec184 100644
+index df6ba540..2e395968 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -483,6 +483,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -486,6 +486,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-     GGML_ASSERT(status);
+     RPC_STATUS_ASSERT(status);
      delete ctx;
 +    delete buffer;
  }
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 0ea72994..ae3a3c33 100644
+index 3992dad0..67503951 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -320,6 +320,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -331,6 +331,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -172,7 +159,7 @@ index 0ea72994..ae3a3c33 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -765,6 +766,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -792,6 +793,7 @@ struct ggml_backend_sycl_split_buffer_context {
  static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
      delete ctx;
@@ -180,7 +167,7 @@ index 0ea72994..ae3a3c33 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1099,6 +1101,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1134,6 +1136,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_sycl_host_free(buffer->context);
@@ -189,10 +176,10 @@ index 0ea72994..ae3a3c33 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index e2b357fd..68768029 100644
+index 4070e248..394a2839 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -8962,6 +8962,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -10209,6 +10209,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -200,7 +187,7 @@ index e2b357fd..68768029 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -9105,6 +9106,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -10352,6 +10353,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);

llama/patches/0002-pretokenizer.patch

@@ -10,10 +10,10 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 9389ca80..806c1b3d 100644
+index f7e03e70..8ebe11cf 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -1503,16 +1503,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1804,16 +1804,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
          if (type == LLAMA_VOCAB_TYPE_BPE) {
              add_space_prefix = false;
              clean_spaces = true;
@@ -31,8 +31,8 @@ index 9389ca80..806c1b3d 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -1651,7 +1642,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
-                 pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+@@ -1975,7 +1966,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+                 pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
                  clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));

llama/patches/0004-clip-unicode.patch → llama/patches/0003-clip-unicode.patch

@@ -10,10 +10,10 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 41ba45a7..cdd8ca44 100644
+index 20c21733..f4f69cfc 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
-@@ -31,6 +31,19 @@
+@@ -28,6 +28,19 @@
  #include <numeric>
  #include <functional>
  
@@ -33,7 +33,7 @@ index 41ba45a7..cdd8ca44 100644
  struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
  
  enum ffn_op_type {
-@@ -2190,7 +2203,29 @@ struct clip_model_loader {
+@@ -2597,7 +2610,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 41ba45a7..cdd8ca44 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -2217,7 +2252,11 @@ struct clip_model_loader {
+@@ -2624,7 +2659,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0003-embeddings.patch

-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: jmorganca <jmorganca@gmail.com>
-Date: Tue, 8 Apr 2025 15:28:34 -0700
-Subject: [PATCH] embeddings
-
-allow a loaded model in llama.cpp to be used for
-both embeddings and causal attention text generation
-instead of forcing one or the error
----
- src/llama-context.cpp | 6 +++---
- 1 file changed, 3 insertions(+), 3 deletions(-)
-
-diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 62246c10..dca22d8b 100644
---- a/src/llama-context.cpp
-+++ b/src/llama-context.cpp
-@@ -901,7 +901,7 @@ int llama_context::decode(llama_batch & inp_batch) {
-     int64_t n_outputs_all = 0;
- 
-     // count outputs
--    if (batch.logits && !embd_pooled) {
-+    if (batch.logits) {
-         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-             n_outputs_all += batch.logits[i] != 0;
-         }
-@@ -982,7 +982,7 @@ int llama_context::decode(llama_batch & inp_batch) {
-         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
-         //}
- 
--        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
-+        auto * t_logits = cparams.causal_attn ? res->get_logits() : nullptr;
-         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
- 
-         if (t_embd && res->get_embd_pooled()) {
-@@ -1151,7 +1151,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
-     const auto n_embd  = hparams.n_embd;
- 
-     // TODO: use a per-batch flag for logits presence instead
--    bool has_logits = !cparams.embeddings;
-+    bool has_logits =  cparams.causal_attn;
-     bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
- 
-     // TODO: hacky enc-dec support

llama/patches/0005-solar-pro.patch → llama/patches/0004-solar-pro.patch

@@ -15,18 +15,18 @@ adds support for the Solar Pro architecture
  7 files changed, 248 insertions(+)
 
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index f2bc8ca7..5ab3f572 100644
+index 18dcc6dd..4b285646 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -69,6 +69,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
-     { LLM_ARCH_GRANITE,          "granite"          },
+@@ -78,6 +78,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
+     { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
 +    { LLM_ARCH_SOLAR,            "solar"            },
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -142,6 +143,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -164,6 +165,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
      { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
@@ -34,7 +34,7 @@ index f2bc8ca7..5ab3f572 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -1502,6 +1504,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1794,6 +1796,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -59,8 +59,8 @@ index f2bc8ca7..5ab3f572 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -1680,6 +1700,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
-     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+@@ -2219,6 +2239,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+     {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
 +    {LLM_TENSOR_BSKCN_TV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -68,18 +68,18 @@ index f2bc8ca7..5ab3f572 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 41a023da..525c1b7d 100644
+index 7af587e7..3ea994c7 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -73,6 +73,7 @@ enum llm_arch {
-     LLM_ARCH_GRANITE,
+@@ -82,6 +82,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE_MOE,
+     LLM_ARCH_GRANITE_HYBRID,
      LLM_ARCH_CHAMELEON,
 +    LLM_ARCH_SOLAR,
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -146,6 +147,7 @@ enum llm_kv {
+@@ -168,6 +169,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
      LLM_KV_ATTENTION_SLIDING_WINDOW,
      LLM_KV_ATTENTION_SCALE,
@@ -87,7 +87,7 @@ index 41a023da..525c1b7d 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -346,6 +348,7 @@ enum llm_tensor {
+@@ -394,6 +396,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -96,11 +96,11 @@ index 41a023da..525c1b7d 100644
      LLM_TENSOR_CONVNEXT_DW,
      LLM_TENSOR_CONVNEXT_NORM,
 diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
-index 90dfe7a7..8a667960 100644
+index 7a06368d..35fc054f 100644
 --- a/src/llama-hparams.cpp
 +++ b/src/llama-hparams.cpp
-@@ -70,6 +70,14 @@ uint32_t llama_hparams::n_embd_v_s() const {
-     return ssm_d_state * ssm_d_inner;
+@@ -146,6 +146,14 @@ uint32_t llama_hparams::n_pos_per_embd() const {
+     return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
  }
  
 +bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
@@ -113,12 +113,12 @@ index 90dfe7a7..8a667960 100644
 +
  bool llama_hparams::is_swa(uint32_t il) const {
      if (il < n_layer) {
-         return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+         return swa_layers[il];
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 7ee6a5b7..48dce407 100644
+index bd231224..29bd9056 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
-@@ -55,6 +55,8 @@ struct llama_hparams {
+@@ -62,6 +62,8 @@ struct llama_hparams {
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
      std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
  
@@ -127,9 +127,9 @@ index 7ee6a5b7..48dce407 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -154,6 +156,9 @@ struct llama_hparams {
-     // dimension of the recurrent state embeddings
-     uint32_t n_embd_v_s() const;
+@@ -220,6 +222,9 @@ struct llama_hparams {
+ 
+     uint32_t n_pos_per_embd() const;
  
 +    // Block skip connection
 +    bool n_bskcn(uint32_t n, uint32_t il) const;
@@ -138,10 +138,10 @@ index 7ee6a5b7..48dce407 100644
  };
  
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 4cce5166..7f6617fa 100644
+index f71c40f8..7eab9b68 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
-@@ -439,6 +439,7 @@ namespace GGUFMeta {
+@@ -465,6 +465,7 @@ namespace GGUFMeta {
      // TODO: this is not very clever - figure out something better
      template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
      template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
@@ -150,10 +150,10 @@ index 4cce5166..7f6617fa 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 3a4e72a3..db62973f 100644
+index 58ca7df7..280129e1 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1402,6 +1402,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1706,6 +1706,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -175,7 +175,7 @@ index 3a4e72a3..db62973f 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -3774,6 +3789,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -4793,6 +4808,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -210,12 +210,12 @@ index 3a4e72a3..db62973f 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                          layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                          layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-@@ -12397,6 +12440,165 @@ struct llm_build_chameleon : public llm_graph_context {
+@@ -15495,6 +15538,165 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
      }
  };
  
 +struct llm_build_solar : public llm_graph_context {
-+    llm_build_solar(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
++    llm_build_solar(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
 +        const int64_t n_embd_head = hparams.n_embd_head_v;
 +        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 +        GGML_ASSERT(n_embd_head == hparams.n_rot);
@@ -270,7 +270,7 @@ index 3a4e72a3..db62973f 100644
 +            // self-attention
 +            {
 +                // rope freq factors for llama3; may return nullptr for llama2 and other models
-+                ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
++                ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
 +
 +                // compute Q and K and RoPE them
 +                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -314,7 +314,7 @@ index 3a4e72a3..db62973f 100644
 +                cb(Kcur, "Kcur", il);
 +                cb(Vcur, "Vcur", il);
 +
-+                cur = build_attn(inp_attn, gf,
++                cur = build_attn(inp_attn,
 +                        model.layers[il].wo, model.layers[il].bo,
 +                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
 +                cb(cur, "attn_out", il);
@@ -373,33 +373,33 @@ index 3a4e72a3..db62973f 100644
 +    }
 +};
 +
- struct llm_build_wavtokenizer_dec : public llm_graph_context {
-     llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
-         ggml_tensor * cur;
-@@ -13157,6 +13359,10 @@ llm_graph_result_ptr llama_model::build_graph(
+ // ref: https://github.com/facebookresearch/chameleon
+ // based on the original build_llama() function, changes:
+ //   * qk-norm
+@@ -18439,6 +18641,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
              {
-                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
+                 llm = std::make_unique<llm_build_chameleon>(*this, params);
              } break;
 +        case LLM_ARCH_SOLAR:
 +            {
-+                llm = std::make_unique<llm_build_solar>(*this, params, gf);
++                llm = std::make_unique<llm_build_solar>(*this, params);
 +            } break;
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
-                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
-@@ -13301,6 +13507,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
-         case LLM_ARCH_GRANITE:
+                 llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
+@@ -18652,6 +18858,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE_MOE:
+         case LLM_ARCH_GRANITE_HYBRID:
          case LLM_ARCH_CHAMELEON:
 +        case LLM_ARCH_SOLAR:
          case LLM_ARCH_BAILINGMOE:
-             return LLAMA_ROPE_TYPE_NORM;
- 
+         case LLM_ARCH_NEO_BERT:
+         case LLM_ARCH_SMOLLM3:
 diff --git a/src/llama-model.h b/src/llama-model.h
-index 6bdec263..43746c7d 100644
+index 6fcd74d5..09964533 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
-@@ -65,6 +65,7 @@ enum llm_type {
+@@ -70,6 +70,7 @@ enum llm_type {
      LLM_TYPE_15B,
      LLM_TYPE_16B,
      LLM_TYPE_20B,
@@ -407,9 +407,9 @@ index 6bdec263..43746c7d 100644
      LLM_TYPE_27B,
      LLM_TYPE_30B,
      LLM_TYPE_32B,
-@@ -315,6 +316,8 @@ struct llama_layer {
-     struct ggml_tensor * ffn_up_scale   = nullptr;
-     struct ggml_tensor * ffn_down_scale = nullptr;
+@@ -367,6 +368,8 @@ struct llama_layer {
+     // openai-moe
+     struct ggml_tensor * attn_sinks = nullptr;
  
 +    struct ggml_tensor * bskcn_tv = nullptr;
 +

llama/patches/0006-fix-deepseek-deseret-regex.patch → llama/patches/0005-fix-deepseek-deseret-regex.patch

@@ -12,10 +12,10 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 806c1b3d..10f34d33 100644
+index 8ebe11cf..c011008f 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
-@@ -298,7 +298,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
+@@ -299,7 +299,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
              case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
                  regex_exprs = {
                      "[\r\n]",
@@ -25,7 +25,7 @@ index 806c1b3d..10f34d33 100644
                      "\\s+$",
                      "[一-龥ࠀ-一가-퟿]+",
 diff --git a/src/unicode.cpp b/src/unicode.cpp
-index e63bb4ab..73cb2b1a 100644
+index 65f36651..ce336a22 100644
 --- a/src/unicode.cpp
 +++ b/src/unicode.cpp
 @@ -2,6 +2,11 @@
@@ -62,7 +62,7 @@ index e63bb4ab..73cb2b1a 100644
  #if defined(__clang__)
      // disable C++17 deprecation warning for std::codecvt_utf8
  #    pragma clang diagnostic push
-@@ -213,6 +233,7 @@ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+@@ -218,6 +238,7 @@ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
  #endif
  
      return conv.from_bytes(s);

llama/patches/0007-maintain-ordering-for-rules-for-grammar.patch → llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index 5b3059c2..656b3eca 100644
+index 637891f5..98b8280f 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
-@@ -349,7 +349,7 @@ private:
+@@ -307,7 +307,7 @@ private:
      friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
      std::function<json(const std::string &)> _fetch_json;
      bool _dotall;

llama/patches/0009-sort-devices-by-score.patch → llama/patches/0007-sort-devices-by-score.patch

@@ -11,10 +11,10 @@ with the fastest acceleration is loaded
  1 file changed, 13 insertions(+), 8 deletions(-)
 
 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index 405d8e31..4e67d243 100644
+index 6c315137..3040b2aa 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
-@@ -157,7 +157,7 @@ struct ggml_backend_reg_entry {
+@@ -162,7 +162,7 @@ struct ggml_backend_reg_entry {
  
  struct ggml_backend_registry {
      std::vector<ggml_backend_reg_entry> backends;
@@ -23,7 +23,7 @@ index 405d8e31..4e67d243 100644
  
      ggml_backend_registry() {
  #ifdef GGML_USE_CUDA
-@@ -202,7 +202,7 @@ struct ggml_backend_registry {
+@@ -207,7 +207,7 @@ struct ggml_backend_registry {
          }
      }
  
@@ -32,7 +32,7 @@ index 405d8e31..4e67d243 100644
          if (!reg) {
              return;
          }
-@@ -213,15 +213,20 @@ struct ggml_backend_registry {
+@@ -218,15 +218,20 @@ struct ggml_backend_registry {
  #endif
          backends.push_back({ reg, std::move(handle) });
          for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
@@ -56,7 +56,7 @@ index 405d8e31..4e67d243 100644
      }
  
      ggml_backend_reg_t load_backend(const fs::path & path, bool silent) {
-@@ -265,7 +270,7 @@ struct ggml_backend_registry {
+@@ -270,7 +275,7 @@ struct ggml_backend_registry {
  
          GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str());
  
@@ -65,7 +65,7 @@ index 405d8e31..4e67d243 100644
  
          return reg;
      }
-@@ -288,7 +293,7 @@ struct ggml_backend_registry {
+@@ -293,7 +298,7 @@ struct ggml_backend_registry {
          // remove devices
          devices.erase(
              std::remove_if(devices.begin(), devices.end(),
@@ -74,7 +74,7 @@ index 405d8e31..4e67d243 100644
              devices.end());
  
          // remove backend
-@@ -346,7 +351,7 @@ size_t ggml_backend_dev_count() {
+@@ -351,7 +356,7 @@ size_t ggml_backend_dev_count() {
  
  ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
      GGML_ASSERT(index < ggml_backend_dev_count());