mtmd : add video input support (ggml-org#24269)

* wip

* ok: lazy bitmap API

* remember to free lazy text

* wip

* add mtmd_helper_video

* support video input on server (base64 input)

* add MTMD_VIDEO config

* add timestamp

* update CLI

* cli: allow auto-completion for video

* add --video arg

* fix build

* update docs

* rename as suggested

Author: ngxson

common/arg.cpp

@@ -2221,8 +2221,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
-        {"--image", "--audio"}, "FILE",
-        "path to an image or audio file. use with multimodal models, use comma-separated values for multiple files\n",
+        {"--image", "--audio", "--video"}, "FILE",
+        "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",
         [](common_params & params, const std::string & value) {
             for (const auto & item : parse_csv_row(value)) {
                 params.image.emplace_back(item);

common/common.h

@@ -571,7 +571,7 @@ struct common_params {
     struct common_params_model mmproj;
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s)
+    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
 

tests/test-mtmd-c-api.c

@@ -2,6 +2,7 @@
 #include <assert.h>
 
 #include "mtmd.h"
+#include "mtmd-helper.h"
 
 int main(void) {
     printf("\n\nTesting libmtmd C API...\n");
@@ -17,6 +18,11 @@ int main(void) {
         return 1;
     }
 
+    // simple test for the helper
+    size_t n_tokens_total = mtmd_helper_get_n_tokens(chunks);
+    printf("Total tokens in chunks: %zu\n", n_tokens_total);
+    assert(n_tokens_total > 0);
+
     size_t n_chunks = mtmd_input_chunks_size(chunks);
     printf("Number of chunks: %zu\n", n_chunks);
     assert(n_chunks > 0);

tools/cli/cli.cpp

@@ -235,14 +235,15 @@ struct cli_context {
 };
 
 // TODO?: Make this reusable, enums, docs
-static const std::array<std::string_view, 7> cmds = {
+static const std::array<std::string_view, 8> cmds = {
     "/audio ",
     "/clear",
     "/exit",
     "/glob ",
     "/image ",
     "/read ",
     "/regen",
+    "/video ",
 };
 
 static std::vector<std::pair<std::string, size_t>> auto_completion_callback(std::string_view line, size_t cursor_byte_pos) {
@@ -457,6 +458,9 @@ int llama_cli(int argc, char ** argv) {
     if (inf.has_inp_audio) {
         console::log("  /audio <file>       add an audio file\n");
     }
+    if (inf.has_inp_video) {
+        console::log("  /video <file>       add a video file\n");
+    }
     console::log("\n");
 
     // interactive loop
@@ -553,7 +557,8 @@ int llama_cli(int argc, char ** argv) {
             continue;
         } else if (
                 (string_starts_with(buffer, "/image ") && inf.has_inp_image) ||
-                (string_starts_with(buffer, "/audio ") && inf.has_inp_audio)) {
+                (string_starts_with(buffer, "/audio ") && inf.has_inp_audio) ||
+                (string_starts_with(buffer, "/video ") && inf.has_inp_video)) {
             // just in case (bad copy-paste for example), we strip all trailing/leading spaces
             std::string fname = string_strip(buffer.substr(7));
             std::string marker = ctx_cli.load_input_file(fname, true);

tools/mtmd/CMakeLists.txt

@@ -1,5 +1,8 @@
 # mtmd
 
+set(MTMD_VIDEO ON CACHE BOOL "enable video support in mtmd (requires ffmpeg binary in PATH)")
+# TODO: add MTMD_VIDEO_METHOD in the future to select between ffmpeg and other backends
+
 find_package(Threads REQUIRED)
 
 add_library(mtmd
@@ -63,6 +66,10 @@ target_include_directories(mtmd PRIVATE ../..)
 target_include_directories(mtmd PRIVATE ../../vendor)
 target_compile_features   (mtmd PRIVATE cxx_std_17)
 
+if (MTMD_VIDEO)
+    target_compile_definitions(mtmd PRIVATE MTMD_VIDEO)
+endif()
+
 if (BUILD_SHARED_LIBS)
     set_target_properties     (mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(mtmd PRIVATE LLAMA_BUILD)

tools/mtmd/mtmd-cli.cpp

@@ -77,6 +77,7 @@ struct mtmd_cli_context {
     int                 n_batch;
 
     mtmd::bitmaps bitmaps;
+    std::vector<mtmd_helper::video_ptr> videos;
 
     // chat template
     common_chat_templates_ptr tmpls;
@@ -166,11 +167,14 @@ struct mtmd_cli_context {
     }
 
     bool load_media(const std::string & fname) {
-        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str(), false));
-        if (!bmp.ptr) {
+        auto res = mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str(), false);
+        if (!res.bitmap) {
             return false;
         }
-        bitmaps.entries.push_back(std::move(bmp));
+        bitmaps.entries.emplace_back(res.bitmap);
+        if (res.video_ctx) {
+            videos.emplace_back(res.video_ctx);
+        }
         return true;
     }
 };
@@ -253,6 +257,7 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
     }
 
     ctx.bitmaps.entries.clear();
+    ctx.videos.clear();
 
     llama_pos new_n_past;
     if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(),
@@ -373,6 +378,9 @@ int main(int argc, char ** argv) {
         if (mtmd_support_audio(ctx.ctx_vision.get())) {
             LOG("\n   /audio <path>    load an audio");
         }
+        if (mtmd_helper_support_video(ctx.ctx_vision.get())) {
+            LOG("\n   /video <path>    load a video");
+        }
         LOG("\n   /clear           clear the chat history");
         LOG("\n   /quit or /exit   exit the program");
         LOG("\n");
@@ -407,14 +415,15 @@ int main(int argc, char ** argv) {
             g_is_generating = true;
             bool is_image = line == "/image" || line.find("/image ") == 0;
             bool is_audio = line == "/audio" || line.find("/audio ") == 0;
-            if (is_image || is_audio) {
+            bool is_video = line == "/video" || line.find("/video ") == 0;
+            if (is_image || is_audio || is_video) {
                 if (line.size() < 8) {
                     LOG_ERR("ERR: Missing media filename\n");
                     continue;
                 }
                 std::string media_path = line.substr(7);
                 if (ctx.load_media(media_path)) {
-                    LOG("%s %s loaded\n", media_path.c_str(), is_image ? "image" : "audio");
+                    LOG("%s %s loaded\n", media_path.c_str(), is_image ? "image" : is_audio ? "audio" : "video");
                     content += mtmd_default_marker();
                 }
                 // else, error is already printed by libmtmd