server : vision support via libmtmd (#12898)

* server : (experimental) vision support via libmtmd

* mtmd : add more api around mtmd_image_tokens

* mtmd : add more api around mtmd_image_tokens

* mtmd : ability to calc image hash

* shared_ptr for mtmd_image_tokens

* move hash to user-define ID (fixed)

* abstract out the batch management

* small fix

* refactor logic adding tokens to batch

* implement hashing image

* use FNV hash, now hash bitmap instead of file data

* allow decoding image embedding to be split into batches

* rm whitespace

* disable some features when mtmd is on

* fix --no-mmproj-offload

* mtmd_context_params no timings

* refactor server_inp to server_tokens

* fix the failing test case

* init

* wip

* working version

* add mtmd::bitmaps

* add test target

* rm redundant define

* test: mtmd_input_chunks_free

* rm outdated comment

* fix merging issue

* explicitly create mtmd::input_chunks

* mtmd_input_chunk_copy

* add clone()

* improve server_input struct

* clip :  fix confused naming ffn_up and ffn_down

* rm ffn_i/o/g naming

* rename n_embd, n_ff

* small fix

* no check n_ff

* fix detokenize

* add const to various places

* add warning about breaking changes

* add c api

* helper: use mtmd_image_tokens_get_n_pos

* fix ctx_shift

* fix name shadowing

* more strict condition

* support remote image_url

* remote image_url log

* add CI test

* do not log base64

* add "has_multimodal" to /props

* remove dangling image

* speculative: use slot.cache_tokens.insert

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <[email protected]>

* rm can_be_detokenized

* on prmpt processing done, assert cache_tokens.size

* handle_completions_impl returns void

* adapt the new web ui

* update docs and hot topics

* rm assert

* small fix (2)

---------

Co-authored-by: Georgi Gerganov <[email protected]>

Author: ngxson

README.md

@@ -16,8 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
+- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim

common/arg.cpp

@@ -40,7 +40,7 @@ using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
     LLAMA_EXAMPLE_LLAVA,
-    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+    LLAMA_EXAMPLE_SERVER,
 };
 
 static std::string read_file(const std::string & fname) {

docs/multimodal.md

@@ -0,0 +1,69 @@
+# Multimodal
+
+llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
+- [llama-mtmd-cli](../tools/mtmd/README.md)
+- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
+
+To enable it, can use use one of the 2 methods below:
+
+- Use `-hf` option with a [supported model](../../docs/multimodal.md)
+    - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
+    - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
+- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
+
+By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload`
+
+For example:
+
+```sh
+# simple usage with CLI
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+
+# simple usage with server
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF
+
+# using local file
+llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf
+
+# no GPU offload
+llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
+```
+
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default.
+
+Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
+
+NOTE: some models may require large context window, for example: `-c 8192`
+
+```sh
+# Gemma 3
+(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF
+(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+(tool_name) -hf ggml-org/pixtral-12b-GGUF
+
+# Qwen 2 VL
+(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
+
+# Qwen 2.5 VL
+(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
+(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
+```

tools/mtmd/README.md

@@ -16,38 +16,7 @@ The naming and structure related to multimodal support have evolved, which might
 
 ## Pre-quantized models
 
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
-
-```sh
-# Gemma 3
-llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
-
-# SmolVLM
-llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
-
-# Pixtral 12B
-llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
-
-# Qwen 2 VL
-llama-mtmd-cli -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
-
-# Qwen 2.5 VL
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
-llama-mtmd-cli -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
-
-# Mistral Small 3.1 24B (IQ2_M quantization)
-llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
-```
+See the list of pre-quantized model [here](../../docs/multimodal.md)
 
 ## How it works and what is `mmproj`?
 

tools/server/CMakeLists.txt

@@ -34,8 +34,9 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
+target_include_directories(${TARGET} PRIVATE ../llava)
 target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
-target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
 if (LLAMA_SERVER_SSL)
     find_package(OpenSSL REQUIRED)

tools/server/README.md

@@ -193,6 +193,12 @@ services:
       LLAMA_ARG_PORT: 8080
 ```
 
+### Multimodal support
+
+Multimodal support was added in [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) and is currently an experimental feature.
+
+For more details, please refer to [multimodal documentation](../../docs/multimodal.md)
+
 ## Build
 
 `llama-server` is built alongside everything else from the root of the project
@@ -749,6 +755,9 @@ This endpoint is public (no API key check). By default, it is read-only. To make
   "total_slots": 1,
   "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
   "chat_template": "...",
+  "modalities": {
+    "vision": false
+  },
   "build_info": "b(build number)-(build commit hash)"
 }
 ```
@@ -757,6 +766,7 @@ This endpoint is public (no API key check). By default, it is read-only. To make
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `model_path` - the path to model file (same with `-m` argument)
 - `chat_template` - the model's original Jinja2 prompt template
+- `modalities` - the list of supported modalities
 
 ### POST `/props`: Change server global properties.
 
@@ -1069,6 +1079,8 @@ print(completion.choices[0].text)
 
 Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
+If model supports multimodal, you can input the media file via `image_url` content part. We support both base64 and remote URL as input. See OAI documentation for more.
+
 *Options:*
 
 See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). llama.cpp `/completion`-specific features such as `mirostat` are also supported.