mtmd : add ultravox audio input

.editorconfig

@@ -48,3 +48,7 @@ end_of_line = unset
 charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
+
+[tools/mtmd/miniaudio.h]
+trim_trailing_whitespace = unset
+insert_final_newline = unset

README.md

@@ -580,3 +580,4 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
 - [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
 - [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
+- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain

common/arg.cpp

@@ -39,7 +39,7 @@
 using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_SERVER,
 };
 
@@ -2233,12 +2233,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
     add_opt(common_arg(
-        {"--image"}, "FILE",
-        "path to an image file. use with multimodal models. Specify multiple times for batching",
+        {"--image", "--audio"}, "FILE",
+        "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",
         [](common_params & params, const std::string & value) {
             params.image.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples({LLAMA_EXAMPLE_MTMD}));
     if (llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
@@ -2868,7 +2868,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.chat_template = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
     add_opt(common_arg(
         {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
         string_format(

common/common.h

@@ -76,7 +76,7 @@ enum llama_example {
     LLAMA_EXAMPLE_SERVER,
     LLAMA_EXAMPLE_CVECTOR_GENERATOR,
     LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
+    LLAMA_EXAMPLE_MTMD,
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,

docs/multimodal.md

@@ -4,7 +4,9 @@ llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools
 - [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:
+Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
+
+To enable it, you can use one of the 2 methods below:
 
 - Use `-hf` option with a supported model (see a list of pre-quantized model below)
     - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
@@ -37,6 +39,8 @@ Replaces the `(tool_name)` with the name of binary you want to use. For example,
 
 NOTE: some models may require large context window, for example: `-c 8192`
 
+**Vision models**:
+
 ```sh
 # Gemma 3
 (tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
@@ -78,3 +82,11 @@ NOTE: some models may require large context window, for example: `-c 8192`
 # Llama 4 Scout
 (tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
 ```
+
+**Audio models**:
+
+```sh
+# Ultravox 0.5
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF
+(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF
+```

gguf-py/gguf/constants.py

@@ -219,10 +219,13 @@ class Adapter:
         TYPE       = "adapter.type"
         LORA_ALPHA = "adapter.lora.alpha"
 
-    class ClipVision:
+    class Clip:
         PROJECTOR_TYPE      = "clip.projector_type"
         HAS_VISION_ENCODER  = "clip.has_vision_encoder"
+        HAS_AUDIO_ENCODER   = "clip.has_audio_encoder"
         HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
+
+    class ClipVision:
         IMAGE_SIZE          = "clip.vision.image_size"
         PATCH_SIZE          = "clip.vision.patch_size"
         EMBEDDING_LENGTH    = "clip.vision.embedding_length"
@@ -243,19 +246,33 @@ class Attention:
         class Projector:
             SCALE_FACTOR    = "clip.vision.projector.scale_factor"
 
+    class ClipAudio:
+        NUM_MEL_BINS        = "clip.audio.num_mel_bins"
+        EMBEDDING_LENGTH    = "clip.audio.embedding_length"
+        FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
+        PROJECTION_DIM      = "clip.audio.projection_dim"
+        BLOCK_COUNT         = "clip.audio.block_count"
+
+        class Attention:
+            HEAD_COUNT      = "clip.audio.attention.head_count"
+            LAYERNORM_EPS   = "clip.audio.attention.layer_norm_epsilon"
+
+        class Projector:
+            STACK_FACTOR    = "clip.audio.projector.stack_factor"
+
 #
 # recommended mapping of model tensor names for storage in gguf
 #
 
 
 class GGUFType:
-    MODEL       = "model"
-    ADAPTER     = "adapter"
-    CLIP_VISION = "clip-vision"
+    MODEL   = "model"
+    ADAPTER = "adapter"
+    MMPROJ  = "mmproj" # dummy, unused for now
 
 
 class MODEL_ARCH(IntEnum):
-    CLIP_VISION      = auto() # dummy arch for clip.cpp
+    MMPROJ           = auto() # dummy arch for clip.cpp
     LLAMA            = auto()
     LLAMA4           = auto()
     DECI             = auto()
@@ -514,10 +531,27 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_QUERY       = auto() # minicpmv
     V_TOK_EMBD_IMG_BREAK = auto() # pixtral
     V_MM_PATCH_MERGER    = auto() # mistral small 3.1
+    # audio (mtmd)
+    A_ENC_EMBD_POS       = auto()
+    A_ENC_CONV1D         = auto()
+    A_PRE_NORM           = auto()
+    A_POST_NORM          = auto()
+    A_ENC_ATTN_Q         = auto()
+    A_ENC_ATTN_K         = auto()
+    A_ENC_ATTN_V         = auto()
+    A_ENC_INPUT_NORM     = auto()
+    A_ENC_OUTPUT         = auto()
+    A_ENC_OUTPUT_NORM    = auto()
+    A_ENC_FFN_UP         = auto()
+    A_ENC_FFN_GATE       = auto()
+    A_ENC_FFN_DOWN       = auto()
+    A_MMPROJ             = auto()
+    A_MM_NORM_PRE        = auto()
+    A_MM_NORM_MID        = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
-    MODEL_ARCH.CLIP_VISION:      "clip", # dummy arch for clip.cpp
+    MODEL_ARCH.MMPROJ:           "clip", # dummy arch for clip.cpp
     MODEL_ARCH.LLAMA:            "llama",
     MODEL_ARCH.LLAMA4:           "llama4",
     MODEL_ARCH.DECI:             "deci",
@@ -776,10 +810,27 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
     MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
     MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
+    # audio (mtmd)
+    MODEL_TENSOR.A_ENC_EMBD_POS:            "a.position_embd",
+    MODEL_TENSOR.A_ENC_CONV1D:              "a.conv1d.{bid}",
+    MODEL_TENSOR.A_PRE_NORM:                "a.pre_ln",
+    MODEL_TENSOR.A_POST_NORM:               "a.post_ln",
+    MODEL_TENSOR.A_ENC_ATTN_Q:              "a.blk.{bid}.attn_q",
+    MODEL_TENSOR.A_ENC_ATTN_K:              "a.blk.{bid}.attn_k",
+    MODEL_TENSOR.A_ENC_ATTN_V:              "a.blk.{bid}.attn_v",
+    MODEL_TENSOR.A_ENC_INPUT_NORM:          "a.blk.{bid}.ln1",
+    MODEL_TENSOR.A_ENC_OUTPUT:              "a.blk.{bid}.attn_out",
+    MODEL_TENSOR.A_ENC_OUTPUT_NORM:         "a.blk.{bid}.ln2",
+    MODEL_TENSOR.A_ENC_FFN_UP:              "a.blk.{bid}.ffn_up",
+    MODEL_TENSOR.A_ENC_FFN_GATE:            "a.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.A_ENC_FFN_DOWN:            "a.blk.{bid}.ffn_down",
+    MODEL_TENSOR.A_MMPROJ:                  "mm.a.mlp.{bid}",
+    MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
+    MODEL_TENSOR.A_MM_NORM_MID:             "mm.a.norm_mid",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
-    MODEL_ARCH.CLIP_VISION: [
+    MODEL_ARCH.MMPROJ: [
         MODEL_TENSOR.V_MMPROJ,
         MODEL_TENSOR.V_MMPROJ_FC,
         MODEL_TENSOR.V_MMPROJ_MLP,
@@ -819,6 +870,23 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.V_RESMPL_QUERY,
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
         MODEL_TENSOR.V_MM_PATCH_MERGER,
+        # audio
+        MODEL_TENSOR.A_ENC_EMBD_POS,
+        MODEL_TENSOR.A_ENC_CONV1D,
+        MODEL_TENSOR.A_PRE_NORM,
+        MODEL_TENSOR.A_POST_NORM,
+        MODEL_TENSOR.A_ENC_ATTN_Q,
+        MODEL_TENSOR.A_ENC_ATTN_K,
+        MODEL_TENSOR.A_ENC_ATTN_V,
+        MODEL_TENSOR.A_ENC_INPUT_NORM,
+        MODEL_TENSOR.A_ENC_OUTPUT,
+        MODEL_TENSOR.A_ENC_OUTPUT_NORM,
+        MODEL_TENSOR.A_ENC_FFN_UP,
+        MODEL_TENSOR.A_ENC_FFN_GATE,
+        MODEL_TENSOR.A_ENC_FFN_DOWN,
+        MODEL_TENSOR.A_MMPROJ,
+        MODEL_TENSOR.A_MM_NORM_PRE,
+        MODEL_TENSOR.A_MM_NORM_MID,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -2186,6 +2254,7 @@ class VisionProjectorType:
     LLAMA4 = "llama4"
     QWEN2VL = "qwen2vl_merger"
     QWEN25VL = "qwen2.5vl_merger"
+    ULTRAVOX = "ultravox"
     INTERNVL = "internvl"
 
 

gguf-py/gguf/gguf_writer.py

@@ -936,12 +936,18 @@ def add_eom_token_id(self, id: int) -> None:
 
     # for vision models
 
+    def add_clip_has_vision_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.Clip.HAS_VISION_ENCODER, value)
+
+    def add_clip_has_audio_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.Clip.HAS_AUDIO_ENCODER, value)
+
+    def add_clip_projector_type(self, value: str) -> None:
+        self.add_string(Keys.Clip.PROJECTOR_TYPE, value)
+
     def add_vision_projection_dim(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)
 
-    def add_vision_has_vision_encoder(self, value: bool) -> None:
-        self.add_bool(Keys.ClipVision.HAS_VISION_ENCODER, value)
-
     def add_vision_patch_size(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.PATCH_SIZE, value)
 
@@ -957,9 +963,6 @@ def add_vision_block_count(self, value: int) -> None:
     def add_vision_head_count(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value)
 
-    def add_vision_projector_type(self, value: str) -> None:
-        self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
-
     def add_vision_attention_layernorm_eps(self, value: float) -> None:
         self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)
 
@@ -987,6 +990,32 @@ def add_vision_projector_scale_factor(self, value: int) -> None:
     def add_vision_n_wa_pattern(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
 
+    # audio models
+
+    def add_audio_projection_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.PROJECTION_DIM, value)
+
+    def add_audio_embedding_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.EMBEDDING_LENGTH, value)
+
+    def add_audio_feed_forward_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.FEED_FORWARD_LENGTH, value)
+
+    def add_audio_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.BLOCK_COUNT, value)
+
+    def add_audio_head_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.Attention.HEAD_COUNT, value)
+
+    def add_audio_attention_layernorm_eps(self, value: float) -> None:
+        self.add_float32(Keys.ClipAudio.Attention.LAYERNORM_EPS, value)
+
+    def add_audio_num_mel_bins(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)
+
+    def add_audio_stack_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -1110,6 +1110,68 @@ class TensorNameMap:
         MODEL_TENSOR.V_MM_PATCH_MERGER: (
             "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
         ),
+
+        # audio (mtmd)
+
+        MODEL_TENSOR.A_ENC_EMBD_POS: (
+            "audio_tower.embed_positions", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_CONV1D: (
+            "audio_tower.conv{bid}", # ultravox
+        ),
+
+        MODEL_TENSOR.A_PRE_NORM: (),
+
+        MODEL_TENSOR.A_POST_NORM: (
+            "audio_tower.layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_Q: (
+            "audio_tower.layers.{bid}.self_attn.q_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_K: (
+            "audio_tower.layers.{bid}.self_attn.k_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_ATTN_V: (
+            "audio_tower.layers.{bid}.self_attn.v_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_INPUT_NORM: (
+            "audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_OUTPUT: (
+            "audio_tower.layers.{bid}.self_attn.out_proj", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_OUTPUT_NORM: (
+            "audio_tower.layers.{bid}.final_layer_norm", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_FFN_UP: (
+            "audio_tower.layers.{bid}.fc1", # ultravox
+        ),
+
+        MODEL_TENSOR.A_ENC_FFN_GATE: (),
+
+        MODEL_TENSOR.A_ENC_FFN_DOWN: (
+            "audio_tower.layers.{bid}.fc2", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MMPROJ: (
+            "audio.multi_modal_projector.linear_{bid}", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MM_NORM_PRE: (
+            "audio.multi_modal_projector.ln_pre", # ultravox
+        ),
+
+        MODEL_TENSOR.A_MM_NORM_MID: (
+            "audio.multi_modal_projector.ln_mid", # ultravox
+        ),
     }
 
     # architecture-specific block mappings