Files
llama.cpp/tools/mtmd/README-dev.md
T
Xuan-Son Nguyen e4b620af19 mtmd: support voxcpm2 (#32443)
* mtmd: support voxcpm

* speed up a bit

* various fixes
2026-11-19 02:39:05 +01:00

14 KiB

libmtmd dev guide

History

Please refer to multimodal.md for a broader context.

In short:

  • libmtmd started as a wrapper around libllava / clip.cpp
  • Various components that used to be in clip.cpp are moved progressively to mtmd. For example, preprocessor is now part of mtmd
  • clip.cpp has since been split into mtmd-model (hparams and weights, one mtmd_model per modality), mtmd-subcontext (backends and compute, one mtmd_subcontext per modality) and mtmd-graph (graph builders, see models/). Only the GGUF format keeps the clip. name (metadata keys and the clip architecture), for compatibility with existing mmproj files

Terminologies

  • mtmd: MulTiMoDal
  • bitmap: representing a raw input data, for example: RGB image, PCM audio
  • tiles / slices: for llava-uhd-style models, the preprocessor breaks a large input into smaller square images called tiles or slices
  • chunk: a mtmd_input_chunk represents a preprocessed input that can then be passed through mtmd_encode()

Structure

Overview of the libmtmd structure:

  • mtmd_context (mtmd.cpp): the public handle
    • mtmd_models: one mmproj file, loaded by mtmd_model_load()
      • mtmd_model v / a / gen_a (mtmd-model.h): one per modality, holds hparams, weights (ctx_data + buf) and requirements
    • mtmd_backends (mtmd-subcontext.h): GPU/CPU backend instances, CPU threadpool and one scheduler, shared by the subcontexts
    • compute_mutex: held by every public function that runs a compute pass, so only one pass at a time runs on the backends
    • mtmd_subcontext ctx_v / ctx_a / ctx_gen_a (mtmd-subcontext.h): one per model, holds a const ref to it and a ref to the backends
      • buf_compute_meta and the resolved flash attn of its graphs
      • mtmd_graph (mtmd-graph.h, models/): built for each encode() call, temporary
    • mtmd_image_preprocessor / mtmd_audio_preprocessor (mtmd-image.h, mtmd-audio.h): read the hparams of their model
  • mtmd_gen_state / mtmd_audio_stream: public, one per sequence / stream
    • mtmd_subctx_gen_state / mtmd_subctx_enc_state: allocated on the backend of a subcontext, see gen_state_init() / enc_state_init()

Rules of thumb:

  • mtmd_model is read-only once loaded. Anything that depends on the backend (flash attn resolution, warmup) belongs to mtmd_subcontext
  • Like the pp/tg graphs of llama_context, the subcontexts share one scheduler: each warmup reserves its worst-case graph on it, so the compute buffers fit the largest one, and switching modality re-plans inside them
  • A model declares its runtime needs (e.g. max_nodes, flash_attn_auto) in load_hparams, the subcontext applies them
  • Models and backends must outlive the subcontexts, mtmd_context declares them first for that reason
  • Models having both audio and vision encoders must have different arch names (projector types), postfixed by a and v. See Gemma 4 as an example (gemma4a / gemma4v)
  • Locking happens only at the public API boundary (mtmd.cpp): a new public function that runs a graph must take compute_mutex. The internals (subcontexts, preprocessors, mtmd_audio_log_mel()) assume it is held and never lock
  • No magic numbers for a model's fixed shapes (layer counts, codebook sizes, window lengths...): define them once as static constexpr in the graph struct in models/models.h, and refer to them from the loader or the subcontext (e.g. mtmd_graph_zonos3_gen::n_codebooks, mtmd_graph_fish_s2_spkenc::n_tfm_layer)

Pipeline

A typical pipeline of the core libmtmd is as follows:

  • A bitmap (RGB image or PCM audio) is created
  • Bitmap and the text prompt is provided to mtmd_tokenize() that breaks the input into chunks
    • Alternatively, mtmd_tokenize_from_parts() takes a list of pre-split text/media parts instead of a marker-based prompt
    • The tokenizer function first expands a "lazy" bitmap if it finds one. Typically, this is used by video, so that one media token corresponds to one input bitmap
    • For models that support "fused" temporal frames like Qwen-VL, the tokenizer tries to merge pair of consecutive frames into one batch. Only bitmaps marked by mtmd_bitmap_set_mergeable() are merged
    • The preprocessor will then be called, which produces a list of chunks
    • Depending on the model itself, special tokens will be injected to separate image chunks (i.e. llava-uhd-style models)
  • Multiple bitmaps may be batched together to form a larger mtmd_batch()
  • Single image or batch is encoded, via mtmd_encode() or mtmd_batch_encode()
  • Get the output embeddings

Helper

We provide a set of helper functions via mtmd_helper to make using libmtmd easier. The helper provides:

  • Image, audio and video file decoding (for example, decode raw JPEG into RGB bitmap)
  • Manage llama_batch and calls to llama_decode

Audio generation support

Audio generation is added to mtmd in PR #26254

Currently, we support the 3-stage pipeline below which should cover most TTS models:

  • Stage 1: Backbone / Semantic Stage: Backbone model accepts text prompt and reference voice as input
  • Stage 2: Acoustic Detail Generator: A model takes the hidden state from backbone and generate audio details (usually as audio codes or mel-spectrogram)
  • Stage 3: Waveform Reconstruction: Convert the semantic and acoustic data from previous stages to the final waveform

For example, Qwen3-TTS:

  • Reference voice is encoded using ECAPA-TDNN speaker encoder (speaker_encoder)
  • Text prompt and reference voice are processed via a backbone (talker.model)
  • A model converts sampled semantic token and hidden state from stage 2 into a list of 15 acoustic codes (talker.code_predictor)
  • 16 generated codes are converted into waveform (code2wav)

API design constraints

Due to wide variety of audio generation pipelines, the mtmd_gen_audio system is designed to be flexible and reusable by new models.

mtmd_gen_audio is split into 2 main API:

  • Core API mtmd.h: handles main inference. Important: mtmd_gen_state holds only the per-sequence data that the mmproj itself writes and reads back (RNG, decoder KV / conv state, output buffers). Data that goes to or comes from the backbone (hidden states, codes, PCM) must stay explicit in mtmd_gen_inp / mtmd_gen_out; the caller drives the backbone and handles audio frame accumulation.
  • Helper API mtmd-helper.h: provides a model-agnostic stateful API. Usage example can be found in the tools/tts directory.

Running a transformer stage on libllama (mtmd-owned llama_context)

Some stages are a full LLM-sized transformer that attends to the keys/values of the backbone, for example the flow-matching NAR of YuE2 (same layer shapes as the backbone, its own weights, bidirectional over its rows, and it reads the backbone KV). Do not re-implement it as an mtmd graph. Instead, mtmd can own a second llama_model + llama_context (see mtmd_yue2_nar_load() and mtmd_gen_flow_yue2(), or mtmd_voxcpm2_res_load() and mtmd_gen_code_voxcpm2() in mtmd.cpp):

  • Conversion: store the layers in the mmproj under a.gen.<stage>. with the libllama tensor names of an existing arch (e.g. a.gen.nar.blk.0.attn_q.weight)
  • Loading: build a synthetic gguf_context with the arch hparams (general.architecture, tokenizer.ggml.model = none, vocab_size = 1, ...) and one tensor info per mmproj tensor (only the type is used, the shapes come from the hparams), then call llama_model_init_from_user() (mtmd_owned_llama_load() does this, the caller only writes the hparams). The set_tensor_data callback reads the data from the mmproj file, so no copy is kept in the mtmd model
    • libllama also creates every optional tensor of the arch (biases, *.scale, *.input_scale, dense_2, ...) for a model built from metadata. The callback must fill them with values that leave the graph unchanged: 0 for biases, 1 for scales and rope_freqs, identity for dense_2. Zeros everywhere silently make every matmul output 0
  • Backbone keys/values: the helper passes the backbone sequence state (llama_state_seq_get_data()) through mtmd_gen_inp, mtmd loads it into its context with llama_state_seq_set_data(). This only checks the layer count and the K/V types and row sizes, so it works across models with the same KV layout. The two contexts must agree on the KV type and on flash attention (v_trans)
  • Inputs and outputs are embedding rows (llama_batch::embd, llama_get_embeddings() with LLAMA_POOLING_TYPE_NONE), the projections around them stay small mtmd graphs
  • Non-causal rows: LLAMA_ATTENTION_TYPE_NON_CAUSAL, all rows in one ubatch (n_ubatch >= rows), they see every cached position of the sequence. Remove them with llama_memory_seq_rm() after each pass
  • Size the context per call (n_ctx, n_ubatch) instead of keeping one for the whole session
  • Causal stage with its own KV cache across calls (the residual LM of VoxCPM2): the mtmd_gen_state owns the llama_context, mtmd_gen_state_reset() clears its memory. Options of the arch may switch parts off, e.g. RoPE with rope.scaling.finetuned = false on granite

Text preprocessing (the tts chat template)

Anything a model needs done to the input text before tokenization (prompt format, text normalization, text to phonemes) goes in tokenizer.chat_template.tts, as a jinja program. It receives the text as messages[0].content, plus language and voice.

Examples:

  • Prompt format: conversion/lfm3.py
  • Text normalization: models/templates/Zyphra-ZONOS3-Mini-tts.jinja
  • Text to phonemes (G2P): models/templates/hexgrad-Kokoro-82M-tts.jinja

Functions available to the template (see docs/tts-norm-patterns.md):

  • tts_norm_scan() / tts_norm_sub(): pattern matching for text normalization
  • tts_dict(): lookup in a dictionary stored in the GGUF, such as a pronunciation lexicon

Add test cases to tests/test-jinja.cpp, which also checks them against the python reference with -py.

Rules for phoneme-based models:

  • C++ only holds the pieces that every such model needs (the functions above). The rules of one model (lexicon handling, stress, suffixes, number spelling, ...) live in its jinja template
  • The reason is binary size: phoneme logic compiled into C++ is shipped to every user, and most of them never run a TTS model. A template only costs the users of that model
  • Do not add a model-specific or language-specific G2P step in C++, and do not add a dependency for it (e.g. espeak-ng)

G2P in jinja has limits, mostly for languages other than English (no lexicon, or rules that need a real morphological analyzer). For now the escape hatch is on the application side: the user runs the text to phonemes step in their own code and sends the phonemes with raw_prompt (see the /tts endpoint of llama-server), which skips the template. Document it in docs/multimodal/tts.md when a model relies on it.

Checklist for porting new audio generation models to mtmd

  1. Make sure to consult merged PRs about adding new TTS models, especially reviewer comments
  2. Establish a list of reusable and missing components from the current mtmd implementation.
  3. For GGUF conversion:
    • Backbone model should be converted to a normal text model (loadable via libllama)
      • If model used hard-coded embedding row ID, append them to token embeddings and assign token name for them (see qwen3tts.py)
      • If model have a specific output logits head for audio codes (usually semantic code), keep the head as-is and pad the logits at inference time (see src/models/qwen3vl.cpp)
    • Sidecar models (code2wav, bigvgan, etc) must live inside the mmproj GGUF (but can be in different mtmd_subcontext if necessary)
      • Note: it should use ggml_build_forward_select to select graphs if multiple graphs living in the same context
    • Reuse existing GGUF metadata key name and tensor name whenever possible; think twice before adding extensive changes to GGUF writer. For example, Qwen3-TTS hard-code part of the hparams in mtmd-model.cpp as they won't likely to change.
    • For tensor naming:
      • Prefixed with a.* for tensors used by speaker encoder pipeline
      • Prefixed with a.gen.* for generation stages (code / mel-spectrogram / PCM generation)
    • For GGUF metadata:
      • Reuse as many existing keys as possible
      • In most cases, you can hard-code model configs in the model graph class, or in mtmd_hparams
      • If some values need to be exposed to the mtmd_helper layer, hard-code them in mtmd_helper and distinguish by pipeline and mtmd_gen_audio_info::model_variant if necessary
      • Do NOT add new GGUF metadata or new fields to mtmd_gen_audio_info unless you can prove that you absolutely need them
    • If the model requires special preprocessing of the input text prompt, see Text preprocessing
  4. Make sure most of the changes happen inside mtmd-helper-gen.cpp. A good PR looks like this:
    • 10-20% changes is to add new backbone (text) model and conversion
    • 60% changes inside mtmd-helper-gen.cpp
    • 10% changes inside libmtmd core (mtmd-model, mtmd-subcontext, models/)
    • The rest downstream code (CLI, server) should have no changes at all
  5. Validate with several concurrent requests on llama-server (e.g. -np 2, one plain and one streaming request at the same time), not only with llama-tts
    • The server interleaves the slots on one llama_context, so a pipeline must copy what it needs (hidden state, logits) right after its own decode, not read it at the next step
  6. Update usage documentation in docs/multimodal/tts.md

IMPORTANT: If your model needs changes that don't fit the existing infrastructure, open an issue first for discussion.

No-go checklist (these will get the PR rejected and require discussion before proceeding):

  • Violating the API design constraints stated above
  • Adding a new model-specific binary: the API and binary surface must stay model-agnostic
  • Adding new flags to conversion script: this script must stay model-agnostic