Files
llama.cpp/tools/mtmd/models/models.h
T
Xuan-Son NguyenandPascal c622f623fc mtmd: support kyutai/hibiki realtime full-duplex (#32491)
* mtmd: support hibiki realtime duplex

* fix and clean up

* fix long session quality

* fix ci

* tests : add the moshi fusion baseline for Metal

The Metal fusion check found 5 RMS_NORM+MUL fusions in the new moshi
arch and no baseline for them.

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-11-21 14:25:19 +01:00

1044 lines
48 KiB
C++

#pragma once
#include "../mtmd-graph.h"
#include <functional>
#include <map>
#include <string>
#include <utility>
#include <vector>
/*
* IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
* We encourage human contributors to ensure the quality and reliability of the codebase.
*/
struct mtmd_graph_siglip : mtmd_graph {
mtmd_graph_siglip(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_gemma4v : mtmd_graph {
mtmd_graph_gemma4v(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const override;
bool support_batch() const override { return true; }
};
struct mtmd_graph_gemma4uv : mtmd_graph {
mtmd_graph_gemma4uv(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_nemotron_transcribe : mtmd_graph {
const mtmd_subctx_enc_state * state; // nullptr during warmup
mtmd_graph_nemotron_transcribe(mtmd_subcontext * ctx, const mtmd_image_f32 & img, const mtmd_subctx_enc_state * state)
: mtmd_graph(ctx, img), state(state) {}
ggml_cgraph * build() override;
ggml_tensor * build_spk_encoder(ggml_tensor * cur, ggml_tensor * pos);
// one recording: encodes its windows in turn, carrying the speaker state between them
static bool encode_windows(mtmd_subcontext * ctx, const mtmd_model & model, mtmd_subctx_encode_params * params);
// NeMo ParallelExpertEncoder online defaults, in output frames: 40 s windows with 4 s of context on each side
static constexpr int window_core = 500;
static constexpr int window_left = 50;
static constexpr int window_right = 50;
};
struct mtmd_graph_pixtral : mtmd_graph {
mtmd_graph_pixtral(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_deepseek4v : mtmd_graph {
mtmd_graph_deepseek4v(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_qwen2vl : mtmd_graph {
mtmd_graph_qwen2vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_qwen3vl : mtmd_graph_qwen2vl {
mtmd_graph_qwen3vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph_qwen2vl(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_magevl : mtmd_graph {
mtmd_graph_magevl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_sensenova_u1 : mtmd_graph {
mtmd_graph_sensenova_u1(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_lfm3vl : mtmd_graph {
mtmd_graph_lfm3vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_ling3vl : mtmd_graph_qwen3vl {
mtmd_graph_ling3vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph_qwen3vl(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_minimax_m3 : mtmd_graph {
mtmd_graph_minimax_m3(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * apply_rope(ggml_tensor * x, ggml_tensor * pos_h, ggml_tensor * pos_w);
};
struct mtmd_graph_mimovl : mtmd_graph {
mtmd_graph_mimovl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
// Force F32 mat-mul accumulation to avoid F16 overflow in the FFN down-proj
// when the mmproj is stored in F16 (the source weights are BF16; downcasting
// to F16 reduces dynamic range below the SwiGLU output magnitude on the last few layers).
ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const override;
};
struct mtmd_graph_step3vl : mtmd_graph {
mtmd_graph_step3vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_youtuvl : mtmd_graph {
mtmd_graph_youtuvl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_yasa2 : mtmd_graph {
mtmd_graph_yasa2(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * layer_norm_channels(ggml_tensor * inp, ggml_tensor * w, ggml_tensor * b, float eps = 1e-6f);
ggml_tensor * convnext_grn(ggml_tensor * inp, ggml_tensor * w, ggml_tensor * b);
};
struct mtmd_graph_minicpmv : mtmd_graph {
mtmd_graph_minicpmv(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_minicpmv4_6 : mtmd_graph {
mtmd_graph_minicpmv4_6(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_internvl : mtmd_graph {
mtmd_graph_internvl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
bool support_batch() const override { return true; }
};
struct mtmd_graph_nemotron_v2_vl : mtmd_graph {
mtmd_graph_nemotron_v2_vl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_nemotron_omni : mtmd_graph {
mtmd_graph_nemotron_omni(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * build_inp_with_temporal_merge();
};
struct mtmd_graph_llama4 : mtmd_graph {
mtmd_graph_llama4(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_kimivl : mtmd_graph {
mtmd_graph_kimivl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_paddleocr : mtmd_graph {
mtmd_graph_paddleocr(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_dotsocr : mtmd_graph {
mtmd_graph_dotsocr(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_dots3note_a : mtmd_graph {
mtmd_graph_dots3note_a(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_cogvlm : mtmd_graph {
mtmd_graph_cogvlm(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_llava : mtmd_graph {
mtmd_graph_llava(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_whisper_enc : mtmd_graph {
mtmd_graph_whisper_enc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
// causal whisper-like encoder with a sliding window KV cache (voxtral realtime)
struct mtmd_graph_voxtral_rt : mtmd_graph {
mtmd_subctx_enc_state * state; // nullptr during warmup, the cache is then a graph input
mtmd_graph_voxtral_rt(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_enc_state * state) : mtmd_graph(ctx, img), state(state) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_deepseekocr : mtmd_graph {
mtmd_graph_deepseekocr(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * build_sam(ggml_tensor * inp); // build the SAM model
bool support_batch() const override { return true; }
};
struct mtmd_graph_deepseekocr2 : mtmd_graph_deepseekocr {
mtmd_graph_deepseekocr2(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph_deepseekocr(ctx, img) {}
ggml_cgraph * build() override; // reuses build_sam() from base
bool support_batch() const override { return true; }
};
struct mtmd_graph_conformer : mtmd_graph {
mtmd_graph_conformer(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
// builders shared by the chatterbox gen and spkenc graphs, defined in chatterbox-gen.cpp
struct mtmd_graph_chatterbox_base : mtmd_graph {
using mtmd_graph::mtmd_graph;
protected:
ggml_tensor * cbx_linear(ggml_tensor * w, ggml_tensor * b, ggml_tensor * x) const;
ggml_tensor * cbx_conv1d(ggml_tensor * k, ggml_tensor * b, ggml_tensor * x,
int stride, int pad_l, int pad_r) const;
ggml_tensor * cbx_conv1d_dil(ggml_tensor * k, ggml_tensor * b, ggml_tensor * x,
int pad, int dil) const;
};
struct mtmd_graph_chatterbox_spkenc : mtmd_graph_chatterbox_base {
mtmd_graph_chatterbox_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph_chatterbox_base(ctx, img) {}
ggml_cgraph * build() override;
private:
ggml_tensor * bn1d(const mtmd_chatterbox::bn & n, ggml_tensor * x);
ggml_tensor * bn2d_relu(const mtmd_chatterbox::bn & n, ggml_tensor * x);
ggml_tensor * res2d(const mtmd_chatterbox::spk_res2d & r, ggml_tensor * x, int stride);
ggml_tensor * cam_layer(const mtmd_chatterbox::spk_cam_layer & l, ggml_tensor * x, int dil, ggml_tensor * segfix);
};
struct mtmd_graph_chatterbox : mtmd_graph_chatterbox_base {
mtmd_subctx_gen_process gen_process = MTMD_SUBCTX_GEN_PROCESS_GEN_CODE;
int n_tokens = 0;
int n_prompt_mel = 0;
int vocode_n_mel = 0;
int vocode_n_stft = 0;
mtmd_graph_chatterbox(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, int n_tokens,
int n_prompt_mel, int vocode_n_mel, int vocode_n_stft)
: mtmd_graph_chatterbox_base(ctx, img), gen_process(gen_process), n_tokens(n_tokens), n_prompt_mel(n_prompt_mel),
vocode_n_mel(vocode_n_mel), vocode_n_stft(vocode_n_stft) {}
ggml_cgraph * build() override;
private:
ggml_tensor * enc_layer(const mtmd_chatterbox::enc_layer & l, ggml_tensor * x, ggml_tensor * pos, int T);
ggml_tensor * causal_block(const mtmd_chatterbox::causal_block & b, ggml_tensor * x);
ggml_tensor * resnet(const mtmd_chatterbox::resnet & r, ggml_tensor * x, ggml_tensor * temb);
ggml_tensor * tfm_block(const mtmd_chatterbox::tfm_block & b, ggml_tensor * x);
ggml_tensor * estimator(ggml_tensor * x_noise, ggml_tensor * mu, ggml_tensor * spks,
ggml_tensor * cond, ggml_tensor * temb, int T);
ggml_tensor * hift_resblock(const mtmd_chatterbox::hift_res & r, ggml_tensor * x);
ggml_cgraph * build_s3tok(int T);
ggml_cgraph * build_vocoder(int n_mel, int n_stft);
};
struct mtmd_graph_kokoro : mtmd_graph {
enum stage_type {
STAGE_TEXT, // tokens -> text encoder CNN
STAGE_F0N, // shared LSTM output -> F0 and energy curves
STAGE_DEC, // aligned features + source STFT -> STFT coefficients
};
static constexpr int n_style = 128; // a voice row is the decoder style, then the predictor style
stage_type stage;
int n_tok = 0;
int n_frames = 0;
int n_har = 0;
mtmd_graph_kokoro(mtmd_subcontext * ctx, const mtmd_image_f32 & img, stage_type stage, int n_tok, int n_frames, int n_har)
: mtmd_graph(ctx, img), stage(stage), n_tok(n_tok), n_frames(n_frames), n_har(n_har) {}
ggml_cgraph * build() override;
// host side, between the graphs
// bert [n_tok, n_embd], text_cnn [n_tok, 512], style is the predictor style
// fills shared and asr as [n_frames, 512], returns n_frames
static int prosody(const mtmd_model & model, const float * bert, int n_embd, const float * text_cnn, int n_tok,
const float * style, float speed, std::vector<float> & shared, std::vector<float> & asr);
// fills har as [n_har, 22], returns n_har
static int har_source(const mtmd_model & model, const std::vector<float> & f0, std::mt19937 & rng, std::vector<float> & har);
static void istft(const std::vector<float> & spec, std::vector<float> & audio);
private:
ggml_tensor * tensor(const std::string & name) const;
bool has_tensor(const std::string & name) const;
ggml_tensor * conv(const std::string & name, ggml_tensor * x, int stride, int pad, int dil) const;
ggml_tensor * adain(const std::string & name, ggml_tensor * x, ggml_tensor * style) const;
ggml_tensor * snake(const std::string & name, ggml_tensor * x) const;
ggml_tensor * interleave(ggml_tensor * a, ggml_tensor * b) const;
ggml_tensor * adain_resblk(const std::string & name, ggml_tensor * x, ggml_tensor * style, bool upsample) const;
ggml_tensor * gen_resblock(const std::string & name, ggml_tensor * x, ggml_tensor * style) const;
ggml_tensor * input(const char * name, ggml_type type, int64_t ne0, int64_t ne1) const;
ggml_cgraph * output(const char * name, ggml_tensor * t) const;
ggml_cgraph * build_text();
ggml_cgraph * build_f0n();
ggml_cgraph * build_dec();
};
struct mtmd_graph_lfm3a : mtmd_graph {
mtmd_graph_lfm3a(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_granite_speech : mtmd_graph {
mtmd_graph_granite_speech(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_vibevoice_asr : mtmd_graph {
mtmd_graph_vibevoice_asr(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
std::map<std::string, ggml_tensor *> state_in;
mutable std::vector<std::pair<std::string, ggml_tensor *>> state_out;
ggml_tensor * with_left_ctx(ggml_tensor * x, const std::string & name) const;
ggml_tensor * causal_conv(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, const std::string & name) const;
ggml_tensor * encode(int e, ggml_tensor * x) const;
};
struct mtmd_graph_muse_spark_audio : mtmd_graph {
mtmd_graph_muse_spark_audio(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_gemma4a : mtmd_graph {
// attention_chunk_size, attention_context_left - 1
static constexpr int64_t chunk_size = 12;
static constexpr int64_t max_past = 12;
mtmd_graph_gemma4a(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const override;
};
struct mtmd_graph_gemma4ua : mtmd_graph {
mtmd_graph_gemma4ua(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_glm4v : mtmd_graph {
mtmd_graph_glm4v(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_hunyuanvl : mtmd_graph {
mtmd_graph_hunyuanvl(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_mobilenetv5 : mtmd_graph {
mtmd_graph_mobilenetv5(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * rms_norm_2d(
ggml_tensor * inp,
ggml_tensor * weight,
float eps = 1e-6f);
ggml_tensor* pad_same_2d(
ggml_tensor* inp,
int kernel_h,
int kernel_w,
int stride_h,
int stride_w,
int dilation_h = 1,
int dilation_w = 1);
ggml_tensor * build_edge_residual(
ggml_tensor * inp,
const mobilenetv5_block & block,
int stride);
ggml_tensor * build_inverted_residual(
ggml_tensor * inp,
const mobilenetv5_block & block,
int stride);
ggml_tensor * build_mobilenet_attn(
ggml_tensor * inp,
const mobilenetv5_block & block);
};
struct mtmd_graph_qwen3a : mtmd_graph {
mtmd_graph_qwen3a(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_mimo_audio : mtmd_graph {
mtmd_graph_mimo_audio(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
// zonos3 speaker encoder: MECT-B2 + ASTP, then the speaker projection of the backbone
struct mtmd_graph_zonos3_spkenc : mtmd_graph {
mtmd_graph_zonos3_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * W(const std::string & name) const;
bool has(const std::string & name) const;
ggml_tensor * chan(ggml_tensor * v) const;
ggml_tensor * bn(ggml_tensor * x, const std::string & p) const;
ggml_tensor * conv(ggml_tensor * x, const std::string & p, int s_t, int s_f, bool bias, int pad) const;
ggml_tensor * grouped3x3(ggml_tensor * x, const std::string & p, int s_t, int s_f) const;
ggml_tensor * res_block(ggml_tensor * x, const std::string & q, int s_f, int s_t) const;
ggml_tensor * linear(ggml_tensor * x, const std::string & p) const;
ggml_tensor * layer_norm(ggml_tensor * x, const std::string & p) const;
ggml_tensor * flatten(ggml_tensor * x) const;
ggml_tensor * unflatten(ggml_tensor * x, int64_t F, int64_t C) const;
ggml_tensor * attention(ggml_tensor * h, const std::string & p) const;
ggml_tensor * moe_ffn(ggml_tensor * h, const std::string & p) const;
ggml_tensor * pos_conv(ggml_tensor * h, const std::string & p) const;
};
struct mtmd_graph_qwen3tts_spkenc : mtmd_graph {
mtmd_graph_qwen3tts_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * conv1d_same(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation) const;
ggml_tensor * res2net(ggml_tensor * x, const mtmd_layer & layer, int dilation, int scale) const;
ggml_tensor * se_block(ggml_tensor * x, const mtmd_layer & layer) const;
ggml_tensor * se_res2net_block(ggml_tensor * x, const mtmd_layer & layer, int dilation, int scale) const;
ggml_tensor * attentive_stats_pool(ggml_tensor * x) const;
};
struct mtmd_graph_qwen3tts_gen : mtmd_graph {
mtmd_graph_qwen3tts_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, int top_k, float top_p)
: mtmd_graph(ctx, img), gen_process(gen_process), top_k(top_k), top_p(top_p) {}
ggml_cgraph * build() override;
// MTMD_SUBCTX_GEN_PROCESS_TOKENIZE: reference clip -> codes and their talker input rows
ggml_cgraph * build_ref_encoder();
static ggml_tensor * nearest_code(ggml_context * ctx0, ggml_tensor * x, ggml_tensor * codebook);
// split RVQ (mimi): the first codebook on x_first, the residual ones on x_rest, cb_rest is stacked on dim 2
// x_first, x_rest: [dim, T] -> one [T] code tensor per codebook; also used by the moshi encoder
static std::vector<ggml_tensor *> rvq_encode(ggml_context * ctx0, ggml_tensor * x_first, ggml_tensor * cb_first,
ggml_tensor * x_rest, ggml_tensor * cb_rest);
// reference encoder transformer, see encoder_config in speech_tokenizer/config.json
static constexpr int ref_enc_n_head = 8;
static constexpr float ref_enc_eps = 1e-5f;
static constexpr float ref_enc_rope_theta = 10000.0f;
// which sub-graph build() constructs, fixed at graph-build time
mtmd_subctx_gen_process gen_process;
// sampling params, fixed at graph-build time (GEN_CODE only)
int top_k;
float top_p;
//
// code_gen: backbone hidden state + sampled code0 -> 16 RVQ codes
// MTP-style code predictor, one token per codebook
//
struct code_gen : mtmd_graph {
code_gen(const mtmd_graph & parent, int top_k, float top_p)
: mtmd_graph(parent), top_k(top_k), top_p(top_p) {}
ggml_cgraph * build() override { GGML_ABORT("call prefill()/step() instead"); }
int top_k;
float top_p;
ggml_tensor * cache_set(ggml_tensor * cache, int row_idx, ggml_tensor * value) const;
ggml_tensor * do_sampling(ggml_tensor * logits, ggml_tensor * inp_rand) const;
ggml_tensor * const_i32(ggml_tensor * anchor, float value) const;
ggml_tensor * causal_mask_row(int64_t n_kv_pad, int pos) const;
ggml_tensor * project_in(ggml_tensor * cur) const;
ggml_tensor * layer_forward(
ggml_tensor * cur,
const mtmd_layer & layer,
ggml_tensor * inp_pos,
ggml_tensor * kq_mask,
ggml_tensor *& k_cache_layer,
ggml_tensor *& v_cache_layer,
int64_t n_kv_pad,
int pos,
int il) const;
void prefill(
std::vector<ggml_tensor *> & k_cache,
std::vector<ggml_tensor *> & v_cache,
ggml_tensor *& out_code_cache,
ggml_tensor * h_state,
ggml_tensor * code0_embd,
ggml_tensor * inp_rand) const;
ggml_tensor * step(
std::vector<ggml_tensor *> & k_cache,
std::vector<ggml_tensor *> & v_cache,
ggml_tensor * out_code_cache,
ggml_tensor * inp_rand,
int step_idx) const;
};
//
// code2wav: RVQ codes -> raw PCM (quantizer + pre_conv + pre_transformer + upsample + DAC).
//
struct code2wav : mtmd_graph {
code2wav(const mtmd_graph & parent) : mtmd_graph(parent) {}
ggml_cgraph * build() override { GGML_ABORT("call decode() instead"); }
// state_in: previous call's persisted state, by slot name (see list_c2w_state_slots())
std::map<std::string, ggml_tensor *> state_in;
// state_out: this call's state to persist, added to the graph outputs by build()
mutable std::vector<std::pair<std::string, ggml_tensor *>> state_out;
// stateful conv ops: read/update their state via state_in/state_out[state_name]
ggml_tensor * causal_conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation, const std::string & state_name) const;
ggml_tensor * causal_conv1d_dw(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, const std::string & state_name, int dilation = 1) const;
ggml_tensor * causal_conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, const std::string & state_name) const;
ggml_tensor * snake(ggml_tensor * x, ggml_tensor * alpha, ggml_tensor * beta) const;
ggml_tensor * quant_decode(ggml_tensor * inp_codes) const;
// window <= 0 uses hparams.wav_tfm_swa
ggml_tensor * tfm_layer_forward(ggml_tensor * cur, const mtmd_layer & layer, int il, int64_t window = 0) const;
ggml_tensor * convnext_block(ggml_tensor * x, const mtmd_code2wav::upsample_block & blk, const std::string & state_prefix) const;
ggml_tensor * dac_res_unit(ggml_tensor * x, const mtmd_code2wav::dac_res & res, int dilation, const std::string & state_name) const;
// drama3: erf GELU in ConvNeXt, not the tanh approximation
bool gelu_erf = false;
// inp_codes [1, n_codes] I32 -> this frame's audio samples [n_samples] F32, clamped to [-1, 1]
ggml_tensor * decode(ggml_tensor * inp_codes) const;
};
};
//
// pocket-tts: SEANet convolution stack, shared by the voice encoder and the mimi decoder.
// stateless unless state_in is populated: convs then pad instead of carrying left-context.
//
struct c2w_state_slot;
struct mtmd_graph_pockettts_seanet : mtmd_graph {
mtmd_graph_pockettts_seanet(const mtmd_graph & parent) : mtmd_graph(parent) {}
ggml_cgraph * build() override { GGML_ABORT("call encode()/decode() instead"); }
// qwen3tts reference codes: the output is quantized, so the im2col is F32 (the kernels must be F32 too)
bool f32_im2col = false;
// qwen3tts reference codes: pad_replicate also repeats the last frame on the right, like MimiConv1d
bool pad_replicate_end = false;
// per-call streaming state, keyed by slot name (see list_pockettts_state_slots)
std::map<std::string, ggml_tensor *> state_in;
mutable std::vector<std::pair<std::string, ggml_tensor *>> state_out;
// the "state_in_<slot>" inputs of the slots, then the "state_out_<slot>" outputs written by the graph
void state_inputs(const std::vector<c2w_state_slot> & slots);
void state_outputs(ggml_cgraph * graph) const;
ggml_tensor * conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, int dilation,
bool pad_replicate = false, const std::string & state_name = "") const;
ggml_tensor * conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride,
const std::string & state_name = "") const;
ggml_tensor * res_unit(ggml_tensor * x, const mtmd_seanet::stage & stage, int dilation,
const std::string & state_prefix = "") const;
// x: [T, C] -> [T / hop, dim], streams when state_in is populated
ggml_tensor * encode(ggml_tensor * x) const;
// x: [T, dim] -> [T * hop, 1], streams when state_in is populated
ggml_tensor * decode(ggml_tensor * x) const;
// mimi transformer, always streams (see list_pockettts_state_slots): x: [dim, n_pos] -> [dim, n_pos]
ggml_tensor * transformer(ggml_tensor * x, const std::vector<mtmd_layer> & layers, int n_head, float eps) const;
// mimi decoder: upsample, transformer, SEANet decoder; x: [dim, n_frames] -> [n_samples, 1]
ggml_tensor * mimi_decode(ggml_tensor * x, int n_head, float eps) const;
};
// mimi encoder + speaker_proj: reference waveform -> voice conditioning rows
// the qwen3tts reference encoder reuses encode() with its own transformer options
struct mtmd_graph_pockettts_spkenc : mtmd_graph {
mtmd_graph_pockettts_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
mtmd_graph_pockettts_spkenc(const mtmd_graph & parent) : mtmd_graph(parent) {}
ggml_cgraph * build() override;
int tfm_n_head = n_head;
float tfm_eps = eps;
float tfm_rope_theta = hparams.rope_theta;
int tfm_rope_type = GGML_ROPE_TYPE_NORMAL;
ffn_op_type tfm_ffn_type = FFN_GELU;
ggml_tensor * tfm_layer_forward(ggml_tensor * cur, const mtmd_layer & layer, ggml_tensor * inp_pos, ggml_tensor * kq_mask, int il) const;
// waveform [n_samples, 1] -> [n_frames, C], at the frame rate after model.downsample_w
ggml_tensor * encode(const mtmd_graph_pockettts_seanet & seanet, ggml_tensor * x, const std::vector<mtmd_layer> & layers);
};
//
// pocket-tts generation:
// GEN_CODE = flow-matching decoder + end-of-speech head, one latent per call
// GEN_WAV = mimi decoder, a window of latents -> PCM
//
struct mtmd_graph_pockettts_gen : mtmd_graph {
mtmd_graph_pockettts_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, int n_step, int n_frames)
: mtmd_graph(ctx, img), gen_process(gen_process), n_step(n_step), n_frames(n_frames) {}
ggml_cgraph * build() override;
mtmd_subctx_gen_process gen_process;
int n_step; // lsd_decode steps, fixed at graph-build time
int n_frames; // GEN_WAV only: number of latents to decode
// AdaLN modulation: x * (1 + scale) + shift
ggml_tensor * modulate(ggml_tensor * x, ggml_tensor * shift, ggml_tensor * scale) const;
ggml_tensor * time_embed(const mtmd_flow_net::time_embd & te, float t) const;
ggml_tensor * flow_forward(ggml_tensor * cond, ggml_tensor * x, float s, float t) const;
};
//
// moshi family (e.g. Hibiki): streaming mimi encoder + RVQ, rows of the input audio codes
// output: [n_embd, n_q * (n_frames + 1)], the embedding of each code of each frame, then of the initial codes
//
struct mtmd_graph_moshi : mtmd_graph {
mtmd_graph_moshi(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
static constexpr int n_samples_per_frame = 1920; // 80 ms at 24 kHz
};
//
// moshi generation:
// GEN_CODE = depformer, the codes of one frame + their backbone embedding; with all codes known, only the embedding
// GEN_WAV = RVQ decode + mimi decoder
//
struct mtmd_graph_moshi_gen : mtmd_graph {
mtmd_graph_moshi_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, bool fb_only,
int n_frames, int top_k, float temp)
: mtmd_graph(ctx, img), gen_process(gen_process), fb_only(fb_only), n_frames(n_frames), top_k(top_k), temp(temp) {}
ggml_cgraph * build() override;
static constexpr int mimi_n_layer = 8;
static constexpr int mimi_n_head = 8;
static constexpr float mimi_eps = 1e-5f;
static constexpr int n_q_max = 32;
mtmd_subctx_gen_process gen_process;
bool fb_only;
int n_frames;
int top_k;
float temp;
ggml_tensor * dep_step(ggml_tensor * x, int64_t iw, std::vector<ggml_tensor *> & k_all, std::vector<ggml_tensor *> & v_all);
ggml_tensor * feedback(const std::vector<ggml_tensor *> & codes);
};
// one persisted state buffer used by code2wav, see qwen3tts-gen.cpp
struct c2w_state_slot {
std::string name;
int64_t ne0;
int64_t ne1;
bool in_place = false; // graph writes the slot directly, no state_in/state_out copy
};
std::vector<c2w_state_slot> list_c2w_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// drama3: tower KV slots, then list_c2w_state_slots()
std::vector<c2w_state_slot> list_drama3_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// same, for the streaming mimi decoder (pocket-tts GEN_WAV)
std::vector<c2w_state_slot> list_pockettts_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// streaming mimi encoder of moshi, "pos" leads
std::vector<c2w_state_slot> list_moshi_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
//
// drama3 generation:
// GEN_CODE = one codebook tower row, each layer attends over the whole sequence (KV in gen state)
// GEN_WAV = Fish DAC decoder, frames -> PCM
//
struct mtmd_graph_drama3_gen : mtmd_graph {
mtmd_graph_drama3_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, mtmd_subctx_gen_state * gen_state,
int n_frames, int top_k, float top_p, float temp)
: mtmd_graph(ctx, img), gen_process(gen_process), gen_state(gen_state), n_frames(n_frames), top_k(top_k), top_p(top_p), temp(temp) {}
ggml_cgraph * build() override;
// tower KV rows, shared by the prompt and the generated frames
static constexpr int n_pos_max = 4096;
mtmd_subctx_gen_process gen_process;
mtmd_subctx_gen_state * gen_state; // nullptr on buffer reserve
int n_frames; // GEN_WAV: frames to decode; GEN_CODE: 2 when the caller gives the codes of the frame (teacher forcing)
int top_k;
float top_p;
float temp;
ggml_cgraph * build_tower();
ggml_cgraph * build_codec();
ggml_tensor * tower_layer(ggml_tensor * x, int il, ggml_tensor * code, ggml_tensor * mask, ggml_tensor * extra,
ggml_tensor * inp_pos, ggml_tensor * k_slot, ggml_tensor * v_slot, int pos);
};
//
// fish_s2 generation:
// GEN_CODE = fast AR, backbone hidden + semantic code -> other codes of the frame + next backbone input
// GEN_WAV = Fish DAC decoder, same as drama3
//
struct mtmd_graph_fish_s2_gen : mtmd_graph_drama3_gen {
using mtmd_graph_drama3_gen::mtmd_graph_drama3_gen;
ggml_cgraph * build() override;
ggml_cgraph * build_fast_ar();
};
// fish_s2 voice cloning: raw 44.1 kHz clip -> codec encoder + RVQ -> one backbone input row per frame
struct mtmd_graph_fish_s2_spkenc : mtmd_graph {
mtmd_graph_fish_s2_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
// codec encoder transformer layers, the rest of model.layers is the quantizer pre_module
static constexpr int n_enc_tfm_layer = 4;
static constexpr int n_tfm_layer = 12;
ggml_tensor * conv(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, int dilation) const;
ggml_tensor * snake(ggml_tensor * x, ggml_tensor * alpha, ggml_tensor * beta) const;
ggml_tensor * convnext(ggml_tensor * x, const mtmd_code2wav::upsample_block & blk) const;
};
//
// zonos3 generation:
// GEN_CODE = 10 codebook heads, or the backbone row of a sampled frame
// GEN_WAV = ZAC codec decoder up to the iSTFT spectra
//
struct mtmd_graph_zonos3_gen : mtmd_graph {
mtmd_graph_zonos3_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, bool all_known, int n_frames)
: mtmd_graph(ctx, img), gen_process(gen_process), all_known(all_known), n_frames(n_frames) {}
ggml_cgraph * build() override;
static constexpr int n_codebooks = 10;
static constexpr int n_cb_vocab = 4096;
static constexpr float softcap = 15.0f;
// ZAC decoder, fixed by the codec config
static constexpr int n_stages = 5;
static constexpr int stage_ctx[n_stages] = { 150, 300, 480, 480, 480 }; // attention window
static constexpr int head_dim = 64;
static constexpr float ln_eps = 1e-5f;
static constexpr float convnext_eps = 1e-6f;
static constexpr int hop = 184; // at 44160 Hz
static constexpr int n_fft[2] = { 736, 2208 };
static constexpr int rs_orig = 736; // 44160 -> 44100 Hz
static constexpr int rs_new = 735;
static constexpr int samples_per_frame = 2940;
mtmd_subctx_gen_process gen_process;
bool all_known; // GEN_CODE: only the backbone row
int n_frames; // GEN_WAV
std::map<std::string, ggml_tensor *> state_in;
std::vector<std::pair<std::string, ggml_tensor *>> state_out;
ggml_cgraph * build_gen_code();
ggml_cgraph * build_gen_wav();
ggml_tensor * convnext(ggml_tensor * x, const mtmd_zonos3_wav::convnext & b, const std::string & state_name);
ggml_tensor * moss_layer(ggml_tensor * x, const mtmd_zonos3_wav::layer & l, int s, int il, ggml_tensor * base);
};
std::vector<c2w_state_slot> list_zonos3_gen_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// graphs that decode with a mtmd_snake_vae (yue2, minimax_music3), see snake-vae.cpp
struct mtmd_graph_snake_vae : mtmd_graph {
using mtmd_graph::mtmd_graph;
// x: [T, C]
ggml_tensor * snake(ggml_tensor * x, const mtmd_snake_vae::snake & s) const;
ggml_tensor * conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation) const;
// x: [T, n_latent] -> [T * ratio, n_out]
ggml_tensor * build_snake_vae(ggml_tensor * x, const mtmd_snake_vae & vae) const;
};
//
// YuE2 generation, the NAR transformer itself runs on a llama_context, see mtmd_gen_flow_yue2():
// LATENT_IN = latent rows -> backbone embd rows
// LATENT_OUT = backbone hidden rows -> velocity rows
// GEN_WAV = Oobleck VAE decoder, latent rows -> interleaved stereo PCM
//
struct mtmd_graph_yue2_gen : mtmd_graph_snake_vae {
mtmd_graph_yue2_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, int n_rows)
: mtmd_graph_snake_vae(ctx, img), gen_process(gen_process), n_rows(n_rows) {}
ggml_cgraph * build() override;
static constexpr int n_latent = 64;
static constexpr int n_freq = 256; // time embedder input
static constexpr int n_channels = 2;
static constexpr int vae_ratio = 1920; // output samples per latent frame
static constexpr int vae_trim = 64; // a decode of n frames has vae_ratio * n - vae_trim samples
static constexpr float rope_theta = 1000000.0f;
static constexpr int n_ctx_max = 24576; // the AR prefix and the NAR rows of one chunk
static constexpr int ode_steps = 32; // midpoint solver
mtmd_subctx_gen_process gen_process;
int n_rows;
};
//
// MiniMax Music 3 generation, see mtmd_gen_wav_music3() for the flow-matching loop:
// GEN_CODE = depth decoder of one frame: backbone hidden rows (cond, then uncond with guidance) + semantic code -> the 8 codes, the next backbone input and the condition row (mix of the 8 hidden states of the frame)
// FLOW_COND = condition rows of a window -> DiT condition, resampled to the latent rate
// FLOW_STEP = one DiT pass, cond and uncond rows as a batch of 2
// GEN_WAV = DAC decoder, latent rows -> interleaved stereo PCM
//
struct mtmd_graph_music3_gen : mtmd_graph_snake_vae {
mtmd_graph_music3_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process,
int n_rows, int n_out, float cfg_scale, float temp)
: mtmd_graph_snake_vae(ctx, img), gen_process(gen_process), n_rows(n_rows), n_out(n_out), cfg_scale(cfg_scale), temp(temp) {}
ggml_cgraph * build() override;
static constexpr int n_latent = 128; // two channels of 64
static constexpr int n_channels = 2;
static constexpr int vae_ratio = 512; // output samples per latent frame
static constexpr int dit_d_head = 64;
static constexpr int dit_n_rot = 32; // partial rotary
static constexpr float dit_theta = 10000.0f;
static constexpr float dit_eps = 1e-5f;
static constexpr int top_k = 50;
// sampling and flow matching of the reference (encoders.py, denoise.py)
static constexpr float depth_cfg = 1.5f;
static constexpr float dit_cfg = 1.7f;
static constexpr int flow_steps = 30;
static constexpr int n_overlap = 172; // latent frames shared by neighboring windows
mtmd_subctx_gen_process gen_process;
int n_rows; // GEN_CODE: backbone rows, FLOW_COND: frames, FLOW_STEP and GEN_WAV: latent frames
int n_out; // FLOW_COND: latent frames
float cfg_scale;
float temp;
static int64_t latent_frames(int64_t n_frames);
ggml_cgraph * build_depth();
ggml_cgraph * build_cond();
ggml_cgraph * build_dit();
ggml_cgraph * build_vocoder();
};
//
// VoxCPM2 generation, the residual LM runs on a llama_context of the gen state, see mtmd_gen_code_voxcpm2():
// RES_INP = backbone row -> FSQ'd hidden state, residual LM input row, stop score
// GEN_CODE = LocDiT flow matching with CFG-Zero* -> next patch, then LocEnc -> next backbone input row
// LATENT_IN = LocEnc, latent patches -> backbone input rows (voice reference)
// GEN_WAV = causal AudioVAE decoder, latent patches -> PCM
//
struct mtmd_graph_voxcpm2_gen : mtmd_graph {
mtmd_graph_voxcpm2_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process,
int n_rows, bool audio_row, int n_steps, float cfg_scale)
: mtmd_graph(ctx, img), gen_process(gen_process), n_rows(n_rows), audio_row(audio_row), n_steps(n_steps), cfg_scale(cfg_scale) {}
ggml_cgraph * build() override;
static constexpr int n_latent = 64; // feat_dim
static constexpr int n_patch = 4; // latents per backbone row
static constexpr int n_dit_tok = 3 + 2 * n_patch; // lm, res, time, cond patch, noisy patch
static constexpr int head_dim = 128; // kv_channels
static constexpr float rope_theta = 10000.0f;
static constexpr int rope_n_ctx_orig = 32768;
static constexpr float fsq_scale = 9.0f;
static constexpr int n_steps_default = 10; // inference_timesteps
static constexpr int n_steps_max = 32;
static constexpr float cfg_default = 2.0f; // cfg_value
static constexpr int n_ctx_res = 8192; // residual LM positions of one sequence, max_length of the reference
static constexpr int dec_ratio = 1920; // decoder samples per latent
static constexpr int res_dilations[3] = { 1, 3, 9 };
// t of the Euler steps with sway sampling, n_steps + 1 values from 1 to 0
static std::vector<float> t_span(int n_steps);
// SinusoidalPosEmb(t * 1000) of the steps, [n_embd, n_steps]
static std::vector<float> time_embd(int n_steps, int n_embd);
mtmd_subctx_gen_process gen_process;
int n_rows;
bool audio_row;
int n_steps;
float cfg_scale;
// bidirectional MiniCPM layers, x: [n_embd, n_tok, n_seq]
ggml_tensor * tfm(ggml_tensor * x, const std::vector<mtmd_layer> & layers, ggml_tensor * norm_w, ggml_tensor * pos) const;
// [n_latent, n_patch * n] -> [n_text_embd, n]
ggml_tensor * loc_enc(ggml_tensor * latents, int n) const;
ggml_tensor * inp_pos(const char * name, int n) const;
void build_res_inp();
void build_flow();
void build_latent_in();
void build_dec();
};
std::vector<c2w_state_slot> list_voxcpm2_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// AudioVAE encoder: raw waveform -> latent patches, LocEnc then runs on the gen model (LATENT_IN)
struct mtmd_graph_voxcpm2_spkenc : mtmd_graph {
mtmd_graph_voxcpm2_spkenc(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
// x: [T, IC], causal: pad zeros on the left
ggml_tensor * conv(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, int dilation, int pad, bool depthwise) const;
};
// tiny-voca: small graphs around the stacks, the stacks run on libllama (see mtmd_gen_wav_tiny_voca())
// LATENT_IN: token states [n_embd, n_rows] -> duration outputs [2, n_rows]
// LATENT_OUT: acoustic states [n_embd, n_rows] -> source-filter params [n_voc, n_rows]
struct mtmd_graph_tiny_voca : mtmd_graph {
mtmd_graph_tiny_voca(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, int n_rows)
: mtmd_graph(ctx, img), gen_process(gen_process), n_rows(n_rows) {}
ggml_cgraph * build() override;
static constexpr int n_prosody_layer = 2;
static constexpr int n_acoustic_layer = 5;
static constexpr int n_max_frames = 3000;
static constexpr int hop = 256; // samples per frame
static constexpr int n_cep_causal = 80;
static constexpr int n_cep_anti = 16;
static constexpr int n_noise_cep = 32;
static constexpr int harm_fft = 512;
static constexpr int harm_hop = 128;
static constexpr int harm_steps = hop / harm_hop;
static constexpr int noise_fft = 256;
static constexpr int noise_hop = 64;
static constexpr int noise_steps = hop / noise_hop;
static constexpr int n_voc = n_cep_causal + n_cep_anti + harm_steps + n_noise_cep + noise_steps;
static constexpr float rope_theta = 10000.0f;
static constexpr float sample_rate = 24000.0f;
static constexpr float harmonic_limit = 0.47f; // highest harmonic, relative to the sample rate
static constexpr float f0_ref = 200.0f; // Hz, log F0 is predicted relative to it
static constexpr float f0_min = 50.0f;
static constexpr float f0_max = 600.0f;
static constexpr float max_dur = 400.0f; // frames per token before the speed
mtmd_subctx_gen_process gen_process;
int n_rows;
// host: frame params [n_frames, n_voc] and F0 [n_frames] in Hz -> waveform [n_frames * hop]
static void synth(const std::vector<float> & voc, const std::vector<float> & f0, std::mt19937 & rng, std::vector<float> & audio);
// unknown-word model: letters (a-z and ') -> CMUdict phonemes separated by spaces, other characters are skipped
static constexpr int n_g2p_slot = 2; // positions per letter
static constexpr int n_g2p_letter = 27;
static constexpr int n_g2p_symbol = 69;
static std::string g2p(const mtmd_tiny_voca & tv, const std::string & word);
private:
ggml_tensor * conv3(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b) const;
};
struct mtmd_graph_soprano : mtmd_graph {
mtmd_graph_soprano(mtmd_subcontext * ctx, const mtmd_image_f32 & img, int n_frames) : mtmd_graph(ctx, img), n_frames(n_frames) {}
ggml_cgraph * build() override;
static void decode_spectrum(const mtmd_model & model, const std::vector<float> & spectrum, std::vector<float> & pcm);
// SopranoDecoder default
static constexpr int upscale = 4;
int n_frames;
};
//
// LFM3-Audio generation:
// GEN_CODE = one MaskGIT pass of the depthformer, or only the feedback embedding once all codes are known
// GEN_WAV = liquid-codec decoder, a window of frames -> polar STFT frames
//
struct mtmd_graph_lfm3a_gen : mtmd_graph {
mtmd_graph_lfm3a_gen(mtmd_subcontext * ctx, const mtmd_image_f32 & img, mtmd_subctx_gen_process gen_process, bool all_known, int n_frames)
: mtmd_graph(ctx, img), gen_process(gen_process), all_known(all_known), n_frames(n_frames) {}
ggml_cgraph * build() override;
mtmd_subctx_gen_process gen_process;
bool all_known; // GEN_CODE only
int n_frames; // GEN_WAV only
ggml_tensor * build_gen_code();
ggml_tensor * build_gen_wav();
ggml_tensor * build_lfm_layer(ggml_tensor * inpL, const mtmd_layer & layer, float eps, int il,
const std::function<ggml_tensor *(ggml_tensor *)> & op);
ggml_tensor * build_lfm_attn(ggml_tensor * cur, const mtmd_layer & layer, int n_head, int n_head_kv,
float eps, float rope_theta, int il, ggml_tensor * inp_pos,
ggml_tensor * kq_mask, ggml_tensor * k_cache, ggml_tensor * v_cache,
ggml_tensor ** k_all, ggml_tensor ** v_all);
ggml_tensor * build_lfm_conv(ggml_tensor * cur, const mtmd_layer & layer, int il,
ggml_tensor * conv_state, ggml_tensor ** conv_state_out);
};
// state slots of the streaming codec decoder (lfm3a_gen GEN_WAV)
std::vector<c2w_state_slot> list_lfm3a_gen_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// state slots of the streaming LFM3-Audio encoder (lfm3a)
std::vector<c2w_state_slot> list_lfm3a_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// state slots of the streaming Gemma 4 audio encoder (gemma4a)
std::vector<c2w_state_slot> list_gemma4a_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
// vibevoice-asr: the left context of every causal conv, a zero state is the causal zero padding
std::vector<c2w_state_slot> list_vibevoice_asr_state_slots(const mtmd_hparams & hparams, const mtmd_model & model);
struct mtmd_graph_kimik25 : mtmd_graph {
mtmd_graph_kimik25(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
};
struct mtmd_graph_parakeet : mtmd_graph {
mtmd_graph_parakeet(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_exaone4_5 : mtmd_graph {
mtmd_graph_exaone4_5(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_granite4_vision : mtmd_graph {
mtmd_graph_granite4_vision(mtmd_subcontext * ctx, const mtmd_image_f32 & img)
: mtmd_graph(ctx, img),
anyres(img.anyres),
n_tiles(img.ny() / img.nx()),
tile_side(img.nx() / patch_size) {}
ggml_cgraph * build() override;
private:
// the input image is a stack of tiles on the Y axis: [overview, tile(0,0), tile(0,1), ...]
const mtmd_image_f32::anyres_info anyres;
const int n_tiles;
const int tile_side; // patches per tile side
ggml_tensor * build_tile_inp();
ggml_tensor * gather(ggml_tensor * src, const std::string & name, int idx_len);
ggml_tensor * interp_down(ggml_tensor * src, int side, int new_side);
ggml_tensor * build_block(const qf_block & blk, ggml_tensor * h, int bid,
int spatial_offset, int image_side, int window_side,
int query_side, float qformer_eps);
ggml_tensor * build_newline_row(ggml_context * ctx0);
ggml_tensor * build_anyres_assembly(ggml_tensor * cur, int out_side);
};
struct mtmd_graph_muse_glimmer : mtmd_graph {
mtmd_graph_muse_glimmer(mtmd_subcontext * ctx, const mtmd_image_f32 & img) : mtmd_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct mtmd_graph_inkling2v : mtmd_graph {
using mtmd_graph::mtmd_graph;
ggml_cgraph * build() override;
};