Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

mtmd: support MiMo-V2.5 audio input (RVQ-based model) (#26190)

Browse Source 
* gguf converter for mimo audio

* fix conv

* cpp impl

* nits

* nits 2
This commit is contained in:
Xuan-Son Nguyen
2026-07-27 23:17:09 +02:00
committed by GitHub
parent 0e4a036223
commit 1cbfd19883
15 changed files with 770 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/CMakeLists.txt
+1
View File
@@ -51,6 +51,7 @@ add_library(mtmd
models/qwen3vl.cpp
models/mimovl.cpp
models/qwen3a.cpp
models/mimo-audio.cpp
models/step3vl.cpp
models/siglip.cpp
models/whisper-enc.cpp
tools/mtmd/clip-graph.h
+8
View File
@@ -13,6 +13,14 @@
struct build_vit_opts {
ggml_tensor * attn_mask = nullptr;
// TODO @ngxson : merge attn_mask and attn_mask_layers into one call
std::vector<ggml_tensor *> attn_mask_layers; // one per layer
// hook at layer output embeddings
std::function<void(ggml_tensor * cur, int il)> callback_layer_out = nullptr;
// whether to skip the automatic post-layernorm (model.post_ln_w) applied at the end
bool skip_post_ln = false;
};
struct clip_graph {
tools/mtmd/clip-impl.h
+26
View File
@@ -82,6 +82,12 @@
#define KEY_A_PROJ_WINDOW_SIZE "clip.audio.projector.window_size"
#define KEY_A_PROJ_DOWNSAMPLE_RATE "clip.audio.projector.downsample_rate"
#define KEY_A_PROJ_HEAD_COUNT "clip.audio.projector.head_count"
#define KEY_A_RVQ_NUM_QUANTIZERS "clip.audio.rvq.num_quantizers" // mimo-audio-tokenizer
#define KEY_A_RVQ_CODEBOOK_SIZE "clip.audio.rvq.codebook_size" // mimo-audio-tokenizer: per-quantizer bin count
#define KEY_A_WA_PATTERN_MODE "clip.audio.wa_pattern_mode" // mimo-audio-tokenizer, per-layer -1 (full) / 0 (windowed)
#define KEY_A_ATTN_WINDOW_SIZE "clip.audio.window_size" // mimo-audio-tokenizer: sliding-window radius
#define KEY_A_LOCAL_BLOCK_COUNT "clip.audio.local_block_count" // mimo-v2.5: input_local_transformer layer count
#define KEY_A_LOCAL_GROUP_SIZE "clip.audio.local_group_size" // mimo-v2.5: input_local_transformer grouping size
//
// tensor name constants
@@ -175,6 +181,24 @@
#define TN_MM_NORM_PRE "mm.a.norm_pre.%s"
#define TN_MM_NORM_MID "mm.a.norm_mid.%s"
// mimo-audio-tokenizer
#define TN_A_DOWNSAMPLE_CONV "a.downsample.conv.%s"
#define TN_A_DOWNSAMPLE_NORM "a.downsample.norm.%s"
#define TN_A_RVQ_CODEBOOK "a.rvq.codebook.%s"
// mimo-v2.5: text-side RVQ code embedding ("text codebook")
#define TN_MM_A_CODE_EMBD "mm.a.code_embd.%s"
// mimo-v2.5: LLM-side connector (input_local_transformer)
#define TN_MM_A_LOCAL_ATTN_Q "mm.a.local_blk.%d.attn_q.%s"
#define TN_MM_A_LOCAL_ATTN_K "mm.a.local_blk.%d.attn_k.%s"
#define TN_MM_A_LOCAL_ATTN_V "mm.a.local_blk.%d.attn_v.%s"
#define TN_MM_A_LOCAL_ATTN_OUT "mm.a.local_blk.%d.attn_out.%s"
#define TN_MM_A_LOCAL_FFN_GATE "mm.a.local_blk.%d.ffn_gate.%s"
#define TN_MM_A_LOCAL_FFN_UP "mm.a.local_blk.%d.ffn_up.%s"
#define TN_MM_A_LOCAL_FFN_DOWN "mm.a.local_blk.%d.ffn_down.%s"
#define TN_MM_A_LOCAL_LN1 "mm.a.local_blk.%d.ln1.%s"
#define TN_MM_A_LOCAL_LN2 "mm.a.local_blk.%d.ln2.%s"
#define TN_MM_A_LOCAL_NORM "mm.a.local_norm.%s"
// cogvlm
#define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
#define TN_MM_H_TO_4H "mm.up.%s"
@@ -374,6 +398,7 @@ enum projector_type {
PROJECTOR_TYPE_MIMOVL,
PROJECTOR_TYPE_MINIMAX_M3,
PROJECTOR_TYPE_GRANITE4_VISION,
PROJECTOR_TYPE_MIMO_AUDIO,
PROJECTOR_TYPE_UNKNOWN,
};
@@ -429,6 +454,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_MIMOVL, "mimovl"},
{ PROJECTOR_TYPE_MINIMAX_M3, "minimax_m3"},
{ PROJECTOR_TYPE_GRANITE4_VISION, "granite4_vision"},
{ PROJECTOR_TYPE_MIMO_AUDIO, "mimo_audio"},
};
static projector_type clip_projector_type_from_string(const std::string & str) {
tools/mtmd/clip-model.h
+22
View File
@@ -124,6 +124,14 @@ struct clip_hparams {
int32_t audio_window_len = -1;
int32_t audio_hop_len = -1;
// mimo-audio-tokenizer: residual vector quantizer
int32_t rvq_num_quantizers = 0;
std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)
// mimo-v2.5: LLM-side connector (input_local_transformer)
int32_t audio_local_n_layer = 0;
int32_t audio_local_group_size = 0;
// legacy
bool has_llava_projector = false;
int minicpmv_version = 0;
@@ -537,6 +545,20 @@ struct clip_model {
ggml_tensor * mm_norm_pre_b = nullptr;
ggml_tensor * mm_norm_mid_w = nullptr;
// mimo-audio-tokenizer: post-transformer downsample + RVQ codebook
ggml_tensor * downsample_conv_w = nullptr; // no bias
ggml_tensor * downsample_norm_w = nullptr;
ggml_tensor * downsample_norm_b = nullptr;
ggml_tensor * rvq_codebook = nullptr; // merged 3D [n_q, max_bins, dim]
// mimo-v2.5: text-side RVQ code embedding ("text codebook")
ggml_tensor * mm_a_code_embd = nullptr; // merged 3D [n_channels, vocab, dim]
// mimo-v2.5: LLM-side connector (input_local_transformer, separate from the
// audio_tokenizer's own encoder `layers`)
std::vector<clip_layer> mm_a_local_layers;
ggml_tensor * mm_a_local_norm_w = nullptr;
// qwen3a
ggml_tensor * conv2d_1_w = nullptr;
ggml_tensor * conv2d_1_b = nullptr;
tools/mtmd/clip.cpp
+165 -2
View File
@@ -340,6 +340,11 @@ ggml_tensor * clip_graph::build_vit(
auto & layer = model.layers[il];
ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states
ggml_tensor * attn_mask = opts.attn_mask;
if (opts.attn_mask_layers.size() > (size_t) il) {
attn_mask = opts.attn_mask_layers[il];
}
// layernorm1
cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il);
cb(cur, "layer_inp_normed", il);
@@ -452,7 +457,7 @@ ggml_tensor * clip_graph::build_vit(
// build_attn returns a flat 2D [n_embd, n_pos*B]
cur = build_attn(layer.o_w, layer.o_b,
Qcur, Kcur, Vcur, opts.attn_mask, kq_scale, il);
Qcur, Kcur, Vcur, attn_mask, kq_scale, il);
cb(cur, "attn_out", il);
}
@@ -471,6 +476,10 @@ ggml_tensor * clip_graph::build_vit(
inpL = cur; // inpL = residual, cur = hidden_states
if (opts.callback_layer_out) {
opts.callback_layer_out(cur, il);
}
cb(cur, "ffn_inp", il);
// layernorm2 (pre-ffn norm)
@@ -519,7 +528,7 @@ ggml_tensor * clip_graph::build_vit(
}
// post-layernorm
if (model.post_ln_w) {
if (model.post_ln_w && !opts.skip_post_ln) {
inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, -1);
}
@@ -1012,6 +1021,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
{
builder = std::make_unique<clip_graph_qwen3a>(ctx, img);
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
builder = std::make_unique<clip_graph_mimo_audio>(ctx, img);
} break;
case PROJECTOR_TYPE_YOUTUVL:
{
builder = std::make_unique<clip_graph_youtuvl>(ctx, img);
@@ -1575,6 +1588,45 @@ struct clip_model_loader {
hparams.audio_window_len = 400;
hparams.audio_hop_len = 160;
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
get_u32(KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers, false);
get_arr_int(KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size, false);
if (hparams.rvq_num_quantizers <= 0) {
throw std::runtime_error(string_format("%s: mimo_audio: missing %s\n", __func__, KEY_A_RVQ_NUM_QUANTIZERS));
}
if ((int) hparams.rvq_codebook_size.size() != hparams.rvq_num_quantizers) {
throw std::runtime_error(string_format(
"%s: mimo_audio: %s length (%zu) must equal %s (%d)\n", __func__,
KEY_A_RVQ_CODEBOOK_SIZE, hparams.rvq_codebook_size.size(),
KEY_A_RVQ_NUM_QUANTIZERS, hparams.rvq_num_quantizers));
}
hparams.ffn_op = FFN_GELU_ERF; // PyTorch F.gelu default (approximate="none")
hparams.rope_theta = 10000.0f;
// audio preprocessing params (mel spectrogram)
hparams.audio_sample_rate = 24000;
hparams.audio_n_fft = 960;
hparams.audio_window_len = 960;
hparams.audio_hop_len = 240;
get_u32(KEY_A_ATTN_WINDOW_SIZE, hparams.attn_window_size);
std::vector<int> wa_pattern;
get_arr_int(KEY_A_WA_PATTERN_MODE, wa_pattern, true);
if ((int) wa_pattern.size() != hparams.n_layer) {
throw std::runtime_error(string_format(
"%s: mimo_audio: %s length (%zu) must equal n_layer (%d)\n", __func__,
KEY_A_WA_PATTERN_MODE, wa_pattern.size(), hparams.n_layer));
}
hparams.wa_pattern_mode.assign(wa_pattern.begin(), wa_pattern.end());
get_u32(KEY_A_LOCAL_BLOCK_COUNT, hparams.audio_local_n_layer);
get_u32(KEY_A_LOCAL_GROUP_SIZE, hparams.audio_local_group_size);
if (hparams.audio_local_group_size <= 0) {
throw std::runtime_error(string_format(
"%s: mimo_audio: %s must be > 0\n", __func__, KEY_A_LOCAL_GROUP_SIZE));
}
} break;
case PROJECTOR_TYPE_PADDLEOCR:
{
hparams.n_merge = 2;
@@ -2444,6 +2496,54 @@ struct clip_model_loader {
model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
model.mm_2_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "bias"));
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 1, "bias"));
model.conv1d_2_w = get_tensor(string_format(TN_CONV1D, 2, "weight"));
model.conv1d_2_b = get_tensor(string_format(TN_CONV1D, 2, "bias"));
model.downsample_conv_w = get_tensor(string_format(TN_A_DOWNSAMPLE_CONV, "weight"));
model.downsample_norm_w = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "weight"));
model.downsample_norm_b = get_tensor(string_format(TN_A_DOWNSAMPLE_NORM, "bias"));
model.rvq_codebook = get_tensor(string_format(TN_A_RVQ_CODEBOOK, "weight"), false);
model.mm_a_code_embd = get_tensor(string_format(TN_MM_A_CODE_EMBD, "weight"), false);
if (!model.rvq_codebook || !model.mm_a_code_embd) {
throw std::runtime_error(string_format("%s: mimo_audio: missing %s or %s\n", __func__,
TN_A_RVQ_CODEBOOK, TN_MM_A_CODE_EMBD));
}
// hparams.rvq_codebook_size comes from GGUF metadata and is independent of the
// tensors' actual shapes - bound it so codebook/code_embd views built from it
// (mimo-audio.cpp) can never read past either tensor's allocated bins/vocab.
for (int32_t bins : hparams.rvq_codebook_size) {
if (bins <= 0 || bins > model.rvq_codebook->ne[1] || bins > model.mm_a_code_embd->ne[1]) {
throw std::runtime_error(string_format(
"%s: mimo_audio: %s entry (%d) out of range for codebook/code_embd tensors\n",
__func__, KEY_A_RVQ_CODEBOOK_SIZE, bins));
}
}
// LLM-side connector: input_local_transformer + projection
model.mm_a_local_layers.resize(hparams.audio_local_n_layer);
for (int il = 0; il < hparams.audio_local_n_layer; il++) {
auto & layer = model.mm_a_local_layers[il];
layer.q_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q, il, "weight"));
layer.q_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_Q, il, "bias"));
layer.k_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K, il, "weight"));
layer.k_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_K, il, "bias"));
layer.v_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V, il, "weight"));
layer.v_b = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_V, il, "bias"));
layer.o_w = get_tensor(string_format(TN_MM_A_LOCAL_ATTN_OUT, il, "weight"));
layer.ff_gate_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_GATE, il, "weight"));
layer.ff_up_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_UP, il, "weight"));
layer.ff_down_w = get_tensor(string_format(TN_MM_A_LOCAL_FFN_DOWN, il, "weight"));
layer.ln_1_w = get_tensor(string_format(TN_MM_A_LOCAL_LN1, il, "weight"));
layer.ln_2_w = get_tensor(string_format(TN_MM_A_LOCAL_LN2, il, "weight"));
}
model.mm_a_local_norm_w = get_tensor(string_format(TN_MM_A_LOCAL_NORM, "weight"));
model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
} break;
case PROJECTOR_TYPE_VOXTRAL:
{
model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -3549,6 +3649,15 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
{
n_patches = img->nx(); // no downsampling: one token per raw waveform frame
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
// conv1(s=1) + conv2(s=2) -> RVQ-encoder downsample conv(k=2,s=2)
int n = img->nx();
n = (n - 1) / 2 + 1; // conv1 + conv2
n = (n - 2) / 2 + 1; // downsample conv
const int group_size = params.audio_local_group_size;
n_patches = (n + group_size - 1) / group_size;
} break;
case PROJECTOR_TYPE_GRANITE_SPEECH:
{
const int ws = ctx->model.hparams.audio_proj_window_size;
@@ -4376,6 +4485,58 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
set_input_f32("pos_emb", pos_emb);
}
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
GGML_ASSERT(imgs.entries.size() == 1);
const int n_frames = imgs.entries.front().nx();
const int n_pos = (n_frames - 1) / 2 + 1; // matches conv1(s=1)+conv2(s=2) output length
std::vector<int32_t> positions(n_pos);
for (int i = 0; i < n_pos; i++) {
positions[i] = i;
}
set_input_i32("mimo_audio_positions", positions);
const int window = hparams.attn_window_size;
GGML_ASSERT(window > 0);
const float neg_inf = std::numeric_limits<float>::lowest();
std::vector<float> full_mask((size_t) n_pos * n_pos);
std::vector<float> window_mask((size_t) n_pos * n_pos);
for (int q = 0; q < n_pos; q++) {
for (int k = 0; k < n_pos; k++) {
const bool causal_ok = k <= q;
full_mask[(size_t) q * n_pos + k] = causal_ok ? 0.0f : neg_inf;
window_mask[(size_t) q * n_pos + k] = (causal_ok && (q - k) <= window) ? 0.0f : neg_inf;
}
}
set_input_f32("mimo_audio_full_mask", full_mask);
set_input_f32("mimo_audio_window_mask", window_mask);
// input_local_transformer: block-diagonal mask + in-group positions
{
const int n_pos_ds = (n_pos - 2) / 2 + 1; // matches downsample conv (k=2,s=2,p=0)
const int group_size = hparams.audio_local_group_size;
GGML_ASSERT(group_size > 0);
const int n_groups = (n_pos_ds + group_size - 1) / group_size;
const int n_padded = n_groups * group_size;
std::vector<int32_t> local_positions(n_padded);
for (int i = 0; i < n_padded; i++) {
local_positions[i] = i % group_size;
}
set_input_i32("mimo_audio_local_positions", local_positions);
std::vector<float> local_mask((size_t) n_padded * n_padded);
for (int q = 0; q < n_padded; q++) {
for (int k = 0; k < n_padded; k++) {
const bool same_group = (q / group_size) == (k / group_size);
local_mask[(size_t) q * n_padded + k] = same_group ? 0.0f : neg_inf;
}
}
set_input_f32("mimo_audio_local_mask", local_mask);
}
} break;
case PROJECTOR_TYPE_LFM2A:
{
GGML_ASSERT(imgs.entries.size() == 1);
@@ -4678,6 +4839,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return ctx->model.qf_proj_blocks.size() * ctx->model.hparams.projection_dim;
case PROJECTOR_TYPE_GLM4V:
return ctx->model.mm_ffn_down_w->ne[1];
case PROJECTOR_TYPE_MIMO_AUDIO:
return ctx->model.mm_2_w->ne[1];
default:
GGML_ABORT("Unknown projector type");
}
tools/mtmd/models/mimo-audio.cpp
+218
View File
@@ -0,0 +1,218 @@
#include "models.h"
ggml_cgraph * clip_graph_mimo_audio::build() {
ggml_tensor * inp = build_inp_raw(1); // [n_frames, n_mel, 1]
ggml_tensor * cur = ggml_conv_1d_ph(ctx0, model.conv1d_1_w, inp, 1, 1);
cur = ggml_add(ctx0, cur, model.conv1d_1_b);
cur = ggml_gelu_erf(ctx0, cur);
cur = ggml_conv_1d_ph(ctx0, model.conv1d_2_w, cur, 2, 1);
cur = ggml_add(ctx0, cur, model.conv1d_2_b);
cur = ggml_gelu_erf(ctx0, cur);
ggml_tensor * inpL = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); // [n_embd, n_pos]
const int64_t n_pos = inpL->ne[1];
cb(inpL, "after_conv1d", -1);
GGML_ASSERT((int) hparams.wa_pattern_mode.size() == n_layer);
ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
ggml_set_name(inp_pos, "mimo_audio_positions");
ggml_set_input(inp_pos);
ggml_tensor * full_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
ggml_set_name(full_mask, "mimo_audio_full_mask");
ggml_set_input(full_mask);
ggml_tensor * window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos);
ggml_set_name(window_mask, "mimo_audio_window_mask");
ggml_set_input(window_mask);
build_vit_opts opts;
opts.attn_mask_layers.resize(n_layer);
for (int il = 0; il < n_layer; il++) {
opts.attn_mask_layers[il] = hparams.wa_pattern_mode[il] == -1 ? full_mask : window_mask;
}
// the skip connection below must be added before the post-transformer norm,
// so build_vit must not apply that norm itself
opts.skip_post_ln = true;
// encoder_skip_layer_id=3 (1-indexed) -> capture output of layer index 2
const int skip_capture_il = 2;
GGML_ASSERT(n_layer > skip_capture_il);
ggml_tensor * skip_hidden = nullptr;
opts.callback_layer_out = [&](ggml_tensor * layer_cur, int il) {
if (il == skip_capture_il) {
skip_hidden = layer_cur;
}
};
auto add_pos = [&](ggml_tensor * x, const clip_layer &) {
return ggml_rope_ext(ctx0, x, inp_pos, nullptr, d_head,
GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
};
inpL = build_vit(inpL, n_pos, NORM_TYPE_NORMAL, hparams.ffn_op, nullptr, add_pos, opts);
inpL = ggml_reshape_2d(ctx0, inpL, n_embd, n_pos); // build_vit restores a (size-1) batch dim
GGML_ASSERT(skip_hidden != nullptr);
inpL = ggml_add(ctx0, inpL, skip_hidden);
inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1);
cb(inpL, "after_transformer", -1);
// downsample: strided conv (no bias) + gelu + layernorm
{
ggml_tensor * ds = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); // [n_pos, n_embd]
ds = ggml_conv_1d(ctx0, model.downsample_conv_w, ds, 2, 0, 1);
ds = ggml_gelu_erf(ctx0, ds);
ds = ggml_cont(ctx0, ggml_transpose(ctx0, ds)); // [n_embd, n_pos/2]
ds = build_norm(ds, model.downsample_norm_w, model.downsample_norm_b, NORM_TYPE_NORMAL, eps, -1);
inpL = ds;
}
cb(inpL, "after_downsample", -1);
// RVQ quantize: codebook ne=[dim, max_bins, n_q]
// quantize input vector to codes (type=I32)
std::vector<ggml_tensor *> codes;
{
GGML_ASSERT(model.rvq_codebook != nullptr);
const int64_t dim = model.rvq_codebook->ne[0];
GGML_ASSERT(dim == inpL->ne[0]);
GGML_ASSERT((int64_t) hparams.rvq_codebook_size.size() == model.rvq_codebook->ne[2]);
ggml_tensor * residual = inpL; // [dim, n_pos_ds]
for (size_t q = 0; q < hparams.rvq_codebook_size.size(); q++) {
const int64_t bins = hparams.rvq_codebook_size[q];
ggml_tensor * codebook_q = ggml_view_2d(ctx0, model.rvq_codebook, dim, bins,
model.rvq_codebook->nb[1], q * model.rvq_codebook->nb[2]);
codebook_q = ggml_cont(ctx0, codebook_q);
ggml_tensor * codebook_norm = ggml_sum_rows(ctx0, ggml_sqr(ctx0, codebook_q)); // [1, bins]
codebook_norm = ggml_cont(ctx0, ggml_transpose(ctx0, codebook_norm)); // [bins, 1]
ggml_tensor * dot = ggml_mul_mat(ctx0, codebook_q, residual); // [bins, n_pos_ds]
ggml_tensor * scores = ggml_sub(ctx0, ggml_scale(ctx0, dot, 2.0f), codebook_norm);
ggml_tensor * idx = ggml_argmax(ctx0, scores); // [n_pos_ds]
codes.push_back(idx);
ggml_tensor * quant = ggml_get_rows(ctx0, codebook_q, idx); // [dim, n_pos_ds]
residual = ggml_sub(ctx0, residual, quant);
cb(idx, "rvq_code", (int) q);
}
}
// convert codes to LLM embeddings
ggml_tensor * code_embd_sum = nullptr;
{
GGML_ASSERT(model.mm_a_code_embd != nullptr);
const int64_t dim = model.mm_a_code_embd->ne[0];
const int64_t vocab = model.mm_a_code_embd->ne[1];
GGML_ASSERT((int64_t) codes.size() == model.mm_a_code_embd->ne[2]);
GGML_ASSERT(dim == inpL->ne[0]);
for (size_t i = 0; i < codes.size(); i++) {
ggml_tensor * table_i = ggml_view_2d(ctx0, model.mm_a_code_embd, dim, vocab,
model.mm_a_code_embd->nb[1], i * model.mm_a_code_embd->nb[2]);
table_i = ggml_cont(ctx0, table_i);
ggml_tensor * embd_i = ggml_get_rows(ctx0, table_i, codes[i]); // [dim, n_pos_ds]
code_embd_sum = code_embd_sum ? ggml_add(ctx0, code_embd_sum, embd_i) : embd_i;
}
cb(code_embd_sum, "code_embd_sum", -1);
}
// input_local_transformer
// groups of `group_size` consecutive downsampled frames are processed together, attending only within their own group.
// Implemented as a block-diagonal mask + in-group-repeating positions
// (rather than a real batch dim) - same technique as the encoder's masks above, and as gemma4a's / deepseekocr2's chunked attention.
// note: hand-rolled here instead of build_vit() because this is a second, independent layer stack
// (own layer array/count, RMSNorm instead of LN, SiLU FFN, own RoPE theta)
ggml_tensor * projected;
{
const int group_size = hparams.audio_local_group_size;
GGML_ASSERT(group_size > 0);
const int64_t n_pos_ds = code_embd_sum->ne[1];
const int64_t n_groups = (n_pos_ds + group_size - 1) / group_size;
const int64_t n_padded = n_groups * group_size;
ggml_tensor * cur_local = code_embd_sum;
if (n_padded != n_pos_ds) {
cur_local = ggml_pad(ctx0, cur_local, 0, (int) (n_padded - n_pos_ds), 0, 0);
}
ggml_tensor * local_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_padded);
ggml_set_name(local_pos, "mimo_audio_local_positions");
ggml_set_input(local_pos);
ggml_tensor * local_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_padded, n_padded);
ggml_set_name(local_mask, "mimo_audio_local_mask");
ggml_set_input(local_mask);
const float local_rope_theta = 640000.0f; // audio_config.rope_theta (differs from the encoder's)
auto apply_local_rope = [&](ggml_tensor * x) {
return ggml_rope_ext(ctx0, x, local_pos, nullptr, d_head,
GGML_ROPE_TYPE_NEOX, 0, local_rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
};
for (int il = 0; il < hparams.audio_local_n_layer; il++) {
auto & layer = model.mm_a_local_layers[il];
ggml_tensor * attn_in = build_norm(cur_local, layer.ln_1_w, nullptr, NORM_TYPE_RMS, eps, il);
ggml_tensor * Qcur = build_mm(layer.q_w, attn_in);
if (layer.q_b) {
Qcur = ggml_add(ctx0, Qcur, layer.q_b);
}
ggml_tensor * Kcur = build_mm(layer.k_w, attn_in);
if (layer.k_b) {
Kcur = ggml_add(ctx0, Kcur, layer.k_b);
}
ggml_tensor * Vcur = build_mm(layer.v_w, attn_in);
if (layer.v_b) {
Vcur = ggml_add(ctx0, Vcur, layer.v_b);
}
Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_padded);
Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_padded);
Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_padded);
Qcur = apply_local_rope(Qcur);
Kcur = apply_local_rope(Kcur);
ggml_tensor * attn_out = build_attn(layer.o_w, nullptr, Qcur, Kcur, Vcur, local_mask, kq_scale, il);
cur_local = ggml_add(ctx0, cur_local, attn_out);
ggml_tensor * ffn_in = build_norm(cur_local, layer.ln_2_w, nullptr, NORM_TYPE_RMS, eps, il);
ggml_tensor * ffn_out = build_ffn(ffn_in,
layer.ff_up_w, nullptr,
layer.ff_gate_w, nullptr,
layer.ff_down_w, nullptr,
FFN_SILU, il);
cur_local = ggml_add(ctx0, cur_local, ffn_out);
}
cur_local = build_norm(cur_local, model.mm_a_local_norm_w, nullptr, NORM_TYPE_RMS, eps, -1);
cb(cur_local, "after_local_transformer", -1);
// flatten each group of `group_size` frames into one (group_size*n_embd)-dim vector
// (matching AudioProjection's flattened input)
ggml_tensor * grouped = ggml_reshape_2d(ctx0, cur_local, n_embd * group_size, n_groups);
// AudioProjection: Linear (no bias) -> GELU -> Linear (no bias)
projected = build_ffn(grouped,
model.mm_1_w, nullptr,
nullptr, nullptr,
model.mm_2_w, nullptr,
FFN_GELU_ERF, -1);
cb(projected, "after_projection", -1);
}
ggml_build_forward_expand(gf, projected);
return gf;
}
tools/mtmd/models/models.h
+5
View File
@@ -210,6 +210,11 @@ struct clip_graph_qwen3a : clip_graph {
ggml_cgraph * build() override;
};
struct clip_graph_mimo_audio : clip_graph {
clip_graph_mimo_audio(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct clip_graph_kimik25 : clip_graph {
clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
tools/mtmd/mtmd-audio.cpp
+66
View File
@@ -725,6 +725,72 @@ bool mtmd_audio_preprocessor_qwen3a::preprocess(const float * sa
return true;
}
//
// mtmd_audio_preprocessor_mimo_audio
//
// Matches torchaudio.transforms.MelSpectrogram(power=1.0, center=True) followed by
// log(clip(spec, min=1e-7)): HTK mel scale, no Slaney area norm, magnitude (not power)
// spectrogram, natural log, reflect-padded by n_fft/2 on each side.
//
void mtmd_audio_preprocessor_mimo_audio::initialize() {
cache.fill_sin_cos_table(hparams.audio_n_fft);
cache.fill_hann_window(hparams.audio_window_len, true);
cache.fill_mel_filterbank_matrix(
hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate,
0.0f, hparams.audio_sample_rate / 2.0f,
/*slaney_area_norm=*/ false,
/*scale=*/ 1.0f,
/*use_htk=*/ true
);
}
bool mtmd_audio_preprocessor_mimo_audio::preprocess(const float * samples,
size_t n_samples,
std::vector<mtmd_audio_mel> & output) {
if (n_samples == 0) {
return false;
}
GGML_ASSERT(!cache.sin_vals.empty());
GGML_ASSERT(!cache.cos_vals.empty());
GGML_ASSERT(!cache.filters.data.empty());
const int pad = hparams.audio_n_fft / 2;
std::vector<float> padded(n_samples + 2 * pad, 0.0f);
for (int i = 0; i < pad; i++) {
int src = pad - i;
padded[i] = (src < (int)n_samples) ? samples[src] : 0.0f;
}
std::copy(samples, samples + n_samples, padded.begin() + pad);
for (int i = 0; i < pad; i++) {
int src = (int)n_samples - 2 - i;
padded[n_samples + pad + i] = (src >= 0) ? samples[src] : 0.0f;
}
filter_params params;
params.n_mel = hparams.n_mel_bins;
params.n_fft_bins = 1 + (hparams.audio_n_fft / 2);
params.hann_window_size = hparams.audio_window_len;
params.hop_length = hparams.audio_hop_len;
params.sample_rate = hparams.audio_sample_rate;
params.no_padding = true; // reflect padding already applied above
params.use_natural_log = true;
params.use_magnitude = true;
params.mel_floor = 1e-7f;
params.norm_per_feature = false;
mtmd_audio_mel out;
bool ok = log_mel_spectrogram(padded.data(), (int)padded.size(), 4, params, cache, out);
if (!ok) {
return false;
}
output.push_back(std::move(out));
return true;
}
//
// mtmd_audio_preprocessor_conformer
//
tools/mtmd/mtmd-audio.h
+9
View File
@@ -111,6 +111,15 @@ struct mtmd_audio_preprocessor_qwen3a : mtmd_audio_preprocessor {
mtmd_audio_cache cache;
};
struct mtmd_audio_preprocessor_mimo_audio : mtmd_audio_preprocessor {
mtmd_audio_preprocessor_mimo_audio(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
void initialize() override;
bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
private:
mtmd_audio_cache cache;
};
//
// streaming ISTFT - converts spectrogram frames back to audio one frame at a time
//
tools/mtmd/mtmd.cpp
+6
View File
@@ -730,6 +730,12 @@ struct mtmd_context {
aud_end = "<audio|>";
audio_preproc = std::make_unique<mtmd_audio_preprocessor_gemma4ua>(ctx_a);
} break;
case PROJECTOR_TYPE_MIMO_AUDIO:
{
aud_beg = "<|mimo_audio_start|>";
aud_end = "<|mimo_audio_end|>";
audio_preproc = std::make_unique<mtmd_audio_preprocessor_mimo_audio>(ctx_a);
} break;
default:
throw std::runtime_error(string_format("%s: unexpected audio projector type %d\n", __func__, proj));
}