Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '11b068d06605288ce7917534b46d52b47823dc13' into concedo_experimental

Browse Source 
# Conflicts:
#	CONTRIBUTING.md
#	docs/backend/SYCL.md
#	docs/install.md
#	docs/speculative.md
#	ggml/src/ggml-hip/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-sycl/common.hpp
#	ggml/src/ggml-sycl/element_wise.cpp
#	ggml/src/ggml-sycl/fattn-onednn.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl
#	ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
#	tests/test-llama-archs.cpp
#	tools/cli/README.md
#	tools/llama-bench/llama-bench.cpp
#	tools/mtmd/CMakeLists.txt
#	tools/server/README.md
This commit is contained in:
Concedo
2026-07-31 17:06:19 +08:00
parent 49ff71991e 11b068d066
commit 98926f27c1
81 changed files with 5314 additions and 518 deletions
Download Patch File Download Diff File Expand all files Collapse all files
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
+35
View File
@@ -82,6 +82,13 @@
#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
#define KEY_AUDIO_SUBSAMPLING_FACTOR "clip.audio.subsampling_factor"
//
// tensor name constants
@@ -175,6 +182,24 @@
#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"
@@ -314,6 +339,12 @@
#define TN_YASA_STAGE_DOWN_CONV "v.stage.%d.down.conv.%s"
#define TN_YASA_STAGE_BLK "v.stage.%d.blk.%d.%s.%s"
// parakeet
#define TN_MEL_FILTERS "a.mel_filters"
#define TN_WINDOW "a.window"
#define TN_CONV_NORM_MEAN "%s.blk.%d.conv_norm_mean"
#define TN_CONV_NORM_VAR "%s.blk.%d.conv_norm_var"
// align x to upper multiple of n
#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
@@ -368,12 +399,14 @@ enum projector_type {
PROJECTOR_TYPE_KIMIK25,
PROJECTOR_TYPE_NEMOTRON_V2_VL,
PROJECTOR_TYPE_HUNYUANVL,
PROJECTOR_TYPE_PARAKEET,
PROJECTOR_TYPE_EXAONE4_5,
PROJECTOR_TYPE_MINICPMV4_6,
PROJECTOR_TYPE_GRANITE_SPEECH,
PROJECTOR_TYPE_MIMOVL,
PROJECTOR_TYPE_MINIMAX_M3,
PROJECTOR_TYPE_GRANITE4_VISION,
PROJECTOR_TYPE_MIMO_AUDIO,
PROJECTOR_TYPE_UNKNOWN,
};
@@ -429,6 +462,8 @@ 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"},
{ PROJECTOR_TYPE_PARAKEET, "parakeet"},
};
static projector_type clip_projector_type_from_string(const std::string & str) {
tools/mtmd/clip-model.h
+38 -8
View File
@@ -110,6 +110,8 @@ struct clip_hparams {
// audio
int32_t n_mel_bins = 0; // whisper preprocessor
int32_t proj_stack_factor = 0; // ultravox
int32_t subsampling_factor = 0; // parakeet
int32_t audio_chunk_size = 0;
int32_t audio_conv_kernel_size = 0;
int32_t audio_max_pos_emb = 0;
@@ -124,6 +126,18 @@ struct clip_hparams {
int32_t audio_window_len = -1;
int32_t audio_hop_len = -1;
// parakeet
std::vector<float> mel_filters;
std::vector<float> window;
// 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;
@@ -237,14 +251,16 @@ struct clip_layer {
ggml_tensor * norm_conv_b = nullptr;
ggml_tensor * linear_pos_w = nullptr;
ggml_tensor * conv_norm_w = nullptr;
ggml_tensor * conv_norm_b = nullptr;
ggml_tensor * conv_dw_w = nullptr;
ggml_tensor * conv_dw_b = nullptr;
ggml_tensor * conv_pw1_w = nullptr;
ggml_tensor * conv_pw1_b = nullptr;
ggml_tensor * conv_pw2_w = nullptr;
ggml_tensor * conv_pw2_b = nullptr;
ggml_tensor * conv_norm_w = nullptr;
ggml_tensor * conv_norm_b = nullptr;
ggml_tensor * conv_norm_mean = nullptr; // parakeet
ggml_tensor * conv_norm_var = nullptr; // parakeet
ggml_tensor * conv_dw_w = nullptr;
ggml_tensor * conv_dw_b = nullptr;
ggml_tensor * conv_pw1_w = nullptr;
ggml_tensor * conv_pw1_b = nullptr;
ggml_tensor * conv_pw2_w = nullptr;
ggml_tensor * conv_pw2_b = nullptr;
// gemma4 audio conformer per-layer
ggml_tensor * attn_pre_norm_w = nullptr;
@@ -537,6 +553,20 @@ 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
+372 -9
View File
@@ -61,11 +61,13 @@
#include "models/minicpmv.cpp"
#include "models/minimax-m3.cpp"
#include "models/paddleocr.cpp"
#include "models/parakeet.cpp"
#include "models/pixtral.cpp"
#include "models/qwen2vl.cpp"
#include "models/qwen3vl.cpp"
#include "models/mimovl.cpp"
#include "models/qwen3a.cpp"
#include "models/mimovl.cpp"
#include "models/mimo-audio.cpp"
#include "models/step3vl.cpp"
#include "models/siglip.cpp"
#include "models/whisper-enc.cpp"
@@ -389,6 +391,11 @@ 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);
@@ -501,7 +508,7 @@ 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);
}
@@ -520,6 +527,10 @@ 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)
@@ -568,7 +579,7 @@ 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);
}
@@ -1061,6 +1072,10 @@ 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);
@@ -1069,6 +1084,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
{
builder = std::make_unique<clip_graph_yasa2>(ctx, img);
} break;
case PROJECTOR_TYPE_PARAKEET:
{
builder = std::make_unique<clip_graph_parakeet>(ctx, img);
} break;
case PROJECTOR_TYPE_GRANITE4_VISION:
{
builder = std::make_unique<clip_graph_granite4_vision>(ctx, img);
@@ -1415,6 +1434,20 @@ struct clip_model_loader {
{
get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
} break;
case PROJECTOR_TYPE_PARAKEET:
{
get_u32(KEY_AUDIO_SUBSAMPLING_FACTOR, hparams.subsampling_factor);
GGML_ASSERT(hparams.subsampling_factor == 8 &&
"subsampling_factor must match the conv strides in clip_graph_parakeet::build()");
get_u32(KEY_A_CONV_KERNEL_SIZE, hparams.audio_conv_kernel_size);
GGML_ASSERT(hparams.audio_conv_kernel_size > 0 && hparams.audio_conv_kernel_size % 2 == 1 &&
"audio_conv_kernel_size must be a positive odd integer");
hparams.audio_chunk_len = 0;
hparams.audio_sample_rate = 16000;
hparams.audio_n_fft = 512;
hparams.audio_window_len = 400;
hparams.audio_hop_len = 160;
} break;
case PROJECTOR_TYPE_IDEFICS3:
{
// use default llava-uhd preprocessing params
@@ -1652,6 +1685,45 @@ 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;
@@ -1923,16 +1995,46 @@ struct clip_model_loader {
return cur;
};
auto get_scalar = [&](const std::string & name, float default_val) {
auto get_vector = [&](const std::string & name) {
std::vector<float> result;
auto it = tensor_offset.find(name);
if (it == tensor_offset.end()) {
return result;
}
const int64_t idx = gguf_find_tensor(ctx_gguf.get(), name.c_str());
if (idx < 0) {
throw std::runtime_error(string_format("%s: failed to find tensor %s\n", __func__, name.c_str()));
}
if (const auto type = gguf_get_tensor_type(ctx_gguf.get(), idx); type != GGML_TYPE_F32) {
throw std::runtime_error(string_format("%s: %s must be %s, was %s\n", __func__,
name.c_str(), ggml_type_name(GGML_TYPE_F32), ggml_type_name(type)));
}
const size_t n_bytes = gguf_get_tensor_size(ctx_gguf.get(), idx);
if (n_bytes == 0) {
throw std::runtime_error(string_format("%s: tensor %s is empty\n", __func__, name.c_str()));
}
const size_t n_elems = n_bytes / sizeof(float);
result.resize(n_elems);
fin.seekg(it->second, std::ios::beg);
fin.read(reinterpret_cast<char*>(result.data()), n_bytes);
return result;
};
auto get_scalar = [&](const std::string & name, float default_val) {
auto v = get_vector(name);
if (v.empty()) {
return default_val;
}
size_t offset = it->second;
fin.seekg(offset, std::ios::beg);
float value;
fin.read(reinterpret_cast<char*>(&value), sizeof(float));
return value;
if (v.size() != 1) {
throw std::runtime_error(string_format("%s: expected scalar tensor '%s' but got %d elements\n",
__func__, name.c_str(), (int) v.size()));
}
return v[0];
};
model.class_embedding = get_tensor(TN_CLASS_EMBD, false);
@@ -2526,6 +2628,54 @@ 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"));
@@ -2782,6 +2932,68 @@ struct clip_model_loader {
layer.conv_pw2_b = get_tensor(string_format(TN_CONV_PW2, prefix, il, "bias"));
}
} break;
case PROJECTOR_TYPE_PARAKEET:
{
hparams.mel_filters = get_vector(TN_MEL_FILTERS);
hparams.window = get_vector(TN_WINDOW);
// Subsampling layers (conv1d)
for (int i : {0, 2, 3, 5, 6}) {
model.pre_encode_conv_X_w[i] = get_tensor(string_format(TN_CONV1D, i, "weight"));
model.pre_encode_conv_X_b[i] = get_tensor(string_format(TN_CONV1D, i, "bias"));
}
model.pre_encode_out_w = get_tensor(string_format(TN_PRE_ENCODE_OUT, "weight"));
model.pre_encode_out_b = get_tensor(string_format(TN_PRE_ENCODE_OUT, "bias"));
// Projection layers
model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"), false);
model.mm_0_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"), false);
model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"), false);
// Encoder layers
for (int il = 0; il < hparams.n_layer; ++il) {
auto & layer = model.layers[il];
// Attention (from shared above)
// Relative position encoding
layer.linear_pos_w = get_tensor(string_format(TN_LINEAR_POS, prefix, il, "weight"));
layer.pos_bias_u = get_tensor(string_format(TN_POS_BIAS_U, prefix, il));
layer.pos_bias_v = get_tensor(string_format(TN_POS_BIAS_V, prefix, il));
// Convolution module
layer.conv_pw1_w = get_tensor(string_format(TN_CONV_PW1, prefix, il, "weight"));
layer.conv_pw1_b = get_tensor(string_format(TN_CONV_PW1, prefix, il, "bias"), false);
layer.conv_dw_w = get_tensor(string_format(TN_CONV_DW, prefix, il, "weight"));
layer.conv_dw_b = get_tensor(string_format(TN_CONV_DW, prefix, il, "bias"), false);
layer.conv_norm_w = get_tensor(string_format(TN_CONV_NORM, prefix, il, "weight"));
layer.conv_norm_b = get_tensor(string_format(TN_CONV_NORM, prefix, il, "bias"));
layer.conv_norm_mean = get_tensor(string_format(TN_CONV_NORM_MEAN, prefix, il));
layer.conv_norm_var = get_tensor(string_format(TN_CONV_NORM_VAR, prefix, il));
layer.conv_pw2_w = get_tensor(string_format(TN_CONV_PW2, prefix, il, "weight"));
layer.conv_pw2_b = get_tensor(string_format(TN_CONV_PW2, prefix, il, "bias"), false);
// Feed-forward networks
layer.ff_norm_w = get_tensor(string_format(TN_FFN_NORM, prefix, il, "weight"));
layer.ff_norm_b = get_tensor(string_format(TN_FFN_NORM, prefix, il, "bias"));
layer.ff_norm_1_w = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "weight"));
layer.ff_norm_1_b = get_tensor(string_format(TN_FFN_NORM_1, prefix, il, "bias"));
layer.ff_up_1_w = get_tensor(string_format(TN_FFN_UP_1, prefix, il, "weight"));
layer.ff_up_1_b = get_tensor(string_format(TN_FFN_UP_1, prefix, il, "bias"), false);
layer.ff_down_1_w = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "weight"));
layer.ff_down_1_b = get_tensor(string_format(TN_FFN_DOWN_1, prefix, il, "bias"), false);
// Layer norms
layer.norm_conv_w = get_tensor(string_format(TN_NORM_CONV, prefix, il, "weight"));
layer.norm_conv_b = get_tensor(string_format(TN_NORM_CONV, prefix, il, "bias"));
}
model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight"));
model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight"));
model.mm_model_mlp_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
} break;
case PROJECTOR_TYPE_GRANITE_SPEECH:
{
model.inp_proj_w = get_tensor(string_format(TN_INP_PROJ, "weight"));
@@ -3627,10 +3839,23 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
}
n_patches = n;
} break;
case PROJECTOR_TYPE_PARAKEET:
{
n_patches = (img->nx() + (params.subsampling_factor - 1)) / params.subsampling_factor;
} break;
case PROJECTOR_TYPE_GEMMA4UA:
{
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;
@@ -4458,6 +4683,58 @@ 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);
@@ -4479,6 +4756,88 @@ 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_PARAKEET:
{
GGML_ASSERT(imgs.entries.size() == 1);
struct ggml_tensor * attn_mask = ggml_graph_get_tensor(gf, "attn_mask");
const int n_q = attn_mask->ne[1];
const int n_k = attn_mask->ne[0];
const int n_frames = imgs.entries.front().nx();
const int n_tokens_real = (n_frames + hparams.subsampling_factor-1) / hparams.subsampling_factor;
const float mask_value = -1e30f;
std::vector<float> mask_data(n_q * n_k);
if (n_k == n_q) {
// full attention: mask keys that are padding
for (int q = 0; q < n_q; ++q) {
for (int k = 0; k < n_k; ++k) {
mask_data[q * n_k + k] = (k >= n_tokens_real) ? mask_value : 0.0f;
}
}
} else {
// local attention: mask keys outside the valid window
const int att_left = n_k / 2;
for (int q = 0; q < n_q; ++q) {
for (int k = 0; k < n_k; ++k) {
const int key = q - att_left + k;
mask_data[q * n_k + k] = (key >= 0 && key < n_tokens_real) ? 0.0f : mask_value;
}
}
}
set_input_f32(attn_mask->name, mask_data);
// local attention skew mask: zeroes out the probs that were
// computed for keys outside the valid sliding window.
if (struct ggml_tensor * local_mask = ggml_graph_get_tensor(gf, "local_mask")) {
const int lm_k = local_mask->ne[0];
const int lm_q = local_mask->ne[1];
const int window_size = lm_k - lm_q + 1;
std::vector<float> lm_data(lm_q * lm_k);
for (int q = 0; q < lm_q; ++q) {
for (int k = 0; k < lm_k; ++k) {
const int rel = k - q;
lm_data[q * lm_k + k] = (rel >= 0 && rel < window_size) ? 1.0f : 0.0f;
}
}
set_input_f32(local_mask->name, lm_data);
}
// Generate rotation frequencies for relative positional encoding.
{
const int n_state = hparams.n_embd;
const int d_half = n_state / 2;
const float log_10000 = logf(10000.0f);
std::vector<float> freqs(d_half);
for (int k = 0; k < d_half; ++k) {
freqs[k] = expf(-(float(k * 2) * log_10000 / float(n_state)));
}
set_input_f32("pos_freqs", freqs);
}
// Generate relative positional distance values which scaled by
// the frequency to produce the angles for sin/cos.
{
// window_size is only known after graph construction since it depends on
// n_time from the conv output, so we read it back from the graph tensor.
struct ggml_tensor * rel_pos = ggml_graph_get_tensor(gf, "rel_positions");
const int window_size = rel_pos->ne[1];
std::vector<float> pos(window_size);
// local attention: window is fixed at [att_left, att_right]
// full attention: window covers the full sequence, centered
if (ggml_graph_get_tensor(gf, "local_mask")) {
const int att_left = window_size / 2;
for (int t = 0; t < window_size; ++t) {
pos[t] = float(att_left - t);
}
} else {
const int n_time = (window_size + 1) / 2;
for (int t = 0; t < window_size; ++t) {
pos[t] = float(n_time - 1 - t);
}
}
set_input_f32(rel_pos->name, pos);
}
} break;
case PROJECTOR_TYPE_GRANITE_SPEECH:
{
const int context_size = ctx->model.hparams.audio_chunk_size;
@@ -4760,6 +5119,10 @@ 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];
case PROJECTOR_TYPE_PARAKEET:
return ctx->model.mm_1_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
+10
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;
@@ -217,6 +222,11 @@ struct clip_graph_kimik25 : clip_graph {
ggml_tensor * resize_position_embeddings_3d(uint32_t interpolation_mode);
};
struct clip_graph_parakeet : clip_graph {
clip_graph_parakeet(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
};
struct clip_graph_exaone4_5 : clip_graph {
clip_graph_exaone4_5(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
tools/mtmd/models/parakeet.cpp
+421
View File
@@ -0,0 +1,421 @@
#include "models.h"
static constexpr int PARAKEET_LOCAL_ATTN_THRESHOLD = 8192;
static constexpr int PARAKEET_LOCAL_ATTN_WINDOW = 128;
// conv subsampling + conformer encoder
ggml_cgraph * clip_graph_parakeet::build() {
// Conv subsampling
ggml_tensor * inp = build_inp_raw(1);
inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp));
// [freq, time, channels, batch]
ggml_tensor * cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[0], inp, 2, 2, 1, 1, 1, 1);
cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[0]);
cb(cur, "pre_conv_0", -1);
cur = ggml_relu(ctx0, cur);
cb(cur, "pre_conv_0_relu", -1);
// [freq, time, channels, batch]
cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[2], cur, 2, 2, 1, 1, 1, 1);
cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[2]);
cb(cur, "pre_conv_2", -1);
// [freq, time, channels, batch]
cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[3], cur, 1, 1, 0, 0, 1, 1);
cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[3]);
cb(cur, "pre_conv_3", -1);
cur = ggml_relu(ctx0, cur);
cb(cur, "pre_conv_3_relu", -1);
// [freq, time, channels, batch]
cur = ggml_conv_2d_dw_direct(ctx0, model.pre_encode_conv_X_w[5], cur, 2, 2, 1, 1, 1, 1);
cb(cur, "pre_conv_5_direct", -1);
cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[5]);
cb(cur, "pre_conv_5", -1);
// [freq, time, channels, batch]
cur = ggml_conv_2d(ctx0, model.pre_encode_conv_X_w[6], cur, 1, 1, 0, 0, 1, 1);
cur = ggml_add(ctx0, cur, model.pre_encode_conv_X_b[6]);
cb(cur, "pre_conv_6", -1);
cur = ggml_relu(ctx0, cur);
cb(cur, "pre_conv_6_relu", -1);
// [freq, time, chan]
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
// [freq, chan, time]
cur = ggml_cont(ctx0, cur);
const int n_freq = cur->ne[0];
const int n_chan = cur->ne[1];
const int n_frames = cur->ne[2];
// [freq, time, chan, batch] -> [(freq * chan), time]
cur = ggml_reshape_2d(ctx0, cur, n_freq * n_chan, n_frames);
cur = build_mm(model.pre_encode_out_w, cur);
cur = ggml_add(ctx0, cur, model.pre_encode_out_b);
ggml_set_name(cur, "pre_enc_out");
// Encoder
const auto & hparams = model.hparams;
const int n_layer = hparams.n_layer;
const int n_state = hparams.n_embd;
const float fc_factor = 0.5f;
const int n_time = cur->ne[1];
const bool local_attn = n_time > PARAKEET_LOCAL_ATTN_THRESHOLD;
const int att_left = local_attn ? PARAKEET_LOCAL_ATTN_WINDOW : n_time - 1;
const int att_right = local_attn ? PARAKEET_LOCAL_ATTN_WINDOW : n_time - 1;
const int window_size = local_attn ? att_left + att_right + 1 : 2 * n_time - 1;
const int d_half = n_state / 2;
const int mask_dim = local_attn ? window_size : n_time;
// mask [key, n_time]
struct ggml_tensor * attn_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, mask_dim, n_time);
ggml_set_name(attn_mask, "attn_mask");
ggml_set_input(attn_mask);
struct ggml_tensor * local_mask = nullptr;
if (local_attn) {
const int chunk = att_left + att_right;
local_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, chunk + window_size - 1, chunk);
ggml_set_name(local_mask, "local_mask");
ggml_set_input(local_mask);
}
struct ggml_tensor * pos_freqs = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, d_half);
ggml_set_name(pos_freqs, "pos_freqs");
ggml_set_input(pos_freqs);
struct ggml_tensor * rel_positions = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, window_size);
ggml_set_name(rel_positions, "rel_positions");
ggml_set_input(rel_positions);
struct ggml_tensor * freqs = ggml_repeat_4d(ctx0, pos_freqs, d_half, window_size, 1, 1);
struct ggml_tensor * theta = ggml_mul(ctx0, freqs, rel_positions);
struct ggml_tensor * sin = ggml_reshape_3d(ctx0, ggml_sin(ctx0, theta), 1, d_half, window_size);
struct ggml_tensor * cos = ggml_reshape_3d(ctx0, ggml_cos(ctx0, theta), 1, d_half, window_size);
struct ggml_tensor * pos_emb = ggml_reshape_2d(ctx0, ggml_cont(ctx0, ggml_concat(ctx0, sin, cos, 0)), n_state, window_size);
ggml_set_name(pos_emb, "pos_emb");
for (int il = 0; il < n_layer; ++il) {
const auto & layer = model.layers[il];
// FFN1
{
struct ggml_tensor * residual = cur;
ggml_format_name(cur, "enc_%d_res", il);
// norm
cur = ggml_norm(ctx0, cur, hparams.eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ff_norm_w), layer.ff_norm_b);
ggml_format_name(cur, "enc_%d_ffn_norm_1", il);
cur = build_ffn(cur, layer.ff_up_w, nullptr, nullptr, nullptr, layer.ff_down_w, nullptr, FFN_SILU, il);
ggml_format_name(cur, "enc_%d_ffn_1", il);
cur = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, fc_factor));
ggml_format_name(cur, "enc_%d_res_ffn", il);
}
// self attention block using relative positional encoding from model.position_embedding.
{
// [feat, time_frames, 1, 1]
struct ggml_tensor * residual = cur;
cur = ggml_norm(ctx0, cur, hparams.eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_1_w), layer.ln_1_b);
ggml_format_name(cur, "enc_%d_attn_norm", il);
const int n_head = hparams.n_head;
const int d_head = n_state / n_head;
// [feat, time_frames, 1, 1]
struct ggml_tensor * Q_cur = build_mm(layer.q_w, cur);
struct ggml_tensor * K_cur = build_mm(layer.k_w, cur);
struct ggml_tensor * V_cur = build_mm(layer.v_w, cur);
// [d_head, n_heads, n_time, 1]
Q_cur = ggml_reshape_3d(ctx0, Q_cur, d_head, n_head, n_time);
K_cur = ggml_reshape_3d(ctx0, K_cur, d_head, n_head, n_time);
V_cur = ggml_reshape_3d(ctx0, V_cur, d_head, n_head, n_time);
// [n_state, window_size]
struct ggml_tensor * pos = build_mm(layer.linear_pos_w, pos_emb);
// [feat, head, window_size, 1]
pos = ggml_reshape_3d(ctx0, pos, d_head, n_head, pos_emb->ne[1]);
// [feat, window_size, head, 1]
pos = ggml_cont(ctx0, ggml_permute(ctx0, pos, 0, 2, 1, 3));
ggml_format_name(pos, "enc_%d_attn_pos", il);
if (local_attn) {
const int chunk = att_left + att_right;
const int n_group = (n_time + chunk - 1) / chunk;
const int n_time_padded = n_group * chunk;
const int n_kv_chunk = chunk + window_size - 1;
const int n_kv_dense = n_kv_chunk * n_group;
const bool need_padding = n_time_padded > n_time;
Q_cur = ggml_cont(ctx0, ggml_permute(ctx0, Q_cur, 0, 2, 1, 3));
K_cur = ggml_cont(ctx0, ggml_permute(ctx0, K_cur, 0, 2, 1, 3));
V_cur = ggml_cont(ctx0, ggml_permute(ctx0, V_cur, 0, 2, 1, 3));
// content bias
struct ggml_tensor * bias_u = ggml_reshape_3d(ctx0, layer.pos_bias_u, d_head, 1, n_head);
struct ggml_tensor * Q_u = ggml_add(ctx0, Q_cur, bias_u);
// position bias
struct ggml_tensor * bias_v = ggml_reshape_3d(ctx0, layer.pos_bias_v, d_head, 1, n_head);
struct ggml_tensor * Q_v = ggml_add(ctx0, Q_cur, bias_v);
// right pad the time dimension
struct ggml_tensor * Q_u_padded = need_padding ?
ggml_pad_ext(ctx0, Q_u, 0, 0, 0, n_time_padded - n_time, 0, 0, 0, 0) : Q_u;
Q_u_padded = ggml_reshape_4d(ctx0, Q_u_padded, d_head, chunk, n_group, n_head);
// pad front and back for the first and last time frames
struct ggml_tensor * K_padded = ggml_pad_ext(ctx0, K_cur, 0, 0, att_left, att_right, 0, 0, 0, 0);
if (n_kv_dense > K_padded->ne[1]) {
K_padded = ggml_pad_ext(ctx0, K_padded, 0, 0, 0, n_kv_dense - K_padded->ne[1], 0, 0, 0, 0);
}
// sliding window view: each group spans n_kv_chunk keys but steps by chunk
struct ggml_tensor * K_chunk = ggml_view_4d(ctx0, K_padded,
d_head, n_kv_chunk, n_group, n_head,
K_padded->nb[1],
(size_t) chunk * K_padded->nb[1],
K_padded->nb[2],
0);
K_chunk = ggml_cont(ctx0, K_chunk);
struct ggml_tensor * content_scores = ggml_mul_mat(ctx0, K_chunk, Q_u_padded);
// trim the dense output down to window_size scores per query
content_scores = ggml_view_4d(ctx0, content_scores,
window_size, chunk, n_group, n_head,
(size_t) (chunk + window_size) * content_scores->nb[0],
content_scores->nb[2],
content_scores->nb[3],
0);
content_scores = ggml_cont(ctx0, content_scores);
// ungroup: [window_size, n_time_padded, n_head]
content_scores = ggml_reshape_3d(ctx0, content_scores, window_size, n_time_padded, n_head);
if (need_padding) {
content_scores = ggml_view_3d(ctx0, content_scores,
window_size, n_time, n_head,
content_scores->nb[1],
content_scores->nb[2],
0);
}
// Q_v: [d_head, time, head]
Q_v = ggml_cont(ctx0, ggml_permute(ctx0, Q_v, 0, 2, 1, 3));
struct ggml_tensor * rel_pos_scores = ggml_mul_mat(ctx0, pos, Q_v);
struct ggml_tensor * attn_scores = ggml_add(ctx0, content_scores, rel_pos_scores);
attn_scores = ggml_soft_max_ext(ctx0, attn_scores, attn_mask, 1.0f / std::sqrt(d_head), 0.0f);
ggml_format_name(attn_scores, "enc_%d_attn_probs", il);
// expand probs back to n_kv_chunk width for the V matmul
struct ggml_tensor * probs_padded = need_padding ?
ggml_pad_ext(ctx0, attn_scores, 0, 0, 0, n_time_padded - n_time, 0, 0, 0, 0) : attn_scores;
probs_padded = ggml_reshape_4d(ctx0, probs_padded, window_size, chunk, n_group, n_head);
probs_padded = ggml_pad_ext(ctx0, probs_padded, 0, chunk, 0, 0, 0, 0, 0, 0);
probs_padded = ggml_view_4d(ctx0, probs_padded,
n_kv_chunk, chunk, n_group, n_head,
(size_t) n_kv_chunk * probs_padded->nb[0],
probs_padded->nb[2],
probs_padded->nb[3],
0);
probs_padded = ggml_cont(ctx0, probs_padded);
probs_padded = ggml_mul(ctx0, probs_padded, local_mask);
struct ggml_tensor * V_padded = ggml_pad_ext(ctx0, V_cur, 0, 0, att_left, att_right, 0, 0, 0, 0);
if (n_kv_dense > V_padded->ne[1]) {
V_padded = ggml_pad_ext(ctx0, V_padded, 0, 0, 0, n_kv_dense - V_padded->ne[1], 0, 0, 0, 0);
}
V_padded = ggml_cont(ctx0, ggml_transpose(ctx0, V_padded));
struct ggml_tensor * V_chunk = ggml_view_4d(ctx0, V_padded,
n_kv_chunk, d_head, n_group, n_head,
V_padded->nb[1],
(size_t) chunk * V_padded->nb[0],
V_padded->nb[2],
0);
V_chunk = ggml_cont(ctx0, V_chunk);
cur = ggml_mul_mat(ctx0, V_chunk, probs_padded);
cur = ggml_reshape_3d(ctx0, cur, d_head, n_time_padded, n_head);
if (need_padding) {
cur = ggml_view_3d(ctx0, cur, d_head, n_time, n_head, cur->nb[1], cur->nb[2], 0);
}
cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));
cur = ggml_reshape_2d(ctx0, cur, n_state, n_time);
cur = build_mm(layer.o_w, cur);
} else {
// full attention
struct ggml_tensor * Q_u = ggml_add(ctx0, Q_cur, layer.pos_bias_u);
ggml_format_name(Q_u, "enc_%d_attn_q_u", il);
struct ggml_tensor * K_prep = ggml_permute(ctx0, K_cur, 0, 2, 1, 3);
struct ggml_tensor * Q_prep = ggml_permute(ctx0, Q_u, 0, 2, 1, 3);
struct ggml_tensor * content_scores = ggml_mul_mat(ctx0, K_prep, Q_prep);
ggml_format_name(content_scores, "enc_%d_attn_content_scores", il);
struct ggml_tensor * Q_v = ggml_add(ctx0, Q_cur, layer.pos_bias_v);
ggml_format_name(Q_v, "enc_%d_attn_q_v", il);
Q_v = ggml_permute(ctx0, Q_v, 0, 2, 1, 3);
Q_v = ggml_cont(ctx0, Q_v);
ggml_format_name(Q_v, "enc_%d_attn_q_v_perm", il);
struct ggml_tensor * rel_pos_scores = ggml_mul_mat(ctx0, pos, Q_v);
ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos", il);
// Relative positional shift
{
const auto pos_window = rel_pos_scores->ne[0];
const auto n_frame = rel_pos_scores->ne[1];
const auto n_head = rel_pos_scores->ne[2];
rel_pos_scores = ggml_pad(ctx0, rel_pos_scores, 1, 0, 0, 0);
rel_pos_scores = ggml_roll(ctx0, rel_pos_scores, 1, 0, 0, 0);
rel_pos_scores = ggml_reshape_3d(ctx0, rel_pos_scores, n_frame, pos_window + 1, n_head);
rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_reshaped", il);
int center = pos_window / 2;
size_t offset = rel_pos_scores->nb[0] * (center+1);
rel_pos_scores = ggml_view_3d(ctx0, rel_pos_scores,
n_frame, pos_window, n_head,
(pos_window) * 4,
rel_pos_scores->nb[2],
offset);
rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_shifted", il);
rel_pos_scores = ggml_view_3d(ctx0, rel_pos_scores,
content_scores->ne[0],
content_scores->ne[1],
rel_pos_scores->ne[2],
rel_pos_scores->nb[1],
rel_pos_scores->nb[2],
0);
rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
ggml_format_name(rel_pos_scores, "enc_%d_attn_rel_pos_shifted_view", il);
}
struct ggml_tensor * attn_scores = ggml_add(ctx0, content_scores, rel_pos_scores);
ggml_format_name(attn_scores, "enc_%d_attn_scores", il);
attn_scores = ggml_scale(ctx0, attn_scores, 1.0f / std::sqrt(d_head));
attn_scores = ggml_add(ctx0, attn_scores, attn_mask);
ggml_format_name(attn_scores, "enc_%d_attn_scores_scaled", il);
struct ggml_tensor * probs = ggml_soft_max(ctx0, attn_scores);
ggml_format_name(probs, "enc_%d_attn_probs", il);
V_cur = ggml_cont(ctx0, ggml_permute(ctx0, V_cur, 1, 2, 0, 3));
ggml_format_name(V_cur, "enc_%d_attn_v_cur", il);
cur = ggml_mul_mat(ctx0, probs, V_cur);
ggml_format_name(cur, "enc_%d_attn_inp", il);
cur = ggml_permute(ctx0, cur, 2, 0, 1, 3);
cur = ggml_cont_2d(ctx0, cur, n_state, n_time);
cur = build_mm(layer.o_w, cur);
}
ggml_format_name(cur, "enc_%d_attn_out", il);
cur = ggml_add(ctx0, residual, cur);
ggml_format_name(cur, "enc_%d_attn_res", il);
}
// Convolution
{
struct ggml_tensor * residual = cur;
ggml_format_name(cur, "enc_%d_residual_conv", il);
cur = ggml_norm(ctx0, cur, hparams.eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.norm_conv_w), layer.norm_conv_b);
ggml_format_name(cur, "enc_%d_norm_conv", il);
// pointwise 1d convolution:
cur = build_mm(layer.conv_pw1_w, cur);
ggml_format_name(cur, "enc_%d_conv_pw1", il);
{
int64_t d = cur->ne[0] / 2;
struct ggml_tensor * signal = ggml_view_2d(ctx0, cur, d, cur->ne[1], cur->nb[1], 0);
struct ggml_tensor * gate = ggml_view_2d(ctx0, cur, d, cur->ne[1], cur->nb[1], d * cur->nb[0]);
cur = ggml_mul(ctx0, signal, ggml_sigmoid(ctx0, gate));
ggml_format_name(cur, "enc_%d_conv_glu", il);
}
cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
// use ggml_ssm_conv for f32 precision
const int dw_pad = (hparams.audio_conv_kernel_size - 1) / 2;
cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
cur = ggml_roll(ctx0, cur, dw_pad, 0, 0, 0);
cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
ggml_format_name(cur, "enc_%d_conv_dw_pad", il);
cur = ggml_ssm_conv(ctx0, cur, layer.conv_dw_w);
ggml_format_name(cur, "enc_%d_conv_1d_dw", il);
cur = ggml_sub(ctx0, cur, layer.conv_norm_mean);
struct ggml_tensor * std = ggml_sqrt(ctx0, layer.conv_norm_var);
cur = ggml_div(ctx0, cur, std);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.conv_norm_w), layer.conv_norm_b);
ggml_format_name(cur, "enc_%d_conv_bn", il);
cur = ggml_silu(ctx0, cur);
ggml_format_name(cur, "enc_%d_conv_silu", il);
cur = build_mm(layer.conv_pw2_w, cur);
ggml_format_name(cur, "enc_%d_conv_pw2", il);
cur = ggml_add(ctx0, residual, cur);
ggml_format_name(cur, "enc_%d_conv_res", il);
}
// FFN2
{
struct ggml_tensor * residual = cur;
cur = ggml_norm(ctx0, cur, hparams.eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ff_norm_1_w), layer.ff_norm_1_b);
ggml_format_name(cur, "enc_%d_ffn_norm_2", il);
cur = build_ffn(cur, layer.ff_up_1_w, nullptr, nullptr, nullptr, layer.ff_down_1_w, nullptr, FFN_SILU, il);
cur = ggml_add(ctx0, residual, ggml_scale(ctx0, cur, 0.5));
ggml_format_name(cur, "enc_%d_ffn_res", il);
}
cur = ggml_norm(ctx0, cur, hparams.eps);
cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.ln_2_w), layer.ln_2_b);
}
cb(cur, "encoder_out", -1);
cur = ggml_rms_norm(ctx0, cur, 1e-6);
cur = ggml_mul(ctx0, cur, model.mm_norm_pre_w);
cb(cur, "sound_projection.norm", -1);
cur = build_ffn(cur, model.mm_0_w, model.mm_0_b, nullptr, nullptr, model.mm_1_w, model.mm_1_b, FFN_RELU_SQR, -1);
cb(cur, "projected", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
tools/mtmd/mtmd-audio.cpp
+269
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
//
@@ -956,6 +1022,209 @@ bool mtmd_audio_preprocessor_gemma4a::preprocess(const float * s
}
//
// mtmd_audio_preprocessor_parakeet implementation
//
void mtmd_audio_preprocessor_parakeet::worker_thread(
int ith,
const float * window_func,
int window_size,
const std::vector<float> & samples,
int n_samples,
int frame_size,
int frame_step,
int n_threads,
int n_fft_bins,
const mtmd_audio_cache & cache,
mtmd_audio_mel & mel) {
std::vector<float> fft_in(frame_size * 2, 0.0);
std::vector<float> fft_out(frame_size * 2 * 2 * 2);
int n_fb = n_fft_bins;
int i = ith;
GGML_ASSERT(n_fb == 1 + (frame_size / 2));
const double eps = 5.960464477539063e-08;
for (; i < std::min(n_samples / frame_step + 1, (int) mel.n_len); i += n_threads) {
const int offset = i * frame_step;
const int window_pad_left = (frame_size - window_size) / 2;
// Zero-pad left.
std::fill(fft_in.begin(), fft_in.begin() + window_pad_left, 0.0f);
// Apply windowed samples in the center.
const int n_to_process = std::min({window_size, n_samples - offset});
for (int j = 0; j < n_to_process; j++) {
fft_in[window_pad_left + j] = window_func[j] * samples[offset + window_pad_left + j];
}
// Zero-pad right.
std::fill(fft_in.begin() + window_pad_left + n_to_process, fft_in.begin() + frame_size, 0.0f);
// FFT.
fft(cache, fft_in.data(), frame_size, fft_out.data());
// Calculate modulus^2 of complex numbers.
for (int j = 0; j < n_fb; j++) {
fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
}
// mel spectrogram.
for (int j = 0; j < mel.n_mel; j++) {
double sum = 0.0;
int k = 0;
for (k = 0; k < n_fb - 3; k += 4) {
sum +=
fft_out[k + 0] * cache.filters.data[j * n_fb + k + 0] +
fft_out[k + 1] * cache.filters.data[j * n_fb + k + 1] +
fft_out[k + 2] * cache.filters.data[j * n_fb + k + 2] +
fft_out[k + 3] * cache.filters.data[j * n_fb + k + 3];
}
for (; k < n_fb; k++) {
sum += fft_out[k] * cache.filters.data[j * n_fb + k];
}
mel.data[j * mel.n_len + i] = std::log(sum + eps);
}
}
// Otherwise fft_out are all zero.
const double empty_sum = std::log(eps);
for (; i < mel.n_len; i += n_threads) {
for (int j = 0; j < mel.n_mel; j++) {
mel.data[j * mel.n_len + i] = empty_sum;
}
}
}
void mtmd_audio_preprocessor_parakeet::initialize() {
cache.fill_sin_cos_table(hparams.audio_n_fft);
const size_t n_fft = hparams.audio_n_fft / 2 + 1;
GGML_ASSERT(hparams.mel_filters.size() == (size_t)hparams.n_mel_bins * n_fft);
cache.filters.n_mel = hparams.n_mel_bins;
cache.filters.n_fft = n_fft;
cache.filters.data = hparams.mel_filters;
GGML_ASSERT(hparams.window.size() == (size_t)hparams.audio_window_len);
GGML_ASSERT(hparams.window.size() <= (size_t) hparams.audio_n_fft);
cache.hann_window = hparams.window;
}
bool mtmd_audio_preprocessor_parakeet::preprocess(const float * samples,
size_t n_samples_in,
std::vector<mtmd_audio_mel> & output) {
if (n_samples_in == 0) {
return false;
}
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;
GGML_ASSERT(!cache.sin_vals.empty());
GGML_ASSERT(!cache.cos_vals.empty());
GGML_ASSERT(!cache.filters.data.empty());
const float * window_func = cache.hann_window.data();
const int window_size = params.hann_window_size;
const int frame_size = (params.n_fft_bins - 1) * 2;
const int frame_step = params.hop_length;
// Apply preemphasis filter (high-pass): x[i] = x[i] - 0.97 * x[i-1]
std::vector<float> samples_preprocessed(samples, samples + n_samples_in);
{
const float preemph = 0.97f;
for (int i = n_samples_in - 1; i > 0; i--) {
samples_preprocessed[i] = samples_preprocessed[i] - preemph * samples_preprocessed[i - 1];
}
}
// Parakeet uses centered constant padding
const size_t pad = (size_t)(frame_size / 2);
std::vector<float> samples_padded(n_samples_in + 2 * pad, 0.0f);
std::copy(samples_preprocessed.begin(), samples_preprocessed.end(), samples_padded.begin() + pad);
mtmd_audio_mel out_full;
out_full.n_mel = params.n_mel;
out_full.n_len = (samples_padded.size() - frame_size) / frame_step + 1;
out_full.n_len_org = out_full.n_len;
out_full.data.resize(out_full.n_mel * out_full.n_len);
const int n_threads = 4;
std::vector<std::thread> workers(n_threads - 1);
for (int iw = 0; iw < n_threads - 1; ++iw) {
workers[iw] = std::thread(
worker_thread, iw + 1,
window_func,
window_size,
std::cref(samples_padded),
samples_padded.size(),
frame_size,
frame_step,
n_threads,
params.n_fft_bins,
std::cref(cache),
std::ref(out_full)
);
}
worker_thread(0,
window_func,
window_size,
samples_padded,
samples_padded.size(),
frame_size,
frame_step,
n_threads,
params.n_fft_bins,
cache,
out_full);
for (int iw = 0; iw < n_threads - 1; ++iw) {
workers[iw].join();
}
// Per-feature normalization (only on valid frames)
{
const double eps = 1e-5;
int valid_frames = n_samples_in / frame_step;
for (int j = 0; j < out_full.n_mel; j++) {
double sum = 0.0;
double sq_diff_sum = 0.0;
// Calculate Mean ONLY on valid audio frames
for (int i = 0; i < valid_frames; i++) {
sum += (double)out_full.data[j * out_full.n_len + i];
}
double mean = sum / valid_frames;
// Calculate Variance ONLY on valid audio frames
for (int i = 0; i < valid_frames; i++) {
double diff = (double)out_full.data[j * out_full.n_len + i] - mean;
sq_diff_sum += diff * diff;
}
double std_dev = std::sqrt(sq_diff_sum / (valid_frames - 1.0));
double denominator = std_dev + eps;
// Apply to ALL frames (including the padded ones)
for (int i = 0; i < out_full.n_len; i++) {
out_full.data[j * out_full.n_len + i] = (float)((out_full.data[j * out_full.n_len + i] - mean) / denominator);
}
}
}
output.push_back(std::move(out_full));
return true;
}
// mtmd_audio_preprocessor_gemma4ua
//
tools/mtmd/mtmd-audio.h
+24
View File
@@ -111,6 +111,30 @@ 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;
};
struct mtmd_audio_preprocessor_parakeet : mtmd_audio_preprocessor {
mtmd_audio_preprocessor_parakeet(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;
static void worker_thread(int ith, const float * window_func, int window_size,
const std::vector<float> & samples, int n_samples,
int frame_size, int frame_step, int n_threads,
int n_fft_bins,
const mtmd_audio_cache & cache, mtmd_audio_mel & mel);
};
//
// streaming ISTFT - converts spectrogram frames back to audio one frame at a time
//
tools/mtmd/mtmd.cpp
+10
View File
@@ -724,12 +724,22 @@ struct mtmd_context {
aud_end = "<audio|>";
audio_preproc = std::make_unique<mtmd_audio_preprocessor_gemma4a>(ctx_a);
} break;
case PROJECTOR_TYPE_PARAKEET:
{
audio_preproc = std::make_unique<mtmd_audio_preprocessor_parakeet>(ctx_a);
} break;
case PROJECTOR_TYPE_GEMMA4UA:
{
aud_beg = "<|audio>";
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));
}