sleepy/llama.cpp
1
0
Fork 0
Code Issues 11 Pull Requests Actions 114 Packages Projects Releases Wiki Activity

model : add ASR support for LFM2-Audio-1.5B (conformer) (#18106)

Browse Source 
* ASR with LFM2-Audio-1.5B

* Set rope_theta

* Fix comment

* Remove rope_theta setting

* Address PR feedback

* rename functions to conformer

* remove some redundant ggml_cont

* fix missing tensor

* add prefix "a." for conv tensors

* remove redundant reshape

* clean up

* add test model

---------

Co-authored-by: Tarek Dakhran <tarek@liquid.ai>
This commit is contained in:
Xuan-Son Nguyen
2025-12-19 00:18:01 +01:00
committed by GitHub
parent f9ec8858ed
commit 8ea958d4d9
17 changed files with 669 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip.cpp
+86
View File
@@ -837,6 +837,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
{
builder = std::make_unique<clip_graph_llava>(ctx, img);
} break;
case PROJECTOR_TYPE_LFM2A:
{
builder = std::make_unique<clip_graph_conformer>(ctx, img);
} break;
case PROJECTOR_TYPE_GLM4V:
{
builder = std::make_unique<clip_graph_glm4v>(ctx, img);
@@ -1187,6 +1191,15 @@ struct clip_model_loader {
hparams.audio_window_len = 400;
hparams.audio_hop_len = 160;
} break;
case PROJECTOR_TYPE_LFM2A:
{
// audio preprocessing params
hparams.audio_chunk_len = 1; // in seconds
hparams.audio_sample_rate = 16000;
hparams.audio_n_fft = 512;
hparams.audio_window_len = 400;
hparams.audio_hop_len = 160;
} break;
default:
break;
}
@@ -1611,6 +1624,52 @@ struct clip_model_loader {
model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
} break;
case PROJECTOR_TYPE_LFM2A:
{
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"));
model.mm_0_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 0, "weight"));
model.mm_0_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 0, "bias"));
model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
model.mm_1_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "bias"));
model.mm_3_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 3, "weight"));
model.mm_3_b = get_tensor(string_format(TN_MM_AUDIO_MLP, 3, "bias"));
for (int il = 0; il < hparams.n_layer; ++il) {
auto & layer = model.layers[il];
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"));
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"));
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));
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"));
layer.linear_pos_w = get_tensor(string_format(TN_LINEAR_POS, prefix, il, "weight"));
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_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"));
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"));
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"));
}
} break;
default:
GGML_ASSERT(false && "unknown projector type");
}
@@ -3004,6 +3063,10 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
{
n_patches += 2; // for BOI and EOI token embeddings
} break;
case PROJECTOR_TYPE_LFM2A:
{
n_patches = ((((img->nx + 1) / 2) + 1) / 2 + 1) / 2;
} break;
default:
GGML_ABORT("unsupported projector type");
}
@@ -3362,6 +3425,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
}
set_input_i32("pos_w", pos_data);
} break;
case PROJECTOR_TYPE_LFM2A:
{
GGML_ASSERT(imgs.entries.size() == 1);
const auto n_frames = clip_n_output_tokens(ctx, imgs.entries.front().get());
auto d_model = 512;
auto seq_len = n_frames * 2 - 1;
std::vector<float> pos_emb(d_model*seq_len);
std::vector<double> inv_freq(d_model / 2);
for (size_t i = 0; i < inv_freq.size(); ++i) {
inv_freq[i] = std::exp(-(std::log(10000.0) / (float)d_model) * (2.0f * (float)(i)));
}
for (int64_t pos = 0; pos < seq_len; ++pos) {
for (size_t i = 0; i < inv_freq.size(); ++i) {
const float ang = (n_frames - pos - 1) * inv_freq[i];
pos_emb[pos*d_model + 2*i + 0] = sinf(ang); // even
pos_emb[pos*d_model + 2*i + 1] = cosf(ang); // odd
}
}
set_input_f32("pos_emb", pos_emb);
} break;
default:
GGML_ABORT("Unknown projector type");
}
@@ -3456,6 +3540,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return ctx->model.mm_2_w->ne[1];
case PROJECTOR_TYPE_COGVLM:
return ctx->model.mm_4h_to_h_w->ne[1];
case PROJECTOR_TYPE_LFM2A:
return ctx->model.position_embeddings->ne[0];
case PROJECTOR_TYPE_GLM4V:
return ctx->model.mm_ffn_down_w->ne[1];
default: