sleepy/llama.cpp
1
0
Fork 0
Code Issues 11 Pull Requests Actions 107 Packages Projects Releases Wiki Activity

model : add glm-asr support (#17901)

Browse Source 
* [model] add glm-asr support

* fix format for ci

* fix convert format for ci

* update glm_asr convert script & use build_ffn for glm_asr clip & use build_stack for padding and review

* check root architecture for convert hf script

* fix conficlt with upstream

* fix convert script for glm asr & format clip-impl

* format

* restore hparams text

* improved conversion

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
piDack
2025-12-15 10:18:46 +08:00
committed by GitHub
parent 52392291b2
commit 745fa0e78b
7 changed files with 160 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip.cpp
+59 -1
View File
@@ -720,6 +720,32 @@ ggml_tensor * clip_graph::build_rope_2d(
return cur;
}
// Generic function to stack frames for audio processing
// Abstracts out the StackAudioFrames logic used by ultravox
ggml_tensor * clip_graph::build_stack(ggml_tensor * cur, int32_t stack_factor, int32_t n_embed) {
if (stack_factor <= 1) {
return cur;
}
int64_t total_elements = ggml_nelements(cur);
int64_t stride = n_embed * stack_factor;
// Calculate padded length
int64_t padded_len = GGML_PAD(total_elements, stride);
int64_t pad = padded_len - total_elements;
if (pad > 0) {
// Pad the tensor to make it divisible by stride
cur = ggml_view_1d(ctx0, cur, total_elements, 0);
cur = ggml_pad(ctx0, cur, pad, 0, 0, 0);
}
// Reshape to [stride, padded_len / stride]
cur = ggml_view_2d(ctx0, cur, stride, padded_len / stride,
ggml_row_size(cur->type, stride), 0);
return cur;
}
// aka pixel_shuffle / pixel_unshuffle / patch_merger (Kimi-VL)
// support dynamic resolution
ggml_tensor * clip_graph::build_patch_merge_permute(ggml_tensor * cur, int scale_factor) {
@@ -796,6 +822,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_VOXTRAL:
case PROJECTOR_TYPE_QWEN2A:
case PROJECTOR_TYPE_GLMA:
{
builder = std::make_unique<clip_graph_whisper_enc>(ctx, img);
} break;
@@ -1136,10 +1163,12 @@ struct clip_model_loader {
} break;
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_QWEN2A:
case PROJECTOR_TYPE_GLMA:
case PROJECTOR_TYPE_VOXTRAL:
{
bool require_stack = model.proj_type == PROJECTOR_TYPE_ULTRAVOX ||
model.proj_type == PROJECTOR_TYPE_VOXTRAL;
model.proj_type == PROJECTOR_TYPE_VOXTRAL ||
model.proj_type == PROJECTOR_TYPE_GLMA;
get_u32(KEY_A_PROJ_STACK_FACTOR, hparams.proj_stack_factor, require_stack);
if (hparams.n_mel_bins != 128) {
throw std::runtime_error(string_format("%s: only 128 mel bins are supported for ultravox\n", __func__));
@@ -1510,6 +1539,21 @@ struct clip_model_loader {
model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight"));
model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias"));
} break;
case PROJECTOR_TYPE_GLMA:
{
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.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_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"));
model.mm_norm_pre_w = get_tensor(string_format(TN_MM_NORM_PRE, "weight"));
model.mm_norm_pre_b = get_tensor(string_format(TN_MM_NORM_PRE, "bias"));
model.mm_boi = get_tensor(string_format(TN_TOK_BOI, "weight"));
model.mm_eoi = get_tensor(string_format(TN_TOK_EOI, "weight"));
} break;
case PROJECTOR_TYPE_LLAMA4:
{
model.mm_model_proj = get_tensor(TN_MM_PROJECTOR);
@@ -2895,6 +2939,16 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
n_patches /= 2;
}
} break;
case PROJECTOR_TYPE_GLMA:
{
n_patches = img->nx;
// whisper downscales input token by half after conv1d
n_patches /= 2;
// reshape by merge_factor
n_patches /= ctx->model.hparams.proj_stack_factor;
// for BOI and EOI token embeddings
n_patches += 2;
} break;
case PROJECTOR_TYPE_COGVLM:
{
n_patches += 2; // for BOI and EOI token embeddings
@@ -3230,6 +3284,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
case PROJECTOR_TYPE_IDEFICS3:
case PROJECTOR_TYPE_INTERNVL:
case PROJECTOR_TYPE_QWEN2A:
case PROJECTOR_TYPE_GLMA:
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_VOXTRAL:
@@ -3340,6 +3395,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return ctx->model.mm_model_proj->ne[1];
case PROJECTOR_TYPE_QWEN2A:
return ctx->model.mm_fc_w->ne[1];
case PROJECTOR_TYPE_GLMA:
return ctx->model.mm_2_w->ne[1];
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_KIMIVL:
return ctx->model.mm_2_w->ne[1];
@@ -3386,6 +3443,7 @@ bool clip_has_audio_encoder(const struct clip_ctx * ctx) {
bool clip_has_whisper_encoder(const struct clip_ctx * ctx) {
return ctx->proj_type() == PROJECTOR_TYPE_ULTRAVOX
|| ctx->proj_type() == PROJECTOR_TYPE_QWEN2A
|| ctx->proj_type() == PROJECTOR_TYPE_GLMA
|| ctx->proj_type() == PROJECTOR_TYPE_VOXTRAL;
}