vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Apply suggestions from code review

Browse source
This commit is contained in:
Xuan-Son Nguyen 2025-01-30 16:17:16 +01:00 • committed by GitHub
parent 0d3ad163f0
commit 0536d004cb
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 16 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

17
examples/llava/clip.cpp
View file

@ -2177,7 +2177,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
return true; return true;
} }
if(ctx->has_glm_projector){ if (ctx->has_glm_projector) {
res_imgs->size = 1; res_imgs->size = 1;
res_imgs->data = new clip_image_f32[res_imgs->size]; res_imgs->data = new clip_image_f32[res_imgs->size];
clip_image_u8 resized_image; clip_image_u8 resized_image;
@ -2376,9 +2376,8 @@ void clip_free(clip_ctx * ctx) {
} }
size_t clip_embd_nbytes(const struct clip_ctx * ctx) { size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
if(ctx->has_glm_projector) int extra_tokens = ctx->has_glm_projector ? 2 : 0;
return (clip_n_patches(ctx)+2) * clip_n_mmproj_embd(ctx) * sizeof(float); return (clip_n_patches(ctx) + extra_tokens) * clip_n_mmproj_embd(ctx) * sizeof(float);
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
} }
size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) { size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
@ -2553,11 +2552,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
if (ctx->has_minicpmv_projector) { if (ctx->has_minicpmv_projector) {
GGML_ASSERT(batch_size == 1); GGML_ASSERT(batch_size == 1);
} }
if(ctx->has_glm_projector) { if (ctx->has_glm_projector) {
GGML_ASSERT(batch_size == 1); GGML_ASSERT(batch_size == 1);
ggml_tensor * boi = ctx->vision_model.boi_w; ggml_tensor * boi = ctx->vision_model.boi_w;
ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi)); ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi));
vec=(float*)(vec+ggml_nelements(boi)); //offset for boi vec = (float*)(vec+ggml_nelements(boi)); //offset for boi
} }
// build the inference graph // build the inference graph
@ -2711,7 +2710,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
free(positions_data); free(positions_data);
if (!ctx->has_glm_projector){ if (!ctx->has_glm_projector) {
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
int* patches_data = (int*)malloc(ggml_nbytes(patches)); int* patches_data = (int*)malloc(ggml_nbytes(patches));
for (int i = 0; i < num_patches; i++) { for (int i = 0; i < num_patches; i++) {
@ -2735,11 +2734,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// copy the embeddings to the location passed by the user // copy the embeddings to the location passed by the user
ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings)); ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
if(ctx->has_glm_projector){ if (ctx->has_glm_projector) {
//eoi //eoi
ggml_tensor * eoi = ctx->vision_model.eoi_w; ggml_tensor * eoi = ctx->vision_model.eoi_w;
int offset = ggml_nelements(embeddings); int offset = ggml_nelements(embeddings);
ggml_backend_tensor_get(eoi,vec+offset,0,ggml_nbytes(eoi)); ggml_backend_tensor_get(eoi, vec+offset, 0, ggml_nbytes(eoi));
} }
return true; return true;

10
src/llama-model.cpp
View file

@ -1093,20 +1093,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
{ {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) { switch (hparams.n_layer) {
case 28:{ case 28: {
if(hparams.n_head(0)==16){ if (hparams.n_head(0)==16) {
type = LLM_TYPE_1_5B; type = LLM_TYPE_1_5B;
}else{ } else {
type = LLM_TYPE_6B; type = LLM_TYPE_6B;
} }
} break; } break;
case 40: { case 40: {
if(hparams.n_head(0)==24){ if(hparams.n_head(0)==24){
type = LLM_TYPE_4B; type = LLM_TYPE_4B;
}else{ } else {
type = LLM_TYPE_9B; type = LLM_TYPE_9B;
} }
}break; } break;
default: type = LLM_TYPE_UNKNOWN; default: type = LLM_TYPE_UNKNOWN;
} }
} break; } break;

6
src/llama.cpp
View file

@ -7215,7 +7215,7 @@ struct llm_build_context {
struct ggml_tensor * Qcur = nullptr; struct ggml_tensor * Qcur = nullptr;
struct ggml_tensor * Kcur = nullptr; struct ggml_tensor * Kcur = nullptr;
struct ggml_tensor * Vcur = nullptr; struct ggml_tensor * Vcur = nullptr;
if(model.type == LLM_TYPE_1_5B|| model.type == LLM_TYPE_4B || model.type == LLM_TYPE_9B) { if (model.type == LLM_TYPE_1_5B || model.type == LLM_TYPE_4B || model.type == LLM_TYPE_9B) {
Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il); cb(Qcur, "Qcur", il);
if (model.layers[il].bq) { if (model.layers[il].bq) {
@ -7234,10 +7234,10 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il); cb(Vcur, "Vcur", il);
} }
}else{ } else {
cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il); cb(cur, "wqkv", il);
if(model.layers[il].bqkv){ if (model.layers[il].bqkv) {
cur = ggml_add(ctx0, cur, model.layers[il].bqkv); cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il); cb(cur, "bqkv", il);
} }