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

fix format

Browse source
This commit is contained in:
liyuhang 2025-01-30 09:19:15 +08:00
parent a7054a11a9
commit 13961b3553
3 changed files with 12 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

14
examples/llava/clip.cpp
View file

@ -177,7 +177,7 @@ enum projector_type {
PROJECTOR_TYPE_LDP,
PROJECTOR_TYPE_LDPV2,
PROJECTOR_TYPE_RESAMPLER,
PROJECTOR_TYPE_ADAPTER,
PROJECTOR_TYPE_GLM_EDGE,
PROJECTOR_TYPE_MERGER,
PROJECTOR_TYPE_UNKNOWN,
};
@ -187,7 +187,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_LDPV2, "ldpv2"},
{ PROJECTOR_TYPE_RESAMPLER, "resampler"},
{ PROJECTOR_TYPE_ADAPTER, "adapter"},
{ PROJECTOR_TYPE_GLM_EDGE, "adapter"},
{ PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
};
@ -1115,7 +1115,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
}
// glm projector
else if(ctx->has_glm_projector){
if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE){
size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
embeddings = ggml_reshape_3d(ctx0,embeddings,gridsz,gridsz,embeddings->ne[1]);
@ -1625,7 +1625,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
}
else if(new_clip->proj_type == PROJECTOR_TYPE_ADAPTER){
else if(new_clip->proj_type == PROJECTOR_TYPE_GLM_EDGE){
vision_model.mm_model_adapter_conv_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "weight"));
vision_model.mm_model_adapter_conv_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "bias"));
vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_LINEAR,"weight"));
@ -2420,7 +2420,7 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i
int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_ADAPTER) {
if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) {
n_patches /= 4;
} else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
if (ctx->minicpmv_version == 2) {
@ -2738,7 +2738,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
if(ctx->has_glm_projector){
//eoi
ggml_tensor * eoi = ctx->vision_model.eoi_w;
int offset=ggml_nelements(eoi)*clip_n_patches(ctx);
int offset = ggml_nelements(embeddings);
ggml_backend_tensor_get(eoi,vec+offset,0,ggml_nbytes(eoi));
}
@ -2903,7 +2903,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return 3584;
}
}
if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE){
return ctx->vision_model.mm_model_mlp_3_w->ne[1];
}
if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {

6
src/llama-arch.cpp
View file

@ -1024,9 +1024,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },

6
src/llama-chat.cpp
View file

@ -115,8 +115,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
}
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
return LLM_CHAT_TEMPLATE_PHI_3;
} else if (tmpl_contains("\n<|assistant|>") && tmpl_contains("<|user|>")) {
return LLM_CHAT_TEMPLATE_FALCON_3;
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
return LLM_CHAT_TEMPLATE_ZEPHYR;
} else if (tmpl_contains("bos_token + message['role']")) {
@ -148,8 +148,6 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_CHATGML_3;
} else if (tmpl_contains("[gMASK]<sop>")) {
return LLM_CHAT_TEMPLATE_CHATGML_4;
} else if (tmpl_contains("<|user|>") && tmpl_contains("<|assistant|>") && !tmpl_contains("<|end|>") && !tmpl_contains("</s>")) {
return LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains(LU8("<用户>"))) {
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
return LLM_CHAT_TEMPLATE_MINICPM;