Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '763d06edb7dd5094ea58bad1d81e2e8d35033e34' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build-linux-cross.yml
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cann/CMakeLists.txt
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-vulkan/CMakeLists.txt
#	tools/mtmd/CMakeLists.txt
#	tools/mtmd/clip.cpp
#	tools/mtmd/mtmd.cpp
#	tools/server/CMakeLists.txt
This commit is contained in:
Concedo
2025-05-31 12:44:18 +08:00
parent 0c108f6054 763d06edb7
commit c987abf9f5
28 changed files with 777 additions and 291 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+18 -13
View File
@@ -455,7 +455,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
}
int64_t llm_graph_context::n_pos_per_embd() const {
return arch == LLM_ARCH_QWEN2VL ? 4 : 1;
return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
}
void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
@@ -1562,20 +1562,25 @@ void llm_graph_context::build_pooling(
ggml_tensor * inp_cls = build_inp_cls();
inp = ggml_get_rows(ctx0, inp, inp_cls);
// classification head
// https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
GGML_ASSERT(cls != nullptr);
GGML_ASSERT(cls_b != nullptr);
if (cls != nullptr && cls_b != nullptr) {
// classification head
// https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
cur = ggml_tanh(ctx0, cur);
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
cur = ggml_tanh(ctx0, cur);
// some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
// https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
if (cls_out) {
// some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
// https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
if (cls_out) {
GGML_ASSERT(cls_out_b != nullptr);
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
}
} else if (cls_out) {
// Single layer classification head (direct projection)
// https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
GGML_ASSERT(cls_out_b != nullptr);
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
} else {
GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
}
} break;
default:
src/llama-kv-cache.cpp
+10 -2
View File
@@ -757,11 +757,19 @@ ggml_tensor * llama_kv_cache_unified::build_rope_shift(
const auto & yarn_beta_slow = cparams.yarn_beta_slow;
const auto & n_rot = hparams.n_rot;
const auto & rope_type = hparams.rope_type;
const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE
// @ngxson : this is a workaround
// for M-RoPE, we want to rotate the whole vector when doing KV shift
// a normal RoPE should work, we just need to use the correct ordering
// ref: https://github.com/ggml-org/llama.cpp/pull/13870
? LLAMA_ROPE_TYPE_NEOX
: hparams.rope_type;
// See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2
? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale))
: cparams.yarn_attn_factor;
ggml_tensor * tmp;