Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	common/common.cpp
#	ggml/CMakeLists.txt
#	scripts/sync-ggml.last
#	scripts/sync_vendor.py
#	src/llama-context.cpp
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tools/cli/README.md
#	tools/completion/README.md
#	tools/server/README.md
This commit is contained in:
Concedo
2026-05-17 11:26:18 +08:00
parent 9d38a9edc0 0253fb21f5
commit 1e828ccabf
66 changed files with 2205 additions and 504 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+27 -3
View File
@@ -2078,6 +2078,12 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
// checks
default:
{
// The MTP head is dense-attention only on hybrid Qwen3.5/3.6, so use a plain
// attention KV cache for the MTP context instead of the hybrid wrapper.
const bool mtp_on_hybrid_qwen35 =
params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
(arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE);
if (llm_arch_is_recurrent(arch)) {
res = new llama_memory_recurrent(
*this,
@@ -2086,8 +2092,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.offload_kqv,
std::max((uint32_t) 1, cparams.n_seq_max),
cparams.n_seq_max,
cparams.n_rs_seq,
nullptr);
} else if (llm_arch_is_hybrid(arch)) {
} else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen35) {
// The main difference between hybrid architectures is the
// layer filters, so pick the right one here
llama_memory_hybrid::layer_filter_cb filter_attn = nullptr;
@@ -2102,6 +2109,14 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
filter_recr = [&](int32_t il) {
return hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
};
} else if (arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE) {
const uint32_t n_main = hparams.n_layer - hparams.nextn_predict_layers;
filter_attn = [&, n_main](int32_t il) {
return (uint32_t)il < n_main && !hparams.is_recurrent(il);
};
filter_recr = [&, n_main](int32_t il) {
return (uint32_t)il < n_main && hparams.is_recurrent(il);
};
}
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
@@ -2119,6 +2134,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
/* recurrent_type_s */ GGML_TYPE_F32,
/* recurrent_rs_size */ std::max((uint32_t) 1, cparams.n_seq_max),
/* n_seq_max */ cparams.n_seq_max,
/* n_rs_seq */ cparams.n_rs_seq,
/* offload */ cparams.offload_kqv,
/* unified */ cparams.kv_unified,
/* filter_attn */ std::move(filter_attn),
@@ -2137,6 +2153,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
/* recurrent_type_v */ GGML_TYPE_F32,
/* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
/* n_seq_max */ cparams.n_seq_max,
/* n_rs_seq */ cparams.n_rs_seq,
/* offload */ cparams.offload_kqv,
/* unified */ cparams.kv_unified,
/* filter_attn */ std::move(filter_attn),
@@ -2144,6 +2161,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
}
} else {
llama_memory_i::layer_reuse_cb reuse = nullptr;
llama_kv_cache::layer_filter_cb filter = nullptr;
if (arch == LLM_ARCH_GEMMA3N || arch == LLM_ARCH_GEMMA4) {
reuse = [&](int32_t il) {
@@ -2155,6 +2173,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
};
}
if (mtp_on_hybrid_qwen35) {
const uint32_t n_main = hparams.n_layer - hparams.nextn_predict_layers;
filter = [n_main](int32_t il) { return (uint32_t)il >= n_main; };
}
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
GGML_ASSERT(hparams.is_swa_any());
@@ -2170,7 +2193,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.n_seq_max,
cparams.n_ubatch,
1,
nullptr,
filter,
reuse);
} else {
GGML_ASSERT(!hparams.is_swa_any());
@@ -2187,7 +2210,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
1,
hparams.n_swa,
hparams.swa_type,
nullptr,
filter,
nullptr);
}
}
@@ -2290,6 +2313,7 @@ int32_t llama_model_n_swa(const llama_model * model) {
return model->hparams.n_swa;
}
uint32_t llama_model_n_cls_out(const struct llama_model * model) {
return model->hparams.n_cls_out;
}