upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-06 18:10:51 -05:00
Code Issues Packages Projects Releases Wiki Activity

llama : MTP support for DeepSeek V3.2 (#26457)

Browse Source 
* llama : MTP support for DeepSeek V3.2

* model : no need to include MTP layers during DeepSeek V3.2 model type discovery

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming
2026-08-03 08:25:01 +02:00
committed by GitHub
parent 1464c62d88
commit 2b63e0610b
4 changed files with 308 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/llama-model.cpp
View File

@@ -2071,24 +2071,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
{
res = nullptr;
} break;
case LLM_ARCH_DEEPSEEK32:
{
res = new llama_kv_cache_dsa(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
1,
hparams.n_swa,
hparams.swa_type,
nullptr,
nullptr);
} break;
case LLM_ARCH_GLM_DSA:
case LLM_ARCH_DEEPSEEK32:
{
if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {
// The NextN/MTP draft head runs dense MLA (no DSA indexer), so the
@@ -2313,7 +2297,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
}
if ((arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_HY_V3 || arch == LLM_ARCH_GLM_DSA ||
arch == LLM_ARCH_MIMO2) &&
arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_DEEPSEEK32) &&
hparams.n_layer_nextn > 0) {
if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP) {
filter = [&](uint32_t il) { return il >= hparams.n_layer(); };