upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-05 09:30:49 -05:00
Code Issues Packages Projects Releases Wiki Activity

DeepseekV4 MTP + DSpark (#25784)

Browse Source
This commit is contained in:
Aman Gupta
2026-08-02 20:55:34 +08:00
committed by GitHub
parent f5919bf458
commit 596a5795bd
14 changed files with 1534 additions and 225 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
src/llama-model.cpp
View File

@@ -2138,6 +2138,75 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
nullptr);
}
} break;
case LLM_ARCH_DEEPSEEK4:
{
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP) {
const llama_memory_i::layer_filter_cb filter_mtp = [&](int32_t il) {
return il >= (int32_t) hparams.n_layer();
};
res = new llama_kv_cache_iswa(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
params.swa_full,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
cparams.n_ubatch,
1,
nullptr,
filter_mtp,
nullptr,
nullptr);
} else {
res = new llama_kv_cache_dsv4(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
params.swa_full,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
cparams.n_ubatch,
1,
cparams.n_rs_seq,
nullptr,
nullptr);
}
} break;
case LLM_ARCH_DFLASH:
{
// DSV4 DSpark stages store a single MLA-style K per position (window = the draft ring)
if (hparams.dsv4_hc_mult > 0) {
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
res = new llama_kv_cache_iswa(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
params.swa_full,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
cparams.n_ubatch,
1,
nullptr,
nullptr,
nullptr,
nullptr);
break;
}
}
[[fallthrough]];
// Models that need standard caching should rely on recurrent/hybrid
// checks
default:
@@ -2253,24 +2322,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
}
}
if (arch == LLM_ARCH_DEEPSEEK4) {
GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
res = new llama_kv_cache_dsv4(
*this,
params.type_k,
params.type_v,
!cparams.flash_attn,
cparams.offload_kqv,
params.swa_full,
cparams.kv_unified,
cparams.n_ctx_seq,
cparams.n_seq_max,
cparams.n_ubatch,
1,
filter,
reuse);
} else if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
GGML_ASSERT(hparams.is_swa_any());
if (arch == LLM_ARCH_GEMMA4_ASSISTANT) {
@@ -2619,9 +2671,12 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_STEP35:
case LLM_ARCH_TALKIE:
case LLM_ARCH_MELLUM:
case LLM_ARCH_DFLASH:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_DFLASH:
// DSV4 DSpark drafters use DeepSeek-V4's normal RoPE; legacy DFlash backbones are NeoX
return model->hparams.dsv4_hc_mult > 0 ? LLAMA_ROPE_TYPE_NORM : LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:
case LLM_ARCH_PADDLEOCR:
return LLAMA_ROPE_TYPE_MROPE;