upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-05 01:20:45 -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

24
src/llama-context.cpp
View File

@@ -120,8 +120,9 @@ llama_context::llama_context(
cparams.no_perf = params.no_perf;
cparams.warmup = false;
cparams.embeddings_layer_inp.resize(hparams.n_layer(), false);
embd_layer_inp.resize(hparams.n_layer());
// +1: id n_layer() taps the output of the last layer ("input" of the head)
cparams.embeddings_layer_inp.resize(hparams.n_layer() + 1, false);
embd_layer_inp.resize(hparams.n_layer() + 1);
cparams.ctx_type = params.ctx_type;
cparams.pooling_type = params.pooling_type;
@@ -1164,7 +1165,7 @@ void llama_context::set_embeddings_nextn(bool value, bool masked) {
void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
LLAMA_LOG_DEBUG("%s: lid = %d, enable = %d\n", __func__, lid, enable);
GGML_ASSERT(lid < model.hparams.n_layer());
GGML_ASSERT(lid <= model.hparams.n_layer());
cparams.embeddings_layer_inp[lid] = enable;
@@ -1716,7 +1717,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;
const int64_t n_vocab = vocab.n_tokens();
const int64_t n_embd = hparams.n_embd_inp();
const bool mtp_embd = cparams.ctx_type == LLAMA_CONTEXT_TYPE_MTP && batch_inp.embd;
const int64_t n_embd = mtp_embd ? hparams.n_embd_out() : hparams.n_embd_inp();
// when computing embeddings, all tokens are output
const bool output_all = cparams.embeddings;
@@ -2275,8 +2277,9 @@ void llama_context::extract_layer_inputs(const llm_graph_result * res, size_t to
}
void llama_context::output_reorder() {
const uint64_t n_vocab = model.vocab.n_tokens();
const uint64_t n_embd = model.hparams.n_embd;
const uint64_t n_vocab = model.vocab.n_tokens();
const uint64_t n_embd = model.hparams.n_embd;
const uint64_t n_embd_out = model.hparams.n_embd_out();
for (size_t s = 0; s < output_swaps.size(); ++s) {
const uint64_t i0 = output_swaps[s].i0;
@@ -2289,14 +2292,14 @@ void llama_context::output_reorder() {
}
if (embd.size > 0) {
for (uint64_t k = 0; k < n_embd; k++) {
std::swap(embd.data[i0*n_embd + k], embd.data[i1*n_embd + k]);
for (uint64_t k = 0; k < n_embd_out; k++) {
std::swap(embd.data[i0*n_embd_out + k], embd.data[i1*n_embd_out + k]);
}
}
if (embd_nextn.size > 0) {
for (uint64_t k = 0; k < n_embd; k++) {
std::swap(embd_nextn.data[i0*n_embd + k], embd_nextn.data[i1*n_embd + k]);
for (uint64_t k = 0; k < n_embd_out; k++) {
std::swap(embd_nextn.data[i0*n_embd_out + k], embd_nextn.data[i1*n_embd_out + k]);
}
}
@@ -2351,6 +2354,7 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
model.arch == LLM_ARCH_QWEN35 ||
model.arch == LLM_ARCH_QWEN35MOE ||
model.arch == LLM_ARCH_DEEPSEEK4 ||
(model.arch == LLM_ARCH_DFLASH && model.hparams.dsv4_hc_mult > 0) ||
model.arch == LLM_ARCH_NANBEIGE ||
model.arch == LLM_ARCH_MINIMAX_M3) {
return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());