furyhawk/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-07-21 02:05:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix #2093

Browse Source
This commit is contained in:
Kawrakow
2026-07-07 16:20:12 +00:00
parent 6c5f047a7c
commit 66d2e14e97
1 changed files with 1 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphs/build_deepseek2.cpp
+1 -4
View File
@@ -331,9 +331,6 @@ static inline int dsa_n_top_k(const llama_context & lctx) {
static inline bool dsa_should_use(const llama_context & lctx) {
return lctx.cparams.mtp_op_type == MTP_OP_NONE && lctx.cparams.dsa && lctx.model.arch == LLM_ARCH_GLM_DSA && !lctx.kv_self.kr_l.empty();
}
static inline bool dsa_should_use(const llama_context & lctx, int il) {
return dsa_should_use(lctx) && lctx.model.layers[il].indexer_attn_q_b && lctx.kv_self.kr_l.size() > (size_t) il && lctx.kv_self.kr_l[il];
}
static inline bool dsa_can_use_fast_path(const llama_context & lctx, int n_tokens, int n_kv) {
auto n_top_k = dsa_n_top_k(lctx);
if (n_tokens == 1 && n_top_k < n_kv && n_top_k%256 == 0 && lctx.cparams.flash_attn && (lctx.cparams.mla_attn == 1 || lctx.cparams.mla_attn == 3)) {
@@ -754,7 +751,7 @@ ggml_tensor * llm_build_context::build_deepseek2_layer_attention(
ggml_tensor * sparse_mask = KQ_mask;
ggml_tensor * sparse_mask_fa = KQ_mask;
bool use_dsa = dsa_should_use(lctx, il);
bool use_dsa = dsa_should_use(lctx);
bool dsa_fast_path = use_dsa && dsa_can_use_fast_path(lctx, n_tokens, n_kv);
// self_attention