Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

graph : create V as a view of K in the k_iswa build_attn (#27392)

Browse Source 
build_attn with the llm_graph_input_attn_k_iswa input was using the cached K
tensor itself as V. Create V as a view of K (the first v_cur->ne[0] elements
of each row), like the other K-only build_attn overloads.

The deepseek4 MTP call site now passes the kv tensor as v_cur.

Assisted-by: pi:llama.cpp/Qwen3.8-27B
This commit is contained in:
Georgi Gerganov
2026-08-20 10:00:35 +03:00
committed by GitHub
parent f466cfa38f
commit 929d47a391
2 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-graph.cpp
+1 -3
View File
@@ -3099,8 +3099,6 @@ ggml_tensor * llm_graph_context::build_attn(
int il) const { int il) const {
const bool is_swa = hparams.is_swa(il); const bool is_swa = hparams.is_swa(il);
GGML_UNUSED(v_cur);
auto * k_rot = is_swa ? inp->self_k_rot_swa : inp->self_k_rot; auto * k_rot = is_swa ? inp->self_k_rot_swa : inp->self_k_rot;
if (k_rot) { if (k_rot) {
@@ -3133,7 +3131,7 @@ ggml_tensor * llm_graph_context::build_attn(
// MLA-style attention: the cached K is used as V // MLA-style attention: the cached K is used as V
ggml_tensor * q = q_cur; ggml_tensor * q = q_cur;
ggml_tensor * k = mctx_cur->get_k(ctx0, il); ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * v = k; ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
src/models/deepseek4.cpp
+1 -1
View File
@@ -1225,7 +1225,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(
if (inp_mtp) { if (inp_mtp) {
out = build_attn(inp_mtp, out = build_attn(inp_mtp,
nullptr, nullptr, nullptr, nullptr, nullptr, nullptr,
q, kv, nullptr, q, kv, kv,
nullptr, layer.attn_sinks, nullptr, nullptr, layer.attn_sinks, nullptr,
1.0f/sqrtf(float(n_embd_head)), il); 1.0f/sqrtf(float(n_embd_head)), il);
cb(out, "attn_raw", il); cb(out, "attn_raw", il);