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

CUDA + ggml: add sparse-fa for DSV4/GLM (#27970)

Browse Source
This commit is contained in:
Aman Gupta
2026-09-02 17:27:37 +03:00
committed by GitHub
parent 0f3a71be15
commit 8e93a9773b
12 changed files with 392 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/deepseek4.cpp
+4 -3
View File
@@ -752,7 +752,8 @@ ggml_tensor * llama_model_deepseek4::graph::build_csa_lid_attention(
ggml_tensor * kq_mask = ggml_concat(ctx0, raw_mask, csa_mask, 0);
cb(kq_mask, "csa_lid_kq_mask", il);
ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
const int64_t n_kv_max = std::min<int64_t>(raw_mask->ne[0], hparams.n_swa) + top_k->ne[0];
ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, n_kv_max, kq_scale, il);
if (k_rot) {
out = llama_mul_mat_hadamard(ctx0, out, k_rot);
}
@@ -807,7 +808,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_hca_attention(
ggml_tensor * kq_mask = ggml_concat(ctx0, raw_mask, hca_mask, 0);
cb(kq_mask, "hca_kq_mask", il);
ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
ggml_tensor * out = build_attn_mha(q, k_all, k_all, nullptr, kq_mask, sinks, nullptr, 0, kq_scale, il);
if (k_rot) {
out = llama_mul_mat_hadamard(ctx0, out, k_rot);
}
@@ -843,7 +844,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_raw_attention(
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * out = build_attn_mha(q, k, k, nullptr, kq_mask, sinks, nullptr, kq_scale, il);
ggml_tensor * out = build_attn_mha(q, k, k, nullptr, kq_mask, sinks, nullptr, 0, kq_scale, il);
if (k_rot) {
out = llama_mul_mat_hadamard(ctx0, out, k_rot);
}