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

llama: refactor fused ops (#24646)

Browse Source
This commit is contained in:
Aman Gupta
2026-07-08 18:18:09 +08:00
committed by GitHub
parent bbebeec4a8
commit 90e0f5cfcb
6 changed files with 122 additions and 131 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/delta-net-base.cpp
+4 -4
View File
@@ -401,9 +401,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
// K=1: output carries the final state only. state s is 4D [S_v, S_v, H_v, n_seqs].
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, /*K=*/1);
if (n_tokens == 1) {
cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
res->add_fused_node({LLM_FUSED_OP_GDN_AR, result, il});
} else {
cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
res->add_fused_node({LLM_FUSED_OP_GDN_CH, result, il});
}
ggml_tensor * output = ggml_view_4d(ctx0, result,
@@ -566,9 +566,9 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
// state s is 4D [S_v, S_v, H_v, n_seqs]; K snapshot slots are written into the output.
ggml_tensor * gdn_out = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, K);
if (n_seq_tokens > 1) {
cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_CH, il);
res->add_fused_node({LLM_FUSED_OP_GDN_CH, gdn_out, il});
} else {
cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_AR, il);
res->add_fused_node({LLM_FUSED_OP_GDN_AR, gdn_out, il});
}
const int64_t attn_score_elems = S_v * H_v * n_seq_tokens * n_seqs;