ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-28 17:14:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

This appears to work

Browse Source
This commit is contained in:
Kawrakow
2026-02-19 16:15:31 +00:00
parent 8737d3d924
commit 07516cec2d
1 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/llama-delta-net.cpp
View File

@@ -578,16 +578,21 @@ ggml_tensor * delta_net::build_layer_attn_linear_core(ggml_context * ctx0, ggml_
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
const int64_t repeat_factor = num_v_heads / num_k_heads;
if (model.layers[il].ssm_beta_alpha) {
const int64_t repeat_factor = num_v_heads / num_k_heads;
ggml_tensor * q_reshaped = ggml_reshape_3d(ctx0, q_conv, head_k_dim, 1, num_k_heads * n_tok);
ggml_tensor * k_reshaped = ggml_reshape_3d(ctx0, k_conv, head_k_dim, 1, num_k_heads * n_tok);
ggml_tensor * q_reshaped = ggml_reshape_3d(ctx0, q_conv, head_k_dim, 1, num_k_heads * n_tok);
ggml_tensor * k_reshaped = ggml_reshape_3d(ctx0, k_conv, head_k_dim, 1, num_k_heads * n_tok);
ggml_tensor * q_repeated = ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads * n_tok, 1);
ggml_tensor * k_repeated = ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads * n_tok, 1);
ggml_tensor * q_repeated = ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads * n_tok, 1);
ggml_tensor * k_repeated = ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads * n_tok, 1);
q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_tok, 1);
k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_tok, 1);
q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_tok, 1);
k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_tok, 1);
} else {
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
}
cb(q_conv, "q_conv_predelta", il);