ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

qwen3next: remove redundant v_conv cont in delta path

Browse Source
This commit is contained in:
yurko
2026-02-07 13:25:34 -08:00
parent a1163d0b68
commit 0e3891b348
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llama-build-context.cpp
View File

@@ -4711,12 +4711,14 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
ggml_tensor * q_conv = ggml_view_2d(ctx0, conv_output_silu, key_dim, n_tok, conv_output_silu->nb[1], 0); ggml_tensor * q_conv = ggml_view_2d(ctx0, conv_output_silu, key_dim, n_tok, conv_output_silu->nb[1], 0);
ggml_tensor * k_conv = ggml_view_2d(ctx0, conv_output_silu, key_dim, n_tok, conv_output_silu->nb[1], ggml_tensor * k_conv = ggml_view_2d(ctx0, conv_output_silu, key_dim, n_tok, conv_output_silu->nb[1],
key_dim * ggml_element_size(conv_output_silu)); key_dim * ggml_element_size(conv_output_silu));
ggml_tensor * v_conv = ggml_view_2d(ctx0, conv_output_silu, value_dim, n_tok, conv_output_silu->nb[1], ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_output_silu, head_v_dim, num_v_heads, n_tok, 1,
ggml_row_size(conv_output_silu->type, head_v_dim),
conv_output_silu->nb[1],
conv_output_silu->nb[1] * n_tok,
2 * key_dim * ggml_element_size(conv_output_silu)); 2 * key_dim * ggml_element_size(conv_output_silu));
q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_tok, 1); q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_tok, 1);
k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_tok, 1); k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_tok, 1);
v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_tok, 1);
if (num_k_heads != num_v_heads) { if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0); GGML_ASSERT(num_v_heads % num_k_heads == 0);