happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

some cleanup and now fails on build attn

This commit is contained in:
ryan-mangeno 2025-09-08 12:21:18 -04:00
parent e101005d1a
commit 044bc7d5cd
2 changed files with 8 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-graph.cpp
View File

@ -1547,7 +1547,7 @@ ggml_tensor * llm_graph_context::build_attn_with_sinks(
// optionally store to KV cache // optionally store to KV cache
if (k_cur) { if (k_cur) {
const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs(); const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
LLAMA_LOG_INFO("k_cur.shape = {%lld, %lld, %lld, %lld}\n", k_cur->ne[0], k_cur->ne[1], k_cur->ne[2], k_cur->ne[3]);
ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il)); ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
} }

16
src/llama-model.cpp
View File

@ -7654,10 +7654,6 @@ struct llm_build_modern_bert : public llm_graph_context {
idx_2d = ggml_cont(ctx0, idx_2d); idx_2d = ggml_cont(ctx0, idx_2d);
if (idx_2d->type != GGML_TYPE_I32) idx_2d = ggml_cast(ctx0, idx_2d, GGML_TYPE_I32); if (idx_2d->type != GGML_TYPE_I32) idx_2d = ggml_cast(ctx0, idx_2d, GGML_TYPE_I32);
Kcur->ne[0], Kcur->ne[1], Kcur->ne[2],
idx_2d->ne[0], idx_2d->ne[1], idx_2d->ne[2], idx_2d->ne[3],
idx_2d->type);
K_work = ggml_get_rows(ctx0, Kcur, idx_2d); K_work = ggml_get_rows(ctx0, Kcur, idx_2d);
V_work = ggml_get_rows(ctx0, Vcur, idx_2d); V_work = ggml_get_rows(ctx0, Vcur, idx_2d);
@ -7679,7 +7675,7 @@ struct llm_build_modern_bert : public llm_graph_context {
// final pos_k to pass to rope // final pos_k to pass to rope
pos_k = pos_rows; pos_k = pos_rows;
LLAMA_LOG_INFO("pos_k final: ne[0]=%lld, type=%d\n", pos_k->ne[0], pos_k->type); LLAMA_LOG_INFO("pos_k final: ne[0]=%lld, ne[1]=%lld type=%d\n", pos_k->ne[0], pos_k->ne[1], pos_k->type);
} }
if( !ggml_is_vector(pos_q) ) { if( !ggml_is_vector(pos_q) ) {
@ -7707,7 +7703,9 @@ struct llm_build_modern_bert : public llm_graph_context {
ext_factor, attn_factor, beta_fast, beta_slow); ext_factor, attn_factor, beta_fast, beta_slow);
// choseing mask, global vs swa // choseing mask, global vs swa
ggml_tensor * kq_b_layer = is_global ? inp_attn->self_kq_mask : inp_attn->self_kq_mask_swa; ggml_tensor * kq_mask = is_global ? inp_attn->self_kq_mask : inp_attn->self_kq_mask_swa;
ggml_tensor * attn_out = build_attn( ggml_tensor * attn_out = build_attn(
inp_attn, inp_attn,
@ -7716,14 +7714,14 @@ struct llm_build_modern_bert : public llm_graph_context {
Qcur, Qcur,
K_work, K_work,
V_work, V_work,
kq_b_layer, kq_mask,
nullptr, nullptr,
1.0f / sqrtf(float(n_embd_head)), 1.0f / sqrtf(float(n_embd_head)),
il il
); );
// residual addition
ggml_tensor * cur_attn = ggml_add(ctx0, attn_out, x); ggml_tensor * cur_attn = ggml_add(ctx0, x, attn_out);
// optional output select // optional output select
if (il == n_layer - 1 && inp_out_ids) { if (il == n_layer - 1 && inp_out_ids) {