vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: Get n_head_kv per-layer in build_bamba

Browse source 
Branch: BambaArchitecture

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart 2024-12-09 13:43:27 -07:00
parent e7b1abbc0a
commit f2478bcab5
1 changed files with 3 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

5
src/llama.cpp
View file

@ -14710,14 +14710,15 @@ struct llm_build_context {
}
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens), inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
LLAMA_LOG_DEBUG("%s[%d]: 9. ggml_rope_ext\n", __func__, il);
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens), inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);