root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

llm : fix llm_build_kqv taking unused tensor (benign, #3837)

Browse Source
This commit is contained in:
Georgi Gerganov 2023-11-01 23:08:30 +02:00
parent 523e49b111
commit c43c2da8af
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 9 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
llama.cpp
View File

@ -3345,7 +3345,6 @@ static struct ggml_tensor * llm_build_ffn(
// if max_alibi_bias > 0 then apply ALiBi // if max_alibi_bias > 0 then apply ALiBi
static struct ggml_tensor * llm_build_kqv( static struct ggml_tensor * llm_build_kqv(
struct ggml_context * ctx, struct ggml_context * ctx,
struct ggml_tensor * cur,
const llama_hparams & hparams, const llama_hparams & hparams,
const llama_kv_cache & kv, const llama_kv_cache & kv,
struct ggml_tensor * wo, struct ggml_tensor * wo,
@ -3411,7 +3410,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
cb(kqv_merged, "kqv_merged", il); cb(kqv_merged, "kqv_merged", il);
cur = ggml_cont_2d(ctx, kqv_merged, n_embd, n_tokens); struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd, n_tokens);
cb(cur, "kqv_merged_cont", il); cb(cur, "kqv_merged_cont", il);
cur = ggml_mul_mat(ctx, wo, cur); cur = ggml_mul_mat(ctx, wo, cur);
@ -3565,7 +3564,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, cur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL, model.layers[il].wo, NULL,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -3677,7 +3676,7 @@ struct llm_build_context {
// apply ALiBi for 13B model // apply ALiBi for 13B model
const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f;
cur = llm_build_kqv(ctx0, cur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL, model.layers[il].wo, NULL,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -3795,7 +3794,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, attn_norm, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL, model.layers[il].wo, NULL,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -3895,7 +3894,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, cur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -4100,7 +4099,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
// TODO: not tested, could be broken // TODO: not tested, could be broken
cur = llm_build_kqv(ctx0, Q, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo, model.layers[il].bo,
Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -4191,7 +4190,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, Qcur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL, model.layers[il].wo, NULL,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -4288,7 +4287,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, Qcur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo, model.layers[il].bo,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);
@ -4382,7 +4381,7 @@ struct llm_build_context {
llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
cur = llm_build_kqv(ctx0, Qcur, hparams, kv_self, cur = llm_build_kqv(ctx0, hparams, kv_self,
model.layers[il].wo, NULL, model.layers[il].wo, NULL,
Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il); Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il);
cb(cur, "kqv_out", il); cb(cur, "kqv_out", il);