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

ci : use Pythia models instead of OpenLlama (#7470)

Browse source 
* ci : start using Pythia models over OpenLlama

ggml-ci

* ci : disable q2_k ppl tests

* ci : use convert-hf-to-gguf.py

* ci : update gg_get_model

* ci : fix convert outfile name

ggml-ci

* llama : gptneox arch use F32 attn prec

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-05-23 15:28:14 +03:00 • committed by GitHub
parent dacfcebd60
commit 55ac3b7aea
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 282 additions and 146 deletions
Download patch file Download diff file Expand all files Collapse all files

4
llama.cpp
View file

@ -6718,7 +6718,7 @@ static struct ggml_tensor * llm_build_kqv(
cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
}
@ -6727,7 +6727,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
ggml_mul_mat_set_prec(kq, GGML_PREC_F32);