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

[conver_hf_to_gguf.py] add phi3 sliding window

Browse source
This commit is contained in:
Shupei Fan 2024-07-24 15:08:01 +08:00
parent b4e3de6b17
commit 19d6ad9db7
2 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

1
convert_hf_to_gguf.py
View file

@ -2078,6 +2078,7 @@ class Phi3MiniModel(Model):
self.gguf_writer.add_rope_dimension_count(rope_dims)
self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
# write rope scaling for long context (128k) model
rope_scaling = self.find_hparam(['rope_scaling'], True)

3
src/llama.cpp
View file

@ -4974,8 +4974,7 @@ static void llm_load_hparams(
} break;
case LLM_ARCH_PHI3:
{
hparams.n_swa = 2048;
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {