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

Add custom kq scaling from Gemma2Attention

Browse source
This commit is contained in:
Andrei Betlen 2024-06-29 10:17:33 -04:00
parent 6f2464e3dd
commit a89427908d
4 changed files with 12 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

1
gguf-py/gguf/constants.py
View file

@ -52,6 +52,7 @@ class Keys:
DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id"
ATTN_LOGIT_SOFTCAPPING = "{arch}.attn_logit_softcapping"
FINAL_LOGIT_SOFTCAPPING = "{arch}.final_logit_softcapping"
QUERY_PRE_ATTN_SCALAR = "{arch}.query_pre_attn_scalar"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"

3
gguf-py/gguf/gguf_writer.py
View file

@ -522,6 +522,9 @@ class GGUFWriter:
def add_final_logit_softcapping(self, value: float) -> None:
self.add_float32(Keys.LLM.FINAL_LOGIT_SOFTCAPPING.format(arch=self.arch), value)
def add_query_pre_attn_scalar(self, value: float) -> None:
self.add_float32(Keys.LLM.QUERY_PRE_ATTN_SCALAR.format(arch=self.arch), value)
def add_expert_count(self, count: int) -> None:
self.add_uint32(Keys.LLM.EXPERT_COUNT.format(arch=self.arch), count)