Skip to content

Commit b9bcdc7

Browse files
authored
Change the load format to pt for Mixtral (#2028)
1 parent 4ff0203 commit b9bcdc7

File tree

1 file changed

+10
-0
lines changed

1 file changed

+10
-0
lines changed

vllm/config.py

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -119,6 +119,16 @@ def _verify_load_format(self) -> None:
119119
# Force ROCm to load from pt weights if nothing specific is set
120120
if load_format == "auto":
121121
load_format = "pt"
122+
123+
# FIXME(woosuk): This is a temporary hack. Support safetensor weights.
124+
architectures = getattr(self.hf_config, "architectures", [])
125+
if "MixtralForCausalLM" in architectures and load_format != "pt":
126+
logger.info(
127+
"Currently, only 'pt' format is supported for Mixtral. "
128+
"Changing the format to 'pt'. This may re-download the "
129+
"weights if you have downloaded the safetensor weights.")
130+
load_format = "pt"
131+
122132
self.load_format = load_format
123133

124134
def _verify_tokenizer_mode(self) -> None:

0 commit comments

Comments
 (0)