[Bugfix] Fix KeyError for Mistral-Large-3 rope_scaling config (#14627)
This commit is contained in:
@@ -358,9 +358,10 @@ class ModelConfig:
|
|||||||
mscale_all_dim = self.hf_config.rope_scaling.get(
|
mscale_all_dim = self.hf_config.rope_scaling.get(
|
||||||
"mscale_all_dim", False
|
"mscale_all_dim", False
|
||||||
)
|
)
|
||||||
scaling_factor = self.hf_config.rope_scaling["factor"]
|
scaling_factor = self.hf_config.rope_scaling.get("factor")
|
||||||
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
if scaling_factor is not None:
|
||||||
self.scaling = self.scaling * mscale * mscale
|
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
||||||
|
self.scaling = self.scaling * mscale * mscale
|
||||||
|
|
||||||
elif "MiniCPM3ForCausalLM" in self.hf_config.architectures:
|
elif "MiniCPM3ForCausalLM" in self.hf_config.architectures:
|
||||||
self.head_dim = 128
|
self.head_dim = 128
|
||||||
|
|||||||
Reference in New Issue
Block a user