rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Better use_cache implementation

Browse Source
This commit is contained in:
Henk
2023-09-07 04:29:28 +02:00
parent dfb63b2340
commit 0d0a671bb9
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
modeling/inference_models/generic_hf_torch/class.py
View File

@@ -107,6 +107,7 @@ class model_backend(HFTorchInferenceModel):
tf_kwargs = {
"low_cpu_mem_usage": True,
"use_cache": True # Workaround for models that accidentally turn cache to false
}
if not hasattr(self.model_config, 'quantization_config'):
@@ -130,8 +131,8 @@ class model_backend(HFTorchInferenceModel):
})
if self.model_type == "gpt2":
# We must disable low_cpu_mem_usage and if using a GPT-2 model
# because GPT-2 is not compatible with this feature yet.
# We must disable low_cpu_mem_usage and quantization if using a GPT-2 model
# because GPT-2 is not compatible with these features yet.
tf_kwargs.pop("low_cpu_mem_usage", None)
tf_kwargs.pop("quantization_config", None)

1
modeling/inference_models/hf.py
View File

@@ -230,7 +230,6 @@ class HFInferenceModel(InferenceModel):
def _post_load(self) -> None:
self.badwordsids = koboldai_settings.badwordsids_default
self.model_type = str(self.model_config.model_type)
self.model.use_cache = True # Workaround for models that accidentally uploaded with False
# These are model specific tokenizer overrides if a model has bad defaults
if self.model_type == "llama":