Offload KQV by default

Author: abetlen

llama_cpp/llama.py

@@ -77,7 +77,7 @@ def __init__(
         mul_mat_q: bool = True,
         logits_all: bool = False,
         embedding: bool = False,
-        offload_kqv: bool = False,
+        offload_kqv: bool = True,
         # Sampling Params
         last_n_tokens_size: int = 64,
         # LoRA Params

llama_cpp/server/settings.py

@@ -90,7 +90,7 @@ class ModelSettings(BaseSettings):
     logits_all: bool = Field(default=True, description="Whether to return logits.")
     embedding: bool = Field(default=True, description="Whether to use embeddings.")
     offload_kqv: bool = Field(
-        default=False, description="Whether to offload kqv to the GPU."
+        default=True, description="Whether to offload kqv to the GPU."
     )
     # Sampling Params
     last_n_tokens_size: int = Field(