removed cache

layers.py

@@ -31,8 +31,6 @@ class Linear(nn.Module):
         out_features: int,
         bias: bool = True,
         dtype: torch.dtype = None,
-        operator_cache: OperatorCache = None,
-        cache_dir: str = None,
         group_size: int = 128,
     ):
         super().__init__()
@@ -51,8 +49,6 @@ class Linear(nn.Module):
                 out_dtype="float16",
                 fast_decoding=True,
                 enable_tuning=True,
-                operator_cache=operator_cache,
-                database_path=cache_dir,
                 group_size=group_size,
             )
         else:

text.py

@@ -101,13 +101,9 @@ def build_text_model(
     )
     qkv_dim = int(config.dim * (1 + 2 * config.n_kv_heads / config.n_heads))
 
-    operator_cache = None
-    cache_dir = None
     group_size = None
     if linear_dtype == torch.int8:
 
-        operator_cache = OperatorCache()
-        cache_dir = "./cache"
         group_size = config.group_size
 
     def create_linear(in_features, out_features, dtype=linear_dtype):
@@ -116,8 +112,6 @@ def build_text_model(
             in_features=in_features,
             out_features=out_features,
             dtype=dtype,
-            operator_cache=operator_cache,
-            cache_dir=cache_dir,
             group_size=group_size,
         )