Replace all the interfaces of Triton with the implementation of ExLlamaV2

server/text_generation_server/layers/gptq/__init__.py

@@ -171,7 +171,7 @@ class GPTQWeightsLoader(WeightsLoader):
             g_idx=g_idx,
             bits=self.bits,
             groupsize=self.groupsize,
-            use_exllama=use_exllama,
+            use_exllama=True,
         )
 
     def get_weights_col_packed(
@@ -227,7 +227,7 @@ class GPTQWeightsLoader(WeightsLoader):
             bits=self.bits,
             groupsize=self.groupsize,
             use_awq_kernel=self.quantize == "awq",
-            use_exllama=False,
+            use_exllama=True,
         )
 
     def get_multi_weights_col(self, weights: Weights, prefixes: List[str], dim: int):
@@ -294,7 +294,7 @@ class GPTQWeightsLoader(WeightsLoader):
             bits=self.bits,
             groupsize=self.groupsize,
             use_awq_kernel=self.quantize == "awq",
-            use_exllama=use_exllama,
+            use_exllama=True,
         )
 
     def get_weights_row(self, weights: Weights, prefix: str):
@@ -394,7 +394,7 @@ class GPTQWeightsLoader(WeightsLoader):
             bits=self.bits,
             groupsize=self.groupsize,
             use_awq_kernel=self.quantize == "awq",
-            use_exllama=use_exllama,
+            use_exllama=True,
         )
 
     def _get_gptq_params(self, weights: Weights):