update VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT

vllm/model_executor/model_loader/utils.py

@@ -288,8 +288,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_LIGHTOP_FILL_MOE_ALIGN'] = '1'
                 if not envs.is_set("VLLM_USE_CAT_MLA"):
                     os.environ['VLLM_USE_CAT_MLA'] = '1'
-                if not envs.is_set("VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT"):
-                    os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
+                # if not envs.is_set("VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT"):
+                #     os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
             else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
                     os.environ['VLLM_USE_PD_SPLIT'] = '1'
@@ -302,8 +302,8 @@ def get_model_architecture(
                         os.environ['VLLM_USE_LIGHTOP_MOE_SUM'] = '1' 
                     if not envs.is_set("VLLM_USE_FUSE_SILU_AND_MUL"):
                         os.environ['VLLM_USE_FUSE_SILU_AND_MUL'] = '1'
-                    if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
-                        os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
+                    # if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
+                    #     os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
                         
             if os.getenv('GEMM_PAD') != '1': 
                 os.environ['GEMM_PAD'] = '0'