update USE_FUSED_RMS_QUANT=1 and USE_FUSED_SILU_MUL_QUANT=1 for dpsk-v3 quantitative models

vllm/model_executor/model_loader/utils.py

@@ -259,12 +259,13 @@ def get_model_architecture(
                     os.environ['VLLM_REJECT_SAMPLE_OPT'] = '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("USE_FUSED_RMS_QUANT"):
-                    os.environ['USE_FUSED_RMS_QUANT'] = '1'
-                if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
-                    os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):
                     os.environ['VLLM_SCHED_ENABLE_MINIMAL_INJECTION'] = '1'
+                if model_config.quantization is not None:
+                    if not envs.is_set("USE_FUSED_RMS_QUANT"):
+                        os.environ['USE_FUSED_RMS_QUANT'] = '1'
+                    if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
+                        os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
             else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
                     os.environ['VLLM_USE_PD_SPLIT'] = '1'
@@ -303,12 +304,13 @@ def get_model_architecture(
                     os.environ['VLLM_REJECT_SAMPLE_OPT'] = '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("USE_FUSED_RMS_QUANT"):
-                    os.environ['USE_FUSED_RMS_QUANT'] = '1'
-                if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
-                    os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):
                     os.environ['VLLM_SCHED_ENABLE_MINIMAL_INJECTION'] = '1'
+                if model_config.quantization is not None:
+                    if not envs.is_set("USE_FUSED_RMS_QUANT"):
+                        os.environ['USE_FUSED_RMS_QUANT'] = '1'
+                    if not envs.is_set("USE_FUSED_SILU_MUL_QUANT"):
+                        os.environ['USE_FUSED_SILU_MUL_QUANT'] = '1'
             else:
                 if not envs.is_set("VLLM_USE_PD_SPLIT"):
                     os.environ['VLLM_USE_PD_SPLIT'] = '1'