[Bugfix] fix custom op GmmSwigluQuantWeightNzTensorList (#4593)

### What this PR does / why we need it?

1. Fixes the environment path used to locate custom op shared libraries.
2. Uses empty tensor initialization for op outputs instead of
zero-initialization for better efficiency.



- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

---------

Signed-off-by: QianChenxi <[email protected]>

Author: ChenxiQ

CMakeLists.txt

@@ -131,6 +131,6 @@ target_link_libraries(
   opapi
 )
 
-target_link_options(vllm_ascend_C PRIVATE "-Wl,-rpath,$ORIGIN:$ORIGIN/lib")
+target_link_options(vllm_ascend_C PRIVATE "-Wl,-rpath,$ORIGIN:$ORIGIN/lib:$ORIGIN/_cann_ops_custom/vendors/vllm-ascend/op_api/lib")
 
 install(TARGETS vllm_ascend_C vllm_ascend_kernels DESTINATION ${VLLM_ASCEND_INSTALL_PATH})

csrc/CMakeLists.txt

@@ -15,7 +15,7 @@ option(BUILD_OPEN_PROJECT         "Build open ascend ops project."  ON)
 option(ENABLE_CCACHE              "Enable ccache capability"        ON)
 set(ASCEND_COMPUTE_UNIT           "ascend910b"                    CACHE   STRING   "soc that need to be compiled")
 set(ASCEND_OP_NAME                "ALL"                           CACHE   STRING   "operators that need to be compiled")
-set(VENDOR_NAME                   "customize"                     CACHE   STRING   "vendor name")
+set(VENDOR_NAME                   "vllm-ascend"                   CACHE   STRING   "vendor name")
 
 include(cmake/config.cmake)
 include(cmake/func.cmake)

csrc/build_aclnn.sh

@@ -31,4 +31,3 @@ bash build.sh -n $CUSTOM_OPS -c $SOC_ARG
 
 # install custom ops to vllm_ascend/_cann_ops_custom
 ./output/CANN-custom_ops*.run --install-path=$ROOT_DIR/vllm_ascend/_cann_ops_custom
-source $ROOT_DIR/vllm_ascend/_cann_ops_custom/vendors/customize/bin/set_env.bash

csrc/torch_binding.cpp

@@ -568,9 +568,9 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> grouped_matmul_swiglu_quant_weigh
     int m = x_size[0];
     int k = x_size[1];
 
-    at::Tensor output = at::zeros({m, n/2}, x.options().dtype(at::kChar));
-    at::Tensor output_scale = at::zeros({m}, x.options().dtype(at::kFloat));
-    at::Tensor output_offset = at::zeros({m}, x.options().dtype(at::kFloat));
+    at::Tensor output = at::empty({m, n/2}, x.options().dtype(at::kChar));
+    at::Tensor output_scale = at::empty({m}, x.options().dtype(at::kFloat));
+    at::Tensor output_offset = at::empty({m}, x.options().dtype(at::kFloat));
 
     EXEC_NPU_CMD(
         aclnnGroupedMatmulSwigluQuantWeightNzTensorList,

vllm_ascend/platform.py

@@ -38,27 +38,6 @@
     prefill_context_parallel_enable, update_aclgraph_sizes,
     update_cudagraph_capture_sizes, update_default_aclgraph_sizes)
 
-# set custom ops path
-CUR_DIR = os.path.dirname(os.path.realpath(__file__))
-CUSTOM_OPP_PATH = os.path.join(CUR_DIR, "vllm_ascend", "_cann_ops_custom",
-                               "vendors", "customize")
-CUSTOM_LIB_PATH = os.path.join(CUSTOM_OPP_PATH, "op_api", "lib")
-
-if os.path.exists(CUSTOM_OPP_PATH):
-    current_cust_opp_path = os.environ.get("ASCEND_CUSTOM_OPP_PATH", "")
-    if current_cust_opp_path:
-        os.environ[
-            "ASCEND_CUSTOM_OPP_PATH"] = f"{CUSTOM_OPP_PATH}:{current_cust_opp_path}"
-    else:
-        os.environ["ASCEND_CUSTOM_OPP_PATH"] = CUSTOM_OPP_PATH
-
-if os.path.exists(CUSTOM_LIB_PATH):
-    current_lib_path = os.environ.get("LD_LIBRARY_PATH", "")
-    if current_lib_path:
-        os.environ["LD_LIBRARY_PATH"] = f"{CUSTOM_LIB_PATH}:{current_lib_path}"
-    else:
-        os.environ["LD_LIBRARY_PATH"] = CUSTOM_LIB_PATH
-
 if TYPE_CHECKING:
     from vllm.config import ModelConfig, VllmConfig
     from vllm.utils import FlexibleArgumentParser

vllm_ascend/utils.py

@@ -247,6 +247,19 @@ def enable_custom_op():
     Ensure that ASCEND_RT_VISIBLE_DEVICES can be dynamically modified before torch.npu.set_device().
     """
     global _CUSTOM_OP_ENABLED
+
+    # set custom ops path
+    CUR_DIR = os.path.dirname(os.path.realpath(__file__))
+    CUSTOM_OPP_PATH = os.path.join(CUR_DIR, "_cann_ops_custom", "vendors",
+                                   "vllm-ascend")
+    if os.path.exists(CUSTOM_OPP_PATH):
+        current_cust_opp_path = os.environ.get("ASCEND_CUSTOM_OPP_PATH", "")
+        if current_cust_opp_path:
+            os.environ[
+                "ASCEND_CUSTOM_OPP_PATH"] = f"{CUSTOM_OPP_PATH}:{current_cust_opp_path}"
+        else:
+            os.environ["ASCEND_CUSTOM_OPP_PATH"] = CUSTOM_OPP_PATH
+
     if _CUSTOM_OP_ENABLED is not None:
         return _CUSTOM_OP_ENABLED
     try: