feat: support the fusion of topk and add operators in the router module of the moe model.

DongheJin · yq33victor · commit 7792c3069869 · 2025-11-05T20:32:13.000+08:00
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -28,20 +28,20 @@ if(USE_NPU)
     if(DEVICE_TYPE STREQUAL "USE_A3")
         message("downloading a3 arm xllm kernels")
         file(DOWNLOAD 
-            "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.6.0/xllm_kernels-1.3.2-Linux.a3.arm.rpm"
+            "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.7.0/xllm_kernels-1.3.2-Linux.a3.arm.rpm"
             "${CMAKE_BINARY_DIR}/xllm_kernels.rpm"
         )
     else()  
       if(DEVICE_ARCH STREQUAL "ARM")
           message("downloading a2 arm xllm_kernels")
           file(DOWNLOAD 
-              "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.6.0/xllm_kernels-1.3.2-Linux.a2.arm.rpm"
+              "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.7.0/xllm_kernels-1.3.2-Linux.a2.arm.rpm"
               "${CMAKE_BINARY_DIR}/xllm_kernels.rpm"
           )
       else()
           message("downloading a2 x86 xllm_kernels")
           file(DOWNLOAD 
-              "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.6.0/xllm_kernels-1.3.2-Linux.a2.x86.rpm"
+              "https://9n-das-tools.s3.cn-north-1.jdcloud-oss.com/xllm-ai/xllm_kernels/0.7.0/xllm_kernels-1.3.2-Linux.a2.x86.rpm"
               "${CMAKE_BINARY_DIR}/xllm_kernels.rpm"
           )
       endif()
diff --git a/third_party/minja b/third_party/minja
@@ -1 +1 @@
-Subproject commit 28097055e81c971e5a229220c0c8ca546fd28d13
+Subproject commit cf9734a25618e19f8e74dc999c283e464500e8f0
diff --git a/xllm/core/runtime/llm_master.cpp b/xllm/core/runtime/llm_master.cpp
@@ -300,8 +300,6 @@ std::shared_ptr<Request> LLMMaster::generate_request(
     }
   }
 
-  // local_prompt_tokens = {151331, 151333, 151336,    198, 109377, 151337};
-  // LOG(INFO)<<local_prompt_tokens;
   COUNTER_ADD(tokenization_latency_seconds, timer.elapsed_seconds());
 
   int32_t max_context_len = model_args_.max_position_embeddings();

Original file line number	Diff line number	Diff line change
`@@ -300,8 +300,6 @@ std::shared_ptr<Request> LLMMaster::generate_request(`
`300`	`300`	`}`
`301`	`301`	`}`
`302`	`302`
`303`		`- // local_prompt_tokens = {151331, 151333, 151336, 198, 109377, 151337};`
`304`		`- // LOG(INFO)<<local_prompt_tokens;`
`305`	`303`	`COUNTER_ADD(tokenization_latency_seconds, timer.elapsed_seconds());`
`306`	`304`
`307`	`305`	`int32_t max_context_len = model_args_.max_position_embeddings();`