diff --git a/.github/workflows/linux-x64-gpu.yml b/.github/workflows/linux-x64-gpu.yml
index ca75f2287..c29940c7c 100644
--- a/.github/workflows/linux-x64-gpu.yml
+++ b/.github/workflows/linux-x64-gpu.yml
@@ -55,6 +55,7 @@ jobs:
             cd /work
             source /opt/conda/bin/activate
             conda activate py38
+            apt install libspdlog-dev -y
             mkdir build && cd build
             bash ../generate.sh make
             make -j$(nproc) && make install
diff --git a/builder/manywheel/entrypoint_build.sh b/builder/manywheel/entrypoint_build.sh
index c9c2cae6e..48e25b485 100755
--- a/builder/manywheel/entrypoint_build.sh
+++ b/builder/manywheel/entrypoint_build.sh
@@ -12,6 +12,7 @@ conda activate $PYTHON_VERSION
 
 cd lmdeploy
 mkdir -p build && cd build && rm -rf *
+apt install libspdlog-dev -y
 bash ../generate.sh make
 make -j$(nproc) && make install
 if [ $? != 0 ]; then
diff --git a/src/turbomind/utils/CMakeLists.txt b/src/turbomind/utils/CMakeLists.txt
index 82f08ef1b..a8b9f4923 100644
--- a/src/turbomind/utils/CMakeLists.txt
+++ b/src/turbomind/utils/CMakeLists.txt
@@ -26,7 +26,11 @@ target_link_libraries(cuda_utils PUBLIC CUDA::cudart)
 add_library(logger STATIC logger.cc)
 set_property(TARGET logger PROPERTY POSITION_INDEPENDENT_CODE  ON)
 set_property(TARGET logger PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS  ON)
-target_link_libraries(logger PUBLIC CUDA::cudart)
+if (UNIX)
+    target_link_libraries(logger PUBLIC CUDA::cudart -lpthread)
+else ()
+    target_link_libraries(logger PUBLIC CUDA::cudart)
+endif ()
 
 add_library(cublasAlgoMap STATIC cublasAlgoMap.cc)
 set_property(TARGET cublasAlgoMap PROPERTY POSITION_INDEPENDENT_CODE  ON)
diff --git a/src/turbomind/utils/logger.cc b/src/turbomind/utils/logger.cc
index 9788a8fad..fba7cfb9f 100644
--- a/src/turbomind/utils/logger.cc
+++ b/src/turbomind/utils/logger.cc
@@ -17,10 +17,22 @@
 #include "src/turbomind/utils/logger.h"
 #include <cuda_runtime.h>
 
+#ifndef _MSC_VER
+#include <spdlog/async.h>
+#include <spdlog/sinks/basic_file_sink.h>
+#include <spdlog/sinks/hourly_file_sink.h>
+#include <spdlog/sinks/rotating_file_sink.h>
+#endif
+
 namespace turbomind {
 
 Logger::Logger()
 {
+#ifndef _MSC_VER
+    // TODO: use config
+    SpdLogger::get_instance().set_log_path("/var/log/lmdeploy.log");
+    SpdLogger::get_instance().init();
+#endif
     char* is_first_rank_only_char = std::getenv("TM_LOG_FIRST_RANK_ONLY");
     bool  is_first_rank_only =
         (is_first_rank_only_char != nullptr && std::string(is_first_rank_only_char) == "ON") ? true : false;
@@ -56,4 +68,33 @@ Logger::Logger()
     }
 }
 
+#ifndef _MSC_VER
+void SpdLogger::init()
+{
+    if (inited_) {
+        return;
+    }
+
+    spdlog::init_thread_pool(8192, 1);
+
+    // rotate 500 MB
+    auto basic_sink = std::make_shared<spdlog::sinks::rotating_file_sink_mt>(path_, 500 * 1024 * 1024, 0);
+    // async
+    auto logger = std::make_shared<spdlog::async_logger>(
+        "logger", basic_sink, spdlog::thread_pool(), spdlog::async_overflow_policy::overrun_oldest);
+    logger->set_level(spdlog::level::trace);
+    // ms, thread_id
+    logger->set_pattern("[%Y-%m-%d %H:%M:%S.%e] [%t] %v");
+
+    logger_ = logger;
+
+    spdlog::register_logger(logger_);
+
+    // real-time refresh
+    logger_->flush_on(spdlog::level::trace);
+
+    inited_ = true;
+}
+#endif
+
 }  // namespace turbomind
diff --git a/src/turbomind/utils/logger.h b/src/turbomind/utils/logger.h
index 3e9f25e9e..a9bbc3f9f 100644
--- a/src/turbomind/utils/logger.h
+++ b/src/turbomind/utils/logger.h
@@ -20,6 +20,10 @@
 #include <map>
 #include <string>
 
+#ifndef _MSC_VER
+#include <spdlog/spdlog.h>
+#endif
+
 #include "src/turbomind/utils/string_utils.h"
 
 namespace turbomind {
@@ -30,11 +34,41 @@ namespace turbomind {
 #undef ERROR
 #endif
 
+#ifndef _MSC_VER
+class SpdLogger {
+public:
+    static SpdLogger& get_instance()
+    {
+        static SpdLogger spdlogger;
+        return spdlogger;
+    }
+    spdlog::logger* get_logger()
+    {
+        return logger_.get();
+    }
+
+    void set_log_path(const std::string& path)
+    {
+        path_ = path;
+    }
+    void init();
+
+private:
+    SpdLogger()                            = default;
+    ~SpdLogger()                           = default;
+    SpdLogger(const SpdLogger&)            = delete;
+    SpdLogger& operator=(const SpdLogger&) = delete;
+
+    std::string                     path_;
+    bool                            inited_ = false;
+    std::shared_ptr<spdlog::logger> logger_;
+};
+#endif
+
 class Logger {
 
 public:
-    enum Level
-    {
+    enum Level {
         TRACE   = 0,
         DEBUG   = 10,
         INFO    = 20,
@@ -47,17 +81,20 @@ class Logger {
         thread_local Logger instance;
         return instance;
     }
-    Logger(Logger const&) = delete;
+    Logger(Logger const&)         = delete;
     void operator=(Logger const&) = delete;
 
     template<typename... Args>
     void log(const Level level, const std::string format, const Args&... args)
     {
         if (level_ <= level) {
-            std::string fmt = getPrefix(level) + format + "\n";
-            // FILE*       out    = level_ < WARNING ? stdout : stderr;
+            std::string fmt    = getPrefix(level) + format + "\n";
             std::string logstr = fmtstr(fmt, args...);
+#ifdef _MSC_VER
             fprintf(stderr, "%s", logstr.c_str());
+#else
+            SpdLogger::get_instance().get_logger()->log(spdlog::level::trace, logstr);
+#endif
         }
     }
 
@@ -65,10 +102,13 @@ class Logger {
     void log(const Level level, const int rank, const std::string format, const Args&... args)
     {
         if (level_ <= level) {
-            std::string fmt = getPrefix(level, rank) + format + "\n";
-            // FILE*       out    = level_ < WARNING ? stdout : stderr;
+            std::string fmt    = getPrefix(level, rank) + format + "\n";
             std::string logstr = fmtstr(fmt, args...);
+#ifdef _MSC_VER
             fprintf(stderr, "%s", logstr.c_str());
+#else
+            SpdLogger::get_instance().get_logger()->log(spdlog::level::trace, logstr);
+#endif
         }
     }