feat: update v5.1 (#281)

NVIDIA · Aug 16, 2022 · bc21406 · bc21406
1 parent e1b931f
commit bc21406
Show file tree

Hide file tree

Showing 604 changed files with 57,827 additions and 29,517 deletions.
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
diff --git a/.gitlab/issue_templates/bug.md b/.gitlab/issue_templates/bug.md
diff --git a/.gitlab/issue_templates/feature.md b/.gitlab/issue_templates/feature.md
diff --git a/.gitlab/merge_request_templates/merge.md b/.gitlab/merge_request_templates/merge.md
diff --git a/3rdparty/INIReader.h b/3rdparty/INIReader.h
@@ -316,7 +316,7 @@ class INIReader
     // Construct INIReader and parse given file. See ini.h for more info
     // about the parsing.
     INIReader(FILE *file);
-
+    ~INIReader();
     // Return the result of ini_parse(), i.e., 0 on success, line number of
     // first error on parse error, or -1 on file open error.
     int ParseError() const;
@@ -384,6 +384,8 @@ inline int INIReader::ParseError() const
     return _error;
 }
 
+inline INIReader::~INIReader() { }
+
 inline const std::set<std::string>& INIReader::Sections() const
 {
     return _sections;

diff --git a/3rdparty/trt_fused_multihead_attention/fused_multihead_attention.h b/3rdparty/trt_fused_multihead_attention/fused_multihead_attention.h
@@ -313,8 +313,15 @@ class TFusedMHAKernelFactory
 
     static TFusedMHAKernelFactory<TFusedMHAKernelList>& Get()
     {
-        static TFusedMHAKernelFactory<TFusedMHAKernelList> s_factory;
-        return s_factory;
+        int device_id;
+        cudaGetDevice(&device_id);
+        static std::unique_ptr<TFusedMHAKernelFactory<TFusedMHAKernelList>> s_factory[32] = {nullptr};
+        if (s_factory[device_id] == nullptr) {
+            assert(device_id <= 32);
+            s_factory[device_id] = std::make_unique<TFusedMHAKernelFactory<TFusedMHAKernelList>>(TFusedMHAKernelFactory<TFusedMHAKernelList>());
+        }
+
+        return *(s_factory[device_id]);
     }
 
 private: