issue/811 use relax graph capture mode, add compile flag for graph instantiate

src/infinicore/graph/graph.cc

@@ -84,7 +84,7 @@ void Graph::instantiate() {
 
     if (infinirtStreamBeginCapture(
             context::getStream(),
-            INFINIRT_STREAM_CAPTURE_MODE_GLOBAL)
+            INFINIRT_STREAM_CAPTURE_MODE_RELAXED)
         != INFINI_STATUS_SUCCESS) {
         return;
     }
@@ -144,7 +144,9 @@ std::shared_ptr<Graph> GraphManager::stop_recording() {
         return nullptr;
     }
     recording_ = false;
+#ifdef USE_INFINIRT_GRAPH
     graph_->instantiate();
+#endif
     return std::exchange(graph_, nullptr);
 }
 

src/infinicore/tensor/view.cc

@@ -2,6 +2,8 @@
 #include "infinicore/dtype.hpp"
 #include "infinicore/tensor.hpp"
 
+#include "../utils.hpp"
+
 #include <spdlog/spdlog.h>
 #include <stdexcept>
 
@@ -62,11 +64,11 @@ Tensor TensorImpl::narrow(const std::vector<TensorSliceParams> &slices) const {
 
 Tensor TensorImpl::permute(const Shape &order) const {
     // Validate input
-    assert(meta_.shape.size() == order.size());
+    INFINICORE_ASSERT(meta_.shape.size() == order.size());
 
     // Check that order contains all indices from 0 to n-1 exactly once
     for (size_t i = 0; i < order.size(); i++) {
-        assert(std::find(order.begin(), order.end(), i) != order.end());
+        INFINICORE_ASSERT(std::find(order.begin(), order.end(), i) != order.end());
     }
 
     // Permute shape and strides

xmake.lua

@@ -205,6 +205,18 @@ if has_config("ninetoothed") then
     add_defines("ENABLE_NINETOOTHED")
 end
 
+-- cuda graph
+option("graph")
+    set_default(false)
+    set_showmenu(true)
+    set_description("Whether to use device graph instantiating feature, such as cuda graph for nvidia")
+option_end()
+
+if has_config("graph") then
+    add_defines("USE_INFINIRT_GRAPH")
+end
+
+
 -- InfiniCCL
 option("ccl")
     set_default(false)