Resetting fusion cache when fusion exceeds the max_fusions limit.

The FusionCache_CUDA test attempts to reset and get a fresh FusionCache.
If the currect cache contains more fusions than what is requested for
the new max_fusions limit it will fail to enforce the new max_fusions
constraint.

Author: mdavis36

python/python_frontend/fusion_cache.cpp

@@ -347,6 +347,27 @@ FusionCache* FusionCache::get(
     if (load_from_default_workspace && fs::exists(file_path)) {
       try {
         singleton_->deserialize(file_path);
+
+        // Check if deserialized cache exceeds max_fusions limit
+        if (singleton_->fusions_.size() > max_fusions) {
+          std::cout
+              << "Warning: Deserialized cache contains "
+              << singleton_->fusions_.size()
+              << " fusions, which exceeds the requested max_fusions limit of "
+              << max_fusions << ". Resetting cache." << std::endl;
+
+          // Delete incompatible workspace
+          std::error_code remove_ec;
+          fs::remove(file_path, remove_ec);
+          if (remove_ec) {
+            std::cout << "Failed to delete common workspace. Exception:\t"
+                      << remove_ec.message() << std::endl;
+          }
+
+          // Reset FusionCache
+          delete singleton_;
+          singleton_ = new FusionCache(max_fusions, selected_device);
+        }
       } catch (const std::exception& deserialize_exception) {
         // The saved workspace can become out-of-date between nvfuser updates.
         // Send warning and delete the incompatible workspace.