diff --git a/transformer_engine/common/init_python.cpp b/transformer_engine/common/init_python.cpp
index c7a2ac5876..79baa7fc3d 100644
--- a/transformer_engine/common/init_python.cpp
+++ b/transformer_engine/common/init_python.cpp
@@ -56,8 +56,28 @@ void warn_initialization_failed() {
       "registration failed. Using CUDA backend as fallback.");
 }
 
+// Global counter to keep track of how many need to pause GC. We can only resume GC when no one is pausing it.
+int num_gc_pauses = 0;
+int gc_was_enabled = 0;
+
 }  // namespace
 
+// Definition of PythonGCPause's methods declared in tvm_ffi_bridge.h.
+// They are defined here because only this file includes Python.h and can use the CPython API
+PythonGCPause::PythonGCPause() : active_(Py_IsInitialized()) {
+  if (!active_) return;
+  const PyGILState_STATE gil = PyGILState_Ensure();
+  if (num_gc_pauses++ == 0) gc_was_enabled = PyGC_Disable();
+  PyGILState_Release(gil);
+}
+
+PythonGCPause::~PythonGCPause() {
+  if (!active_) return;
+  const PyGILState_STATE gil = PyGILState_Ensure();
+  if (--num_gc_pauses == 0 && gc_was_enabled) PyGC_Enable();
+  PyGILState_Release(gil);
+}
+
 // Initialize the Python interpreter and import the CuTeDSL backend module. This is only compiled
 // with NVTE_WITH_CUTEDSL=ON in CMake and will be ignored otherwise
 bool initialize_python_cutedsl_backend() {
diff --git a/transformer_engine/common/tvm_ffi_bridge.h b/transformer_engine/common/tvm_ffi_bridge.h
index d57fd8f03d..5305e56b4f 100644
--- a/transformer_engine/common/tvm_ffi_bridge.h
+++ b/transformer_engine/common/tvm_ffi_bridge.h
@@ -35,6 +35,40 @@ namespace tvm_ffi_bridge {
 
 bool initialize_python_cutedsl_backend();
 
+// A helper class to pause python GC temporarily. This is needed for CuTeDSL kernels compilation
+// which is performed by python, which would trigger GC if it's not paused.
+
+// This is a workaround for the issue that occurs due to the following factors:
+// 1. In TE JAX, we do a temporary CUDA graph capture for operations to trigger their "lazy load"
+//    before executing them. See https://github.com/NVIDIA/TransformerEngine/pull/2219 for context.
+// 2. This process only happens in C++ until we introduce CuTeDSL kernels, where we exit C++ and
+//    enter Python again to compile them.
+// 3. During this process, Python GC can be triggered if too many objects are created to compile
+//    the kernel.
+// 4. This GC cleans up JAX objects from earlier that are no longer used, which ultimately may cause
+//    the destruction of a `GpuExecutable` and its "thunks".
+// Sidenote: JAX executables are compiled and executed by XLA, which has a "thunk"-based runtime,
+//    where a thunk represents a single kernel, cuBLAS call, cuDNN call, custom call, etc.
+// 5. The problem is that one of these thunks (for example, `CollectiveKernelThunk`) may own some
+//    CUDA virtual memory, and when it's destroyed, it will attempt to free that memory.
+// 6. Before this virtual memory is deallocated, it must wait for in-flight GPU operations that may
+//    still use it, so it calls `cuCtxSynchronize` to wait for all work in that GPU's context.
+// 7. Synchronizing the GPU during a CUDA graph capture is invalid and triggers
+//    `CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED`, which also invalidates the capture.
+// Note: TE's graph capture is not visible to XLA, so XLA can't know it's unsafe to free its virtual
+// memory at this point.
+class PythonGCPause {
+ public:
+  // The class method definitions are in init_python.cpp.
+  PythonGCPause();
+  ~PythonGCPause();
+  PythonGCPause(const PythonGCPause &) = delete;
+  PythonGCPause &operator=(const PythonGCPause &) = delete;
+
+ private:
+  bool active_;
+};
+
 inline DLDataType convert_to_dltype(NVTEDType type) {
   switch (type) {
     case kNVTEFloat32:
@@ -258,8 +292,13 @@ class TVMFFICentral {
       return std::nullopt;
     }
     const std::string key = cfg.to_key();
+    bool success = false;
+    {
+      PythonGCPause gc_pause;
+      success = cfg.retrieve_func_from_python(key);
+    }
     std::optional<tvm::ffi::Function> fn =
-        cfg.retrieve_func_from_python(key) ? tvm::ffi::Function::GetGlobal(key) : std::nullopt;
+        success ? tvm::ffi::Function::GetGlobal(key) : std::nullopt;
     if (!fn) {
       maybe_warn_not_chosen("no TVM-FFI kernel is registered for config `", key, "`.");
     }
diff --git a/transformer_engine/jax/__init__.py b/transformer_engine/jax/__init__.py
index 45379dbf0d..7505a53ab6 100644
--- a/transformer_engine/jax/__init__.py
+++ b/transformer_engine/jax/__init__.py
@@ -32,12 +32,6 @@
 
 from transformer_engine.common import load_framework_extension
 
-if os.getenv("NVTE_ENABLE_CUTEDSL_BACKEND", "0") != "0":
-    warnings.warn(
-        "TE/JAX is not yet compatible with CuTeDSL due to XLA, TE/JAX, and Python GC conflicts. "
-        "We are working on a fix; in the meantime NVTE_ENABLE_CUTEDSL_BACKEND is forced off."
-    )
-os.environ["NVTE_ENABLE_CUTEDSL_BACKEND"] = "0"
 
 load_framework_extension("jax")