diff --git a/megatron/core/inference/contexts/dynamic_context.py b/megatron/core/inference/contexts/dynamic_context.py
index 4267f9d0952..78f412ea5ab 100644
--- a/megatron/core/inference/contexts/dynamic_context.py
+++ b/megatron/core/inference/contexts/dynamic_context.py
@@ -1609,6 +1609,13 @@ def add_request(self, req: DynamicInferenceRequest, chunk_length: Optional[int]
         metadata_types = req.get_metadata_types()
         for m, m_type in zip(metadata, metadata_types):
             label, _, _ = m_type
+            if not isinstance(m, torch.Tensor):
+                m = torch.as_tensor(
+                    m,
+                    device=self.request_metadata[label].device,
+                    dtype=self.request_metadata[label].dtype,
+                )
+
             self.request_metadata[label][current_id] = m
 
         # Handle length and block assignments.
diff --git a/megatron/core/inference/engines/dynamic_engine.py b/megatron/core/inference/engines/dynamic_engine.py
index 1bb4ac73f44..86501ab3063 100644
--- a/megatron/core/inference/engines/dynamic_engine.py
+++ b/megatron/core/inference/engines/dynamic_engine.py
@@ -1,6 +1,7 @@
 # Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 
 import asyncio
+import concurrent.futures
 import logging
 import multiprocessing
 import os
@@ -1358,11 +1359,29 @@ async def async_step(
         # Keep for compatibility with current test suite.
         return ret
 
+    def _run_coroutine_sync(self, coro):
+        """Run a coroutine synchronously, handling the case when already in an event loop.
+
+        This method safely runs an async coroutine from synchronous code, even when
+        called from within an already running event loop (e.g., when used with async
+        frameworks like pytriton).
+        """
+        try:
+            # Check if there's already a running event loop
+            asyncio.get_running_loop()
+            # We're inside a running loop - run in a separate thread
+            with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
+                future = executor.submit(asyncio.run, coro)
+                return future.result()
+        except RuntimeError:
+            # No running loop - safe to use run_until_complete
+            return self._loop.run_until_complete(coro)
+
     def step_modern(
         self,
     ) -> Tuple[List[DynamicInferenceRequest], List[DynamicInferenceRequest], float]:
         """Synchronous wrapper for `self.async_step`."""
-        return self._loop.run_until_complete(self.async_step())
+        return self._run_coroutine_sync(self.async_step())
 
     def step_legacy(
         self, sampling_params: SamplingParams
@@ -1373,7 +1392,7 @@ def step_legacy(
             "0.16. Please use `step_modern()` going forward, which will eventually "
             "be renamed to `step()`."
         )
-        result = self._loop.run_until_complete(self.async_step())
+        result = self._run_coroutine_sync(self.async_step())
         active_requests = [self.get_request(i) for i in result["active_request_ids"]]
         finished_requests = [r.merge() for r in result["finished_request_records"]]
         return active_requests, finished_requests, result["step_time"]