diff --git a/olmocr/pipeline.py b/olmocr/pipeline.py
index c6f24f9..e4d5d14 100644
--- a/olmocr/pipeline.py
+++ b/olmocr/pipeline.py
@@ -624,12 +624,12 @@ async def vllm_server_task(model_name_or_path, args, semaphore, unknown_args=Non
 
     # Shared variables between tasks
     last_running_req, last_queue_req = 0, 0
-    prev_running_req_at_release = 0  # Track running requests at last semaphore release
+    running_reqs_decreased = False
     server_printed_ready_message = False
     last_semaphore_release = time.time()
 
     async def process_line(line):
-        nonlocal last_running_req, last_queue_req, last_semaphore_release, server_printed_ready_message
+        nonlocal last_running_req, last_queue_req, running_reqs_decreased, last_semaphore_release, server_printed_ready_message
         server_logger.info(line)
 
         if "Detected errors during sampling" in line:
@@ -640,12 +640,15 @@ async def vllm_server_task(model_name_or_path, args, semaphore, unknown_args=Non
             server_printed_ready_message = True
             last_semaphore_release = time.time()
 
-        match = re.search(r"Running: (\d+)", line)
-        if match:
-            last_running_req = int(match.group(1))
+        if match := re.search(r"Running: (\d+)", line):
+            current_running = int(match.group(1))
+            # Check for negative derivative (decrease in running requests), to not overload VLLM
+            if current_running < last_running_req:
+                running_reqs_decreased = True
+                logger.info(f"Running requests decreased: {last_running_req} -> {current_running}")
+            last_running_req = current_running
 
-        match = re.search(r"(?:Waiting|Pending):\s*(\d+)", line)
-        if match:
+        if match := re.search(r"(?:Waiting|Pending):\s*(\d+)", line):
             last_queue_req = int(match.group(1))
             logger.info(f"vllm running req: {last_running_req} queue req: {last_queue_req}")
 
@@ -661,25 +664,25 @@ async def vllm_server_task(model_name_or_path, args, semaphore, unknown_args=Non
                 logger.warning(f"Got {ex} when reading log line from inference server, skipping")
 
     async def timeout_task():
-        nonlocal last_running_req, last_queue_req, last_semaphore_release, prev_running_req_at_release
+        nonlocal last_running_req, last_queue_req, last_semaphore_release, running_reqs_decreased
         try:
             while True:
                 await asyncio.sleep(1)
-                
+
                 # Check if we should release the semaphore
                 should_release = (
                     server_printed_ready_message
                     and last_queue_req == 0
                     and time.time() - last_semaphore_release > 30
                     and semaphore.locked()
-                    and (last_running_req == 0 or last_running_req < prev_running_req_at_release)
+                    and (last_running_req == 0 or running_reqs_decreased)
                 )
-                
+
                 if should_release:
                     semaphore.release()
-                    prev_running_req_at_release = last_running_req
+                    running_reqs_decreased = False  # Reset flag after release
                     last_semaphore_release = time.time()
-                    logger.info(f"Semaphore released, allowing a worker to proceed. Running requests: {last_running_req} (prev: {prev_running_req_at_release})")
+                    logger.info(f"Semaphore released, allowing a worker to proceed. Running requests: {last_running_req}")
         except asyncio.CancelledError:
             pass  # Clean up if the task is cancelled