diff --git a/paxml/programs.py b/paxml/programs.py
index a85fabdab..029f0adb3 100644
--- a/paxml/programs.py
+++ b/paxml/programs.py
@@ -376,7 +376,12 @@ def run(self, state: TrainState, step: int) -> TrainProgramOutput:
     logging.log_first_n(
         logging.INFO, '[PAX STATUS]:  Writing summaries (attempt).', 5
     )
-    steps_per_sec = self._maybe_write_summaries(step, new_step, train_outputs)
+    steps_per_sec = self._maybe_write_summaries(
+	step,
+    	new_step,
+	train_outputs,
+	train_p.compute_steps_per_sec_interval_steps
+    )
 
     # Run eval at regular step interval.
     # While the eval ones below are post-model weight updates, hence we use the
@@ -456,11 +461,12 @@ def train_input_partition_spec(
     """Returns the partition spec for the model training inputs."""
 
   def _maybe_write_summaries(
-      self, step: int, new_step: int, train_outputs: StepFnOutput
+      self,
+      step: int,
+      new_step: int,
+      train_outputs: StepFnOutput,
+      compute_steps_per_sec_interval_steps: int
   ) -> float | None:
-    # Compute steps/sec every this many steps, revisit when necessary.
-    compute_steps_per_sec_interval_steps = 10
-
     steps_per_sec = None
     should_compute_steps_per_sec = (
         new_step % compute_steps_per_sec_interval_steps == 0
diff --git a/paxml/tasks_lib.py b/paxml/tasks_lib.py
index 2b51f64ca..811dc54d4 100644
--- a/paxml/tasks_lib.py
+++ b/paxml/tasks_lib.py
@@ -1240,6 +1240,8 @@ class Train:
       decode_start_after_n_steps: Starts decoder after N steps.
       decode_use_ema_states: If True, use ema states to run decode during train,
         note that in this case ema MUST be enabled in the learner.
+      compute_steps_per_sec_interval_steps: number of steps to average over when
+        computing steps/sec.
       profiler_num_steps: The number of steps to be captured by the profiler
         based on the step time estimate.
       profiler_min_duration_sec: The minimum duration to be captured by the
@@ -1302,6 +1304,7 @@ class Train:
     decode_start_after_n_steps: int = 0
     # TODO(zhishuai): verify this for a pjit model.
     decode_use_ema_states: bool = False
+    compute_steps_per_sec_interval_steps: int = 10
     profiler_num_steps: int = 2
     profiler_min_duration_sec: float = 1.0
     profiler_capture_step: int | None = None