Edge inference times are EMA-smoothed twice, with a hard-coded alpha=0.2

[ff225]

Priority: High

Edge per-layer inference times get an exponential moving average applied twice, and the first application ignores the configured offloading_algo.ema_alpha.

What happens

1. model_manager.py:212-216 smooths self.inference_times with a hard-coded alpha = 0.2:

   if layer_key in self.inference_times:
       alpha = 0.2  # Weight for new measurement
       self.inference_times[layer_key] = (
           alpha * elapsed_time + (1 - alpha) * self.inference_times[layer_key]
       )

2. That dict is device_profiles[device_id]["edge_inference_times"] — it is passed by reference (edge_initialization.py:140,152 → model_manager.py:285), as the comment at model_manager.py:231 notes.
3. request_handler.py then applies the EMA again to the same dict, this time with the configured alpha, using the times returned by Edge.run_inference.

Why it matters

• Over-smoothing: edge times are smoothed twice, device times once. The two sides of the offloading cost comparison respond to change at different rates, so the edge estimate lags and the offloading point adapts slower than intended (relates to Verify that the offloading point adapts to changing conditions #33, Verify correctness of offloading algorithm #32).
• Config is bypassed: Make EMA alpha configurable instead of hard-coded 0.5 #9 (closed by Make EMA alpha configurable instead of hard-coded 0.5 #56) made the EMA alpha configurable, but the edge path still hard-codes 0.2. Changing ema_alpha in settings.yaml only affects the second application.

Suggested fix

Decide which layer owns the smoothing. Most likely model_manager should report the raw elapsed_time and leave all EMA work to the state manager (DeviceStateManager.update_edge_times), which already reads the configured alpha.

Files

• src/server/models/model_manager.py:211-218
• src/server/edge/edge_initialization.py:140,152
• src/server/communication/device_state.py (update_edge_times)

Found while refactoring RequestHandler for #11 / #45. The refactor preserves the current behaviour exactly; this is a pre-existing bug.