The KV transfer metrics log line currently shows raw numpy scalar reprs when the stats come from a connector inside a MultiConnector:

(APIServer pid=3547631) INFO 09-15 19:44:52 [metrics.py:107] KV Transfer metrics: NixlPullConnector={'Num successful transfers': 8, 'Avg xfer time (ms)': np.float64(2.338), 'P90 xfer time (ms)': 2.772, 'Avg post time (ms)': np.float64(1.111), 'P90 post time (ms)': 1.88, 'Avg MB per transfer': np.float64(25.441), 'Throughput (MB/s)': np.float64(10883.285), 'Avg number of descriptors': np.float64(133.5)}

The single-connector path was fine because the log line formats each value with an f-string (str()), which renders np.float64 as a plain number. With a MultiConnector, reduce() returns a nested dict and the whole dict gets rendered with str(), which formats every value with repr() — and numpy 2 renders repr(np.float64(2.338)) as np.float64(2.338).

The fix is to have the nixl/mooncake/hf3fs reduce() implementations return plain Python scalars via .item(), which is what their dict[str, int | float] annotation already promised. After the fix the same line reads:

KV Transfer metrics: NixlPullConnector={'Num successful transfers': 8, 'Avg xfer time (ms)': 2.338, 'P90 xfer time (ms)': 2.772, 'Avg post time (ms)': 1.111, 'P90 post time (ms)': 1.88, 'Avg MB per transfer': 25.441, 'Throughput (MB/s)': 10883.285, 'Avg number of descriptors': 133.5}

Assisted-by: Mistral Vibe