MIE05E010010

Fault Symptom

The MindIE inference request fails. No parameter cache is found in the decode phase.

Fault Cause

The upper-layer application does not send a prefill request or continues to send a decode request after the request is stopped.

Fault Location

Based on the two possible causes, perform the following operations:

1. When calling the Generator.generate_token method, check whether a certain input_metadata contains an abnormal request_id. That is, check whether the value of is_prefill is False when the Generator.generate_token method is passed for the first time.

2. When calling the Generator.generate_token method, check whether a certain input_metadata contains request_id with inference completed. That is, check whether the output of the last iteration of request_id contains a non-zero EOS flag (0 indicates that the inference is not completed).

In the multi-device inference scenario, if the EOS identifiers of each device are inconsistent, this fault may also occur. Therefore, you need to check the input_metadata and EOS identifiers of each process.