tft_register_mindx_callback

This API is called by MindCluster to register the callback function of the repair process with MindIO TFT.

mindio_ttp.controller_ttp.tft_register_mindx_callback(action: str, func: Callable)

Parameter	Mandatory/Optional	Description	Value
action	Mandatory	Name of the action to be registered by the callback function.	The value is of the string type. The following action names are supported: report_fault_ranks report_stop_complete report_strategies report_result
func	Mandatory	Function to be registered.	The callback function cannot be empty. For details about the input parameters of the callback function, see Table 1 to Table 4.

Parameter

Mandatory/Optional

Description

Value

action

Mandatory

Name of the action to be registered by the callback function.

The value is of the string type. The following action names are supported:

func

Mandatory

Function to be registered.

The callback function cannot be empty. For details about the input parameters of the callback function, see Table 1 to Table 4.

**Table 1** Callback function parameters when action is set to report_fault_ranks
Parameter	Mandatory/Optional	Description	Value
error_rank_dict	-	Information about the faulty NPU.	<int key, int errorType> dictionary: key: rank ID of the faulty NPU errorType: fault type 0: UCE 1: non-UCE fault

**Table 2** Callback function parameters when action is set to report_stop_complete
Parameter	Mandatory/Optional	Description	Value
code	-	Action execution result.	0: success 400: common error 401: MindCluster task ID not found 402: model error 403: sequence error 404: Processor not ready
msg	-	Message indicating whether the training stops.	String
error_rank_dict	-	Information about the faulty NPU.	<int key, int errorType> dictionary: key: rank ID of the faulty NPU errorType: fault type 0: UCE 1: non-UCE fault

**Table 3** Callback function parameters when action is set to report_strategies
Parameter	Mandatory/Optional	Description	Value
error_rank_dict	-	Information about the faulty NPU.	<int key, int errorType> dictionary: key: rank ID of the faulty NPU errorType: fault type 0: UCE 1: non-UCE fault
strategy_list	-	List of repair policies supported by MindIO TFT based on the current available replica information.	The value is of the list type. The supported repair policies (string) are as follows: retry: UCE repair recover: ARF repair dump: dying gasp exit

**Table 4** Callback function parameters when action is set to report_result
Parameter	Mandatory/Optional	Description	Value
code	-	Action execution result.	0: The repair is successful. 405: The retry policy fails. The recover, dump, and exit policies are supported. 406: The repair fails. The dump or exit repair policy is supported. 499: The repair fails. Only the exit policy is supported.
msg	-	Message indicating repair success or failure.	String
error_rank_dict	-	Information about the faulty NPU.	<int key, int errorType> dictionary: key: rank ID of the faulty NPU errorType: fault type 0: UCE 1: non-UCE fault
curr_strategy	-	Current repair policy.	The value is of the string type. For details about the value, see strategy_list in Table 3.

Parent topic: API Reference