UnlinkLlmClusters
Applicability
Product |
Supported (Yes/No) |
|---|---|
Atlas 350 Accelerator Card |
No |
Yes |
|
Yes |
|
No |
|
No |
|
No |
Note: For the
Function Description
Disconnects links between devices.
Prototype
1 | Status UnlinkLlmClusters(const std::vector<ClusterInfo> &clusters, std::vector<Status> &rets, int32_t timeout_in_millis = 1000, bool force_flag = false) |
Parameters
Parameter |
Input/Output |
Description |
|---|---|---|
clusters |
Input |
Information about the clusters whose links are to be disconnected. The type is ClusterInfo. |
rets |
Output |
Link disconnection result of each cluster. |
timeout_in_millis |
Input |
Link disconnection timeout interval, in milliseconds. The default value is 1000. |
force_flag |
Input |
Whether to perform forcible disconnection. By default, this feature is disabled. For forcible disconnection, only the local link is forcibly disconnected; therefore, this API must be called on both sides. Non-forcible disconnection is initiated by the Decode side. If no link fault occurs, the links on both ends will be disconnected automatically. In case of a link fault, the disconnection process takes a long time and requires an additional call to be initiated on the Prompt side. |
Returns
- SUCCESS: A success message is returned only when links are successfully disconnected for all clusters.
- Other values: Link disconnection failed. Check the link disconnection result of each cluster in rets.
Exception Handling
- LLM_PROCESSING_LINK: The API is protected by a lock. When multiple threads of an LLM-DataDist instance call the link disconnection API, the calls are executed serially. If the waiting time of other threads exceeds the configured timeout period, an error is reported and the threads exit.
- LLM_UNLINK_FAILED: Link disconnection failed.
Constraints
This API must be called after the Initialize API is called to complete initialization.