Skip to content

Latest commit

 

History

History
178 lines (149 loc) · 8.38 KB

model_management.md

File metadata and controls

178 lines (149 loc) · 8.38 KB
Raw

Model Management

Triton provides model management APIs are part of the HTTP/REST and GRPC protocols, and as part of the C API. Triton operates in one of three model control modes: NONE, EXPLICIT or POLL. The model control mode determines how changes to the model repository are handled by Triton and which of these protocols and APIs are available.

Model Control Mode NONE

Triton attempts to load all models in the model repository at startup. Models that Triton is not able to load will be marked as UNAVAILABLE and will not be available for inferencing.

Changes to the model repository while the server is running will be ignored. Model load and unload requests using the model control protocol will have no affect and will return an error response.

This model control mode is selected by specifying --model-control-mode=none when starting Triton. This is the default model control mode. Changing the model repository while Triton is running must be done carefully, as explained in Modifying the Model Repository.

Model Control Mode EXPLICIT

At startup, Triton loads only those models specified explicitly with the --load-model command-line option. If --load-model is not specified then no models are loaded at startup. Models that Triton is not able to load will be marked as UNAVAILABLE and will not be available for inferencing.

After startup, all model load and unload actions must be initiated explicitly by using the model control protocol. The response status of the model control request indicates success or failure of the load or unload action. When attempting to reload an already loaded model, if the reload fails for any reason the already loaded model will be unchanged and will remain loaded. If the reload succeeds, the newly loaded model will replace the already loaded model without any loss in availability for the model.

This model control mode is enabled by specifying --model-control-mode=explicit. Changing the model repository while Triton is running must be done carefully, as explained in Modifying the Model Repository.

Model Control Mode POLL

Triton attempts to load all models in the model repository at startup. Models that Triton is not able to load will be marked as UNAVAILABLE and will not be available for inferencing.

Changes to the model repository will be detected and Triton will attempt to load and unload models as necessary based on those changes. When attempting to reload an already loaded model, if the reload fails for any reason the already loaded model will be unchanged and will remain loaded. If the reload succeeds, the newly loaded model will replace the already loaded model without any loss of availability for the model.

Changes to the model repository may not be detected immediately because Triton polls the repository periodically. You can control the polling interval with the --repository-poll-secs option. The console log or the model ready protocol or the index operation of the model control protocol can be used to determine when model repository changes have taken effect.

WARNING: There is no synchronization between when Triton polls the model repository and when you make any changes to the repository. As a result Triton could observe partial and incomplete changes that lead to unexpected behavior. For this reason POLL mode is not recommended for use in production environments.

Model load and unload requests using the model control protocol will have no affect and will return an error response.

This model control mode is enabled by specifying --model-control-mode=poll and by setting --repository-poll-secs to a non-zero value when starting Triton. Changing the model repository while Triton is running must be done carefully, as explained in Modifying the Model Repository.

In POLL mode Triton responds to the following model repository changes:

  • Versions may be added and removed from models by adding and removing the corresponding version subdirectory. Triton will allow in-flight requests to complete even if they are using a removed version of the model. New requests for a removed model version will fail. Depending on the model's version policy, changes to the available versions may change which model version is served by default.

  • Existing models can be removed from the repository by removing the corresponding model directory. Triton will allow in-flight requests to any version of the removed model to complete. New requests for a removed model will fail.

  • New models can be added to the repository by adding a new model directory.

  • The model configuration file (config.pbtxt) can be changed and Triton will unload and reload the model to pick up the new model configuration.

  • Label(s) files providing labels for outputs that represent classifications can be added, removed, or modified and Triton will unload and reload the model to pick up the new labels. If a label file is added or removed the corresponding edit to the label_filename property of the output it corresponds to in the model configuration must be performed at the same time.

Modifying the Model Repository

Each model in a model repository resides in its own sub-directory. The activity allowed on the contents of a model's sub-directory varies depending on how Triton is using that model. The state of a model can be determined by using the model metadata or repository index APIs.

  • If the model is actively loading or unloading, no files or directories within that sub-directory must be added, removed or modified.

  • If the model has never been loaded or has been completely unloaded, then the entire model sub-directory can be removed or any of its contents can be added, removed or modified.

  • If the model has been completely loaded then any files or directories within that sub-directory can be added, removed or modified; except for shared libraries implementing the model's backend. Triton uses the backend shared libraries while the model is loading so removing or modifying them will likely cause Triton to crash. To update a model's backend you must first unload the model completely, modify the backend shared libraries, and then reload the model. On some OSes it may also be possible to simply move the existing shared-libraries to another location outside of the model repository, copy in the new shared libraries, and then reload the model.