Is this page helpful?

Release Notes for NVIDIA NIM for Object Detection#

This documentation contains the release notes for NVIDIA NIM for Object Detection.

Release 2.0.1#

This patch release includes the following updates:

  • Adds the com.nvidia.nim.model_download_protocol=native-v1 container label for NIM Operator deployment compatibility (requires NIM Operator >= 3.1.2).

  • The non-default RUST_LOG configuration is removed so that logging is controlled by the product default or explicit user configuration.

Release 2.0#

This release represents a major runtime upgrade for the nvidia/nemotron-object-detection NIM that includes a new purpose-built runtime. The release provides the choice of both latency-optimized and throughput-optimized runtime configurations.

Release 2.0 supports the following models:

  • nvidia/nemotron-page-elements-v3

  • nvidia/nemotron-table-structure-v1

The nvidia/nemotron-graphic-elements-v1 model is not included in this release.

Highlights#

This release contains the following key changes:

  • Added support for multiple object detection models in a single NIM image.

    • Each container serves one configured model.

    • By default, the container loads nvidia/nemotron-page-elements-v3 from /model/page-elements.

    • To serve nvidia/nemotron-table-structure-v1 instead, set NIM_ENGINE_MODEL_NAME to nvidia/nemotron-table-structure-v1 and NIM_ENGINE_MODEL_PATH to /model/table-structure.

  • Added HTTP-endpoint-based routing:

    • POST /v1/page-elements

    • POST /v1/table-structure

  • Added the combined OpenAPI specification for the object detection server. For details, refer to API Reference for NVIDIA NIM for Object Detection.

  • Added a Docker health check that uses /v1/health/ready and supports custom server ports and TLS deployments.

  • Replaced model-profile selection with explicit runtime settings. No manual profile selection steps are required.

  • Added new environment variables. For details, refer to Environment Variables for NVIDIA NIM for Object Detection.

    • Added NIM_ENGINE_MODEL_NAME – Selects the supported object detection model to load.

    • Added NIM_PERFORMANCE_MODE – Selects latency-oriented defaults when set to 0 or false (default), and throughput-oriented defaults when set to 1 or true.

    • Added Hugging Face and NGC model download providers with automatic credential-based selection.

    • Added controls for engine count, batching, stage-specific warmup, GPU preprocessing, and more.

    • Added model-download-only and precompile-only startup modes for staging artifacts and precompiled CUDA kernels.

    • Added optional KServe V2 gRPC, HTTPS, queue, timeout, and structured logging configuration.

Support Matrix Updates#

For details, refer to Support Matrix for NVIDIA NIM for Object Detection.

  • Added support for NVIDIA GB200 and GB10 systems.

Compatibility Updates#

Most environment variables beginning with NIM_TRITON_ are not supported by the 2.0 runtime. The runtime emits a deprecation warning when an alias is used. For new deployments, update your code to use the current environment variables.

The following deprecated compatibility aliases remain available for migration.

Canonical variable

Deprecated compatibility aliases

NIM_SERVER_MAX_WAIT_MS

NIM_TRITON_DYNAMIC_BATCHING_MAX_QUEUE_DELAY_MICROSECONDS

NIM_ENGINE_COUNT

NIM_TRITON_MODEL_INSTANCE_COUNT

NIM_PIPELINE_MAX_BATCH_SIZE

NIM_TRITON_DATA_MAX_BATCH_SIZE, NIM_TRITON_MAX_BATCH_SIZE, NIM_TRITON_MODEL_MAX_BATCH_SIZE

NIM_SERVER_MAX_QUEUE_SIZE

NIM_TRITON_PIPELINE_MAX_BATCH_SIZE

NIM_SERVER_REQUEST_TIMEOUT_S

NIM_TRITON_REQUEST_TIMEOUT

NIM_PIPELINE_TIMING_TELEMETRY

NIM_TRITON_ENABLE_PIPELINE_TIMING

NIM_ENGINE_MODEL_PATH

NIM_OBJECT_DETECTION_MODEL_PATH

Note

Compatibility alias values are parsed by using the units of the current environment variable. In particular, the value of NIM_TRITON_DYNAMIC_BATCHING_MAX_QUEUE_DELAY_MICROSECONDS is interpreted as milliseconds by the 2.0 runtime.

All Known Issues#

The known issues for Object Detection NIM are the following:

  • Helm nim.modelName is not propagated. Use the appropriate runtime model-selection environment variable directly. For object detection, use envVars.NIM_ENGINE_MODEL_NAME.

  • Use persistence.storageClass not persistence.class.

  • Do not combine persistence.enabled=true with hostPath.enabled=true. Currently, the chart mounts the PVC and ignores hostPath.

  • NIM_PIPELINE_MAX_BATCH_SIZE supports values from 1 through 255. Values higher than 255 are rejected at startup.

Previous Versions#