Inference Builder 1.0.0

Summary

Release Inference Builder 1.0.0 with agent integrations, remotely hosted MCP support, hardware-accelerated video output, expanded backend templates, DeepStream 9.1 platform updates, and reliability improvements.

New features

• Add Agent Skill support for Claude Code and OpenAI Codex, including installation tooling, configuration, documentation, and evaluations.
• Add Claude plugin metadata and repository-specific claude.md guidance.
• Expand generated API support with NIM and Triton responder and route templates.
• Add remotely hosted MCP service support:
  • Streamable HTTP transport
  • Optional Bearer-token authentication
  • Per-client session workspaces
  • Artifact and log downloads
  • Server start, stop, status, and log management
• Add hardware-accelerated H.264 MP4 output using NVENC:
  • Return encoded video as a file or registered asset
  • Import generated videos through AssetManager
  • Add schema validation and output-type handling
  • Add unit and integration coverage
• Add force_cpu output configuration so DeepStream tensors can remain on GPU unless CPU transfer is explicitly required.
• Add improved Unicode string handling.
• Expand samples and validation coverage for video output, asset management, concurrency, frame sampling, output types, and Unicode inputs.
• Update the release version to 1.0.0.

Fixes and improvements

• Fix high-batch workloads failing due to low file-descriptor limits by raising the soft limit in container entrypoints.
• Avoid unnecessary DeepStream GPU-to-CPU tensor transfers for improved inference performance.
• Fix CUDA 13.2 configuration in Jetson and DGX Spark Dockerfiles.
• Fix Tegra HPC-X UCX/UCC library-path compatibility.
• Prevent numerical overflow in GDINO mask sigmoid processing.
• Update DeepStream containers to public 9.1 images for x86_64, Jetson, and DGX Spark.
• Update the NVCLIP optimizer to Triton Server 26.03.
• Upgrade Protobuf to 6.33.5 and Pillow to 12.2.0 for compatibility and security fixes.
• Add Sonar configuration, third-party license documentation, and expanded validation tests.
• Improve model-download, platform, Jetson storage, and sample configuration documentation.
  Before opening the PR, note:

Inference Builder 0.0.2

New Features:

vLLM backend — New backend support with Cosmos and Qwen3-VL sample configs, including NVDEC HW decode variants
CRADIO sample — New DeepStream-based radio detection pipeline with ONNX export and TensorRT engine prep
MCP server — Model Context Protocol integration for AI-assisted config authoring (mcp/)
JSON Schema validation — Full schema definitions for all backends and config validation (schemas/)
Error handling framework — Comprehensive error module (lib/errors.py, +1163 lines)
Live streaming support — New TAO live stream config and add_live_stream responder
Asset management & concurrency tests — New test suites for asset management and concurrent requests
DGX Spark Dockerfiles — New Dockerfile.dgxspark for ds_app and vllm samples

Enhancements:

Major refactor of lib/inference.py (+1,420 / -300) — improved pipeline handling
Enhanced builder/validate.py (+365) — stronger config validation
Updated lib/asset_manager.py and lib/utils.py
Expanded test framework
Updated Dockerfiles and configs across changenet, ds_app, nvclip, qwen, tao samples
DeepStream backend template significantly expanded.

Fixes:

Security vulnerability fixes
Removed internal inferencemodeltoolkit dependency and inaccessible urm.nvidia.com references (Fixes Changenet example dockerfile build failure #4)
Jetson VIC compose failure fix
Type error fix for source_config
Model download tools updated

Initial Releae

Initial release of the Inference Builder tools.