Release v1.1.6

Added

• An example workflow script to operate X nvcf function to run M jobs
• Upgrade vllm to 0.11.0
• Upgrade transformers to 4.57.0
• Agent context files for Codex, Claude, and Gemini
• Runner abstraction for pipeline execution
• Increase test coverage

Fixed

• Allow extra environment variables to be passed to the pixi runtime env
• Let slurm env setting override defaults inside container
• Remove dependency on pynvml
• Remove max_seq_len_to_capture from vLLM engine creation
• Improve the speed for final summary generation
• Downgrade click dep version to fix ray and revive e2e nvcf test

Release v1.1.5

Known Issues

• Caption ordering bug: Inflight batching (enabled by default) can assign captions to the wrong videos. Fixed in
  v1.2.0.

Added

• Upgrade to cosmos-xenna 0.1.6 for improved performance.

Changed

• Update default parameters for stages' cpu resource requests for higher throughput

Release v1.1.4

Added

• Add gpt-oss-20b as an option for EnhanceCaption stage.
• Enable batching for internvideo2 embedding stage for improved throughput.
• Upgrade to cosmos-xenna 0.1.5 for improved performance and stability.

Release v1.1.3

Added

• Release Grafana dashboard for pipeline monitoring.
• Add inflight batching for VLM captioning throughput.

Changed

• Merge video-splitting env into unified env.
• Improve Slurm instructions.

Release v1.1.2

Added

• Upgrade to cosmos-xenna 0.1.3 for improved scalability and observability.
• Enable Semantic Deduplication on Ray and improve IO efficiency for improved throughput.

Release v1.1.1

Added

• Add stage2 caption support to VLLMCaptionStage
• Add Nsight Systems for CUDA profiling

Fixed

• Avoid unnecessary post-install docker layers
• Pin Ray to the same version for both pixi and poetry
• Update slurm cli to work with pixi

Release v1.1.0

Added

• Use pixi to manage environments inside container image
• Use absolute URL for cosmos-xenna submodule; PLEASE run git submodule sync after pulling update
• Support for Cosmos-Reason1 as an alternative model for captioning
• Support for running Phi-4 with vLLM

Fixed

• Suppress warnings to make log more readable
• Make /dev/shm (and hence Ray object store) a fraction of system memory in local mode.

Release v1.0.2

Added

• Support for using multiple GPUs in captioning stage to enable large models
• Support for generating dataset to post-train Cosmos-Predict2
• Support for Phi-4 as an alternative model for captioning

Fixed

• PyNvCodec path for video decoding by fixing NVIDIA_DRIVER_CAPABILITIES env var
• CLI to import existing NVCF functions

Release v1.0.1

Added

• Multi-camera AV video split and caption pipelines
• Semantic-deduplication pipeline
• Support for Cosmos-Embed1 embedding model
• Support for using pre-signed URLs as input and output paths

Fixed

• Splitting & transcoding accuracy for MPEG-TS files

Changed

• Update required python version from 3.10.14 to 3.10.18

Security

• Upgrade base image and packages to mitigate security vulnerabilities

Release v1.0.0

Added

• Initial version