ReaderLM-v2 Jetson runtime archives
Fixed arm64 Docker/OCI archives for LZC AI Pod ReaderLM v2 0.1.1. No ReaderLM weights are included. Download the GGUF or NVFP4 model separately from manateelazycat/ReaderLM-v2-GGUF or manateelazycat/ReaderLM-v2-NVFP4.
- Orin: JetPack 6.2 / R36.4.3 / CUDA 12.6 / SM87, pinned llama.cpp native runtime.
- Thor: JetPack 7.2 / R39.2 / CUDA 13.2 / SM110, vLLM source commit 635dd6aae656. NVFP4 requires
--quantization modelopt_fp4 --linear-backend cutlass. BF16 KV, 32K maximum context, max 4 sequences, 1.5 GiB shared KV.
These archives contain third-party NVIDIA, CUDA, Python and inference framework components. Each component retains its own license. See LICENSES.md and the license notices inside the archive. They are not uniformly licensed as the ReaderLM model. NVIDIA components require compatible NVIDIA Jetson hardware and applicable NVIDIA agreements. The ReaderLM-v2 weights are separately CC-BY-NC-4.0.
Thor release metadata removes inherited build-publishing environment variables, command history and build attestations. Inference layer digests are unchanged from the measured image. The small compressed_tensors hadamards.safetensors file is a public framework transform constant, not ReaderLM weights.
SHA256SUMS and runtimes.json fix archive identities, sizes and SHA-256. The application locks immutable revisions of both ModelScope and Hugging Face and verifies every download. Large archive imports need extra temporary storage; cached images only skip transfer after exact identity verification.