BLIP AI Pod runtimes

Verified ARM64 native vLLM Docker-load archives for AGX Orin and AGX Thor (runtime 0.2.0), published for BLIP LPK 0.2.0. Salesforce/blip-image-captioning-large model weights are downloaded separately and mounted read-only. Verify size and SHA256 in checksums.json before docker load, then verify the image identity. These device-specific archives are not interchangeable. The archives contain Python, PyTorch, CUDA libraries, Transformers and other inference dependencies. The native adapter and framework patch sources are provided in blip-native-source-0.2.0.tar.gz. Each component retains its own license and license notices, including NVIDIA runtime notices, inside the archive.

Orin JetPack 7.2.1 / BLIP 0.2.2

The JP7 release uses native vLLM 0.20.0 on the NVIDIA Orin CUDA 13 runtime. The model weights remain separate and are mounted read-only. The following files are available from this public repository:

  • runtime-orin-jp7-0.2.2.tar.gz: complete ARM64 Docker-load image for registry.lazycat.cloud/catdogai/blip-orin-jp7:0.2.2 (9,176,903,837 bytes, SHA-256 54241509d03bfed4c98236a916576f3511419c3a373e2a486d7e8a570f597f99).
  • blip-native-source-orin-jp7-0.2.2.tar.gz: matching Dockerfile, BLIP adapter sources, device overrides, deployment profile and pinned manifest (26,955 bytes, SHA-256 a2e0e85dbfe52e30bc0737a0321f75fd80719573b4ce5599e24f6b723912023d).

Verify the image checksum, then load it with docker load -i runtime-orin-jp7-0.2.2.tar.gz. The source bundle is generated from the BLIP repository commit 7118c69. The JP6 and Thor 0.2.0 archives above remain separate and should not be substituted for JP7.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support