Blackwell support was originally slated for enablement for
1.5, and support was introduced while enabling vllm-0.7.2. At the same
time we attempted to upgrade CUDA 12.8, where Blackwell support was added
[1], but found that CUDA 12.8 was not supported by several other packages.
As a result, we downgraded CUDA 12.8 to CUDA 12.6, and eventually CUDA
12.4 in 1.5.

We need to support the NVIDIA Blackwell architecture in an AIPCC release.
In order to do this we must update torch to a newer version that enables
the support of Blackwell. The latest version of torch, torch-2.7,
supports CUDA 12.8 [1] which enables blackwell [2].

[1] https://pytorch.org/get-started/locally/
[2] https://en.wikipedia.org/wiki/CUDA#GPUs_supported