Bug: Broken FlashInfer on Bristen

[xzyaoi]

RuntimeError: CUDA error: named symbol not found
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

[xzyaoi]

This seems to be likely an issue with PyTorch in the base image (weird it's working on todi and all other environments), since the triton backend fails too.

[xzyaoi]

Resolved with torch-24.05 image from NGC.