6 Commits
Author SHA1 Message Date
Lucas ManningandgVisor bot c38ce1feb1 Create vllm TPU serving image.
This image is useful for benchmarking. It will be used for more automated
testing and benchmarking in the future.

PiperOrigin-RevId: 699286661
2024-11-22 14:25:40 -08:00
Lucas ManningandgVisor bot e22dc29727 Update pytorch image python module versions.
The switch to a newer numpy was breaking guitar tests.

PiperOrigin-RevId: 695514200
2024-11-11 16:10:37 -08:00
Lucas ManningandgVisor bot 8b5a0b0614 Internal change.
PiperOrigin-RevId: 693458549
2024-11-05 13:28:01 -08:00
Lucas ManningandgVisor bot bbbecc35cc Add support for v5pod and fix TPU v5 bugs.
V5 support was broken in a few different ways:
- tpu device files (/dev/vfio/X) were created with incorrect minor device nums.
- PCI paths on bus' other than 0000:00 were not supported.
- VFIO unmap was broken and not properly added to the seccomp allowlist.
- The VFIO main device file (/dev/vfio/vfio) did not account for overlapping
  device address ranges that correspond to different VFIO container groups.

Previously TPU support was tested on machines with single TPUs, which masked
most of these issues.

All these issues should be fixed by this change. Tested manually on GKE.

PiperOrigin-RevId: 656037853
2024-07-25 12:12:30 -07:00
Jing ChenandgVisor bot 7e9c01f12a Add Dockerfile and K8S deployments for a basic TPU pytorch workflow.
PiperOrigin-RevId: 620968670
2024-04-01 15:25:12 -07:00
Lucas ManningandgVisor bot fb487fbb2f Create a TPU hello world image for use in gVisor tests.
PiperOrigin-RevId: 574506997
2023-10-18 09:57:07 -07:00