JAX cannot find TPU metadata inside a container

[pawalt]

Description

When a TPU container is initialized, it's missing some environment variables that JAX needs in order to initialize. In the absence of these variables, JAX attempts to look up their values over the network. This fails as the container may not have direct access to the network.

I have also tried this with network=host to no avail.

Steps to reproduce

Run a jax image with --tpuproxy:

Runsc command:

sudo runsc --debug \
    --debug-log=/home/peyton/tputesting/logs/ \
    --strace \
    --root=/home/peyton/tputesting/runroot \
    --tpuproxy \
    run \
    --bundle=/home/peyton/tputesting \
    my-jax-container

Start the container:

peyton@t1v-n-1f714773-w-0:~/tputesting$ ./start.sh 
# python
Python 3.11.9 (main, Aug 13 2024, 02:18:20) [GCC 12.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import jax
>>> jax.device_count()                                                                                                                      
Failed to get TPU metadata (tpu-env) from instance metadata for variable CHIPS_PER_HOST_BOUNDS: INTERNAL: Couldn't connect to server
=== Source Location Trace: ===
learning/45eac/tfrc/runtime/gcp_metadata_utils.cc:99
learning/45eac/tfrc/runtime/env_var_utils.cc:50

I've built this image by exporting the following dockerfile:

FROM python:3.11

RUN pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html

runsc version

runsc -version
runsc version release-20240826.0-81-g4bcbb55fcba5
spec: 1.1.0-rc.1

docker version (if using docker)

I'm not using docker

uname

uname -a Linux t1v-n-1f714773-w-0 5.19.0-1022-gcp #24~22.04.1-Ubuntu SMP Sun Apr 23 09:51:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux

kubectl (if using Kubernetes)

No response

repo state (if built from source)

git describe release-20240826.0-81-g4bcbb55fc

runsc debug logs (if available)

No response

[pawalt]

None of these env variables are present on the host - it populates them dynamically

[manninglucas]

Seems like the sandbox cannot reach the GCE metadata server for some reason. I'm surprised that --network=host does not fix this. I will investigate and try to repro myself. If you can share your strace logs that will help me debug the issue as well.

[pawalt]

runsc.log.20240918-170153.986977.boot.txt Here are the strace logs - had to upload a file as it's quite a lot of logs.

[manninglucas]

It looks like --network=host is not true in the logs you sent.

From runsc.log.20240918-170153.986977.boot.txt:

D0918 17:01:54.094908       1 config.go:439] Config.Network (--network): sandbox

You'll also want to run runsc in the host's network namespace to allow proper host network access. Right now it looks like you're running in a separate network namespace.

From runsc.log.20240918-170153.986977.boot.txt:

  "linux": {
    "namespaces": [
      {
        "type": "pid"
      },
      {
        "type": "network"
      },
      {
        "type": "ipc"
      },
      {
        "type": "uts"
      },
      {
        "type": "mount"
      }
    ]
  }

FWIW I tested this on my own V5 GCE VM and it worked.

Here was my /dev/vfio directory:

$ stat /dev/vfio/*
  File: /dev/vfio/0
  Size: 0               Blocks: 0          IO Block: 4096   character special file
Device: 5h/5d   Inode: 334         Links: 1     Device type: ef,0
Access: (0666/crw-rw-rw-)  Uid: (    0/    root)   Gid: (    0/    root)
Access: 2024-09-18 21:11:48.884000071 +0000
Modify: 2024-09-18 21:11:48.884000071 +0000
Change: 2024-09-18 21:11:48.888000071 +0000
 Birth: 2024-09-18 21:11:48.884000071 +0000
  File: /dev/vfio/vfio
  Size: 0               Blocks: 0          IO Block: 4096   character special file
Device: 5h/5d   Inode: 141         Links: 1     Device type: a,c4
Access: (0666/crw-rw-rw-)  Uid: (    0/    root)   Gid: (    0/    root)
Access: 2024-09-18 21:11:48.832000070 +0000
Modify: 2024-09-18 21:11:48.832000070 +0000
Change: 2024-09-18 21:11:48.832000070 +0000
 Birth: 2024-09-18 21:11:45.235863372 +0000

Here was my config.json: https://gist.github.com/manninglucas/14de68aab7abaab02cf41553f900782e My command was: sudo ./runsc --debug --debug-log=debug.txt --network=host --tpuproxy run bash In the container, I ran:

$ pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
$ python
>>> import jax
>>> jax.device_count()

[manninglucas]

I'll just add as a note that these environment variables are piped into the configs automatically in GKE. In GCE you'll either have to add the environment variables to your spec yourself (maybe by fetching them from the metadata server before starting a sandbox) or give the sandbox at least some host network access.

[manninglucas]

@pawalt were you able to get something working for your use case? It would be great to be able to close this bug if so.

[pawalt]

@manninglucas I haven't had time to get this working, but that's not a GVisor issue - we've just been swamped with other work. You're welcome to close out this ticket since it'll probably go stale. Your env var fixes look reasonable anyway. I'm happy to reopen this or open another ticket if I get a chance to try this out again and run into the same issues.