[SYCL] pass SYCL CI

[airMeng]

fix some issues in norm and permute GEMM

There are instability within the cloud instance, so only keep the basic functionality in SYCL CI to save the machine time

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[airMeng]

@ggerganov how to trigger ggml-7-sycl?

[ggerganov]

Add the string ggml-ci somewhere in the commit message. You can git commit --amend, add ggml-ci and then force push to trigger the CI.

[airMeng]

@arthw it is wield that test-quantize-fns is a pure host side executable and it can pass under debug build but failed under debug model. Is it possible that it is an ICX/ICPX issue?

[airMeng]

@ggerganov only ggml-7-sycl not triggered, is it offline?

[ggerganov]

Looks like the same connectivity issues as before:

# when trying to SSH to the instance
channel 0: open failed: connect failed: No route to host
stdio forwarding failed
Connection closed by UNKNOWN port 65535

The instance is shown as "Ready" in the Intel Cloud.

[airMeng]

Looks like the same connectivity issues as before:

# when trying to SSH to the instance
channel 0: open failed: connect failed: No route to host
stdio forwarding failed
Connection closed by UNKNOWN port 65535

The instance is shown as "Ready" in the Intel Cloud.

Please let us know if this happens from time to time? Sorry for inconvenience, ITDC is still in developments.

[ggerganov]

It seemed to happen quite often - usually about 10 minutes after I manage to SSH. I've reported at least 3 occurrences to Intel support on 2 different machines, but no root cause was found. Yesterday it looked like the issue was gone because I was able to SSH and setup the CI successfully without disconnects. But now it seems the issue is back. Note that I wasn't connected or doing anything on the machine. The last CI run was ~4 hours ago from the commits in this PR: https://github.com/ggml-org/ci/commits/results/. Will let you know if the connection comes back again.

[airMeng]

seems the machine hang there and now it is back.

@ggerganov could you share your command, for example, a script ro run llama.cpp CI with one click that the cloud engineers could debug it, either here or through the email? we think this should be a simple issue but one we hadn't considered before

[ggerganov]

Yes, will send you an email in about half an hour

[airMeng]

@ggerganov solved the connection issue, you might need to re-request another instance. Sorry for the inconvenience, you can use this PR to test the CI again.

[ggerganov]

Thank you. Will try to set it up today. Will let you know.

[ggerganov]

@airMeng I just configured the new instance today and as I was just finishing the deployment of ggml-ci I again got disconnected and can no longer use the instance:

channel 0: open failed: connect failed: No route to host
stdio forwarding failed
Connection closed by UNKNOWN port 65535