Skip test_mxfp on A770 #3811
Draft
Skip test_mxfp on A770 #3811
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
It looks like the problem might be elsewhere: |
etiotto
approved these changes
Apr 2, 2025
|
trying again with new approach (not using skiplist): https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14228029469/job/39872117025 |
pbchekin
reviewed
Apr 2, 2025
| @@ -316,6 +316,8 @@ def fp8e8m0_to_float32(scale): | |||
| @pytest.mark.parametrize("NUM_STAGES", [1, 3]) | |||
| @pytest.mark.parametrize("NUM_WARPS", [4, 8]) | |||
| @pytest.mark.parametrize("nonKDim", ([0, 16, 32] if is_hip_cdna() else [0])) | |||
| @pytest.mark.skipif(is_xpu() and not torch.xpu.get_device_capability()['has_subgroup_matrix_multiply_accumulate'], | |||
There was a problem hiding this comment.
Just for the case when has_subgroup_matrix_multiply_accumulate not in capabilities.
Suggested change
| @pytest.mark.skipif(is_xpu() and not torch.xpu.get_device_capability()['has_subgroup_matrix_multiply_accumulate'], | |
| @pytest.mark.skipif(is_xpu() and not torch.xpu.get_device_capability().get('has_subgroup_matrix_multiply_accumulate', False), |
In general, this decorator is executed in import-time, what is not convenient and Python best practice is minimize import-time logic. IMHO it is better to move this conditional skip to the test body.
Please note this run uses the default skip list: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14228029469/job/39872117025#step:11:38 We do not (yet) autodetect a skip list based on the selected runner, so i would recommend specifying the skip list explicitly. The workflow "Build and test GPU" requires setting "Runner label" and "Skip list" inputs, so it is better suited for such runs. |
|
ok - what is the syntax for the skiplist parameter? is this documented somewhere? |
It is not documented yet. For "Build and test" the default value is "default", for "Build and test GPU" it is empty by default, so we have to specify it. Basically it is the last directory name in the path to a skip list: https://github.com/intel/intel-xpu-backend-for-triton/tree/main/scripts/skiplist. For example: |
Also a specified skip list can be identified in a successful run, in the step "Print inputs". For example: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14167030129/job/39682308605#step:2:35 |
Specifying a skip list via workflow's input was considered as a temporary measure. The plan was to identify it automatically based on the selected runner. We still want this some day, just not enough capacity at the moment. |
|
resubmitted: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14231903364/job/39884095474 |
Yes, looks good: correct runner and skip list. The input "device" is also a temporary input (with max1100 as a default value), will be defined by a runner when we implement everything right. |
|
We could also use the same approach to skip |
|
Running tests in main on A770 on a runner with |
|
@alexbaden you can now just specify |
alexbaden
force-pushed
the
alex/update_client_skiplist
branch
from
843992c to
236116c
Compare
|
A770 run with blanket |
anmyachev
approved these changes
Apr 8, 2025
thanks! FYI: pre-commit checks failed |
|
yup - but the test is already running so let's just let it finish. |
|
With the skip list from 1cc2bd3 the workflow run took ~2.5h: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14286539636. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The A770 job recently increased to 5+ hours. See if skipping mxfp brings the job back to taking a reasonable amount of time to complete.