Skip to content

Support nested outputs correctness check and use Xprof time instead of wall time by default#62

Merged
shangkunwang01 merged 5 commits into
mainfrom
shangkun-fix-evaluation
Jul 7, 2026
Merged

Support nested outputs correctness check and use Xprof time instead of wall time by default#62
shangkunwang01 merged 5 commits into
mainfrom
shangkun-fix-evaluation

Conversation

@shangkunwang01

Copy link
Copy Markdown
Collaborator
  1. The evaluation cannot work if the there are multiple outputs from the test function. This pr fix this by comparing all the outputs
  2. Use Xprof XLA Module time as the default time for calculating speedups
  3. Poison HBM memory to prevent kernels from inheriting residuals

@shangkunwang01 shangkunwang01 requested a review from NinaCai July 6, 2026 21:43
Comment thread MaxKernel/evaluation/harness_code.py Outdated
dummy.block_until_ready()
del dummy
except Exception:
pass

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a silent failure, could you explicitly show a log that it fails to dirty the HBM memory?

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added logging to the harness and pass it to result.

Comment thread MaxKernel/evaluation/harness_code.py Outdated
else:
val = 123 # Fits within int8/uint8 and all larger integer dtypes

dummy = jax.device_put(jnp.full(leaf.shape, val, dtype=leaf.dtype))

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

device_put() is redundant, it should be removed.
Should be

dummy = jnp.full(leaf.shape, val, dtype=leaf.dtype)

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks, removed.

@shangkunwang01 shangkunwang01 merged commit 61b420a into main Jul 7, 2026
7 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants