[pallas/pallas_mgpu] Discharging run_scoped should not be discharging the intermediates #25639
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Outdated suggestions cannot be applied.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.
[pallas/pallas_mgpu] Discharging run_scoped should not be discharging the intermediates
When we do run_scoped[jaxpr, R1,R2], it can't be assumed that references
corresponding to R1 and R2 can be safely discharged. Sometimes they can (eg
Accumulator) but sometimes they can't (eg SMEM scratch). It should be up to the
lowering rule to do such discharging.
This further means that during lowering there is no guarantee that the
references will not be used/returned by nested scoped blocks so we also remove
that check.