Skip to content

Reduce peak memory during compiled KumoTabular fitting #1244

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
aw471 merged 4 commits into from
Oct 23, 2026
Merged

Reduce peak memory during compiled KumoTabular fitting #1244

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8512b81
Compile compact ICL cache copies as nested regions
aw471 Oct 22, 2026
1940095
Pass cache-copy tensors by keyword
aw471 Oct 22, 2026
3f307cd
Explain nested compilation compatibility check
aw471 Oct 23, 2026
d4bd43e
Clarify why nested compilation needs a compatibility guard
aw471 Oct 23, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 19 additions & 3 deletions sdm/models/kumo/tabular/icl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,9 +98,12 @@ def forward(

if cache is not None and cache.is_recording:
x, (key, value) = result
if self.kv_heads is not None:
key = key[..., : self.kv_heads, :].contiguous()
value = value[..., : self.kv_heads, :].contiguous()
kv_heads = self.kv_heads
if kv_heads is not None and kv_heads < key.size(-2):
key, value = _copy_key_value(
key=key[..., :kv_heads, :],
value=value[..., :kv_heads, :],
)
cache[cache_key] = KVCacheEntry(key, value)
del key, value
else:
Expand Down Expand Up @@ -134,3 +137,16 @@ def forward(
x = torch.cat([x_context, x_query], -2)

return self.head(self.norm(x))


def _copy_key_value(key: Tensor, value: Tensor) -> tuple[Tensor, Tensor]:
return (
key.clone(memory_format=torch.contiguous_format),
value.clone(memory_format=torch.contiguous_format),
)


# Keep the guard for PyTorch 2.7; 2.8+ can use the decorator directly.
if hasattr(torch.compiler, "nested_compile_region"):
# Compile cache copies separately to free each layer's full K/V tensors.
_copy_key_value = torch.compiler.nested_compile_region(_copy_key_value)
Loading