[Core] Reduce mm scheduler, get_num_embed overhead (#40143)

Signed-off-by: milesial <milesial@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
milesial
2026-04-18 11:25:49 +08:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent 993859ceb0
commit b0755523dc
2 changed files with 9 additions and 13 deletions
+3 -6
View File
@@ -26,11 +26,8 @@ def test_placeholder_range_get_num_embeds(is_embed, expected):
"is_embed,expected",
[
(None, None),
(
torch.tensor([False, True, False, True, True]),
torch.tensor([0, 1, 1, 2, 3]),
),
(torch.tensor([True, True, True]), torch.tensor([1, 2, 3])),
(torch.tensor([False, True, False, True, True]), [0, 1, 1, 2, 3]),
(torch.tensor([True, True, True]), [1, 2, 3]),
],
)
def test_placeholder_range_embeds_cumsum(is_embed, expected):
@@ -41,6 +38,6 @@ def test_placeholder_range_embeds_cumsum(is_embed, expected):
assert pr.embeds_cumsum is None
return
assert torch.equal(pr.embeds_cumsum, expected)
assert pr.embeds_cumsum == expected
# cached_property should return the same object on repeated access
assert pr.embeds_cumsum is pr.embeds_cumsum