[v1] Add PrefixLM support to FlexAttention backend (#27938)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-12-07 15:51:36 +00:00
committed by GitHub
parent 541a2ef892
commit b952f4d3c3
16 changed files with 173 additions and 25 deletions
+25
View File
@@ -175,6 +175,31 @@ class PlaceholderRange:
return int(self.is_embed.sum().item())
def extract_embeds_range(self) -> list[tuple[int, int]]:
"""Extract the start and end indices of the embedded region in prompt.
For example, given `PlaceholderRange(offset=2, length=5)` and
`is_embed = [False, True, False, True, True]`, the output is
`[(1 + offset, 1 + offset), (3 + offset, 4 + offset)]`.
Returns:
A tuple `(start, end)` representing the start and end
indices (inclusive) of the embedded region.
Returns full placeholder range if `is_embed` is `None`.
"""
if self.is_embed is None:
return [(self.offset, self.offset + self.length)]
mask_i = self.is_embed.int()
starts = torch.nonzero(
torch.diff(mask_i, prepend=mask_i.new_zeros(1)) == 1
).flatten()
ends = torch.nonzero(
torch.diff(mask_i, append=mask_i.new_zeros(1)) == -1
).flatten()
ranges = torch.stack((starts, ends), dim=1) + self.offset
return [tuple(x) for x in ranges.tolist()]
def __eq__(self, other: object) -> bool:
if not isinstance(other, self.__class__):
return False