[Bugfix][RAINFUSION_ATTN] Fix padding-mask rejection on packed [real, pad] layouts and reuse query tail as output padding (#7235)

Signed-off-by: brandneway <gyuan4892@gmail.com>
This commit is contained in:
ooooooye authored and GitHub committed 2026-09-20 10:56:51 +08:00
1 parent fa422a3f44
commit a9abd563f9
4 files changed
+84 -5

No files matched your search

@@ -135,6 +135,11 @@ class RainFusionPlan:
class RainFusionAttentionBackend(AttentionBackend):
accept_output_buffer: bool = True
supported_platforms: tuple[str, ...] = ("npu",)
# The impl trims [real, pad] packed tensors to the valid prefix with
# [:, :used] itself, so models may run the padded layout without
# materializing a padding mask (a forward that carries one anyway falls
# back dense in _resolve_plan).
supports_prefix_kv_slicing: bool = True
@classmethod
def validate_available(cls) -> None:
@@ -509,7 +514,12 @@ class RainFusionAttentionImpl(AttentionImpl):
logger.info_once("RainFusion uses MindIE-SD sparse attention, precision=%s.", precision)
used = plan.used_len
q, k, v = (tensor[:, :used] for tensor in (query, key, value))
# Split query once: the valid prefix drives the kernel, the padding
# tail is reused verbatim as the output padding below. All views,
# zero-copy. split_with_sizes always yields both parts, including an
# empty tail when the sequence needs no padding.
q, q_pad = torch.split(query, [used, query.shape[1] - used], dim=1)
k, v = key[:, :used], value[:, :used]
# Ulysses has already gathered the full sequence onto this rank and split
# the heads, so read the head count off the tensor rather than num_heads.
common_kwargs: dict[str, object] = {
@@ -543,6 +553,9 @@ class RainFusionAttentionImpl(AttentionImpl):
out = sparse_attention(q, k, v, **common_kwargs)
if used == query.shape[1]:
return out
padded = torch.zeros_like(query)
padded[:, :used] = out
return padded
# Only the first ``used`` rows are read back downstream, so the padding
# payload is free. Reuse the query tail split off above instead of
# zeroing a fresh buffer: a single cat write, no extra allocation, and
# always finite values (uninitialized memory could hold NaN bit
# patterns).
return torch.cat([out, q_pad], dim=1)