[Bugfix] Preserve tensor causal metadata for grouped attention (#48135)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Lucas Wilkinson
2026-07-09 17:57:53 +00:00
committed by GitHub
co-authored by Codex
parent 67e7ea8977
commit e08a915146
+4 -2
View File
@@ -573,7 +573,7 @@ def build_attn_metadata(
mm_req_doc_ranges: dict[int, list[tuple[int, int]]] | None = None,
model_specific_attn_metadata: ModelSpecificAttnMetadata | None = None,
for_cudagraph_capture: bool = False,
causal: bool | Mapping[int, bool] = True,
causal: bool | torch.Tensor | Mapping[int, bool] = True,
rswa_prefix_lens: torch.Tensor | None = None,
) -> dict[str, Any]:
seq_lens = seq_lens[:num_reqs]
@@ -588,7 +588,9 @@ def build_attn_metadata(
block_table = block_tables[i]
slot_mapping = slot_mappings[i]
# Per-group causal for hybrid drafters (mixed SWA/full attention).
group_causal = causal if isinstance(causal, bool) else causal.get(i, True)
group_causal = (
causal if isinstance(causal, (bool, torch.Tensor)) else causal.get(i, True)
)
common_attn_metadata_extra_kwargs = (
model_specific_attn_metadata.get_extra_common_attn_kwargs(i, num_reqs)