[Bugfix][ROCm] Alias SparseAttnIndexerKpool.forward_cuda to forward_native (GLM-5.3-Flash boot crash) (#57425)
Signed-off-by: Mustafa YILDIRIM <mustafa@character.ai> Co-authored-by: Hermes Agent <hermes@nousresearch.com> Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Mustafa YILDIRIM committed
Sep 17, 2026 at 20:36 UTC
d12c2768530a4da8d7c7d4a4f5d6adfa351675cd
1 file changed
+24
-1
vllm/models/glm5next/amd/sparse_indexer.py
+24
-1
@@ -665,7 +665,7 @@ class SparseAttnIndexerKpool(CustomOp):
665
self.skip_k_cache_insert = skip_k_cache_insert
666
self.use_fp4_cache = use_fp4_cache
667
668
- def forward_native(
668
+ def forward_hip(
669
self,
670
hidden_states: torch.Tensor,
671
q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
@@ -727,3 +727,26 @@ class SparseAttnIndexerKpool(CustomOp):
727
self.tail_cache.kv_cache if self.tail_cache is not None else None,
728
self.tail_cache.prefix if self.tail_cache is not None else None,
729
)
730
+
731
+ def forward_native(
732
+ self,
733
+ hidden_states: torch.Tensor,
734
+ q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
735
+ k: torch.Tensor,
736
+ weights: torch.Tensor,
737
+ *,
738
+ gate_score: torch.Tensor | None = None,
739
+ compress_ape: torch.Tensor | None = None,
740
+ index_kpool: int = 1,
741
+ positions: torch.Tensor | None = None,
742
+ ) -> torch.Tensor:
743
+ return self.forward_hip(
744
+ hidden_states,
745
+ q_quant,
746
+ k,
747
+ weights,
748
+ gate_score=gate_score,
749
+ compress_ape=compress_ape,
750
+ index_kpool=index_kpool,
751
+ positions=positions,
752
+ )