From d12c2768530a4da8d7c7d4a4f5d6adfa351675cd Mon Sep 17 00:00:00 2001 From: Mustafa YILDIRIM <2341474+mustafayildirim@users.noreply.github.com> Date: Thu, 17 Sep 2026 20:36:55 -0700 Subject: [PATCH] [Bugfix][ROCm] Alias SparseAttnIndexerKpool.forward_cuda to forward_native (GLM-5.3-Flash boot crash) (#57425) Signed-off-by: Mustafa YILDIRIM Co-authored-by: Hermes Agent Co-authored-by: Tyler Michael Smith --- vllm/models/glm5next/amd/sparse_indexer.py | 25 +++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/vllm/models/glm5next/amd/sparse_indexer.py b/vllm/models/glm5next/amd/sparse_indexer.py index f47c9164bc..c9b68e973f 100644 --- a/vllm/models/glm5next/amd/sparse_indexer.py +++ b/vllm/models/glm5next/amd/sparse_indexer.py @@ -665,7 +665,7 @@ class SparseAttnIndexerKpool(CustomOp): self.skip_k_cache_insert = skip_k_cache_insert self.use_fp4_cache = use_fp4_cache - def forward_native( + def forward_hip( self, hidden_states: torch.Tensor, q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor], @@ -727,3 +727,26 @@ class SparseAttnIndexerKpool(CustomOp): self.tail_cache.kv_cache if self.tail_cache is not None else None, self.tail_cache.prefix if self.tail_cache is not None else None, ) + + def forward_native( + self, + hidden_states: torch.Tensor, + q_quant: torch.Tensor | tuple[torch.Tensor, torch.Tensor], + k: torch.Tensor, + weights: torch.Tensor, + *, + gate_score: torch.Tensor | None = None, + compress_ape: torch.Tensor | None = None, + index_kpool: int = 1, + positions: torch.Tensor | None = None, + ) -> torch.Tensor: + return self.forward_hip( + hidden_states, + q_quant, + k, + weights, + gate_score=gate_score, + compress_ape=compress_ape, + index_kpool=index_kpool, + positions=positions, + )