From 16e91176cf77bf0f40ae48da22365a5e21b517af Mon Sep 17 00:00:00 2001 From: Itay Alroy <75032521+itayalroy@users.noreply.github.com> Date: Tue, 16 Jun 2026 01:50:18 +0300 Subject: [PATCH] [EP] Query NIXL EP top-k index dtype (#45298) Signed-off-by: Itay Alroy --- .../layers/fused_moe/prepare_finalize/nixl_ep.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/nixl_ep.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/nixl_ep.py index ce44cd6a3f..89571278c6 100644 --- a/vllm/model_executor/layers/fused_moe/prepare_finalize/nixl_ep.py +++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/nixl_ep.py @@ -28,6 +28,8 @@ logger = init_logger(__name__) # NIXL EP kernels quantize dispatch inputs in 128 element chunks. NIXL_EP_QUANT_BLOCK_SIZE = 128 NIXL_EP_QUANT_BLOCK_SHAPE = [NIXL_EP_QUANT_BLOCK_SIZE, NIXL_EP_QUANT_BLOCK_SIZE] +NIXL_EP_TOPK_INDICES_DTYPE = getattr(nixl_ep, "topk_idx_t", torch.int64) +assert isinstance(NIXL_EP_TOPK_INDICES_DTYPE, torch.dtype) def dequant_fp8( @@ -151,7 +153,7 @@ class NixlEPPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular): all2all_manager.commit_staged_state() def topk_indices_dtype(self) -> torch.dtype | None: - return torch.int64 + return NIXL_EP_TOPK_INDICES_DTYPE def _map_global_to_physical_ids(self, topk_ids: torch.Tensor) -> torch.Tensor: if self.global_to_physical is None: