From e4a0bf3b718f81c18c91587181d7c09ad17d87eb Mon Sep 17 00:00:00 2001 From: drbh Date: Wed, 31 Jul 2024 22:15:52 +0000 Subject: [PATCH] fix: allocate tmp based on sgmv kernel if available --- server/text_generation_server/utils/sgmv.py | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/server/text_generation_server/utils/sgmv.py b/server/text_generation_server/utils/sgmv.py index e0aec25f..2d0a73a5 100644 --- a/server/text_generation_server/utils/sgmv.py +++ b/server/text_generation_server/utils/sgmv.py @@ -151,13 +151,17 @@ def get_tmp_expand_size(size: int) -> int: def get_tmp_tensors( nsegments: int, lora_rank: int, device: torch.device ) -> Tuple[torch.Tensor, torch.Tensor]: - if use_cutlass_shrink(lora_rank) and has_sgmv(): + use_cutlass = use_cutlass_shrink(lora_rank) and has_sgmv() + has_sgmv_available = has_sgmv() + + if use_cutlass: tmp = get_tmp_tensor_for_size(nsegments, device) return tmp, tmp + elif has_sgmv_available: + return get_tmp_tensor(device), get_tmp_tensor_for_size(nsegments, device) else: - tmp_shrink = get_tmp_tensor(device) - tmp_expand = get_tmp_tensor_for_size_no_kernels(nsegments, device) - return tmp_shrink, tmp_expand + tmp = get_tmp_tensor_for_size(nsegments, device) + return tmp, tmp def lora_a_sgmv_cutlass(