From fdf282b9e08560e230565b47de68513a89261050 Mon Sep 17 00:00:00 2001 From: Konrad Zawora Date: Wed, 15 May 2024 16:40:02 +0300 Subject: [PATCH] WA: Disable cumsum in HPU _prepare_prompt --- vllm/worker/habana_model_runner.py | 9 --------- 1 file changed, 9 deletions(-) diff --git a/vllm/worker/habana_model_runner.py b/vllm/worker/habana_model_runner.py index a8f801d62cc3d..e306ef0ae12cb 100644 --- a/vllm/worker/habana_model_runner.py +++ b/vllm/worker/habana_model_runner.py @@ -402,15 +402,6 @@ def _prepare_prompt( dtype=torch.int32, device=self.device) - torch.cumsum(query_lens_tensor, - dim=0, - dtype=subquery_start_loc.dtype, - out=subquery_start_loc[1:]) - - torch.cumsum(seq_lens_tensor, - dim=0, - dtype=seq_start_loc.dtype, - out=seq_start_loc[1:]) attn_metadata = self.attn_backend.make_metadata( is_prompt=True, seq_lens=seq_lens,