From beb4327c46ef34dfdb8e2c44f7f68b547e028ec8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Eldar=20Kurti=C4=87?= <8884008+eldarkurtic@users.noreply.github.com> Date: Tue, 7 Jul 2026 16:24:14 +0200 Subject: [PATCH] Enable causal masking for SWA in vllm-project/speculators models (#47745) Signed-off-by: Eldar Kurtic <8884008+eldarkurtic@users.noreply.github.com> --- vllm/transformers_utils/configs/speculators/algos.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/transformers_utils/configs/speculators/algos.py b/vllm/transformers_utils/configs/speculators/algos.py index f4b0f73f62..e034cec974 100644 --- a/vllm/transformers_utils/configs/speculators/algos.py +++ b/vllm/transformers_utils/configs/speculators/algos.py @@ -120,6 +120,10 @@ def update_dflash(config_dict: dict, pre_trained_config: dict) -> None: "mask_token_id": config_dict["mask_token_id"], "target_layer_ids": [i - 1 for i in aux_layer_ids], } + # Enable causal masking in SWA for vllm-project/speculators models + pre_trained_config["dflash_config"]["causal"] = not config_dict.get( + "sliding_window_non_causal", True + ) @register_speculator("dspark")