Enable causal masking for SWA in vllm-project/speculators models (#47745)
Signed-off-by: Eldar Kurtic <8884008+eldarkurtic@users.noreply.github.com>
Eldar Kurtić committed
Jul 7, 2026 at 16:24 UTC
beb4327c46ef34dfdb8e2c44f7f68b547e028ec8
1 file changed
+4
vllm/transformers_utils/configs/speculators/algos.py
+4
@@ -120,6 +120,10 @@ def update_dflash(config_dict: dict, pre_trained_config: dict) -> None:
120
"mask_token_id": config_dict["mask_token_id"],
121
"target_layer_ids": [i - 1 for i in aux_layer_ids],
122
}
123
+ # Enable causal masking in SWA for vllm-project/speculators models
124
+ pre_trained_config["dflash_config"]["causal"] = not config_dict.get(
125
+ "sliding_window_non_causal", True
126
+ )
127
128
129
@register_speculator("dspark")