Enable causal masking for SWA in vllm-project/speculators models (#47745)

Signed-off-by: Eldar Kurtic <8884008+eldarkurtic@users.noreply.github.com>
2026-07-07 10:24:14 -04:00
parent c46ced1ee3
commit beb4327c46
1 changed files with 4 additions and 0 deletions
@@ -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")