diff --git a/chandra/model/hf.py b/chandra/model/hf.py
index 6bedf49..0803f7d 100644
--- a/chandra/model/hf.py
+++ b/chandra/model/hf.py
@@ -73,7 +73,7 @@ def load_model():
         "dtype": settings.TORCH_DTYPE,
         "device_map": device_map,
     }
-    if settings.TORCH_ATTN:
+    if settings.TORCH_ATTN_IMPLEMENTATION:
         kwargs["attn_implementation"] = settings.TORCH_ATTN_IMPLEMENTATION
 
     model = Qwen3VLForConditionalGeneration.from_pretrained(
diff --git a/chandra/settings.py b/chandra/settings.py
index 2c59ec3..67bb55d 100644
--- a/chandra/settings.py
+++ b/chandra/settings.py
@@ -13,7 +13,7 @@ class Settings(BaseSettings):
     MODEL_CHECKPOINT: str = "datalab-to/chandra"
     TORCH_DEVICE: str | None = None
     MAX_OUTPUT_TOKENS: int = 8192
-    TORCH_ATTN: str | None = None
+    TORCH_ATTN_IMPLEMENTATION: str | None = None
 
     # vLLM server settings
     VLLM_API_KEY: str = "EMPTY"