logits_to_keep & use_query_position

Author: rebel-thkim

src/optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py

@@ -145,13 +145,10 @@ def prepare_forward_args(self, *args):
         local_block_tables = args.pop(0) if self.rbln_config.use_local_attention else None
         query_position = (
             args.pop(0)
-            # query_position usage: 1. causal_lm prefill or 2. sliding_window cache_position
+            # query_position usage: 1. prefill & logits_to_keep == 1 or 2. sliding_window cache_position
             if (
-                "prefill" in self.phase
-                and (
-                    (self.is_causal_lm and self.rbln_config.logits_to_keep == 1)
-                    or self.rbln_config.use_local_attention
-                )
+                ("prefill" in self.phase and self.rbln_config.logits_to_keep == 1)
+                or self.rbln_config.use_local_attention
             )
             else None
         )

src/optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -33,7 +33,7 @@
     Qwen2_5_VLRotaryEmbedding,
 )
 
-from ....configuration_utils import RBLNCompileConfig, RBLNModelConfig
+from ....configuration_utils import RBLNCompileConfig
 from ....modeling import RBLNModel
 from ....utils.logging import get_logger
 from ...modeling_outputs import RBLNDecoderOnlyOutput
@@ -387,7 +387,6 @@ def prefill_output_size(self):
             hidden_size,
         )
 
-
     @classmethod
     def get_input_info(
         cls,