@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 14e2f1231e
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 14, 05:30 UTC

[Bugfix] Make `get_mrope_input_positions` instance methods (#27342)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Cyrus Leung committed Oct 22, 2025 at 23:38 UTC 14e2f1231e93e956348a1647d8cb858067ce9dcc
7 files changed +6 -13
vllm/model_executor/models/ernie45_vl.py
+1 -2
@@ -1404,9 +1404,8 @@ class Ernie4_5_VLMoeForConditionalGeneration(
1404 else:
1405 self.visual_token_mask = None
1406
1407 - @classmethod
1407 def get_mrope_input_positions(
1409 - cls,
1408 + self,
1409 input_tokens: list[int],
1410 hf_config: PretrainedConfig,
1411 image_grid_thw: list[list[int]] | torch.Tensor,
vllm/model_executor/models/glm4v.py
+1 -2
@@ -619,9 +619,8 @@ class GLM4VForCausalLM(
619
620 return self.transformer.vision(pixel_values)
621
622 - @classmethod
622 def get_mrope_input_positions(
624 - cls,
623 + self,
624 input_tokens: list[int],
625 hf_config: PretrainedConfig,
626 image_grid_thw: list[list[int]] | torch.Tensor,
vllm/model_executor/models/keye_vl1_5.py
+1 -2
@@ -594,9 +594,8 @@ class KeyeVL1_5ForConditionalGeneration(
594 new_video_embeds.append(video_embeds[start:end])
595 return tuple(new_video_embeds)
596
597 - @classmethod
597 def get_mrope_input_positions(
599 - cls,
598 + self,
599 input_tokens: list[int],
600 hf_config: PretrainedConfig,
601 image_grid_thw: list[list[int]] | torch.Tensor,
vllm/model_executor/models/qwen2_5_omni_thinker.py
+1 -2
@@ -986,9 +986,8 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
986 def get_language_model(self) -> torch.nn.Module:
987 return self.language_model
988
989 - @classmethod
989 def get_mrope_input_positions(
991 - cls,
990 + self,
991 input_tokens: list[int],
992 hf_config: PretrainedConfig,
993 image_grid_thw: list[list[int]] | torch.Tensor,
vllm/model_executor/models/qwen2_5_vl.py
+1 -2
@@ -1078,9 +1078,8 @@ class Qwen2_5_VLForConditionalGeneration(
1078
1079 supports_encoder_tp_data = True
1080
1081 - @classmethod
1081 def get_mrope_input_positions(
1083 - cls,
1082 + self,
1083 input_tokens: list[int],
1084 hf_config: PretrainedConfig,
1085 image_grid_thw: list[list[int]] | torch.Tensor,
vllm/model_executor/models/qwen3_omni_moe_thinker.py
-1
@@ -1421,7 +1421,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
1421
1422 return loaded_weights
1423
1424 - @classmethod
1424 def get_mrope_input_positions(
1425 self,
1426 input_tokens: list[int],
vllm/model_executor/models/qwen3_vl.py
+1 -2
@@ -1480,9 +1480,8 @@ class Qwen3VLForConditionalGeneration(
1480 )
1481 return mm_input_by_modality
1482
1483 - @classmethod
1483 def get_mrope_input_positions(
1485 - cls,
1484 + self,
1485 input_tokens: list[int],
1486 hf_config: PretrainedConfig,
1487 image_grid_thw: list[list[int]] | torch.Tensor,