d0e186c16f · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/model_executor/models/ernie45_vl.py

-3

@@ -1435,8 +1435,6 @@ class Ernie4_5_VLMoeForConditionalGeneration(

1435 hf_config: PretrainedConfig,

1436 image_grid_thw: list[list[int]] | torch.Tensor,

1437 video_grid_thw: list[list[int]] | torch.Tensor,

1438 - context_len: int = 0,

1439 - seq_len: int | None = None,

1438 second_per_grid_ts: list[float] | None = None,

1439 audio_feature_lengths: torch.Tensor | None = None,

1440 use_audio_in_video: bool = False,

@@ -1569,7 +1567,6 @@ class Ernie4_5_VLMoeForConditionalGeneration(

1567 llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1))

1568

1569 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1572 - llm_positions = llm_positions[:, context_len:seq_len]

1570 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1571 return llm_positions, mrope_position_delta

1572

vllm/model_executor/models/glm4_1v.py

-3

@@ -1622,8 +1622,6 @@ class Glm4vForConditionalGeneration(

1622 image_grid_thw: list[list[int]] | torch.Tensor | None,

1623 video_grid_thw: list[list[int]] | torch.Tensor | None,

1624 second_per_grid_ts: list[float] | None = None,

1625 - context_len: int = 0,

1626 - seq_len: int | None = None,

1625 audio_feature_lengths: torch.Tensor | None = None,

1626 use_audio_in_video: bool = False,

1627 ) -> tuple[torch.Tensor, int]:

@@ -1754,7 +1752,6 @@ class Glm4vForConditionalGeneration(

1752 llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1))

1753

1754 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1757 - llm_positions = llm_positions[:, context_len:seq_len]

1755 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1756 return llm_positions, mrope_position_delta

1757

vllm/model_executor/models/glm4v.py

-3

@@ -625,8 +625,6 @@ class GLM4VForCausalLM(

625 hf_config: PretrainedConfig,

626 image_grid_thw: list[list[int]] | torch.Tensor,

627 video_grid_thw: list[list[int]] | torch.Tensor,

628 - context_len: int = 0,

629 - seq_len: int | None = None,

628 second_per_grid_ts: list[float] | None = None,

629 audio_feature_lengths: torch.Tensor | None = None,

630 use_audio_in_video: bool = False,

@@ -758,7 +756,6 @@ class GLM4VForCausalLM(

756 llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1))

757

758 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

761 - llm_positions = llm_positions[:, context_len:seq_len]

759 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

760 return llm_positions, mrope_position_delta

761

vllm/model_executor/models/interfaces.py

-4

@@ -995,8 +995,6 @@ class SupportsMRoPE(Protocol):

995 image_grid_thw: list[list[int]] | torch.Tensor | None,

996 video_grid_thw: list[list[int]] | torch.Tensor | None,

997 second_per_grid_ts: list[float] | None = None,

998 - context_len: int = 0,

999 - seq_len: int | None = None,

998 audio_feature_lengths: torch.Tensor | None = None,

999 use_audio_in_video: bool = False,

1000 ) -> tuple[torch.Tensor, int]:

@@ -1012,8 +1010,6 @@ class SupportsMRoPE(Protocol):

1010 image_grid_thw: Image grid dimensions (t, h, w)

1011 video_grid_thw: Video grid dimensions (t, h, w)

1012 second_per_grid_ts: Seconds per grid timestep for videos

1015 - context_len: Context length

1016 - seq_len: Sequence length

1013 audio_feature_lengths: Audio feature lengths for multimodal models

1014 use_audio_in_video: Whether to use audio in video for interleaving

1015

vllm/model_executor/models/keye.py

-3

@@ -1630,8 +1630,6 @@ class KeyeForConditionalGeneration(

1630 hf_config: PretrainedConfig,

1631 image_grid_thw: list[list[int]] | torch.Tensor,

1632 video_grid_thw: list[list[int]] | torch.Tensor,

1633 - context_len: int = 0,

1634 - seq_len: int | None = None,

1633 second_per_grid_ts: list[float] | None = None,

1634 audio_feature_lengths: torch.Tensor | None = None,

1635 use_audio_in_video: bool = False,

@@ -1759,6 +1757,5 @@ class KeyeForConditionalGeneration(

1757

1758 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1759 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1762 - llm_positions = llm_positions[:, context_len:seq_len]

1760

1761 return llm_positions, mrope_position_delta

vllm/model_executor/models/keye_vl1_5.py

-3

@@ -600,8 +600,6 @@ class KeyeVL1_5ForConditionalGeneration(

600 hf_config: PretrainedConfig,

601 image_grid_thw: list[list[int]] | torch.Tensor,

602 video_grid_thw: list[list[int]] | torch.Tensor,

603 - context_len: int = 0,

604 - seq_len: int | None = None,

603 second_per_grid_ts: list[float] | None = None,

604 audio_feature_lengths: torch.Tensor | None = None,

605 use_audio_in_video: bool = False,

@@ -729,6 +727,5 @@ class KeyeVL1_5ForConditionalGeneration(

727

728 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

729 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

732 - llm_positions = llm_positions[:, context_len:seq_len]

730

731 return llm_positions, mrope_position_delta

vllm/model_executor/models/paddleocr_vl.py

-3

@@ -1179,8 +1179,6 @@ class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, Support

1179 image_grid_thw: list[list[int]] | torch.Tensor,

1180 video_grid_thw: list[list[int]] | torch.Tensor,

1181 second_per_grid_ts: list[float],

1182 - context_len: int = 0,

1183 - seq_len: int | None = None,

1182 audio_feature_lengths: torch.Tensor | None = None,

1183 use_audio_in_video: bool = False,

1184 ) -> tuple[torch.Tensor, int]:

@@ -1293,7 +1291,6 @@ class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, Support

1291

1292 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1293 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1296 - llm_positions = llm_positions[:, context_len:seq_len]

1294

1295 return llm_positions, mrope_position_delta

1296

vllm/model_executor/models/qwen2_5_omni_thinker.py

-3

@@ -927,8 +927,6 @@ class Qwen2_5OmniThinkerForConditionalGeneration(

927 image_grid_thw: list[list[int]] | torch.Tensor,

928 video_grid_thw: list[list[int]] | torch.Tensor,

929 second_per_grid_ts: list[float] | None = None,

930 - context_len: int = 0,

931 - seq_len: int | None = None,

930 audio_feature_lengths: torch.Tensor | None = None,

931 use_audio_in_video: bool = False,

932 ) -> tuple[torch.Tensor, int]:

@@ -1125,7 +1123,6 @@ class Qwen2_5OmniThinkerForConditionalGeneration(

1123 mrope_position_delta = (

1124 torch.cat(llm_pos_ids_list, dim=1).max() + 1 - len(src_item)

1125 )

1128 - llm_positions = llm_positions[:, context_len:seq_len]

1126

1127 return llm_positions, mrope_position_delta

1128

vllm/model_executor/models/qwen2_5_vl.py

-3

@@ -1118,8 +1118,6 @@ class Qwen2_5_VLForConditionalGeneration(

1118 image_grid_thw: list[list[int]] | torch.Tensor,

1119 video_grid_thw: list[list[int]] | torch.Tensor,

1120 second_per_grid_ts: list[float],

1121 - context_len: int = 0,

1122 - seq_len: int | None = None,

1121 audio_feature_lengths: torch.Tensor | None = None,

1122 use_audio_in_video: bool = False,

1123 ) -> tuple[torch.Tensor, int]:

@@ -1232,7 +1230,6 @@ class Qwen2_5_VLForConditionalGeneration(

1230

1231 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1232 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1235 - llm_positions = llm_positions[:, context_len:seq_len]

1233

1234 return llm_positions, mrope_position_delta

1235

vllm/model_executor/models/qwen2_vl.py

-3

@@ -1240,8 +1240,6 @@ class Qwen2VLForConditionalGeneration(

1240 image_grid_thw: list[list[int]] | torch.Tensor | None,

1241 video_grid_thw: list[list[int]] | torch.Tensor | None,

1242 second_per_grid_ts: list[float] | None = None,

1243 - context_len: int = 0,

1244 - seq_len: int | None = None,

1243 audio_feature_lengths: torch.Tensor | None = None,

1244 use_audio_in_video: bool = False,

1245 ) -> tuple[torch.Tensor, int]:

@@ -1360,7 +1358,6 @@ class Qwen2VLForConditionalGeneration(

1358

1359 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1360 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1363 - llm_positions = llm_positions[:, context_len:seq_len]

1361

1362 return llm_positions, mrope_position_delta

1363

vllm/model_executor/models/qwen3_omni_moe_thinker.py

-2

@@ -1417,8 +1417,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(

1417 image_grid_thw: list[list[int]] | torch.Tensor | None,

1418 video_grid_thw: list[list[int]] | torch.Tensor | None,

1419 second_per_grid_ts: list[float] | None = None,

1420 - context_len: int = 0,

1421 - seq_len: int | None = None,

1420 audio_feature_lengths: torch.Tensor | None = None,

1421 use_audio_in_video: bool = False,

1422 ) -> tuple[torch.Tensor, int]:

vllm/model_executor/models/qwen3_vl.py

+1 -3

@@ -1419,8 +1419,6 @@ class Qwen3VLForConditionalGeneration(

1419 hf_config: PretrainedConfig,

1420 image_grid_thw: list[list[int]] | torch.Tensor,

1421 video_grid_thw: list[list[int]] | torch.Tensor,

1422 - context_len: int = 0,

1423 - seq_len: int | None = None,

1422 second_per_grid_ts: list[float] | None = None,

1423 audio_feature_lengths: torch.Tensor | None = None,

1424 use_audio_in_video: bool = False,

@@ -1519,7 +1517,7 @@ class Qwen3VLForConditionalGeneration(

1517

1518 llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)

1519 mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

1522 - llm_positions = llm_positions[:, context_len:seq_len]

1520 +

1521 return llm_positions, mrope_position_delta

1522

1523 def get_language_model(self) -> torch.nn.Module:

vllm/model_executor/models/transformers/multimodal.py

+1 -3

@@ -371,8 +371,6 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):

371 image_grid_thw: list[list[int]] | torch.Tensor | None,

372 video_grid_thw: list[list[int]] | torch.Tensor | None,

373 second_per_grid_ts: list[float] | None = None,

374 - context_len: int = 0,

375 - seq_len: int | None = None,

374 audio_feature_lengths: torch.Tensor | None = None,

375 use_audio_in_video: bool = False,

376 ) -> tuple[torch.Tensor, int]:

@@ -390,7 +388,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):

388 video_grid_thw=video_grid_thw,

389 )

390

393 - mrope_positions = mrope_positions[:, 0, context_len:seq_len]

391 + mrope_positions = mrope_positions[:, 0]

392 mrope_position_delta = mrope_position_delta[0].item()

393

394 return mrope_positions, mrope_position_delta

[V0 Deprecation] Remove unused `context_len` and `seq_len` from M-RoPE (#28395)