@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 57f94e88ea

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

[Models] Optimise and simplify `_validate_and_reshape_mm_tensor` (#24742)

Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>

Lukas Geiger committed Sep 12, 2025 at 16:37 UTC 57f94e88ea1ed2e48ea8ea9b01e9591f0d79557b

9 files changed +12 -10

vllm/model_executor/models/ernie45_vl.py

+1 -1

@@ -1338,7 +1338,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(nn.Module, SupportsMultiModal,

1338 raise ValueError(f"{name} should be 2D or batched 3D tensor. "

1339 f"Got ndim: {mm_input.ndim} "

1340 f"(shape={mm_input.shape})")

1341 - return torch.concat(list(mm_input))

1341 + return mm_input.reshape(-1, mm_input.shape[-1])

1342 else:

1343 return torch.concat(mm_input)

1344

vllm/model_executor/models/glm4_1v.py

+1 -1

@@ -1378,7 +1378,7 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,

1378 raise ValueError(f"{name} should be 2D or batched 3D tensor. "

1379 f"Got ndim: {mm_input.ndim} "

1380 f"(shape={mm_input.shape})")

1381 - return torch.concat(list(mm_input))

1381 + return mm_input.reshape(-1, mm_input.shape[-1])

1382 else:

1383 return torch.concat(mm_input)

1384

vllm/model_executor/models/keye.py

+2 -2

@@ -1611,12 +1611,12 @@ class KeyeForConditionalGeneration(BaseKeyeModule, SupportsMultiModal,

1611 raise ValueError(f"{name} should be 2D or batched 3D tensor. "

1612 f"Got ndim: {mm_input.ndim} "

1613 f"(shape={mm_input.shape})")

1614 - return torch.concat(list(mm_input))

1614 + return mm_input.reshape(-1, mm_input.shape[-1])

1615 elif is_list_of(mm_input, torch.Tensor):

1616 if all(p.dim() == 4 for p in mm_input) or all(p.dim() == 2

1617 for p in mm_input):

1618 return mm_input

1619 - return torch.concat(list(mm_input))

1619 + return torch.concat(mm_input)

1620

1621 def _parse_and_validate_image_input(

1622 self, **kwargs: object) -> Optional[KeyeImageInputs]:

vllm/model_executor/models/keye_vl1_5.py

+2 -2

@@ -491,14 +491,14 @@ class KeyeVL1_5ForConditionalGeneration(BaseKeyeModule, SupportsMultiModal,

491 if mm_input.ndim == expected_dim:

492 return mm_input

493 elif mm_input.ndim == expected_dim + 1:

494 - return torch.concat(list(mm_input))

494 + return mm_input.reshape(-1, *mm_input.shape[2:])

495 else:

496 raise ValueError(

497 f"{name} should be {expected_dim}D or "

498 f"batched {expected_dim}D tensor."

499 f"Got ndim: {mm_input.ndim} (shape={mm_input.shape})")

500 else:

501 - return torch.concat(list(mm_input))

501 + return torch.concat(mm_input)

502

503 def _parse_and_validate_image_input(

504 self, **kwargs: object) -> Optional[KeyeVL1_5ImageInputs]:

vllm/model_executor/models/midashenglm.py

+1 -1

@@ -669,7 +669,7 @@ class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):

669 raise ValueError(f"Incorrect type of {name}. "

670 f"Got type: {type(mm_input)}")

671 if isinstance(mm_input, torch.Tensor):

672 - return torch.concat(list(mm_input))

672 + return mm_input.reshape(-1, *mm_input.shape[2:])

673 else:

674 return torch.concat(mm_input)

675

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -551,6 +551,8 @@ class Qwen2_5OmniConditionalGenerationMixin:

551 raise ValueError(f"Incorrect type of {name}. "

552 f"Got type: {type(mm_input)}")

553 if isinstance(mm_input, torch.Tensor):

554 + if dim == 0:

555 + return mm_input.reshape(-1, *mm_input.shape[2:])

556 return torch.concat(list(mm_input), dim=dim)

557 else:

558 return torch.concat(mm_input, dim=dim)

vllm/model_executor/models/qwen2_5_vl.py

+1 -1

@@ -986,7 +986,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,

986 raise ValueError(f"{name} should be 2D or batched 3D tensor. "

987 f"Got ndim: {mm_input.ndim} "

988 f"(shape={mm_input.shape})")

989 - return torch.concat(list(mm_input))

989 + return mm_input.reshape(-1, mm_input.shape[-1])

990 else:

991 return torch.concat(mm_input)

992

vllm/model_executor/models/qwen2_audio.py

+1 -1

@@ -342,7 +342,7 @@ class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal,

342 raise ValueError(f"Incorrect type of {name}. "

343 f"Got type: {type(mm_input)}")

344 if isinstance(mm_input, torch.Tensor):

345 - return torch.concat(list(mm_input))

345 + return mm_input.reshape(-1, *mm_input.shape[2:])

346 else:

347 return torch.concat(mm_input)

348

vllm/model_executor/models/qwen2_vl.py

+1 -1

@@ -1167,7 +1167,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,

1167 raise ValueError(f"{name} should be 2D or batched 3D tensor. "

1168 f"Got ndim: {mm_input.ndim} "

1169 f"(shape={mm_input.shape})")

1170 - return torch.concat(list(mm_input))

1170 + return mm_input.reshape(-1, mm_input.shape[-1])

1171 else:

1172 return torch.concat(mm_input)

1173