Skip to content

vllm.v1.worker.utils

gather_mm_placeholders 

gather_mm_placeholders(
    placeholders: Tensor, is_embed: Optional[Tensor]
) -> Tensor

Reconstructs the embeddings from the placeholder tokens.

This is the operation of [scatter_mm_placeholders][].

Source code in vllm/v1/worker/utils.py 
62
63
64
65
66
67
68
69
70
71
72
73
74
def gather_mm_placeholders(
    placeholders: torch.Tensor,
    is_embed: Optional[torch.Tensor],
) -> torch.Tensor:
    """
    Reconstructs the embeddings from the placeholder tokens.

    This is the operation of [scatter_mm_placeholders][].
    """
    if is_embed is None:
        return placeholders

    return placeholders[is_embed]

sanity_check_mm_encoder_outputs 

sanity_check_mm_encoder_outputs(
    mm_embeddings: object, expected_num_items: int
) -> None

Perform sanity checks for the result of vllm.model_executor.models.SupportsMultiModal.get_multimodal_embeddings.

Source code in vllm/v1/worker/utils.py 
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
def sanity_check_mm_encoder_outputs(
    mm_embeddings: object,
    expected_num_items: int,
) -> None:
    """
    Perform sanity checks for the result of
    [`vllm.model_executor.models.SupportsMultiModal.get_multimodal_embeddings`][].
    """
    assert isinstance(mm_embeddings, (list, tuple, torch.Tensor)), (
        "Expected multimodal embeddings to be a list/tuple of 2D tensors, "
        f"or a single 3D tensor, but got {type(mm_embeddings)} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `get_multimodal_embeddings` method.")

    assert len(mm_embeddings) == expected_num_items, (
        "Expected number of multimodal embeddings to match number of "
        f"input items: {expected_num_items}, but got {len(mm_embeddings)=} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `get_multimodal_embeddings` method.")

    assert all(e.ndim == 2 for e in mm_embeddings), (
        "Expected multimodal embeddings to be a sequence of 2D tensors, "
        f"but got tensors with shapes {[e.shape for e in mm_embeddings]} "
        "instead. This is most likely due to incorrect implementation "
        "of the model's `get_multimodal_embeddings` method.")

scatter_mm_placeholders 

scatter_mm_placeholders(
    embeds: Tensor, is_embed: Optional[Tensor]
) -> Tensor

Scatter the multimodal embeddings into a contiguous tensor that represents the placeholder tokens.

vllm.multimodal.processing.PromptUpdateDetails.is_embed.

Parameters:

Name Type Description Default
embeds Tensor

The multimodal embeddings. Shape: (num_embeds, embed_dim)

 required
is_embed Optional[Tensor]

A boolean mask indicating which positions in the placeholder tokens need to be filled with multimodal embeddings. Shape: (num_placeholders, num_embeds)

 required
Source code in vllm/v1/worker/utils.py 
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
def scatter_mm_placeholders(
    embeds: torch.Tensor,
    is_embed: Optional[torch.Tensor],
) -> torch.Tensor:
    """
    Scatter the multimodal embeddings into a contiguous tensor that represents
    the placeholder tokens.

    [`vllm.multimodal.processing.PromptUpdateDetails.is_embed`][].

    Args:
        embeds: The multimodal embeddings.
          Shape: `(num_embeds, embed_dim)`
        is_embed: A boolean mask indicating which positions in the placeholder
          tokens need to be filled with multimodal embeddings.
          Shape: `(num_placeholders, num_embeds)`
    """
    if is_embed is None:
        return embeds

    placeholders = embeds.new_full(
        (is_embed.shape[0], embeds.shape[-1]),
        fill_value=torch.nan,
    )
    placeholders[is_embed] = embeds
    return placeholders