Skip to content

vllm.model_executor.models.glmasr_utils

_calculate_conv_output_length(input_length, padding, kernel_size, stride)

Calculate Conv1d output length using standard formula.

Source code in vllm/model_executor/models/glmasr_utils.py 
17
18
19
20
21
22
def _calculate_conv_output_length(
    input_length: torch.Tensor, padding: int, kernel_size: int, stride: int
) -> torch.Tensor:
    """Calculate Conv1d output length using standard formula."""
    # in sync with `hf_processor._get_audio_token_length`
    return (input_length + 2 * padding - (kernel_size - 1) - 1) // stride + 1

_get_audio_output_lengths_for_tower(audio_tower, audio_lengths, merge_factor, conv_params)

Calculate the output lengths after audio processing.

The output length accounts for: 1. Convolution layers (downsampling) 2. Merge factor (further downsampling during projection)

Parameters:

  • audio_tower

    (Module) –

    The audio encoder module

  • audio_lengths

    (Tensor) –

    Input feature lengths [batch_size]

  • merge_factor

    (int) –

    Factor for merging adjacent features

  • conv_params

    (list[tuple[int, int, int]]) –

    List of (padding, kernel_size, stride) for each conv layer

Returns:

  • Tensor –

    Output lengths after all processing [batch_size]

Source code in vllm/model_executor/models/glmasr_utils.py 
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
def _get_audio_output_lengths_for_tower(
    audio_tower: nn.Module,
    audio_lengths: torch.Tensor,
    merge_factor: int,
    conv_params: list[tuple[int, int, int]],
) -> torch.Tensor:
    """
    Calculate the output lengths after audio processing.

    The output length accounts for:
    1. Convolution layers (downsampling)
    2. Merge factor (further downsampling during projection)

    Args:
        audio_tower: The audio encoder module
        audio_lengths: Input feature lengths [batch_size]
        merge_factor: Factor for merging adjacent features
        conv_params: List of (padding, kernel_size, stride) for each conv layer

    Returns:
        Output lengths after all processing [batch_size]
    """
    # First, calculate the output length after convolutions
    if hasattr(audio_tower, "_get_feat_extract_output_lengths"):
        _, conv_output_lengths = audio_tower._get_feat_extract_output_lengths(
            audio_lengths
        )
    else:
        conv_output_lengths = audio_lengths
        for padding, kernel_size, stride in conv_params:
            conv_output_lengths = _calculate_conv_output_length(
                conv_output_lengths, padding, kernel_size, stride
            )

    # Then, apply merge_factor to get final output length
    # Formula: (conv_output_lengths - merge_factor) // merge_factor + 1
    return (conv_output_lengths - merge_factor) // merge_factor + 1