Skip to content

vllm.lora.ops.torch_ops.lora_ops

bgmv_expand 

bgmv_expand(
    inputs: Tensor,
    lora_b_weights: Tensor,
    output_tensor: Tensor,
    lora_indices_tensor: Tensor,
    add_inputs: bool = True,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def bgmv_expand(
    inputs: torch.Tensor,
    lora_b_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    add_inputs: bool = True,
):
    selected_loras = lora_b_weights[lora_indices_tensor].to(dtype=output_tensor.dtype)
    if len(selected_loras.shape) == 4:
        selected_loras = selected_loras.squeeze(dim=1)
    inputs = inputs.to(dtype=output_tensor.dtype)
    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)

    limit = output_tensor.shape[0]
    if outputs.shape[0] == 1 and output_tensor.shape[0] != 1:
        limit = 1

    # LoRA adapter and model may add different amounts of padding to output
    common_len = min(outputs.shape[1], output_tensor.shape[1])

    if add_inputs:
        output_tensor[:, :common_len] += outputs[:limit, :common_len]
    else:
        output_tensor[:, :common_len] = outputs[:limit, :common_len]

bgmv_expand_slice 

bgmv_expand_slice(
    inputs: Tensor,
    lora_b_weights: Tensor,
    output_tensor: Tensor,
    lora_indices_tensor: Tensor,
    slice_offset: int,
    slice_size: int,
    add_inputs: bool = True,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
def bgmv_expand_slice(
    inputs: torch.Tensor,
    lora_b_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    slice_offset: int,
    slice_size: int,
    add_inputs: bool = True,
):
    selected_loras = lora_b_weights[lora_indices_tensor].to(dtype=output_tensor.dtype)
    inputs = inputs.to(dtype=output_tensor.dtype)
    if len(selected_loras.shape) == 4:
        selected_loras = selected_loras.squeeze(dim=1)
    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)

    if add_inputs:
        output_tensor[:, slice_offset : slice_offset + slice_size] += outputs[:]
    else:
        output_tensor[:, slice_offset : slice_offset + slice_size] = outputs[:]

bgmv_shrink 

bgmv_shrink(
    inputs: Tensor,
    lora_b_weights: Tensor,
    output_tensor: Tensor,
    lora_indices_tensor: Tensor,
    scaling: float = 1.0,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
67
68
69
70
71
72
73
74
75
76
77
78
79
80
def bgmv_shrink(
    inputs: torch.Tensor,
    lora_b_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    scaling: float = 1.0,
):
    selected_loras = lora_b_weights[lora_indices_tensor].to(dtype=output_tensor.dtype)
    if len(selected_loras.shape) == 4:
        selected_loras = selected_loras.squeeze(dim=1)
    inputs = inputs.to(dtype=output_tensor.dtype)
    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)

    output_tensor[:, : outputs.shape[1]] = scaling * outputs[:]

sgmv_expand 

sgmv_expand(
    inputs: Tensor,
    lora_b_weights: Tensor,
    output_tensor: Tensor,
    b_seq_start_loc: Tensor,
    seq_len_tensor: Tensor,
    lora_indices_tensor: Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    add_inputs: bool = False,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
def sgmv_expand(
    inputs: torch.Tensor,
    lora_b_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    b_seq_start_loc: torch.Tensor,
    seq_len_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    add_inputs: bool = False,
):
    exploded_indices = torch.repeat_interleave(lora_indices_tensor, seq_len_tensor)

    bgmv_expand(inputs, lora_b_weights, output_tensor, exploded_indices, add_inputs)

sgmv_expand_slice 

sgmv_expand_slice(
    inputs: Tensor,
    lora_b_weights: Tensor,
    output_tensor: Tensor,
    b_seq_start_loc: Tensor,
    seq_len_tensor: Tensor,
    lora_indices_tensor: Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    slice_offset: int,
    slice_size: int,
    add_inputs: bool = False,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def sgmv_expand_slice(
    inputs: torch.Tensor,
    lora_b_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    b_seq_start_loc: torch.Tensor,
    seq_len_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    slice_offset: int,
    slice_size: int,
    add_inputs: bool = False,
):
    exploded_indices = torch.repeat_interleave(lora_indices_tensor, seq_len_tensor)

    bgmv_expand_slice(
        inputs,
        lora_b_weights,
        output_tensor,
        exploded_indices,
        slice_offset,
        slice_size,
        add_inputs,
    )

sgmv_shrink 

sgmv_shrink(
    inputs: Tensor,
    lora_a_weights: Tensor,
    output_tensor: Tensor,
    b_seq_start_loc: Tensor,
    seq_len_tensor: Tensor,
    lora_indices_tensor: Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    scaling: float,
)
Source code in vllm/lora/ops/torch_ops/lora_ops.py 
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
def sgmv_shrink(
    inputs: torch.Tensor,
    lora_a_weights: torch.Tensor,
    output_tensor: torch.Tensor,
    b_seq_start_loc: torch.Tensor,
    seq_len_tensor: torch.Tensor,
    lora_indices_tensor: torch.Tensor,
    batches: int,
    max_seq_length: int,
    token_nums: int,
    scaling: float,
):
    exploded_indices = torch.repeat_interleave(lora_indices_tensor, seq_len_tensor)

    bgmv_shrink(inputs, lora_a_weights, output_tensor, exploded_indices, scaling)