Skip to content

vllm.v1.pool.metadata

pin_memory module-attribute 

pin_memory = is_pin_memory_available()

PoolingCursor dataclass

Source code in vllm/v1/pool/metadata.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
@dataclass
class PoolingCursor:
    index: list[int]
    first_token_indices_gpu: torch.Tensor
    last_token_indices_gpu: torch.Tensor
    prompt_lens_cpu: torch.Tensor
    num_scheduled_tokens_cpu: torch.Tensor

    def __getitem__(self, indices: slice):
        return PoolingCursor(
            index=self.index[indices],
            first_token_indices_gpu=self.first_token_indices_gpu[indices],
            last_token_indices_gpu=self.last_token_indices_gpu[indices],
            prompt_lens_cpu=self.prompt_lens_cpu[indices],
            num_scheduled_tokens_cpu=self.num_scheduled_tokens_cpu[indices],
        )

    def is_partial_prefill(self):
        return not torch.all(self.prompt_lens_cpu == self.num_scheduled_tokens_cpu)

first_token_indices_gpu instance-attribute 

first_token_indices_gpu: Tensor

index instance-attribute 

index: list[int]

last_token_indices_gpu instance-attribute 

last_token_indices_gpu: Tensor

num_scheduled_tokens_cpu instance-attribute 

num_scheduled_tokens_cpu: Tensor

prompt_lens_cpu instance-attribute 

prompt_lens_cpu: Tensor

__getitem__ 

__getitem__(indices: slice)
Source code in vllm/v1/pool/metadata.py 
21
22
23
24
25
26
27
28
def __getitem__(self, indices: slice):
    return PoolingCursor(
        index=self.index[indices],
        first_token_indices_gpu=self.first_token_indices_gpu[indices],
        last_token_indices_gpu=self.last_token_indices_gpu[indices],
        prompt_lens_cpu=self.prompt_lens_cpu[indices],
        num_scheduled_tokens_cpu=self.num_scheduled_tokens_cpu[indices],
    )

__init__ 

__init__(
    index: list[int],
    first_token_indices_gpu: Tensor,
    last_token_indices_gpu: Tensor,
    prompt_lens_cpu: Tensor,
    num_scheduled_tokens_cpu: Tensor,
) -> None

is_partial_prefill 

is_partial_prefill()
Source code in vllm/v1/pool/metadata.py 
30
31
def is_partial_prefill(self):
    return not torch.all(self.prompt_lens_cpu == self.num_scheduled_tokens_cpu)

PoolingMetadata dataclass

Tensors for pooling.

Source code in vllm/v1/pool/metadata.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
60
@dataclass
class PoolingMetadata:
    """Tensors for pooling."""

    prompt_lens: torch.Tensor  # CPU Tensor
    prompt_token_ids: torch.Tensor | None
    pooling_params: list[PoolingParams]
    pooling_cursor: PoolingCursor | None = None

    def __getitem__(self, indices: slice):
        return PoolingMetadata(
            prompt_lens=self.prompt_lens[indices],
            prompt_token_ids=None
            if self.prompt_token_ids is None
            else self.prompt_token_ids[indices],
            pooling_params=self.pooling_params[indices],
            pooling_cursor=None
            if self.pooling_cursor is None
            else self.pooling_cursor[indices],
        )

    def build_pooling_cursor(
        self, num_scheduled_tokens: list[int], device: torch.device
    ):
        self.pooling_cursor = build_pooling_cursor(
            num_scheduled_tokens, self.prompt_lens, device
        )

pooling_cursor class-attribute instance-attribute 

pooling_cursor: PoolingCursor | None = None

pooling_params instance-attribute 

pooling_params: list[PoolingParams]

prompt_lens instance-attribute 

prompt_lens: Tensor

prompt_token_ids instance-attribute 

prompt_token_ids: Tensor | None

__getitem__ 

__getitem__(indices: slice)
Source code in vllm/v1/pool/metadata.py 
43
44
45
46
47
48
49
50
51
52
53
def __getitem__(self, indices: slice):
    return PoolingMetadata(
        prompt_lens=self.prompt_lens[indices],
        prompt_token_ids=None
        if self.prompt_token_ids is None
        else self.prompt_token_ids[indices],
        pooling_params=self.pooling_params[indices],
        pooling_cursor=None
        if self.pooling_cursor is None
        else self.pooling_cursor[indices],
    )

__init__ 

__init__(
    prompt_lens: Tensor,
    prompt_token_ids: Tensor | None,
    pooling_params: list[PoolingParams],
    pooling_cursor: PoolingCursor | None = None,
) -> None

build_pooling_cursor 

build_pooling_cursor(
    num_scheduled_tokens: list[int], device: device
)
Source code in vllm/v1/pool/metadata.py 
55
56
57
58
59
60
def build_pooling_cursor(
    self, num_scheduled_tokens: list[int], device: torch.device
):
    self.pooling_cursor = build_pooling_cursor(
        num_scheduled_tokens, self.prompt_lens, device
    )

build_pooling_cursor 

build_pooling_cursor(
    num_scheduled_tokens: list[int],
    prompt_lens: Tensor,
    device: device,
)
Source code in vllm/v1/pool/metadata.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def build_pooling_cursor(
    num_scheduled_tokens: list[int], prompt_lens: torch.Tensor, device: torch.device
):
    assert len(prompt_lens) == len(num_scheduled_tokens)

    n_seq = len(num_scheduled_tokens)
    index = list(range(n_seq))
    num_scheduled_tokens = torch.tensor(num_scheduled_tokens, device="cpu")
    cumsum = torch.zeros(
        n_seq + 1, dtype=torch.int64, pin_memory=pin_memory, device="cpu"
    )
    torch.cumsum(num_scheduled_tokens, dim=0, out=cumsum[1:])
    cumsum = cumsum.to(device, non_blocking=True)
    return PoolingCursor(
        index=index,
        first_token_indices_gpu=cumsum[:n_seq],
        last_token_indices_gpu=cumsum[1:] - 1,
        prompt_lens_cpu=prompt_lens,
        num_scheduled_tokens_cpu=num_scheduled_tokens,
    )