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 
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
@dataclass
class PoolingCursor:
    index: list[int]
    first_token_indices_gpu: torch.Tensor
    last_token_indices_gpu: torch.Tensor
    prompt_lens_cpu: torch.Tensor
    seq_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],
            seq_lens_cpu=self.seq_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)

    def is_finished(self):
        return self.prompt_lens_cpu == self.seq_lens_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

seq_lens_cpu instance-attribute 

seq_lens_cpu: Tensor

__getitem__ 

__getitem__(indices: slice)
Source code in vllm/v1/pool/metadata.py 
23
24
25
26
27
28
29
30
31
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],
        seq_lens_cpu=self.seq_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,
    seq_lens_cpu: Tensor,
    num_scheduled_tokens_cpu: Tensor,
) -> None

is_finished 

is_finished()
Source code in vllm/v1/pool/metadata.py 
36
37
def is_finished(self):
    return self.prompt_lens_cpu == self.seq_lens_cpu

is_partial_prefill 

is_partial_prefill()
Source code in vllm/v1/pool/metadata.py 
33
34
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 
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 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
@dataclass
class PoolingMetadata:
    """Tensors for pooling."""

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

    def __post_init__(self) -> None:
        pooling_params = self.pooling_params

        tasks: list[PoolingTask] = [
            task
            for pooling_param in pooling_params
            if (task := pooling_param.task) is not None
        ]
        assert len(pooling_params) == len(tasks)

        self.tasks = tasks

    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_states=self.pooling_states[indices],
            pooling_cursor=None
            if self.pooling_cursor is None
            else self.pooling_cursor[indices],
        )

    def get_prompt_token_ids(self) -> list[torch.Tensor]:
        prompt_token_ids = self.prompt_token_ids
        assert prompt_token_ids is not None, (
            "Please set `requires_token_ids=True` in `get_pooling_updates`"
        )

        return [prompt_token_ids[i, :num] for i, num in enumerate(self.prompt_lens)]

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

pooling_cursor class-attribute instance-attribute 

pooling_cursor: PoolingCursor | None = None

pooling_params instance-attribute 

pooling_params: list[PoolingParams]

pooling_states instance-attribute 

pooling_states: list[PoolingStates]

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 
71
72
73
74
75
76
77
78
79
80
81
82
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_states=self.pooling_states[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_states: list[PoolingStates],
    pooling_cursor: PoolingCursor | None = None,
) -> None

__post_init__ 

__post_init__() -> None
Source code in vllm/v1/pool/metadata.py 
59
60
61
62
63
64
65
66
67
68
69
def __post_init__(self) -> None:
    pooling_params = self.pooling_params

    tasks: list[PoolingTask] = [
        task
        for pooling_param in pooling_params
        if (task := pooling_param.task) is not None
    ]
    assert len(pooling_params) == len(tasks)

    self.tasks = tasks

build_pooling_cursor 

build_pooling_cursor(
    num_scheduled_tokens: list[int],
    seq_lens_cpu: Tensor,
    device: device,
)
Source code in vllm/v1/pool/metadata.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
def build_pooling_cursor(
    self,
    num_scheduled_tokens: list[int],
    seq_lens_cpu: torch.Tensor,
    device: torch.device,
):
    self.pooling_cursor = build_pooling_cursor(
        num_scheduled_tokens, seq_lens_cpu, self.prompt_lens, device
    )

get_prompt_token_ids 

get_prompt_token_ids() -> list[Tensor]
Source code in vllm/v1/pool/metadata.py 
84
85
86
87
88
89
90
def get_prompt_token_ids(self) -> list[torch.Tensor]:
    prompt_token_ids = self.prompt_token_ids
    assert prompt_token_ids is not None, (
        "Please set `requires_token_ids=True` in `get_pooling_updates`"
    )

    return [prompt_token_ids[i, :num] for i, num in enumerate(self.prompt_lens)]

PoolingStates

Source code in vllm/v1/pool/metadata.py 
40
41
42
43
44
45
46
class PoolingStates:
    def __init__(self):
        # for chunked prefill with ALL pooling
        self.hidden_states_cache: list[torch.Tensor] = []

    def clean(self):
        self.hidden_states_cache.clear()

hidden_states_cache instance-attribute 

hidden_states_cache: list[Tensor] = []

__init__ 

__init__()
Source code in vllm/v1/pool/metadata.py 
41
42
43
def __init__(self):
    # for chunked prefill with ALL pooling
    self.hidden_states_cache: list[torch.Tensor] = []

clean 

clean()
Source code in vllm/v1/pool/metadata.py 
45
46
def clean(self):
    self.hidden_states_cache.clear()

build_pooling_cursor 

build_pooling_cursor(
    num_scheduled_tokens: list[int],
    seq_lens_cpu: Tensor,
    prompt_lens: Tensor,
    device: device,
)
Source code in vllm/v1/pool/metadata.py 
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
def build_pooling_cursor(
    num_scheduled_tokens: list[int],
    seq_lens_cpu: torch.Tensor,
    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_cpu = 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_cpu, 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,
        seq_lens_cpu=seq_lens_cpu,
        num_scheduled_tokens_cpu=num_scheduled_tokens_cpu,
    )