wehub-resource-sync
241 lines
8.2 KiB
Python
241 lines
8.2 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
import dataclasses
|
|
from collections.abc import Mapping, Set
|
|
from itertools import groupby
|
|
|
|
import torch
|
|
|
|
from vllm.config import PoolerConfig
|
|
from vllm.model_executor.layers.pooler import PoolingParamsUpdate
|
|
from vllm.tasks import PoolingTask
|
|
from vllm.v1.pool.metadata import PoolingMetadata
|
|
|
|
from .abstract import Pooler, PoolerOutput
|
|
from .common import ClassifierFn
|
|
from .seqwise import (
|
|
SequencePoolingFn,
|
|
SequencePoolingMethod,
|
|
pooler_for_classify,
|
|
pooler_for_embed,
|
|
)
|
|
from .tokwise import AllPool, pooler_for_token_classify, pooler_for_token_embed
|
|
|
|
|
|
class DispatchPooler(Pooler):
|
|
"""Dispatches calls to a sub-pooler based on the pooling task."""
|
|
|
|
@classmethod
|
|
def for_embedding(cls, pooler_config: PoolerConfig):
|
|
return cls(
|
|
{
|
|
"token_embed": pooler_for_token_embed(pooler_config),
|
|
"embed": pooler_for_embed(pooler_config),
|
|
},
|
|
)
|
|
|
|
@classmethod
|
|
def for_seq_cls(
|
|
cls,
|
|
pooler_config: PoolerConfig,
|
|
*,
|
|
pooling: SequencePoolingMethod | SequencePoolingFn | None = None,
|
|
classifier: ClassifierFn | None = None,
|
|
):
|
|
return cls(
|
|
{
|
|
"token_classify": pooler_for_token_classify(
|
|
pooler_config,
|
|
pooling=AllPool(),
|
|
classifier=classifier,
|
|
),
|
|
"classify": pooler_for_classify(
|
|
pooler_config,
|
|
pooling=pooling,
|
|
classifier=classifier,
|
|
),
|
|
}
|
|
)
|
|
|
|
def __init__(self, poolers_by_task: Mapping[PoolingTask, Pooler]) -> None:
|
|
super().__init__()
|
|
|
|
for task, pooler in poolers_by_task.items():
|
|
if task not in pooler.get_supported_tasks():
|
|
raise ValueError(
|
|
f"{pooler=} does not support {task=}. "
|
|
f"Supported tasks: {pooler.get_supported_tasks()}"
|
|
)
|
|
|
|
self.poolers_by_task = poolers_by_task
|
|
|
|
def get_supported_tasks(self) -> Set[PoolingTask]:
|
|
return set(self.poolers_by_task)
|
|
|
|
def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
|
|
return self.poolers_by_task[task].get_pooling_updates(task)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
pooling_metadata: PoolingMetadata,
|
|
) -> PoolerOutput:
|
|
poolers_by_task = self.poolers_by_task
|
|
cursor = pooling_metadata.pooling_cursor
|
|
|
|
outputs = list[torch.Tensor | None]()
|
|
offset = 0
|
|
token_offset = 0
|
|
for task, group in groupby(pooling_metadata.tasks):
|
|
if not (pooler := poolers_by_task.get(task)):
|
|
raise ValueError(
|
|
f"Unsupported task: {task!r} "
|
|
f"Supported tasks: {self.get_supported_tasks()}"
|
|
)
|
|
|
|
num_items = len(list(group))
|
|
group_metadata = pooling_metadata[offset : offset + num_items]
|
|
if cursor is None:
|
|
group_hidden_states = hidden_states
|
|
else:
|
|
# Slice out this group's tokens so sub-poolers see only their
|
|
# portion of the batch. Token offset is computed from the CPU
|
|
# `num_scheduled_tokens_cpu` to avoid a GPU->CPU sync.
|
|
group_cursor = group_metadata.pooling_cursor
|
|
assert group_cursor is not None
|
|
num_group_tokens = int(group_cursor.num_scheduled_tokens_cpu.sum())
|
|
group_hidden_states = hidden_states[
|
|
token_offset : token_offset + num_group_tokens
|
|
]
|
|
if token_offset:
|
|
# Shift first/last indices to be relative to the slice
|
|
# so seqwise poolers (which index `hidden_states` directly)
|
|
# remain correct.
|
|
pooling_cursor = dataclasses.replace(
|
|
group_cursor,
|
|
first_token_indices_gpu=(
|
|
group_cursor.first_token_indices_gpu - token_offset
|
|
),
|
|
last_token_indices_gpu=(
|
|
group_cursor.last_token_indices_gpu - token_offset
|
|
),
|
|
)
|
|
group_metadata = dataclasses.replace(
|
|
group_metadata, pooling_cursor=pooling_cursor
|
|
)
|
|
token_offset += num_group_tokens
|
|
|
|
group_output: PoolerOutput = pooler(group_hidden_states, group_metadata)
|
|
|
|
outputs.extend(group_output)
|
|
offset += num_items
|
|
|
|
return outputs
|
|
|
|
def extra_repr(self) -> str:
|
|
s = f"supported_task={self.get_supported_tasks()}"
|
|
return s
|
|
|
|
|
|
class IdentityPooler(Pooler):
|
|
def get_supported_tasks(self) -> Set[PoolingTask]:
|
|
return {"plugin"}
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
pooling_metadata: PoolingMetadata,
|
|
) -> PoolerOutput:
|
|
return hidden_states
|
|
|
|
|
|
class BOSEOSFilter(Pooler):
|
|
"""Filters the BOS and EOS token results from outputs."""
|
|
|
|
def __init__(
|
|
self,
|
|
pooler: Pooler,
|
|
bos_token_id: int = -1, # -1 disables the filtering
|
|
eos_token_id: int = -1,
|
|
) -> None:
|
|
super().__init__()
|
|
|
|
self.pooler = pooler
|
|
self.bos_token_id = bos_token_id
|
|
self.eos_token_id = eos_token_id
|
|
|
|
def extra_repr(self) -> str:
|
|
return f"bos_token_id={self.bos_token_id}, eos_token_id={self.eos_token_id}"
|
|
|
|
def get_supported_tasks(self) -> Set[PoolingTask]:
|
|
return self.pooler.get_supported_tasks()
|
|
|
|
def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
|
|
return PoolingParamsUpdate(requires_token_ids=True)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor | list[torch.Tensor],
|
|
pooling_metadata: PoolingMetadata,
|
|
) -> PoolerOutput:
|
|
pooled_outputs = self.pooler(hidden_states, pooling_metadata)
|
|
assert isinstance(pooled_outputs, list)
|
|
prompt_token_ids = pooling_metadata.get_prompt_token_ids_cpu()
|
|
|
|
for i, (prompt_len, token_ids) in enumerate(
|
|
zip(pooling_metadata.prompt_lens, prompt_token_ids)
|
|
):
|
|
pooled_data = pooled_outputs[i]
|
|
assert (
|
|
isinstance(pooled_data, torch.Tensor)
|
|
and pooled_data.shape[0] == prompt_len
|
|
)
|
|
if int(token_ids[0]) == self.bos_token_id:
|
|
pooled_data = pooled_data[1:]
|
|
if int(token_ids[-1]) == self.eos_token_id:
|
|
pooled_data = pooled_data[:-1]
|
|
pooled_outputs[i] = pooled_data.squeeze(-1)
|
|
|
|
return pooled_outputs
|
|
|
|
|
|
class BgeM3Pooler(Pooler):
|
|
def __init__(self, token_classify_pooler: Pooler, embed_pooler: Pooler) -> None:
|
|
super().__init__()
|
|
self.token_classify_pooler = token_classify_pooler
|
|
self.embed_pooler = embed_pooler
|
|
|
|
def forward(
|
|
self, hidden_states: torch.Tensor, pooling_metadata: PoolingMetadata
|
|
) -> PoolerOutput:
|
|
embed_outputs = self.embed_pooler(hidden_states, pooling_metadata)
|
|
token_classify_outputs = self.token_classify_pooler(
|
|
hidden_states, pooling_metadata
|
|
)
|
|
pooler_outputs: list[torch.Tensor] = []
|
|
for embed_output, token_classify_output in zip(
|
|
embed_outputs, token_classify_outputs
|
|
):
|
|
pooler_outputs.append(
|
|
torch.cat(
|
|
[embed_output.view(-1), token_classify_output.view(-1)], dim=-1
|
|
)
|
|
)
|
|
|
|
return pooler_outputs
|
|
|
|
def get_supported_tasks(self) -> Set[PoolingTask]:
|
|
return {"embed&token_classify"}
|
|
|
|
def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
|
|
return self.embed_pooler.get_pooling_updates(
|
|
"embed"
|
|
) | self.token_classify_pooler.get_pooling_updates("token_classify")
|
|
|
|
def extra_repr(self) -> str:
|
|
s = f"supported_task={self.get_supported_tasks()}"
|
|
return s
|
|
|
|
|
|
__all__ = ["BOSEOSFilter", "DispatchPooler", "IdentityPooler", "BgeM3Pooler"]
|