[V0 Deprecation] Remove pooling model support in V0 (#23434)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Maximilien de Bayser
2025-08-29 00:04:02 -07:00
committed by GitHub
co-authored by Woosuk Kwon
parent 934bebf192
commit 2554b27baa
38 changed files with 99 additions and 808 deletions
+7 -25
View File
@@ -13,17 +13,12 @@ import torch.nn.functional as F
from transformers import PretrainedConfig
from vllm.config import ModelConfig, PoolerConfig
from vllm.model_executor.pooling_metadata import ( # noqa: E501
PoolingMetadata as V0PoolingMetadata)
from vllm.model_executor.pooling_metadata import PoolingTensors
from vllm.pooling_params import PoolingParams
from vllm.sequence import PoolerOutput, PoolingSequenceGroupOutput
from vllm.tasks import PoolingTask
from vllm.utils import current_stream, resolve_obj_by_qualname
from vllm.v1.pool.metadata import PoolingCursor
from vllm.v1.pool.metadata import PoolingMetadata as V1PoolingMetadata
from vllm.v1.pool.metadata import PoolingCursor, PoolingMetadata
PoolingMetadata = Union[V0PoolingMetadata, V1PoolingMetadata]
PoolingFn = Callable[
[Union[torch.Tensor, list[torch.Tensor]], PoolingMetadata],
Union[torch.Tensor, list[torch.Tensor]]]
@@ -127,36 +122,23 @@ def get_prompt_lens(
hidden_states: Union[torch.Tensor, list[torch.Tensor]],
pooling_metadata: PoolingMetadata,
) -> torch.Tensor:
if isinstance(pooling_metadata, V1PoolingMetadata):
return pooling_metadata.prompt_lens
return PoolingTensors.from_pooling_metadata(
pooling_metadata, hidden_states[0].device).prompt_lens
return pooling_metadata.prompt_lens
def get_prompt_token_ids(
pooling_metadata: PoolingMetadata) -> list[torch.Tensor]:
if isinstance(pooling_metadata, V1PoolingMetadata):
assert pooling_metadata.prompt_token_ids is not None, (
"Please set `requires_token_ids=True` in `get_pooling_updates`")
return [
pooling_metadata.prompt_token_ids[i, :num]
for i, num in enumerate(pooling_metadata.prompt_lens)
]
assert pooling_metadata.prompt_token_ids is not None, (
"Please set `requires_token_ids=True` in `get_pooling_updates`")
return [
torch.tensor(seq_data_i.prompt_token_ids)
for seq_data_i in pooling_metadata.seq_data.values()
pooling_metadata.prompt_token_ids[i, :num]
for i, num in enumerate(pooling_metadata.prompt_lens)
]
def get_pooling_params(
pooling_metadata: PoolingMetadata) -> list[PoolingParams]:
if isinstance(pooling_metadata, V0PoolingMetadata):
pooling_params = [p for _, p in pooling_metadata.seq_groups]
else:
pooling_params = pooling_metadata.pooling_params
pooling_params = pooling_metadata.pooling_params
return pooling_params
+1 -1
View File
@@ -24,9 +24,9 @@ from vllm.model_executor.layers.pooler import (ClassifierPooler,
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding)
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.sequence import IntermediateTensors
from vllm.tasks import PoolingTask
from vllm.v1.pool.metadata import PoolingMetadata
from .interfaces import SupportsCrossEncoding, SupportsQuant
from .interfaces_base import default_pooling_type
+1 -1
View File
@@ -15,10 +15,10 @@ from vllm.model_executor.layers.pooler import (DispatchPooler, Pooler,
build_output, get_prompt_lens,
get_prompt_token_ids)
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.sequence import PoolerOutput
from vllm.tasks import PoolingTask
from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
from vllm.v1.pool.metadata import PoolingMetadata
from .interfaces_base import default_pooling_type
+1 -1
View File
@@ -22,9 +22,9 @@ from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.sequence import IntermediateTensors
from vllm.tasks import PoolingTask
from vllm.v1.pool.metadata import PoolingMetadata
from .interfaces import SupportsCrossEncoding
from .interfaces_base import default_pooling_type
-90
View File
@@ -1,90 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, Optional
import torch
from vllm.pooling_params import PoolingParams
from vllm.utils import is_pin_memory_available
from vllm.v1.pool.metadata import PoolingCursor, build_pooling_cursor
class PoolingMetadata:
"""Metadata for pooling operations in the Pooler layer.
This class holds the necessary information for pooling operations,
providing context for how to perform pooling and other related operations.
Attributes:
seq_groups: List of (seq_ids, pooling_params).
seq_data: A mapping of sequence ID to additional sequence data.
prompt_lens: List of the lengths of each prompt.
"""
def __init__(
self,
seq_groups: list[tuple[list[int], PoolingParams]],
seq_data: dict[int, Any], # Specific data related to sequences
prompt_lens: list[int],
pooling_cursor: Optional[PoolingCursor] = None) -> None:
self.seq_groups = seq_groups
self.seq_data = seq_data
self.prompt_lens = prompt_lens
self.pooling_cursor: Optional[PoolingCursor] = pooling_cursor
def __repr__(self) -> str:
return ("PoolingMetadata("
f"seq_groups={self.seq_groups}, "
f"seq_data={self.seq_data}, "
f"prompt_lens={self.prompt_lens})")
def __getitem__(self, indices: slice):
return PoolingMetadata(
seq_groups=self.seq_groups[indices],
seq_data=dict(list(self.seq_data.items())[indices]),
prompt_lens=self.prompt_lens[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):
prompt_lens = torch.tensor(self.prompt_lens, device="cpu")
self.pooling_cursor = build_pooling_cursor(num_scheduled_tokens,
prompt_lens,
device=device)
@dataclass
class PoolingTensors:
"""Tensors for pooling."""
prompt_lens: torch.Tensor
@classmethod
def from_pooling_metadata(
cls,
pooling_metadata: "PoolingMetadata",
device: torch.device,
) -> "PoolingTensors":
"""
Create PoolingTensors from PoolingMetadata.
Args:
pooling_metadata: PoolingMetadata instance to convert.
device: Device to store the tensors.
"""
# Convert prompt lengths to tensor
pin_memory = is_pin_memory_available()
prompt_lens_t = torch.tensor(
pooling_metadata.prompt_lens,
device="cpu",
dtype=torch.long,
pin_memory=pin_memory,
)
return cls(prompt_lens=prompt_lens_t.to(device=device,
non_blocking=True), )