sgl-project--sglang/python/sglang/srt/models/qwen2_classification.py

# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

from typing import Iterable, Optional, Tuple

import torch
from torch import nn
from transformers import Qwen2Config

from sglang.srt.layers.pooler import (
    EmbeddingPoolerOutput,
    Pooler,
    PoolingType,
    score_and_pool,
)
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.models.qwen2 import Qwen2ForCausalLM, Qwen2Model
from sglang.srt.utils import add_prefix


class Qwen2ForSequenceClassification(nn.Module):
    def __init__(
        self,
        config: Qwen2Config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.model = Qwen2Model(
            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
        )
        self.score = nn.Linear(config.hidden_size, config.num_labels)
        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=False)

        self.eos_token_id = config.eos_token_id

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
        get_embedding: bool = True,
    ) -> EmbeddingPoolerOutput:
        assert (
            get_embedding
        ), "Qwen2ForSequenceClassification is only used for embedding"

        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
        return score_and_pool(
            self.score, self.pooler, hidden_states, forward_batch, input_ids
        )

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        # Filter out lm_head weights of Qwen2ForCausalLM
        filtered_weights = [
            (name, w) for name, w in weights if not name.startswith("lm_head")
        ]
        return Qwen2ForCausalLM.load_weights(self, filtered_weights)


EntryClass = [
    Qwen2ForSequenceClassification,
]