sgl-project--sglang/python/sglang/srt/models/gpt2.py

# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt2/modeling_gpt2.py
# Copyright 2023 The vLLM team.
# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# 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.
"""Inference-only GPT-2 model compatible with HuggingFace weights."""

from typing import Iterable, Optional, Tuple, Type

import torch
from torch import nn
from transformers import GPT2Config

from sglang.srt.layers.activation import NewGELU
from sglang.srt.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.runtime_context import get_parallel
from sglang.srt.utils import add_prefix


class GPT2Attention(nn.Module):

    def __init__(
        self,
        layer_id: int,
        config: GPT2Config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        total_num_heads = config.num_attention_heads
        tensor_model_parallel_world_size = get_parallel().tp_size
        assert total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = total_num_heads // tensor_model_parallel_world_size
        self.head_dim = self.hidden_size // total_num_heads
        self.scale = self.head_dim**-0.5

        self.c_attn = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            total_num_heads,
            bias=True,
            quant_config=quant_config,
            prefix=add_prefix("c_attn", prefix),
        )
        self.c_proj = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=True,
            quant_config=quant_config,
            prefix=add_prefix("c_proj", prefix),
        )
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            scaling=self.scale,
            num_kv_heads=total_num_heads,
            layer_id=layer_id,
            quant_config=quant_config,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        qkv, _ = self.c_attn(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        attn_output = self.attn(q, k, v, forward_batch)
        attn_output, _ = self.c_proj(attn_output)
        return attn_output


class GPT2MLP(nn.Module):

    def __init__(
        self,
        intermediate_size: int,
        config: GPT2Config,
        act_layer: Type[nn.Module] = NewGELU,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        hidden_size = config.hidden_size
        self.c_fc = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            bias=True,
            quant_config=quant_config,
            prefix=add_prefix("c_fc", prefix),
        )
        self.c_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=True,
            quant_config=quant_config,
            prefix=add_prefix("c_proj", prefix),
        )
        self.act = act_layer()

    def forward(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        hidden_states, _ = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.c_proj(hidden_states)
        return hidden_states


class GPT2Block(nn.Module):

    def __init__(
        self,
        layer_id: int,
        config: GPT2Config,
        act_layer: Type[nn.Module] = NewGELU,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        hidden_size = config.hidden_size
        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size

        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.attn = GPT2Attention(
            layer_id, config, quant_config, prefix=add_prefix("attn", prefix)
        )
        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
        self.mlp = GPT2MLP(
            inner_dim,
            config,
            act_layer=act_layer,
            quant_config=quant_config,
            prefix=add_prefix("mlp", prefix),
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_output = self.attn(
            hidden_states=hidden_states,
            forward_batch=forward_batch,
        )
        # residual connection
        hidden_states = attn_output + residual

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.mlp(hidden_states)
        # residual connection
        hidden_states = residual + feed_forward_hidden_states
        return hidden_states


class GPT2Model(nn.Module):

    def __init__(
        self,
        config: GPT2Config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        assert not config.add_cross_attention
        assert not config.scale_attn_by_inverse_layer_idx
        assert not config.reorder_and_upcast_attn
        self.embed_dim = config.hidden_size
        self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
        self.h = nn.ModuleList(
            [
                GPT2Block(
                    i,
                    config,
                    quant_config=quant_config,
                    prefix=add_prefix(f"h.{i}", prefix),
                )
                for i in range(config.num_hidden_layers)
            ]
        )

        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        for i in range(len(self.h)):
            layer = self.h[i]
            hidden_states = layer(hidden_states, forward_batch)

        hidden_states = self.ln_f(hidden_states)
        return hidden_states


class GPT2LMHeadModel(nn.Module):

    def __init__(
        self,
        config: GPT2Config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.transformer = GPT2Model(
            config, quant_config, prefix=add_prefix("transformer", prefix)
        )
        self.lm_head = self.transformer.wte

        self.logits_processor = LogitsProcessor(config)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        hidden_states = self.transformer(input_ids, positions, forward_batch)
        return self.logits_processor(
            input_ids, hidden_states, self.lm_head, forward_batch
        )

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        for name, loaded_weight in weights:
            if "lm_head.weight" in name:
                # GPT-2 ties the weights of the embedding layer and the final
                # linear layer.
                continue
            if ".attn.bias" in name or ".attn.masked_bias" in name:
                # Skip attention mask.
                # NOTE: "c_attn.bias" should not be skipped.
                continue
            if not name.startswith("transformer."):
                name = "transformer." + name

            param = params_dict[name]
            # The HF's GPT-2 implementation uses Conv1D instead of Linear.
            # Because of this, we need to transpose the weights.
            # Note(zhuohan): the logic below might break quantized models.
            for conv1d_weight_name in ["c_attn", "c_proj", "c_fc"]:
                if conv1d_weight_name not in name:
                    continue
                if not name.endswith(".weight"):
                    continue
                loaded_weight = loaded_weight.t()
            weight_loader = getattr(param, "weight_loader", default_weight_loader)
            weight_loader(param, loaded_weight)


EntryClass = GPT2LMHeadModel