srt/models/bailing_moe_nextn.py

# coding=utf-8
# Copyright 2023 Antgroup and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.
"""SGLang BailingMoENextN model."""
import logging
from typing import Iterable, Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from sglang.srt.distributed import get_tensor_model_parallel_world_size
from sglang.srt.layers.dp_attention import is_dp_attention_enabled
from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.models.bailing_moe import BailingMoEBlock, BailingMoEForCausalLM
from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils import add_prefix

LoraConfig = None
logger = logging.getLogger(__name__)


class BailingMoEModelNextN(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
            logger.warning(
                "Overriding DeepseekV3ForCausalLMNextN quant config for modelopt_fp4 Deepseek model."
            )
            quant_config = None

        self.vocab_size = config.vocab_size

        self.word_embeddings = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            enable_tp=not is_dp_attention_enabled(),
            prefix=add_prefix("word_embeddings", prefix),
        )

        self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.eh_proj = nn.Linear(2 * config.hidden_size, config.hidden_size, bias=False)

        self.decoder = BailingMoEBlock(
            config,
            0,
            quant_config=quant_config,
            # is_nextn=True,
            prefix=add_prefix("decoder", prefix),
        )

        self.shared_head = nn.Module()
        self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:

        if input_embeds is None:
            hidden_states = self.word_embeddings(input_ids)
        else:
            hidden_states = input_embeds

        if hidden_states.shape[0] > 0:
            hidden_states = self.eh_proj(
                torch.cat(
                    (
                        self.enorm(hidden_states),
                        self.hnorm(forward_batch.spec_info.hidden_states),
                    ),
                    dim=-1,
                )
            )

        residual = None
        hidden_states, residual = self.decoder(
            positions, hidden_states, forward_batch, residual
        )

        if not forward_batch.forward_mode.is_idle():
            if residual is not None:
                hidden_states, _ = self.final_layernorm(hidden_states, residual)
            else:
                hidden_states = self.final_layernorm(hidden_states)

        return hidden_states


class BailingMoeForCausalLMNextN(BailingMoEForCausalLM):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        nn.Module.__init__(self)
        self.config = config
        self.tp_size = get_tensor_model_parallel_world_size()
        self.quant_config = quant_config
        if hasattr(self, "determine_num_fused_shared_experts"):
            # Asystem has determine_num_fused_shared_experts but theta does not.
            self.determine_num_fused_shared_experts("BailingMoeForCausalLMNextN")

        self.model = BailingMoEModelNextN(
            config, quant_config, prefix=add_prefix("model", prefix)
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=add_prefix("model.shared_head.head", prefix),
            use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
        )
        self.logits_processor = LogitsProcessor(config)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        hidden_states = self.model(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]]):
        super().load_weights(weights, is_nextn=True)


EntryClass = [BailingMoeForCausalLMNextN]