srt/models/apertus.py

# Copyright 2025 The SwissAI Initiative
# 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.
# ==============================================================================

# Adapted from
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/llama.py#L1
"""Inference-only Apertus model compatible with HuggingFace weights."""

import logging
from typing import Any, Dict, Iterable, List, Optional, Tuple, Union

import torch
from torch import nn
from transformers import ApertusConfig

from sglang.srt.distributed import (
    get_pp_group,
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
)
from sglang.srt.layers.activation import XIELU
from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.linear import (
    ColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
from sglang.srt.layers.pooler import Pooler, PoolingType
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.rotary_embedding import get_rope
from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
from sglang.srt.layers.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
from sglang.srt.model_loader.weight_utils import (
    default_weight_loader,
    kv_cache_scales_loader,
    maybe_remap_kv_scale_name,
)
from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils import add_prefix, make_layers

logger = logging.getLogger(__name__)


class ApertusMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        prefix: str = "",
        reduce_results: bool = True,
    ) -> None:
        super().__init__()
        self.up_proj = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("up_proj", prefix),
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
            reduce_results=reduce_results,
        )
        if hidden_act != "xielu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only xIELU is supported for now."
            )
        self.act_fn = XIELU()

    def forward(
        self,
        x,
        forward_batch=None,
        use_reduce_scatter: bool = False,
    ):
        # note: with xielu, there's no gate_proj
        x, _ = self.up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(
            x,
            skip_all_reduce=use_reduce_scatter,
        )
        return x


class ApertusAttention(nn.Module):
    def __init__(
        self,
        config: ApertusConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        layer_id: int = 0,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        rope_is_neox_style: bool = True,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        bias: bool = False,
        bias_o_proj: bool = False,
    ) -> None:
        super().__init__()
        self.layer_id = layer_id
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
        self.head_dim = getattr(
            config, "head_dim", self.hidden_size // self.total_num_heads
        )
        partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
        self.rotary_dim = int(partial_rotary_factor * self.head_dim)
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=bias_o_proj,
            quant_config=quant_config,
            prefix=add_prefix("o_proj", prefix),
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=rope_is_neox_style,
        )
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            layer_id=layer_id,
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
        )
        self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q = self.q_norm(q.contiguous().view(-1, self.head_dim)).view_as(q)
        k = self.k_norm(k.contiguous().view(-1, self.head_dim)).view_as(k)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, forward_batch)
        output, _ = self.o_proj(attn_output)
        return output


class ApertusDecoderLayer(nn.Module):
    def __init__(
        self,
        config: ApertusConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
            config, "original_max_position_embeddings", None
        ):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings
            )
        rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        # Support llamafy/Qwen-Qwen2.5-7B-Instruct-llamafied with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
            config, "bias", False
        )
        bias_o_proj = attention_bias
        # support internlm/internlm3-8b with qkv_bias
        if hasattr(config, "qkv_bias"):
            attention_bias = config.qkv_bias
        self.self_attn = ApertusAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            layer_id=layer_id,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            rope_is_neox_style=rope_is_neox_style,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            prefix=add_prefix("self_attn", prefix),
            bias=attention_bias,
            bias_o_proj=bias_o_proj,
        )
        self.mlp = ApertusMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            bias=getattr(config, "mlp_bias", False),
            prefix=add_prefix("mlp", prefix),
        )
        self.attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.feedforward_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        residual: Optional[torch.Tensor],
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.attention_layernorm(hidden_states)
        else:
            hidden_states, residual = self.attention_layernorm(hidden_states, residual)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
        )

        # Fully Connected
        hidden_states, residual = self.feedforward_layernorm(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


class ApertusModel(nn.Module):
    def __init__(
        self,
        config: ApertusConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.quant_config = quant_config
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.org_vocab_size = config.vocab_size
        self.pp_group = get_pp_group()
        if self.pp_group.is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=add_prefix("embed_tokens", prefix),
            )
        else:
            self.embed_tokens = PPMissingLayer()

        self.layers, self.start_layer, self.end_layer = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: ApertusDecoderLayer(
                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
            ),
            pp_rank=self.pp_group.rank_in_group,
            pp_size=self.pp_group.world_size,
            prefix="model.layers",
        )

        if self.pp_group.is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer(return_tuple=True)
        self.layers_to_capture = []

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]], PPProxyTensors]:
        if self.pp_group.is_first_rank:
            if input_embeds is None:
                hidden_states = self.embed_tokens(input_ids)
            else:
                hidden_states = input_embeds
            residual = None
        else:
            assert pp_proxy_tensors is not None
            # FIXME(@ying): reduce the number of proxy tensors by not fusing layer norms
            hidden_states = pp_proxy_tensors["hidden_states"]
            residual = pp_proxy_tensors["residual"]
            deferred_norm = None

        aux_hidden_states = []
        for i in range(self.start_layer, self.end_layer):
            if i in self.layers_to_capture:
                aux_hidden_states.append(hidden_states + residual)
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                forward_batch,
                residual,
            )

        if not self.pp_group.is_last_rank:
            return PPProxyTensors(
                {
                    "hidden_states": hidden_states,
                    "residual": residual,
                }
            )
        else:
            hidden_states, _ = self.norm(hidden_states, residual)

        if len(aux_hidden_states) == 0:
            return hidden_states

        return hidden_states, aux_hidden_states

    # If this function is called, it should always initialize KV cache scale
    # factors (or else raise an exception). Thus, handled exceptions should
    # make sure to leave KV cache scale factors in a known good (dummy) state
    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
        tp_size = get_tensor_model_parallel_world_size()
        tp_rank = get_tensor_model_parallel_rank()
        for layer_idx, scaling_factor in kv_cache_scales_loader(
            quantization_param_path,
            tp_rank,
            tp_size,
            self.config.num_hidden_layers,
            self.config.__class__.model_type,
        ):
            if not isinstance(self.layers[layer_idx], nn.Identity):
                layer_self_attn = self.layers[layer_idx].self_attn

            if hasattr(layer_self_attn.attn, "k_scale"):
                layer_self_attn.attn.k_scale = scaling_factor
                layer_self_attn.attn.v_scale = scaling_factor
            else:
                raise RuntimeError(
                    "Self attention has no KV cache scaling " "factor attribute!"
                )


class ApertusForCausalLM(nn.Module):
    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]
    # BitandBytes specific attributes
    default_bitsandbytes_target_modules = [
        ".down_proj.",
        ".up_proj.",
        ".q_proj.",
        ".k_proj.",
        ".v_proj.",
        ".o_proj.",
    ]
    # in TP, these weights are partitioned along the column dimension (dim=-1)
    column_parallel_weights_modules = [".down_proj.", ".o_proj."]
    bitsandbytes_stacked_params_mapping = {
        # shard_name, weight_name, index
        ".q_proj": (".qkv_proj", 0),
        ".k_proj": (".qkv_proj", 1),
        ".v_proj": (".qkv_proj", 2),
    }

    def __init__(
        self,
        config: ApertusConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.pp_group = get_pp_group()
        self.config = config
        self.quant_config = quant_config
        self.model = self._init_model(config, quant_config, add_prefix("model", prefix))
        if self.config.tie_word_embeddings:
            self.lm_head = self.model.embed_tokens
        else:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=add_prefix("lm_head", prefix),
                use_attn_tp_group=get_global_server_args().enable_dp_lm_head,
            )
        self.logits_processor = LogitsProcessor(config)
        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
        self.stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
        ]

        self.capture_aux_hidden_states = False

    def _init_model(
        self,
        config: ApertusConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        return ApertusModel(config, quant_config=quant_config, prefix=prefix)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
        get_embedding: bool = False,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
    ) -> LogitsProcessorOutput:
        hidden_states = self.model(
            input_ids,
            positions,
            forward_batch,
            input_embeds,
            pp_proxy_tensors=pp_proxy_tensors,
        )

        aux_hidden_states = None
        if self.capture_aux_hidden_states:
            hidden_states, aux_hidden_states = hidden_states

        if self.pp_group.is_last_rank:
            if not get_embedding:
                return self.logits_processor(
                    input_ids,
                    hidden_states,
                    self.lm_head,
                    forward_batch,
                    aux_hidden_states,
                )
            else:
                return self.pooler(hidden_states, forward_batch)
        else:
            return hidden_states

    @torch.no_grad()
    def forward_split_prefill(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        split_interval: Tuple[int, int],  # [start, end) 0-based
        input_embeds: torch.Tensor = None,
    ) -> Optional[LogitsProcessorOutput]:
        start, end = split_interval
        # embed
        if start == 0:
            if input_embeds is None:
                forward_batch.hidden_states = self.model.embed_tokens(input_ids)
            else:
                forward_batch.hidden_states = input_embeds
        # decoder layer
        for i in range(start, end):
            layer = self.model.layers[i]
            forward_batch.hidden_states, forward_batch.residual = layer(
                positions,
                forward_batch.hidden_states,
                forward_batch,
                forward_batch.residual,
            )

        if end == self.model.config.num_hidden_layers:
            # norm
            hidden_states, _ = self.model.norm(
                forward_batch.hidden_states, forward_batch.residual
            )
            forward_batch.hidden_states = hidden_states
            # logits process
            result = self.logits_processor(
                input_ids, forward_batch.hidden_states, self.lm_head, forward_batch
            )
        else:
            result = None

        return result

    @property
    def start_layer(self):
        return self.model.start_layer

    @property
    def end_layer(self):
        return self.model.end_layer

    def get_input_embeddings(self) -> nn.Embedding:
        return self.model.embed_tokens

    def get_module_name_from_weight_name(self, name):
        for param_name, weight_name, shard_id, num_shard in self.stacked_params_mapping:
            if weight_name in name:
                return (
                    name.replace(weight_name, param_name)[: -len(".weight")],
                    num_shard,
                )
        return name[: -len(".weight")], 1

    def get_num_params(self):
        params_dict = dict(self.named_parameters())
        return len(params_dict)

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
        ]

        params_dict = dict(self.named_parameters())

        for name, buffer in self.named_buffers():
            if name.endswith(".beta") or name.endswith(".eps"):
                params_dict[name] = buffer

        for name, loaded_weight in weights:
            layer_id = get_layer_id(name)
            if (
                layer_id is not None
                and hasattr(self.model, "start_layer")
                and (
                    layer_id < self.model.start_layer
                    or layer_id >= self.model.end_layer
                )
            ):
                continue
            if "rotary_emb.inv_freq" in name or "projector" in name:
                continue
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                continue
            if name.startswith("model.vision_tower") and name not in params_dict:
                continue
            if self.config.tie_word_embeddings and "lm_head.weight" in name:
                continue
            # Handle FP8 kv-scale remapping
            if "scale" in name:
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip loading kv_scale from ckpts towards new design.
                if name.endswith(".kv_scale") and name not in params_dict:
                    continue
                if name in params_dict.keys():
                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)
                else:
                    logger.warning(f"Parameter {name} not found in params_dict")

    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_embed_and_head(self, embed, head):
        del self.model.embed_tokens.weight
        del self.lm_head.weight
        self.model.embed_tokens.weight = embed
        self.lm_head.weight = head
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    def get_embed(self):
        return self.model.embed_tokens.weight

    def set_embed(self, embed):
        # NOTE: If draft hidden size != target hidden size, the embed weight cannot be shared for EAGLE3
        if (
            hasattr(self.config, "target_hidden_size")
            and self.config.target_hidden_size != self.config.hidden_size
        ):
            return
        del self.model.embed_tokens.weight
        self.model.embed_tokens.weight = embed
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
        self.model.load_kv_cache_scales(quantization_param_path)

    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
        if not self.pp_group.is_last_rank:
            return

        if layer_ids is None:
            self.capture_aux_hidden_states = True
            num_layers = self.config.num_hidden_layers
            self.model.layers_to_capture = [2, num_layers // 2, num_layers - 3]
        else:
            self.capture_aux_hidden_states = True
            # we plus 1 here because in sglang, for the ith layer, it takes the output
            # of the (i-1)th layer as aux hidden state
            self.model.layers_to_capture = [val + 1 for val in layer_ids]


EntryClass = [ApertusForCausalLM]