sgl-project--sglang/python/sglang/srt/models/exaone4.py

from collections.abc import Iterable
from typing import Any, List, Optional, Tuple, Union

import torch
from torch import nn
from transformers import Exaone4Config

from sglang.srt.distributed import get_pp_group
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.linear import (
    MergedColumnParallelLinear,
    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,
    maybe_remap_kv_scale_name,
)
from sglang.srt.runtime_context import get_parallel, get_server_args
from sglang.srt.utils import add_prefix, make_layers
from sglang.utils import get_exception_traceback, logger


# Aligned with HF's implementation, using sliding window inclusive with the last token
# SGLang assumes exclusive
def get_attention_sliding_window_size(config):
    if getattr(config, "sliding_window", None) is not None:
        return config.sliding_window - 1
    else:
        return None


class Exaone4GatedMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class Exaone4Attention(nn.Module):
    def __init__(
        self,
        config,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        layer_id: int = 0,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-06,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = False,
        bias_o_proj: bool = False,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_parallel().tp_size

        attn_tp_rank = get_parallel().attn_tp_rank
        attn_tp_size = get_parallel().attn_tp_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)

        self.head_dim = head_dim or hidden_size // self.total_num_heads
        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=hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )

        self.o_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=hidden_size,
            bias=bias_o_proj,
            quant_config=quant_config,
            prefix=add_prefix("o_proj", prefix),
            tp_rank=attn_tp_rank,
            tp_size=attn_tp_size,
        )

        is_neox_style = True
        if quant_config is not None and quant_config.get_name() == "gguf":
            is_neox_style = False

        interleaved_sliding_window = get_attention_sliding_window_size(config)
        self.sliding_window_pattern = getattr(config, "sliding_window_pattern", None)

        self.is_sliding = False
        if self.sliding_window_pattern:
            if (layer_id + 1) % len(self.sliding_window_pattern) != 0:
                self.is_sliding = True

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=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,
            sliding_window_size=(
                interleaved_sliding_window if self.is_sliding else None
            ),
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
        )

        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
        self.k_norm = RMSNorm(self.head_dim, eps=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)

        # Add qk-norm
        q_shape = q.shape
        q = q.reshape(-1, self.head_dim)
        q = self.q_norm(q)
        q = q.reshape(q_shape)

        k_shape = k.shape
        k = k.reshape(-1, self.head_dim)
        k = self.k_norm(k)
        k = k.reshape(k_shape)

        if not self.sliding_window_pattern or self.is_sliding:
            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 Exaone4DecoderLayer(nn.Module):
    def __init__(
        self,
        config: Exaone4Config,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.layer_id = layer_id
        self.hidden_size = config.hidden_size

        rope_theta = getattr(config, "rope_theta", 1000000)
        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
            )

        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)

        self.attn_tp_size = get_parallel().attn_tp_size
        self.attn_tp_rank = get_parallel().attn_tp_rank

        self.self_attn = Exaone4Attention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=getattr(
                config, "num_key_value_heads", config.num_key_value_heads
            ),
            layer_id=layer_id,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            prefix=add_prefix("self_attn", prefix),
        )
        self.mlp = Exaone4GatedMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=add_prefix("mlp", prefix),
        )
        self.post_attention_layernorm = RMSNorm(
            self.hidden_size, eps=config.rms_norm_eps
        )
        self.post_feedforward_layernorm = RMSNorm(
            self.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]:

        if residual is None:
            residual = hidden_states

        # Self Attention
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            forward_batch=forward_batch,
        )

        # Use post-LN
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = hidden_states + residual
        residual = hidden_states

        # Fully Connected
        hidden_states = self.mlp(hidden_states)

        # Use post-LN
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = hidden_states + residual
        residual = hidden_states

        return hidden_states, residual


class Exaone4Model(nn.Module):
    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.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: Exaone4DecoderLayer(
                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=add_prefix("layers", prefix),
        )
        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)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: Optional[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.get_input_embeddings(input_ids)
            else:
                hidden_states = input_embeds
            residual = None
        else:
            assert pp_proxy_tensors is not None
            hidden_states = pp_proxy_tensors["hidden_states"]
            residual = pp_proxy_tensors["residual"]

        for i in range(len(self.layers)):
            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)
        return hidden_states


class Exaone4ForCausalLM(nn.Module):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    base_model_prefix = "language_model"

    # BitandBytes specific attributes
    default_bitsandbytes_target_modules = [
        ".gate_proj.",
        ".down_proj.",
        ".up_proj.",
        ".q_proj.",
        ".k_proj.",
        ".v_proj.",
        ".o_proj.",
    ]
    bitsandbytes_stacked_params_mapping = {
        ".q_proj": (".qkv_proj", 0),
        ".k_proj": (".qkv_proj", 1),
        ".v_proj": (".qkv_proj", 2),
        ".gate_proj": (".gate_up_proj", 0),
        ".up_proj": (".gate_up_proj", 1),
    }

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        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))
        # Exaone-4.0 32B set tie_word_embeddins to False
        # Exaone-4.0 1.2B set tie_word_embeddins to True
        if 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_server_args().enable_dp_lm_head,
            )

        self.logits_processor = LogitsProcessor(config)
        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)

    def _init_model(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        return Exaone4Model(config, quant_config=quant_config, prefix=prefix)

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

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: Optional[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,
        )

        if self.pp_group.is_last_rank:
            if not get_embedding:
                return self.logits_processor(
                    input_ids,
                    hidden_states,
                    self.lm_head,
                    forward_batch,
                )
            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,
    ):
        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_attention_sliding_window_size(self):
        return get_attention_sliding_window_size(self.config)

    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"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]

        params_dict = dict(self.named_parameters())

        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_weights_by_name(
        self, name: str, truncate_size: int = 100, tp_size: int = 1
    ) -> Optional[torch.Tensor]:
        """Get the weights of the parameter by its name. Similar to `get_parameter` in Hugging Face.

        Only used for unit test with an unoptimized performance.
        For optimized performance, please use torch.save and torch.load.
        """
        try:
            if name == "lm_head.weight" and self.config.tie_word_embeddings:
                logger.info(
                    "word embedding is tied for this model, return embed_tokens.weight as lm_head.weight."
                )
                return (
                    self.model.embed_tokens.weight.cpu()
                    .to(torch.float32)
                    .numpy()
                    .tolist()[:truncate_size]
                )

            mapped_name = name
            mapped_shard_id = None
            for param_name, weight_name, shard_id in self.stacked_params_mapping:
                if weight_name in name:
                    mapped_name = name.replace(weight_name, param_name)
                    mapped_shard_id = shard_id
                    break
            params_dict = dict(self.named_parameters())
            param = params_dict[mapped_name]
            if mapped_shard_id is not None:
                if mapped_shard_id in ["q", "k", "v"]:
                    num_heads = self.config.num_attention_heads // tp_size
                    num_kv_heads = self.config.num_key_value_heads // tp_size
                    head_dim = (
                        self.config.hidden_size // self.config.num_attention_heads
                    )
                    if mapped_shard_id == "q":
                        offset = 0
                        size = num_heads * head_dim
                    elif mapped_shard_id == "k":
                        offset = num_heads * head_dim
                        size = num_kv_heads * head_dim
                    elif mapped_shard_id == "v":
                        offset = (num_heads + num_kv_heads) * head_dim
                        size = num_kv_heads * head_dim
                    weight = param.data.narrow(0, offset, size)
                elif mapped_shard_id in [0, 1]:
                    intermediate_size = self.config.intermediate_size
                    slice_size = intermediate_size // tp_size
                    if mapped_shard_id == 0:  # gate_proj
                        offset = 0
                        size = slice_size
                    elif mapped_shard_id == 1:  # up_proj
                        offset = slice_size
                        size = slice_size

                    weight = param.data.narrow(0, offset, size)
                else:
                    weight = param.data
            else:
                weight = param.data
            if tp_size > 1 and ("o_proj" in name or "down_proj" in name):
                gathered_weights = [torch.zeros_like(weight) for _ in range(tp_size)]
                torch.distributed.all_gather(gathered_weights, weight)
                weight = torch.cat(gathered_weights, dim=1)
            return weight.cpu().to(torch.float32).numpy().tolist()[:truncate_size]

        except Exception:
            logger.error(
                f"Error getting weights by name {name} in Exaone4ForCausalLM: {get_exception_traceback()}"
            )
            return None

    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)


EntryClass = Exaone4ForCausalLM