aphrodite-engine/aphrodite/modeling/models/baichuan.py

# coding=utf-8
# Copyright 2022 EleutherAI 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.
"""Inference-only BaiChuan model compatible with HuggingFace weights."""

import math
from typing import List, Optional, Tuple

import torch
from torch import nn

from aphrodite.attention import Attention, AttentionMetadata
from aphrodite.modeling.layers.activation import SiluAndMul
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (
    LinearMethodBase,
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
    ColumnParallelLinear,
)
from aphrodite.modeling.layers.rotary_embedding import get_rope
from aphrodite.modeling.layers.logits_processor import LogitsProcessor
from aphrodite.modeling.layers.sampler import Sampler
from aphrodite.modeling.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding,
    ParallelLMHead,
)
from aphrodite.distributed import (
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
)
from aphrodite.modeling.sampling_metadata import SamplingMetadata
from aphrodite.modeling.hf_downloader import (
    default_weight_loader,
    hf_model_weights_iterator,
)
from aphrodite.common.sequence import SamplerOutput
from aphrodite.transformers_utils.configs.baichuan import BaiChuanConfig


def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
    closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
    base = torch.tensor(
        2**(-(2**-(math.log2(closest_power_of_2) - 3))),
        dtype=torch.float32,
    )
    powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != total_num_heads:
        extra_base = torch.tensor(
            2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
            dtype=torch.float32,
        )
        num_remaining_heads = min(closest_power_of_2,
                                  total_num_heads - closest_power_of_2)
        extra_powers = torch.arange(start=1,
                                    end=1 + 2 * num_remaining_heads,
                                    step=2,
                                    dtype=torch.int32)
        slopes = torch.cat(
            [slopes, torch.pow(extra_base, extra_powers)], dim=0)
    return slopes


class BaiChuanMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        if (linear_method is not None
                and not linear_method.quant_config.merge_weight()):
            self.merge_weight = False
            self.gate_proj = ColumnParallelLinear(
                hidden_size,
                intermediate_size,
                bias=False,
                linear_method=linear_method,
            )
            self.up_proj = ColumnParallelLinear(
                hidden_size,
                intermediate_size,
                bias=False,
                linear_method=linear_method,
            )
        else:
            self.merge_weight = True
            self.gate_up_proj = MergedColumnParallelLinear(
                hidden_size,
                [intermediate_size] * 2,
                bias=False,
                linear_method=linear_method,
            )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            linear_method=linear_method,
        )
        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):
        if self.merge_weight:
            gate_up, _ = self.gate_up_proj(x)
        else:
            up, _ = self.up_proj(x)
            gate, _ = self.gate_proj(x)
            gate_up = torch.cat([gate, up], dim=-1)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class BaiChuanAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        position_embedding: str,
        rope_theta: float = 10000,
        max_position_embeddings: int = 8192,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.hidden_size = hidden_size
        tensor_model_parallel_world_size = (
            get_tensor_model_parallel_world_size())
        self.total_num_heads = num_heads
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = (self.total_num_heads //
                          tensor_model_parallel_world_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.postion_embedding = position_embedding
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        # pylint: disable=invalid-name
        self.W_pack = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_heads,
            bias=False,
            linear_method=linear_method,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            linear_method=linear_method,
        )
        # Create the alibi slopes and slice them.
        if self.postion_embedding == "ALIBI":
            tp_rank = get_tensor_model_parallel_rank()
            head_start = tp_rank * self.num_heads
            head_end = (tp_rank + 1) * self.num_heads
            alibi_slopes = _get_alibi_slopes(self.total_num_heads)
            alibi_slopes = alibi_slopes[head_start:head_end].tolist()

            scaling = self.head_dim**-0.5
            self.attn = Attention(
                self.num_heads,
                self.head_dim,
                scaling,
                alibi_slopes=alibi_slopes,
            )
        else:
            is_neox_style = (True if linear_method is None
                             or linear_method.quant_config.rope_style() is None
                             else linear_method.quant_config.rope_style())
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.head_dim,
                max_position=self.max_position_embeddings,
                base=self.rope_theta,
                is_neox_style=is_neox_style,
            )
            self.scaling = self.head_dim**-0.5
            self.attn = Attention(self.num_heads, self.head_dim, self.scaling)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        qkv, _ = self.W_pack(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        if self.postion_embedding != "ALIBI":
            q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.o_proj(attn_output)
        return output


class BaiChuanDecoderLayer(nn.Module):

    def __init__(
        self,
        config: BaiChuanConfig,
        position_embedding: str,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = BaiChuanAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            position_embedding=position_embedding,
            rope_theta=rope_theta,
            max_position_embeddings=max_position_embeddings,
            linear_method=linear_method,
        )
        self.mlp = BaiChuanMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            linear_method=linear_method,
        )
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)

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

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


class BaiChuanModel(nn.Module):

    def __init__(
        self,
        config: BaiChuanConfig,
        position_embedding: str,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                                   config.hidden_size,
                                                   linear_method=linear_method)
        self.layers = nn.ModuleList([
            BaiChuanDecoderLayer(config, position_embedding, linear_method)
            for _ in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        hidden_states = self.embed_tokens(input_ids)
        residual = None
        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                kv_caches[i],
                attn_metadata,
                residual,
            )
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class BaiChuanBaseForCausalLM(nn.Module):

    def __init__(
        self,
        config,
        position_embedding: str,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.model = BaiChuanModel(config, position_embedding, linear_method)
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      linear_method=linear_method)
        self.logits_processor = LogitsProcessor(config.vocab_size,
                                                config.tokenizer_vocab_size)
        self.sampler = Sampler()

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
                                   attn_metadata)
        return hidden_states

    def compute_logits(self, hidden_states: torch.Tensor,
                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        next_tokens = self.sampler(logits, sampling_metadata)
        return next_tokens

    def load_weights(
        self,
        model_name_or_path: str,
        cache_dir: Optional[str] = None,
        load_format: str = "auto",
        revision: Optional[str] = None,
    ):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        if (self.linear_method is not None
                and not self.linear_method.quant_config.merge_weight()):
            stacked_params_mapping = []
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in hf_model_weights_iterator(
                model_name_or_path, cache_dir, load_format, revision,
                self.config):
            if "rotary_emb.inv_freq" in name:
                continue
            if name == "lm_head.weight":
                # Unlike Baichuan, Baichuan2 normalizes the head weights. Ref.:
                # https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat/blob/84603cde5ebffb6084e476cfaeceaf0b8b91fe54/modeling_baichuan.py#L508
                # Distinguish between Baichuan and Baichuan2 by checking the
                # vocab size.
                is_baichuan2 = self.config.vocab_size == 125696
                if is_baichuan2:
                    loaded_weight = torch.nn.functional.normalize(
                        loaded_weight)

            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
                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
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)


class BaichuanForCausalLM(BaiChuanBaseForCausalLM):
    """Baichuan 13B and Baichuan2 7B/13B."""

    def __init__(self,
                 config,
                 linear_method: Optional[LinearMethodBase] = None):
        if config.hidden_size == 4096:  # baichuan2 7b
            super().__init__(config, "ROPE", linear_method)
        else:  # baichuan 13b, baichuan2 13b
            super().__init__(config, "ALIBI", linear_method)


class BaiChuanForCausalLM(BaiChuanBaseForCausalLM):
    """Baichuan 7B."""

    def __init__(self,
                 config,
                 linear_method: Optional[LinearMethodBase] = None):
        super().__init__(config, "ROPE", linear_method)