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

# coding=utf-8
# Adapted from
# https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_phi.py
# Copyright 2023 The PygmalionAI team.
# Copyright 2023 The vLLM team.
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
#
# BSD 3-Clause License
#
# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# * Redistributions of source code must retain the above copyright notice, this
#   list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above copyright notice,
#   this list of conditions and the following disclaimer in the documentation
#   and/or other materials provided with the distribution.
#
# * Neither the name of the copyright holder nor the names of its
#   contributors may be used to endorse or promote products derived from
#   this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""Inference-only Phi model compatible with HuggingFace weights."""

from typing import List, Optional

import torch
from torch import nn
from transformers import PretrainedConfig

from aphrodite.attention import Attention, AttentionMetadata
from aphrodite.modeling.layers.activation import get_act_fn
from aphrodite.modeling.layers.linear import (
    ColumnParallelLinear,
    LinearMethodBase,
    QKVParallelLinear,
    RowParallelLinear,
)
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_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


class PhiAttention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.total_num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.total_num_heads

        tensor_model_parallel_world_size = (
            get_tensor_model_parallel_world_size())
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = (self.total_num_heads //
                          tensor_model_parallel_world_size)

        # pylint: disable=C0103
        if (linear_method is not None
                and not linear_method.quant_config.merge_weight()):
            self.merge_weight = False
            self.q_proj = ColumnParallelLinear(
                self.hidden_size,
                self.hidden_size,
                bias=True,
                linear_method=linear_method,
            )
            self.k_proj = ColumnParallelLinear(
                self.hidden_size,
                self.hidden_size,
                bias=True,
                linear_method=linear_method,
            )
            self.v_proj = ColumnParallelLinear(
                self.hidden_size,
                self.hidden_size,
                bias=True,
                linear_method=linear_method,
            )
        else:
            self.merge_weight = True
            self.qkv_proj = QKVParallelLinear(
                self.hidden_size,
                self.head_size,
                self.total_num_heads,
                bias=True,
                linear_method=linear_method,
            )
        self.dense = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            linear_method=linear_method,
        )

        scaling = self.head_size**-0.5
        rotary_dim = int(config.partial_rotary_factor *
                         (config.hidden_size // config.num_attention_heads))
        assert rotary_dim % 2 == 0

        # pylint: disable=C0301
        # Refer to:
        # https://huggingface.co/microsoft/phi-1_5/blob/d212a789620c380ff32ca1d1ee9943a777360987/modeling_phi.py#L518
        rope_theta = 10000
        max_position_embeddings = getattr(config, "n_positions", 2048)
        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_size,
            rotary_dim=rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            is_neox_style=is_neox_style,
        )
        self.attn = Attention(self.num_heads, self.head_size, scaling)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        if self.merge_weight:
            qkv, _ = self.qkv_proj(hidden_states)
            q, k, v = qkv.chunk(chunks=3, dim=-1)
        else:
            q, _ = self.q_proj(hidden_states)
            k, _ = self.k_proj(hidden_states)
            v, _ = self.v_proj(hidden_states)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.dense(attn_output)
        return output


class PhiMLP(nn.Module):

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

        n_inner = getattr(config, "n_inner", None)
        n_inner = n_inner if n_inner is not None else 4 * config.hidden_size

        self.fc1 = ColumnParallelLinear(
            config.hidden_size,
            n_inner,
            linear_method=linear_method,
        )
        self.fc2 = RowParallelLinear(
            n_inner,
            config.hidden_size,
            linear_method=linear_method,
        )
        quant_config = getattr(linear_method, "quant_config", None)
        self.act = get_act_fn(config.hidden_act, quant_config, n_inner)

    def forward(self, hidden_states):
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)
        return hidden_states


class PhiLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.layer_norm_eps)
        self.self_attn = PhiAttention(config, linear_method)
        self.mlp = PhiMLP(config, linear_method)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        attn_outputs = self.self_attn(
            position_ids=position_ids,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            attn_metadata=attn_metadata,
        )
        feed_forward_hidden_states = self.mlp(hidden_states)
        hidden_states = attn_outputs + feed_forward_hidden_states + residual
        return hidden_states


class PhiModel(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                                   config.hidden_size,
                                                   linear_method=linear_method)
        self.layers = nn.ModuleList([
            PhiLayer(config, linear_method)
            for _ in range(config.num_hidden_layers)
        ])
        self.final_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.layer_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)
        for i in range(self.config.num_hidden_layers):
            layer = self.layers[i]
            hidden_states = layer(
                positions,
                hidden_states,
                kv_caches[i],
                attn_metadata,
            )

        hidden_states = self.final_layernorm(hidden_states)

        return hidden_states


class PhiForCausalLM(nn.Module):

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

        self.model = PhiModel(config, linear_method)

        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            bias=True,
            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, self.lm_head.bias)
        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)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
        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

            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
                # pylint: disable=E1136

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