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

# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The PygmalionAI team.
# Copyright 2023 The vLLM team.
# 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 Mistral model compatible with HuggingFace weights."""
from typing import List, Optional, Tuple

import torch
from torch import nn
from transformers import MistralConfig

from aphrodite.modeling.metadata import InputMetadata
from aphrodite.modeling.layers.activation import SiluAndMul
from aphrodite.modeling.layers.attention import PagedAttention
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (LinearMethodBase,
                                              MergedColumnParallelLinear,
                                              QKVParallelLinear,
                                              RowParallelLinear)
from aphrodite.modeling.layers.rotary_embedding import get_rope
from aphrodite.modeling.layers.sampler import Sampler
from aphrodite.modeling.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding, ParallelLMHead)
from aphrodite.modeling.megatron.parallel_state 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

KVCache = Tuple[torch.Tensor, torch.Tensor]


class MistralMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        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):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class MistralAttention(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 num_heads: int,
                 num_kv_heads: int,
                 max_position: int = 4096 * 32,
                 rope_theta: float = 10000,
                 linear_method: Optional[LinearMethodBase] = None,
                 sliding_window: Optional[int] = None) -> None:
        super().__init__()
        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)
        self.head_dim = 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.sliding_window = sliding_window

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_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,
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=self.rope_theta,
        )
        self.attn = PagedAttention(self.num_heads,
                                   self.head_dim,
                                   self.scaling,
                                   num_kv_heads=self.num_kv_heads,
                                   sliding_window=self.sliding_window)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
    ) -> 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, k = self.rotary_emb(positions, q, k)
        k_cache, v_cache = kv_cache
        attn_output = self.attn(q, k, v, k_cache, v_cache, input_metadata,
                                cache_event)
        output, _ = self.o_proj(attn_output)
        return output


class MistralDecoderLayer(nn.Module):

    def __init__(
        self,
        config: MistralConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 10000)
        self.self_attn = MistralAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            linear_method=linear_method,
            sliding_window=config.sliding_window)
        self.mlp = MistralMLP(
            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: KVCache,
        input_metadata: InputMetadata,
        cache_event: Optional[torch.cuda.Event],
        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,
            input_metadata=input_metadata,
            cache_event=cache_event,
        )

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


class MistralModel(nn.Module):

    def __init__(
        self,
        config: MistralConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> 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,
        )
        self.layers = nn.ModuleList([
            MistralDecoderLayer(config, 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[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
    ) -> torch.Tensor:
        hidden_states = self.embed_tokens(input_ids)
        residual = None
        for i in range(len(self.layers)):
            cache_event = None if cache_events is None else cache_events[i]
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                kv_caches[i],
                input_metadata,
                cache_event,
                residual,
            )
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class MistralForCausalLM(nn.Module):

    def __init__(
        self,
        config: MistralConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.model = MistralModel(config, linear_method)
        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
        self.sampler = Sampler(config.vocab_size)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[KVCache],
        input_metadata: InputMetadata,
        cache_events: Optional[List[torch.cuda.Event]],
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, kv_caches,
                                   input_metadata, cache_events)
        return hidden_states

    def sample(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> SamplerOutput:
        next_tokens = self.sampler(self.lm_head.weight, hidden_states,
                                   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"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in hf_model_weights_iterator(
                model_name_or_path, cache_dir, load_format, revision):
            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
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)