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

# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2024 The X-AI team.
# Copyright 2024 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 Grok model."""

from typing import List, Optional, Tuple

import numpy as np

import torch
import torch.nn.functional as F

from torch import nn

from aphrodite.modeling.metadata import InputMetadata
from aphrodite.modeling.layers.attention import PagedAttention
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (
    LinearMethodBase,
    ReplicatedLinear,
    QKVParallelLinear,
    RowParallelLinear,
    ColumnParallelLinear,
)
from aphrodite.modeling.layers.rotary_embedding import get_rope
from aphrodite.modeling.layers.sampler import Sampler, QuantSampler
from aphrodite.modeling.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding,
    ParallelLMHead,
)
from aphrodite.modeling.megatron.communication_op import (
    tensor_model_parallel_all_reduce, )
from aphrodite.modeling.megatron.parallel_state 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 import GrokConfig

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


class GrokMLP(nn.Module):

    def __init__(
        self,
        num_experts: int,
        hidden_size: int,
        intermediate_size: int,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.num_experts = num_experts
        self.ffn_dim = intermediate_size
        self.hidden_dim = hidden_size

        self.linear = ReplicatedLinear(
            self.hidden_dim,
            self.ffn_dim,
            bias=False,
            linear_method=linear_method,
        )
        self.linear_1 = ReplicatedLinear(
            self.ffn_dim,
            self.hidden_dim,
            bias=False,
            linear_method=linear_method,
        )
        self.linear_v = ReplicatedLinear(
            self.hidden_dim,
            self.ffn_dim,
            bias=False,
            linear_method=linear_method,
        )

        # TODO: Use Aphrodite's SiluAndMul
        self.act_fn = nn.SiLU()

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        linear_out, _ = self.linear(hidden_states)
        linear_out = self.act_fn(linear_out)
        linear_v_out, _ = self.linear_v(hidden_states)
        current_hidden_states = linear_out * linear_v_out
        current_hidden_states, _ = self.linear_1(current_hidden_states)
        return current_hidden_states


class GrokMoE(nn.Module):

    def __init__(
        self,
        config: GrokConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ):
        super().__init__()
        self.config = config
        self.rank = get_tensor_model_parallel_rank()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.num_total_experts = config.num_local_experts
        self.top_k = config.num_experts_per_tok
        if self.tp_size > self.num_total_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {self.num_total_experts}.")
        # Split experts equally between ranks
        self.expert_indicies = np.array_split(range(
            self.num_total_experts), self.tp_size)[self.rank].tolist()
        if not self.expert_indicies:
            raise ValueError(
                f"Rank {self.rank} has no experts assigned to it.")

        self.experts = nn.ModuleList([
            GrokMLP(
                self.num_total_experts,
                config.hidden_size,
                config.intermediate_size,
                linear_method=linear_method,
            ) if idx in self.expert_indicies else None
            for idx in range(self.num_total_experts)
        ])
        self.gate = ReplicatedLinear(
            config.hidden_size,
            self.num_total_experts,
            bias=False,
            linear_method=None,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits, _ = self.gate(hidden_states)

        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights,
                                                       self.top_k,
                                                       dim=-1)
        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)

        final_hidden_states = None
        for expert_idx in self.expert_indicies:
            expert_layer = self.experts[expert_idx]
            expert_mask = selected_experts == expert_idx
            expert_weights = (routing_weights * expert_mask).sum(dim=-1,
                                                                 keepdim=True)

            current_hidden_states = expert_layer(hidden_states).mul_(
                expert_weights)
            if final_hidden_states is None:
                final_hidden_states = current_hidden_states
            else:
                final_hidden_states.add_(current_hidden_states)

        return tensor_model_parallel_all_reduce(final_hidden_states).view(
            batch_size, sequence_length, hidden_dim)


class GrokAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 2,
        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

        if (linear_method is not None
                and not linear_method.quant_config.merge_weight()):
            self.merge_weight = False
            self.query = ColumnParallelLinear(
                hidden_size,
                self.q_size,
                bias=False,
                linear_method=linear_method,
            )
            self.key = ColumnParallelLinear(
                hidden_size,
                self.kv_size,
                bias=False,
                linear_method=linear_method,
            )
            self.value = ColumnParallelLinear(
                hidden_size,
                self.kv_size,
                bias=False,
                linear_method=linear_method,
            )
        else:
            self.merge_weight = True
            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.linear = 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=int(self.rope_theta),
            is_neox_style=True,
        )
        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,
    ) -> torch.Tensor:
        if self.merge_weight:
            qkv, _ = self.qkv_proj(hidden_states)
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size],
                                dim=-1)
        else:
            q, _ = self.query(hidden_states)
            k, _ = self.key(hidden_states)
            v, _ = self.value(hidden_states)
        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)
        # NOTE: This is a hack, figure out a better way to do this.
        attn_output = 30 * torch.tanh(attn_output / 30)
        output, _ = self.linear(attn_output)
        return output


class GrokDecoderLayer(nn.Module):

    def __init__(
        self,
        config: GrokConfig,
        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 = GrokAttention(
            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,
            sliding_window=config.sliding_window,
            linear_method=linear_method,
        )
        self.block_sparse_moe = GrokMoE(config=config,
                                        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,
        residual: Optional[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,
        )

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


class GrokModel(nn.Module):

    def __init__(
        self,
        config: GrokConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        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([
            GrokDecoderLayer(config, linear_method=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,
    ) -> 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], input_metadata,
                                            residual)
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class GrokForCausalLM(nn.Module):

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

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

    def sample(
        self,
        hidden_states: Optional[torch.Tensor],
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        if (self.linear_method is not None
                and not self.linear_method.quant_config.merge_weight()):
            next_tokens = self.quant_sampler(self.lm_head(hidden_states),
                                             sampling_metadata)
        else:
            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", "query", "q"),
            ("qkv_proj", "key", "k"),
            ("qkv_proj", "value", "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,
                fall_back_to_pt=False,
        ):
            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
                # Skip experts that are not assigned to this worker.
                if ("moe." in name 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)