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

# coding=utf-8
# Adapted from
# https://huggingface.co/Qwen/Qwen-7B/blob/main/modeling_qwen.py
# Copyright (c) Alibaba Cloud.
# LICENSE: https://huggingface.co/Qwen/Qwen-7B/blob/main/LICENSE
"""Inference-only QWen model compatible with HuggingFace weights."""
from typing import Any, Dict, Iterable, List, Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from aphrodite.attention import Attention, AttentionMetadata
from aphrodite.common.config import CacheConfig
from aphrodite.common.sequence import IntermediateTensors, SamplerOutput
from aphrodite.common.utils import progress_bar
from aphrodite.distributed import get_tensor_model_parallel_world_size
from aphrodite.modeling.layers.activation import SiluAndMul
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (MergedColumnParallelLinear,
                                              QKVParallelLinear,
                                              RowParallelLinear)
from aphrodite.modeling.layers.logits_processor import LogitsProcessor
from aphrodite.modeling.layers.rotary_embedding import get_rope
from aphrodite.modeling.layers.sampler import Sampler
from aphrodite.modeling.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
from aphrodite.modeling.model_loader.weight_utils import default_weight_loader
from aphrodite.modeling.sampling_metadata import SamplingMetadata
from aphrodite.quantization.base_config import QuantizationConfig


class QWenMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str = "silu",
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config)
        self.c_proj = RowParallelLinear(intermediate_size,
                                        hidden_size,
                                        bias=False,
                                        quant_config=quant_config)
        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.c_proj(x)
        return x


class QWenAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        max_position_embeddings: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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.c_attn = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            bias=True,
            quant_config=quant_config,
        )
        self.c_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )
        self.scaling = self.head_dim**-0.5

        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,
        )
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              cache_config=cache_config,
                              quant_config=quant_config)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        qkv, _ = self.c_attn(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.c_proj(attn_output)
        return output


class QWenBlock(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        self.attn = QWenAttention(config.hidden_size,
                                  config.num_attention_heads,
                                  config.max_position_embeddings,
                                  rope_theta=rope_theta,
                                  rope_scaling=rope_scaling,
                                  cache_config=cache_config,
                                  quant_config=quant_config)

        self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        self.mlp = QWenMLP(config.hidden_size,
                           config.intermediate_size // 2,
                           quant_config=quant_config)

    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.ln_1(hidden_states)
        else:
            hidden_states, residual = self.ln_1(hidden_states, residual)
        hidden_states = self.attn(
            positions=positions,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            attn_metadata=attn_metadata,
        )

        # Fully Connected
        hidden_states, residual = self.ln_2(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


class QWenModel(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.config = config
        self.vocab_size = config.vocab_size

        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.h = nn.ModuleList([
            QWenBlock(config, cache_config, quant_config)
            for _ in range(config.num_hidden_layers)
        ])
        self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

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


class QWenLMHeadModel(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.transformer = QWenModel(config, cache_config, quant_config)
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.sampler = Sampler()

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> torch.Tensor:
        hidden_states = self.transformer(input_ids, positions, kv_caches,
                                         attn_metadata)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[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, weights: Iterable[Tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "w2", 0),
            ("gate_up_proj", "w1", 1),
        ]
        params_dict = dict(self.named_parameters())
        weights_list = list(weights)
        for name, loaded_weight in progress_bar(weights_list,
                                                desc="Loading modules..."):
            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)