aphrodite-engine/aphrodite/modeling/models/llama.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 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 LLaMA model compatible with HuggingFace weights."""

from typing import Any, Dict, List, Optional, Tuple

import torch
from torch import nn
from transformers import LlamaConfig

from aphrodite.attention import Attention, AttentionMetadata
from aphrodite.common.config import LoRAConfig
from aphrodite.common.sequence import SamplerOutput
from aphrodite.common.utils import is_hip
from aphrodite.modeling.hf_downloader import (default_weight_loader,
                                              hf_model_weights_iterator,
                                              kv_cache_scales_loader)
from aphrodite.modeling.layers.activation import SiluAndMul
from aphrodite.modeling.layers.layernorm import RMSNorm
from aphrodite.modeling.layers.linear import (ColumnParallelLinear,
                                              LinearMethodBase,
                                              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 (
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
from aphrodite.distributed import (get_tensor_model_parallel_rank,
                                   get_tensor_model_parallel_world_size)
from aphrodite.modeling.sampling_metadata import SamplingMetadata


class LlamaMLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> 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 LlamaAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        linear_method: Optional[LinearMethodBase] = None,
        bias: bool = False,
        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.max_position_embeddings = max_position_embeddings

        # This will be overwritten by model initialization if we are using it.
        # N.B. currently we only support per tensor scalar scaling factors
        # & only applicable to ROCm (AMD GPU).
        # The scaling factor convention we are assuming is
        # quantized_value * scaling_factor ~= true_value
        # which is consistent with the practice of setting
        # scaling_factor = tensor_amax / FPtype_max
        self.kv_scale = 1.0

        if (linear_method is not None
                and not linear_method.quant_config.merge_weight()):
            self.merge_weight = False
            self.q_proj = ColumnParallelLinear(hidden_size,
                                               self.q_size,
                                               bias=bias,
                                               linear_method=linear_method)
            self.k_proj = ColumnParallelLinear(
                hidden_size,
                self.kv_size,
                bias=bias,
                linear_method=linear_method,
            )
            self.v_proj = ColumnParallelLinear(
                hidden_size,
                self.kv_size,
                bias=bias,
                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=bias,
                linear_method=linear_method,
            )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=bias,
            linear_method=linear_method,
        )

        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,
            is_neox_style=True,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            sliding_window=sliding_window,
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
        # kv_quant_param: List[float],
    ) -> 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.q_proj(hidden_states)
            k, _ = self.k_proj(hidden_states)
            v, _ = self.v_proj(hidden_states)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata,
                                self.kv_scale)
        output, _ = self.o_proj(attn_output)
        return output


class LlamaDecoderLayer(nn.Module):

    def __init__(
        self,
        config: LlamaConfig,
        linear_method: Optional[LinearMethodBase] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        sliding_window = getattr(config, "sliding_window", None)
        # Support abacusai/Smaug-72B-v0.1 with attention_bias
        # Support internlm/internlm-7b with bias
        attention_bias = getattr(config, "attention_bias", False) or getattr(
            config, "bias", False)
        self.self_attn = LlamaAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=getattr(config, "num_key_value_heads",
                                 config.num_attention_heads),
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            linear_method=linear_method,
            bias=attention_bias,
            sliding_window=sliding_window,
        )
        self.mlp = LlamaMLP(
            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)
        if config.model_type == "Yi":
            # Some old Yi finetunes and quants have not been llama-fied
            self.ln1 = self.input_layernorm
            self.ln2 = self.post_attention_layernorm

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
        residual: Optional[torch.Tensor],
        # kv_quant_param: List[float],
    ) -> 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,
            # kv_quant_param=kv_quant_param,
        )

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


class LlamaModel(nn.Module):

    def __init__(
        self,
        config: LlamaConfig,
        linear_method: Optional[LinearMethodBase] = None,
        lora_config: Optional[LoRAConfig] = None,
    ) -> None:
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        lora_vocab = ((lora_config.lora_extra_vocab_size *
                       (lora_config.max_loras or 1)) if lora_config else 0)
        self.vocab_size = config.vocab_size + lora_vocab
        self.org_vocab_size = config.vocab_size
        self.embed_tokens = VocabParallelEmbedding(
            self.vocab_size,
            config.hidden_size,
            linear_method=linear_method,
            org_num_embeddings=config.vocab_size,
        )
        self.layers = nn.ModuleList([
            LlamaDecoderLayer(config, linear_method)
            for _ in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(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,
                # attn_metadata.kv_quant_params[i]
                # if attn_metadata.kv_quant_params is not None else None,
            )
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class LlamaForCausalLM(nn.Module):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    # LoRA specific attributes
    supported_lora_modules = [
        "qkv_proj",
        "o_proj",
        "gate_up_proj",
        "down_proj",
        "embed_tokens",
        "lm_head",
    ]
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    def __init__(
        self,
        config: LlamaConfig,
        linear_method: Optional[LinearMethodBase] = None,
        lora_config: Optional[LoRAConfig] = None,
    ) -> None:
        super().__init__()
        self.config = config
        self.linear_method = linear_method
        self.model = LlamaModel(config, linear_method, lora_config=lora_config)
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            linear_method=linear_method,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config else lora_config.lora_vocab_padding_size,
        )
        logit_scale = getattr(config, "logit_scale", 1.0)
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size, logit_scale)
        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)
            ("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),
        ]
        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 ("rotary_emb.cos_cached" in name
                    or "rotary_emb.sin_cached" in name):
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                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)

    # If this function is called, it should always initialize KV cache scale
    # factors (or else raise an exception). Thus, handled exceptions should
    # make sure to leave KV cache scale factors in a known good (dummy) state
    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
        tp_size = get_tensor_model_parallel_world_size()
        tp_rank = get_tensor_model_parallel_rank()
        for layer_idx, scaling_factor in kv_cache_scales_loader(
                quantization_param_path, tp_rank, tp_size,
                self.config.num_hidden_layers,
                self.config.__class__.model_type):
            layer_self_attn = self.model.layers[layer_idx].self_attn

            if is_hip():
                # The scaling factor convention we are assuming is
                # quantized_value * scaling_factor ~= true_value
                # which is consistent with the practice of setting
                # scaling_factor = tensor_amax / FPtype_max
                scaling_factor *= 2
            if hasattr(layer_self_attn, "kv_scale"):
                layer_self_attn.kv_scale = scaling_factor
            else:
                raise RuntimeError("Self attention has no KV cache scaling "
                                   "factor attribute!")