mixtral.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 Mixtral model."""
from typing import Iterable, List, Optional, Tuple

import torch
from torch import nn
from transformers import MixtralConfig

from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig, LoRAConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
                                               ReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
    DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors, SamplerOutput
from vllm.utils import print_warning_once

from .interfaces import SupportsLoRA


class MixtralMoE(nn.Module):
    """A tensor-parallel MoE implementation for Mixtral that shards each expert
    across all ranks.

    Each expert's weights are sharded across all ranks and a fused MoE
    kernel is used for the forward pass, and finally we reduce the outputs
    across ranks.
    """

    def __init__(self,
                 num_experts: int,
                 top_k: int,
                 hidden_size: int,
                 intermediate_size: int,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 tp_size: Optional[int] = None):
        super().__init__()
        self.hidden_size = hidden_size

        # Gate always runs at half / full precision for now.
        self.gate = ReplicatedLinear(hidden_size,
                                     num_experts,
                                     bias=False,
                                     params_dtype=params_dtype,
                                     quant_config=None)

        self.experts = FusedMoE(num_experts=num_experts,
                                top_k=top_k,
                                hidden_size=hidden_size,
                                intermediate_size=intermediate_size,
                                params_dtype=params_dtype,
                                reduce_results=True,
                                renormalize=True,
                                quant_config=quant_config,
                                tp_size=tp_size)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_size = hidden_states.shape
        hidden_states = hidden_states.view(-1, self.hidden_size)
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = self.experts(hidden_states, router_logits)
        return final_hidden_states.view(num_tokens, hidden_size)


class MixtralAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )
        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 = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              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.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)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        output, _ = self.o_proj(attn_output)
        return output


class MixtralDecoderLayer(nn.Module):

    def __init__(
        self,
        config: MixtralConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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 = MixtralAttention(
            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,
            cache_config=cache_config,
            quant_config=quant_config)
        self.block_sparse_moe = MixtralMoE(
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            quant_config=quant_config)
        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: torch.Tensor,
        attn_metadata: AttentionMetadata,
        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,
            attn_metadata=attn_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 MixtralModel(nn.Module):

    def __init__(
        self,
        config: MixtralConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ) -> None:
        super().__init__()
        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,
            org_num_embeddings=config.vocab_size,
        )
        self.layers = nn.ModuleList([
            MixtralDecoderLayer(config,
                                cache_config,
                                quant_config=quant_config)
            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[torch.Tensor],
        attn_metadata: AttentionMetadata,
    ) -> 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], attn_metadata,
                                            residual)
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class MixtralForCausalLM(nn.Module, SupportsLoRA):
    fall_back_to_pt_during_load = False

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
    }

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

    def __init__(
        self,
        config: MixtralConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ) -> None:
        super().__init__()

        self.config = config
        self.lora_config = lora_config

        self.model = MixtralModel(config,
                                  cache_config,
                                  quant_config,
                                  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,
            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,
        )
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                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.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.weight, hidden_states,
                                       sampling_metadata)
        return logits

    def sample(
        self,
        logits: Optional[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)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]

        expert_params_mapping = [
            # These are the weight scales for the experts
            # (param_name, weight_name, expert_id, shard_id)
            ("experts.w13_scale"
             if weight_name in ["w1", "w3"] else "experts.w2_scale",
             f"experts.{expert_id}.{weight_name}.weight_scale", expert_id,
             shard_id) for expert_id in range(self.config.num_local_experts)
            for shard_id, weight_name in enumerate(["w1", "w2", "w3"])
        ] + [
            # These are the weights for the experts
            # (param_name, weight_name, expert_id)
            ("experts.w13_weight"
             if weight_name in ["w1", "w3"] else "experts.w2_weight",
             f"experts.{expert_id}.{weight_name}.weight", expert_id, shard_id)
            for expert_id in range(self.config.num_local_experts)
            for shard_id, weight_name in enumerate(["w1", "w2", "w3"])
        ] + [
            # These are the activation scales for the experts
            # (param_name, weight_name, expert_id)
            ("experts.a13_scale"
             if weight_name in ["w1", "w3"] else "experts.a2_scale",
             f"experts.{expert_id}.{weight_name}.input_scale", expert_id,
             shard_id) for expert_id in range(self.config.num_local_experts)
            for shard_id, weight_name in enumerate(["w1", "w2", "w3"])
        ]

        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            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:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    weight_loader(param,
                                  loaded_weight,
                                  weight_name,
                                  shard_id=shard_id,
                                  expert_id=expert_id)
                    break
                else:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    # Remapping the name of FP8 kv-scale.
                    if name.endswith("kv_scale"):
                        remapped_kv_scale_name = name.replace(
                            ".kv_scale", ".attn.kv_scale")
                        if remapped_kv_scale_name not in params_dict:
                            print_warning_once(
                                "Found kv scale in the checkpoint "
                                f"(e.g. {name}), but not found the expected "
                                f"name in the model "
                                f"(e.g. {remapped_kv_scale_name}). "
                                "kv-scale is not loaded.")
                            continue
                        else:
                            name = remapped_kv_scale_name
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)