internlm2.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from collections.abc import Iterable
from functools import partial
from itertools import islice
from typing import Any, Optional, Union

import torch
from torch import nn
from transformers import PretrainedConfig

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              split_tensor_along_last_dim,
                              tensor_model_parallel_all_gather)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.pooler import DispatchPooler, Pooler
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    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

from .interfaces import SupportsLoRA, SupportsPP
from .interfaces_base import default_pooling_type
from .utils import (is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)


class InternLM2MLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.w2 = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.w2",
        )
        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.w2(x)
        return x


class InternLM2Attention(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,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.total_num_heads = num_heads
        assert self.total_num_heads % self.tp_size == 0
        self.num_heads = self.total_num_heads // self.tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= self.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 % self.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 self.tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // self.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.key_value_groups = int(self.num_heads / self.num_kv_heads)
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        self.wqkv = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wqkv",
        )
        self.wo = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wo",
        )

        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,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

    def split_qkv(self, qkv: torch.Tensor):
        seq_len = qkv.shape[0]
        if self.tp_size > 1:
            qkv_map = [self.q_size, self.kv_size, self.kv_size] * self.tp_size
            qkv = tensor_model_parallel_all_gather(qkv)
            qkv = torch.split(qkv, qkv_map, dim=-1)
            qkv = qkv[::3] + qkv[1::3] + qkv[2::3]
            qkv = torch.cat(qkv, dim=-1)

        qkv = qkv.view(seq_len, self.total_num_kv_heads,
                       self.key_value_groups + 2, self.head_dim)
        q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=-2)
        q = q.reshape(seq_len, self.q_size * self.tp_size)
        k = k.reshape(seq_len, self.kv_size * self.tp_size)
        v = v.reshape(seq_len, self.kv_size * self.tp_size)

        if self.tp_size > 1:
            splitter = partial(split_tensor_along_last_dim,
                               num_partitions=self.tp_size)
            q = splitter(q)[self.tp_rank]
            k = splitter(k)[self.tp_rank]
            v = splitter(v)[self.tp_rank]
        return q, k, v

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.wqkv(hidden_states)
        q, k, v = self.split_qkv(qkv)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.wo(attn_output)
        return output


class InternLMDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> 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)
        self.attention = InternLM2Attention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attention",
        )
        self.feed_forward = InternLM2MLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.feed_forward",
        )
        self.attention_norm = RMSNorm(config.hidden_size,
                                      eps=config.rms_norm_eps)
        self.ffn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.attention_norm(hidden_states)
        else:
            hidden_states, residual = self.attention_norm(
                hidden_states, residual)
        hidden_states = self.attention(
            positions=positions,
            hidden_states=hidden_states,
        )

        # Fully Connected
        hidden_states, residual = self.ffn_norm(hidden_states, residual)
        hidden_states = self.feed_forward(hidden_states)
        return hidden_states, residual


@support_torch_compile
class InternLM2Model(nn.Module):

    def __init__(
            self,
            *,
            vllm_config: VllmConfig,
            prefix: str = "",
            layer_type: type[InternLMDecoderLayer] = InternLMDecoderLayer):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.config = config
        self.vocab_size = config.vocab_size
        self.tok_embeddings = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: layer_type(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers")
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states, residual = layer(positions, hidden_states, residual)
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class InternLM2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
    packed_modules_mapping = {
        "wqkv": ["wqkv"],
        "gate_up_proj": ["w1", "w3"],
    }

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 model_type: type[InternLM2Model] = InternLM2Model):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.quant_config = quant_config
        self.lora_config = lora_config

        self.model = model_type(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "model"))
        self.output = ParallelLMHead(config.vocab_size,
                                     config.hidden_size,
                                     quant_config=quant_config,
                                     prefix=maybe_prefix(prefix, "output"))
        if self.config.tie_word_embeddings:
            self.output.weight = self.model.tok_embeddings.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.output, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "w1", 0),
            ("gate_up_proj", "w3", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        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
                if is_pp_missing_parameter(name, self):
                    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
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


@default_pooling_type("ALL")
class InternLM2ForRewardModel(InternLM2ForCausalLM):

    is_pooling_model = True

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        model_type: type[InternLM2Model] = InternLM2Model,
    ):
        super().__init__(vllm_config=vllm_config,
                         prefix=prefix,
                         model_type=model_type)

        for attr in ("output", "logits_processor"):
            delattr(self, attr)

        config = vllm_config.model_config.hf_config
        self.head_dtype = vllm_config.model_config.head_dtype

        self.v_head = RowParallelLinear(config.hidden_size,
                                        1,
                                        bias=False,
                                        input_is_parallel=False,
                                        params_dtype=self.head_dtype,
                                        prefix=maybe_prefix(prefix, "v_head"),
                                        return_bias=False)

        pooler_config = vllm_config.model_config.pooler_config
        assert pooler_config is not None

        self.pooler = DispatchPooler(
            {"encode": Pooler.for_encode(pooler_config)}, )

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        hidden_states = hidden_states.to(self.head_dtype)
        logits = self.v_head(hidden_states)
        return logits