vllm/vllm/model_executor/models/gpt_j.py

# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gptj/modeling_gptj.py
# Copyright 2023 The vLLM team.
# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
#
# 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 GPT-J model compatible with HuggingFace weights."""
from typing import Iterable, List, Optional, Set, Tuple, Union

import torch
from torch import nn
from transformers import GPTJConfig

from vllm.attention import Attention, AttentionMetadata
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_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors

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


class GPTJAttention(nn.Module):

    def __init__(
        self,
        config: GPTJConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.total_num_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.total_num_heads

        self.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_size,
            self.total_num_heads,
            bias=False,
            quant_config=quant_config,
        )
        self.out_proj = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
        )

        tp_world_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tp_world_size == 0
        self.num_heads = self.total_num_heads // tp_world_size

        scaling = self.head_size**-0.5
        assert getattr(config, "rotary", True)
        assert config.rotary_dim % 2 == 0
        rope_theta = getattr(config, "rope_theta", 10000)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.rotary_emb = get_rope(
            self.head_size,
            rotary_dim=config.rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            is_neox_style=False,
        )
        self.attn = Attention(self.num_heads,
                              self.head_size,
                              scaling,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")

    def forward(
        self,
        position_ids: 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.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        attn_output, _ = self.out_proj(attn_output)
        return attn_output


class GPTJMLP(nn.Module):

    def __init__(
        self,
        intermediate_size: int,
        config: GPTJConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        hidden_size = config.n_embd
        self.fc_in = ColumnParallelLinear(
            hidden_size,
            intermediate_size,
            quant_config=quant_config,
        )
        self.fc_out = RowParallelLinear(
            intermediate_size,
            hidden_size,
            quant_config=quant_config,
        )
        self.act = get_act_fn(config.activation_function)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.fc_in(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.fc_out(hidden_states)
        return hidden_states


class GPTJBlock(nn.Module):

    def __init__(
        self,
        config: GPTJConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        inner_dim = (4 * config.n_embd
                     if config.n_inner is None else config.n_inner)
        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = GPTJAttention(config,
                                  cache_config,
                                  quant_config,
                                  prefix=f"{prefix}.attn")
        self.mlp = GPTJMLP(inner_dim, config, quant_config)

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_output = self.attn(
            position_ids=position_ids,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            attn_metadata=attn_metadata,
        )
        mlp_output = self.mlp(hidden_states)
        hidden_states = attn_output + mlp_output + residual
        return hidden_states


@support_torch_compile
class GPTJModel(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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.embed_dim = config.n_embd
        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            self.embed_dim,
        )
        self.start_layer, self.end_layer, self.h = make_layers(
            config.n_layer,
            lambda prefix: GPTJBlock(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.h",
        )
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.n_embd))

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

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors],
        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)
        else:
            hidden_states = intermediate_tensors["hidden_states"]
        for i in range(self.start_layer, self.end_layer):
            layer = self.h[i]
            hidden_states = layer(
                position_ids,
                hidden_states,
                kv_caches[i - self.start_layer],
                attn_metadata,
            )
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        hidden_states = self.ln_f(hidden_states)
        return hidden_states


class GPTJForCausalLM(nn.Module, SupportsPP):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config
        assert not config.tie_word_embeddings
        self.transformer = GPTJModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(
                                         prefix, "transformer"))
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.n_embd,
            bias=True,
            quant_config=quant_config,
        )
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.sampler = get_sampler()
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.transformer(input_ids, positions, kv_caches,
                                         attn_metadata, 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.lm_head, hidden_states,
                                       sampling_metadata, self.lm_head.bias)
        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]]) -> Set[str]:
        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),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: Set[str] = set()
        for name, loaded_weight in weights:
            if "attn.bias" in name or "attn.masked_bias" in name:
                continue

            if (self.quant_config is not None and
                (scale_name := self.quant_config.get_cache_scale(name))):
                # Loading kv cache scales for quark and
                # compressed-tensors quantization
                param = params_dict[scale_name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
                                 loaded_weight[0])
                weight_loader(param, loaded_weight)
                loaded_params.add(scale_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:
                name = maybe_remap_kv_scale_name(name, params_dict)
                if name is None:
                    continue
                # 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