starcoder2.py

# coding=utf-8
# Copyright 2024 BigCode 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.
""" PyTorch Starcoder2 model."""
from typing import Iterable, List, Optional, Tuple, Union

import torch
from torch import nn
from transformers import Starcoder2Config

from vllm.attention import Attention, AttentionMetadata
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig
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 Sampler, SamplerOutput
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

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


class Starcoder2Attention(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__()
        self.config = config

        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_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 = self.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 = config.rope_theta
        self.max_position_embeddings = config.max_position_embeddings
        self.use_bias = config.use_bias

        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=self.use_bias,
            quant_config=quant_config,
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            self.hidden_size,
            bias=self.use_bias,
            quant_config=quant_config,
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            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 Starcoder2MLP(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__()
        self.c_fc = ColumnParallelLinear(
            config.hidden_size,
            config.intermediate_size,
            bias=config.use_bias,
            quant_config=quant_config,
        )
        self.c_proj = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            bias=config.use_bias,
            quant_config=quant_config,
        )
        self.act = get_act_fn(config.hidden_act, quant_config,
                              config.intermediate_size)

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


class Starcoder2DecoderLayer(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = Starcoder2Attention(config,
                                             cache_config,
                                             quant_config=quant_config)
        self.mlp = Starcoder2MLP(config, quant_config=quant_config)
        self.input_layernorm = nn.LayerNorm(config.hidden_size,
                                            eps=config.norm_epsilon)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
                                                     eps=config.norm_epsilon)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        # Self Attention
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            kv_cache=kv_cache,
            attn_metadata=attn_metadata,
        )
        hidden_states = residual + hidden_states

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

        return hidden_states


@support_torch_compile
class Starcoder2Model(nn.Module):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        # TODO: consider padding_idx (currently removed)
        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                                   config.hidden_size)
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Starcoder2DecoderLayer(
                config, cache_config, quant_config=quant_config),
            prefix=f"{prefix}.layers",
        )
        self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.hidden_size))

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors],
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            hidden_states = self.embed_tokens(input_ids)
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
        for i in range(self.start_layer, self.end_layer):
            layer = self.layers[i]
            hidden_states = layer(positions, 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.norm(hidden_states)
        return hidden_states


class Starcoder2ForCausalLM(nn.Module, SupportsPP):

    def __init__(self,
                 config: Starcoder2Config,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__()
        self.config = config
        self.model = Starcoder2Model(config,
                                     cache_config,
                                     quant_config=quant_config)
        self.vocab_size = config.vocab_size
        self.unpadded_vocab_size = config.vocab_size
        if config.tie_word_embeddings:
            self.lm_head = self.model.embed_tokens
        else:
            self.unpadded_vocab_size = config.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,
                quant_config=quant_config,
            )
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.sampler = Sampler()
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

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

        params_dict = dict(self.named_parameters(remove_duplicate=False))
        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)
                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:
                if self.config.tie_word_embeddings and "lm_head.weight" in name:
                    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)