phi.py

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

# Adapted from
# https://huggingface.co/microsoft/phi-1_5/blob/main/modeling_phi.py
# Copyright 2023 The vLLM team.
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
#
# BSD 3-Clause License
#
# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# * Redistributions of source code must retain the above copyright notice, this
#   list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above copyright notice,
#   this list of conditions and the following disclaimer in the documentation
#   and/or other materials provided with the distribution.
#
# * Neither the name of the copyright holder nor the names of its
#   contributors may be used to endorse or promote products derived from
#   this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""Inference-only Phi-1.5 model compatible with HuggingFace weights."""

from collections.abc import Iterable
from itertools import islice

import torch
from torch import nn
from transformers import PhiConfig

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_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.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.sequence import IntermediateTensors

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


class PhiAttention(nn.Module):
    def __init__(
        self,
        config: PhiConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = 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

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

        # pylint: disable=C0103
        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_size,
            self.total_num_heads,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.dense = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.dense",
        )

        scaling = self.head_size**-0.5
        rotary_dim = int(
            config.partial_rotary_factor
            * (config.hidden_size // config.num_attention_heads)
        )
        assert rotary_dim % 2 == 0

        # pylint: disable=C0301
        # Refer to:
        # https://huggingface.co/microsoft/phi-1_5/blob/d212a789620c380ff32ca1d1ee9943a777360987/modeling_phi.py#L518
        rope_theta = getattr(config, "rope_theta", 10000.0)
        max_position_embeddings = getattr(config, "max_position_embeddings", 2048)
        self.rotary_emb = get_rope(
            self.head_size,
            rotary_dim=rotary_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
        )
        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,
    ) -> 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)
        output, _ = self.dense(attn_output)
        return output


class PhiMLP(nn.Module):
    def __init__(
        self,
        config: PhiConfig,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()

        n_inner = getattr(config, "n_inner", None)
        n_inner = n_inner if n_inner is not None else 4 * config.hidden_size

        self.fc1 = ColumnParallelLinear(
            config.hidden_size,
            n_inner,
            quant_config=quant_config,
            prefix=f"{prefix}.fc1",
        )
        self.fc2 = RowParallelLinear(
            n_inner,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.fc2",
        )
        self.act = get_act_fn(config.hidden_act)

    def forward(self, hidden_states):
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)
        return hidden_states


class PhiLayer(nn.Module):
    def __init__(
        self,
        config: PhiConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.input_layernorm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        self.self_attn = PhiAttention(
            config, cache_config, quant_config, prefix=f"{prefix}.self_attn"
        )
        self.mlp = PhiMLP(config, quant_config, prefix=f"{prefix}.mlp")

    def forward(
        self,
        position_ids: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        attn_outputs = self.self_attn(
            position_ids=position_ids,
            hidden_states=hidden_states,
        )
        feed_forward_hidden_states = self.mlp(hidden_states)
        hidden_states = attn_outputs + feed_forward_hidden_states + residual
        return hidden_states


@support_torch_compile
class PhiModel(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.quant_config = quant_config
        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: PhiLayer(config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers",
        )
        self.final_layernorm = nn.LayerNorm(
            config.hidden_size, eps=config.layer_norm_eps
        )
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states"], config.hidden_size
        )

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states = layer(positions, hidden_states)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})

        hidden_states = self.final_layernorm(hidden_states)

        return hidden_states

    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"),
        ]
        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
                # pylint: disable=E1136

                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


class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ]
    }

    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
        # lm_head use bias, cannot share word embeddings
        assert not config.tie_word_embeddings

        self.quant_config = quant_config

        self.model = PhiModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )

        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            bias=True,
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

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

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

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        logits = self.logits_processor(self.lm_head, hidden_states, self.lm_head.bias)
        return logits

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)