gemma.py

# Adapted from:
# https://github.com/vllm-project/vllm/blob/d65fac2738f0287a41955b45df76a2d5a919bff6/vllm/model_executor/models/gemma.py
"""Inference-only Gemma model compatible with HuggingFace weights."""
from typing import Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.config import LoRAConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
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.vocab_parallel_embedding import VocabParallelEmbedding

from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.managers.router.model_runner import InputMetadata
from sglang.srt.weight_utils import default_weight_loader, hf_model_weights_iterator


class GemmaMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )
        self.act_fn = GeluAndMul()

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class GemmaAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        head_dim: int,
        layer_id: int = 0,
        max_position_embeddings: int = 8192,
        rope_theta: float = 10000,
        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 = head_dim
        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_embeddings,
            base=self.rope_theta,
            is_neox_style=True,
        )
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            layer_id=layer_id,
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        input_metadata: InputMetadata,
    ) -> 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, input_metadata)
        output, _ = self.o_proj(attn_output)
        return output


class GemmaDecoderLayer(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = GemmaAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            head_dim=config.head_dim,
            layer_id=layer_id,
            max_position_embeddings=config.max_position_embeddings,
            rope_theta=config.rope_theta,
            quant_config=quant_config,
        )
        self.mlp = GemmaMLP(
            hidden_size=self.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,
        input_metadata: InputMetadata,
        residual: Optional[torch.Tensor],
    ) -> Tuple[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,
            input_metadata=input_metadata,
        )

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


class GemmaModel(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ) -> None:
        super().__init__()
        self.config = config

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.layers = nn.ModuleList(
            [
                GemmaDecoderLayer(config, i, quant_config=quant_config)
                for i 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,
        input_metadata: InputMetadata,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds

        # Normalize the embedding by sqrt(hidden_size)
        hidden_states *= self.config.hidden_size**0.5

        residual = None
        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                input_metadata,
                residual,
            )
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


class GemmaForCausalLM(nn.Module):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    # LoRA specific attributes
    supported_lora_modules = [
        "qkv_proj",
        "o_proj",
        "gate_up_proj",
        "down_proj",
    ]
    # Gemma does not apply LoRA to the embedding layer.
    embedding_modules = {}
    embedding_padding_modules = []

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
    ) -> None:
        del lora_config  # Unused.
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.model = GemmaModel(config, quant_config=quant_config)
        self.logits_processor = LogitsProcessor(config)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        input_metadata: InputMetadata,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
        return self.logits_processor(
            input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
        )

    def load_weights(
        self,
        model_name_or_path: str,
        cache_dir: Optional[str] = None,
        load_format: str = "auto",
        revision: Optional[str] = None,
    ):
        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()
        for name, loaded_weight in hf_model_weights_iterator(
            model_name_or_path, cache_dir, load_format, revision
        ):
            for param_name, shard_name, shard_id in stacked_params_mapping:
                if shard_name not in name:
                    continue
                name = name.replace(shard_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:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # GemmaRMSNorm is different from Llama's in that it multiplies
                # (1 + weight) to the output, instead of just weight.
                if "norm.weight" in name:
                    loaded_weight += 1.0
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        unloaded_params = params_dict.keys() - loaded_params
        if unloaded_params:
            raise RuntimeError(
                "Some weights are not initialized from checkpoints: "
                f"{unloaded_params}"
            )


EntryClass = GemmaForCausalLM