sgl-project--sglang/python/sglang/srt/models/stablelm.py

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

# Copyright 2023-2024 SGLang Team
# 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.
# ==============================================================================

# Adapted from:
# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/stablelm.py#L1
"""
Inference-only StableLM-2 (https://huggingface.co/stabilityai/stablelm-2-1_6b)
model compatible with HuggingFace weights.
"""

from typing import Iterable, Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.layers.rotary_embedding import get_rope
from sglang.srt.layers.vocab_parallel_embedding import (
    ParallelLMHead,
    VocabParallelEmbedding,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.runtime_context import get_parallel
from sglang.srt.utils import add_prefix, is_npu

_is_npu = is_npu()


class StablelmMLP(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_up_proj = MergedColumnParallelLinear(
            config.hidden_size,
            [config.intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
        )
        self.down_proj = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
        )
        self.act_fn = SiluAndMul()

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


class StablelmAttention(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        tp_size = get_parallel().tp_size
        self.total_num_heads = config.num_attention_heads
        self.num_heads = self.total_num_heads // tp_size

        self.total_num_key_value_heads = config.num_key_value_heads
        if self.total_num_key_value_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_key_value_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_key_value_heads == 0
        self.num_key_value_heads = max(1, self.total_num_key_value_heads // tp_size)
        self.head_dim = self.hidden_size // self.total_num_heads
        self.max_position_embeddings = config.max_position_embeddings
        rope_pct = getattr(
            config, "rope_pct", getattr(config, "partial_rotary_factor", 1)
        )
        self.rotary_ndims = int(self.head_dim * rope_pct)
        self.scaling = self.head_dim**-0.5
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_key_value_heads * self.head_dim
        self.qkv_bias = getattr(config, "use_qkv_bias", False)
        if (self.head_dim * self.num_heads * tp_size) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads "
                f"(got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )

        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_key_value_heads,
            self.qkv_bias,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("o_proj", prefix),
        )
        if not _is_npu:
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.rotary_ndims,
                max_position=self.config.max_position_embeddings,
                base=self.config.rope_parameters["rope_theta"],
            )
        else:
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.rotary_ndims,
                max_position=self.config.max_position_embeddings,
                base=self.config.rope_parameters["rope_theta"],
                dtype=torch.float32,
            )
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_key_value_heads,
            layer_id=layer_id,
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        if not _is_npu:
            q, k = self.rotary_emb(positions, q, k)
        else:
            odtype = q.dtype
            q, k = self.rotary_emb(positions, q.to(torch.float32), k.to(torch.float32))
            q, k = q.to(odtype), k.to(odtype)
        attn_output = self.attn(q, k, v, forward_batch)
        output, _ = self.o_proj(attn_output)
        return output


class StablelmDecoderLayer(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.self_attn = StablelmAttention(
            config, layer_id=layer_id, prefix=add_prefix("self_attn", prefix)
        )
        self.mlp = StablelmMLP(
            config, quant_config=quant_config, prefix=add_prefix("mlp", prefix)
        )
        norm_eps = getattr(config, "norm_eps", getattr(config, "layer_norm_eps", 1e-05))
        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=norm_eps)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> Tuple[torch.Tensor, 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,
            forward_batch=forward_batch,
        )
        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, residual


class StableLMEpochModel(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            prefix=add_prefix("embed_tokens", prefix),
        )
        self.layers = nn.ModuleList(
            [
                StablelmDecoderLayer(
                    config,
                    i,
                    quant_config=quant_config,
                    prefix=add_prefix(f"layers.{i}", prefix),
                )
                for i in range(config.num_hidden_layers)
            ]
        )
        norm_eps = getattr(config, "norm_eps", getattr(config, "layer_norm_eps", 1e-05))
        self.norm = nn.LayerNorm(config.hidden_size, eps=norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds
        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                forward_batch,
            )
        hidden_states = self.norm(hidden_states)
        return hidden_states


class StableLmForCausalLM(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.model = StableLMEpochModel(
            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size, config.hidden_size, prefix=add_prefix("lm_head", prefix)
        )
        self.logits_processor = LogitsProcessor(config)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: torch.Tensor = None,
    ) -> torch.Tensor:
        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
        return self.logits_processor(
            input_ids, hidden_states, self.lm_head, forward_batch
        )

    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"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                # Models trained using ColossalAI may include these tensors in
                # the checkpoint. Skip them.
                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
                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
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)


EntryClass = StableLmForCausalLM