sgl-project--sglang/python/sglang/srt/models/olmo2.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/main/vllm/model_executor/models/olmo2.py
"""Inference-only OLMo2 model compatible with HuggingFace weights."""

from functools import partial
from typing import Iterable, Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from sglang.srt.distributed import (
    split_tensor_along_last_dim,
    tensor_model_parallel_all_gather,
)
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.layernorm import RMSNorm
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_executor.runner import get_is_capture_mode
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.runtime_context import get_parallel, get_stream
from sglang.srt.utils import add_prefix, is_cuda, make_layers

_is_cuda = is_cuda()


# Aligned with HF's implementation, using sliding window inclusive with the last token
# SGLang assumes exclusive
def get_attention_sliding_window_size(config):
    return config.sliding_window - 1 if hasattr(config, "sliding_window") else None


class Olmo2Attention(nn.Module):
    """
    This is the attention block where the output is computed as
    Attention(LN(x)) in MLP(LN(x + Attention(LN(x))))
    (plus another skip connection).
    """

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.tp_size = get_parallel().tp_size
        self.total_num_heads = config.num_attention_heads

        assert self.hidden_size % self.total_num_heads == 0
        assert self.total_num_heads % self.tp_size == 0

        self.num_heads = self.total_num_heads // self.tp_size
        self.total_num_kv_heads = self.config.num_key_value_heads

        if self.total_num_kv_heads >= self.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 % self.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 self.tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // self.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.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_parameters["rope_theta"]

        # Attention input projection. Projects x -> (q, k, v)
        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=config.attention_bias,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
        )
        self.tp_rank = get_parallel().tp_rank
        self.alt_stream = alt_stream

        self.k_norm = RMSNorm(
            self.total_num_kv_heads * self.head_dim,
            eps=self.config.rms_norm_eps,
        )
        self.q_norm = RMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps)

        sliding_window = None
        if (
            layer_types := getattr(self.config, "layer_types", None)
        ) is not None and layer_types[layer_id] == "sliding_attention":
            sliding_window = get_attention_sliding_window_size(self.config)

        # Rotary embeddings. Rope scaling is only applied on full attention
        # layers.
        self.rope_scaling = (
            self.config.rope_scaling
            if sliding_window is None
            else {"rope_type": "default"}
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=self.rope_scaling,
        )
        self.scaling = self.head_dim**-0.5
        self.attn = RadixAttention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            layer_id=layer_id,
            sliding_window_size=sliding_window,
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
        )

        # Attention output projection.
        self.o_proj = RowParallelLinear(
            self.head_dim * self.total_num_heads,
            self.hidden_size,
            bias=config.attention_bias,
            quant_config=quant_config,
            prefix=add_prefix("o_proj", prefix),
        )

    def _apply_qk_norm(
        self, q: torch.Tensor, k: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        if self.tp_size > 1:
            q = tensor_model_parallel_all_gather(q.contiguous())
            k = tensor_model_parallel_all_gather(k.contiguous())

        if self.alt_stream is not None and get_is_capture_mode():
            current_stream = torch.cuda.current_stream()
            self.alt_stream.wait_stream(current_stream)

            q_shape = q.shape
            k_shape = k.shape

            q_by_last = q.reshape(-1, q_shape[-1])
            q_by_last = self.q_norm(q_by_last)

            with torch.cuda.stream(self.alt_stream):
                k_by_last = k.reshape(-1, k_shape[-1])
                k_by_last = self.k_norm(k_by_last)

            current_stream.wait_stream(self.alt_stream)

            q = q_by_last.view(q_shape)
            k = k_by_last.view(k_shape)
        else:
            q = self.q_norm.forward_native(q)
            k = self.k_norm.forward_native(k)

        if self.tp_size > 1:
            splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size)
            q = splitter(q)[self.tp_rank]
            k = splitter(k)[self.tp_rank]
        return q, k

    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)
        q, k = self._apply_qk_norm(q, k)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, forward_batch)
        output, _ = self.o_proj(attn_output)
        return output


class Olmo2MLP(nn.Module):
    """
    This is the MLP block where the output is computed as
    MLP(x) in LN(MLP(x + LN(Attention(x))))
    (plus another skip connection).
    """

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size

        # Feed-forward input projection.
        self.gate_up_proj = MergedColumnParallelLinear(
            self.hidden_size,
            [self.intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
        )

        # Activation function.
        self.act_fn = SiluAndMul()

        # Feed-forward output projection.
        self.down_proj = RowParallelLinear(
            self.intermediate_size,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
        )

    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 Olmo2DecoderLayer(nn.Module):
    """
    This is a typical transformer block where the output is
    computed as MLP(LN(x + Attention(LN(x))))
    (plus another skip connection).
    """

    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ):
        super().__init__()
        self.layer_id = layer_id
        self.alt_stream = alt_stream
        # Attention block.
        self.self_attn = Olmo2Attention(
            config,
            layer_id,
            quant_config,
            prefix=add_prefix("self_attn", prefix),
            alt_stream=alt_stream,
        )

        # MLP block.
        self.mlp = Olmo2MLP(config, quant_config, prefix=add_prefix("mlp", prefix))

        # RMSNorm
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

        self.post_feedforward_layernorm = RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        # Attention block.
        residual = hidden_states
        hidden_states = self.self_attn(positions, hidden_states, forward_batch)
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = hidden_states + residual

        # MLP block.
        residual = hidden_states
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states


class Olmo2Model(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ):
        super().__init__()
        self.config = config
        if alt_stream is None and _is_cuda:
            alt_stream = get_stream("alt")
        self.alt_stream = alt_stream

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            prefix=add_prefix("embed_tokens", prefix),
        )
        self.layers = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: Olmo2DecoderLayer(
                config=config,
                layer_id=idx,
                quant_config=quant_config,
                prefix=prefix,
                alt_stream=self.alt_stream,
            ),
            prefix=add_prefix("layers", prefix),
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """
        :param input_ids: A tensor of shape `(batch_size, seq_len)`.
        """
        # Get embeddings of input.
        # shape: (batch_size, seq_len, d_model)

        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds

        # Apply blocks one-by-one.
        for layer_id, decoder_layer in enumerate(self.layers):
            # shape: (batch_size, seq_len, d_model)
            hidden_states = decoder_layer(
                positions,
                hidden_states,
                forward_batch,
            )

        # Apply final layer norm.
        # shape: (batch_size, seq_len or 1, d_model)
        hidden_states = self.norm(hidden_states)
        return hidden_states


class Olmo2ForCausalLM(nn.Module):
    """
    Extremely barebones HF model wrapper.
    """

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        alt_stream: Optional[torch.cuda.Stream] = None,
    ):
        super().__init__()
        self.config = config
        self.model = Olmo2Model(
            config,
            quant_config,
            prefix=add_prefix("model", prefix),
            alt_stream=alt_stream,
        )
        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,
                quant_config=quant_config,
                prefix=add_prefix("lm_head", prefix),
            )
        self.logits_processor = LogitsProcessor(config)

    def get_attention_sliding_window_size(self):
        return get_attention_sliding_window_size(self.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=input_ids,
            positions=positions,
            forward_batch=forward_batch,
            input_embeds=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(remove_duplicate=False))
        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
            # With tie_word_embeddings, we can skip lm_head.weight
            # The weight might appear unnecessarily in the files if the model is
            # processed with quantization, LoRA, fine-tuning, etc.
            if self.config.tie_word_embeddings and "lm_head.weight" 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
                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 = Olmo2ForCausalLM