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

# 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.
# ==============================================================================
"""Inference-only LoopCoder model compatible with HuggingFace weights."""

import logging
from typing import Iterable, Optional, Tuple

import torch
from torch import nn
from transformers import PretrainedConfig

from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.linear import (
    ColumnParallelLinear,
    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.models.llama import LlamaMLP as LoopCoderMLP
from sglang.srt.runtime_context import get_parallel
from sglang.srt.utils import add_prefix, make_layers
from sglang.srt.utils.hf_transformers_utils import get_rope_config

logger = logging.getLogger(__name__)


class LoopGateProjection(nn.Module):
    """Gate projection for mixed attention in Loop 2+.

    Computes: g = sigmoid(linear(Q)) for each head independently.
    This gate determines how much to use Loop1's KV (global) vs current loop's KV (local).

    Supports tensor parallelism: each GPU handles a subset of heads.
    The weight matrix has shape [num_heads, head_dim] and is split along the head dimension.
    """

    def __init__(
        self,
        total_num_heads: int,
        head_dim: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.total_num_heads = total_num_heads
        self.head_dim = head_dim
        tp_size = get_parallel().tp_size
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size

        self.gate_proj = ColumnParallelLinear(
            head_dim,
            self.total_num_heads,
            bias=True,
            gather_output=False,
            quant_config=quant_config,
            prefix=add_prefix("gate_proj", prefix),
        )

    def forward(self, query: torch.Tensor) -> torch.Tensor:
        """Compute gate values from query tensor.

        Args:
            query: [num_heads, num_tokens, head_dim]
                where num_heads is the number of heads on this TP rank
                and num_tokens = batch * seq_len

        Returns:
            gate: [num_tokens, num_heads * head_dim] (flattened format matching q shape)
        """
        num_heads, num_tokens, head_dim = query.shape

        assert (
            num_heads == self.num_heads
        ), f"Expected {self.num_heads} heads, got {num_heads}"

        query_flat = query.reshape(-1, head_dim)

        gate_logits_flat, _ = self.gate_proj(query_flat)

        gate_logits = gate_logits_flat.reshape(num_heads, num_tokens, self.num_heads)

        # Extract diagonal: each head h's query should use output column h
        gate_logits = torch.diagonal(gate_logits, dim1=0, dim2=2)
        gate_logits = gate_logits.transpose(0, 1)
        gate_logits = gate_logits.unsqueeze(-1)

        # Apply sigmoid
        gate = torch.sigmoid(gate_logits)

        # Expand and reshape to match q shape: [num_tokens, num_heads * head_dim]
        gate = gate.transpose(0, 1)
        gate = gate.expand(-1, -1, head_dim)
        gate = gate.reshape(num_tokens, num_heads * head_dim)

        return gate


class LoopCoderAttention(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        layer_id: int = 0,
        max_position: int = 4096 * 32,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.layer_id = layer_id
        self.hidden_size = hidden_size
        tp_size = get_parallel().tp_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:
            assert self.total_num_kv_heads % tp_size == 0
        else:
            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 = 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

        # Get loop_num from config, default to 2 if not specified
        self.loop_num = getattr(config, "loop_num", 2)
        self.loop_window_size = getattr(config, "loop_window_size", 64)

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
        )
        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=add_prefix("o_proj", prefix),
        )

        rope_theta, rope_scaling = get_rope_config(config)
        max_position_embeddings = getattr(
            config, "max_position_embeddings", max_position
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
        )

        # Create attention instances for each loop
        # Loop 0: global attention without sliding window for full context
        # Loop 1+: local attention with sliding window for recent tokens
        # Each loop needs a unique layer_id to avoid KV cache conflicts
        self.attn = nn.ModuleList()
        total_layers = getattr(config, "num_hidden_layers", 24)
        for loop_idx in range(self.loop_num):
            sliding_window = -1 if loop_idx == 0 else self.loop_window_size
            # Use unique layer_id for each loop: loop_idx * total_layers + layer_id
            # This ensures each loop has its own KV cache space
            unique_layer_id = loop_idx * total_layers + layer_id

            self.attn.append(
                RadixAttention(
                    self.num_heads,
                    self.head_dim,
                    self.scaling,
                    num_kv_heads=self.num_kv_heads,
                    layer_id=unique_layer_id,  # Unique layer_id for each loop
                    sliding_window_size=sliding_window,
                    quant_config=quant_config,
                    prefix=add_prefix(f"attn.{loop_idx}", prefix),
                )
            )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        loop_idx: int,
        gate_proj: Optional[LoopGateProjection] = None,
    ) -> 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)

        if loop_idx == 0:
            # First loop: standard global attention, save KV to cache
            attn_output = self.attn[0](q, k, v, forward_batch)
        else:
            # Loop 2+: mixed attention with learned gating
            # Global attention: read from Loop 0's KV cache without updating (save_kv_cache=False)
            # This provides full context information
            # Pass k=None, v=None to read from KV cache instead of recomputing
            global_attn_output = self.attn[0](
                q, None, None, forward_batch, save_kv_cache=False
            )

            # Local attention: use current loop's KV with sliding window
            # This focuses on recent tokens within the window
            local_attn_output = self.attn[loop_idx](q, k, v, forward_batch)

            # Compute gating weights using query-dependent projection
            assert gate_proj is not None, "gate_proj must be provided for loop_idx > 0"
            num_tokens = q.shape[0]
            q_reshaped = q.view(num_tokens, self.num_heads, self.head_dim).transpose(
                0, 1
            )
            gate = gate_proj(q_reshaped)

            # Mix global and local attention outputs with learned gate
            # gate controls the balance between global context and local focus
            attn_output = global_attn_output * gate + local_attn_output * (1 - gate)

        output, _ = self.o_proj(attn_output)
        return output


class LoopCoderDecoderLayer(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        layer_id: int = 0,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        self.layer_id = layer_id

        self.self_attn = LoopCoderAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            layer_id=layer_id,
            max_position=getattr(config, "max_position_embeddings", 4096 * 32),
            quant_config=quant_config,
            prefix=add_prefix("self_attn", prefix),
        )
        self.mlp = LoopCoderMLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=add_prefix("mlp", prefix),
        )
        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,
        forward_batch: ForwardBatch,
        loop_idx: int,
        gate_proj: Optional[LoopGateProjection] = None,
    ) -> 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,
            loop_idx=loop_idx,
            gate_proj=gate_proj,
        )
        hidden_states = hidden_states + residual

        # MLP
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = hidden_states + residual

        return hidden_states


class IQuestLoopCoderModel(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.vocab_size = config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=add_prefix("embed_tokens", prefix),
        )

        self.loop_num = getattr(self.config, "loop_num", 2)
        self.window_size = getattr(self.config, "loop_window_size", 64)

        # Gate projections for Loop 2+ (one per layer)
        head_dim = config.hidden_size // config.num_attention_heads
        gate_projections = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: LoopGateProjection(
                total_num_heads=config.num_attention_heads,
                head_dim=head_dim,
                quant_config=quant_config,
                prefix=prefix,
            ),
            prefix=add_prefix("gate_projections", prefix),
        )
        if isinstance(gate_projections, tuple):
            self.start_layer, self.end_layer, self.gate_projections = gate_projections
        else:
            self.start_layer, self.end_layer = 0, config.num_hidden_layers
            self.gate_projections = gate_projections

        layers = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: LoopCoderDecoderLayer(
                config=config,
                layer_id=idx,
                quant_config=quant_config,
                prefix=prefix,
            ),
            prefix=add_prefix("layers", prefix),
        )
        if isinstance(layers, tuple):
            self.start_layer, self.end_layer, self.layers = layers
        else:
            self.start_layer, self.end_layer = 0, config.num_hidden_layers
            self.layers = layers

        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: torch.Tensor = None,
    ) -> torch.Tensor:
        if input_embeds is not None:
            hidden_states = input_embeds
        else:
            hidden_states = self.embed_tokens(input_ids)

        # Multi-loop forward pass
        for loop_idx in range(self.loop_num):
            for layer_idx in range(self.start_layer, self.end_layer):
                layer = self.layers[layer_idx]
                # Get gate_proj for this layer (only for loop_idx > 0)
                gate_proj = self.gate_projections[layer_idx] if loop_idx > 0 else None
                hidden_states = layer(
                    positions, hidden_states, forward_batch, loop_idx, gate_proj
                )

        hidden_states = self.norm(hidden_states)
        return hidden_states


class IQuestLoopCoderForCausalLM(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config

        self.model = IQuestLoopCoderModel(
            config=config,
            quant_config=quant_config,
            prefix=add_prefix("model", prefix),
        )

        if config.tie_word_embeddings:
            self.lm_head = self.model.embed_tokens
        else:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                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,
    ):
        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 = [
            ("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

            # Handle gate_projections weights
            if name.startswith("gate_projections."):
                if name.endswith(".weight"):
                    sglang_name = name.replace(".weight", ".gate_proj.weight")
                elif name.endswith(".bias"):
                    sglang_name = name.replace(".bias", ".gate_proj.bias")
                else:
                    continue

                if sglang_name in params_dict:
                    param = params_dict[sglang_name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)
                continue

            # Handle stacked parameters
            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 name.endswith(".bias") and name not in params_dict:
                    continue
                if name in params_dict:
                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Handle regular parameters
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if name in params_dict:
                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)


# Entry class for model registration
EntryClass = IQuestLoopCoderForCausalLM