lightseekorg--tokenspeed/python/tokenspeed/runtime/models/qwen2.py

# Copyright (c) 2026 LightSeek Foundation
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.

"""Inference-only Qwen2 model compatible with HuggingFace weights."""

from __future__ import annotations

from collections.abc import Iterable
from typing import Any

import torch
from torch import nn

from tokenspeed.runtime.configs.qwen2_config import Qwen2Config
from tokenspeed.runtime.configs.utils import get_rope_theta
from tokenspeed.runtime.distributed.comm_ops import all_reduce
from tokenspeed.runtime.distributed.mapping import Mapping
from tokenspeed.runtime.execution.context import ForwardContext
from tokenspeed.runtime.layers.activation import SiluAndMul
from tokenspeed.runtime.layers.layernorm import RMSNorm
from tokenspeed.runtime.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
from tokenspeed.runtime.layers.paged_attention import PagedAttention
from tokenspeed.runtime.layers.quantization.base_config import QuantizationConfig
from tokenspeed.runtime.layers.rotary_embedding import get_rope
from tokenspeed.runtime.layers.utils import get_layer_id
from tokenspeed.runtime.layers.vocab_parallel_embedding import VocabParallelEmbedding
from tokenspeed.runtime.model_loader.weight_utils import (
    default_weight_loader,
    kv_cache_scales_loader,
)
from tokenspeed.runtime.models.base import BaseCausalLM
from tokenspeed.runtime.models.utils import validate_attention_partition
from tokenspeed.runtime.utils import add_prefix, make_layers
from tokenspeed.runtime.utils.env import global_server_args_dict


class Qwen2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: QuantizationConfig | None = None,
        tp_rank: int | None = None,
        tp_size: int | None = None,
        tp_group: tuple[int, ...] | None = None,
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            tp_rank=tp_rank,
            tp_size=tp_size,
            tp_group=tp_group,
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=False,
            tp_rank=tp_rank,
            tp_size=tp_size,
            tp_group=tp_group,
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. "
                "Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

    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 Qwen2Attention(nn.Module):
    def __init__(
        self,
        config: Qwen2Config,
        mapping: Mapping,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        layer_id: int = 0,
        rope_theta: float = 1000000,
        rope_scaling: dict[str, Any] | None = None,
        head_dim: int | None = None,
        max_position_embeddings: int = 32768,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.mapping = mapping
        self.hidden_size = hidden_size
        self.tp_rank = self.mapping.attn.tp_rank
        self.tp_size = self.mapping.attn.tp_size
        self.total_num_heads = num_heads
        self.total_num_kv_heads = num_kv_heads
        validate_attention_partition(
            self.total_num_heads,
            self.total_num_kv_heads,
            self.tp_size,
        )
        self.num_heads = self.total_num_heads // self.tp_size
        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
        self.head_dim = head_dim or 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
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        # Qwen2 uses biases on Q/K/V projections but not on the output projection.
        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=True,
            quant_config=quant_config,
            prefix=add_prefix("qkv_proj", prefix),
            tp_rank=self.mapping.attn.tp_rank,
            tp_size=self.mapping.attn.tp_size,
            tp_group=self.mapping.attn.tp_group,
        )
        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),
            reduce_results=False,
            tp_rank=self.mapping.attn.tp_rank,
            tp_size=self.mapping.attn.tp_size,
            tp_group=self.mapping.attn.tp_group,
        )

        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,
        )
        self.attn = PagedAttention(
            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,
        ctx: ForwardContext,
        out_cache_loc: torch.Tensor,
        cos_sin: tuple[torch.Tensor, torch.Tensor] | None = 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)
        attn_output = self.attn(q, k, v, ctx, out_cache_loc)
        if len(attn_output.size()) == 3:
            attn_output = attn_output.reshape(attn_output.shape[0], -1)
        output, _ = self.o_proj(attn_output)
        return output


class Qwen2DecoderLayer(nn.Module):
    def __init__(
        self,
        config: Qwen2Config,
        mapping: Mapping,
        layer_id: int = 0,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.mapping = mapping
        if self.mapping.attn.tp_size != self.mapping.dense.tp_size:
            raise ValueError(
                "Qwen2 does not use CommManager and assumes attn_tp_size == dense_tp_size"
            )
        self.hidden_size = config.hidden_size
        rope_theta = get_rope_theta(config, 1000000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
        head_dim = getattr(config, "head_dim", None)
        self.self_attn = Qwen2Attention(
            config=config,
            mapping=self.mapping,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            layer_id=layer_id,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            head_dim=head_dim,
            max_position_embeddings=max_position_embeddings,
            quant_config=quant_config,
            prefix=add_prefix("self_attn", prefix),
        )
        self.mlp = Qwen2MLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            tp_rank=self.mapping.dense.tp_rank,
            tp_size=self.mapping.dense.tp_size,
            tp_group=self.mapping.dense.tp_group,
        )
        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,
        ctx: ForwardContext,
        out_cache_loc: torch.Tensor,
        residual: torch.Tensor | None,
        cos_sin: tuple[torch.Tensor, torch.Tensor] | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        elif (
            ctx.input_num_tokens > global_server_args_dict["comm_fusion_max_num_tokens"]
        ):
            hidden_states = all_reduce(hidden_states, self.mapping.dense.tp_group)
            hidden_states, residual = self.input_layernorm(hidden_states, residual)
        else:
            hidden_states, residual, *_ = (
                self.input_layernorm.forward_with_allreduce_fusion(
                    self.mapping.dense.tp_rank,
                    self.mapping.dense.tp_group,
                    hidden_states,
                    residual,
                )
            )

        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            ctx=ctx,
            out_cache_loc=out_cache_loc,
            cos_sin=cos_sin,
        )

        if ctx.input_num_tokens > global_server_args_dict["comm_fusion_max_num_tokens"]:
            hidden_states = all_reduce(hidden_states, self.mapping.attn.tp_group)
            hidden_states, residual = self.post_attention_layernorm(
                hidden_states, residual
            )
        else:
            hidden_states, residual, *_ = (
                self.post_attention_layernorm.forward_with_allreduce_fusion(
                    self.mapping.attn.tp_rank,
                    self.mapping.attn.tp_group,
                    hidden_states,
                    residual,
                )
            )
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


class Qwen2Model(nn.Module):
    def __init__(
        self,
        config: Qwen2Config,
        mapping: Mapping,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        decoder_layer_type: type[nn.Module] = None,
    ) -> None:
        super().__init__()
        self.mapping = mapping
        self.config = config
        self.padding_idx = getattr(config, "pad_token_id", None)
        self.vocab_size = config.vocab_size
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            tp_rank=self.mapping.attn.tp_rank,
            tp_size=self.mapping.attn.tp_size,
            tp_group=self.mapping.attn.tp_group,
        )
        decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
        self.layers = make_layers(
            config.num_hidden_layers,
            lambda idx, prefix: decoder_layer_type(
                config=config,
                mapping=self.mapping,
                layer_id=idx,
                quant_config=quant_config,
            ),
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        if hasattr(self.config, "scale_emb"):
            return self.embed_tokens(input_ids) * self.config.scale_emb
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        ctx: ForwardContext,
        out_cache_loc: torch.Tensor,
        input_embeds: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, None]:
        if input_embeds is None:
            hidden_states = self.embed_tokens(input_ids)
        else:
            hidden_states = input_embeds
        residual = None

        for i in range(len(self.layers)):
            layer = self.layers[i]
            hidden_states, residual = layer(
                positions,
                hidden_states,
                ctx,
                out_cache_loc,
                residual,
                cos_sin=None,
            )
        if ctx.input_num_tokens > global_server_args_dict["comm_fusion_max_num_tokens"]:
            hidden_states = all_reduce(hidden_states, self.mapping.dense.tp_group)
            hidden_states, _ = self.norm(hidden_states, residual)
        else:
            hidden_states, *_ = self.norm.forward_with_allreduce_fusion(
                self.mapping.dense.tp_rank,
                self.mapping.dense.tp_group,
                hidden_states,
                residual,
            )
        return hidden_states, None

    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
        tp_size = self.mapping.attn.tp_size
        tp_rank = self.mapping.attn.tp_rank
        for layer_idx, scaling_factor in kv_cache_scales_loader(
            quantization_param_path,
            tp_rank,
            tp_size,
            self.config.num_hidden_layers,
            self.config.__class__.model_type,
        ):
            if not isinstance(self.layers[layer_idx], nn.Identity):
                layer_self_attn = self.layers[layer_idx].self_attn
            if hasattr(layer_self_attn.attn, "k_scale"):
                layer_self_attn.attn.k_scale = scaling_factor
                layer_self_attn.attn.v_scale = scaling_factor
            else:
                raise RuntimeError(
                    "Self attention has no KV cache scaling " "factor attribute!"
                )


class Qwen2ForCausalLM(BaseCausalLM):
    model_cls = Qwen2Model

    default_bitsandbytes_target_modules = [
        ".gate_proj.",
        ".down_proj.",
        ".up_proj.",
        ".q_proj.",
        ".k_proj.",
        ".v_proj.",
        ".o_proj.",
    ]
    bitsandbytes_stacked_params_mapping = {
        "q_proj": ("qkv_proj", 0),
        "k_proj": ("qkv_proj", 1),
        "v_proj": ("qkv_proj", 2),
        "gate_proj": ("gate_up_proj", 0),
        "up_proj": ("gate_up_proj", 1),
    }

    def __init__(
        self,
        config: Qwen2Config,
        mapping: Mapping,
        quant_config: QuantizationConfig | None = None,
    ) -> None:
        super().__init__(
            config=config,
            mapping=mapping,
            quant_config=quant_config,
        )

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

    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 "Embedding" in self.config.name_or_path:
                name = add_prefix(name, "model")
            layer_id = get_layer_id(name)
            if (
                layer_id is not None
                and hasattr(self.model, "start_layer")
                and (
                    layer_id < self.model.start_layer
                    or layer_id >= self.model.end_layer
                )
            ):
                continue

            if "rotary_emb.inv_freq" in name or "projector" in name:
                continue
            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
                continue
            if self.config.tie_word_embeddings and "lm_head.weight" in name:
                continue
            if name.startswith("model.vision_tower") and name not in params_dict:
                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)
                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:
                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)

    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_embed_and_head(self, embed, head):
        del self.model.embed_tokens.weight
        del self.lm_head.weight
        self.model.embed_tokens.weight = embed
        self.lm_head.weight = head
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
        self.model.load_kv_cache_scales(quantization_param_path)


EntryClass = Qwen2ForCausalLM