qwen3.py

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

# Copyright 2024 The Qwen team.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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 Qwen3 model compatible with HuggingFace weights."""
from collections.abc import Iterable
from typing import Any, Optional, Union

import torch
from torch import nn
from transformers import Qwen3Config

from vllm.attention import Attention, AttentionType
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .qwen2 import Qwen2MLP as Qwen3MLP
from .qwen2 import Qwen2Model

from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
                    maybe_prefix)
import vllm.envs as envs
from vllm.utils import direct_register_custom_op

logger = init_logger(__name__)


class Qwen3Attention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        max_position: int = 4096 * 32,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-06,
        qkv_bias: bool = False,
        rope_theta: float = 10000,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        rope_scaling: Optional[tuple] = None,
        prefix: str = "",
        attn_type: str = AttentionType.DECODER,
        dual_chunk_attention_config: Optional[dict[str, Any]] = 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 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.dual_chunk_attention_config = dual_chunk_attention_config

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

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position,
            base=self.rope_theta,
            rope_scaling=rope_scaling,
            dual_chunk_attention_config=dual_chunk_attention_config,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            attn_type=attn_type,
            **{
                "layer_idx": extract_layer_index(prefix),
                "dual_chunk_attention_config": dual_chunk_attention_config,
            } if dual_chunk_attention_config else {},
        )
        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
        self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
        
    def rms_rotary_embedding_fuse(
        positions: torch.Tensor,
        query: torch.Tensor,
        key: Optional[torch.Tensor],
        head_size: int,
        cos_sin_cache: torch.Tensor,
        is_neox_style: bool,
        q_weight: torch.Tensor,
        k_weight: torch.Tensor,
        q_bias: Optional[torch.Tensor],
        k_bias: Optional[torch.Tensor],
        epsilon: float,
    ) -> None:
        from lightop import rms_rotary_embedding_fuse as fused_kernel
        fused_kernel(
            positions,
            query,
            key,
            head_size,
            cos_sin_cache,
            is_neox_style,
            q_weight,
            k_weight,
            q_bias,
            k_bias,
            epsilon,
        )

    def rms_rotary_embedding_fuse_fake(
        positions: torch.Tensor,
        query: torch.Tensor,
        key: Optional[torch.Tensor],
        head_size: int,
        cos_sin_cache: torch.Tensor,
        is_neox_style: bool,
        q_weight: torch.Tensor,
        k_weight: torch.Tensor,
        q_bias: Optional[torch.Tensor],
        k_bias: Optional[torch.Tensor],
        epsilon: float,
    ) -> None:
        # Fake impl intentionally left as no-op for graph tracing modes.
        pass

    if not hasattr(torch.ops.vllm, "rms_rotary_embedding_fuse"):
        direct_register_custom_op(
            op_name="rms_rotary_embedding_fuse",
            op_func=rms_rotary_embedding_fuse,
            mutates_args=["query", "key"],
            fake_impl=rms_rotary_embedding_fuse_fake,
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> 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 envs.VLLM_USE_FUSED_RMS_ROPE:
            # Fused RMSNorm + RoPE path through custom op.
            cos_sin_cache = self.rotary_emb.cos_sin_cache
            if (cos_sin_cache.device != q.device
                    or cos_sin_cache.dtype != q.dtype):
                cos_sin_cache = cos_sin_cache.to(q.device,
                                                 dtype=q.dtype,
                                                 non_blocking=True)
                # Persist the converted cache so we don't re-copy/re-allocate
                # on every forward when the original buffer starts on CPU.
                self.rotary_emb.cos_sin_cache = cos_sin_cache
            q = q.contiguous()
            k = k.contiguous()
            torch.ops.vllm.rms_rotary_embedding_fuse(
                positions,
                q,
                k,
                self.head_dim,
                cos_sin_cache,
                self.rotary_emb.is_neox_style,
                self.q_norm.weight,
                self.k_norm.weight,
                None,
                None,
                self.q_norm.variance_epsilon,
            )
        else:
            # Add qk-norm
            q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
                            self.head_dim)
            if envs.VLLM_USE_APEX_RN:
                q_by_head = self.q_norm.forward_apex(q_by_head)
            else:
                q_by_head = self.q_norm.forward_cuda(q_by_head)
            q = q_by_head.view(q.shape)
            k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
                            self.head_dim)
            if envs.VLLM_USE_APEX_RN:
                k_by_head = self.k_norm.forward_apex(k_by_head)
            else:
                k_by_head = self.k_norm.forward_cuda(k_by_head)
            k = k_by_head.view(k.shape)
            q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output


class Qwen3DecoderLayer(nn.Module):

    def __init__(
        self,
        config: Qwen3Config,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        # Requires transformers > 4.32.0
        rope_theta = getattr(config, "rope_theta", 1000000)
        rope_scaling = getattr(config, "rope_scaling", None)
        dual_chunk_attention_config = getattr(config,
                                              "dual_chunk_attention_config",
                                              None)

        # By default, Qwen3 uses causal attention as it is a decoder-only model.
        # You can override the HF config with `is_causal=False` to enable
        # bidirectional attention, which is used in some embedding models
        # (e.g. Alibaba-NLP/gte-Qwen3-7B-instruct)
        if getattr(config, "is_causal", True):
            attn_type = AttentionType.DECODER
        else:
            attn_type = AttentionType.ENCODER_ONLY

        self.self_attn = Qwen3Attention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            max_position=config.max_position_embeddings,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=getattr(config, 'attention_bias', False),
            head_dim=getattr(config, 'head_dim', None),
            cache_config=cache_config,
            quant_config=quant_config,
            rope_scaling=rope_scaling,
            prefix=f"{prefix}.self_attn",
            attn_type=attn_type,
            dual_chunk_attention_config=dual_chunk_attention_config,
        )
        self.mlp = Qwen3MLP(
            hidden_size=self.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
        )
        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,
        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,
        )

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


ALL_DECODER_LAYER_TYPES = {
    "attention": Qwen3DecoderLayer,
}


@support_torch_compile(
    dynamic_arg_dims={
        "input_ids": 0,
        # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
        # otherwise (seq_len, ).
        "positions": -1,
        "intermediate_tensors": 0,
        "inputs_embeds": 0,
    })
class Qwen3Model(Qwen2Model):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config,
                         prefix=prefix,
                         decoder_layer_type=Qwen3DecoderLayer)


class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.lora_config = lora_config

        self.quant_config = quant_config
        self.model = Qwen3Model(vllm_config=vllm_config,
                                prefix=maybe_prefix(prefix, "model"))

        if get_pp_group().is_last_rank:
            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=maybe_prefix(
                                                  prefix, "lm_head"))
        else:
            self.lm_head = PPMissingLayer()

        self.logits_processor = LogitsProcessor(config.vocab_size)

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        self.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        num_layers = len(self.model.layers)
        return (2, num_layers // 2, num_layers - 3)

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."]
                           if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(weights)