vllm-project--vllm/vllm/model_executor/models/qwen3_moe.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 Qwen3MoE model compatible with HuggingFace weights."""

from collections.abc import Iterable
from itertools import islice
from typing import Any

import torch
import torch.nn.functional as F
from torch import nn

from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
    get_ep_group,
    get_pp_group,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_gather,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    ReplicatedLinear,
    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,
    VocabParallelEmbedding,
)
from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.sequence import IntermediateTensors

from .interfaces import (
    EagleModelMixin,
    MixtureOfExperts,
    SupportsEagle,
    SupportsEagle3,
    SupportsLoRA,
    SupportsPP,
    SupportsQuant,
)
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    WeightsMapper,
    extract_layer_index,
    make_empty_intermediate_tensors_factory,
    make_layers,
    maybe_prefix,
)

logger = init_logger(__name__)


class Qwen3MoeMLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: QuantizationConfig | None = None,
        reduce_results: bool = True,
        expert_gate: torch.nn.Linear | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()
        self.expert_gate = expert_gate

    def forward(self, x):
        gate_up, _ = self.gate_up_proj(x)
        out = self.act_fn(gate_up)
        out, _ = self.down_proj(out)

        if self.expert_gate is not None:
            out = F.sigmoid(self.expert_gate(x)[0]) * out

        return out


class Qwen3MoeSparseMoeBlock(nn.Module):
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str = "",
    ):
        super().__init__()

        config = vllm_config.model_config.hf_text_config
        parallel_config = vllm_config.parallel_config
        quant_config = vllm_config.quant_config

        self.tp_size = get_tensor_model_parallel_world_size()

        self.ep_group = get_ep_group().device_group
        self.ep_rank = get_ep_group().rank_in_group
        self.ep_size = self.ep_group.size()
        self.n_routed_experts = config.num_experts

        self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe

        if self.tp_size > config.num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.num_experts}."
            )

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        eplb_config = vllm_config.parallel_config.eplb_config
        self.enable_eplb = parallel_config.enable_eplb

        self.n_logical_experts = self.n_routed_experts
        self.n_redundant_experts = eplb_config.num_redundant_experts
        self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
        self.physical_expert_end = (
            self.physical_expert_start + self.n_local_physical_experts
        )

        self.gate = ReplicatedLinear(
            config.hidden_size,
            config.num_experts,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate",
        )

        shared_expert_intermediate_size = getattr(
            config, "shared_expert_intermediate_size", 0
        )
        if shared_expert_intermediate_size > 0:
            self.shared_expert_gate = ReplicatedLinear(
                config.hidden_size,
                1,
                bias=False,
                quant_config=None,
                prefix=f"{prefix}.shared_expert_gate",
            )
            self.shared_expert = Qwen3MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=shared_expert_intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                reduce_results=False,
                expert_gate=self.shared_expert_gate,
                prefix=f"{prefix}.shared_expert",
            )
        else:
            self.shared_expert_gate = None
            self.shared_expert = None

        self.experts = FusedMoE(
            shared_experts=self.shared_expert,
            gate=self.gate,
            num_experts=self.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            prefix=f"{prefix}.experts",
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts,
            is_sequence_parallel=self.is_sequence_parallel,
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        assert hidden_states.dim() <= 2, (
            "Qwen3MoeSparseMoeBlock only supports 1D or 2D inputs"
        )
        is_input_1d = hidden_states.dim() == 1
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

        if self.is_sequence_parallel:
            hidden_states = sequence_parallel_chunk(hidden_states)

        if self.experts.is_internal_router:
            # In this case, the gate/router runs inside the FusedMoE class
            final_hidden_states = self.experts(
                hidden_states=hidden_states, router_logits=hidden_states
            )
        else:
            # Actually this will be dead code, since we always pass gate into
            # FusedMoE in the current implementation. But we keep this code
            # here for clarity and future flexibility.
            router_logits, _ = self.gate(hidden_states)
            final_hidden_states = self.experts(
                hidden_states=hidden_states, router_logits=router_logits
            )

        if self.is_sequence_parallel:
            final_hidden_states = tensor_model_parallel_all_gather(
                final_hidden_states, 0
            )
            final_hidden_states = final_hidden_states[:num_tokens]

        # return to 1d if input is 1d
        return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states


class Qwen3MoeAttention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_parameters: dict[str, Any],
        max_position_embeddings: int = 8192,
        head_dim: int | None = None,
        rms_norm_eps: float = 1e-06,
        qkv_bias: bool = False,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        dual_chunk_attention_config: dict[str, Any] | None = 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.max_position_embeddings = max_position_embeddings
        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,
            max_position=max_position_embeddings,
            rope_parameters=rope_parameters,
            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",
            **{
                "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 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)
        # Add qk-norm
        q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
        q_by_head = self.q_norm(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)
        k_by_head = self.k_norm(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 Qwen3MoeDecoderLayer(nn.Module):
    def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()

        config = vllm_config.model_config.hf_text_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.hidden_size = config.hidden_size
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        dual_chunk_attention_config = getattr(
            config, "dual_chunk_attention_config", None
        )
        self.self_attn = Qwen3MoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_parameters=config.rope_parameters,
            max_position_embeddings=max_position_embeddings,
            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,
            prefix=f"{prefix}.self_attn",
            dual_chunk_attention_config=dual_chunk_attention_config,
        )

        # `mlp_only_layers` in the config.
        layer_idx = extract_layer_index(prefix)
        mlp_only_layers = (
            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
        )
        if (layer_idx not in mlp_only_layers) and (
            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
        ):
            self.mlp = Qwen3MoeSparseMoeBlock(
                vllm_config=vllm_config, prefix=f"{prefix}.mlp"
            )
        else:
            self.mlp = Qwen3MoeMLP(
                hidden_size=config.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: torch.Tensor | None,
    ) -> 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


@support_torch_compile
class Qwen3MoeModel(nn.Module, EagleModelMixin):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_stacked={
            # weight_name: (param_name, shard_id)
            ".q_proj": (".qkv_proj", "q"),
            ".k_proj": (".qkv_proj", "k"),
            ".v_proj": (".qkv_proj", "v"),
            # .experts.gate_up_proj must be handled by MoERunner.load_weights for EP
            ".mlp.gate_proj": (".mlp.gate_up_proj", 0),
            ".mlp.up_proj": (".mlp.gate_up_proj", 1),
            ".shared_expert.gate_proj": (".shared_expert.gate_up_proj", 0),
            ".shared_expert.up_proj": (".shared_expert.gate_up_proj", 1),
        }
    )

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        decoder_layer_type: type[torch.nn.Module] = Qwen3MoeDecoderLayer,
    ):
        super().__init__()

        config = vllm_config.model_config.hf_text_config
        quant_config = vllm_config.quant_config
        parallel_config = vllm_config.parallel_config
        eplb_config = parallel_config.eplb_config
        self.num_redundant_experts = eplb_config.num_redundant_experts

        self.vocab_size = config.vocab_size
        self.config = config
        self.quant_config = quant_config
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.embed_tokens",
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: decoder_layer_type(vllm_config=vllm_config, prefix=prefix),
            prefix=f"{prefix}.layers",
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        aux_hidden_states = self._maybe_add_hidden_state(
            [], self.start_layer, hidden_states, residual
        )
        for layer_idx, layer in enumerate(
            islice(self.layers, self.start_layer, self.end_layer),
            start=self.start_layer,
        ):
            hidden_states, residual = layer(positions, hidden_states, residual)
            self._maybe_add_hidden_state(
                aux_hidden_states, layer_idx + 1, hidden_states, residual
            )

        if not get_pp_group().is_last_rank:
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
        hidden_states, _ = self.norm(hidden_states, residual)

        # Return auxiliary hidden states if collected
        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            ignore_unexpected_suffixes=[
                ".bias",
                "_bias",
                ".weight_scale",
                "_weight_scale",
                ".input_scale",
                "_input_scale",
            ],
        )
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)


class Qwen3MoeForCausalLM(
    nn.Module,
    SupportsPP,
    SupportsLoRA,
    SupportsEagle,
    SupportsEagle3,
    MixtureOfExperts,
    SupportsQuant,
):
    hf_to_vllm_mapper = Qwen3MoeModel.hf_to_vllm_mapper
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ]
    }

    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }

    fall_back_to_pt_during_load = False

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_text_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config
        # Only perform the following mapping when Qwen3MoeMLP exists
        if getattr(config, "mlp_only_layers", []):
            self.packed_modules_mapping["gate_up_proj"] = ["gate_proj", "up_proj"]
        self.model = Qwen3MoeModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

        # Set MoE hyperparameters

        self.moe_layers = []
        example_layer = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue

            assert isinstance(layer, Qwen3MoeDecoderLayer)
            if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock):
                example_layer = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_layer is None:
            raise RuntimeError("No Qwen3MoE layer found in the model.layers.")

        self.num_moe_layers = len(self.moe_layers)
        self.num_expert_groups = 1
        self.num_shared_experts = 0
        self.num_logical_experts = example_layer.n_logical_experts
        self.num_physical_experts = example_layer.n_physical_experts
        self.num_local_physical_experts = example_layer.n_local_physical_experts
        self.num_routed_experts = example_layer.n_routed_experts
        self.num_redundant_experts = example_layer.n_redundant_experts

    def update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = num_physical_experts - self.num_logical_experts
        for layer in self.model.layers:
            if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock):
                moe = layer.mlp
                moe.n_local_physical_experts = num_local_physical_experts
                moe.n_physical_experts = num_physical_experts
                moe.n_redundant_experts = self.num_redundant_experts
                moe.experts.update_expert_map()

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

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> 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,
    ) -> torch.Tensor | None:
        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)
        return loader.load_weights(weights)