vllm.model_executor.models.qwen3_next ¶

class Qwen3NextAttention(nn.Module):

    def __init__(
        self,
        config: Qwen3NextConfig,
        model_config: Optional[ModelConfig] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = config.num_key_value_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 = config.head_dim or (self.hidden_size // self.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.dual_chunk_attention_config = getattr(
            config, "dual_chunk_attention_config", None)
        self.attn_output_gate = getattr(config, "attn_output_gate", True)

        self.qkv_proj = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            self.total_num_heads * (1 + self.attn_output_gate),
            self.total_num_kv_heads,
            bias=getattr(config, "qkv_bias", False),
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

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

        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            rotary_dim=self.head_dim,
            max_position=config.max_position_embeddings,
            base=config.rope_theta,
            rope_scaling=config.rope_scaling,
            partial_rotary_factor=config.partial_rotary_factor,
            dual_chunk_attention_config=self.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":
                self.dual_chunk_attention_config,
            } if self.dual_chunk_attention_config else {},
        )

        self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        output: torch.Tensor,
        hidden_states: torch.Tensor,
    ):
        qkv, _ = self.qkv_proj(hidden_states)

        if self.attn_output_gate:
            q_gate, k, v = qkv.split(
                [self.q_size * 2, self.kv_size, self.kv_size], dim=-1)
            orig_shape = q_gate.shape[:-1]
            q_gate = q_gate.view(*orig_shape, self.num_heads, -1)
            q, gate = torch.chunk(q_gate, 2, dim=-1)
            q = q.reshape(*orig_shape, -1)
            gate = gate.reshape(*orig_shape, -1)
        else:
            q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size],
                                dim=-1)

        q = self.q_norm(q.view(-1, self.num_heads, self.head_dim)).view(
            -1, self.num_heads * self.head_dim)
        k = self.k_norm(k.view(-1, self.num_kv_heads, self.head_dim)).view(
            -1, self.num_kv_heads * self.head_dim)

        q, k = self.rotary_emb(positions, q, k)

        attn_output = self.attn(q, k, v)

        if self.attn_output_gate:
            gate = torch.sigmoid(gate)
            attn_output = attn_output * gate

        output[:], _ = self.o_proj(attn_output)

class Qwen3NextDecoderLayer(nn.Module):

    def __init__(
        self,
        vllm_config: VllmConfig,
        layer_type: str,
        prefix: str = "",
    ) -> None:
        super().__init__()

        config = vllm_config.model_config.hf_config
        model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        speculative_config = vllm_config.speculative_config

        self.layer_type = layer_type
        self.layer_idx = extract_layer_index(prefix)

        if self.layer_type == "linear_attention":
            self.linear_attn = Qwen3NextGatedDeltaNet(
                config,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
                speculative_config=speculative_config,
                prefix=f'{prefix}.linear_attn')
        elif self.layer_type == "full_attention":
            self.self_attn = Qwen3NextAttention(
                config,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=f'{prefix}.self_attn',
            )
        else:
            raise ValueError(f"Invalid layer_type {self.layer_type}")

        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
                           config.mlp_only_layers)
        if (self.layer_idx not in mlp_only_layers) and (
                config.num_experts > 0 and
            (self.layer_idx + 1) % config.decoder_sparse_step == 0):
            self.mlp = Qwen3NextSparseMoeBlock(
                vllm_config=vllm_config,
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = Qwen3NextMLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
            )

        self.input_layernorm = Qwen3NextRMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen3NextRMSNorm(
            config.hidden_size, eps=config.rms_norm_eps)

        self.layer_scale = getattr(config, "layer_scale", False)
        if self.layer_scale:
            self.attn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
                    config.hidden_size,
                    dtype=config.torch_dtype,
                ), )
            self.ffn_layer_scale = torch.nn.Parameter(
                torch.zeros(
                    1,
                    1,
                    config.hidden_size,
                    dtype=config.torch_dtype,
                ), )

    def forward(
        self,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        positions: torch.Tensor = None,
        **kwargs: object,
    ):
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)

        self_attention_output = torch.empty_like(hidden_states)
        if self.layer_type == "linear_attention":
            self.linear_attn(
                hidden_states=hidden_states,
                output=self_attention_output,
            )
        elif self.layer_type == "full_attention":
            self.self_attn(
                hidden_states=hidden_states,
                output=self_attention_output,
                positions=positions,
            )
        else:
            raise ValueError("Invalid layer_type")
        hidden_states = self_attention_output

        if self.layer_scale:
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states * (
                    self.attn_layer_scale.to(hidden_states.dtype)[0] + 1)
            else:
                hidden_states = hidden_states * (
                    self.attn_layer_scale.to(hidden_states.dtype) + 1)

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

        if self.layer_scale:
            if len(hidden_states.shape) == 2:
                hidden_states = hidden_states * (
                    self.ffn_layer_scale.to(hidden_states.dtype)[0] + 1)
            else:
                assert len(hidden_states.shape) == len(
                    self.ffn_layer_scale.shape
                ), f'shape must be the same {len(hidden_states.shape)}, {len(self.ffn_layer_scale.shape)}'  # noqa: E501
                hidden_states = hidden_states * (
                    self.ffn_layer_scale.to(hidden_states.dtype) + 1)

        return hidden_states, residual

class Qwen3NextForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,
                           MixtureOfExperts, IsHybrid):
    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 = ""):
        config = vllm_config.model_config.hf_config
        self.vllm_config = vllm_config
        self.model_config = vllm_config.model_config
        cache_config = vllm_config.cache_config
        lora_config = vllm_config.lora_config
        scheduler_config = vllm_config.scheduler_config
        assert not cache_config.enable_prefix_caching, \
            "Qwen3Next currently does not support prefix caching"
        self.quant_config = vllm_config.quant_config

        super().__init__()
        self.config = config
        self.scheduler_config = scheduler_config
        self.model = Qwen3NextModel(vllm_config=vllm_config,
                                    prefix=maybe_prefix(prefix, "model"))
        self.unpadded_vocab_size = config.vocab_size
        if lora_config:
            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
        self.lm_head = ParallelLMHead(
            self.unpadded_vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            padding_size=DEFAULT_VOCAB_PADDING_SIZE
            # We need bigger padding if using lora for kernel
            # compatibility
            if not lora_config else lora_config.lora_vocab_padding_size,
            prefix=maybe_prefix(prefix, "lm_head"))
        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

        # Set MoE hyperparameters
        self.expert_weights = []

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

            assert isinstance(layer, Qwen3NextDecoderLayer)
            if isinstance(layer.mlp, Qwen3NextSparseMoeBlock):
                example_layer = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_layer is None:
            raise RuntimeError("No Qwen3Next 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 set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    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, Qwen3NextSparseMoeBlock):
                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 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,
        **kwargs: object,
    ):
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)

        return hidden_states

    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
            vllm_config.model_config.dtype,
            vllm_config.cache_config.mamba_cache_dtype)

    @classmethod
    def get_mamba_state_shape_from_config(
            cls, vllm_config: "VllmConfig"
    ) -> tuple[tuple[int, int], tuple[int, int]]:
        parallel_config = vllm_config.parallel_config
        hf_config = vllm_config.model_config.hf_config
        tp_size = parallel_config.tensor_parallel_size
        num_spec = (vllm_config.speculative_config.num_speculative_tokens
                    if vllm_config.speculative_config else 0)
        return MambaStateShapeCalculator.gated_delta_net_state_shape(
            tp_size, hf_config.linear_num_key_heads,
            hf_config.linear_num_value_heads, hf_config.linear_key_head_dim,
            hf_config.linear_value_head_dim, hf_config.linear_conv_kernel_dim,
            num_spec)

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

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=["mtp."],
        )
        return loader.load_weights(weights)

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()