vllm.model_executor.models.qwen3_5 ¶

@MULTIMODAL_REGISTRY.register_processor(
    Qwen3VLMultiModalProcessor,
    info=Qwen3_5ProcessingInfo,
    dummy_inputs=Qwen3VLDummyInputsBuilder,
)
class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        # protocols have not __init__ method, so we need to use nn.Module.__init__
        nn.Module.__init__(self)
        config: Qwen3_5Config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
        self.video_pruning_rate = multimodal_config.video_pruning_rate
        self.is_multimodal_pruning_enabled = (
            multimodal_config.is_multimodal_pruning_enabled()
        )

        with self._mark_tower_model(vllm_config, {"image", "video"}):
            self.visual = Qwen3_VisionTransformer(
                config.vision_config,
                norm_eps=getattr(config, "rms_norm_eps", 1e-6),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "visual"),
            )

        with self._mark_language_model(vllm_config):
            self.language_model = Qwen3_5ForCausalLM(
                vllm_config=vllm_config, prefix=maybe_prefix(prefix, "language_model")
            )

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors
        )

    def embed_input_ids(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: MultiModalEmbeddings | None = None,
        *,
        is_multimodal: torch.Tensor | None = None,
        handle_oov_mm_token: bool = False,
    ) -> torch.Tensor:
        inputs_embeds = self._embed_text_input_ids(
            input_ids,
            self.language_model.embed_input_ids,
            is_multimodal=is_multimodal,
            handle_oov_mm_token=handle_oov_mm_token,
        )

        if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
            return inputs_embeds

        is_multimodal = _require_is_multimodal(is_multimodal)

        inputs_embeds = _merge_multimodal_embeddings(
            inputs_embeds=inputs_embeds,
            multimodal_embeddings=multimodal_embeddings,
            is_multimodal=is_multimodal,
        )

        return inputs_embeds

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        **kwargs: object,
    ) -> torch.Tensor | IntermediateTensors:
        """Run forward pass for Qwen3.5.

        Args:
            input_ids: Flattened (concatenated) input_ids corresponding to a
                batch.
            positions: Flattened (concatenated) position ids corresponding to a
                batch.
                **NOTE**: If mrope is enabled (default setting for Qwen3VL
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,).
            intermediate_tensors: Intermediate tensors from previous pipeline
                stages.
            inputs_embeds: Pre-computed input embeddings.
            **kwargs: Additional keyword arguments including:
                - pixel_values: Pixel values to be fed to a model.
                    `None` if no images are passed.
                - image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in
                    LLM. `None` if no images are passed.
                - pixel_values_videos: Pixel values of videos to be fed to a
                    model. `None` if no videos are passed.
                - video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in
                    LLM. `None` if no videos are passed.
        """

        if intermediate_tensors is not None:
            inputs_embeds = None

        hidden_states = self.language_model.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
        )

        return 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, mapper=self.hf_to_vllm_mapper)

    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, torch.dtype]:
        mamba_ssm_dtype = vllm_config.model_config.hf_text_config.mamba_ssm_dtype
        return MambaStateDtypeCalculator.gated_delta_net_state_dtype(
            vllm_config.model_config.dtype, mamba_ssm_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_text_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,
        )

    @classmethod
    def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]:
        return MambaStateCopyFuncCalculator.gated_delta_net_state_copy_func()

class Qwen3_5GatedDeltaNet(Qwen3NextGatedDeltaNet):
    def __init__(
        self,
        config: Qwen3_5TextConfig | Qwen3_5MoeTextConfig,
        model_config: ModelConfig | None = None,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        speculative_config: SpeculativeConfig | None = None,
        prefix: str = "",
    ) -> None:
        super(Qwen3NextGatedDeltaNet, self).__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.hidden_size = config.hidden_size
        self.num_v_heads = config.linear_num_value_heads
        self.num_k_heads = config.linear_num_key_heads
        self.head_k_dim = config.linear_key_head_dim
        self.head_v_dim = config.linear_value_head_dim
        self.key_dim = self.head_k_dim * self.num_k_heads
        self.value_dim = self.head_v_dim * self.num_v_heads

        self.conv_kernel_size = config.linear_conv_kernel_dim
        self.layer_idx = extract_layer_index(prefix)
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.layer_norm_epsilon = config.rms_norm_eps
        self.prefix = prefix

        self.config = config
        self.model_config = model_config
        self.cache_config = cache_config
        self.quant_config = quant_config
        self.speculative_config = speculative_config
        self.num_spec = (
            self.speculative_config.num_speculative_tokens
            if self.speculative_config
            else 0
        )

        # QKV
        self.conv_dim = self.key_dim * 2 + self.value_dim
        self.conv1d = ColumnParallelLinear(
            input_size=self.conv_kernel_size,
            output_size=self.conv_dim,
            bias=False,
            prefix=f"{prefix}.conv1d",
        )
        self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)

        self.in_proj_qkv = MergedColumnParallelLinear(
            input_size=self.hidden_size,
            output_sizes=[self.key_dim, self.key_dim, self.value_dim],
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.in_proj_qkv",
        )
        self.in_proj_z = ColumnParallelLinear(
            input_size=self.hidden_size,
            output_size=self.value_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.in_proj_z",
        )
        self.in_proj_b = ColumnParallelLinear(
            input_size=self.hidden_size,
            output_size=self.num_v_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.in_proj_b",
        )
        self.in_proj_a = ColumnParallelLinear(
            input_size=self.hidden_size,
            output_size=self.num_v_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.in_proj_a",
        )

        query_key_settings = (self.key_dim, 0, False)
        value_settings = (self.value_dim, 0, False)

        delattr(self.conv1d.weight, "weight_loader")
        set_weight_attrs(
            self.conv1d.weight,
            {
                "weight_loader": mamba_v2_sharded_weight_loader(
                    [
                        query_key_settings,
                        query_key_settings,
                        value_settings,
                    ],
                    self.tp_size,
                    self.tp_rank,
                )
            },
        )

        # selective projection used to make dt, B and C input dependant

        # time step projection (discretization)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(
            torch.ones(self.num_v_heads // self.tp_size),
        )
        self.A_log = nn.Parameter(
            torch.empty(
                divide(self.num_v_heads, self.tp_size),
            )
        )

        set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
        set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})

        self.norm = RMSNormGated(
            self.head_v_dim,
            eps=self.layer_norm_epsilon,
            group_size=None,
            norm_before_gate=True,
            device=current_platform.current_device(),
            dtype=config.dtype,
        )

        self.out_proj = RowParallelLinear(
            self.value_dim,
            self.hidden_size,
            bias=False,
            input_is_parallel=True,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )

        self.chunk_gated_delta_rule = ChunkGatedDeltaRule()

        compilation_config = get_current_vllm_config().compilation_config
        if prefix in compilation_config.static_forward_context:
            raise ValueError(f"Duplicate layer name: {prefix}")
        compilation_config.static_forward_context[prefix] = self

    def fix_query_key_value_ordering(
        self,
        mixed_qkv,
        z,
        b,
        a,
    ):
        raise NotImplementedError(
            "Qwen3.5 Series dont need to fix query key value ordering"
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        output: torch.Tensor,
    ):
        """
        Forward pass with three parts:
        1. Input projection
        2. Core attention (custom op)
        3. Output projection
        """
        num_tokens = hidden_states.size(0)

        # ============================================================
        # Part 1: Input Projection
        # ============================================================
        mixed_qkv, _ = self.in_proj_qkv(hidden_states)
        z, _ = self.in_proj_z(hidden_states)
        z = z.reshape(z.size(0), -1, self.head_v_dim)
        b, _ = self.in_proj_b(hidden_states)
        a, _ = self.in_proj_a(hidden_states)

        b = b.contiguous()
        a = a.contiguous()

        # ============================================================
        # Part 2: Core Attention (Custom Op)
        # ============================================================
        # Note: we should not use torch.empty here like other attention backends,
        # see discussions in https://github.com/vllm-project/vllm/pull/28182
        core_attn_out = torch.zeros(
            (num_tokens, self.num_v_heads // self.tp_size, self.head_v_dim),
            dtype=hidden_states.dtype,
            device=hidden_states.device,
        )

        torch.ops.vllm.gdn_attention_core(
            mixed_qkv,
            b,
            a,
            core_attn_out,
            self.prefix,
        )

        # ============================================================
        # Part 3: Output Projection
        # ============================================================
        z_shape_og = z.shape
        # Reshape input data into 2D tensor
        core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
        z = z.reshape(-1, z.shape[-1])
        core_attn_out = self.norm(core_attn_out, z)
        core_attn_out = core_attn_out.reshape(z_shape_og)
        core_attn_out = rearrange(core_attn_out, "... h d -> ... (h d)")
        output[:num_tokens], _ = self.out_proj(core_attn_out)