vllm.model_executor.models.deepseek_v2 ¶

class DeepseekV2Attention(nn.Module):
    def __init__(
        self,
        vllm_config: VllmConfig,
        config: DeepseekV2Config | DeepseekV3Config,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
        rope_scaling: dict[str, Any] | None = None,
        max_position_embeddings: int = 8192,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
        assert topk_indices_buffer is None, (
            "topk_indices_buffer is not \
        supported for DeepseekV2Attention"
        )

        if self.q_lora_rank is not None:
            self.q_a_proj = ReplicatedLinear(
                self.hidden_size,
                self.q_lora_rank,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.q_a_proj",
            )
            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(
                q_lora_rank,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.q_b_proj",
            )
        else:
            self.q_proj = ColumnParallelLinear(
                self.hidden_size,
                self.num_heads * self.qk_head_dim,
                bias=False,
                quant_config=quant_config,
                prefix=f"{prefix}.q_proj",
            )

        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa",
        )
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj",
        )
        # O projection.
        self.o_proj = RowParallelLinear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
        if rope_scaling:
            rope_scaling["rope_type"] = "deepseek_yarn"

        self.rotary_emb = get_rope(
            qk_rope_head_dim,
            rotary_dim=qk_rope_head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=False,
        )

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

        self.attn = Attention(
            self.num_local_heads,
            self.qk_head_dim,
            self.scaling,
            num_kv_heads=self.num_local_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(
                -1, self.num_local_heads, self.qk_head_dim
            )
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a)
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_pe = latent_cache[:, :, self.kv_lora_rank :]

        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)

        q[..., self.qk_nope_head_dim :] = q_pe
        k = torch.empty_like(q)
        k[..., : self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim :] = k_pe
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
            v, [0, self.qk_head_dim - self.v_head_dim], value=0
        ).view(-1, self.num_local_heads * self.qk_head_dim)
        attn_output = self.attn(q, k, v)
        attn_output = attn_output.view(-1, self.num_local_heads, self.qk_head_dim)[
            ..., : self.v_head_dim
        ].reshape(-1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output

class DeepseekV2DecoderLayer(nn.Module):
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
        config: DeepseekV2Config | None = None,
        topk_indices_buffer: torch.Tensor | None = None,
    ) -> None:
        super().__init__()

        if config is None:
            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
        parallel_config = vllm_config.parallel_config

        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep=".")[-1])
        self.layer_idx = layer_idx
        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
            vllm_config=vllm_config,
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            topk_indices_buffer=topk_indices_buffer,
        )

        if (
            config.n_routed_experts is not None
            and layer_idx >= config.first_k_dense_replace
            and layer_idx % config.moe_layer_freq == 0
        ):
            self.mlp = DeepseekV2MoE(
                config=config,
                parallel_config=parallel_config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = DeepseekV2MLP(
                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
        )
        self.routed_scaling_factor = config.routed_scaling_factor

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> torch.Tensor:
        # Self Attention
        if residual is None:
            residual = hidden_states.clone()
            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,
        )

        if hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # We scale both hidden_states and residual before
            # rmsnorm, and rmsnorm result would not affect by scale.
            hidden_states *= 1.0 / self.routed_scaling_factor
            if self.layer_idx == 0:
                # The residual is shared by all layers, we only scale it on
                # first layer.
                residual *= 1.0 / self.routed_scaling_factor

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

        if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16:
            # Fix FP16 overflow
            # Scaling the DeepseekV2MLP output, it is the input of
            # input_layernorm of next decoder layer.
            # The scaling of DeepseekV2MOE output would be done in the forward
            # of DeepseekV2MOE
            hidden_states *= 1.0 / self.routed_scaling_factor

        return hidden_states, residual

class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts, SupportsLoRA):
    packed_modules_mapping = {
        "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
        self.config = config
        self.quant_config = quant_config

        # `packed_modules_mapping` needs to be modified before
        # initializing DeepseekV2Model, as it is passed inplace to
        # quantization config init and may be used to select the
        # quant_method for relevant layers during initialization.
        self.fuse_qkv_a_proj = (
            hasattr(config, "q_lora_rank") and config.q_lora_rank is not None
        )
        if self.fuse_qkv_a_proj:
            self.packed_modules_mapping["fused_qkv_a_proj"] = [
                "q_a_proj",
                "kv_a_proj_with_mqa",
            ]

        self.model = DeepseekV2Model(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        if get_pp_group().is_last_rank:
            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
        )
        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
        self.num_expert_groups = config.n_group

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

            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_moe is None:
            raise RuntimeError("No DeepseekV2MoE layer found in model.layers.")

        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.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, DeepseekV2MoE):
                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: 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 get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        return SharedFusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts,
            num_redundant_experts=0,
        )

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
            ("fused_qkv_a_proj", "q_a_proj", 0),
            ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
        ]

        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts
            + (
                self.config.n_shared_experts
                if is_rocm_aiter_fusion_shared_expert_enabled()
                else 0
            ),
            num_redundant_experts=self.num_redundant_experts,
        )

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is not None:
                continue  # skip spec decode layers for main model

            is_fuse_shared_experts_layer = (
                is_rocm_aiter_fusion_shared_expert_enabled()
                and ("mlp.shared_experts" in name)
            )

            for param_name, weight_name, shard_id in stacked_params_mapping:
                # Skip non-stacked layers and experts (experts handled below).
                if weight_name not in name:
                    continue
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if ("mlp.experts." in name) and name not in params_dict:
                    continue
                if is_fuse_shared_experts_layer:
                    continue
                name_mapped = name.replace(weight_name, param_name)

                # QKV fusion is optional, fall back to normal
                # weight loading if it's not enabled
                # if go with fusion option, then update name
                if (
                    param_name == "fused_qkv_a_proj"
                ) and name_mapped not in params_dict:
                    continue
                else:
                    name = name_mapped
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue

                if is_pp_missing_parameter(name, self):
                    continue

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                is_expert_weight = False

                # Special handling: when AITER fusion_shared_experts is enabled,
                # checkpoints may provide a single widened shared_experts tensor
                # without explicit expert indices
                # (e.g. ...mlp.shared_experts.gate_proj.weight).
                # For models with multiple shared experts, split that tensor
                # evenly into per-shared-expert slices and load them into
                # appended expert slots mlp.experts.{n_routed_experts + j}.*
                # accordingly.
                num_chunks = 1
                if is_fuse_shared_experts_layer:
                    num_chunks = getattr(self.config, "n_shared_experts", 1) or 1
                    # Determine split axis based on op type
                    # gate/up: ColumnParallel → split along dim 0
                    # down: RowParallel → split along dim 1
                    split_dim = 1 if "down_proj.weight" in name else 0
                    total = loaded_weight.shape[split_dim]
                    assert total % num_chunks == 0, (
                        f"Shared expert weight dim {total} "
                        f"not divisible by num_chunks {num_chunks}"
                    )
                    chunk_size = total // num_chunks

                for j in range(num_chunks):
                    chunk_name = name
                    weight_to_load = loaded_weight

                    if is_fuse_shared_experts_layer:
                        if split_dim == 0:
                            weight_to_load = loaded_weight[
                                j * chunk_size : (j + 1) * chunk_size, :
                            ]
                        else:
                            weight_to_load = loaded_weight[
                                :, j * chunk_size : (j + 1) * chunk_size
                            ]
                        # Synthesize an expert-style name so expert mapping
                        # can route it
                        chunk_name = name.replace(
                            "mlp.shared_experts",
                            f"mlp.experts.{self.config.n_routed_experts + j}",
                        )

                    # Use expert_params_mapping to locate the destination
                    # param and delegate to its expert-aware weight_loader
                    # with expert_id.
                    for mapping in expert_params_mapping:
                        param_name, weight_name, expert_id, shard_id = mapping
                        if weight_name not in chunk_name:
                            continue

                        # Anyway, this is an expert weight and should not be
                        # attempted to load as other weights later
                        is_expert_weight = True

                        # Do not modify `name` since the loop may continue here
                        # Instead, create a new variable
                        name_mapped = chunk_name.replace(weight_name, param_name)

                        if is_pp_missing_parameter(name_mapped, self):
                            continue

                        param = params_dict[name_mapped]
                        # We should ask the weight loader to return success or
                        # not here since otherwise we may skip experts with
                        # other available replicas.
                        weight_loader = typing.cast(
                            Callable[..., bool], param.weight_loader
                        )
                        success = weight_loader(
                            param,
                            weight_to_load,
                            name_mapped,
                            shard_id=shard_id,
                            expert_id=expert_id,
                            return_success=True,
                        )
                        if success:
                            if not is_fuse_shared_experts_layer:
                                name = name_mapped
                            else:
                                loaded_params.add(name_mapped)
                            break
                    else:
                        if is_expert_weight:
                            # We've checked that this is an expert weight
                            # However it's not mapped locally to this rank
                            # So we simply skip it
                            continue

                        # Skip loading extra bias for GPTQ models.
                        if name.endswith(".bias") and name not in params_dict:
                            continue

                        # Remapping the name of FP8 kv-scale.
                        name = maybe_remap_kv_scale_name(name, params_dict)
                        if name is None:
                            continue

                        if is_pp_missing_parameter(name, self):
                            continue

                        param = params_dict[name]
                        weight_loader = getattr(
                            param, "weight_loader", default_weight_loader
                        )
                        weight_loader(param, loaded_weight)
            if not is_fuse_shared_experts_layer:
                loaded_params.add(name)

        return loaded_params