vllm.model_executor.models.deepseek_v2 ¶

class DeepseekAttention(nn.Module):
    """Normal MHA implementation used by Deepseek v1."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        config: DeepseekV2Config | DeepseekV3Config,
        hidden_size: int,
        num_heads: int,
        rope_theta: float = 10000,
        rope_parameters: dict[str, Any] | None = None,
        max_position_embeddings: int = 8192,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        **kwargs,
    ) -> 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 = 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 = 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.max_position_embeddings = max_position_embeddings

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_parameters=rope_parameters,
        )
        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",
        )

    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)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class DeepseekForCausalLM(DeepseekV2ForCausalLM):
    pass

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_parameters: 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_parameters:
            rope_parameters["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_parameters=rope_parameters,
            is_neox_style=False,
        )

        if rope_parameters:
            mscale_all_dim = rope_parameters.get("mscale_all_dim", False)
            scaling_factor = rope_parameters["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
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
        moe_layer_freq = getattr(config, "moe_layer_freq", 1)
        # 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

        # verify MLA attention specific fields
        qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0)
        qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0)
        v_head_dim = getattr(config, "v_head_dim", 0)
        kv_lora_rank = getattr(config, "kv_lora_rank", 0)
        use_mha = config.model_type == "deepseek" or all(
            dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)
        )

        if use_mha:
            attn_cls = DeepseekAttention
        elif 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=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=kv_lora_rank,
            rope_theta=config.rope_parameters["rope_theta"],
            rope_parameters=config.rope_parameters,
            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 % 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 = getattr(config, "routed_scaling_factor", 1.0)

    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 (
            not isinstance(self.self_attn, DeepseekAttention)
            and 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