vllm.model_executor.models.bailing_moe ¶

class BailingAttention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.total_num_heads = config.num_attention_heads
        self.total_kv_heads = config.num_key_value_heads
        tp_size = get_tensor_model_parallel_world_size()

        assert self.total_num_heads % tp_size == 0
        assert self.total_kv_heads % tp_size == 0
        assert self.total_num_heads >= self.total_kv_heads

        self.num_heads = self.total_num_heads // tp_size
        self.head_dim = config.head_dim or (self.hidden_size //
                                            self.total_num_heads)
        self.q_size_per_rank = self.head_dim * self.num_heads
        self.num_kv_heads = self.total_kv_heads // tp_size
        self.kv_size_per_rank = self.num_kv_heads * self.head_dim
        self.scale = self.head_dim**-0.5
        self.use_qk_norm = getattr(config, "use_qk_norm", False)
        self.use_rmsnorm = getattr(config, "use_rmsnorm", False)

        self.query_key_value = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_kv_heads,
            bias=(config.use_bias or config.use_qkv_bias),
            quant_config=quant_config,
            prefix=f"{prefix}.query_key_value",
        )

        if self.use_qk_norm:
            self.query_layernorm = (RMSNorm(
                self.head_dim, eps=config.rms_norm_eps) if self.use_rmsnorm
                                    else nn.LayerNorm(self.head_dim, eps=1e-6))
            self.key_layernorm = (RMSNorm(
                self.head_dim, eps=config.rms_norm_eps) if self.use_rmsnorm
                                  else nn.LayerNorm(self.head_dim, eps=1e-6))

        self.dense = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            self.hidden_size,
            bias=config.use_bias,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.dense",
        )

        self.partial_rotary_factor = getattr(config, "partial_rotary_factor",
                                             1.0)

        self.rotary_dim = getattr(config, "rotary_dim", self.head_dim)

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.rotary_dim,
            max_position=config.max_position_embeddings,
            base=config.rope_theta,
            is_neox_style=True,
            rope_scaling=config.rope_scaling,
            partial_rotary_factor=self.partial_rotary_factor,
        )

        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scale,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            prefix=f"{prefix}.attn",
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:

        qkv, _ = self.query_key_value(hidden_states)
        q, k, v = qkv.split([
            self.q_size_per_rank, self.kv_size_per_rank, self.kv_size_per_rank
        ],
                            dim=-1)

        if self.use_qk_norm:
            q = q.view(-1, self.num_heads, self.head_dim)
            k = k.view(-1, self.num_kv_heads, self.head_dim)
            q = self.query_layernorm(q)
            k = self.key_layernorm(k)
            q = q.view(-1, self.q_size_per_rank)
            k = k.view(-1, self.kv_size_per_rank)

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

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

        attn_output, _ = self.dense(context_layer)
        return attn_output

class BailingMLP(nn.Module):

    def __init__(
        self,
        intermediate_size: int,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: Optional[bool] = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            config.hidden_size,
            [intermediate_size] * 2,
            bias=config.use_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            config.hidden_size,
            bias=config.use_bias,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        self.act_fn = SiluAndMul()

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

class BailingMoE(nn.Module):

    def __init__(
        self,
        intermediate_size: int,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: Optional[bool] = True,
        prefix: str = "",
    ):
        super().__init__()

        self.tp_size = get_tensor_model_parallel_world_size()
        self.tp_rank = get_tensor_model_parallel_rank()
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_expert_prob = config.norm_topk_prob
        self.hidden_size = config.hidden_size
        self.quant_config = quant_config
        self.num_shared_experts = config.num_shared_experts
        self.score_function = getattr(config, "score_function", None)
        self.n_group = getattr(config, "n_group", None)
        self.topk_group = getattr(config, "topk_group", None)
        self.use_grouped_topk = (self.n_group is not None
                                 and self.topk_group is not None)
        self.routed_scaling_factor = getattr(config, "routed_scaling_factor",
                                             1.0)

        router_dtype = getattr(config, "router_dtype", None)
        if router_dtype is None:
            self.router_dtype = None
        elif router_dtype == "fp32":
            self.router_dtype = torch.float32
        else:
            self.router_dtype = torch.bfloat16

        self.gate = nn.Linear(
            self.hidden_size,
            self.num_experts,
            bias=False,
            dtype=self.router_dtype,
        )

        if getattr(config, "moe_router_enable_expert_bias", False):
            self.gate.expert_bias = nn.Parameter(
                torch.empty((config.num_experts, ), dtype=torch.float32))
        else:
            self.gate.expert_bias = None

        self.correction_bias = (self.gate.expert_bias.data
                                if self.gate.expert_bias is not None else None)

        if self.score_function is not None:
            assert (
                self.score_function == "softmax"
                and self.correction_bias is None
            ) or (
                self.score_function == "sigmoid"
                and self.correction_bias is not None
            ), "score_function and correction_bias should be in 2 combination (softmax, None) or (sigmoid, not None)"  # noqa: E501
        else:
            # default value for scoring_func
            self.score_function = "softmax"

        self.experts = FusedMoE(
            num_experts=self.num_experts,
            top_k=self.top_k,
            hidden_size=self.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=self.norm_expert_prob,
            quant_config=quant_config,
            prefix=f"{prefix}.experts",
            scoring_func=self.score_function,
            e_score_correction_bias=self.gate.expert_bias,
            num_expert_group=self.n_group,
            topk_group=self.topk_group,
            use_grouped_topk=self.use_grouped_topk,
        )

        if self.num_shared_experts > 0:
            if hasattr(config, "moe_shared_expert_intermediate_size"):
                intermediate_size = config.moe_shared_expert_intermediate_size
            else:
                intermediate_size = config.moe_intermediate_size
            intermediate_size *= config.num_shared_experts
            self.shared_experts = BailingMLP(
                intermediate_size=intermediate_size,
                config=config,
                quant_config=quant_config,
                reduce_results=False,
                prefix=f"{prefix}.shared_experts")
        else:
            self.shared_experts = None

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_size = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_size)
        if self.shared_experts:
            shared_output = self.shared_experts(hidden_states)
        # router_logits: (num_tokens, n_experts)
        router_logits = self.gate(hidden_states.to(self.router_dtype))
        router_logits = router_logits.to(hidden_states.dtype)

        final_hidden_states = self.experts(hidden_states=hidden_states,
                                           router_logits=router_logits)

        final_hidden_states *= self.routed_scaling_factor

        if self.shared_experts:
            final_hidden_states = final_hidden_states + shared_output

        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(
                final_hidden_states)
        return final_hidden_states.view(num_tokens, hidden_size)

class BailingMoeBlock(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        layer_idx = int(prefix.split('.')[-1])
        self.config = config
        hidden_size = config.hidden_size
        intermediate_size = config.intermediate_size

        self.input_layernorm = RMSNorm(hidden_size, eps=config.rms_norm_eps)
        self.attention = BailingAttention(config,
                                          cache_config,
                                          quant_config,
                                          prefix=f"{prefix}.attention")

        self.post_attention_layernorm = RMSNorm(hidden_size,
                                                eps=config.rms_norm_eps)

        # Choose MLP class based on the number of experts and layer index
        if layer_idx < config.first_k_dense_replace:
            mlp_class = BailingMLP
        else:
            mlp_class = BailingMoE
        self.mlp = mlp_class(intermediate_size,
                             config,
                             quant_config,
                             True,
                             prefix=f"{prefix}.mlp")

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
        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.attention(
            hidden_states=hidden_states,
            position_ids=position_ids,
        )

        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual