vllm.model_executor.models.lfm2_moe ¶

class Lfm2MoeAttention(nn.Module):
    def __init__(
        self,
        config: Lfm2MoeConfig,
        layer_idx: int,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.layer_idx = layer_idx
        self.hidden_size = hidden_size
        self.num_kv_heads = num_kv_heads
        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 = num_kv_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 = self.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.hidden_size,
            head_size=self.head_dim,
            total_num_heads=self.total_num_heads,
            total_num_kv_heads=self.total_num_kv_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.out_proj = RowParallelLinear(
            input_size=self.total_num_heads * self.head_dim,
            output_size=self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=self.max_position_embeddings,
            base=self.rope_theta,
            rope_scaling=rope_scaling,
            is_neox_style=True,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            prefix=f"{prefix}.attn",
        )
        self.q_layernorm = RMSNorm(self.head_dim, eps=config.norm_eps)
        self.k_layernorm = RMSNorm(self.head_dim, eps=config.norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        n_tokens, _ = hidden_states.shape
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q = q.view(n_tokens, self.num_heads, self.head_dim).contiguous()
        k = k.view(n_tokens, self.num_kv_heads, self.head_dim).contiguous()
        q = self.q_layernorm(q)
        k = self.k_layernorm(k)
        q, k = self.rotary_emb(positions, q, k)
        q = q.view(n_tokens, self.num_heads * self.head_dim)
        k = k.view(n_tokens, self.num_kv_heads * self.head_dim)
        attn_output = self.attn(q, k, v)
        output, _ = self.out_proj(attn_output)
        return output

class Lfm2MoeAttentionDecoderLayer(nn.Module):
    def __init__(
        self,
        config: Lfm2MoeConfig,
        layer_idx: int,
        model_config: Optional[ModelConfig] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ) -> None:
        super().__init__()
        self.prefix = prefix
        self.config = config
        self.layer_idx = layer_idx

        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        if rope_scaling is not None and getattr(
            config, "original_max_position_embeddings", None
        ):
            rope_scaling["original_max_position_embeddings"] = (
                config.original_max_position_embeddings
            )
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)

        self.self_attn = Lfm2MoeAttention(
            config=config,
            layer_idx=layer_idx,
            hidden_size=config.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            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",
        )

        if layer_idx < config.num_dense_layers:
            self.feed_forward = Lfm2MoeMlp(
                dim=config.hidden_size,
                ff_dim=config.intermediate_size,
                quant_config=quant_config,
                prefix=f"{prefix}.feed_forward",
            )
        else:
            self.feed_forward = Lfm2MoeSparseMoeBlock(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.feed_forward",
                enable_eplb=enable_eplb,
            )

        self.operator_norm = RMSNorm(config.hidden_size, eps=config.norm_eps)
        self.ffn_norm = RMSNorm(config.hidden_size, eps=config.norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if residual is None:
            residual = hidden_states
            hidden_states = self.operator_norm(hidden_states)
        else:
            hidden_states, residual = self.operator_norm(hidden_states, residual)
        hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
        hidden_states, residual = self.ffn_norm(hidden_states, residual)
        return self.feed_forward(hidden_states), residual

class Lfm2MoeForCausalLM(
    nn.Module,
    HasInnerState,
    SupportsLoRA,
    SupportsPP,
    IsHybrid,
    SupportsQuant,
    MixtureOfExperts,
):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "w1": [
            "w1",
            "w3",
        ],
    }

    # LoRA specific attributes
    embedding_modules = {
        "embed_tokens": "input_embeddings",
        "lm_head": "output_embeddings",
    }
    embedding_padding_modules = ["lm_head"]

    @classmethod
    def get_mamba_state_dtype_from_config(
        cls,
        vllm_config: "VllmConfig",
    ) -> tuple[torch.dtype, ...]:
        return MambaStateDtypeCalculator.short_conv_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]]:
        """Calculate shapes for LFM2's convolutional cache.

        Args:
            vllm_config: vLLM config

        Returns:
            Tuple containing:
            - conv_state_shape: Shape for convolutional state cache
        """
        parallel_config = vllm_config.parallel_config
        hf_config = vllm_config.model_config.hf_config

        return MambaStateShapeCalculator.short_conv_state_shape(
            tp_world_size=parallel_config.tensor_parallel_size,
            intermediate_size=hf_config.hidden_size,
            conv_kernel=hf_config.conv_L_cache,
        )

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        cache_config = vllm_config.cache_config
        lora_config = vllm_config.lora_config
        assert not cache_config.enable_prefix_caching, (
            "Lfm2Moe currently does not support prefix caching"
        )

        super().__init__()
        self.config = config
        self.model = Lfm2MoeModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )

        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = self.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
                ),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
            self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
        else:
            self.lm_head = PPMissingLayer()

        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, (Lfm2MoeAttentionDecoderLayer, Lfm2MoeShortConvDecoderLayer)
            )
            if isinstance(layer.feed_forward, Lfm2MoeSparseMoeBlock):
                example_layer = layer.feed_forward
                self.moe_layers.append(layer.feed_forward.experts)

        if example_layer is None:
            raise RuntimeError(
                "No Lfm2MoeSparseMoeBlock 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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    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.feed_forward, Lfm2MoeSparseMoeBlock):
                moe = layer.feed_forward
                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 forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
        return hidden_states

    def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor:
        logits = self.logits_processor(self.lm_head, hidden_states)
        return logits

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
        )
        return loader.load_weights(weights)

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