vllm.model_executor.models.mimo_v2_flash ¶

class MiMoV2Attention(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        head_dim: int,
        v_head_dim: int | None = None,
        v_scale: float | None = None,
        sliding_window_size: int = -1,
        attention_bias: bool = False,
        add_swa_attention_sink_bias: bool = False,
        layer_id: int = 0,
        rope_theta: float = 1000000,
        max_position_embeddings: int = 32768,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        partial_rotary_factor: float = 1.0,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.layer_id = layer_id
        tp_size = get_tensor_model_parallel_world_size()

        self.total_num_heads = num_heads
        self.num_heads = self.total_num_heads // tp_size

        self.total_num_kv_heads = num_kv_heads
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)

        self.head_dim = head_dim

        self.v_head_dim = v_head_dim if v_head_dim is not None else head_dim

        self.q_size = self.num_heads * self.head_dim
        self.k_size = self.num_kv_heads * self.head_dim
        self.v_size = self.num_kv_heads * self.v_head_dim

        self.v_scale = v_scale
        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=attention_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
            v_head_size=self.v_head_dim,
        )

        self.o_proj = RowParallelLinear(
            self.total_num_heads * self.v_head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=True,
            prefix=f"{prefix}.o_proj",
        )

        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            max_position=max_position_embeddings,
            rope_parameters={
                "rope_type": "default",
                "rope_theta": rope_theta,
                "partial_rotary_factor": partial_rotary_factor,
            },
        )

        self.attention_sink_bias = (
            torch.nn.Parameter(torch.empty(self.num_heads), requires_grad=False)
            if add_swa_attention_sink_bias
            else None
        )

        sliding_window = sliding_window_size if sliding_window_size > -1 else None
        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,
            per_layer_sliding_window=sliding_window,
            attn_type=AttentionType.DECODER,
            prefix=f"{prefix}.attn",
            sinks=self.attention_sink_bias,
        )

    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.k_size, self.v_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)

        # Apply v_scale before attention
        if self.v_scale is not None:
            v = v * self.v_scale

        v = v.view(-1, self.num_kv_heads, self.v_head_dim)
        v = torch.nn.functional.pad(v, [0, self.head_dim - self.v_head_dim], value=0)
        v = v.view(-1, self.num_kv_heads * self.head_dim)

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

        attn_output = attn_output.view(-1, self.num_heads, self.head_dim)[
            ..., : self.v_head_dim
        ].reshape(-1, self.num_heads * self.v_head_dim)

        output, _ = self.o_proj(attn_output)
        return output

class MiMoV2FlashDecoderLayer(nn.Module):
    def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
        super().__init__()
        config = vllm_config.model_config.hf_text_config
        quant_config = vllm_config.quant_config
        layer_id = extract_layer_index(prefix)

        self.hidden_size = config.hidden_size
        self.config = config
        self.layer_id = layer_id

        rope_theta = getattr(config, "rope_theta", 1000000)
        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)

        v_scale = getattr(config, "attention_value_scale", None)

        if self.is_compressed_softmax_layer():
            self.self_attn = MiMoV2Attention(
                hidden_size=self.hidden_size,
                num_heads=config.swa_num_attention_heads,
                num_kv_heads=config.swa_num_key_value_heads,
                head_dim=config.swa_head_dim,
                v_head_dim=getattr(config, "swa_v_head_dim", None),
                v_scale=v_scale,
                sliding_window_size=config.sliding_window_size,
                attention_bias=config.attention_bias,
                add_swa_attention_sink_bias=getattr(
                    config, "add_swa_attention_sink_bias", False
                ),
                layer_id=layer_id,
                rope_theta=getattr(config, "swa_rope_theta", rope_theta),
                max_position_embeddings=max_position_embeddings,
                quant_config=quant_config,
                partial_rotary_factor=getattr(config, "partial_rotary_factor", 1.0),
                prefix=f"{prefix}.self_attn",
            )
        else:
            self.self_attn = MiMoV2Attention(
                hidden_size=self.hidden_size,
                num_heads=config.num_attention_heads,
                num_kv_heads=config.num_key_value_heads,
                head_dim=config.head_dim,
                v_head_dim=getattr(config, "v_head_dim", None),
                v_scale=v_scale,
                sliding_window_size=-1,  # normal attention
                attention_bias=config.attention_bias,
                layer_id=layer_id,
                rope_theta=rope_theta,
                max_position_embeddings=max_position_embeddings,
                quant_config=quant_config,
                partial_rotary_factor=getattr(config, "partial_rotary_factor", 1.0),
                prefix=f"{prefix}.self_attn",
            )

        self.is_layer_sparse = self.is_moe_layer(layer_id)
        if self.is_layer_sparse:
            self.mlp = MiMoV2MoE(
                vllm_config=vllm_config,
                prefix=f"{prefix}.mlp",
            )
        else:
            self.mlp = MiMoV2MLP(
                hidden_size=self.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.layernorm_epsilon)
        self.post_attention_layernorm = RMSNorm(
            config.hidden_size, eps=config.layernorm_epsilon
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> tuple[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.self_attn(
            positions=positions,
            hidden_states=hidden_states,
        )

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

    def is_moe_layer(self, layer_idx: int) -> bool:
        return (
            hasattr(self.config, "moe_layer_freq")
            and layer_idx >= 0
            and not isinstance(self.config.moe_layer_freq, int)
            and self.config.moe_layer_freq[layer_idx]
        )

    def is_compressed_softmax_layer(self) -> bool:
        return self.config.hybrid_layer_pattern[self.layer_id] == 1

class MiMoV2FlashForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
    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
        self.model = MiMoV2Model(
            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
        )

    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        self.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        num_layers = len(self.model.layers)
        return (2, num_layers // 2, num_layers - 3)

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(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]]:
        return self.model.get_expert_mapping()

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

class MiMoV2MLP(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: QuantizationConfig | None = None,
        reduce_results: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(
            intermediate_size,
            hidden_size,
            bias=False,
            quant_config=quant_config,
            reduce_results=reduce_results,
            prefix=f"{prefix}.down_proj",
        )
        if hidden_act != "silu":
            raise ValueError(
                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
            )
        self.act_fn = SiluAndMul()

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