vllm.model_executor.models.longcat_flash ¶

class FlashConfig(PretrainedConfig):
    """Flash model configuration."""

    model_type = "longcat_flash"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=131072,
        hidden_size=4096,
        intermediate_size=8192,
        num_layers=28,
        num_hidden_layers=None,
        num_attention_heads=96,
        num_key_value_heads=128,
        ep_size=1,
        kv_lora_rank=512,
        q_lora_rank=1536,
        qk_rope_head_dim=64,
        v_head_dim=128,
        qk_nope_head_dim=128,
        num_experts_per_tok=None,
        norm_topk_prob=False,
        max_position_embeddings=8192,
        initializer_range=0.02,
        rms_norm_eps=1e-05,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=100000,
        eos_token_id=100001,
        pretraining_tp=1,
        tie_word_embeddings=False,
        rope_parameters=None,
        attention_bias=False,
        attention_dropout=0.0,
        mla_scale_q_lora=False,
        mla_scale_kv_lora=False,
        dtype="bfloat16",
        params_dtype="bfloat16",
        router_dtype="float32",
        router_bias=False,
        topk_method=None,
        routed_scaling_factor=1.0,
        zero_expert_num=0,
        zero_expert_type=None,
        nextn_use_scmoe=False,
        **kwargs,
    ):
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            dtype=dtype,
            params_dtype=params_dtype,
            router_dtype=router_dtype,
            topk_method=topk_method,
            router_bias=router_bias,
            nextn_use_scmoe=nextn_use_scmoe,
            **kwargs,
        )
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        # Longcat Flash HF configs use num_layers as the actual layer count.
        # num_hidden_layers is often derived (e.g., 2 * num_layers) and should
        # not drive vLLM layer construction.
        if num_layers is not None:
            self.num_hidden_layers = num_layers
        else:
            self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.ep_size = ep_size
        self.kv_lora_rank = kv_lora_rank
        self.q_lora_rank = q_lora_rank
        self.qk_rope_head_dim = qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.qk_nope_head_dim = qk_nope_head_dim
        self.num_experts_per_tok = num_experts_per_tok
        self.norm_topk_prob = norm_topk_prob
        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.pretraining_tp = pretraining_tp
        self.use_cache = use_cache
        # Try to set `rope_scaling` if available, otherwise use `rope_parameters`
        rope_scaling = kwargs.pop("rope_scaling", None)
        rope_parameters = rope_scaling or rope_parameters or {"rope_type": "default"}
        rope_theta = kwargs.pop("rope_theta", 1000000.0)
        if "rope_theta" not in rope_parameters:
            rope_parameters["rope_theta"] = rope_theta
        self.rope_parameters = rope_parameters
        self.attention_bias = attention_bias
        self.attention_dropout = attention_dropout
        self.mla_scale_q_lora = mla_scale_q_lora
        self.mla_scale_kv_lora = mla_scale_kv_lora
        self.zero_expert_num = zero_expert_num
        self.zero_expert_type = zero_expert_type
        self.routed_scaling_factor = routed_scaling_factor
        self.hidden_act = "silu"
        self.intermediate_size = (
            self.ffn_hidden_size
            if hasattr(self, "ffn_hidden_size")
            else intermediate_size
        )
        if hasattr(self, "moe_intermediate_size"):
            self.moe_intermediate_size = self.moe_intermediate_size
        elif hasattr(self, "expert_ffn_hidden_size"):
            self.moe_intermediate_size = self.expert_ffn_hidden_size
        else:
            self.moe_intermediate_size = self.intermediate_size

class FlashDecoderLayer(nn.Module):
    """Flash decoder layer with dual attention and MLP structure."""

    def __init__(
        self,
        vllm_config: VllmConfig,
        config: FlashConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ) -> None:
        super().__init__()
        self.layer_idx = int(prefix.split(sep=".")[-1])
        self.hidden_size = config.hidden_size
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)

        # Dual attention structure
        self.self_attn = nn.ModuleList(
            [
                DeepseekV2MLAAttention(
                    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,
                    max_position_embeddings=max_position_embeddings,
                    cache_config=cache_config,
                    quant_config=None
                    if "self_attn" in getattr(config, "disable_quant_module", [])
                    else quant_config,
                    prefix=f"{prefix}.self_attn.{i}",
                )
                for i in range(2)
            ]
        )
        self.input_layernorm = nn.ModuleList(
            [RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for i in range(2)]
        )
        self.post_attention_layernorm = nn.ModuleList(
            [RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for i in range(2)]
        )

        # Dual MLP structure
        self.mlps = nn.ModuleList(
            [
                FlashMLP(
                    hidden_size=self.hidden_size,
                    intermediate_size=config.intermediate_size,
                    hidden_act=config.hidden_act,
                    quant_config=None
                    if "mlps" in getattr(config, "disable_quant_module", [])
                    else quant_config,
                    prefix=f"{prefix}.mlps.{i}",
                )
                for i in range(2)
            ]
        )

        self.mlp = LongcatMoe(
            config=config,
            num_experts=config.n_routed_experts
            if hasattr(config, "n_routed_experts")
            else config.num_experts[self.layer_idx],
            top_k=config.moe_topk
            if hasattr(config, "moe_topk")
            else config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            quant_config=quant_config,
            prefix=(f"{prefix}.mlp"),
        )

    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[0](hidden_states)
        else:
            hidden_states, residual = self.input_layernorm[0](hidden_states, residual)

        hidden_states = self.self_attn[0](
            positions=positions,
            hidden_states=hidden_states,
            llama_4_scaling=None,
        )

        hidden_states, residual = self.post_attention_layernorm[0](
            hidden_states, residual
        )

        # moe
        hidden_states_copy = hidden_states.clone()
        moe_hidden_states = self.mlp(hidden_states_copy)

        # first mlp
        hidden_states = self.mlps[0](hidden_states)

        hidden_states, residual = self.input_layernorm[1](hidden_states, residual)

        # second_attn
        hidden_states = self.self_attn[1](
            positions=positions,
            hidden_states=hidden_states,
            llama_4_scaling=None,
        )
        hidden_states, residual = self.post_attention_layernorm[1](
            hidden_states, residual
        )

        # second_mlp
        hidden_states = self.mlps[1](hidden_states)

        hidden_states = hidden_states + moe_hidden_states

        return hidden_states, residual

class FlashMLP(nn.Module):
    """Flash MLP layer."""

    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: torch.Tensor) -> torch.Tensor:
        if x.numel() == 0:
            return x

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

@support_torch_compile
class FlashModel(nn.Module):
    """Flash model."""

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = FlashConfig(**vllm_config.model_config.hf_config.__dict__)
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        self.config = config

        self.padding_idx = getattr(config, "pad_token_id", None)
        self.vocab_size = config.vocab_size

        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
                prefix=maybe_prefix(prefix, "embed_tokens"),
            )
        else:
            self.embed_tokens = PPMissingLayer()
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: FlashDecoderLayer(
                vllm_config,
                config,
                cache_config=cache_config,
                quant_config=quant_config,
                prefix=prefix,
            ),
            prefix=f"{prefix}.layers",
        )
        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        for layer in islice(self.layers, self.start_layer, self.end_layer):
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )

        if not get_pp_group().is_last_rank:
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )

        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states