vllm.model_executor.models.longcat_flash

Inference-only Flash model compatible with HuggingFace weights.

Classes:

• FlashDecoderLayer –

  Flash decoder layer with dual attention and MLP structure.

• FlashMLP –

  Flash MLP layer.

• FlashModel –

  Flash model.

• LongcatFlashForCausalLM –

  Flash model for causal language modeling.

Functions:

• scale_mla_lora_norms_on_load –

  Fold LongCat's MLA LoRA scaling into the q/kv_a_layernorm weight loaders.

FlashDecoderLayer

Bases: Module

Flash decoder layer with dual attention and MLP structure.

Source code in vllm/model_executor/models/longcat_flash.py

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

    def __init__(
        self,
        vllm_config: VllmConfig,
        config: LongcatFlashConfig,
        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,
                    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)
            ]
        )
        for self_attn in self.self_attn:
            scale_mla_lora_norms_on_load(self_attn, config)
        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,
            top_k=config.moe_topk,
            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

FlashMLP

Bases: Module

Flash MLP layer.

Source code in vllm/model_executor/models/longcat_flash.py

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

FlashModel

Bases: Module

Flash model.

Source code in vllm/model_executor/models/longcat_flash.py

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

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

        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_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

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
        return fused_moe_make_expert_params_mapping(
            self,
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
            num_experts=self.config.n_routed_experts,
        )

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            ("fused_qkv_a_proj", "q_a_proj", 0),
            ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]

        expert_params_mapping = self.get_expert_mapping()
        loaded_params: set[str] = set()

        params_dict = dict(self.named_parameters())
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            # The MTP layers are loaded by LongCatFlashMTP.
            if name.startswith("mtp."):
                continue
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                if "mlp" in name and "mlps" not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if (
                    name.endswith(".bias") or name.endswith("_bias")
                ) and name not in params_dict:
                    continue
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                is_expert_weight = False
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, expert_shard_id = mapping
                    if weight_name not in name:
                        continue
                    is_expert_weight = True
                    name_mapped = name.replace(weight_name, param_name)
                    if (
                        name_mapped.endswith(".bias") or name_mapped.endswith("_bias")
                    ) and name not in params_dict:
                        continue
                    if is_pp_missing_parameter(name, self):
                        continue
                    param = params_dict[name_mapped]
                    weight_loader = param.weight_loader
                    weight_loader = typing.cast(
                        Callable[..., bool], param.weight_loader
                    )
                    success = weight_loader(
                        param,
                        loaded_weight,
                        name_mapped,
                        shard_id=expert_shard_id,
                        expert_id=expert_id,
                        return_success=True,
                    )
                    if success:
                        name = name_mapped
                        break
                else:
                    if is_expert_weight:
                        # We've checked that this is an expert weight
                        # However it's not mapped locally to this rank
                        # So we simply skip it
                        continue
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    # Skip loading kv_scale from ckpts towards new design.
                    if name.endswith(".kv_scale") and name not in params_dict:
                        continue
                    if name is None:
                        continue
                    if is_pp_missing_parameter(name, self):
                        continue
                    param = params_dict[name]
                    weight_loader = getattr(
                        param, "weight_loader", default_weight_loader
                    )
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

LongcatFlashForCausalLM

Bases: Module, SupportsLoRA, SupportsPP

Flash model for causal language modeling.

Source code in vllm/model_executor/models/longcat_flash.py

class LongcatFlashForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
    """Flash model for causal language modeling."""

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    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 = FlashModel(
            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 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 | None,
        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)

scale_mla_lora_norms_on_load(self_attn, config)

Fold LongCat's MLA LoRA scaling into the q/kv_a_layernorm weight loaders.

Source code in vllm/model_executor/models/longcat_flash.py

def scale_mla_lora_norms_on_load(self_attn: nn.Module, config) -> None:
    """Fold LongCat's MLA LoRA scaling into the q/kv_a_layernorm weight loaders."""
    for norm, rank, enabled in (
        (self_attn.q_a_layernorm, config.q_lora_rank, config.mla_scale_q_lora),
        (self_attn.kv_a_layernorm, config.kv_lora_rank, config.mla_scale_kv_lora),
    ):
        if enabled:
            scale = (config.hidden_size / rank) ** 0.5
            set_weight_attrs(
                norm.weight, {"weight_loader": _scaled_weight_loader(scale)}
            )