mlx-community/LongCat-Flash-Chat-4bit

1# -*- coding: utf-8 -*-2# Copyright (c) 2025 Meituan3# This code is licensed under the MIT License, for details, see the ./LICENSE file. 4 5from typing import Callable, Optional, Union6 7import torch8import torch.nn.functional as F9from torch import nn10 11from transformers.activations import ACT2FN12from transformers.cache_utils import Cache, DynamicCache13from transformers.generation import GenerationMixin14from transformers.integrations import use_kernel_forward_from_hub15from transformers.masking_utils import create_causal_mask16from transformers.modeling_flash_attention_utils import FlashAttentionKwargs17from transformers.modeling_layers import GradientCheckpointingLayer18from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast19from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update20from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel21from transformers.processing_utils import Unpack22from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple23from transformers.utils.generic import check_model_inputs24from .configuration_longcat_flash import LongcatFlashConfig25 26 27@use_kernel_forward_from_hub("RMSNorm")28class LongcatFlashRMSNorm(nn.Module):29    def __init__(self, hidden_size, eps=1e-6):30        """31        LongcatFlashRMSNorm is equivalent to T5LayerNorm32        """33        super().__init__()34        self.weight = nn.Parameter(torch.ones(hidden_size))35        self.variance_epsilon = eps36 37    def forward(self, hidden_states):38        input_dtype = hidden_states.dtype39        hidden_states = hidden_states.to(torch.float32)40        variance = hidden_states.pow(2).mean(-1, keepdim=True)41        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)42        return self.weight * hidden_states.to(input_dtype)43 44    def extra_repr(self):45        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"46 47 48class LongcatFlashRotaryEmbedding(nn.Module):49    def __init__(self, config: LongcatFlashConfig, device=None):50        super().__init__()51        # BC: "rope_type" was originally "type"52        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):53            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))54        else:55            self.rope_type = "default"56        self.max_seq_len_cached = config.max_position_embeddings57        self.original_max_seq_len = config.max_position_embeddings58 59        self.config = config60        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]61 62        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)63        self.register_buffer("inv_freq", inv_freq, persistent=False)64        self.original_inv_freq = self.inv_freq65 66    @torch.no_grad()67    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)68    def forward(self, x, position_ids):69        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)70        position_ids_expanded = position_ids[:, None, :].float()71 72        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"73        with torch.autocast(device_type=device_type, enabled=False):  # Force float3274            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)75            emb = torch.cat((freqs, freqs), dim=-1)76            cos = emb.cos() * self.attention_scaling77            sin = emb.sin() * self.attention_scaling78 79        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)80 81 82class LongcatFlashMLP(nn.Module):83    def __init__(self, config, hidden_size=None, intermediate_size=None):84        super().__init__()85        self.config = config86        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size87        self.intermediate_size = config.ffn_hidden_size if intermediate_size is None else intermediate_size88 89        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)90        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)91        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)92        self.act_fn = ACT2FN[config.hidden_act]93 94    def forward(self, x):95        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))96        return down_proj97 98 99class LongcatFlashTopkRouter(nn.Module):100    def __init__(self, config):101        super().__init__()102        self.config = config103        self.top_k = config.moe_topk104        self.n_routed_experts = (105            config.n_routed_experts106            if config.zero_expert_num is None107            else config.n_routed_experts + config.zero_expert_num108        )109        self.routed_scaling_factor = config.routed_scaling_factor110        self.norm_topk_prob = config.norm_topk_prob111        self.router_bias = config.router_bias112 113        self.classifier = nn.Linear(config.hidden_size, self.n_routed_experts, bias=self.router_bias)114        self.register_buffer("e_score_correction_bias", torch.zeros((self.n_routed_experts)))115 116    @torch.no_grad()117    def get_topk_indices(self, scores):118        scores_for_choice = scores.view(-1, self.n_routed_experts) + self.e_score_correction_bias.unsqueeze(0)119        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]120        return topk_indices121 122    def forward(self, hidden_states):123        hidden_states = hidden_states.view(-1, self.config.hidden_size)124        router_logits = F.linear(hidden_states.type(torch.float32), self.classifier.weight.type(torch.float32))125        scores = router_logits.softmax(dim=-1)126        topk_indices = self.get_topk_indices(scores)127        topk_weights = scores.gather(1, topk_indices)128        if self.norm_topk_prob:129            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20130            topk_weights /= denominator131        topk_weights = topk_weights * self.routed_scaling_factor132        return topk_indices, topk_weights133 134 135class LongcatFlashMoE(nn.Module):136    """137    moe module.138    """139 140    def __init__(self, config):141        super().__init__()142        self.config = config143        self.experts = nn.ModuleList(144            [145                LongcatFlashMLP(config, intermediate_size=config.expert_ffn_hidden_size)146                for _ in range(config.n_routed_experts)147            ]148        )149        self.router = LongcatFlashTopkRouter(config)150        self.zero_expert_num = config.zero_expert_num151        self.zero_expert_type = config.zero_expert_type152 153    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):154        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)155        total_experts = len(self.experts) if self.zero_expert_num is None else len(self.experts) + self.zero_expert_num156 157        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=total_experts)158        expert_mask = expert_mask.permute(2, 0, 1)159 160        for expert_idx in range(total_experts):161            expert = self.experts[expert_idx] if expert_idx < len(self.experts) else None162            mask = expert_mask[expert_idx]163            token_indices, weight_indices = torch.where(mask)164 165            if token_indices.numel() > 0:166                expert_weights = topk_weights[token_indices, weight_indices]167                expert_input = hidden_states[token_indices]168 169                if self.zero_expert_num is None or expert_idx < len(self.experts):170                    expert_output = expert(expert_input)171                elif self.zero_expert_type == "identity":172                    expert_output = expert_input173                else:174                    raise ValueError("Unknown condition")175 176                weighted_output = expert_output * expert_weights.unsqueeze(-1)177                final_hidden_states.index_add_(0, token_indices, weighted_output)178 179        return final_hidden_states.type(hidden_states.dtype)180 181    def forward(self, hidden_states):182        orig_shape = hidden_states.shape183        topk_indices, topk_weights = self.router(hidden_states)184        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])185        hidden_states = self.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)186        return hidden_states187 188 189def rotate_half(x):190    """Rotates half the hidden dims of the input."""191    x1 = x[..., : x.shape[-1] // 2]192    x2 = x[..., x.shape[-1] // 2 :]193    return torch.cat((-x2, x1), dim=-1)194 195 196def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:197    """198    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,199    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)200    """201    batch, num_key_value_heads, slen, head_dim = hidden_states.shape202    if n_rep == 1:203        return hidden_states204    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)205    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)206 207 208def eager_attention_forward(209    module: nn.Module,210    query: torch.Tensor,211    key: torch.Tensor,212    value: torch.Tensor,213    attention_mask: Optional[torch.Tensor],214    scaling: float,215    dropout: float = 0.0,216    **kwargs: Unpack[TransformersKwargs],217):218    key_states = repeat_kv(key, module.num_key_value_groups)219    value_states = repeat_kv(value, module.num_key_value_groups)220 221    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling222    if attention_mask is not None:223        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]224        attn_weights = attn_weights + causal_mask225 226    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)227    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)228    attn_output = torch.matmul(attn_weights, value_states)229    attn_output = attn_output.transpose(1, 2).contiguous()230 231    return attn_output, attn_weights232 233 234def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1, use_mla=False):235    """Applies Rotary Position Embedding to the query and key tensors.236 237    Args:238        q (`torch.Tensor`): The query tensor.239        k (`torch.Tensor`): The key tensor.240        cos (`torch.Tensor`): The cosine part of the rotary embedding.241        sin (`torch.Tensor`): The sine part of the rotary embedding.242        position_ids (`torch.Tensor`, *optional*):243            Deprecated and unused.244        unsqueeze_dim (`int`, *optional*, defaults to 1):245            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and246            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note247            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and248            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes249            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have250            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.251    Returns:252        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.253    """254    cos = cos.unsqueeze(unsqueeze_dim)255    sin = sin.unsqueeze(unsqueeze_dim)256 257    if use_mla:258        b, h, s, d = q.shape259        q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)260 261        b, h, s, d = k.shape262        k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)263 264    q_embed = (q * cos) + (rotate_half(q) * sin)265    k_embed = (k * cos) + (rotate_half(k) * sin)266    return q_embed, k_embed267 268 269class LongcatFlashMLA(nn.Module):270    """Modified from Deepseek MLA"""271 272    def __init__(self, config: LongcatFlashConfig, layer_idx: int):273        super().__init__()274        self.config = config275        self.layer_idx = layer_idx276        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads277        self.attention_dropout = config.attention_dropout278        self.num_heads = config.num_attention_heads279        self.rope_theta = config.rope_theta280        self.q_lora_rank = config.q_lora_rank281        self.qk_rope_head_dim = config.qk_rope_head_dim282        self.kv_lora_rank = config.kv_lora_rank283        self.v_head_dim = config.v_head_dim284        self.qk_nope_head_dim = config.qk_nope_head_dim285        self.qk_head_dim = config.qk_head_dim286 287        self.is_causal = True288        if self.q_lora_rank is None:289            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)290        else:291            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)292            self.q_a_layernorm = LongcatFlashRMSNorm(config.q_lora_rank)293            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)294 295        self.kv_a_proj_with_mqa = nn.Linear(296            config.hidden_size,297            self.kv_lora_rank + self.qk_rope_head_dim,298            bias=config.attention_bias,299        )300        self.kv_a_layernorm = LongcatFlashRMSNorm(self.kv_lora_rank)301        self.kv_b_proj = nn.Linear(302            self.kv_lora_rank,303            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),304            bias=False,305        )306 307        self.o_proj = nn.Linear(308            self.num_heads * self.v_head_dim,309            config.hidden_size,310            bias=config.attention_bias,311        )312 313        if config.mla_scale_q_lora:314            self.mla_scale_q_lora = (config.hidden_size / self.q_lora_rank) ** 0.5315        if config.mla_scale_kv_lora:316            self.mla_scale_kv_lora = (config.hidden_size / self.kv_lora_rank) ** 0.5317        self.scaling = self.qk_head_dim ** (-0.5)318 319    def forward(320        self,321        hidden_states: torch.Tensor,322        position_embeddings: tuple[torch.Tensor, torch.Tensor],323        attention_mask: Optional[torch.Tensor],324        past_key_value: Optional[Cache] = None,325        cache_position: Optional[torch.LongTensor] = None,326        **kwargs: Unpack[FlashAttentionKwargs],327    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:328        batch_size, seq_length = hidden_states.shape[:-1]329        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)330        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)331 332        q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states))).view(query_shape).transpose(1, 2)333        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)334 335        # apply q_lora scaling336        if self.mla_scale_q_lora is not None:337            q_pass = q_pass * self.mla_scale_q_lora338            q_rot = q_rot * self.mla_scale_q_lora339 340        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)341        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)342        k_pass = self.kv_a_layernorm(k_pass)343 344        # apply kv_lora scaling345        if self.mla_scale_kv_lora is not None:346            k_pass = k_pass * self.mla_scale_kv_lora347 348        k_pass = self.kv_b_proj(k_pass).view(key_shape).transpose(1, 2)349        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)350 351        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)352 353        cos, sin = position_embeddings354        q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin, use_mla=True)355        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)356 357        query_states = torch.cat((q_pass, q_rot), dim=-1)358        key_states = torch.cat((k_pass, k_rot), dim=-1)359 360        if past_key_value is not None:361            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}362            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)363 364        if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:365            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])366 367        attention_interface: Callable = eager_attention_forward368        if self.config._attn_implementation != "eager":369            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]370 371        attn_output, attn_weights = attention_interface(372            self,373            query_states,374            key_states,375            value_states,376            attention_mask,377            dropout=0.0 if not self.training else self.attention_dropout,378            scaling=self.scaling,379            **kwargs,380        )381 382        if self.config._attn_implementation == "flash_attention_2" and self.qk_head_dim != self.v_head_dim:383            attn_output = attn_output[:, :, :, : self.v_head_dim]384 385        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()386        attn_output = self.o_proj(attn_output)387        return attn_output, attn_weights388 389 390def create_attention_block(class_name, *args, **kwargs):391    attention_mapping = {"MLA": LongcatFlashMLA}392 393    chosen_class = attention_mapping.get(class_name)394    if not chosen_class:395        raise ValueError(f"No class found for name: {class_name}")396 397    return chosen_class(*args, **kwargs)398 399 400class LongcatFlashDecoderLayer(GradientCheckpointingLayer):401    def __init__(self, config: LongcatFlashConfig, layer_idx: int):402        super().__init__()403        self.layer_idx = layer_idx404        self.hidden_size = config.hidden_size405        self.mlp = LongcatFlashMoE(config)406 407        self_attn = []408        mlps = []409        input_layernorm = []410        post_attention_layernorm = []411        for i in range(2):412            self_attn.append(413                create_attention_block(config.attention_method, config=config, layer_idx=layer_idx * 2 + i)414            )415            mlps.append(LongcatFlashMLP(config))416            input_layernorm.append(LongcatFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps))417            post_attention_layernorm.append(LongcatFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps))418 419        self.self_attn = nn.ModuleList(self_attn)420        self.mlps = nn.ModuleList(mlps)421        self.input_layernorm = nn.ModuleList(input_layernorm)422        self.post_attention_layernorm = nn.ModuleList(post_attention_layernorm)423 424    def forward(425        self,426        hidden_states: torch.Tensor,427        attention_mask: Optional[torch.Tensor] = None,428        position_ids: Optional[torch.LongTensor] = None,429        past_key_value: Optional[Cache] = None,430        use_cache: Optional[bool] = False,431        cache_position: Optional[torch.LongTensor] = None,432        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,433        **kwargs: Unpack[FlashAttentionKwargs],434    ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]:435        for i in range(2):436            residual = hidden_states437 438            hidden_states = self.input_layernorm[i](hidden_states)439 440            hidden_states, _ = self.self_attn[i](441                hidden_states=hidden_states,442                attention_mask=attention_mask,443                position_ids=position_ids,444                past_key_value=past_key_value,445                use_cache=use_cache,446                cache_position=cache_position,447                position_embeddings=position_embeddings,448                **kwargs,449            )450            hidden_states = residual + hidden_states451 452            residual = hidden_states453            hidden_states = self.post_attention_layernorm[i](hidden_states)454 455            if i == 0:456                shortcut_mlp_output = self.mlp(hidden_states)  # shortcut output (MoE output)457 458            hidden_states = self.mlps[i](hidden_states)459            hidden_states = residual + hidden_states460            if i == 1:461                hidden_states = hidden_states + shortcut_mlp_output462 463        return hidden_states464 465 466@auto_docstring467class LongcatFlashPreTrainedModel(PreTrainedModel):468    config: LongcatFlashConfig469    base_model_prefix = "model"470    supports_gradient_checkpointing = True471    _no_split_modules = ["LongcatFlashDecoderLayer"]472    _skip_keys_device_placement = ["past_key_values"]473    _supports_flash_attn = True474    _supports_sdpa = True475    _supports_flex_attn = True476    _can_compile_fullgraph = True477    _supports_attention_backend = True478    _can_record_outputs = {479        "hidden_states": LongcatFlashDecoderLayer,480        "attentions": LongcatFlashMLA,481    }482 483 484@auto_docstring485class LongcatFlashModel(LongcatFlashPreTrainedModel):486    _keys_to_ignore_on_load_unexpected = [r"model\.mtp.*"]487 488    def __init__(self, config: LongcatFlashConfig):489        super().__init__(config)490        self.padding_idx = config.pad_token_id491        self.vocab_size = config.vocab_size492 493        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)494        self.layers = nn.ModuleList(495            [LongcatFlashDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]496        )497        self.norm = LongcatFlashRMSNorm(config.hidden_size, eps=config.rms_norm_eps)498        self.rotary_emb = LongcatFlashRotaryEmbedding(config=config)499        self.gradient_checkpointing = False500 501        # Initialize weights and apply final processing502        self.post_init()503 504    @check_model_inputs505    @auto_docstring506    def forward(507        self,508        input_ids: Optional[torch.LongTensor] = None,509        attention_mask: Optional[torch.Tensor] = None,510        position_ids: Optional[torch.LongTensor] = None,511        past_key_values: Optional[Cache] = None,512        inputs_embeds: Optional[torch.FloatTensor] = None,513        cache_position: Optional[torch.LongTensor] = None,514        use_cache: Optional[bool] = None,515        **kwargs: Unpack[TransformersKwargs],516    ) -> BaseModelOutputWithPast:517        if (input_ids is None) ^ (inputs_embeds is not None):518            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")519 520        if inputs_embeds is None:521            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)522 523        if use_cache and past_key_values is None:524            past_key_values = DynamicCache()525 526        if cache_position is None:527            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0528            cache_position: torch.Tensor = torch.arange(529                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device530            )531 532        if position_ids is None:533            position_ids = cache_position.unsqueeze(0)534 535        causal_mask = create_causal_mask(536            config=self.config,537            input_embeds=inputs_embeds,538            attention_mask=attention_mask,539            cache_position=cache_position,540            past_key_values=past_key_values,541            position_ids=position_ids,542        )543 544        hidden_states = inputs_embeds545        position_embeddings = self.rotary_emb(hidden_states, position_ids)546 547        for decoder_layer in self.layers[: self.config.num_hidden_layers]:548            hidden_states = decoder_layer(549                hidden_states,550                attention_mask=causal_mask,551                position_ids=position_ids,552                past_key_value=past_key_values,553                cache_position=cache_position,554                position_embeddings=position_embeddings,555                **kwargs,556            )557 558        hidden_states = self.norm(hidden_states)559        return BaseModelOutputWithPast(560            last_hidden_state=hidden_states,561            past_key_values=past_key_values,562        )563 564 565@auto_docstring566class LongcatFlashForCausalLM(LongcatFlashPreTrainedModel, GenerationMixin):567    _tied_weights_keys = ["lm_head.weight"]568    _tp_plan = {"lm_head": "colwise_rep"}569    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}570    _keys_to_ignore_on_load_unexpected = [r"model\.mtp.*"]571 572    def __init__(self, config):573        super().__init__(config)574        self.model = LongcatFlashModel(config)575        self.vocab_size = config.vocab_size576        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)577 578        # Initialize weights and apply final processing579        self.post_init()580 581    def set_decoder(self, decoder):582        self.model = decoder583 584    def get_decoder(self):585        return self.model586 587    @can_return_tuple588    @auto_docstring589    def forward(590        self,591        input_ids: Optional[torch.LongTensor] = None,592        attention_mask: Optional[torch.Tensor] = None,593        position_ids: Optional[torch.LongTensor] = None,594        past_key_values: Optional[Cache] = None,595        inputs_embeds: Optional[torch.FloatTensor] = None,596        labels: Optional[torch.LongTensor] = None,597        use_cache: Optional[bool] = None,598        cache_position: Optional[torch.LongTensor] = None,599        logits_to_keep: Union[int, torch.Tensor] = 0,600        **kwargs: Unpack[TransformersKwargs],601    ) -> CausalLMOutputWithPast:602        r"""603        Example:604 605        ```python606        >>> from transformers import AutoTokenizer, LongcatFlashForCausalLM607 608        >>> model = LongcatFlashForCausalLM.from_pretrained("meta-longcat_flash/LongcatFlash-2-7b-hf")609        >>> tokenizer = AutoTokenizer.from_pretrained("meta-longcat_flash/LongcatFlash-2-7b-hf")610 611        >>> prompt = "Hey, are you conscious? Can you talk to me?"612        >>> inputs = tokenizer(prompt, return_tensors="pt")613 614        >>> # Generate615        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)616        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]617        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."618        ```"""619        outputs: BaseModelOutputWithPast = self.model(620            input_ids=input_ids,621            attention_mask=attention_mask,622            position_ids=position_ids,623            past_key_values=past_key_values,624            inputs_embeds=inputs_embeds,625            use_cache=use_cache,626            cache_position=cache_position,627            **kwargs,628        )629 630        hidden_states = outputs.last_hidden_state631        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss632        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep633        logits = self.lm_head(hidden_states[:, slice_indices, :])634 635        loss = None636        if labels is not None:637            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)638 639        return CausalLMOutputWithPast(640            loss=loss,641            logits=logits,642            past_key_values=outputs.past_key_values,643            hidden_states=outputs.hidden_states,644            attentions=outputs.attentions,645        )646 647 648__all__ = ["LongcatFlashPreTrainedModel", "LongcatFlashModel", "LongcatFlashForCausalLM"]649