""" LFM2-MoE (Liquid Foundation Model 2 - Mixture of Experts) implementation for SGLang. This is a hybrid architecture with attention, ShortConv, and MoE layers: - Attention layers use standard KV cache (RadixAttention) - Conv layers use MambaPool for state caching (via HybridReqToTokenPool) - First `num_dense_layers` use dense MLP, rest use MoE with sigmoid routing Key MoE characteristics: - Sigmoid routing (not softmax) - auxiliary-loss-free style - Expert bias (fp32) affects selection but not weighting - Post-hoc normalization of top-k weights """ from typing import Iterable, Optional, Set, Tuple import torch from torch import nn from sglang.srt.configs.lfm2_moe import Lfm2MoeConfig from sglang.srt.distributed import get_pp_group from sglang.srt.layers.activation import SiluAndMul from sglang.srt.layers.attention.mamba.causal_conv1d import ( causal_conv1d_fn, causal_conv1d_update, ) from sglang.srt.layers.layernorm import RMSNorm from sglang.srt.layers.linear import ( MergedColumnParallelLinear, QKVParallelLinear, ReplicatedLinear, RowParallelLinear, ) from sglang.srt.layers.logits_processor import LogitsProcessor from sglang.srt.layers.moe.fused_moe_triton import FusedMoE from sglang.srt.layers.moe.topk import TopK from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.rotary_embedding import get_rope from sglang.srt.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding, ) from sglang.srt.model_executor.forward_batch_info import ForwardBatch from sglang.srt.model_executor.forward_context import get_attn_backend from sglang.srt.model_loader.weight_utils import ( default_weight_loader, sharded_weight_loader, ) from sglang.srt.runtime_context import get_parallel from sglang.srt.utils import add_prefix, make_layers, set_weight_attrs class Lfm2MoeMLP(nn.Module): """Dense MLP for first N layers (before MoE kicks in).""" def __init__( self, config: Lfm2MoeConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() # Use MergedColumnParallelLinear for w1/w3 (gate/up projections) self.gate_up_proj = MergedColumnParallelLinear( config.hidden_size, [config.intermediate_size] * 2, bias=False, quant_config=quant_config, prefix=add_prefix("gate_up_proj", prefix), ) self.down_proj = RowParallelLinear( config.intermediate_size, config.hidden_size, bias=False, quant_config=quant_config, prefix=add_prefix("down_proj", prefix), ) self.act_fn = SiluAndMul() def forward(self, x: torch.Tensor) -> torch.Tensor: gate_up, _ = self.gate_up_proj(x) x = self.act_fn(gate_up) out, _ = self.down_proj(x) return out class Lfm2MoeSparseMoeBlock(nn.Module): """ Sparse MoE block with sigmoid routing using optimized FusedMoE. Key features: - Sigmoid scoring (not softmax) - auxiliary-loss-free style - Expert bias (fp32) for load balancing - Bias affects selection only, not weighting - Uses FusedMoE for efficient batched expert computation """ def __init__( self, config: Lfm2MoeConfig, layer_idx: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.tp_size = get_parallel().tp_size self.routed_scaling_factor = config.routed_scaling_factor if self.tp_size > config.num_experts: raise ValueError( f"Tensor parallel size {self.tp_size} is greater than " f"the number of experts {config.num_experts}." ) # Gate (router) - outputs logits for each expert self.gate = ReplicatedLinear( config.hidden_size, config.num_experts, bias=False, quant_config=None, prefix=add_prefix("gate", prefix), ) # Expert bias (fp32) - affects selection but not weighting if config.use_expert_bias: self.expert_bias = nn.Parameter( torch.zeros(config.num_experts, dtype=torch.float32) ) else: self.register_parameter("expert_bias", None) # TopK selector with sigmoid scoring self.topk = TopK( top_k=config.num_experts_per_tok, layer_id=layer_idx, renormalize=config.norm_topk_prob, scoring_func="sigmoid", correction_bias=self.expert_bias if config.use_expert_bias else None, ) # FusedMoE for efficient batched expert computation # Note: We intentionally do NOT pass routed_scaling_factor to FusedMoE. # While FusedMoE supports it, passing it there increases numerical # differences vs HuggingFace (likely due to different code paths in the # Triton runner when scaling_factor != None). We apply it manually below. self.experts = FusedMoE( num_experts=config.num_experts, top_k=config.num_experts_per_tok, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, layer_id=layer_idx, reduce_results=True, quant_config=quant_config, prefix=add_prefix("experts", prefix), ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: """Optimized expert forward pass using FusedMoE.""" # Get router logits router_logits, _ = self.gate(hidden_states) # Select top-k experts with sigmoid scoring topk_output = self.topk(hidden_states, router_logits) # Run fused expert computation final_hidden_states = self.experts(hidden_states, topk_output) # Apply routed scaling factor (see __init__ comment for why not in FusedMoE) return final_hidden_states * self.routed_scaling_factor class Lfm2MoeAttention(nn.Module): """Grouped-query attention with RoPE and Q/K layernorm.""" def __init__( self, config: Lfm2MoeConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() self.hidden_size = config.hidden_size self.total_num_heads = config.num_attention_heads self.total_num_kv_heads = config.num_key_value_heads self.head_dim = self.hidden_size // self.total_num_heads self.scaling = self.head_dim**-0.5 rope_parameters = getattr(config, "rope_parameters", None) if rope_parameters is not None and "rope_theta" in rope_parameters: rope_theta = rope_parameters["rope_theta"] else: rope_theta = getattr(config, "rope_theta", 1000000.0) self.rotary_emb = get_rope( head_size=self.head_dim, rotary_dim=self.head_dim, max_position=getattr(config, "max_position_embeddings", 128000), rope_scaling=rope_parameters or getattr(config, "rope_scaling", None), base=rope_theta, is_neox_style=True, dtype=torch.get_default_dtype(), ) self.qkv_proj = QKVParallelLinear( self.hidden_size, self.head_dim, self.total_num_heads, self.total_num_kv_heads, bias=False, quant_config=quant_config, prefix=add_prefix("qkv_proj", prefix), ) self.out_proj = RowParallelLinear( self.total_num_heads * self.head_dim, self.hidden_size, bias=False, quant_config=quant_config, prefix=add_prefix("out_proj", prefix), ) self.q_layernorm = RMSNorm(self.head_dim, eps=config.norm_eps) self.k_layernorm = RMSNorm(self.head_dim, eps=config.norm_eps) self.num_local_q_heads = self.qkv_proj.num_heads self.num_local_kv_heads = self.qkv_proj.num_kv_heads self.attn = RadixAttention( num_heads=self.num_local_q_heads, head_dim=self.head_dim, scaling=self.scaling, num_kv_heads=self.num_local_kv_heads, layer_id=layer_id, prefix=add_prefix("attn", prefix), ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ) -> torch.Tensor: T = hidden_states.shape[0] qkv, _ = self.qkv_proj(hidden_states) q_size = self.num_local_q_heads * self.head_dim kv_size = self.num_local_kv_heads * self.head_dim q, k, v = torch.split(qkv, [q_size, kv_size, kv_size], dim=-1) q = q.reshape(T, self.num_local_q_heads, self.head_dim) k = k.reshape(T, self.num_local_kv_heads, self.head_dim) q = self.q_layernorm(q.reshape(-1, self.head_dim)).reshape( T, self.num_local_q_heads, self.head_dim ) k = self.k_layernorm(k.reshape(-1, self.head_dim)).reshape( T, self.num_local_kv_heads, self.head_dim ) q, k = self.rotary_emb(positions, q, k) attn_out = self.attn(q.reshape(T, -1), k.reshape(T, -1), v, forward_batch) out, _ = self.out_proj(attn_out) return out class Lfm2MoeShortConv(nn.Module): """ Gated short convolution layer using optimized causal_conv1d kernels. Architecture: in_proj -> split(B, C, x) -> Bx -> conv1d -> C*conv_out -> out_proj - Supports tensor parallelism: hidden dimension is sharded across TP ranks """ def __init__( self, config: Lfm2MoeConfig, layer_idx: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.layer_idx = layer_idx self.conv_kernel = int(config.conv_L_cache) self.use_bias = bool(config.conv_bias) self.hidden_size = config.hidden_size # Get tensor parallel size for sharding self.tp_size = get_parallel().tp_size self.hidden_size_per_partition = self.hidden_size // self.tp_size # Use MergedColumnParallelLinear so each output (B, C, x) is sharded separately self.in_proj = MergedColumnParallelLinear( config.hidden_size, [config.hidden_size] * 3, # B, C, x each get hidden_size bias=self.use_bias, quant_config=quant_config, prefix=f"{prefix}.in_proj", ) self.out_proj = RowParallelLinear( config.hidden_size, config.hidden_size, bias=self.use_bias, input_is_parallel=True, quant_config=quant_config, prefix=f"{prefix}.out_proj", ) # Conv weights sharded along hidden dimension: (hidden_size/tp, kernel_size) self.conv_weight = nn.Parameter( torch.empty(self.hidden_size_per_partition, self.conv_kernel) ) set_weight_attrs(self.conv_weight, {"weight_loader": sharded_weight_loader(0)}) if self.use_bias: self.conv_bias = nn.Parameter(torch.empty(self.hidden_size_per_partition)) set_weight_attrs( self.conv_bias, {"weight_loader": sharded_weight_loader(0)} ) else: self.register_parameter("conv_bias", None) def forward( self, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ) -> torch.Tensor: if forward_batch.forward_mode.is_idle(): return hidden_states # The backend owns the per-request conv-state plumbing (slot indices, # prefix mask, cu-seqlens, cuda-graph buffers); this layer just runs its # depthwise conv against the returned handle. meta = get_attn_backend().conv_state_metadata(self.layer_idx, forward_batch) conv_state = meta.layer_cache.conv[0] proj, _ = self.in_proj(hidden_states) B_gate, C_gate, x = proj.chunk(3, dim=-1) Bx = B_gate * x if forward_batch.forward_mode.is_decode(): conv_out = causal_conv1d_update( Bx, conv_state, self.conv_weight, self.conv_bias, activation=None, conv_state_indices=meta.cache_indices, ) else: Bx_t = Bx.transpose(0, 1).contiguous() conv_out = causal_conv1d_fn( Bx_t, self.conv_weight, self.conv_bias, query_start_loc=meta.query_start_loc, cache_indices=meta.cache_indices, has_initial_state=meta.has_initial_state, conv_states=conv_state, activation=None, ).transpose(0, 1) output, _ = self.out_proj(C_gate * conv_out) return output class Lfm2MoeDecoderLayer(nn.Module): """ Decoder layer with attention/conv and dense MLP or MoE. - Layers 0 to num_dense_layers-1: use Lfm2MoeMLP (dense) - Layers num_dense_layers+: use Lfm2MoeSparseMoeBlock (MoE) """ def __init__( self, config: Lfm2MoeConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.layer_type = config.layer_types[layer_id] self.is_attention_layer = self.layer_type == "full_attention" self.operator_norm = RMSNorm(config.hidden_size, eps=config.norm_eps) self.ffn_norm = RMSNorm(config.hidden_size, eps=config.norm_eps) # Attention or Conv if self.is_attention_layer: self.self_attn = Lfm2MoeAttention( config=config, layer_id=layer_id, quant_config=quant_config, prefix=add_prefix("self_attn", prefix), ) else: self.conv = Lfm2MoeShortConv( config=config, layer_idx=layer_id, quant_config=quant_config, prefix=add_prefix("conv", prefix), ) # Dense MLP or MoE if layer_id < config.num_dense_layers: self.feed_forward = Lfm2MoeMLP( config=config, quant_config=quant_config, prefix=add_prefix("feed_forward", prefix), ) else: self.feed_forward = Lfm2MoeSparseMoeBlock( config=config, layer_idx=layer_id, quant_config=quant_config, prefix=add_prefix("feed_forward", prefix), ) def forward( self, layer_id: int, positions: torch.Tensor, hidden_states: torch.Tensor, residual: Optional[torch.Tensor], forward_batch: ForwardBatch, **kwargs, ) -> Tuple[torch.Tensor, torch.Tensor]: if not forward_batch.forward_mode.is_idle(): residual = hidden_states normed = self.operator_norm(hidden_states) if self.is_attention_layer: hidden_states = self.self_attn(positions, normed, forward_batch) else: hidden_states = self.conv(normed, forward_batch) hidden_states = hidden_states + residual hidden_states = hidden_states + self.feed_forward( self.ffn_norm(hidden_states) ) return hidden_states, residual class Lfm2MoeModel(nn.Module): def __init__( self, config: Lfm2MoeConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.config = config self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, org_num_embeddings=config.vocab_size, prefix=add_prefix("embed_tokens", prefix), ) # Count attention layers for KV cache sizing self.num_attention_layers = sum( 1 for lt in config.layer_types if lt == "full_attention" ) def get_layer(idx: int, prefix: str, **kwargs): return Lfm2MoeDecoderLayer( config=config, layer_id=idx, quant_config=quant_config, prefix=prefix, ) self.layers = make_layers( config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers" ) self.embedding_norm = RMSNorm(config.hidden_size, eps=config.norm_eps) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, inputs_embeds: Optional[torch.Tensor] = None, ) -> torch.Tensor: hidden_states = ( inputs_embeds if inputs_embeds is not None else self.embed_tokens(input_ids) ) residual = None for i in range(len(self.layers)): hidden_states, residual = self.layers[i]( layer_id=i, positions=positions, hidden_states=hidden_states, residual=residual, forward_batch=forward_batch, ) return self.embedding_norm(hidden_states) class Lfm2MoeForCausalLM(nn.Module): """LFM2-MoE for causal language modeling.""" fall_back_to_pt_during_load = False def __init__( self, config: Lfm2MoeConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() self.config = config self.pp_group = get_pp_group() assert self.pp_group.is_first_rank and self.pp_group.is_last_rank self.quant_config = quant_config self.model = Lfm2MoeModel( config, quant_config, prefix=add_prefix("model", prefix) ) self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, org_num_embeddings=config.vocab_size, prefix=add_prefix("lm_head", prefix), ) self.logits_processor = LogitsProcessor(config) self.num_attention_layers = self.model.num_attention_layers def get_num_kv_cache_layers(self) -> int: return self.num_attention_layers @torch.no_grad() def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, inputs_embeds: Optional[torch.Tensor] = None, **kwargs, ): hidden_states = self.model(input_ids, positions, forward_batch, inputs_embeds) return self.logits_processor( input_ids, hidden_states, self.lm_head, forward_batch ) def load_weights( self, weights: Iterable[Tuple[str, torch.Tensor]], is_mtp: bool = False ) -> Set[str]: """Load weights with FusedMoE expert format.""" stacked_params_mapping = [ # (param_name, weight_name, shard_id) ("qkv_proj", "q_proj", "q"), ("qkv_proj", "k_proj", "k"), ("qkv_proj", "v_proj", "v"), # Dense MLP w1/w3 -> gate_up_proj ("gate_up_proj", "w1", 0), ("gate_up_proj", "w3", 1), ] # FusedMoE expert params mapping # HF format: experts.{expert_id}.w{1,2,3}.weight # FusedMoE format: experts.w13_weight, experts.w2_weight expert_params_mapping = FusedMoE.make_expert_params_mapping( ckpt_gate_proj_name="w1", ckpt_down_proj_name="w2", ckpt_up_proj_name="w3", num_experts=self.config.num_experts, ) params_dict = dict(self.named_parameters()) loaded_params: Set[str] = set() embed_tokens_weight = None for name, loaded_weight in weights: if "rotary_emb.inv_freq" in name: continue if "embed_tokens.weight" in name: embed_tokens_weight = loaded_weight # Handle conv weight/bias naming: HF uses conv.conv, we use conv_weight/conv_bias if ".conv.conv.weight" in name: name = name.replace(".conv.conv.weight", ".conv.conv_weight") loaded_weight = loaded_weight.squeeze(1) # (D, 1, K) -> (D, K) if ".conv.conv.bias" in name: name = name.replace(".conv.conv.bias", ".conv.conv_bias") # Handle dense MLP w2 -> down_proj if "feed_forward.w2" in name and "experts" not in name: name = name.replace("feed_forward.w2", "feed_forward.down_proj") # Transformers >= v5.0 packs MoE expert weights into a single 3D tensor # per projection (experts.gate_up_proj / experts.down_proj) instead of # per-expert weights (experts.{i}.w{1,2,3}.weight). This is the layout an # in-memory Transformers model exposes -- e.g. the update_weights_from_tensor # / RLHF weight-sync path -- so map the packed tensors onto the fused # FusedMoE params (w13_weight / w2_weight) per expert. LFM2-MoE packs # out-features-major (gate_up_proj as [num_experts, 2 * intermediate, # hidden], down_proj as [num_experts, hidden, intermediate]), matching the # FusedMoE layout, so no transpose is needed. if "feed_forward.experts.gate_up_proj" in name: fused_name = name if fused_name.endswith(".weight"): fused_name = fused_name[: -len(".weight")] fused_name = fused_name.replace( "feed_forward.experts.gate_up_proj", "feed_forward.experts.w13_weight", ) if fused_name in params_dict: if loaded_weight.dim() != 3: raise ValueError( f"Expected a 3D packed tensor for {name}, got " f"{loaded_weight.dim()}D {tuple(loaded_weight.shape)}" ) param = params_dict[fused_name] weight_loader = param.weight_loader if loaded_weight.shape[1] % 2 != 0: raise ValueError( f"Invalid gate_up_proj shape for {name}: " f"{tuple(loaded_weight.shape)}" ) w1, w3 = loaded_weight.chunk(2, dim=1) for expert_id in range(w1.shape[0]): weight_loader( param, w1[expert_id], fused_name, shard_id="w1", expert_id=expert_id, ) weight_loader( param, w3[expert_id], fused_name, shard_id="w3", expert_id=expert_id, ) loaded_params.add(fused_name) continue if "feed_forward.experts.down_proj" in name: fused_name = name if fused_name.endswith(".weight"): fused_name = fused_name[: -len(".weight")] fused_name = fused_name.replace( "feed_forward.experts.down_proj", "feed_forward.experts.w2_weight", ) if fused_name in params_dict: if loaded_weight.dim() != 3: raise ValueError( f"Expected a 3D packed tensor for {name}, got " f"{loaded_weight.dim()}D {tuple(loaded_weight.shape)}" ) param = params_dict[fused_name] weight_loader = param.weight_loader for expert_id in range(loaded_weight.shape[0]): weight_loader( param, loaded_weight[expert_id], fused_name, shard_id="w2", expert_id=expert_id, ) loaded_params.add(fused_name) continue # Handle stacked params (QKV, dense MLP gate_up) for param_name, weight_name, shard_id in stacked_params_mapping: if weight_name not in name: continue # Skip expert weights (handled below) if "experts" in name: continue name = name.replace(weight_name, param_name) if name.endswith(".bias") and name not in params_dict: break if name not in params_dict: break param = params_dict[name] weight_loader = getattr(param, "weight_loader") weight_loader(param, loaded_weight, shard_id) loaded_params.add(name) break else: # Handle MoE expert weights using FusedMoE format # HF format: model.layers.X.feed_forward.experts.Y.wZ.weight # FusedMoE format: model.layers.X.feed_forward.experts.w13_weight/w2_weight for ( param_name, weight_name, expert_id, shard_id, ) in expert_params_mapping: if weight_name not in name: continue # Build our parameter name name = name.replace(weight_name, param_name) if name not in params_dict: continue param = params_dict[name] weight_loader = param.weight_loader weight_loader( param, loaded_weight, name, shard_id=shard_id, expert_id=expert_id, ) loaded_params.add(name) break else: # Handle regular weights if name.endswith(".bias") and name not in params_dict: continue if name not in params_dict: continue param = params_dict[name] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) loaded_params.add(name) # Handle tied lm_head weight if "lm_head.weight" not in loaded_params and "lm_head.weight" in params_dict: if embed_tokens_weight is not None: param = params_dict["lm_head.weight"] weight_loader = getattr(param, "weight_loader", default_weight_loader) weight_loader(param, embed_tokens_weight) loaded_params.add("lm_head.weight") return loaded_params EntryClass = [Lfm2MoeForCausalLM]