# coding=utf-8 """ SGLang SDARMoeModelLM (block diffusion / dLLM-style forward) with MoE MLP. """ import logging from typing import Iterable, Optional, Tuple, Union import torch from torch import nn from transformers import PretrainedConfig from sglang.srt.distributed import ( get_pp_group, tensor_model_parallel_all_reduce, ) from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation from sglang.srt.eplb.expert_location_dispatch import ExpertLocationDispatchInfo from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes from sglang.srt.layers.dp_attention import ( is_dp_attention_enabled, ) from sglang.srt.layers.layernorm import RMSNorm from sglang.srt.layers.linear import ( QKVParallelLinear, ReplicatedLinear, RowParallelLinear, ) from sglang.srt.layers.logits_processor import LogitsProcessor from sglang.srt.layers.moe import ( get_moe_a2a_backend, should_skip_post_experts_all_reduce, ) from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE from sglang.srt.layers.moe.topk import TopK from sglang.srt.layers.moe.utils import ( RoutingMethodType, filter_moe_weight_param_global_expert, ) from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.layers.radix_attention import AttentionType, RadixAttention from sglang.srt.layers.rotary_embedding import get_rope from sglang.srt.layers.utils import PPMissingLayer, get_layer_id from sglang.srt.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding, ) from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors from sglang.srt.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name, ) from sglang.srt.models.utils import ( apply_qk_norm, create_fused_set_kv_buffer_arg, enable_fused_set_kv_buffer, ) from sglang.srt.runtime_context import ( get_forward, get_parallel, get_server_args, get_stream, ) from sglang.srt.utils import LazyValue, add_prefix, is_cuda, make_layers logger = logging.getLogger(__name__) _is_cuda = is_cuda() class SDARMoeSparseMoeBlock(nn.Module): """ Qwen3MoE-style sparse MoE block: - gate: ReplicatedLinear(hidden, num_experts) - topk routing: TopK - experts: get_moe_impl_class(quant_config)(...) """ def __init__( self, layer_id: int, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.layer_id = layer_id self.tp_size = get_parallel().tp_size if self.tp_size > config.num_experts: raise ValueError( f"Tensor parallel size {self.tp_size} > num_experts {config.num_experts}." ) self.topk = TopK( top_k=config.num_experts_per_tok, renormalize=config.norm_topk_prob, use_grouped_topk=False, layer_id=layer_id, ) self.experts = get_moe_impl_class(quant_config)( num_experts=config.num_experts + get_server_args().ep_num_redundant_experts, top_k=config.num_experts_per_tok, layer_id=layer_id, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, quant_config=quant_config, prefix=add_prefix("experts", prefix), routing_method_type=RoutingMethodType.Renormalize, ) self.gate = ReplicatedLinear( config.hidden_size, config.num_experts, bias=False, quant_config=None, prefix=add_prefix("gate", prefix), ) # Deepep / FuseEP support if get_moe_a2a_backend().is_deepep(): self.ep_size = get_parallel().moe_ep_size self.num_experts = ( config.num_experts + get_server_args().ep_num_redundant_experts ) self.top_k = config.num_experts_per_tok def forward( self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None, ) -> torch.Tensor: if ( not get_moe_a2a_backend().is_deepep() and not get_moe_a2a_backend().is_ascend_fuseep() ): return self.forward_normal(hidden_states) else: assert forward_batch is not None, "deepep/fuseep MoE needs forward_batch" return self.forward_deepep(hidden_states, forward_batch) def forward_normal( self, hidden_states: torch.Tensor, ) -> torch.Tensor: num_tokens, hidden_dim = hidden_states.shape hidden_states = hidden_states.view(-1, hidden_dim) router_logits, _ = self.gate(hidden_states) # (T, E) topk_output = self.topk(hidden_states, router_logits) out = self.experts(hidden_states, topk_output) # (T, H) if self.tp_size > 1 and not should_skip_post_experts_all_reduce( is_tp_path=True, ): out = tensor_model_parallel_all_reduce(out) return out.view(num_tokens, hidden_dim) def forward_deepep(self, hidden_states: torch.Tensor, forward_batch: ForwardBatch): if hidden_states.shape[0] > 0: router_logits, _ = self.gate(hidden_states) topk_output = self.topk( hidden_states, router_logits, num_token_non_padded=forward_batch.num_token_non_padded, expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new( layer_id=self.layer_id ), ) else: topk_output = self.topk.empty_topk_output(hidden_states.device) out = self.experts(hidden_states=hidden_states, topk_output=topk_output) return out def get_moe_weights(self): return [ p.data for name, p in self.experts.named_parameters() if name not in ["correction_bias"] and filter_moe_weight_param_global_expert( name, p, self.experts.num_local_experts ) ] class SDARMoeAttention(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, reduce_results: bool = True, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ): super().__init__() self.layer_id = layer_id self.hidden_size = config.hidden_size self.total_num_heads = config.num_attention_heads attn_tp_rank = get_parallel().attn_tp_rank attn_tp_size = get_parallel().attn_tp_size assert self.total_num_heads % attn_tp_size == 0 self.num_heads = self.total_num_heads // attn_tp_size self.total_num_kv_heads = config.num_key_value_heads if self.total_num_kv_heads >= attn_tp_size: assert self.total_num_kv_heads % attn_tp_size == 0 else: assert attn_tp_size % self.total_num_kv_heads == 0 self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size) self.head_dim = getattr( config, "head_dim", self.hidden_size // self.total_num_heads ) self.q_size = self.num_heads * self.head_dim self.kv_size = self.num_kv_heads * self.head_dim self.scale = self.head_dim**-0.5 self.qkv_proj = QKVParallelLinear( self.hidden_size, self.head_dim, self.total_num_heads, self.total_num_kv_heads, bias=getattr(config, "attention_bias", False), quant_config=quant_config, prefix=add_prefix("qkv_proj", prefix), ) self.o_proj = RowParallelLinear( self.total_num_heads * self.head_dim, self.hidden_size, bias=getattr(config, "attention_bias", False), quant_config=quant_config, reduce_results=reduce_results, tp_rank=attn_tp_rank, tp_size=attn_tp_size, prefix=add_prefix("o_proj", prefix), ) self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) rope_theta = getattr(config, "rope_theta", 10000.0) rope_scaling = getattr(config, "rope_scaling", None) max_pos = getattr(config, "max_position_embeddings", 32768) self.rotary_dim = self.head_dim self.rotary_emb = get_rope( self.head_dim, rotary_dim=self.rotary_dim, max_position=max_pos, base=rope_theta, rope_scaling=rope_scaling, ) self.attn = RadixAttention( self.num_heads, self.head_dim, self.scale, num_kv_heads=self.num_kv_heads, layer_id=layer_id, attn_type=AttentionType.ENCODER_ONLY, prefix=add_prefix("attn", prefix), ) self.alt_stream = alt_stream def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ) -> torch.Tensor: if get_server_args().rl_on_policy_target is not None: hidden_states = hidden_states.bfloat16() qkv, _ = self.qkv_proj(hidden_states) q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) q, k = apply_qk_norm( q=q, k=k, q_norm=self.q_norm, k_norm=self.k_norm, head_dim=self.head_dim, alt_stream=self.alt_stream, ) q, k = self.rotary_emb( positions, q, k, fused_set_kv_buffer_arg=( create_fused_set_kv_buffer_arg( value=v, layer=self.attn, forward_batch=forward_batch, ) if enable_fused_set_kv_buffer(forward_batch) else None ), ) if get_server_args().rl_on_policy_target is not None: q = q.to(torch.bfloat16) k = k.to(torch.bfloat16) context = self.attn( q, k, v, forward_batch, save_kv_cache=not enable_fused_set_kv_buffer(forward_batch), ) out, _ = self.o_proj(context) return out class SDARMoeBlock(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ): super().__init__() self.config = config self.hidden_size = config.hidden_size self.layer_id = layer_id norm_kwargs = ( dict( weight_dtype=torch.float32, cast_x_before_out_mul=True, override_orig_dtype=torch.float32, fp32_residual=True, ) if get_server_args().rl_on_policy_target is not None else {} ) self.input_layernorm = RMSNorm( self.hidden_size, eps=config.rms_norm_eps, **norm_kwargs ) self.post_attention_layernorm = RMSNorm( self.hidden_size, eps=config.rms_norm_eps, **norm_kwargs ) self.self_attn = SDARMoeAttention( config=config, layer_id=layer_id, quant_config=quant_config, reduce_results=False, prefix=add_prefix("self_attn", prefix), alt_stream=alt_stream, ) self.mlp = SDARMoeSparseMoeBlock( layer_id=layer_id, config=config, quant_config=quant_config, prefix=add_prefix("mlp", prefix), ) self.layer_scatter_modes = LayerScatterModes.init_new( layer_id=layer_id, num_layers=config.num_hidden_layers, is_layer_sparse=True, is_previous_layer_sparse=True, is_next_layer_sparse=True, ) self.layer_communicator = LayerCommunicator( layer_scatter_modes=self.layer_scatter_modes, input_layernorm=self.input_layernorm, post_attention_layernorm=self.post_attention_layernorm, allow_reduce_scatter=True, is_last_layer=(layer_id == config.num_hidden_layers - 1), ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], ) -> Tuple[torch.Tensor, torch.Tensor]: hidden_states, residual = self.layer_communicator.prepare_attn( hidden_states, residual, forward_batch ) if hidden_states.shape[0] != 0: hidden_states = self.self_attn( positions=positions, hidden_states=hidden_states, forward_batch=forward_batch, ) hidden_states, residual = self.layer_communicator.prepare_mlp( hidden_states, residual, forward_batch ) fuse_mlp_allreduce = ( self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer( forward_batch ) ) mlp_reduce_scatter = self.layer_communicator.should_use_reduce_scatter( forward_batch ) with get_forward().scoped( fuse_mlp_allreduce=fuse_mlp_allreduce, mlp_reduce_scatter=mlp_reduce_scatter, ): hidden_states = self.mlp( hidden_states, forward_batch=forward_batch, ) if fuse_mlp_allreduce: hidden_states._sglang_needs_allreduce_fusion = True else: hidden_states, residual = self.layer_communicator.postprocess_layer( hidden_states, residual, forward_batch ) return hidden_states, residual class SDARMoeModel(nn.Module): def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ): super().__init__() self.config = config self.vocab_size = config.vocab_size self.embed_dim = config.hidden_size self.pp_group = get_pp_group() if self.pp_group.is_first_rank: self.embed_tokens = VocabParallelEmbedding( self.vocab_size, self.embed_dim, quant_config=quant_config, use_attn_tp_group=is_dp_attention_enabled(), prefix=add_prefix("embed_tokens", prefix), ) else: self.embed_tokens = PPMissingLayer() self.layers, self.start_layer, self.end_layer = make_layers( config.num_hidden_layers, lambda idx, prefix: SDARMoeBlock( config=config, layer_id=idx, quant_config=quant_config, prefix=prefix, alt_stream=alt_stream, ), pp_rank=self.pp_group.rank_in_group, pp_size=self.pp_group.world_size, prefix=add_prefix("layers", prefix), ) if self.pp_group.is_last_rank: norm_kwargs = ( dict( weight_dtype=torch.float32, cast_x_before_out_mul=True, override_orig_dtype=torch.float32, fp32_residual=True, ) if get_server_args().rl_on_policy_target is not None else {} ) self.norm = RMSNorm(self.embed_dim, eps=config.rms_norm_eps, **norm_kwargs) else: self.norm = PPMissingLayer(return_tuple=True) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, input_embeds: torch.Tensor = None, pp_proxy_tensors: Optional[PPProxyTensors] = None, ) -> Union[torch.Tensor, PPProxyTensors]: if self.pp_group.is_first_rank: hidden_states = ( self.embed_tokens(input_ids) if input_embeds is None else input_embeds ) residual = None else: assert pp_proxy_tensors is not None hidden_states = pp_proxy_tensors["hidden_states"] residual = pp_proxy_tensors.get("residual", None) for i in range(self.start_layer, self.end_layer): layer = self.layers[i] with get_global_expert_distribution_recorder().with_current_layer(i): hidden_states, residual = layer( positions, hidden_states, forward_batch, residual ) if not self.pp_group.is_last_rank: return PPProxyTensors( {"hidden_states": hidden_states, "residual": residual} ) if not forward_batch.forward_mode.is_idle(): hidden_states, residual = self.norm(hidden_states, residual) return hidden_states class SDARMoeForCausalLM(nn.Module): fall_back_to_pt_during_load = False def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.pp_group = get_pp_group() assert self.pp_group.world_size == 1, ( f"SDARMoeForCausalLM does not support pipeline parallel (pp_size={self.pp_group.world_size}). " "Please set pp_size=1." ) self.pp_group = get_pp_group() self.config = config self.quant_config = quant_config alt_stream = get_stream("alt") if _is_cuda else None self.model = SDARMoeModel( config, quant_config=quant_config, prefix=add_prefix("model", ""), alt_stream=alt_stream, ) if self.pp_group.is_last_rank: tp_size = get_parallel().tp_size if ( self.pp_group.world_size == 1 and getattr(config, "tie_word_embeddings", False) and tp_size == 1 ): self.lm_head = self.model.embed_tokens else: self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, use_attn_tp_group=get_server_args().enable_dp_lm_head, prefix=add_prefix("lm_head", prefix), ) else: self.lm_head = PPMissingLayer() self.logits_processor = LogitsProcessor(config, return_full_logits=True) @property def start_layer(self): return self.model.start_layer @property def end_layer(self): return self.model.end_layer @torch.no_grad() def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, input_embeds: Optional[torch.Tensor] = None, pp_proxy_tensors: Optional[PPProxyTensors] = None, ) -> torch.Tensor: hidden_states = self.model( input_ids=input_ids, positions=positions, forward_batch=forward_batch, input_embeds=input_embeds, pp_proxy_tensors=pp_proxy_tensors, ) if self.pp_group.is_last_rank: return self.logits_processor( input_ids, hidden_states, self.lm_head, forward_batch ) return hidden_states def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): stacked_params_mapping = [ ("qkv_proj", "q_proj", "q"), ("qkv_proj", "k_proj", "k"), ("qkv_proj", "v_proj", "v"), ("gate_up_proj", "gate_proj", 0), ("gate_up_proj", "up_proj", 1), ] expert_params_mapping = FusedMoE.make_expert_params_mapping( ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", ckpt_up_proj_name="up_proj", num_experts=self.config.num_experts, ) if not hasattr(self, "_cached_params_dict"): self._cached_params_dict = dict(self.named_parameters()) params_dict = self._cached_params_dict for name, loaded_weight in weights: if not name.startswith("model.") and ( name.startswith("layers.") or name.startswith("embed_tokens.") or name.startswith("norm.") ): name = add_prefix(name, "model") if name == "model.embed_tokens.weight": if self.pp_group.is_last_rank and getattr( self.config, "tie_word_embeddings", False ): if "lm_head.weight" in params_dict: param = params_dict["lm_head.weight"] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) layer_id = get_layer_id(name) if ( layer_id is not None and hasattr(self.model, "start_layer") and ( layer_id < self.model.start_layer or layer_id >= self.model.end_layer ) ): continue if "rotary_emb.inv_freq" in name or "projector" in name: continue if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: continue if "scale" in name: name = maybe_remap_kv_scale_name(name, params_dict) if name is None: continue for param_name, weight_name, shard_id in stacked_params_mapping: if weight_name not in name: continue if "mlp.experts" in name: continue name2 = name.replace(weight_name, param_name) if name2.endswith(".bias") and name2 not in params_dict: continue if name2 not in params_dict: continue param = params_dict[name2] weight_loader = getattr(param, "weight_loader", default_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, shard_id = mapping if weight_name not in name: continue is_expert_weight = True name2 = name.replace(weight_name, param_name) if name2 not in params_dict: continue param = params_dict[name2] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader( param, loaded_weight, name2, shard_id=shard_id, expert_id=expert_id, ) break else: if is_expert_weight: continue # 3) regular params 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) if not hasattr(self, "routed_experts_weights_of_layer"): self.routed_experts_weights_of_layer = LazyValue( lambda: { lid: self.model.layers[lid].mlp.get_moe_weights() for lid in range(self.start_layer, self.end_layer) if isinstance(self.model.layers[lid].mlp, SDARMoeSparseMoeBlock) } ) @classmethod def get_model_config_for_expert_location(cls, config): return ModelConfigForExpertLocation( num_layers=config.num_hidden_layers, num_logical_experts=config.num_experts, num_groups=None, ) EntryClass = SDARMoeForCausalLM