# Copyright 2025-2026 SGLang Team # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Inference-only GLM-4.5, GLM-4.6 and GLM-4.7 model compatible with HuggingFace weights""" import logging import re from typing import Any, Dict, Iterable, List, Optional, Tuple, Union import torch import torch.nn.functional as F from torch import nn from transformers import PretrainedConfig from sglang.srt.batch_overlap.single_batch_overlap import SboFlags from sglang.srt.batch_overlap.two_batch_overlap import model_forward_maybe_tbo from sglang.srt.distributed import ( get_pp_group, get_pp_indices, parallel_state, tensor_model_parallel_all_reduce, ) from sglang.srt.distributed.device_communicators.pynccl_allocator import ( use_symmetric_memory, ) from sglang.srt.environ import envs 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.activation import SiluAndMul from sglang.srt.layers.communicator import ( LayerCommunicator, LayerScatterModes, enable_moe_dense_fully_dp, ) from sglang.srt.layers.dp_attention import ( is_allocation_symmetric, is_dp_attention_enabled, ) from sglang.srt.layers.layernorm import RMSNorm from sglang.srt.layers.linear import ( MergedColumnParallelLinear, QKVParallelLinear, 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, should_use_flashinfer_cutlass_moe_fp4_allgather, ) 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.kt_ep_wrapper import KTEPWrapperMethod 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.quantization.fp8_kernel import is_fp8_fnuz from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.rotary_embedding import get_rope from sglang.srt.layers.utils import PPMissingLayer 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_executor.runner import get_is_capture_mode from sglang.srt.model_loader.weight_utils import default_weight_loader from sglang.srt.models.deepseek_nextn import DeepseekV3ForCausalLMNextN from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM from sglang.srt.models.utils import WeightsMapper, apply_qk_norm from sglang.srt.runtime_context import ( get_forward, get_parallel, get_server_args, get_stream, ) from sglang.srt.utils import ( add_prefix, cpu_has_amx_support, get_bool_env_var, get_device_sm, is_cpu, is_cuda, is_hip, is_non_idle_and_non_empty, is_npu, log_info_on_rank0, make_layers, ) from sglang.srt.utils.hf_transformers_utils import get_rope_config _is_hip = is_hip() _is_cuda = is_cuda() _is_fp8_fnuz = is_fp8_fnuz() _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip _is_cpu_amx_available = cpu_has_amx_support() _is_cpu = is_cpu() _is_npu = is_npu() _device_sm = get_device_sm() logger = logging.getLogger(__name__) if _is_npu: from sgl_kernel_npu.norm.split_qkv_rmsnorm_rope import split_qkv_rmsnorm_rope from sglang.srt.hardware_backend.npu.utils import ( process_shared_expert, wait_share_stream, ) class Glm4MoeMLP(nn.Module): def __init__( self, hidden_size: int, intermediate_size: int, hidden_act: str, quant_config: Optional[QuantizationConfig] = None, reduce_results: bool = True, prefix: str = "", tp_rank: Optional[int] = None, tp_size: Optional[int] = None, ) -> None: super().__init__() self.tp_size = tp_size self.gate_up_proj = MergedColumnParallelLinear( hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config, prefix=add_prefix("gate_up_proj", prefix), tp_rank=tp_rank, tp_size=tp_size, ) self.down_proj = RowParallelLinear( intermediate_size, hidden_size, bias=False, quant_config=quant_config, reduce_results=reduce_results, prefix=add_prefix("down_proj", prefix), tp_rank=tp_rank, tp_size=tp_size, ) 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, forward_batch=None, ): if (self.tp_size == 1) and x.shape[0] == 0: return x gate_up, _ = self.gate_up_proj(x) x = self.act_fn(gate_up) x, _ = self.down_proj(x) return x class Glm4MoeAttention(nn.Module): def __init__( self, hidden_size: int, num_heads: int, num_kv_heads: int, layer_id: int = 0, start_layer: int = 0, rope_theta: float = 1000000, partial_rotary_factor: float = 0.5, rope_scaling: Optional[Dict[str, Any]] = None, max_position_embeddings: int = 8192, head_dim: Optional[int] = None, rms_norm_eps: float = 1e-05, attention_bias: bool = True, quant_config: Optional[QuantizationConfig] = None, use_qk_norm: bool = False, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ) -> None: super().__init__() self.hidden_size = hidden_size self.start_layer = start_layer attn_tp_rank = get_parallel().attn_tp_rank attn_tp_size = get_parallel().attn_tp_size self.total_num_heads = num_heads assert self.total_num_heads % attn_tp_size == 0 self.num_heads = self.total_num_heads // attn_tp_size self.total_num_kv_heads = num_kv_heads if self.total_num_kv_heads >= attn_tp_size: # Number of KV heads is greater than TP size, so we partition # the KV heads across multiple tensor parallel GPUs. assert self.total_num_kv_heads % attn_tp_size == 0 else: # Number of KV heads is less than TP size, so we replicate # the KV heads across multiple tensor parallel GPUs. 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 = head_dim or 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.scaling = self.head_dim**-0.5 self.rope_theta = rope_theta self.use_qk_norm = use_qk_norm self.max_position_embeddings = max_position_embeddings self.tp_rank = get_parallel().tp_rank self.qkv_proj = QKVParallelLinear( hidden_size, self.head_dim, self.total_num_heads, self.total_num_kv_heads, bias=attention_bias, quant_config=quant_config, tp_rank=attn_tp_rank, tp_size=attn_tp_size, prefix=add_prefix("qkv_proj", prefix), ) self.o_proj = RowParallelLinear( self.total_num_heads * self.head_dim, hidden_size, bias=False, quant_config=quant_config, tp_rank=attn_tp_rank, tp_size=attn_tp_size, reduce_results=False, prefix=add_prefix("o_proj", prefix), ) self.rotary_emb = get_rope( self.head_dim, rotary_dim=self.head_dim, max_position=max_position_embeddings, partial_rotary_factor=partial_rotary_factor, base=rope_theta, rope_scaling=rope_scaling, ) self.attn = RadixAttention( self.num_heads, self.head_dim, self.scaling, num_kv_heads=self.num_kv_heads, layer_id=layer_id, prefix=add_prefix("attn", prefix), ) if self.use_qk_norm: self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) self.alt_stream = alt_stream def op_prepare(self, state): state.attn_intermediate_state = self.forward_prepare( positions=state.positions, hidden_states=state.pop("hidden_states_after_comm_pre_attn"), forward_batch=state.forward_batch, ) def op_core(self, state): state.hidden_states_after_attn = self.forward_core( state.pop("attn_intermediate_state") ) def forward_prepare( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ): # hidden_states can be a (fp8_tensor, scale) tuple from fused RMSNorm+Quant hs = hidden_states[0] if isinstance(hidden_states, tuple) else hidden_states if hs.shape[0] == 0: return hidden_states, forward_batch, None qkv, _ = self.qkv_proj(hidden_states) if ( not _is_npu or forward_batch.forward_mode.is_extend_or_draft_extend_or_mixed() ): q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) if self.use_qk_norm: 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) else: if self.attn.layer_id == self.start_layer: self.rotary_emb.get_cos_sin_with_position(positions) if self.use_qk_norm: eps = self.q_norm.variance_epsilon q_weight = self.q_norm.weight k_weight = self.k_norm.weight q_bias = getattr(self.q_norm, "bias", None) k_bias = getattr(self.k_norm, "bias", None) else: eps = None q_weight = None k_weight = None q_bias = None k_bias = None q, k, v = split_qkv_rmsnorm_rope( qkv, self.rotary_emb.position_sin, self.rotary_emb.position_cos, self.q_size, self.kv_size, self.head_dim, eps=eps, q_weight=q_weight, k_weight=k_weight, q_bias=q_bias, k_bias=k_bias, ) inner_state = q, k, v, forward_batch return None, forward_batch, inner_state def forward_core(self, intermediate_state): hidden_states, forward_batch, inner_state = intermediate_state if inner_state is None: return hidden_states attn_output = self.attn(*inner_state) output, _ = self.o_proj(attn_output) return output def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, ) -> torch.Tensor: s = self.forward_prepare( positions=positions, hidden_states=hidden_states, forward_batch=forward_batch, ) return self.forward_core(s) class Glm4MoeGate(nn.Module): def __init__( self, config, prefix: str = "", ): super().__init__() self.weight = nn.Parameter( torch.empty((config.n_routed_experts, config.hidden_size)) ) self.e_score_correction_bias = nn.Parameter( torch.empty((config.n_routed_experts), dtype=torch.float32) ) # GLM requires FP32 gate projection; cache to avoid per-forward cast. # FIXME: if gate weight is updated at runtime (e.g. expert rebalancing), _weight_fp32 must be invalidated. self.register_buffer("_weight_fp32", None, persistent=False) def forward(self, hidden_states): if self._weight_fp32 is None: self._weight_fp32 = self.weight.data.to(torch.float32) logits = F.linear(hidden_states.to(torch.float32), self._weight_fp32, None) return logits class Glm4MoeSparseMoeBlock(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ): nn.Module.__init__(self) self.top_k = config.num_experts_per_tok self.tp_size = get_parallel().tp_size self.moe_ep_size = get_parallel().moe_ep_size self.routed_scaling_factor = config.routed_scaling_factor self.n_shared_experts = config.n_shared_experts self.num_fused_shared_experts = ( 0 if get_server_args().disable_shared_experts_fusion else config.n_shared_experts ) self.config = config self.layer_id = layer_id self.alt_stream = alt_stream if self.tp_size > config.n_routed_experts: raise ValueError( f"Tensor parallel size {self.tp_size} is greater than " f"the number of experts {config.n_routed_experts}." ) if config.hidden_act != "silu": raise ValueError( f"Unsupported activation: {config.hidden_act}. " "Only silu is supported for now." ) self.gate = Glm4MoeGate(config=config, prefix=add_prefix("gate", prefix)) self.experts = get_moe_impl_class(quant_config)( num_experts=config.n_routed_experts + self.num_fused_shared_experts, num_fused_shared_experts=self.num_fused_shared_experts, top_k=self.top_k + self.num_fused_shared_experts, layer_id=self.layer_id, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, quant_config=quant_config, routed_scaling_factor=self.routed_scaling_factor, routing_method_type=RoutingMethodType.DeepSeekV3, prefix=add_prefix("experts", prefix), ) self.topk = TopK( top_k=self.top_k + self.num_fused_shared_experts, layer_id=self.layer_id, renormalize=config.norm_topk_prob, use_grouped_topk=True, num_expert_group=config.n_group, topk_group=config.topk_group, correction_bias=self.gate.e_score_correction_bias, routed_scaling_factor=self.routed_scaling_factor, num_fused_shared_experts=self.num_fused_shared_experts, apply_routed_scaling_factor_on_output=getattr( self.experts, "should_fuse_routed_scaling_factor_in_topk", False ), fused_shared_experts_scaling_factor=1, ) self.shared_experts_is_int8 = False self.shared_experts_is_fp8 = False self.shared_experts_weight_block_size = None if config.n_shared_experts is not None and self.num_fused_shared_experts == 0: intermediate_size = config.moe_intermediate_size * config.n_shared_experts # disable tp for shared experts when enable deepep moe, or with fp4 allgather self.shared_experts = Glm4MoeMLP( hidden_size=config.hidden_size, intermediate_size=intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, reduce_results=False, prefix=add_prefix("shared_experts", prefix), **( dict(tp_rank=0, tp_size=1) if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake() or get_moe_a2a_backend().is_nixl() or get_moe_a2a_backend().is_mori() or get_moe_a2a_backend().is_ascend_fuseep() or get_moe_a2a_backend().is_flashinfer() or should_use_flashinfer_cutlass_moe_fp4_allgather() else {} ), ) is_packed_weight = hasattr( self.shared_experts.gate_up_proj.quant_method, "quant_config" ) and self.shared_experts.gate_up_proj.quant_method.quant_config.get_name() in { "awq", "awq_marlin", "moe_wna16", } self.shared_experts_is_int8 = ( not is_packed_weight and self.shared_experts.gate_up_proj.weight.dtype == torch.int8 ) self.shared_experts_is_fp8 = ( not is_packed_weight and self.shared_experts.gate_up_proj.weight.dtype == torch.float8_e4m3fn ) if self.shared_experts_is_fp8: if ( _use_aiter and config.quantization_config.get("quant_method") == "compressed-tensors" ): # For compressed-tensors ptpc model, don't need to check the weight_block_size pass else: assert ( self.shared_experts.gate_up_proj.quant_method.quant_config.weight_block_size == self.shared_experts.down_proj.quant_method.quant_config.weight_block_size ) self.shared_experts_weight_block_size = ( self.shared_experts.gate_up_proj.quant_method.quant_config.weight_block_size ) self.top_k = config.num_experts_per_tok if ( get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake() or get_moe_a2a_backend().is_nixl() or get_moe_a2a_backend().is_mori() or get_moe_a2a_backend().is_ascend_fuseep() ): # TODO: we will support tp < ep in the future self.ep_size = get_parallel().moe_ep_size self.num_experts = ( config.n_routed_experts + get_server_args().ep_num_redundant_experts ) self.renormalize = config.norm_topk_prob self.topk_group = config.topk_group self.num_expert_group = config.n_group self.correction_bias = ( self.gate.e_score_correction_bias.data if self.gate.e_score_correction_bias is not None else None ) self._enable_a2a_moe = ( get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake() or get_moe_a2a_backend().is_nixl() or get_moe_a2a_backend().is_mori() or get_moe_a2a_backend().is_ascend_fuseep() or get_moe_a2a_backend().is_flashinfer() ) self._fuse_shared_experts_inside_sbo = SboFlags.fuse_shared_experts_inside_sbo() def get_moe_weights(self): return [ x.data for name, x in self.experts.named_parameters() if name not in ["correction_bias"] and filter_moe_weight_param_global_expert( name, x, self.experts.num_local_experts ) ] def forward( self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None, ) -> torch.Tensor: if not self._enable_a2a_moe: if ( self.alt_stream is not None and self.num_fused_shared_experts == 0 and hidden_states.shape[0] > 0 and get_is_capture_mode() ): return self.forward_normal_dual_stream(hidden_states) else: return self.forward_normal(hidden_states) else: return self.forward_deepep(hidden_states, forward_batch) def forward_normal_dual_stream( self, hidden_states: torch.Tensor, ) -> torch.Tensor: current_stream = torch.cuda.current_stream() self.alt_stream.wait_stream(current_stream) shared_output = self._forward_shared_experts(hidden_states) with torch.cuda.stream(self.alt_stream): # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) topk_output = self.topk(hidden_states, router_logits) final_hidden_states = self.experts(hidden_states, topk_output) if not _is_cuda or isinstance(self.experts.quant_method, KTEPWrapperMethod): final_hidden_states *= self.routed_scaling_factor current_stream.wait_stream(self.alt_stream) final_hidden_states += shared_output if self.tp_size > 1 and not should_skip_post_experts_all_reduce( is_tp_path=True, ): final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states) return final_hidden_states def forward_normal( self, hidden_states: torch.Tensor, ) -> torch.Tensor: if hidden_states.shape[0] > 0: shared_output = self._forward_shared_experts(hidden_states) # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) topk_output = self.topk(hidden_states, router_logits) else: shared_output = None topk_output = self.topk.empty_topk_output(hidden_states.device) final_hidden_states = self.experts(hidden_states, topk_output) if not _is_cuda and not _use_aiter: final_hidden_states *= self.routed_scaling_factor if shared_output is not None: with use_symmetric_memory( parallel_state.get_tp_group(), disabled=not is_allocation_symmetric() ): final_hidden_states_out = torch.empty_like(final_hidden_states) torch.add(final_hidden_states, shared_output, out=final_hidden_states_out) final_hidden_states = final_hidden_states_out if self.tp_size > 1 and not should_skip_post_experts_all_reduce( is_tp_path=True, ): final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states) return final_hidden_states def forward_deepep( self, hidden_states: torch.Tensor, forward_batch: ForwardBatch ) -> torch.Tensor: shared_output = None enable_npu_dual_stream = ( _is_npu and ( forward_batch.forward_mode.is_extend() or forward_batch.forward_mode.is_target_verify() ) and envs.SGLANG_NPU_USE_MULTI_STREAM.get() ) if hidden_states.shape[0] > 0: # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) if enable_npu_dual_stream: shared_output = process_shared_expert( hidden_states, self._forward_shared_experts ) else: shared_output = self._forward_shared_experts(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) final_hidden_states = self.experts( hidden_states=hidden_states, topk_output=topk_output, ) if enable_npu_dual_stream: wait_share_stream() if shared_output is not None: x = shared_output if self.experts.should_fuse_routed_scaling_factor_in_topk: x.add_(final_hidden_states) else: x.add_(final_hidden_states, alpha=self.routed_scaling_factor) final_hidden_states = x else: if not self.experts.should_fuse_routed_scaling_factor_in_topk: final_hidden_states *= self.routed_scaling_factor return final_hidden_states def _forward_shared_experts(self, hidden_states: torch.Tensor): if (hidden_states.shape[0] > 0) and (self.num_fused_shared_experts == 0): return self.shared_experts(hidden_states) else: return None def op_gate(self, state): if is_non_idle_and_non_empty( state.forward_batch.forward_mode, state.hidden_states_mlp_input ): # router_logits: (num_tokens, n_experts) state.router_logits = self.gate(state.hidden_states_mlp_input) else: state.router_logits = None def op_select_experts(self, state): router_logits = state.pop("router_logits") hidden_states = state.hidden_states_mlp_input if router_logits is not None: with get_global_expert_distribution_recorder().with_current_layer( self.layer_id ): state.topk_output = self.topk( hidden_states=hidden_states, router_logits=router_logits, num_token_non_padded=state.forward_batch.num_token_non_padded, expert_location_dispatch_info=ExpertLocationDispatchInfo.init_new( layer_id=self.layer_id, ), ) else: state.topk_output = self.topk.empty_topk_output(hidden_states.device) def op_dispatch_a(self, state): if self.ep_size > 1: self.experts.dispatcher.dispatch_a( hidden_states=state.hidden_states_mlp_input, topk_output=state.pop("topk_output"), tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_dispatch_b(self, state): if self.ep_size > 1: with get_global_expert_distribution_recorder().with_current_layer( self.layer_id ): state.dispatch_output = self.experts.dispatcher.dispatch_b( tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_experts(self, state): state.combine_input = self.experts.run_moe_core( dispatch_output=state.dispatch_output, ) def op_combine_a(self, state): if self.ep_size > 1: self.experts.dispatcher.combine_a( combine_input=state.pop("combine_input"), tbo_subbatch_index=state.get("tbo_subbatch_index"), ) state.pop("dispatch_output") def op_combine_b(self, state): if self.ep_size > 1: state.hidden_states_after_combine = self.experts.dispatcher.combine_b( tbo_subbatch_index=state.get("tbo_subbatch_index"), ) def op_output(self, state): final_hidden_states = state.pop("hidden_states_after_combine") if (shared_output := state.pop("shared_output")) is not None: x = shared_output x.add_(final_hidden_states, alpha=self.routed_scaling_factor) final_hidden_states = x else: final_hidden_states *= self.routed_scaling_factor state.hidden_states_mlp_output = final_hidden_states class Glm4MoeDecoderLayer(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, start_layer: int = 0, quant_config: Optional[QuantizationConfig] = None, is_nextn: bool = False, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ) -> None: nn.Module.__init__(self) self.hidden_size = config.hidden_size self.config = config rope_theta, rope_scaling = get_rope_config(config) partial_rotary_factor = (rope_scaling or {}).get("partial_rotary_factor") if partial_rotary_factor is None: partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5) max_position_embeddings = getattr(config, "max_position_embeddings", 8192) head_dim = getattr( config, "head_dim", config.hidden_size // config.num_attention_heads ) rms_norm_eps = config.rms_norm_eps attention_bias = config.attention_bias self.layer_id = layer_id use_qk_norm = config.use_qk_norm if hasattr(config, "use_qk_norm") else False self.self_attn = Glm4MoeAttention( hidden_size=self.hidden_size, num_heads=config.num_attention_heads, num_kv_heads=config.num_key_value_heads, layer_id=layer_id, start_layer=start_layer, rope_theta=rope_theta, rope_scaling=rope_scaling, partial_rotary_factor=partial_rotary_factor, max_position_embeddings=max_position_embeddings, head_dim=head_dim, rms_norm_eps=rms_norm_eps, attention_bias=attention_bias, quant_config=quant_config, prefix=add_prefix("self_attn", prefix), use_qk_norm=use_qk_norm, alt_stream=alt_stream, ) self.is_layer_sparse = self._is_layer_sparse(layer_id, is_nextn=is_nextn) is_previous_layer_sparse = self._is_layer_sparse(layer_id - 1, is_nextn=False) is_next_layer_sparse = self._is_layer_sparse(layer_id + 1, is_nextn=False) self.layer_scatter_modes = LayerScatterModes.init_new( layer_id=layer_id, num_layers=1 if is_nextn else config.num_hidden_layers, is_layer_sparse=self.is_layer_sparse, is_previous_layer_sparse=is_previous_layer_sparse, is_next_layer_sparse=is_next_layer_sparse, ) if self.is_layer_sparse: self.mlp = Glm4MoeSparseMoeBlock( config=config, quant_config=quant_config, prefix=add_prefix("mlp", prefix), layer_id=self.layer_id, alt_stream=alt_stream, ) else: if enable_moe_dense_fully_dp(): mlp_tp_rank, mlp_tp_size = 0, 1 else: mlp_tp_rank, mlp_tp_size = None, None self.mlp = Glm4MoeMLP( hidden_size=config.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, prefix=add_prefix("mlp", prefix), tp_rank=mlp_tp_rank, tp_size=mlp_tp_size, ) self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = RMSNorm( config.hidden_size, eps=config.rms_norm_eps ) 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=( is_nextn or (self.layer_id == self.config.num_hidden_layers - 1) ), ) # Detect if QKV uses aiter FP8 per-token quant so we can fuse # RMSNorm + FP8 quant into a single kernel in prepare_attn self.attn_quant_format = "" self._detect_attn_quant_format() def _detect_fp8_per_token_quant(self, linear_layer, label: str) -> str: """Check if a linear layer uses aiter FP8 per-token quantization.""" from sglang.srt.utils import get_bool_env_var, is_hip if not (get_bool_env_var("SGLANG_USE_AITER") and is_hip()): return "" if not hasattr(linear_layer, "quant_method"): return "" scheme = getattr(linear_layer, "scheme", None) or getattr( linear_layer.quant_method, "scheme", None ) if scheme is not None: from compressed_tensors.quantization import QuantizationStrategy from sglang.srt.layers.quantization.compressed_tensors.schemes.compressed_tensors_w8a8_fp8 import ( CompressedTensorsW8A8Fp8, ) if ( isinstance(scheme, CompressedTensorsW8A8Fp8) and scheme.strategy == QuantizationStrategy.CHANNEL ): logger.info( "layer_%d Fused RMSNorm+Quant %s: ENABLED (fp8_per_token)", self.layer_id, label, ) return "fp8_per_token" logger.info( "layer_%d Fused RMSNorm+Quant %s: skipped", self.layer_id, label, ) return "" def _detect_attn_quant_format(self): self.attn_quant_format = self._detect_fp8_per_token_quant( self.self_attn.qkv_proj, "attn" ) def _is_layer_sparse(self, layer_id: int, is_nextn: bool) -> bool: return is_nextn or ( self.config.n_routed_experts is not None and layer_id >= self.config.first_k_dense_replace ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], ) -> torch.Tensor: hidden_states, residual = self.layer_communicator.prepare_attn( hidden_states, residual, forward_batch, quant_format=self.attn_quant_format, ) 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 ) ) # For DP with padding, reduce scatter can be used instead of all-reduce. 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) 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 def op_comm_prepare_attn( self, state, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], tbo_subbatch_index: Optional[int] = None, ): state.hidden_states_after_comm_pre_attn, state.residual_after_input_ln = ( self.layer_communicator.prepare_attn( hidden_states, residual, forward_batch, quant_format=self.attn_quant_format, ) ) state.update( dict( forward_batch=forward_batch, positions=positions, tbo_subbatch_index=tbo_subbatch_index, ) ) def op_comm_prepare_mlp(self, state): state.hidden_states_mlp_input, state.residual_after_comm_pre_mlp = ( self.layer_communicator.prepare_mlp( state.pop("hidden_states_after_attn"), state.pop("residual_after_input_ln"), state.forward_batch, ) ) def op_comm_postprocess_layer(self, state): hidden_states, residual = self.layer_communicator.postprocess_layer( state.pop("hidden_states_mlp_output"), state.pop("residual_after_comm_pre_mlp"), state.forward_batch, ) output = dict( positions=state.positions, hidden_states=hidden_states, residual=residual, forward_batch=state.forward_batch, tbo_subbatch_index=state.tbo_subbatch_index, ) state.clear( expect_keys={ "positions", "forward_batch", "tbo_subbatch_index", } ) return output class Glm4MoeModel(nn.Module): def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.pp_group = get_pp_group() self.config = config self.vocab_size = config.vocab_size self.first_k_dense_replace = config.first_k_dense_replace self.embed_dim = config.hidden_size if self.pp_group.is_first_rank: self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, use_attn_tp_group=is_dp_attention_enabled(), ) else: self.embed_tokens = PPMissingLayer() self.alt_stream = get_stream("alt") if _is_cuda else None pp_start_layer, _ = get_pp_indices( config.num_hidden_layers, self.pp_group.rank_in_group, self.pp_group.world_size, ) self.layers, self.start_layer, self.end_layer = make_layers( config.num_hidden_layers, lambda idx, prefix: Glm4MoeDecoderLayer( layer_id=idx, start_layer=pp_start_layer, config=config, quant_config=quant_config, prefix=prefix, alt_stream=self.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: self.norm = RMSNorm(self.embed_dim, eps=config.rms_norm_eps) else: self.norm = PPMissingLayer(return_tuple=True) self.layers_to_capture = [] def get_input_embeddings(self) -> torch.Tensor: return self.embed_tokens 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: if input_embeds is None: hidden_states = self.embed_tokens(input_ids) else: hidden_states = input_embeds residual = None else: assert pp_proxy_tensors is not None hidden_states = pp_proxy_tensors["hidden_states"] residual = pp_proxy_tensors["residual"] normal_start_layer = self.start_layer normal_end_layer = self.end_layer if forward_batch.can_run_tbo: if ( self.first_k_dense_replace > normal_start_layer and self.first_k_dense_replace < normal_end_layer ): normal_end_layer = self.first_k_dense_replace elif self.first_k_dense_replace < normal_start_layer: normal_end_layer = normal_start_layer = 0 aux_hidden_states = [] for i in range(normal_start_layer, normal_end_layer): with get_global_expert_distribution_recorder().with_current_layer(i): if i in self.layers_to_capture: aux_hidden_states.append(hidden_states + residual) layer = self.layers[i] hidden_states, residual = layer( positions, hidden_states, forward_batch, residual, ) if normal_end_layer != self.end_layer: hidden_states, residual = model_forward_maybe_tbo( layers=self.layers[normal_end_layer : self.end_layer], enable_tbo=True, positions=positions, forward_batch=forward_batch, hidden_states=hidden_states, residual=residual, input_data_scatter_mode=self.layers[ normal_end_layer - 1 ].layer_scatter_modes.layer_output_mode, ) if not self.pp_group.is_last_rank: return PPProxyTensors( { "hidden_states": hidden_states, "residual": residual, } ) else: if not forward_batch.forward_mode.is_idle(): if residual is None: hidden_states = self.norm(hidden_states) else: hidden_states, _ = self.norm(hidden_states, residual) if len(aux_hidden_states) == 0: return hidden_states return hidden_states, aux_hidden_states class Glm4MoeForCausalLM(nn.Module): def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: nn.Module.__init__(self) self.pp_group = get_pp_group() self.config = config self.tp_size = get_parallel().tp_size self.quant_config = quant_config self.num_fused_shared_experts = 0 self.determine_num_fused_shared_experts() self.model = Glm4MoeModel( config, quant_config, prefix=add_prefix("model", prefix) ) self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=add_prefix("lm_head", prefix), use_attn_tp_group=get_server_args().enable_dp_lm_head, ) self.logits_processor = LogitsProcessor(config) # For EAGLE3 support self.capture_aux_hidden_states = False def determine_num_fused_shared_experts(self): if get_server_args().disable_shared_experts_fusion: return disable_reason = None if (not _is_cuda or torch.cuda.get_device_capability("cuda") < (8, 0)) and ( not _is_hip or torch.cuda.get_device_capability("cuda") < (9, 4) ): disable_reason = ( "Only GLM-4.5 on NV-platform with capability >= 80 " "or AMD-platform with capability >= gfx942(MI30x) can use shared experts fusion optimization." ) elif get_parallel().moe_ep_size > 1 and ( not _is_hip or torch.cuda.get_device_capability("cuda") < (9, 4) ): disable_reason = "Only GLM-4.5 on AMD-platform with capability >= gfx942(MI30x) can use shared experts fusion optimization under expert parallelism." elif disable_reason is None and ( get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mori() ): disable_reason = "GLM-4.5 cannot use shared experts fusion optimization under deepep expert parallelism." elif self.quant_config and self.quant_config.get_name() == "w4afp8": disable_reason = "GLM-4.5 W4AFP8 model uses different quant method for routed experts and shared experts." if disable_reason is not None: from sglang.srt.arg_groups.overrides import declare_load_time_override declare_load_time_override( "Glm4MoeForCausalLM.determine_num_fused_shared_experts", {"disable_shared_experts_fusion": True}, ) self.num_fused_shared_experts = 0 log_info_on_rank0( logger, f"{disable_reason} Shared experts fusion optimization is disabled.", ) return self.num_fused_shared_experts = self.config.n_shared_experts def get_input_embeddings(self) -> nn.Embedding: return self.model.embed_tokens @torch.no_grad() def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, forward_batch: ForwardBatch, input_embeds: torch.Tensor = None, pp_proxy_tensors: Optional[PPProxyTensors] = None, ) -> torch.Tensor: hidden_states = self.model( input_ids, positions, forward_batch, input_embeds, pp_proxy_tensors ) aux_hidden_states = None if self.capture_aux_hidden_states: hidden_states, aux_hidden_states = hidden_states if self.pp_group.is_last_rank: return self.logits_processor( input_ids, hidden_states, self.lm_head, forward_batch, aux_hidden_states ) else: return hidden_states @property def start_layer(self): return self.model.start_layer @property def end_layer(self): return self.model.end_layer def load_weights( self, weights: Iterable[Tuple[str, torch.Tensor]], is_nextn=False, params_dict=None, ): if is_nextn: if hasattr(self.config, "num_nextn_predict_layers"): num_nextn_layers = self.config.num_nextn_predict_layers assert num_nextn_layers == 1, "Only 1 nextn layer is supported" # compatible with old design nextn_layer_id = ( 0 if self.config.num_hidden_layers == 1 else self.config.num_hidden_layers ) else: raise ValueError("num_nextn_predict_layers is not in the config") stacked_params_mapping = [ # (param_name, shard_name, shard_id) ("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), ] if self.num_fused_shared_experts > 0: assert self.num_fused_shared_experts == 1 def iter_weights_with_fused_shared_experts( weights: Iterable[Tuple[str, torch.Tensor]], ) -> Iterable[Tuple[str, torch.Tensor]]: pattern = re.compile( r"^model\.layers\.(\d+)\.mlp\.shared_experts\.(.+)$" ) for name, weight in weights: match = pattern.match(name) if match: layer_id = int(match.group(1)) suffix = match.group(2) name = f"model.layers.{layer_id}.mlp.experts.{self.config.n_routed_experts}.{suffix}" yield name, weight weights = iter_weights_with_fused_shared_experts(weights) # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) 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.n_routed_experts + self.num_fused_shared_experts, ) if is_nextn: nextn_layer_prefix = f"model.layers.{nextn_layer_id}" nextn_spec_weight_names = [ "shared_head.norm", "eh_proj", "enorm", "hnorm", ] else: nextn_layer_prefix = None nextn_spec_weight_names = [] if params_dict is None: params_dict = dict(self.named_parameters()) weight_names = [] for name, loaded_weight in weights: weight_names.append(name) if not is_nextn: if hasattr(self.config, "num_nextn_predict_layers"): num_nextn_layers = self.config.num_nextn_predict_layers if num_nextn_layers > 0 and name.startswith("model.layers"): name_list = name.split(".") if ( len(name_list) >= 3 and int(name_list[2]) >= self.config.num_hidden_layers ): continue else: if nextn_layer_prefix and not name.startswith(nextn_layer_prefix): continue if nextn_layer_prefix is not None: # mtp # Use shared head and embed weights from target model if "shared_head.head" in name or "embed_tokens" in name: continue is_decoder = True # For nextn specific weights for weight_name in nextn_spec_weight_names: if weight_name in name: name = name.replace(nextn_layer_prefix, "model") is_decoder = False break # For decoder layer weights if is_decoder: name = name.replace(nextn_layer_prefix, "model.decoder") if "rotary_emb.inv_freq" in name: continue for param_name, weight_name, shard_id in stacked_params_mapping: # Skip non-stacked layers and experts (experts handled below). if weight_name not in name: continue # We have mlp.experts[0].gate_proj in the checkpoint. # Since we handle the experts below in expert_params_mapping, # we need to skip here BEFORE we update the name, otherwise # name will be updated to mlp.experts[0].gate_up_proj, which # will then be updated below in expert_params_mapping # for mlp.experts[0].gate_gate_up_proj, which breaks load. if "mlp.experts" in name: continue name = name.replace(weight_name, param_name) # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue if name not in params_dict: continue param = params_dict[name] weight_loader = param.weight_loader weight_loader(param, loaded_weight, shard_id) break else: # Track if this is an expert weight to enable early skipping 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 # Mark as expert weight regardless of whether we can process it is_expert_weight = True name = name.replace(weight_name, param_name) if name not in params_dict: # Expert weight not on this rank, will be skipped below continue param = params_dict[name] weight_loader = param.weight_loader weight_loader( param, loaded_weight, name, shard_id=shard_id, expert_id=expert_id, ) break else: if is_expert_weight: # This is an expert weight but not mapped to this rank, skip all remaining processing continue # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue if name not in params_dict: continue if name in params_dict.keys(): param = params_dict[name] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) else: logger.warning(f"Parameter {name} not found in params_dict") def get_embed_and_head(self): return self.model.embed_tokens.weight, self.lm_head.weight def set_embed_and_head(self, embed, head): del self.model.embed_tokens.weight del self.lm_head.weight self.model.embed_tokens.weight = embed self.lm_head.weight = head torch.cuda.empty_cache() torch.cuda.synchronize() @classmethod def get_model_config_for_expert_location(cls, config): return ModelConfigForExpertLocation( num_layers=config.num_hidden_layers, num_logical_experts=config.n_routed_experts, num_groups=config.n_group, ) def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None): if not self.pp_group.is_last_rank: return if layer_ids is None: self.capture_aux_hidden_states = True num_layers = self.config.num_hidden_layers self.model.layers_to_capture = [2, num_layers // 2, num_layers - 3] else: self.capture_aux_hidden_states = True # we plus 1 here because in sglang, for the ith layer, it takes the output # of the (i-1)th layer as aux hidden state self.model.layers_to_capture = [val + 1 for val in layer_ids] class GlmMoeDsaForCausalLM(DeepseekV2ForCausalLM): def determine_num_fused_shared_experts(self): super().determine_num_fused_shared_experts("GlmMoeDsaForCausalLM") class GlmMoeDsaForCausalLMNextN(DeepseekV3ForCausalLMNextN): # GLM-5.2's MTP layer index differs from DeepSeek's (61), so the inherited # substr mapping would wrongly rewrite GLM's real layer-61 weights. # exclude_layers remapping for the MTP layer is handled explicitly in # _resolve_nextn_quant_config below instead. hf_to_sglang_mapper = WeightsMapper() _NEXTN_SPEC_WEIGHT_NAMES = ("shared_head.norm", "eh_proj", "enorm", "hnorm") @classmethod def _map_mtp_ckpt_name(cls, name: str, layer_prefix: str) -> str: # Keep this mapping in sync with DeepseekV2WeightLoaderMixin's # NextN rule: MTP-specific weights live under model.*, while the # decoder block weights live under model.decoder.*. if any(part in name for part in cls._NEXTN_SPEC_WEIGHT_NAMES): return name.replace(layer_prefix, "model", 1) return name.replace(layer_prefix, "model.decoder", 1) def _resolve_nextn_quant_config(self, config, quant_config): if quant_config is None or quant_config.get_name() != "quark": return quant_config layer_prefix = f"model.layers.{config.num_hidden_layers}" # Quark's per-module scheme selection (e.g. MTP self_attn in PTPC-FP8 # while MTP MoE is MXFP4) is keyed by "layer_quant_config" patterns # using the checkpoint's "model.layers..*" naming. SGLang queries # schemes by the runtime "model.*"/"model.decoder.*" prefix, so those # keys need the same remap as exclude_layers below, or they silently # fall back to the wrong (layer-type/global) scheme. layer_quant_config = quant_config.quant_config.get("layer_quant_config") if layer_quant_config: quant_config.quant_config["layer_quant_config"] = { ( self._map_mtp_ckpt_name(pattern, layer_prefix) if pattern.startswith(layer_prefix + ".") else pattern ): pattern_config for pattern, pattern_config in layer_quant_config.items() } mtp_excluded = [ name for name in quant_config.exclude_layers if name.startswith(layer_prefix + ".") ] if not mtp_excluded: return quant_config names = set(quant_config.exclude_layers) for name in mtp_excluded: names.add(self._map_mtp_ckpt_name(name, layer_prefix)) # Fused routed experts are queried by the coarse module prefix # "model.decoder.mlp.experts". Expanded per-expert leaf excludes do not # match that prefix, so add the coarse prefix when any routed expert in # the MTP layer is excluded. This keeps only that fused MoE module bf16 # while allowing the remaining draft modules to use their quant config. if any(".mlp.experts." in name for name in mtp_excluded): names.add("model.decoder.mlp.experts") import copy quant_config = copy.copy(quant_config) quant_config.exclude_layers = list(names) return quant_config EntryClass = [Glm4MoeForCausalLM, GlmMoeDsaForCausalLM, GlmMoeDsaForCausalLMNextN]