# Copyright 2026-2027 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-4.7-Flash model compatible with HuggingFace weights.""" import logging import re from typing import 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, parallel_state, tensor_model_parallel_all_reduce, ) from sglang.srt.distributed.device_communicators.pynccl_allocator import ( use_symmetric_memory, ) 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, get_attn_tp_context, ) 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, 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, TopKOutputFormat from sglang.srt.layers.moe.utils import filter_moe_weight_param_global_expert from sglang.srt.layers.quantization.base_config import QuantizationConfig 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_common.deepseek_weight_loader import ( DeepseekV2WeightLoaderMixin, ) from sglang.srt.models.deepseek_common.utils import _is_cuda, _use_aiter from sglang.srt.models.deepseek_v2 import DeepseekV2AttentionMLA from sglang.srt.runtime_context import ( get_forward, get_parallel, get_server_args, get_stream, ) from sglang.srt.utils import ( BumpAllocator, LazyValue, add_prefix, is_non_idle_and_non_empty, log_info_on_rank0, make_layers, ) from sglang.srt.utils.hf_transformers_utils import get_rope_config logger = logging.getLogger(__name__) class Glm4MoeLiteMLP(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 Glm4MoeLiteGate(nn.Module): def __init__( self, config, prefix: str = "", is_nextn: bool = False, ): super().__init__() self.is_nextn = is_nextn 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 Glm4MoeLiteSparseMoeBlock(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, is_nextn: bool = False, ): super().__init__() self.tp_size = get_parallel().tp_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 self.is_nextn = is_nextn 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 = Glm4MoeLiteGate( config=config, prefix=add_prefix("gate", prefix), is_nextn=is_nextn ) self.experts = get_moe_impl_class(quant_config)( num_experts=config.n_routed_experts + self.num_fused_shared_experts + get_server_args().ep_num_redundant_experts, num_fused_shared_experts=self.num_fused_shared_experts, top_k=config.num_experts_per_tok + self.num_fused_shared_experts, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, layer_id=self.layer_id, quant_config=quant_config, routed_scaling_factor=self.routed_scaling_factor, prefix=add_prefix("experts", prefix), ) self.topk = TopK( top_k=config.num_experts_per_tok + 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, num_fused_shared_experts=self.num_fused_shared_experts, topk_group=config.topk_group, correction_bias=self.gate.e_score_correction_bias, quant_config=quant_config, routed_scaling_factor=self.routed_scaling_factor, apply_routed_scaling_factor_on_output=self.experts.should_fuse_routed_scaling_factor_in_topk, # Some Fp4 MoE backends require the output format to be bypassed but the MTP layers are unquantized # and requires the output format to be standard. We use quant_config to determine the output format. output_format=TopKOutputFormat.STANDARD if quant_config is None else None, ) self.shared_experts_is_int8 = False self.shared_experts_is_fp8 = False self.shared_experts_weight_block_size = None self._shared_expert_tp1 = False 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 = Glm4MoeLiteMLP( 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 should_use_flashinfer_cutlass_moe_fp4_allgather() else {} ), ) is_packed_weight = hasattr( self.shared_experts.gate_up_proj.quant_method, "quant_config" ) 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 ) self.top_k = config.num_experts_per_tok if get_moe_a2a_backend().is_deepep() or get_moe_a2a_backend().is_mooncake(): # 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() ) 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 if hidden_states.shape[0] > 0: # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states) 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 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_shared_experts(self, state): hidden_states_mlp_input = state.pop("hidden_states_mlp_input") if (self.num_fused_shared_experts == 0) and is_non_idle_and_non_empty( state.forward_batch.forward_mode, hidden_states_mlp_input ): state.shared_output = self.shared_experts(hidden_states_mlp_input) else: state.shared_output = 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 get_moe_a2a_backend().is_mori(): num_tokens = state.pop("num_tokens") final_hidden_states = final_hidden_states[:num_tokens] if (shared_output := state.pop("shared_output")) is not None: x = shared_output if _use_aiter: x.add_(final_hidden_states) else: x.add_(final_hidden_states, alpha=self.routed_scaling_factor) final_hidden_states = x elif _use_aiter: # fused in aiter_biased_grouped_topk so we can skip here pass else: final_hidden_states *= self.routed_scaling_factor state.hidden_states_mlp_output = final_hidden_states class Glm4MoeLiteDecoderLayer(nn.Module): def __init__( self, config: PretrainedConfig, layer_id: int, quant_config: Optional[QuantizationConfig] = None, is_nextn: bool = False, prefix: str = "", alt_stream: Optional[torch.cuda.Stream] = None, ) -> None: super().__init__() # Required for MTP: Glm4MoeLiteModelNextN bypasses Glm4MoeLiteForCausalLM.__init__ config.moe_layer_freq = 1 self.hidden_size = config.hidden_size self.config = config rope_theta, rope_scaling = get_rope_config(config) max_position_embeddings = getattr(config, "max_position_embeddings", 202752) self.layer_id = layer_id self.is_nextn = is_nextn self.self_attn = DeepseekV2AttentionMLA( config=config, hidden_size=config.hidden_size, num_heads=config.num_attention_heads, qk_nope_head_dim=config.qk_nope_head_dim, qk_rope_head_dim=config.qk_rope_head_dim, v_head_dim=config.v_head_dim, q_lora_rank=config.q_lora_rank, kv_lora_rank=config.kv_lora_rank, rope_theta=rope_theta, rope_scaling=rope_scaling, max_position_embeddings=max_position_embeddings, quant_config=quant_config, reduce_results=False, layer_id=layer_id, prefix=add_prefix("self_attn", prefix), ) 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 = Glm4MoeLiteSparseMoeBlock( config=config, quant_config=quant_config, prefix=add_prefix("mlp", prefix), layer_id=self.layer_id, alt_stream=alt_stream, is_nextn=is_nextn, ) 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 = Glm4MoeLiteMLP( 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._gfx95_quant_format = self._detect_gfx95_quant_format() 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) ), qkv_latent_func=self.self_attn.prepare_qkv_latent, ) def _detect_gfx95_quant_format(self) -> str: from sglang.srt.models.deepseek_common.utils import _is_gfx95_supported if not _is_gfx95_supported: return "" weight = getattr( getattr(self.self_attn, "fused_qkv_a_proj_with_mqa", None), "weight", None ) if weight is None: return "" if weight.dtype == torch.uint8: return "mxfp4" if weight.dtype == getattr(torch, "float8_e4m3fn", None): return "fp8" return "" 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 and layer_id % self.config.moe_layer_freq == 0 ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, forward_batch: ForwardBatch, residual: Optional[torch.Tensor], zero_allocator: BumpAllocator, ) -> torch.Tensor: hidden_states, residual = self.layer_communicator.prepare_attn( hidden_states, residual, forward_batch, getattr(self, "_gfx95_quant_format", ""), ) hidden_states = self.self_attn( positions=positions, hidden_states=hidden_states, forward_batch=forward_batch, zero_allocator=zero_allocator, layer_scatter_modes=self.layer_scatter_modes, ) if isinstance(hidden_states, tuple): hidden_states = hidden_states[0] get_attn_tp_context().clear_attn_inputs() 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], zero_allocator: BumpAllocator, 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) ) if get_moe_a2a_backend().is_mori(): state.num_tokens = hidden_states.shape[0] state.update( dict( forward_batch=forward_batch, positions=positions, zero_allocator=zero_allocator, 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, zero_allocator=state.zero_allocator, tbo_subbatch_index=state.tbo_subbatch_index, ) state.clear( expect_keys={ "positions", "forward_batch", "zero_allocator", "tbo_subbatch_index", } ) return output class Glm4MoeLiteModel(nn.Module): fall_back_to_pt_during_load = False def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ): super().__init__() self.padding_id = config.pad_token_id self.vocab_size = config.vocab_size self.first_k_dense_replace = config.first_k_dense_replace self.pp_group = get_pp_group() 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 self.layers, self.start_layer, self.end_layer = make_layers( config.num_hidden_layers, lambda idx, prefix: Glm4MoeLiteDecoderLayer( config=config, layer_id=idx, 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(config.hidden_size, 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]: total_num_layers = self.end_layer - self.start_layer 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"] device = hidden_states.device zero_allocator = BumpAllocator( buffer_size=total_num_layers * 2 * (2 if forward_batch.can_run_tbo else 1), dtype=torch.float32, device=device, ) 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, zero_allocator, ) 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, zero_allocator=zero_allocator, ) 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 Glm4MoeLiteForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin): # for quark model load packed_modules_mapping = {} def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() config.moe_layer_freq = 1 self.config = config self.tp_size = get_parallel().tp_size self.quant_config = quant_config self.pp_group = get_pp_group() self.determine_num_fused_shared_experts("Glm4MoeLiteForCausalLM") self.model = Glm4MoeLiteModel( 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) self._routed_experts_weights_of_layer = LazyValue( lambda: { layer_id: layer.mlp.get_moe_weights() for layer_id, layer in enumerate(self.model.layers) if isinstance(layer.mlp, Glm4MoeLiteSparseMoeBlock) } ) self.capture_aux_hidden_states = False @property def routed_experts_weights_of_layer(self): return self._routed_experts_weights_of_layer.value def determine_num_fused_shared_experts( self, architecture: str = "Glm4MoeLiteForCausalLM" ): self.num_fused_shared_experts = 0 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) or self.config.architectures[0] != architecture or self.config.n_shared_experts != 1 ): disable_reason = "Only GLM-4.5 or GLM-4.6 on NV-platform with capability >= 80 can use shared experts fusion optimization." elif get_parallel().moe_ep_size > 1: disable_reason = "GLM-4.5 or GLM-4.6 cannot use shared experts fusion optimization under expert parallelism." if disable_reason is not None: from sglang.srt.arg_groups.overrides import declare_load_time_override declare_load_time_override( "Glm4MoeLiteForCausalLM.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: with get_attn_tp_context().maybe_input_scattered(forward_batch): 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 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 # TODO (Qiaolin-Yu): check if other draft models need similar layer id # adjustment if layer_ids and layer_ids[0] == 1: self.model.layers_to_capture = [val + 1 for val in layer_ids] else: self.model.layers_to_capture = list(layer_ids) def set_dflash_layers_to_capture(self, layer_ids: List[int]): if not self.pp_group.is_last_rank: return if layer_ids is None: raise ValueError( "DFLASH requires explicit layer_ids for aux hidden capture." ) self.capture_aux_hidden_states = True self.model.layers_to_capture = [val + 1 for val in layer_ids] 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, ) # Fuse q_a_proj and kv_a_proj_with_mqa along output dimension when q_lora_rank is not None fuse_qkv_a_proj = hasattr(self.config, "q_lora_rank") and ( self.config.q_lora_rank is not None ) cached_a_proj = {} if fuse_qkv_a_proj else None 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 # GLM NOTE: for MLA if fuse_qkv_a_proj and ( "q_a_proj" in name or "kv_a_proj_with_mqa" in name ): cached_a_proj[name] = loaded_weight q_a_proj_name = ( name if "q_a_proj" in name else name.replace("kv_a_proj_with_mqa", "q_a_proj") ) kv_a_proj_name = ( name if "kv_a_proj_with_mqa" in name else name.replace("q_a_proj", "kv_a_proj_with_mqa") ) # When both q_a_proj and kv_a_proj_with_mqa has been cached, load the fused weight to parameter if ( q_a_proj_name in cached_a_proj and kv_a_proj_name in cached_a_proj ): q_a_proj_weight = cached_a_proj[q_a_proj_name] kv_a_proj_weight = cached_a_proj[kv_a_proj_name] fused_weight = torch.cat( [q_a_proj_weight, kv_a_proj_weight], dim=0 ) param_name = ( name.replace("q_a_proj", "fused_qkv_a_proj_with_mqa") if "q_a_proj" in name else name.replace( "kv_a_proj_with_mqa", "fused_qkv_a_proj_with_mqa" ) ) if param_name not in params_dict: continue param = params_dict[param_name] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, fused_weight) cached_a_proj.pop(q_a_proj_name) cached_a_proj.pop(kv_a_proj_name) else: if ( "k_scale" in name or "v_scale" in name ) and name not in params_dict: # modelopt attn kv scale is named differently if any(scale in name for scale in ["k_scale", "v_scale"]): name = name.replace("_proj", "attn_mqa") else: logger.warning( f"Unknown scale found in checkpoint: {name}" ) 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") # DeepseekV2AttentionMLA.forward_* expects post_load_weights() to populate # per-layer packed weights like `w_kc`/`w_vc` (used during CUDA graph capture). # GLM-4.7-Flash configs not set `config.mla`, but this model always uses # DeepseekV2AttentionMLA, so we must run the post-load processing. # Use weight_names=None to ensure we always process all layers. Some checkpoints / # naming schemes may not include "kv_b_proj" in `weight_names`, but `w_kc`/`w_vc` # are still required by DeepseekV2AttentionMLA at runtime. self.post_load_weights(is_nextn=is_nextn, weight_names=None) EntryClass = [Glm4MoeLiteForCausalLM]