# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # Copyright 2025 The ZhipuAI Team. # Copyright 2023 The vLLM team. # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. # # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX # and OPT implementations in this library. It has been modified from its # original forms to accommodate minor architectural differences compared # to GPT-NeoX and OPT used by the Meta AI team that trained the model. # # 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, GLM-4.7 model compatible with HuggingFace weights.""" import typing from collections.abc import Callable, Iterable from itertools import islice import torch from torch import nn from transformers.models.glm4_moe import Glm4MoeConfig from vllm.attention.layer import Attention from vllm.compilation.decorators import support_torch_compile from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config from vllm.distributed import ( get_ep_group, get_pp_group, get_tensor_model_parallel_world_size, ) from vllm.logger import init_logger from vllm.model_executor.layers.activation import SiluAndMul from vllm.model_executor.layers.fused_moe import SharedFusedMoE from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ( MergedColumnParallelLinear, QKVParallelLinear, RowParallelLinear, ) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding, ) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name, ) from vllm.sequence import IntermediateTensors from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP from .utils import ( AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix, ) logger = init_logger(__name__) class Glm4MoeMLP(nn.Module): def __init__( self, hidden_size: int, intermediate_size: int, hidden_act: str, quant_config: QuantizationConfig | None = None, reduce_results: bool = True, prefix: str = "", ) -> None: super().__init__() self.gate_up_proj = MergedColumnParallelLinear( hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config, prefix=f"{prefix}.gate_up_proj", ) self.down_proj = RowParallelLinear( intermediate_size, hidden_size, bias=False, quant_config=quant_config, reduce_results=reduce_results, prefix=f"{prefix}.down_proj", ) 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): gate_up, _ = self.gate_up_proj(x) x = self.act_fn(gate_up) x, _ = self.down_proj(x) return x class Glm4MoE(nn.Module): def __init__( self, config: Glm4MoeConfig, quant_config: QuantizationConfig | None = None, prefix: str = "", enable_eplb: bool = False, ): super().__init__() self.tp_size = get_tensor_model_parallel_world_size() self.routed_scaling_factor = config.routed_scaling_factor self.ep_group = get_ep_group().device_group self.ep_rank = get_ep_group().rank_in_group self.ep_size = self.ep_group.size() self.n_routed_experts: int = config.n_routed_experts self.n_shared_experts: int = config.n_shared_experts if config.hidden_act != "silu": raise ValueError( f"Unsupported activation: {config.hidden_act}. " "Only silu is supported for now." ) # NOTE In the transformers implementation, the gate isn't an nn.Linear, # so we cannot use ReplicatedLinear here. # See: https://github.com/huggingface/transformers/blob/v4.55.1/src/transformers/models/glm4_moe/modeling_glm4_moe.py#L260 self.gate = nn.Linear( config.hidden_size, config.n_routed_experts, bias=False, dtype=torch.float32, ) self.gate.e_score_correction_bias = nn.Parameter( torch.empty(config.n_routed_experts, dtype=torch.float32) ) # Load balancing settings. vllm_config = get_current_vllm_config() eplb_config = vllm_config.parallel_config.eplb_config self.enable_eplb = enable_eplb self.n_redundant_experts = eplb_config.num_redundant_experts self.n_logical_experts = self.n_routed_experts self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts self.n_local_physical_experts = self.n_physical_experts // self.ep_size self.physical_expert_start = self.ep_rank * self.n_local_physical_experts self.physical_expert_end = ( self.physical_expert_start + self.n_local_physical_experts ) if config.n_shared_experts is not None: intermediate_size = config.moe_intermediate_size * config.n_shared_experts 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=f"{prefix}.shared_experts", ) else: self.shared_experts = None self.experts = SharedFusedMoE( shared_experts=self.shared_experts, num_experts=config.n_routed_experts, top_k=config.num_experts_per_tok, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, reduce_results=False, renormalize=config.norm_topk_prob, quant_config=quant_config, use_grouped_topk=True, num_expert_group=config.n_group, topk_group=config.topk_group, prefix=f"{prefix}.experts", scoring_func="sigmoid", # we do scaling outside, set factor to 1.0 to avoid double mul routed_scaling_factor=1.0, e_score_correction_bias=self.gate.e_score_correction_bias, enable_eplb=self.enable_eplb, num_redundant_experts=self.n_redundant_experts, router_logits_dtype=torch.float32, ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: num_tokens, hidden_dim = hidden_states.shape hidden_states = hidden_states.view(-1, hidden_dim) # router_logits: (num_tokens, n_experts) router_logits = self.gate(hidden_states.to(dtype=torch.float32)) fused_moe_out = self.experts( hidden_states=hidden_states, router_logits=router_logits ) if self.shared_experts is not None: shared_output, final_hidden_states = fused_moe_out assert shared_output is not None final_hidden_states = ( final_hidden_states * self.routed_scaling_factor + shared_output ) else: final_hidden_states = fused_moe_out * self.routed_scaling_factor if self.tp_size > 1: final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( final_hidden_states ) return final_hidden_states.view(num_tokens, hidden_dim) class Glm4MoeAttention(nn.Module): def __init__( self, config: Glm4MoeConfig, hidden_size: int, num_heads: int, num_kv_heads: int, max_position_embeddings: int = 131072, head_dim: int | None = None, rms_norm_eps: float = 1e-05, qkv_bias: bool = False, use_qk_norm: bool = False, cache_config: CacheConfig | None = None, quant_config: QuantizationConfig | None = None, prefix: str = "", ) -> None: super().__init__() self.hidden_size = hidden_size tp_size = get_tensor_model_parallel_world_size() self.total_num_heads = num_heads assert self.total_num_heads % tp_size == 0 self.num_heads = self.total_num_heads // tp_size self.total_num_kv_heads = num_kv_heads if self.total_num_kv_heads >= 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 % 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 tp_size % self.total_num_kv_heads == 0 self.num_kv_heads = max(1, self.total_num_kv_heads // 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.max_position_embeddings = max_position_embeddings self.use_qk_norm = use_qk_norm self.qkv_proj = QKVParallelLinear( hidden_size, self.head_dim, self.total_num_heads, self.total_num_kv_heads, bias=qkv_bias, quant_config=quant_config, prefix=f"{prefix}.qkv_proj", ) self.o_proj = RowParallelLinear( self.total_num_heads * self.head_dim, hidden_size, bias=False, quant_config=quant_config, prefix=f"{prefix}.o_proj", ) config.rope_parameters.setdefault("partial_rotary_factor", 0.5) self.rotary_emb = get_rope( self.head_dim, max_position=max_position_embeddings, rope_parameters=config.rope_parameters, ) self.attn = Attention( self.num_heads, self.head_dim, self.scaling, num_kv_heads=self.num_kv_heads, cache_config=cache_config, quant_config=quant_config, prefix=f"{prefix}.attn", ) 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) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, ) -> torch.Tensor: qkv, _ = self.qkv_proj(hidden_states) q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) if self.use_qk_norm: q = self.q_norm(q.reshape(-1, self.num_heads, self.head_dim)).reshape( q.shape ) k = self.k_norm(k.reshape(-1, self.num_kv_heads, self.head_dim)).reshape( k.shape ) q, k = self.rotary_emb(positions, q, k) attn_output = self.attn(q, k, v) output, _ = self.o_proj(attn_output) return output class Glm4MoeDecoderLayer(nn.Module): def __init__( self, config: Glm4MoeConfig, cache_config: CacheConfig | None = None, quant_config: QuantizationConfig | None = None, prefix: str = "", enable_eplb: bool = False, ) -> None: super().__init__() self.hidden_size = config.hidden_size max_position_embeddings = getattr(config, "max_position_embeddings", 131072) # DecoderLayers are created with `make_layers` which passes the prefix # with the layer's index. layer_idx = int(prefix.split(sep=".")[-1]) self.layer_idx = layer_idx self.self_attn = Glm4MoeAttention( config=config, hidden_size=self.hidden_size, num_heads=config.num_attention_heads, num_kv_heads=config.num_key_value_heads, max_position_embeddings=max_position_embeddings, head_dim=config.head_dim, rms_norm_eps=config.rms_norm_eps, qkv_bias=config.attention_bias, cache_config=cache_config, quant_config=quant_config, prefix=f"{prefix}.self_attn", use_qk_norm=config.use_qk_norm, ) if ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace ): self.mlp = Glm4MoE( config=config, quant_config=quant_config, prefix=f"{prefix}.mlp", enable_eplb=enable_eplb, ) else: self.mlp = Glm4MoeMLP( hidden_size=config.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, prefix=f"{prefix}.mlp", ) 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.routed_scaling_factor = config.routed_scaling_factor def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, residual: torch.Tensor | None, ) -> tuple[torch.Tensor, torch.Tensor]: if residual is None: residual = hidden_states hidden_states = self.input_layernorm(hidden_states) else: hidden_states, residual = self.input_layernorm(hidden_states, residual) hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states) hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) hidden_states = self.mlp(hidden_states) return hidden_states, residual @support_torch_compile( dynamic_arg_dims={ "input_ids": 0, "positions": -1, "intermediate_tensors": 0, "inputs_embeds": 0, } ) class Glm4MoeModel(nn.Module): def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config enable_eplb = vllm_config.parallel_config.enable_eplb self.config = config self.vocab_size = config.vocab_size if get_pp_group().is_first_rank: self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens" ) else: self.embed_tokens = PPMissingLayer() self.start_layer, self.end_layer, self.layers = make_layers( config.num_hidden_layers, lambda prefix: Glm4MoeDecoderLayer( config=config, cache_config=cache_config, quant_config=quant_config, prefix=prefix, enable_eplb=enable_eplb, ), prefix=f"{prefix}.layers", ) if get_pp_group().is_last_rank: self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.norm = PPMissingLayer() self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( ["hidden_states", "residual"], config.hidden_size ) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: return self.embed_tokens(input_ids) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None = None, inputs_embeds: torch.Tensor | None = None, ) -> torch.Tensor | IntermediateTensors: if get_pp_group().is_first_rank: if inputs_embeds is not None: hidden_states = inputs_embeds else: hidden_states = self.embed_input_ids(input_ids) residual = None else: assert intermediate_tensors is not None hidden_states = intermediate_tensors["hidden_states"] residual = intermediate_tensors["residual"] for layer in islice(self.layers, self.start_layer, self.end_layer): hidden_states, residual = layer(positions, hidden_states, residual) if not get_pp_group().is_last_rank: return IntermediateTensors( {"hidden_states": hidden_states, "residual": residual} ) hidden_states, _ = self.norm(hidden_states, residual) return hidden_states def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) return SharedFusedMoE.make_expert_params_mapping( self, 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, ) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: 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), ] params_dict = dict(self.named_parameters()) loaded_params: set[str] = set() expert_params_mapping = self.get_expert_mapping() for name, loaded_weight in weights: spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) if spec_layer is not None: 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) and name not in params_dict: 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 is_pp_missing_parameter(name, self): continue param = params_dict[name] weight_loader = param.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 # Anyway, this is an expert weight and should not be # attempted to load as other weights later is_expert_weight = True # Do not modify `name` since the loop may continue here # Instead, create a new variable name_mapped = name.replace(weight_name, param_name) if is_pp_missing_parameter(name_mapped, self): continue param = params_dict[name_mapped] # We should ask the weight loader to return success or not # here since otherwise we may skip experts with other # available replicas. weight_loader = typing.cast( Callable[..., bool], param.weight_loader ) success = weight_loader( param, loaded_weight, name_mapped, shard_id=shard_id, expert_id=expert_id, return_success=True, ) if success: name = name_mapped break else: if is_expert_weight: # We've checked that this is an expert weight # However it's not mapped locally to this rank # So we simply skip it continue # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue # Remapping the name of FP8 kv-scale. name = maybe_remap_kv_scale_name(name, params_dict) if name is None: continue if is_pp_missing_parameter(name, self): continue param = params_dict[name] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) loaded_params.add(name) return loaded_params class Glm4MixtureOfExperts(MixtureOfExperts): def extract_moe_parameters(self, example_moe: Glm4MoE | None) -> None: if example_moe is None: raise RuntimeError("No Glm4MoE layer found in model.layers.") else: self.num_logical_experts = example_moe.n_logical_experts self.num_physical_experts = example_moe.n_physical_experts self.num_local_physical_experts = example_moe.n_local_physical_experts self.num_routed_experts = example_moe.n_routed_experts self.num_shared_experts = example_moe.n_shared_experts self.num_redundant_experts = example_moe.n_redundant_experts def update_physical_experts_metadata( self, num_physical_experts: int, num_local_physical_experts: int, ) -> None: assert self.num_local_physical_experts == num_local_physical_experts self.num_physical_experts = num_physical_experts self.num_local_physical_experts = num_local_physical_experts self.num_redundant_experts = num_physical_experts - self.num_logical_experts for moe in self.moe_mlp_layers: moe.n_local_physical_experts = num_local_physical_experts moe.n_physical_experts = num_physical_experts moe.n_redundant_experts = self.num_redundant_experts moe.experts.update_expert_map() class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, Glm4MixtureOfExperts): packed_modules_mapping = { "qkv_proj": [ "q_proj", "k_proj", "v_proj", ], "gate_up_proj": [ "gate_proj", "up_proj", ], } fall_back_to_pt_during_load = False def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config self.config = config self.quant_config = quant_config self.model = Glm4MoeModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) if get_pp_group().is_last_rank: self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=maybe_prefix(prefix, "lm_head"), ) else: self.lm_head = PPMissingLayer() self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors ) self.expert_weights = [] # Set MoE hyperparameters self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace self.num_expert_groups = config.n_group self.moe_layers = [] self.moe_mlp_layers: list[Glm4MoE] = [] example_moe = None for layer in self.model.layers: if isinstance(layer, PPMissingLayer): continue assert isinstance(layer, Glm4MoeDecoderLayer) if isinstance(layer.mlp, Glm4MoE): # Pick last one layer since the first ones may be dense layers. example_moe = layer.mlp self.moe_mlp_layers.append(layer.mlp) self.moe_layers.append(layer.mlp.experts) self.extract_moe_parameters(example_moe) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: return self.model.embed_input_ids(input_ids) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None = None, inputs_embeds: torch.Tensor | None = None, ) -> torch.Tensor | IntermediateTensors: hidden_states = self.model( input_ids, positions, intermediate_tensors, inputs_embeds ) return hidden_states def compute_logits( self, hidden_states: torch.Tensor, ) -> torch.Tensor | None: logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: loader = AutoWeightsLoader(self) return loader.load_weights(weights) def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: return self.model.get_expert_mapping() def get_spec_layer_idx_from_weight_name( config: Glm4MoeConfig, weight_name: str ) -> int | None: if hasattr(config, "num_nextn_predict_layers") and ( config.num_nextn_predict_layers > 0 ): layer_idx = config.num_hidden_layers for i in range(config.num_nextn_predict_layers): if f"layers.{layer_idx + i}." in weight_name: return layer_idx + i return None