# # Copyright (c) 2025 Baidu, Inc. All Rights Reserved. # Adapted from vllm/model_executor/models/qwen3.py # Copyright 2023 The vLLM team. # # This file is a part of the vllm-kunlun project. # # 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 Qwen3 model compatible with HuggingFace weights.""" from collections.abc import Iterable from typing import Optional, Union import xtorch_ops import torch import os from torch import nn from transformers import Qwen3Config from vllm.attention import AttentionType, AttentionMetadata from vllm_kunlun.ops.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_pp_group, get_tensor_model_parallel_world_size from vllm.logger import init_logger from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import (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 from vllm_kunlun.ops.vocab_parallel_embedding import VocabParallelEmbedding from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm import envs from vllm.model_executor.models.adapters import as_seq_cls_model from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP from .qwen2 import Qwen2MLP as Qwen3MLP from vllm.model_executor.models.utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index, is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix) from vllm.forward_context import ForwardContext, get_forward_context from vllm.platforms import current_platform from vllm_kunlun.ops.rotary_embedding import Split_Norm_Rope logger = init_logger(__name__) class Qwen3Attention(nn.Module): def __init__(self, hidden_size: int, num_heads: int, num_kv_heads: int, max_position: int = 4096 * 32, head_dim: Optional[int] = None, rms_norm_eps: float = 1e-06, qkv_bias: bool = False, rope_theta: float = 10000, cache_config: Optional[CacheConfig] = None, quant_config: Optional[QuantizationConfig] = None, rope_scaling: Optional[tuple] = None, prefix: str = "", attn_type: str = AttentionType.DECODER) -> 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.rope_theta = rope_theta self.max_position = max_position if rope_scaling is not None: scaling_factor = rope_scaling["factor"] self.max_position = int(self.max_position * scaling_factor) 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", ) self.rotary_emb = get_rope( self.head_dim, rotary_dim=self.head_dim, max_position=self.max_position, base=self.rope_theta, rope_scaling=rope_scaling, ) 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", attn_type=attn_type) 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, attn_metadata: AttentionMetadata, residual: Optional[torch.Tensor], ) -> torch.Tensor: qkv, _ = self.qkv_proj(hidden_states) # TODO: Supports both original Rope and Kunlun Rope fusion operators if os.getenv('FUSED_QK_ROPE_OP') == "1": # Rope fusion operators q, k, v = Split_Norm_Rope(qkv, self.rotary_emb.cos_sin_cache, self.q_norm.weight, self.k_norm.weight, positions, self.max_position, self.num_heads, self.num_kv_heads, self.head_dim, ) else: q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) # Add qk-norm q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim) q_by_head = self.q_norm(q_by_head) q = q_by_head.view(q.shape) k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim) k_by_head = self.k_norm(k_by_head) k = k_by_head.view(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 Qwen3DecoderLayer(nn.Module): def __init__( self, config: Qwen3Config, cache_config: Optional[CacheConfig] = None, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__() self.hidden_size = config.hidden_size # Requires transformers > 4.32.0 rope_theta = getattr(config, "rope_theta", 1000000) rope_scaling = getattr(config, "rope_scaling", None) # By default, Qwen3 uses causal attention as it is a decoder-only model. # You can override the HF config with `is_causal=False` to enable # bidirectional attention, which is used in some embedding models # (e.g. Alibaba-NLP/gte-Qwen3-7B-instruct) if getattr(config, "is_causal", True): attn_type = AttentionType.DECODER else: attn_type = AttentionType.ENCODER_ONLY self.self_attn = Qwen3Attention( hidden_size=self.hidden_size, num_heads=config.num_attention_heads, max_position=config.max_position_embeddings, num_kv_heads=config.num_key_value_heads, rope_theta=rope_theta, rms_norm_eps=config.rms_norm_eps, qkv_bias=getattr(config, 'attention_bias', False), head_dim=getattr(config, 'head_dim', None), cache_config=cache_config, quant_config=quant_config, rope_scaling=rope_scaling, prefix=f"{prefix}.self_attn", attn_type=attn_type, ) self.mlp = Qwen3MLP( hidden_size=self.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) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, attn_metadata: AttentionMetadata, residual: Optional[torch.Tensor], ) -> tuple[torch.Tensor, torch.Tensor]: # Self Attention 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, attn_metadata=attn_metadata, residual=residual, ) # Fully Connected hidden_states, residual = self.post_attention_layernorm( hidden_states, residual) hidden_states = self.mlp(hidden_states) return hidden_states, residual ALL_DECODER_LAYER_TYPES = { "attention": Qwen3DecoderLayer, } @support_torch_compile( dynamic_arg_dims={ "input_ids": 0, # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, # otherwise (seq_len, ). "positions": -1, "intermediate_tensors": 0, "inputs_embeds": 0, }) class Qwen3Model(nn.Module): """Qwen3Model""" def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", decoder_layer_type: type[nn.Module] = Qwen3DecoderLayer): super().__init__() config = vllm_config.model_config.hf_config cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config # TODO (@robertgshaw2): see if this can be moved out if (cache_config.sliding_window is not None and hasattr(config, "max_window_layers")): assert config.max_window_layers == config.num_hidden_layers, ( "Sliding window for some but all layers is not supported. " "This model uses sliding window but `max_window_layers` = {} " "is less than `num_hidden_layers` = {}. Please open an issue " "to discuss this feature.".format( config.max_window_layers, config.num_hidden_layers, )) self.config = config self.quant_config = quant_config self.vocab_size = config.vocab_size if get_pp_group().is_first_rank or (config.tie_word_embeddings and get_pp_group().is_last_rank): self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=f"{prefix}.embed_tokens", ) else: self.embed_tokens = PPMissingLayer() # Use the provided decoder layer type or default to Qwen2DecoderLayer decoder_layer_type = decoder_layer_type or Qwen3DecoderLayer self.start_layer, self.end_layer, self.layers = make_layers( config.num_hidden_layers, lambda prefix: decoder_layer_type(config=config, cache_config=cache_config, quant_config=quant_config, prefix=prefix), prefix=f"{prefix}.layers", ) self.make_empty_intermediate_tensors = ( make_empty_intermediate_tensors_factory( ["hidden_states", "residual"], config.hidden_size)) if get_pp_group().is_last_rank: self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.norm = PPMissingLayer() def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: """get_input_embeddings""" return self.embed_tokens(input_ids) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: Optional[IntermediateTensors] = None, inputs_embeds: Optional[torch.Tensor] = None, ) -> Union[torch.Tensor, IntermediateTensors]: """ Args: input_ids (torch.Tensor): Input sequence of shape `(batch, seq_len)`. Indices are expected to be in the range `[0, config.vocab_size]`. positions (torch.Tensor): Positional tensor of shape `(batch, seq_len)`. intermediate_tensors (Optional[IntermediateTensors], optional): Intermediate tensors from previous forward pass. Defaults to `None`. inputs_embeds (Optional[torch.Tensor], optional): Optionally, instead of positional embeddings, you can choose to provide your own embedding lookup matrix of shape `(batch, seq_len, emb_dim)`. If None, the model will create one on its own using the input ids. Defaults to `None`. Returns: Union[torch.Tensor, IntermediateTensors]: If `intermediate_tensors` is not None, returns a IntermediateTensors object. Otherwise, returns a tensor of shape `(batch, seq_len, hidden_size)` representing the output of the last transformer encoder layer. """ forward_context: ForwardContext = get_forward_context() attn_metadata = forward_context.attn_metadata if get_pp_group().is_first_rank: if inputs_embeds is not None: hidden_states = inputs_embeds else: hidden_states = self.get_input_embeddings(input_ids) residual = None else: assert intermediate_tensors is not None hidden_states = intermediate_tensors["hidden_states"] residual = intermediate_tensors["residual"] for i, layer in enumerate(self.layers[self.start_layer:self.end_layer], start=self.start_layer): hidden_states, residual = layer( positions, hidden_states, attn_metadata, 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 load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: """Load model weights. Args: weights (Iterable[tuple[str, torch.Tensor]]): An iterator containing weight names and their corresponding values. Returns (set[str]): A set of already loaded weight names. Exceptions: None. """ 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(remove_duplicate=False)) loaded_params: set[str] = set() for name, loaded_weight in weights: if "rotary_emb.inv_freq" in name: continue if (self.quant_config is not None and (scale_name := self.quant_config.get_cache_scale(name))): # Loading kv cache quantization scales param = params_dict[scale_name] weight_loader = getattr(param, "weight_loader", default_weight_loader) loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]) weight_loader(param, loaded_weight) loaded_params.add(scale_name) continue for (param_name, weight_name, shard_id) in stacked_params_mapping: if weight_name not 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 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: # 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 Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): packed_modules_mapping = { "qkv_proj": [ "q_proj", "k_proj", "v_proj", ], "gate_up_proj": [ "gate_proj", "up_proj", ], } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config lora_config = vllm_config.lora_config self.config = config self.lora_config = lora_config self.quant_config = quant_config self.model = Qwen3Model(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) if get_pp_group().is_last_rank: if config.tie_word_embeddings: self.lm_head = self.model.embed_tokens else: 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) def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: return self.model.get_input_embeddings(input_ids) def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: Optional[IntermediateTensors] = None, inputs_embeds: Optional[torch.Tensor] = None, kv_caches: list[torch.Tensor] = None ) -> Union[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, sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: logits = self.logits_processor(self.lm_head, hidden_states, sampling_metadata) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: loader = AutoWeightsLoader( self, skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None), ) return loader.load_weights(weights) Qwen3ForSequenceClassification = as_seq_cls_model(Qwen3ForCausalLM)