qwen3_moe.py

# SPDX-License-Identifier: Apache-2.0

# Copyright 2024 The Qwen 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 Qwen3MoE model compatible with HuggingFace weights."""
from typing import Any, Dict, Iterable, Optional, Set, Tuple, Union

import os
import re
import torch
from torch import nn
from transformers import PretrainedConfig

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (get_pp_group,
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               ReplicatedLinear,
                                               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
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsPP
from .utils import (AutoWeightsLoader, extract_layer_index,
                    is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf
from vllm.utils import W8a8GetCacheJSON

logger = init_logger(__name__)


class Qwen3MoeMLP(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 = "",
    ) -> 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 Qwen3MoeSparseMoeBlock(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()

        if self.tp_size > config.num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.num_experts}.")

        self.experts = FusedMoE(num_experts=config.num_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,
                                prefix=f"{prefix}.experts")

        self.gate = ReplicatedLinear(config.hidden_size,
                                     config.num_experts,
                                     bias=False,
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # NOTE: hidden_states can have either 1D or 2D shape.
        orig_shape = hidden_states.shape
        hidden_dim = hidden_states.shape[-1]
        hidden_states = hidden_states.view(-1, hidden_dim)

        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
        final_hidden_states = self.experts(hidden_states=hidden_states,
                                           router_logits=router_logits)
        final_hidden_states = final_hidden_states
        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(
                final_hidden_states)

        return final_hidden_states.view(orig_shape)


class Qwen3MoeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[Dict[str, Any]] = None,
        max_position_embeddings: int = 8192,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-06,
        qkv_bias: bool = False,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = 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.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        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=max_position_embeddings,
            base=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")

        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)
        # 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.forward_native(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.forward_native(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 Qwen3MoeDecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
        self.self_attn = Qwen3MoeAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            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,
            prefix=f"{prefix}.self_attn",
        )

        # `mlp_only_layers` in the config.
        layer_idx = extract_layer_index(prefix)
        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
                           config.mlp_only_layers)
        if (layer_idx not in mlp_only_layers) and (
                config.num_experts > 0 and
            (layer_idx + 1) % config.decoder_sparse_step == 0):
            self.mlp = Qwen3MoeSparseMoeBlock(config=config,
                                              quant_config=quant_config,
                                              prefix=f"{prefix}.mlp")
        else:
            self.mlp = Qwen3MoeMLP(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)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[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,
        )

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual


@support_torch_compile
class Qwen3MoeModel(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

        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.config = config
        if self.config.quantization_config["bits"] == 4:
            os.environ['LLAMA_NN'] = '0'
            os.environ['LM_NN'] = '0'  
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            prefix=f"{prefix}.embed_tokens")
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: Qwen3MoeDecoderLayer(config=config,
                                                cache_config=cache_config,
                                                quant_config=quant_config,
                                                prefix=prefix),
            prefix=f"{prefix}.layers",
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config
        self.tritonsingleton= W8a8GetCacheJSON()
            
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
        self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
        self.use_fa_pad = os.environ.get('FA_PAD') == '1'
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '1'))

    def get_input_embeddings(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: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        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 in range(self.start_layer, self.end_layer):
            layer = self.layers[i]
            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 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 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.num_experts)

        params_dict = dict(self.named_parameters())
        loaded_params: Set[str] = set()
        for name, loaded_weight in weights:
            if self.use_llama_nn:
                current_count = loaded_weight.current_count 
                total_count = loaded_weight.total_count
            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") or name.endswith("_bias"))
                        and name not in params_dict):
                    continue
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    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:
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        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:
                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Remapping the name of FP8 kv-scale.
                    if name.endswith("kv_scale"):
                        remapped_kv_scale_name = name.replace(
                            ".kv_scale", ".attn.kv_scale")
                        if remapped_kv_scale_name not in params_dict:
                            logger.warning_once(
                                "Found kv scale in the checkpoint "
                                f"(e.g. {name}), but not found the expected "
                                f"name in the model "
                                f"(e.g. {remapped_kv_scale_name}). "
                                "kv-scale is not loaded.")
                            continue
                        else:
                            name = remapped_kv_scale_name
                    param = params_dict[name]
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        
        if self.use_llama_nn and self.quant_method is None and current_count==total_count:
            lay_key_words = [
                "gate_up_proj.weight",
                "down_proj.weight",
                "mlp.gate.weight",
                "self_attn.qkv_proj.weight",
                "self_attn.o_proj.weight",
                "lm_head.weight",
            ]
            combined_words = "|".join(lay_key_words)
            
            # lay_qkv_words = ["self_attn.qkv_proj.weight"]   
            # qkv_words = "|".join(lay_qkv_words)  
            
            # lay_qkv_bias_words = ["self_attn.qkv_proj.bias"]   
            # qkv_bias_words = "|".join(lay_qkv_bias_words) 
            
            # for layername in loaded_params:
            for layername in params_dict.keys():
                weight = params_dict[layername]
                os.environ['LM_NN'] = '0' 
                # if self.use_fa_pad and (re.findall(qkv_bias_words, layername)):
                #     weight.data = pad_weight(weight.data, 32)
                    
                matches = re.findall(combined_words, layername)
                if matches:   
                    # if self.use_gemm_pad and gemm_bank_conf(weight.data.shape[0]):
                    #     weight.data = pad_weight(weight.data, 32)  
                    
                    # if self.use_fa_pad and (re.findall(qkv_words, layername)):
                    #     if not gemm_bank_conf(weight.data.shape[0]):
                    #         weight.data = pad_weight(weight.data, 32)
                        
                    _weight = torch.zeros_like(weight.data)
                    ori_shape =_weight.shape
                    
                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)
                    
                    weight.data=weight.data.reshape(ori_shape[1],-1)
        return loaded_params


class Qwen3MoeForCausalLM(nn.Module, SupportsPP):

    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 = Qwen3MoeModel(vllm_config=vllm_config,
                                   prefix=maybe_prefix(prefix, "model"))
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        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,
    ) -> 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=(["rotary_emb.inv_freq"]),
        )
        return loader.load_weights(weights)