""" # Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved. # # 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. """ from typing import Optional import paddle from paddle import nn from paddleformers.utils.log import logger from fastdeploy import envs from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce from fastdeploy.model_executor.layers.utils import get_tensor from fastdeploy.model_executor.utils import h2d_copy, slice_fn from fastdeploy.platforms import current_platform from fastdeploy.worker.experts_manager import RedundantExpertManger try: from fastdeploy.model_executor.ops.gpu import noaux_tc, noaux_tc_redundant except: logger.warning("import noaux_tc Failed!") import numpy as np def get_moe_method(): """ return moe method based on device platform """ if current_platform.is_cuda() or current_platform.is_iluvatar(): from .fused_moe_cutlass_backend import CutlassMoEMethod return CutlassMoEMethod(None) elif current_platform.is_xpu(): from fastdeploy.model_executor.layers.backends import XPUMoEMethod return XPUMoEMethod(None) elif current_platform.is_gcu(): from fastdeploy.model_executor.layers.backends import GCUFusedMoeMethod return GCUFusedMoeMethod(None) elif current_platform.is_intel_hpu(): from fastdeploy.model_executor.layers.backends import HpuMoEMethod return HpuMoEMethod(None) # return HpuTensorWiseFP8MoEMethod(None) elif current_platform.is_maca(): from fastdeploy.model_executor.layers.backends import ( MetaxCutlassUnquantizedFusedMoEMethod, ) return MetaxCutlassUnquantizedFusedMoEMethod(None) return None def get_moe_scores( gating_output: paddle.Tensor, n_group, topk_group, top_k, routed_scaling_factor, e_score_correction_bias, renormalize: bool = False, expert_id_to_ep_rank_array: paddle.Tensor = None, expert_in_rank_num_list: paddle.Tensor = None, tokens_per_expert_stats_list: paddle.Tensor = None, redundant_ep_rank_num_plus_one: int = 1, ) -> paddle.Tensor: """ compute moe scores using e_score_correction_bias. """ scores = paddle.nn.functional.sigmoid(gating_output) assert e_score_correction_bias is not None, "e_score_correction_bias is none!" scores_with_bias = scores + e_score_correction_bias if expert_id_to_ep_rank_array is None: scores, topk_values, topk_idx = noaux_tc( scores, scores_with_bias, n_group if n_group > 0 else 1, topk_group if topk_group > 0 else 1, top_k, renormalize, routed_scaling_factor, ) else: scores, topk_values, topk_idx = noaux_tc_redundant( scores, scores_with_bias, expert_id_to_ep_rank_array, expert_in_rank_num_list, tokens_per_expert_stats_list, n_group if n_group > 0 else 1, topk_group if topk_group > 0 else 1, top_k, renormalize, routed_scaling_factor, redundant_ep_rank_num_plus_one, ) return scores, topk_values, topk_idx class FusedMoE(nn.Layer): """ FusedMoE is a layer that performs MoE (Mixture of Experts) computation. """ def __init__( self, fd_config, reduce_results: bool = True, renormalize: bool = False, moe_intermediate_size: int = -1, num_experts: int = -1, expert_id_offset: int = 0, top_k: int = -1, topk_method: str = "", topk_group: int = -1, n_group: int = -1, routed_scaling_factor: float = 1.0, layer_idx: int = -1, moe_tag: str = "", gate_correction_bias=None, redundant_table_manger: RedundantExpertManger = None, weight_key_map: dict = {}, with_bias: bool = False, activation="swiglu", model_format: Optional[str] = None, ): """ Initialize the Moe layer with given parameters. Args: fd_config (FDConfig): Arguments related to inference, containing attributes such as weight_dtype, act_dtype, mp_size, hidden_size, head_dim, num_attention_heads, and ffn_hidden_size. """ super().__init__() self.fd_config = fd_config self.layer_idx = layer_idx self.reduce_results = reduce_results self.renormalize = renormalize self.tp_rank = fd_config.parallel_config.tensor_parallel_rank self.tp_size = fd_config.parallel_config.tensor_parallel_size self.ep_size = fd_config.parallel_config.expert_parallel_size self.ep_rank = fd_config.parallel_config.expert_parallel_rank self.tp_group = fd_config.parallel_config.tp_group # NOTE(Zhenyu Li): just supports tp_size = 1 when ep_size > 1 in MOE now. if self.ep_size > 1: self.tp_size = 1 self.tp_rank = 0 assert (self.tp_size >= 1 and self.ep_size == 1) or ( self.tp_size == 1 and self.ep_size > 1 ), "MoE only support parallelism on TP or EP dimension." self.hidden_size = fd_config.model_config.hidden_size self.num_experts = num_experts self.num_local_experts = self.num_experts // self.ep_size self.moe_intermediate_size = moe_intermediate_size // self.tp_size self.top_k = top_k self.weight_key_map = weight_key_map self.use_method = envs.FD_MOE_BACKEND.lower() self.moe_tag = moe_tag self.with_bias = with_bias self.activation = activation if self.ep_size > 1: expert_id_offset = expert_id_offset + self.ep_rank * self.num_local_experts self.expert_id_offset = expert_id_offset self.gate_correction_bias_key = self.weight_key_map.get("gate_correction_bias_key", None) if self.gate_correction_bias_key is not None: self.moe_use_gate_correction_bias = True else: self.moe_use_gate_correction_bias = False # used for deepseek_v3 self.topk_method = topk_method self.topk_group = topk_group self.n_group = n_group self.routed_scaling_factor = routed_scaling_factor self._dtype = self._helper.get_default_dtype() self.weight_dtype = self._dtype self.is_quantized = fd_config.model_config.is_quantized and not ( fd_config.quant_config.name() == "mix_quant" and fd_config.quant_config.moe_quant_type is None ) moe_quant_config = fd_config.quant_config self.moe_quant_config = moe_quant_config self.moe_quant_type = None if moe_quant_config and moe_quant_config.get_quant_method(self): self.quant_method = moe_quant_config.get_quant_method(self) self.moe_quant_type = moe_quant_config.name() else: # unquantized quant_method self.quant_method = get_moe_method() assert self.quant_method is not None, "self.quant_method should not be None" self.redundant_table_manger = redundant_table_manger self.is_rearrange = False if self.ep_size > 1: self.quant_method.init_ep(self) # Merge normal and RL build model if gate_correction_bias is not None: self.gate_correction_bias = gate_correction_bias else: self.gate_correction_bias = None self.quant_method.create_weights( self, weight_loader=self.weight_loader, model_format=fd_config.model_config.model_format if model_format is None else model_format, num_experts=self.num_local_experts if self.ep_size > 1 else self.num_experts, hidden_size=self.hidden_size, moe_intermediate_size=self.moe_intermediate_size, ) logger.info( f"{moe_tag}MoE config is {num_experts=}[{expert_id_offset}, {expert_id_offset + self.num_local_experts}), \ {top_k=}, hidden_size={self.hidden_size}, {moe_intermediate_size=}, \ , ep_size={self.ep_size}, \ tp_size={self.tp_size}." ) def weight_loader( self, param, loaded_weight, expert_id, shard_id: Optional[str] = None, source: Optional[str] = None ): """ source:Avoid redundant transpose of fused weights when weight_loader is called iteratively """ if expert_id is None and shard_id is None: # MoE experts has been fused in disk self._load_fused_experts_weight(param, loaded_weight) return if hasattr(param, "SHARD_ID_TO_SHARDED_DIM"): SHARD_ID_TO_SHARDED_DIM = param.SHARD_ID_TO_SHARDED_DIM elif current_platform.is_cuda() or current_platform.is_iluvatar() or current_platform.is_maca(): SHARD_ID_TO_SHARDED_DIM = {"gate": 1, "down": 0, "up": 1} else: SHARD_ID_TO_SHARDED_DIM = {"gate": 0, "down": 1, "up": 0} if not param._is_initialized(): param.initialize() if not (expert_id - self.expert_id_offset >= 0 and expert_id - self.expert_id_offset < self.num_local_experts): return weight_need_transpose = getattr(param, "weight_need_transpose", False) if shard_id is None: # 1.gate up fused in disk if weight_need_transpose: loaded_weight = get_tensor(loaded_weight) loaded_weight = loaded_weight.transpose([1, 0]) output_size = param[expert_id - self.expert_id_offset].shape[SHARD_ID_TO_SHARDED_DIM["gate"]] shard_offsets = [ # (shard_id, shard_offset, shard_size) ("gate", 0, output_size // 2 * self.tp_size), ("up", output_size // 2 * self.tp_size, output_size // 2 * self.tp_size), ] for shard_id, shard_offset, shard_size in shard_offsets: loaded_weight_shard = slice_fn( loaded_weight, SHARD_ID_TO_SHARDED_DIM[shard_id], shard_offset, shard_offset + shard_size ) self.weight_loader(param, loaded_weight_shard, expert_id, shard_id, "fused") else: if weight_need_transpose and source != "fused": loaded_weight = get_tensor(loaded_weight) loaded_weight = loaded_weight.transpose([1, 0]) # 2.gate up splited in disk assert shard_id in ["gate", "down", "up"] self._load_expert_weight( param=param, expert_id=expert_id, loaded_weight=loaded_weight, shard_id=shard_id, shard_dim=SHARD_ID_TO_SHARDED_DIM[shard_id], ) def _load_gate_up_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None, is_sharded=False): if self.tp_size > 1 and not is_sharded: tp_shard_dim = shard_dim weight_dim = -1 if tp_shard_dim else 0 size = loaded_weight.shape[weight_dim] block_size = size // self.tp_size shard_offset = self.tp_rank * block_size shard_size = (self.tp_rank + 1) * block_size loaded_weight = slice_fn(loaded_weight, tp_shard_dim, shard_offset, shard_size) expert_param = param[expert_id - self.expert_id_offset] dim = -1 if shard_dim else 0 param_shard_size = expert_param.shape[dim] // 2 if shard_id == "gate": param_shard_offset = 0 else: # shard_id == "up": param_shard_offset = param_shard_size expert_param = slice_fn( expert_param, shard_dim, start=param_shard_offset, end=param_shard_offset + param_shard_size ) if hasattr(param, "tensor_track"): # for dyn quant param.tensor_track.mark( start=param_shard_offset, end=param_shard_offset + param_shard_size, batch_id=expert_id - self.expert_id_offset, ) # To ensure compatibility across backends, apply an extra transpose for GCU and XPU if expert_param.shape != loaded_weight.shape: loaded_weight = loaded_weight.transpose([1, 0]) assert expert_param.shape == loaded_weight.shape, ( f"Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({expert_param.shape})" ) if expert_param.dtype != loaded_weight.dtype: if loaded_weight.dtype == paddle.int8 and expert_param.dtype == paddle.float8_e4m3fn: loaded_weight = loaded_weight.view(expert_param.dtype) else: loaded_weight = loaded_weight.cast(expert_param.dtype) h2d_copy(dst=expert_param, src=loaded_weight) def _load_down_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None): if self.tp_size > 1 and shard_dim is not None: tp_shard_dim = shard_dim dim = -1 if tp_shard_dim else 0 size = loaded_weight.shape[dim] block_size = size // self.tp_size shard_offset = self.tp_rank * block_size shard_size = (self.tp_rank + 1) * block_size loaded_weight = slice_fn(loaded_weight, tp_shard_dim, shard_offset, shard_size) expert_param = param[expert_id - self.expert_id_offset] if hasattr(param, "tensor_track"): # for dyn quant param.tensor_track.mark(start=0, batch_id=expert_id - self.expert_id_offset) # To ensure compatibility across backends, apply an extra transpose for GCU and XPU and opensource weight if expert_param.shape != loaded_weight.shape: loaded_weight = loaded_weight.transpose([1, 0]) assert expert_param.shape == loaded_weight.shape, ( f"Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({expert_param.shape})" ) if expert_param.dtype != loaded_weight.dtype: if loaded_weight.dtype == paddle.int8 and expert_param.dtype == paddle.float8_e4m3fn: loaded_weight = loaded_weight.view(expert_param.dtype) else: loaded_weight = loaded_weight.cast(expert_param.dtype) h2d_copy(dst=expert_param, src=loaded_weight) def _load_fused_experts_weight(self, param, loaded_weight): if self.tp_size > 1: dim = -1 if isinstance(loaded_weight, (np.ndarray, paddle.Tensor)): size = loaded_weight.shape[dim] else: size = loaded_weight.get_shape()[dim] block_size = size // self.tp_size shard_offset = self.tp_rank * block_size shard_size = (self.tp_rank + 1) * block_size loaded_weight = slice_fn(loaded_weight, dim, shard_offset, shard_size) assert param.shape == loaded_weight.shape, ( f"Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({param.shape})" ) h2d_copy(dst=param, src=loaded_weight) if hasattr(param, "tensor_track"): for i in range(self.num_local_experts): param.tensor_track.mark(start=0, batch_id=i) def _load_expert_weight( self, param, expert_id, loaded_weight, shard_id, shard_dim=None, ): if shard_id == "down": self._load_down_weight(param, expert_id, loaded_weight, shard_id, shard_dim) elif shard_id in ["gate", "up"]: self._load_gate_up_weight(param, expert_id, loaded_weight, shard_id, shard_dim) @classmethod def make_expert_params_mapping( cls, num_experts: int, ckpt_gate_proj_name: Optional[str] = None, ckpt_up_proj_name: Optional[str] = None, ckpt_down_proj_name: Optional[str] = None, ckpt_gate_up_proj_name: Optional[str] = None, param_gate_up_proj_name: Optional[str] = None, param_down_proj_name: Optional[str] = None, ckpt_expert_key_name: str = "experts", experts_offset: int = 0, num_experts_start_offset: int = 0, ) -> list[tuple[str, str, int, str]]: param_name_maping = [] if ckpt_gate_up_proj_name: param_name_maping.append((None, ckpt_gate_up_proj_name)) if ckpt_gate_proj_name: param_name_maping.append(("gate", ckpt_gate_proj_name)) if ckpt_down_proj_name: param_name_maping.append(("down", ckpt_down_proj_name)) if ckpt_up_proj_name: param_name_maping.append(("up", ckpt_up_proj_name)) return [ # (param_name, weight_name, expert_id, shard_id) ( ( param_gate_up_proj_name if weight_name in [ckpt_gate_proj_name, ckpt_up_proj_name, ckpt_gate_up_proj_name] else param_down_proj_name ), f"{ckpt_expert_key_name}.{expert_id}.{weight_name}.", expert_id, shard_id, ) for expert_id in range( experts_offset + num_experts_start_offset, experts_offset + num_experts_start_offset + num_experts ) for shard_id, weight_name in param_name_maping ] def load_experts_weight( self, state_dict: dict, up_gate_proj_expert_weight_key: str, down_proj_expert_weight_key: str, is_rearrange: bool = False, ): """ Load experts weight from state_dict. Args: state_dict (dict): The state_dict of model. up_gate_proj_expert_weight_key (str): The key of up_gate_proj expert weight. down_proj_expert_weight_key (str): The key of down_proj expert weight. """ logical_expert_ids = [ i for i in range( self.expert_id_offset, self.expert_id_offset + self.num_local_experts, ) ] ep_rank_to_expert_id_list = [i for i in range(self.num_experts)] if self.redundant_table_manger is not None and is_rearrange is True: ( ep_rank_to_expert_id_list, expert_id_to_ep_rank_array, expert_in_rank_num_list, tokens_per_expert_stats_list, ) = self.redundant_table_manger.get_ep_rank_to_expert_id_list_by_layer(self.layer_idx) logical_expert_ids = ep_rank_to_expert_id_list[ self.expert_id_offset : self.expert_id_offset + self.num_local_experts ] up_gate_proj_weights = [] down_proj_weights = [] if isinstance(state_dict, list): state_dict = dict(state_dict) is_ffn_merged = ( up_gate_proj_expert_weight_key.format(logical_expert_ids[0] if is_rearrange else self.expert_id_offset) in state_dict ) if is_ffn_merged: for expert_idx in logical_expert_ids: down_proj_expert_weight_key_name = down_proj_expert_weight_key.format(expert_idx) up_gate_proj_expert_weight_key_name = up_gate_proj_expert_weight_key.format(expert_idx) up_gate_proj_weights.append( get_tensor( ( state_dict.pop(up_gate_proj_expert_weight_key_name) if up_gate_proj_expert_weight_key_name in state_dict else up_gate_proj_expert_weight_key_name ), self.fd_config.model_config.model, ) ) down_proj_weights.append( get_tensor( ( state_dict.pop(down_proj_expert_weight_key_name) if down_proj_expert_weight_key_name in state_dict else down_proj_expert_weight_key_name ), self.fd_config.model_config.model, ) ) else: gate_expert_weight_key = up_gate_proj_expert_weight_key.replace("up_gate_proj", "gate_proj") up_expert_weight_key = up_gate_proj_expert_weight_key.replace("up_gate_proj", "up_proj") for expert_idx in logical_expert_ids: gate_expert_weight_key_name = gate_expert_weight_key.format(expert_idx) up_expert_weight_key_name = up_expert_weight_key.format(expert_idx) down_proj_expert_weight_key_name = down_proj_expert_weight_key.format(expert_idx) gate = get_tensor( ( state_dict.pop(gate_expert_weight_key_name) if gate_expert_weight_key_name in state_dict else gate_expert_weight_key_name ), self.fd_config.model_config.model, ) up = get_tensor( ( state_dict.pop(up_expert_weight_key_name) if up_expert_weight_key_name in state_dict else up_expert_weight_key_name ), self.fd_config.model_config.model, ) up_gate_proj_weights.append(paddle.concat([gate, up], axis=-1)) down_proj_weights.append( get_tensor( ( state_dict.pop(down_proj_expert_weight_key_name) if down_proj_expert_weight_key_name in state_dict else down_proj_expert_weight_key_name ), self.fd_config.model_config.model, ) ) return up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list def extract_moe_ffn_weights(self, state_dict: dict): """ Extract MoE FFN weights from state dict based on weight key mapping. Args: state_dict (dict): Model state dictionary containing the weights. Returns: tuple: A tuple containing two lists: - up_gate_proj_weights: List of tensors for first FFN layer weights - down_proj_weights: List of tensors for second FFN layer weights Raises: AssertionError: If required weight keys are missing or number of weights doesn't match number of local experts. """ up_gate_proj_expert_weight_key = self.weight_key_map.get("up_gate_proj_expert_weight_key", None) down_proj_expert_weight_key = self.weight_key_map.get("down_proj_expert_weight_key", None) assert up_gate_proj_expert_weight_key is not None, "up_gate_proj_expert_weight_key should not be none." assert down_proj_expert_weight_key is not None, "down_proj_expert_weight_key should not be none." up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list = ( self.load_experts_weight( state_dict, up_gate_proj_expert_weight_key, down_proj_expert_weight_key, ) ) assert ( len(up_gate_proj_weights) == self.num_local_experts ), "up_gate_proj_weights length should be equal to num_local_experts." assert ( len(down_proj_weights) == self.num_local_experts ), "down_proj_weights length should be equal to num_local_experts." return up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list def extract_gate_correction_bias(self, gate_correction_bias_key, state_dict): """ extract_gate_correction_bias function. """ gate_correction_bias_tensor = get_tensor(state_dict.pop(gate_correction_bias_key)).astype("float32") return gate_correction_bias_tensor def load_state_dict(self, state_dict, is_rearrange: bool = False): """ load_state_dict function. """ if self.is_quantized or self.fd_config.model_config.is_moe_quantized: if getattr(self.fd_config.quant_config, "is_permuted", True): self.quant_method.process_prequanted_weights(self, state_dict, is_rearrange) else: self.quant_method.process_loaded_weights(self, state_dict) else: self.quant_method.process_loaded_weights(self, state_dict) def forward_split_allgather(self, x: paddle.Tensor, gate: nn.Layer): """ Forward split allgather function. """ token_num = x.shape[0] token_num_per_rank = (token_num + self.tp_size - 1) // self.tp_size # AllGather will hang when the data shapes on multi-ranks are different! part_x = paddle.zeros(shape=[token_num_per_rank, x.shape[1]], dtype=x.dtype) start_offset = self.tp_rank * token_num_per_rank end_offset = (self.tp_rank + 1) * token_num_per_rank if start_offset >= token_num: start_offset = token_num if end_offset > token_num: end_offset = token_num part_x[: (end_offset - start_offset), :] = x[start_offset:end_offset, :] out = self.quant_method.apply(self, part_x, gate) multi_outs = paddle.zeros([token_num_per_rank * self.tp_size, x.shape[1]], dtype=x.dtype) paddle.distributed.all_gather(multi_outs, out, self.tp_group) out = multi_outs[:token_num, :] return out def forward(self, x: paddle.Tensor, gate: nn.Layer): """ Defines the forward computation of the moe layer. Args: x (Tensor): Input tensor to the moe layer. Returns: Tensor: Output tensor.s """ token_num = x.shape[0] if ( self.ep_size > 1 and self.tp_size > 1 and (not self.fd_config.parallel_config.use_sequence_parallel_moe) and token_num >= self.tp_size ): out = self.forward_split_allgather(x, gate) else: out = self.quant_method.apply(self, x, gate) if self.reduce_results and self.tp_size > 1: out = tensor_model_parallel_all_reduce(out, self.tp_group) return out