mirror of
https://github.com/Vision-CAIR/MiniGPT-4.git
synced 2025-04-19 04:00:46 +00:00
281 lines
13 KiB
Python
281 lines
13 KiB
Python
import copy
|
||
import pickle
|
||
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
|
||
class MoELayer(nn.Module):
|
||
def __init__(self, hidden_size, expert, num_experts, route_method, topk=1, use_balance_loss=True, weight_type='l2_norm'):
|
||
# remove hash list
|
||
nn.Module.__init__(self)
|
||
self.num_experts = num_experts
|
||
self.experts = nn.ModuleList([copy.deepcopy(expert) for i in range(num_experts)])
|
||
self.route_method = route_method
|
||
self.topk = topk
|
||
self.use_balance_loss = use_balance_loss
|
||
self.weight_type = weight_type
|
||
|
||
if route_method in ["gate-token", "gate-sentence"]:
|
||
self.gate = nn.Linear(hidden_size, num_experts, bias=False).float()
|
||
elif route_method in ["gate-sentence-post"]:
|
||
gate = nn.Linear(hidden_size, 1, bias=False).float()
|
||
# self.gates = nn.ModuleList([copy.deepcopy(gate) for i in range(num_experts)])
|
||
self.gate = gate
|
||
else:
|
||
raise KeyError("Routing method not supported.")
|
||
|
||
def _balancing_loss(self, prob_gate, num_tokens):
|
||
# From MOEBERT
|
||
# compute the load balancing loss
|
||
# prob_gate，是 [bz, num_expert]，每个样本被分配给每个expert的概率
|
||
# 等价于 VMOE 中 _gshard_auxiliary_loss
|
||
P = prob_gate.mean(0) # torch.Size([num_expert]) 每个expert被分配到样本的平均概率
|
||
temp = num_tokens.float()
|
||
f = temp / temp.sum(0, keepdim=True) # 每个expert被分配的sample比例
|
||
balance_loss = self.num_experts * torch.sum(P * f)
|
||
return balance_loss
|
||
|
||
def _importance_auxiliary_loss(self, prob_gate):
|
||
# From VMOE
|
||
# _importance_auxiliary_loss
|
||
axis = tuple(range(prob_gate.ndim - 1)) # All except last.
|
||
importance_per_expert = torch.sum(prob_gate, dim=axis)
|
||
std_importance_per_expert = torch.std(importance_per_expert)
|
||
mean_importance_per_expert = torch.mean(importance_per_expert)
|
||
# Compute coefficient of variation (i.e. std/mean) squared.
|
||
return (std_importance_per_expert / mean_importance_per_expert)**2
|
||
|
||
def _forward_gate_token(self, x):
|
||
bsz, seq_len, dim = x.size()
|
||
|
||
x = x.view(-1, dim)
|
||
logits_gate = self.gate(x)
|
||
prob_gate = F.softmax(logits_gate, dim=-1)
|
||
gate = torch.argmax(prob_gate, dim=-1)
|
||
|
||
order = gate.argsort(0)
|
||
num_tokens = F.one_hot(gate, self.num_experts).gt(0).sum(0)
|
||
gate_load = num_tokens.clone()
|
||
x = x[order] # reorder according to expert number
|
||
x = x.split(num_tokens.tolist(), dim=0) # a list of length self.num_experts
|
||
|
||
# compute the load balancing loss
|
||
P = prob_gate.mean(0)
|
||
temp = num_tokens.float()
|
||
f = temp / temp.sum(0, keepdim=True)
|
||
balance_loss = self.num_experts * torch.sum(P * f)
|
||
|
||
prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
|
||
prob_gate = prob_gate[order]
|
||
prob_gate = prob_gate.split(num_tokens.tolist(), dim=0)
|
||
|
||
def forward_expert(input_x, prob_x, expert_idx):
|
||
input_x = self.experts[expert_idx].forward(input_x)
|
||
input_x = input_x * prob_x
|
||
return input_x
|
||
|
||
x = [forward_expert(x[i], prob_gate[i], i) for i in range(self.num_experts)]
|
||
x = torch.vstack(x)
|
||
x = x[order.argsort(0)] # restore original order
|
||
x = x.view(bsz, seq_len, dim)
|
||
|
||
return x, balance_loss, gate_load
|
||
|
||
def _forward_gate_sentence_top1_raw(self, x, attention_mask):
|
||
"""
|
||
x: query_attention_output , torch.Size([bz, 32, 768])
|
||
attention_mask: torch.ones([bz, 32])
|
||
|
||
### Notice:
|
||
the raw version of expert_attention_mask is the extended_attention_mask,
|
||
which will be add to attention_score directly
|
||
the values of extended_attention_mask are -0.0 or -10000
|
||
it should be adjust to 1/0 version to be processed by experts
|
||
"""
|
||
attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
|
||
x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
|
||
logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
|
||
prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
|
||
gate = torch.argmax(prob_gate, dim=-1) # torch.Size([bz])
|
||
|
||
order = gate.argsort(0)
|
||
num_sentences = F.one_hot(gate, self.num_experts).gt(0).sum(0)
|
||
gate_load = num_sentences.clone()
|
||
x = x[order] # reorder according to expert number
|
||
x = x.split(num_sentences.tolist(), dim=0) # a list of length self.num_experts
|
||
|
||
# compute the load balancing loss
|
||
P = prob_gate.mean(0)
|
||
temp = num_sentences.float()
|
||
f = temp / temp.sum(0, keepdim=True)
|
||
balance_loss = self.num_experts * torch.sum(P * f)
|
||
|
||
prob_gate = prob_gate.gather(dim=1, index=gate.unsqueeze(1))
|
||
prob_gate = prob_gate[order]
|
||
prob_gate = prob_gate.split(num_sentences.tolist(), dim=0)
|
||
|
||
def forward_expert(input_x, prob_x, expert_idx):
|
||
input_x = self.experts[expert_idx].forward(input_x)
|
||
input_x = input_x * prob_x.unsqueeze(-1)
|
||
return input_x
|
||
|
||
result = []
|
||
for i in range(self.num_experts):
|
||
if x[i].size(0) > 0:
|
||
result.append(forward_expert(x[i], prob_gate[i], i))
|
||
result = torch.vstack(result)
|
||
result = result[order.argsort(0)] # restore original order
|
||
|
||
return result, balance_loss, gate_load
|
||
|
||
def _forward_gate_sentence_post(self, x, attention_mask):
|
||
"""
|
||
x: query_attention_output; torch.Size([bz, 32, 768])
|
||
attention_mask: torch.ones([bz, 32])
|
||
bz = 4
|
||
x = torch.randn(bz,32,768)
|
||
attention_mask = torch.ones([bz, 32])
|
||
|
||
"""
|
||
attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
|
||
|
||
def forward_expert(input_x, expert_idx):
|
||
# input_x += torch.randn(4,32,768)
|
||
# return input_x
|
||
output_x = self.experts[expert_idx].forward(input_x)
|
||
return output_x
|
||
|
||
outputs = list()
|
||
logits_gate_lst = list()
|
||
for expert_idx in range(self.num_experts):
|
||
output_x = forward_expert(x_masked, expert_idx)
|
||
outputs.append(output_x.unsqueeze(0))
|
||
|
||
output_x_aver = output_x.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
|
||
# gate_acore = self.gates[expert_idx](output_x_aver)
|
||
gate_score = self.gate(output_x_aver)
|
||
logits_gate_lst.append(gate_score)
|
||
|
||
candidate_output = torch.cat(outputs) # torch.Size([num_expert, bz, 32, 768])
|
||
logits_gate = torch.cat(logits_gate_lst,dim=1)# torch.Size([bz, num_expert])
|
||
prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
|
||
topk_values, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
|
||
num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
|
||
gate_load = num_sentences.clone()
|
||
|
||
# load balancing loss
|
||
if self.use_balance_loss:
|
||
balance_loss = self._balancing_loss(prob_gate, num_sentences)
|
||
else:
|
||
balance_loss = 0.0
|
||
|
||
# importance loss
|
||
importance_loss = self._importance_auxiliary_loss(prob_gate)
|
||
|
||
# output_average = candidate_output.sum(2) / candidate_attn_mask.unsqueeze(-1).sum(2) # torch.Size([num_expert, bz, 768])
|
||
# output_average = torch.permute(output_average, (1, 0, 2)) # torch.Size([bz, num_expert, 768])
|
||
# logits_gate = self.gate(output_average) # torch.Size([bz, num_experts, 1])
|
||
|
||
prob_gate_topk = torch.zeros_like(prob_gate)
|
||
prob_gate_topk.scatter_(1, gate, topk_values)
|
||
prob_gate_normalized = prob_gate_topk / prob_gate_topk.sum(dim=1, keepdim=True) # torch.Size([bz, num_expert])
|
||
candidate_output_ad = torch.permute(candidate_output, (1, 0, 2, 3)) # torch.Size([bz, num_expert, 32, 768])
|
||
results = prob_gate_normalized.unsqueeze(-1).unsqueeze(-1) * candidate_output_ad # torch.Size([bz, num_expert, 32, 768])
|
||
moe_result = torch.sum(results, dim=1) # torch.Size([bz, 32, 768])
|
||
# import pdb;pdb.set_trace()
|
||
|
||
return moe_result, (balance_loss+importance_loss), prob_gate_normalized
|
||
|
||
|
||
def _forward_gate_sentence(self, x, attention_mask):
|
||
"""
|
||
x: query_attention_output , torch.Size([bz, 32, 768])
|
||
attention_mask: torch.ones([bz, 32])
|
||
|
||
### Notice:
|
||
the raw version of expert_attention_mask is the extended_attention_mask,
|
||
which will be add to attention_score directly
|
||
the values of extended_attention_mask are -0.0 or -10000
|
||
it should be adjust to 1/0 version to be processed by experts
|
||
"""
|
||
attention_mask = torch.ones(attention_mask.shape[0], attention_mask.shape[1]).to(x.device)
|
||
x_masked = x * attention_mask.unsqueeze(-1) # torch.Size([bz, 32, 768])
|
||
x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1) # torch.Size([bz, 768])
|
||
logits_gate = self.gate(x_average) # torch.Size([bz, num_experts])
|
||
prob_gate = F.softmax(logits_gate, dim=-1) # torch.Size([bz, num_experts])
|
||
select_prob_gate, gate = torch.topk(prob_gate, self.topk, dim=1) # gate, 每个样本被分配的expert: torch.Size([bz, topk])
|
||
|
||
# 这里用l2 norm 去加权
|
||
if self.weight_type == 'l2_norm':
|
||
normalized_tensor = torch.nn.functional.normalize(select_prob_gate, p=2, dim=0) # L2 Normalization torch.Size([bz, topk])
|
||
elif self.weight_type == 'average':
|
||
normalized_tensor = select_prob_gate / select_prob_gate.sum(dim=1, keepdim=True)
|
||
|
||
num_sentences = F.one_hot(gate, self.num_experts).sum(1).gt(0).sum(0) # 每个expert被分配的样本数 torch.Size([num_expert])
|
||
gate_load = num_sentences.clone()
|
||
|
||
# load balancing loss
|
||
if self.use_balance_loss:
|
||
balance_loss = self._balancing_loss(prob_gate, num_sentences)
|
||
else:
|
||
balance_loss = 0.0
|
||
|
||
# importance loss
|
||
importance_loss = self._importance_auxiliary_loss(prob_gate)
|
||
|
||
# forward experts
|
||
def forward_expert(input_x, expert_idx):
|
||
input_x = self.experts[expert_idx].forward(input_x)
|
||
return input_x
|
||
|
||
result_lst = list()
|
||
for i in range(self.topk):
|
||
# top1、top2... 分别为一组，进行gate分组之后过expert，然后乘以概率后相加
|
||
tmp_gate = gate[:,i]
|
||
tmp_prob = normalized_tensor[:,i].unsqueeze(-1).unsqueeze(-1)
|
||
order = tmp_gate.argsort(0)
|
||
num_sentences_t = F.one_hot(tmp_gate, self.num_experts).gt(0).sum(0)
|
||
x1 = x[order] # reorder according to expert number
|
||
x1 = x1.split(num_sentences_t.tolist(), dim=0) # a list of length self.num_experts
|
||
|
||
result = []
|
||
for i in range(self.num_experts):
|
||
if x1[i].size(0) > 0:
|
||
result.append(forward_expert(x1[i], i))
|
||
result = torch.vstack(result)
|
||
result = result[order.argsort(0)] # restore original order
|
||
# result_lst.append(result * tmp_prob) # result * prob
|
||
result_lst.append(result) # result * prob # add_1212
|
||
|
||
moe_result = sum(result_lst)
|
||
# import pdb;pdb.set_trace()
|
||
return moe_result, (balance_loss+importance_loss), gate
|
||
|
||
def _forward_sentence_single_expert(self, x, attention_mask):
|
||
x_masked = x * attention_mask.unsqueeze(-1)
|
||
x_average = x_masked.sum(1) / attention_mask.unsqueeze(-1).sum(1)
|
||
logits_gate = self.gate(x_average)
|
||
prob_gate = F.softmax(logits_gate, dim=-1)
|
||
gate = torch.argmax(prob_gate, dim=-1)
|
||
|
||
gate_load = F.one_hot(gate, self.num_experts).gt(0).sum(0)
|
||
x = self.experts[gate.cpu().item()].forward(x)
|
||
return x, 0.0, gate_load
|
||
|
||
def forward(self, x, attention_mask):
|
||
if self.route_method == "gate-token":
|
||
x, balance_loss, gate_load = self._forward_gate_token(x)
|
||
elif self.route_method == "gate-sentence":
|
||
if x.size(0) == 1:
|
||
x, balance_loss, gate_load = self._forward_sentence_single_expert(x, attention_mask)
|
||
else:
|
||
x, balance_loss, gate_load = self._forward_gate_sentence(x, attention_mask)
|
||
elif self.route_method == "gate-sentence-post":
|
||
x, balance_loss, gate_load = self._forward_gate_sentence_post(x, attention_mask)
|
||
else:
|
||
raise KeyError("Routing method not supported.")
|
||
|
||
return x, balance_loss, gate_load
|