Update modeling_densebackward_olmoe0125.py
Browse files- modeling_densebackward_olmoe0125.py +105 -141
modeling_densebackward_olmoe0125.py
CHANGED
@@ -6,15 +6,46 @@ import torch.nn.functional as F
|
|
6 |
|
7 |
# 导入官方实现(注意根据你的 transformers 版本调整导入路径)
|
8 |
from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM, OlmoeSparseMoeBlock, OlmoeMLP
|
9 |
-
from
|
10 |
|
11 |
|
12 |
class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
def forward(self, hidden_states: torch.Tensor):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
batch_size, seq_length, hidden_dim = hidden_states.shape
|
15 |
flat_hidden = hidden_states.view(-1, hidden_dim) # (B*seq_len, hidden_dim)
|
16 |
|
17 |
-
# 计算路由 logits
|
18 |
router_logits = self.gate(flat_hidden) # (B*seq_len, num_experts)
|
19 |
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) # (B*seq_len, num_experts)
|
20 |
|
@@ -25,130 +56,90 @@ class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock):
|
|
25 |
routing_weights_topk = routing_weights_topk.to(flat_hidden.dtype)
|
26 |
|
27 |
# ---------- 稀疏计算部分 ----------
|
28 |
-
#
|
29 |
sparse_output = torch.zeros((flat_hidden.size(0), hidden_dim), dtype=flat_hidden.dtype, device=flat_hidden.device)
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
expert_outputs_dict = {i: {} for i in range(self.num_experts)}
|
34 |
-
|
35 |
-
# one-hot 编码 top-k 专家
|
36 |
expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts) # (B*seq_len, top_k, num_experts)
|
37 |
expert_mask = expert_mask.permute(2, 1, 0) # (num_experts, top_k, B*seq_len)
|
38 |
|
39 |
-
# 保存每个token激活的专家列表
|
40 |
-
token_to_experts = [[] for _ in range(flat_hidden.size(0))]
|
41 |
-
|
42 |
-
# 稀疏计算
|
43 |
for expert_idx in range(self.num_experts):
|
44 |
expert_layer = self.experts[expert_idx]
|
45 |
idx, top_x = torch.where(expert_mask[expert_idx])
|
46 |
-
|
47 |
if top_x.numel() > 0:
|
48 |
-
# 记录该专家被哪些token激活
|
49 |
-
top_x_list = top_x.tolist()
|
50 |
-
token_indices_per_expert[expert_idx].extend(top_x_list)
|
51 |
-
|
52 |
-
# 记录每个token激活了哪些专家
|
53 |
-
for token_idx in top_x_list:
|
54 |
-
token_to_experts[token_idx].append(expert_idx)
|
55 |
-
|
56 |
-
# 标准MoE前向计算
|
57 |
current_state = flat_hidden[top_x] # (n, hidden_dim)
|
58 |
current_output = expert_layer(current_state) # (n, hidden_dim)
|
59 |
weight = routing_weights_topk[top_x, idx].unsqueeze(-1) # (n, 1)
|
60 |
weighted_output = current_output * weight
|
61 |
sparse_output.index_add_(0, top_x, weighted_output.to(flat_hidden.dtype))
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
# ----------
|
68 |
-
#
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
# 我们使用一个小的缩放参数来确保前向计算几乎不受影响
|
90 |
-
scale = 1e-4
|
91 |
-
|
92 |
-
# 计算一个残差连接,确保梯度能够流向所有专家
|
93 |
-
for token_idx in range(flat_hidden.size(0)):
|
94 |
-
token_input = flat_hidden[token_idx:token_idx+1] # 保持维度
|
95 |
-
|
96 |
-
# 对每个未激活的专家创建伪梯度路径
|
97 |
-
for expert_idx in range(self.num_experts):
|
98 |
-
if not computed_mask[token_idx, expert_idx]:
|
99 |
-
# 查找是否有token激活了这个专家
|
100 |
-
if token_indices_per_expert[expert_idx]:
|
101 |
-
# 从激活该专家的token中选择一个
|
102 |
-
# 优先选择与当前token共享其他专家的token
|
103 |
-
similar_tokens = []
|
104 |
-
for other_idx in token_indices_per_expert[expert_idx]:
|
105 |
-
# 检查是否有共同激活的专家
|
106 |
-
common_experts = set(token_to_experts[token_idx]) & set(token_to_experts[other_idx])
|
107 |
-
if common_experts:
|
108 |
-
similar_tokens.append(other_idx)
|
109 |
-
|
110 |
-
if similar_tokens:
|
111 |
-
# 使用相似token的平均输出
|
112 |
-
similar_outputs = [expert_outputs_dict[expert_idx][t] for t in similar_tokens]
|
113 |
-
estimated_output = torch.stack(similar_outputs).mean(0)
|
114 |
-
else:
|
115 |
-
# 使用所有激活该专家的token的平均输出
|
116 |
-
all_outputs = [expert_outputs_dict[expert_idx][t] for t in token_indices_per_expert[expert_idx]]
|
117 |
-
estimated_output = torch.stack(all_outputs).mean(0)
|
118 |
-
|
119 |
-
# 添加微小的直接计算以维持梯度流
|
120 |
-
direct_output = self.experts[expert_idx](token_input).squeeze(0)
|
121 |
-
|
122 |
-
# 组合估计输出和直接计算
|
123 |
-
# 前向使用估计输出,反向使用直接计算
|
124 |
-
combined = estimated_output.detach() + scale * (direct_output - direct_output.detach())
|
125 |
-
dense_outputs[token_idx, expert_idx] = combined
|
126 |
-
else:
|
127 |
-
# 如果没有token激活该专家,直接进行计算但使用小缩放
|
128 |
-
# 这保证了梯度流而对前向几乎无影响
|
129 |
-
direct_output = scale * self.experts[expert_idx](token_input).squeeze(0)
|
130 |
-
dense_outputs[token_idx, expert_idx] = direct_output
|
131 |
-
|
132 |
-
# ---------- 组合输出 ----------
|
133 |
-
# 使用路由权重对每个token的所有专家输出进行加权
|
134 |
-
dense_combined = torch.zeros_like(sparse_output)
|
135 |
-
|
136 |
-
for token_idx in range(flat_hidden.size(0)):
|
137 |
-
# 对该token的所有专家输出进行加权
|
138 |
-
token_experts_output = dense_outputs[token_idx] # (num_experts, hidden_dim)
|
139 |
-
token_routing_weights = routing_weights[token_idx].unsqueeze(-1) # (num_experts, 1)
|
140 |
-
weighted_experts = token_experts_output * token_routing_weights
|
141 |
-
token_output = weighted_experts.sum(0) # (hidden_dim)
|
142 |
-
dense_combined[token_idx] = token_output
|
143 |
-
|
144 |
-
# 使用直通梯度技巧:前向使用sparse_output,反向使用dense_combined
|
145 |
-
# 增大直通梯度系数以增强梯度流
|
146 |
-
straight_through_scale = 1.0 # 可以尝试不同的值
|
147 |
-
final_flat = sparse_output.detach() + straight_through_scale * (dense_combined - dense_combined.detach())
|
148 |
-
|
149 |
final_output = final_flat.view(batch_size, seq_length, hidden_dim)
|
150 |
return final_output, router_logits
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM):
|
154 |
"""
|
@@ -161,48 +152,21 @@ class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM):
|
|
161 |
base_model_prefix = "olmoe"
|
162 |
|
163 |
def __init__(self, config):
|
164 |
-
# 首先调用父类初始化方法
|
165 |
super().__init__(config)
|
166 |
-
|
167 |
-
# 不要尝试重新赋值self,而是从预训练模型加载并更新当前模型
|
168 |
-
pretrained_model = OlmoeForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0125")
|
169 |
-
|
170 |
-
# 复制预训练模型的状态到当前模型
|
171 |
-
self.config = pretrained_model.config
|
172 |
-
self.model = pretrained_model.model
|
173 |
-
self.vocab_size = pretrained_model.vocab_size
|
174 |
-
self.router_aux_loss_coef = pretrained_model.router_aux_loss_coef
|
175 |
-
self.num_experts = pretrained_model.num_experts
|
176 |
-
self.lm_head = pretrained_model.lm_head
|
177 |
-
|
178 |
# 遍历���型中所有 decoder 层,替换每个 OlmoeSparseMoeBlock 为 DenseBackward 版本
|
179 |
# 此处假设官方模型在 self.model.layers 中组织 decoder 层,
|
180 |
# 且每层中 mlp 模块包含属性 sparse_moe_block。
|
181 |
for layer in self.model.layers:
|
182 |
-
if hasattr(layer.mlp, "
|
183 |
-
|
184 |
-
orig_block = layer.mlp
|
185 |
# 通过直接复制原版属性创建新的块
|
186 |
new_block = DenseBackwardOlmoeSparseMoeBlock(config) # 或其他适当参数
|
187 |
# 然后手动复制需要共享的属性:
|
188 |
new_block.gate = orig_block.gate
|
189 |
new_block.experts = orig_block.experts
|
|
|
190 |
new_block.num_experts = orig_block.num_experts
|
191 |
new_block.top_k = orig_block.top_k
|
192 |
new_block.norm_topk_prob = orig_block.norm_topk_prob
|
193 |
-
layer.mlp = new_block
|
194 |
-
print(type(layer.mlp))
|
195 |
|
196 |
-
def main():
|
197 |
-
config = DenseBackwardOLMoEConfig( # 官方模型参数
|
198 |
-
model_marker="DenseBackward_olmoe_marker",
|
199 |
-
)
|
200 |
-
# 创建自定义模型实例
|
201 |
-
model = DenseBackwardOLMoEForCausalLM(config)
|
202 |
-
print(type(model))
|
203 |
-
print(type(model.model))
|
204 |
-
print(type(model.model.layers[0]))
|
205 |
-
print(type(model.model.layers[0].mlp))
|
206 |
-
print(type(model.model.layers[0].mlp.experts))
|
207 |
-
if __name__ == "__main__":
|
208 |
-
main()
|
|
|
6 |
|
7 |
# 导入官方实现(注意根据你的 transformers 版本调整导入路径)
|
8 |
from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM, OlmoeSparseMoeBlock, OlmoeMLP
|
9 |
+
from configuration_custom import DenseBackwardOLMoEConfig
|
10 |
|
11 |
|
12 |
class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock):
|
13 |
+
"""
|
14 |
+
继承自官方 OlmoeSparseMoeBlock,实现 dense backward 功能:
|
15 |
+
前向输出依旧保持与官方相同(即稀疏计算结果),
|
16 |
+
但在反向传播时,通过直通梯度让 dense 计算的梯度传递回来,
|
17 |
+
dense 输出通过对每个专家在所有 token 上进行计算,并利用全 routing 权重加权获得。
|
18 |
+
|
19 |
+
输入:
|
20 |
+
hidden_states: Tensor, shape (batch_size, sequence_length, hidden_dim)
|
21 |
+
输出:
|
22 |
+
final_output: Tensor, shape (batch_size, sequence_length, hidden_dim)
|
23 |
+
router_logits: Tensor, shape (batch_size * sequence_length, num_experts)
|
24 |
+
"""
|
25 |
def forward(self, hidden_states: torch.Tensor):
|
26 |
+
"""
|
27 |
+
输入:
|
28 |
+
hidden_states: Tensor, shape (batch_size, sequence_length, hidden_dim)
|
29 |
+
输出:
|
30 |
+
final_output: Tensor, shape (batch_size, sequence_length, hidden_dim)
|
31 |
+
router_logits: Tensor, shape (batch_size * sequence_length, num_experts)
|
32 |
+
实现思路:
|
33 |
+
1. 将输入展平为 (B*seq_len, hidden_dim),通过 self.gate 得到 router_logits,
|
34 |
+
并计算全专家的 routing 权重(softmax 后)。
|
35 |
+
2. 对 routing 权重取 top-k,得到 routing_weights_topk 与 selected_experts;
|
36 |
+
如配置要求,归一化 top-k 概率。
|
37 |
+
3. 稀疏计算部分:仅计算每个 token 对于 top-k 专家的输出,
|
38 |
+
并累加得到 sparse_output(保留原版计算流程,同时记录激活专家的实际输出)。
|
39 |
+
4. Dense 估计部分:先计算所有专家对所有 token 的输出(all_expert_outputs),
|
40 |
+
再逐 token 调用 estimate_dense_output 得到 dense 输出(dense_estimated)。
|
41 |
+
5. 使用直通梯度技巧:前向输出用 sparse_output,但梯度来源于 dense_estimated。
|
42 |
+
6. 最后 reshape 为 (batch_size, sequence_length, hidden_dim) 并返回 final_output 及 router_logits.
|
43 |
+
"""
|
44 |
+
#determine the shape of hidden_states
|
45 |
batch_size, seq_length, hidden_dim = hidden_states.shape
|
46 |
flat_hidden = hidden_states.view(-1, hidden_dim) # (B*seq_len, hidden_dim)
|
47 |
|
48 |
+
# 计算路由 logits 和全专家 routing 权重
|
49 |
router_logits = self.gate(flat_hidden) # (B*seq_len, num_experts)
|
50 |
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) # (B*seq_len, num_experts)
|
51 |
|
|
|
56 |
routing_weights_topk = routing_weights_topk.to(flat_hidden.dtype)
|
57 |
|
58 |
# ---------- 稀疏计算部分 ----------
|
59 |
+
# 初始化稀疏输出,shape: (B*seq_len, hidden_dim)
|
60 |
sparse_output = torch.zeros((flat_hidden.size(0), hidden_dim), dtype=flat_hidden.dtype, device=flat_hidden.device)
|
61 |
+
# 用于记录每个 token 对激活专家的实际输出
|
62 |
+
activated_outputs = [{} for _ in range(flat_hidden.size(0))]
|
63 |
+
# one-hot 编码 top-k 专家,shape: (B*seq_len, top_k, num_experts)
|
|
|
|
|
|
|
64 |
expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts) # (B*seq_len, top_k, num_experts)
|
65 |
expert_mask = expert_mask.permute(2, 1, 0) # (num_experts, top_k, B*seq_len)
|
66 |
|
|
|
|
|
|
|
|
|
67 |
for expert_idx in range(self.num_experts):
|
68 |
expert_layer = self.experts[expert_idx]
|
69 |
idx, top_x = torch.where(expert_mask[expert_idx])
|
|
|
70 |
if top_x.numel() > 0:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
current_state = flat_hidden[top_x] # (n, hidden_dim)
|
72 |
current_output = expert_layer(current_state) # (n, hidden_dim)
|
73 |
weight = routing_weights_topk[top_x, idx].unsqueeze(-1) # (n, 1)
|
74 |
weighted_output = current_output * weight
|
75 |
sparse_output.index_add_(0, top_x, weighted_output.to(flat_hidden.dtype))
|
76 |
+
# 保存当前 token 对该专家的实际输出
|
77 |
+
for pos, token_idx in enumerate(top_x.tolist()):
|
78 |
+
activated_outputs[token_idx][expert_idx] = current_output[pos]
|
79 |
+
# ---------- 稀疏计算结束 ----------
|
80 |
+
|
81 |
+
# ---------- Dense估计部分 ----------
|
82 |
+
# 计算所有专家对所有 token 的 dense 输出,shape: (B*seq_len, num_experts, hidden_dim)
|
83 |
+
all_expert_outputs = torch.stack([expert(flat_hidden) for expert in self.experts], dim=1)
|
84 |
+
# 将 selected_experts 转换为 list,每个 token 的激活专家列表
|
85 |
+
all_routing = selected_experts.tolist() # 长度为 (B*seq_len)
|
86 |
+
|
87 |
+
dense_outputs = []
|
88 |
+
for i in range(flat_hidden.size(0)):
|
89 |
+
dense_est = self.estimate_dense_output(
|
90 |
+
token_idx=i,
|
91 |
+
activated=all_routing[i], # 当前 token 激活的专家列表,例如 [a, b]
|
92 |
+
gate_prob=routing_weights[i], # 当前 token 的完整 routing 权重 (num_experts,)
|
93 |
+
activated_outputs=activated_outputs[i], # 当前 token 对激活专家的实际输出
|
94 |
+
all_routing=all_routing, # 全 batch 每个 token 的激活专家列表(list of lists)
|
95 |
+
all_expert_outputs=all_expert_outputs # (B*seq_len, num_experts, hidden_dim)
|
96 |
+
)
|
97 |
+
dense_outputs.append(dense_est.unsqueeze(0))
|
98 |
+
dense_outputs = torch.cat(dense_outputs, dim=0) # (B*seq_len, hidden_dim)
|
99 |
+
# ---------- Dense估计结束 ----------
|
100 |
+
|
101 |
+
# 使用直通梯度:前向输出用稀疏结果,但反向传播时梯度来源于 dense 估计
|
102 |
+
final_flat = sparse_output.detach() + (dense_outputs - dense_outputs.detach())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
103 |
final_output = final_flat.view(batch_size, seq_length, hidden_dim)
|
104 |
return final_output, router_logits
|
105 |
|
106 |
+
def estimate_dense_output(self, token_idx, activated, gate_prob, activated_outputs, all_routing, all_expert_outputs):
|
107 |
+
"""
|
108 |
+
对于当前 token,根据 mini-batch 中的信息估计 dense 输出。
|
109 |
+
参数:
|
110 |
+
token_idx: 当前 token 的索引(标量)
|
111 |
+
activated: 当前 token 激活的专家列表,例如 [1, 3]
|
112 |
+
gate_prob: 当前 token 的 routing 权重,形状 (num_experts,)
|
113 |
+
activated_outputs: dict,当前 token 对激活专家的实际输出,形状 (hidden_dim,)
|
114 |
+
all_routing: list,每个 token 的激活专家列表(长度为 N,每个元素为 list)
|
115 |
+
all_expert_outputs: Tensor, (N, num_experts, hidden_dim)
|
116 |
+
返回:
|
117 |
+
estimated_dense: Tensor, (hidden_dim,)
|
118 |
+
"""
|
119 |
+
num_experts = gate_prob.size(0)
|
120 |
+
dense_parts = {}
|
121 |
+
# 对于激活的专家,直接使用其实际输出
|
122 |
+
for idx in activated:
|
123 |
+
dense_parts[idx] = activated_outputs[idx]
|
124 |
+
# 对于未激活的专家,使用 mini-batch 中其他 token 的输出估计
|
125 |
+
non_activated = [i for i in range(num_experts) if i not in activated]
|
126 |
+
for i in non_activated:
|
127 |
+
indices = []
|
128 |
+
for idx, r_dec in enumerate(all_routing):
|
129 |
+
if (i in r_dec) and (len(set(r_dec) & set(activated)) > 0):
|
130 |
+
indices.append(idx)
|
131 |
+
if indices:
|
132 |
+
selected_outputs = all_expert_outputs[indices, i, :] # (n, hidden_dim)
|
133 |
+
estimated = selected_outputs.mean(dim=0)
|
134 |
+
else:
|
135 |
+
estimated = all_expert_outputs[:, i, :].mean(dim=0)
|
136 |
+
dense_parts[i] = estimated
|
137 |
+
# 按 gate_prob 加权求和各专家输出
|
138 |
+
estimated_dense = 0
|
139 |
+
for i in range(num_experts):
|
140 |
+
estimated_dense += gate_prob[i] * dense_parts[i]
|
141 |
+
return estimated_dense
|
142 |
+
|
143 |
|
144 |
class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM):
|
145 |
"""
|
|
|
152 |
base_model_prefix = "olmoe"
|
153 |
|
154 |
def __init__(self, config):
|
|
|
155 |
super().__init__(config)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
# 遍历���型中所有 decoder 层,替换每个 OlmoeSparseMoeBlock 为 DenseBackward 版本
|
157 |
# 此处假设官方模型在 self.model.layers 中组织 decoder 层,
|
158 |
# 且每层中 mlp 模块包含属性 sparse_moe_block。
|
159 |
for layer in self.model.layers:
|
160 |
+
if hasattr(layer.mlp, "sparse_moe_block"):
|
161 |
+
orig_block = layer.mlp.sparse_moe_block
|
|
|
162 |
# 通过直接复制原版属性创建新的块
|
163 |
new_block = DenseBackwardOlmoeSparseMoeBlock(config) # 或其他适当参数
|
164 |
# 然后手动复制需要共享的属性:
|
165 |
new_block.gate = orig_block.gate
|
166 |
new_block.experts = orig_block.experts
|
167 |
+
new_block.router = orig_block.router
|
168 |
new_block.num_experts = orig_block.num_experts
|
169 |
new_block.top_k = orig_block.top_k
|
170 |
new_block.norm_topk_prob = orig_block.norm_topk_prob
|
171 |
+
layer.mlp.sparse_moe_block = new_block
|
|
|
172 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|