|
|
|
|
|
import torch |
|
import torch.nn as nn |
|
import torch.nn.functional as F |
|
|
|
|
|
from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM, OlmoeSparseMoeBlock, OlmoeMLP |
|
from configuration_custom import DenseBackwardOLMoEConfig |
|
|
|
|
|
class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock): |
|
""" |
|
继承自官方 OlmoeSparseMoeBlock,实现 dense backward 功能: |
|
前向输出依旧保持与官方相同(即稀疏计算结果), |
|
但在反向传播时,通过直通梯度让 dense 计算的梯度传递回来, |
|
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。 |
|
|
|
输入: |
|
hidden_states: Tensor, shape (batch_size, sequence_length, hidden_dim) |
|
输出: |
|
final_output: Tensor, shape (batch_size, sequence_length, hidden_dim) |
|
router_logits: Tensor, shape (batch_size * sequence_length, num_experts) |
|
""" |
|
def forward(self, hidden_states: torch.Tensor): |
|
""" |
|
输入: |
|
hidden_states: Tensor, shape (batch_size, sequence_length, hidden_dim) |
|
输出: |
|
final_output: Tensor, shape (batch_size, sequence_length, hidden_dim) |
|
router_logits: Tensor, shape (batch_size * sequence_length, num_experts) |
|
实现思路: |
|
1. 将输入展平为 (B*seq_len, hidden_dim),通过 self.gate 得到 router_logits, |
|
并计算全专家的 routing 权重(softmax 后)。 |
|
2. 对 routing 权重取 top-k,得到 routing_weights_topk 与 selected_experts; |
|
如配置要求,归一化 top-k 概率。 |
|
3. 稀疏计算部分:仅计算每个 token 对于 top-k 专家的输出, |
|
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)。 |
|
4. Dense 估计部分:先计算所有专家对所有 token 的输出(all_expert_outputs), |
|
再逐 token 调用 estimate_dense_output 得到 dense 输出(dense_estimated)。 |
|
5. 使用直通梯度技巧:前向输出用 sparse_output,但梯度来源于 dense_estimated。 |
|
6. 最后 reshape 为 (batch_size, sequence_length, hidden_dim) 并返回 final_output 及 router_logits. |
|
""" |
|
|
|
batch_size, seq_length, hidden_dim = hidden_states.shape |
|
flat_hidden = hidden_states.view(-1, hidden_dim) |
|
|
|
|
|
router_logits = self.gate(flat_hidden) |
|
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float) |
|
|
|
|
|
routing_weights_topk, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1) |
|
if self.norm_topk_prob: |
|
routing_weights_topk = routing_weights_topk / routing_weights_topk.sum(dim=-1, keepdim=True) |
|
routing_weights_topk = routing_weights_topk.to(flat_hidden.dtype) |
|
|
|
|
|
|
|
sparse_output = torch.zeros((flat_hidden.size(0), hidden_dim), dtype=flat_hidden.dtype, device=flat_hidden.device) |
|
|
|
activated_outputs = [{} for _ in range(flat_hidden.size(0))] |
|
|
|
expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts) |
|
expert_mask = expert_mask.permute(2, 1, 0) |
|
|
|
for expert_idx in range(self.num_experts): |
|
expert_layer = self.experts[expert_idx] |
|
idx, top_x = torch.where(expert_mask[expert_idx]) |
|
if top_x.numel() > 0: |
|
current_state = flat_hidden[top_x] |
|
current_output = expert_layer(current_state) |
|
weight = routing_weights_topk[top_x, idx].unsqueeze(-1) |
|
weighted_output = current_output * weight |
|
sparse_output.index_add_(0, top_x, weighted_output.to(flat_hidden.dtype)) |
|
|
|
for pos, token_idx in enumerate(top_x.tolist()): |
|
activated_outputs[token_idx][expert_idx] = current_output[pos] |
|
|
|
|
|
|
|
|
|
all_expert_outputs = torch.stack([expert(flat_hidden) for expert in self.experts], dim=1) |
|
|
|
all_routing = selected_experts.tolist() |
|
|
|
dense_outputs = [] |
|
for i in range(flat_hidden.size(0)): |
|
dense_est = self.estimate_dense_output( |
|
token_idx=i, |
|
activated=all_routing[i], |
|
gate_prob=routing_weights[i], |
|
activated_outputs=activated_outputs[i], |
|
all_routing=all_routing, |
|
all_expert_outputs=all_expert_outputs |
|
) |
|
dense_outputs.append(dense_est.unsqueeze(0)) |
|
dense_outputs = torch.cat(dense_outputs, dim=0) |
|
|
|
|
|
|
|
final_flat = sparse_output.detach() + (dense_outputs - dense_outputs.detach()) |
|
final_output = final_flat.view(batch_size, seq_length, hidden_dim) |
|
return final_output, router_logits |
|
|
|
def estimate_dense_output(self, token_idx, activated, gate_prob, activated_outputs, all_routing, all_expert_outputs): |
|
""" |
|
对于当前 token,根据 mini-batch 中的信息估计 dense 输出。 |
|
参数: |
|
token_idx: 当前 token 的索引(标量) |
|
activated: 当前 token 激活的专家列表,例如 [1, 3] |
|
gate_prob: 当前 token 的 routing 权重,形状 (num_experts,) |
|
activated_outputs: dict,当前 token 对激活专家的实际输出,形状 (hidden_dim,) |
|
all_routing: list,每个 token 的激活专家列表(长度为 N,每个元素为 list) |
|
all_expert_outputs: Tensor, (N, num_experts, hidden_dim) |
|
返回: |
|
estimated_dense: Tensor, (hidden_dim,) |
|
""" |
|
num_experts = gate_prob.size(0) |
|
dense_parts = {} |
|
|
|
for idx in activated: |
|
dense_parts[idx] = activated_outputs[idx] |
|
|
|
non_activated = [i for i in range(num_experts) if i not in activated] |
|
for i in non_activated: |
|
indices = [] |
|
for idx, r_dec in enumerate(all_routing): |
|
if (i in r_dec) and (len(set(r_dec) & set(activated)) > 0): |
|
indices.append(idx) |
|
if indices: |
|
selected_outputs = all_expert_outputs[indices, i, :] |
|
estimated = selected_outputs.mean(dim=0) |
|
else: |
|
estimated = all_expert_outputs[:, i, :].mean(dim=0) |
|
dense_parts[i] = estimated |
|
|
|
estimated_dense = 0 |
|
for i in range(num_experts): |
|
estimated_dense += gate_prob[i] * dense_parts[i] |
|
return estimated_dense |
|
|
|
|
|
class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM): |
|
""" |
|
自定义的 Olmoe ForCausalLM 模型,使用新的 DenseBackwardOlmoeSparseMoeBlock 替换原版的 MoE 模块, |
|
以实现 dense backward 功能。 |
|
|
|
配置类:DenseBackwardOLMoEConfig |
|
""" |
|
config_class = DenseBackwardOLMoEConfig |
|
base_model_prefix = "olmoe" |
|
|
|
def __init__(self, config): |
|
|
|
super().__init__(config) |
|
|
|
|
|
pretrained_model = OlmoeForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924", torch_dtype=torch.bfloat16) |
|
|
|
|
|
self.config = pretrained_model.config |
|
self.model = pretrained_model.model |
|
self.vocab_size = pretrained_model.vocab_size |
|
self.router_aux_loss_coef = pretrained_model.router_aux_loss_coef |
|
self.num_experts = pretrained_model.num_experts |
|
self.lm_head = pretrained_model.lm_head |
|
|
|
|
|
|
|
|
|
for layer in self.model.layers: |
|
if hasattr(layer.mlp, "gate"): |
|
print("111") |
|
orig_block = layer.mlp |
|
|
|
new_block = DenseBackwardOlmoeSparseMoeBlock(config) |
|
|
|
new_block.gate = orig_block.gate |
|
new_block.experts = orig_block.experts |
|
new_block.num_experts = orig_block.num_experts |
|
new_block.top_k = orig_block.top_k |
|
new_block.norm_topk_prob = orig_block.norm_topk_prob |
|
layer.mlp = new_block |
|
print(type(layer.mlp)) |
|
|
|
test_param = self.model.layers[0].mlp.experts[0].up_proj.weight.data[0, 0].item() |
|
print(f"权重示例值(前): {test_param}") |
|
self.post_init() |
|
|
|
test_param_after = self.model.layers[0].mlp.experts[0].up_proj.weight.data[0, 0].item() |
|
print(f"权重示例值(后): {test_param_after}") |
|
|
|
def main(): |
|
config = DenseBackwardOLMoEConfig( |
|
model_marker="DenseBackward_olmoe_marker", |
|
torch_dtype="bfloat16" |
|
) |
|
|
|
model = DenseBackwardOLMoEForCausalLM(config) |
|
print(type(model)) |
|
print(type(model.model)) |
|
print(type(model.model.layers[0])) |
|
print(type(model.model.layers[0].mlp)) |
|
print(type(model.model.layers[0].mlp.experts)) |
|
if __name__ == "__main__": |
|
main() |