revert: MoE k=1 → k=2(PCOC 从 1.059 炸到 2.075,Top-1 破坏输出校准)

保留 inference_mode + torch.compile(default)
This commit is contained in:
2026-06-13 11:50:30 +08:00
parent 47c89cc76d
commit faedab5245
+23 -31
View File
@@ -320,7 +320,7 @@ class TopKGate(nn.Module):
return topk_idx, topk_score, probs return topk_idx, topk_score, probs
class SMoE(nn.Module): class SMoE(nn.Module):
def __init__(self, d_model, dim_ff, num_experts, k=1): def __init__(self, d_model, dim_ff, num_experts, k=2):
super().__init__() super().__init__()
self.num_experts = num_experts self.num_experts = num_experts
self.k = k self.k = k
@@ -337,39 +337,31 @@ class SMoE(nn.Module):
topk_idx, topk_score, probs = self.gate(x) topk_idx, topk_score, probs = self.gate(x)
# flatten: [B, S, k] → [B*S, k] out = torch.zeros_like(x)
x_flat = x.reshape(-1, D)
idx_flat = topk_idx.reshape(-1, self.k) # flatten
x_flat = x.reshape(-1, D) # [B*S, D]
idx_flat = topk_idx.reshape(-1, self.k) # [B*S, k]
score_flat = topk_score.reshape(-1, self.k) score_flat = topk_score.reshape(-1, self.k)
if self.k == 1: for i in range(self.num_experts):
# Top-1 快速路径:无需二维 mask 和加权累加 # 找到被路由到 expert i 的 token
idx_flat = idx_flat.squeeze(-1) # [B*S] mask = (idx_flat == i) # [B*S, k]
score_flat = score_flat.squeeze(-1) # [B*S]
out = torch.zeros_like(x_flat)
for i in range(self.num_experts): if not mask.any():
mask = (idx_flat == i) # [B*S] continue
if not mask.any():
continue
selected_x = x_flat[mask]
expert_out = self.experts[i](selected_x)
out[mask] = expert_out * score_flat[mask].unsqueeze(-1)
out = out.reshape(B, S, D) # 哪些 token 命中了 expert i
else: token_idx, k_idx = mask.nonzero(as_tuple=True)
# Top-K 通用路径(k > 1
out = torch.zeros_like(x) selected_x = x_flat[token_idx] # [N, D]
for i in range(self.num_experts):
mask = (idx_flat == i) # [B*S, k] expert_out = self.experts[i](selected_x) # [N, D]
if not mask.any():
continue weight = score_flat[token_idx, k_idx].unsqueeze(-1)
token_idx, k_idx = mask.nonzero(as_tuple=True)
selected_x = x_flat[token_idx] out_flat = out.reshape(-1, D)
expert_out = self.experts[i](selected_x) out_flat[token_idx] += expert_out * weight
weight = score_flat[token_idx, k_idx].unsqueeze(-1)
out_flat = out.reshape(-1, D)
out_flat[token_idx] += expert_out * weight
importance = probs.sum(dim=(0,1)) # [E] importance = probs.sum(dim=(0,1)) # [E]
moe_loss = (importance.std() / (importance.mean() + 1e-6)) moe_loss = (importance.std() / (importance.mean() + 1e-6))
@@ -396,7 +388,7 @@ class TransformerEncoder(nn.Module):
self.act = getattr(F, act) self.act = getattr(F, act)
self.attention_fn = attention_fn self.attention_fn = attention_fn
self.moe = nn.ModuleList([ self.moe = nn.ModuleList([
SMoE(d_model, dim_ff, num_experts=8, k=1) # Top-1 gating: 每个 token 仅激活 1 个 expert SMoE(d_model, dim_ff, num_experts=8, k=2)
for _ in range(num_layers) for _ in range(num_layers)
]) ])