Crystalcareai
commited on
Update modeling_gemmoe.py
Browse files- modeling_gemmoe.py +29 -37
modeling_gemmoe.py
CHANGED
@@ -629,32 +629,23 @@ GEMMOE_ATTENTION_CLASSES = {
|
|
629 |
"sdpa": GemmoeSdpaAttention,
|
630 |
}
|
631 |
|
632 |
-
|
633 |
-
|
634 |
-
class ParallelLinear(nn.Module):
|
635 |
-
def __init__(self, in_features, out_features, num_experts):
|
636 |
super().__init__()
|
637 |
-
self.
|
638 |
-
self.
|
639 |
-
|
640 |
-
self.
|
641 |
-
nn.
|
642 |
-
|
643 |
-
|
644 |
-
|
645 |
-
|
646 |
-
|
647 |
-
|
648 |
-
|
649 |
-
|
650 |
-
|
651 |
-
output = output.view(-1, top_k * self.out_features)
|
652 |
-
output = output[torch.argsort(ordering)]
|
653 |
-
|
654 |
-
if routing_weights is not None:
|
655 |
-
output = output.view(-1, top_k, self.out_features)
|
656 |
-
output = torch.bmm(routing_weights.unsqueeze(1), output).squeeze(1)
|
657 |
-
return output
|
658 |
|
659 |
class GemmoeSparseMoeBlock(nn.Module):
|
660 |
def __init__(self, config):
|
@@ -663,14 +654,11 @@ class GemmoeSparseMoeBlock(nn.Module):
|
|
663 |
self.ffn_dim = config.intermediate_size
|
664 |
self.num_experts = config.num_local_experts
|
665 |
self.top_k = 2
|
666 |
-
self.use_scatter2scatter = config.use_scatter2scatter if hasattr(config, 'use_scatter2scatter') else False
|
667 |
|
668 |
# gating
|
669 |
self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
670 |
|
671 |
-
self.
|
672 |
-
self.expert_mlp2 = ParallelLinear(self.ffn_dim, self.hidden_dim, self.num_experts)
|
673 |
-
self.activation = nn.GELU()
|
674 |
|
675 |
def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
676 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
@@ -679,18 +667,22 @@ class GemmoeSparseMoeBlock(nn.Module):
|
|
679 |
# router_logits: (batch * sequence_length, n_experts)
|
680 |
router_logits = self.gate(hidden_states)
|
681 |
routing_weights = F.softmax(router_logits, dim=1)
|
682 |
-
|
683 |
-
topk_weight = routing_weights.gather(1, topk_idx)
|
684 |
topk_weight /= topk_weight.sum(dim=-1, keepdim=True)
|
685 |
|
686 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
687 |
|
688 |
-
|
689 |
-
hidden_states = self.expert_mlp1(hidden_states, ordering, self.top_k, grouped_in=False, grouped_out=True, use_scatter2scatter=self.use_scatter2scatter)
|
690 |
-
hidden_states = self.activation(hidden_states)
|
691 |
-
hidden_states = self.expert_mlp2(hidden_states, ordering, self.top_k, topk_weight, grouped_in=True, grouped_out=False, use_scatter2scatter=self.use_scatter2scatter)
|
692 |
|
693 |
-
final_hidden_states =
|
694 |
return final_hidden_states.to(hidden_states.dtype), router_logits.to(hidden_states.dtype)
|
695 |
|
696 |
|
|
|
629 |
"sdpa": GemmoeSdpaAttention,
|
630 |
}
|
631 |
|
632 |
+
class GemmoeBlockSparseTop2MLP(nn.Module):
|
633 |
+
def __init__(self, config: GemmoeConfig):
|
|
|
|
|
634 |
super().__init__()
|
635 |
+
self.ffn_dim = config.intermediate_size
|
636 |
+
self.hidden_dim = config.hidden_size
|
637 |
+
|
638 |
+
self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
639 |
+
self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
|
640 |
+
self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
641 |
+
|
642 |
+
self.act_fn = approx_gelu
|
643 |
+
|
644 |
+
def forward(self, hidden_states):
|
645 |
+
current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
|
646 |
+
current_hidden_states = self.w2(current_hidden_states)
|
647 |
+
return current_hidden_states.to(hidden_states.dtype)
|
648 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
649 |
|
650 |
class GemmoeSparseMoeBlock(nn.Module):
|
651 |
def __init__(self, config):
|
|
|
654 |
self.ffn_dim = config.intermediate_size
|
655 |
self.num_experts = config.num_local_experts
|
656 |
self.top_k = 2
|
|
|
657 |
|
658 |
# gating
|
659 |
self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
660 |
|
661 |
+
self.experts = nn.ModuleList([GemmoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
|
|
|
|
|
662 |
|
663 |
def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
664 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
|
|
667 |
# router_logits: (batch * sequence_length, n_experts)
|
668 |
router_logits = self.gate(hidden_states)
|
669 |
routing_weights = F.softmax(router_logits, dim=1)
|
670 |
+
topk_weight, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
|
|
|
671 |
topk_weight /= topk_weight.sum(dim=-1, keepdim=True)
|
672 |
|
673 |
+
hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
|
674 |
+
|
675 |
+
y = torch.empty_like(hidden_states)
|
676 |
+
|
677 |
+
flat_topk_idx = topk_idx.view(-1)
|
678 |
+
for i in range(self.num_experts):
|
679 |
+
expert = self.experts[i]
|
680 |
+
expert_output = expert(hidden_states[flat_topk_idx == i])
|
681 |
+
y[flat_topk_idx == i] = expert_output
|
682 |
|
683 |
+
y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
|
|
|
|
|
|
|
684 |
|
685 |
+
final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
|
686 |
return final_hidden_states.to(hidden_states.dtype), router_logits.to(hidden_states.dtype)
|
687 |
|
688 |
|