Delete shared_moe_projector.py
Browse files- shared_moe_projector.py +0 -182
shared_moe_projector.py
DELETED
|
@@ -1,182 +0,0 @@
|
|
| 1 |
-
import torch
|
| 2 |
-
import torch.nn as nn
|
| 3 |
-
import torch.nn.functional as F # noqa: N812
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
class SwiGLUExpert(nn.Module):
|
| 7 |
-
"""SwiGLU expert MLP (used for both shared and routed experts)."""
|
| 8 |
-
|
| 9 |
-
def __init__(self, input_dim: int, hidden_dim: int, output_dim: int):
|
| 10 |
-
super().__init__()
|
| 11 |
-
self.gate_proj = nn.Linear(input_dim, hidden_dim, bias=False)
|
| 12 |
-
self.up_proj = nn.Linear(input_dim, hidden_dim, bias=False)
|
| 13 |
-
self.down_proj = nn.Linear(hidden_dim, output_dim, bias=False)
|
| 14 |
-
self.act = nn.SiLU()
|
| 15 |
-
|
| 16 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 17 |
-
return self.down_proj(self.act(self.gate_proj(x)) * self.up_proj(x))
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
class SharedMoEBlock(nn.Module):
|
| 21 |
-
"""MoE block with shared expert + sparse routed experts."""
|
| 22 |
-
|
| 23 |
-
def __init__(
|
| 24 |
-
self,
|
| 25 |
-
input_dim: int,
|
| 26 |
-
hidden_dim: int,
|
| 27 |
-
output_dim: int,
|
| 28 |
-
num_experts: int = 4,
|
| 29 |
-
top_k: int = 2,
|
| 30 |
-
):
|
| 31 |
-
super().__init__()
|
| 32 |
-
self.num_experts = num_experts
|
| 33 |
-
self.top_k = top_k
|
| 34 |
-
self.output_dim = output_dim
|
| 35 |
-
|
| 36 |
-
# Router: zero-initialized for natural learning
|
| 37 |
-
self.router = nn.Linear(input_dim, num_experts, bias=False)
|
| 38 |
-
nn.init.zeros_(self.router.weight)
|
| 39 |
-
|
| 40 |
-
# Shared expert (always active)
|
| 41 |
-
self.shared_expert = SwiGLUExpert(input_dim, hidden_dim, output_dim)
|
| 42 |
-
|
| 43 |
-
# Routed experts (sparse)
|
| 44 |
-
self.experts = nn.ModuleList(
|
| 45 |
-
[SwiGLUExpert(input_dim, hidden_dim, output_dim) for _ in range(num_experts)]
|
| 46 |
-
)
|
| 47 |
-
|
| 48 |
-
# For auxiliary loss (cached to avoid recomputation)
|
| 49 |
-
self.last_router_logits = None
|
| 50 |
-
self.last_router_probs = None
|
| 51 |
-
|
| 52 |
-
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 53 |
-
batch_size, seq_len, dim = hidden_states.shape
|
| 54 |
-
|
| 55 |
-
# Shared expert output (all tokens)
|
| 56 |
-
shared_out = self.shared_expert(hidden_states)
|
| 57 |
-
|
| 58 |
-
# Routing
|
| 59 |
-
flat_hidden = hidden_states.view(-1, dim)
|
| 60 |
-
router_logits = self.router(flat_hidden)
|
| 61 |
-
router_probs = F.softmax(router_logits.float(), dim=-1)
|
| 62 |
-
|
| 63 |
-
# Cache for aux loss
|
| 64 |
-
self.last_router_logits = router_logits
|
| 65 |
-
self.last_router_probs = router_probs
|
| 66 |
-
|
| 67 |
-
# Top-k selection and renormalization
|
| 68 |
-
top_k_weights, top_k_indices = torch.topk(router_probs, self.top_k, dim=-1)
|
| 69 |
-
top_k_weights = top_k_weights / top_k_weights.sum(dim=-1, keepdim=True)
|
| 70 |
-
top_k_weights = top_k_weights.to(hidden_states.dtype)
|
| 71 |
-
|
| 72 |
-
# Routed expert output via token dispatch
|
| 73 |
-
routed_out = self._dispatch_experts(flat_hidden, top_k_indices, top_k_weights)
|
| 74 |
-
routed_out = routed_out.view(batch_size, seq_len, -1)
|
| 75 |
-
|
| 76 |
-
# Combine: shared expert baseline + routed experts (grow in via zero-init down_proj)
|
| 77 |
-
return shared_out + routed_out
|
| 78 |
-
|
| 79 |
-
def _dispatch_experts(
|
| 80 |
-
self,
|
| 81 |
-
hidden_states: torch.Tensor,
|
| 82 |
-
top_k_indices: torch.Tensor,
|
| 83 |
-
top_k_weights: torch.Tensor,
|
| 84 |
-
) -> torch.Tensor:
|
| 85 |
-
"""Token dispatch - gather tokens per expert, process, scatter back."""
|
| 86 |
-
num_tokens = hidden_states.shape[0]
|
| 87 |
-
output = torch.zeros(
|
| 88 |
-
num_tokens, self.output_dim, device=hidden_states.device, dtype=hidden_states.dtype
|
| 89 |
-
)
|
| 90 |
-
|
| 91 |
-
for expert_idx, expert in enumerate(self.experts):
|
| 92 |
-
expert_mask = top_k_indices == expert_idx
|
| 93 |
-
if not expert_mask.any():
|
| 94 |
-
continue
|
| 95 |
-
|
| 96 |
-
token_indices, slot_indices = torch.where(expert_mask)
|
| 97 |
-
expert_input = hidden_states[token_indices]
|
| 98 |
-
expert_output = expert(expert_input)
|
| 99 |
-
weights = top_k_weights[token_indices, slot_indices].unsqueeze(-1)
|
| 100 |
-
output.index_add_(0, token_indices, expert_output * weights)
|
| 101 |
-
|
| 102 |
-
return output
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
def load_balancing_loss(router_probs: torch.Tensor, num_experts: int, top_k: int) -> torch.Tensor:
|
| 106 |
-
"""Auxiliary loss to encourage balanced expert usage."""
|
| 107 |
-
_, selected = torch.topk(router_probs, top_k, dim=-1)
|
| 108 |
-
expert_mask = F.one_hot(selected, num_experts).float()
|
| 109 |
-
tokens_per_expert = expert_mask.mean(dim=(0, 1))
|
| 110 |
-
prob_per_expert = router_probs.mean(dim=0)
|
| 111 |
-
return (tokens_per_expert * prob_per_expert).sum() * num_experts
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
def z_loss(router_logits: torch.Tensor) -> torch.Tensor:
|
| 115 |
-
"""Z-loss to prevent router logits from growing too large."""
|
| 116 |
-
return torch.logsumexp(router_logits.float(), dim=-1).square().mean()
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
class SharedMoEAudioProjector(nn.Module):
|
| 120 |
-
def __init__(self, config):
|
| 121 |
-
super().__init__()
|
| 122 |
-
|
| 123 |
-
# Temporal downsampling
|
| 124 |
-
self.k = getattr(config, "projector_pool_stride", 4)
|
| 125 |
-
|
| 126 |
-
# Dimensions
|
| 127 |
-
encoder_dim = config.encoder_dim
|
| 128 |
-
in_dim = encoder_dim * self.k
|
| 129 |
-
out_dim = config.llm_dim
|
| 130 |
-
hidden_dim = getattr(config, "projector_hidden_dim", None) or in_dim
|
| 131 |
-
|
| 132 |
-
# MoE config
|
| 133 |
-
self.num_experts = getattr(config, "num_experts", 4)
|
| 134 |
-
self.top_k = getattr(config, "num_experts_per_tok", 2)
|
| 135 |
-
self.aux_loss_coef = getattr(config, "router_aux_loss_coef", 0.02)
|
| 136 |
-
self.z_loss_coef = getattr(config, "router_z_loss_coef", 0.001)
|
| 137 |
-
|
| 138 |
-
# Layers
|
| 139 |
-
self.moe = SharedMoEBlock(in_dim, hidden_dim, out_dim, self.num_experts, self.top_k)
|
| 140 |
-
|
| 141 |
-
# Init
|
| 142 |
-
self._init_weights(in_dim)
|
| 143 |
-
|
| 144 |
-
def _init_weights(self, in_dim: int):
|
| 145 |
-
with torch.no_grad():
|
| 146 |
-
# Shared expert - orthogonal init for stable condition numbers
|
| 147 |
-
nn.init.orthogonal_(self.moe.shared_expert.gate_proj.weight)
|
| 148 |
-
nn.init.orthogonal_(self.moe.shared_expert.up_proj.weight)
|
| 149 |
-
nn.init.orthogonal_(self.moe.shared_expert.down_proj.weight, gain=0.5)
|
| 150 |
-
|
| 151 |
-
# Routed experts - orthogonal for gate/up, tiny orthogonal for down (grow-in)
|
| 152 |
-
# gain=0.01 gives ~1% initial contribution while maintaining good conditioning
|
| 153 |
-
for expert in self.moe.experts:
|
| 154 |
-
nn.init.orthogonal_(expert.gate_proj.weight)
|
| 155 |
-
nn.init.orthogonal_(expert.up_proj.weight)
|
| 156 |
-
nn.init.orthogonal_(expert.down_proj.weight, gain=0.01)
|
| 157 |
-
|
| 158 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 159 |
-
batch_size, seq_len, dim = x.size()
|
| 160 |
-
|
| 161 |
-
target_dtype = self.moe.shared_expert.gate_proj.weight.dtype
|
| 162 |
-
if x.dtype != target_dtype:
|
| 163 |
-
x = x.to(target_dtype)
|
| 164 |
-
|
| 165 |
-
# Pad for pooling (at most k-1 frames -> 1 extra token, negligible impact)
|
| 166 |
-
if seq_len % self.k:
|
| 167 |
-
x = F.pad(x, (0, 0, 0, self.k - seq_len % self.k))
|
| 168 |
-
|
| 169 |
-
# Temporal pooling
|
| 170 |
-
x = x.view(batch_size, -1, dim * self.k)
|
| 171 |
-
|
| 172 |
-
return self.moe(x)
|
| 173 |
-
|
| 174 |
-
def get_aux_loss(self) -> torch.Tensor:
|
| 175 |
-
"""Get auxiliary losses (call after forward)."""
|
| 176 |
-
if self.moe.last_router_logits is None:
|
| 177 |
-
return torch.tensor(0.0, device=self.moe.router.weight.device)
|
| 178 |
-
|
| 179 |
-
balance = load_balancing_loss(self.moe.last_router_probs, self.num_experts, self.top_k)
|
| 180 |
-
z = z_loss(self.moe.last_router_logits)
|
| 181 |
-
|
| 182 |
-
return self.aux_loss_coef * balance + self.z_loss_coef * z
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|