File size: 10,379 Bytes
82f6f0e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 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 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 | # SPDX-License-Identifier: Apache-2.0
# MegaBlocks CPU Fused MoE Implementation
#
# This is a pure Python/PyTorch implementation for CPU.
# For better performance, consider using the C++ kernel implementation.
#
import torch
import torch.nn.functional as F
def swigluoai_activation(gate: torch.Tensor, up: torch.Tensor,
alpha: float = 1.702, limit: float = 7.0) -> torch.Tensor:
"""
SwigluOAI activation function used in GptOss models.
Formula:
gate = clamp(gate, max=limit)
up = clamp(up, -limit, limit)
glu = gate * sigmoid(gate * alpha)
output = (up + 1) * glu
Args:
gate: Gate tensor from gate projection
up: Up tensor from up projection
alpha: Scaling factor for sigmoid (default: 1.702)
limit: Clamp limit (default: 7.0)
Returns:
Activated tensor
"""
gate = gate.clamp(max=limit)
up = up.clamp(min=-limit, max=limit)
glu = gate * torch.sigmoid(gate * alpha)
return (up + 1) * glu
def silu_and_mul_activation(gate: torch.Tensor, up: torch.Tensor) -> torch.Tensor:
"""
SiLU (Swish) activation with element-wise multiplication.
Formula:
output = silu(gate) * up
Args:
gate: Gate tensor
up: Up tensor
Returns:
Activated tensor
"""
return F.silu(gate) * up
def route_tokens_cpu(
x: torch.Tensor,
router_weight: torch.Tensor,
router_bias: torch.Tensor | None,
moe_top_k: int,
moe_num_experts: int,
moe_normalize_expert_weights: int | None = None,
) -> tuple:
"""
Route tokens to experts and compute expert weights and indices (CPU version).
Args:
x: Input tensor [batch, seq, hidden] or [tokens, hidden]
router_weight: Router weight [num_experts, hidden]
router_bias: Router bias [num_experts] or None
moe_top_k: Number of experts per token
moe_num_experts: Total number of experts
moe_normalize_expert_weights: Normalization order or None
Returns:
Tuple of (logits, expert_weights, expert_indices)
"""
x_flat = x.view(-1, x.shape[-1])
logits = F.linear(x_flat, router_weight, router_bias)
if moe_top_k == 1:
expert_weights, expert_indices = logits.max(dim=-1, keepdim=True)
else:
expert_weights, expert_indices = torch.topk(logits, moe_top_k, dim=-1)
expert_weights = expert_weights.softmax(dim=-1)
if moe_normalize_expert_weights is not None:
expert_weights = expert_weights / torch.norm(
expert_weights,
p=moe_normalize_expert_weights,
dim=-1,
keepdim=True,
)
return logits, expert_weights, expert_indices
def cpu_fused_moe(
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
w1_bias: torch.Tensor | None = None,
w2_bias: torch.Tensor | None = None,
activation: str = "silu",
alpha: float = 1.702,
limit: float = 7.0,
is_interleaved: bool = True,
) -> torch.Tensor:
"""
CPU Fused MoE using PyTorch operations.
This implementation processes all experts in parallel using batched operations
instead of sequential for loops, which is more efficient on CPU.
Args:
hidden_states: [num_tokens, hidden_size]
w1: [num_experts, hidden_size, 2*inter_size] - gate_up_proj weights
w2: [num_experts, inter_size, hidden_size] - down_proj weights
topk_weights: [num_tokens, topk] - routing weights
topk_ids: [num_tokens, topk] - expert indices
w1_bias: [num_experts, 2*inter_size] or None
w2_bias: [num_experts, hidden_size] or None
activation: "silu" or "swigluoai"
alpha: swigluoai alpha parameter
limit: swigluoai limit parameter
is_interleaved: whether gate_up is interleaved [g0,u0,g1,u1,...] (True for GptOss)
Returns:
output: [num_tokens, hidden_size]
"""
num_tokens, hidden_size = hidden_states.shape
num_experts = w1.shape[0]
inter_size = w2.shape[1]
topk = topk_weights.shape[1]
# Initialize output
output = torch.zeros_like(hidden_states)
# Build expert mask: which tokens go to which expert
# expert_mask[expert_id] contains indices of (token_idx, topk_pos) pairs
for expert_idx in range(num_experts):
# Find tokens assigned to this expert
# mask shape: [num_tokens, topk], True where topk_ids == expert_idx
mask = (topk_ids == expert_idx)
if not mask.any():
continue
# Get token indices and topk positions
token_indices, topk_positions = torch.where(mask)
if len(token_indices) == 0:
continue
# Gather input tokens for this expert
# current_hidden: [num_selected_tokens, hidden_size]
current_hidden = hidden_states[token_indices]
# Get weights for this expert
# w1[expert_idx]: [hidden_size, 2*inter_size]
# w2[expert_idx]: [inter_size, hidden_size]
expert_w1 = w1[expert_idx] # [hidden_size, 2*inter_size]
expert_w2 = w2[expert_idx] # [inter_size, hidden_size]
# First projection: hidden @ w1 -> [num_selected, 2*inter_size]
gate_up = current_hidden @ expert_w1
# Add bias if present
if w1_bias is not None:
gate_up = gate_up + w1_bias[expert_idx]
# Split gate and up projections
if is_interleaved:
# GptOss uses interleaved layout: [g0, u0, g1, u1, ...]
gate = gate_up[..., ::2] # [num_selected, inter_size]
up = gate_up[..., 1::2] # [num_selected, inter_size]
else:
# Standard layout: [gate_all, up_all]
gate = gate_up[..., :inter_size]
up = gate_up[..., inter_size:]
# Apply activation
if activation == "swigluoai":
activated = swigluoai_activation(gate, up, alpha, limit)
else: # silu
activated = silu_and_mul_activation(gate, up)
# Second projection: activated @ w2 -> [num_selected, hidden_size]
expert_out = activated @ expert_w2
# Add bias if present
if w2_bias is not None:
expert_out = expert_out + w2_bias[expert_idx]
# Apply routing weights and accumulate
# weights shape: [num_selected]
weights = topk_weights[token_indices, topk_positions].unsqueeze(-1)
weighted_out = expert_out * weights
# Accumulate to output
output.index_add_(0, token_indices, weighted_out.to(output.dtype))
return output
class MegaBlocksMoeMLP(torch.nn.Module):
"""
CPU MoE MLP module that can be used as a drop-in replacement for
the transformers GptOssMLP when using @use_kernel_forward_from_hub.
"""
can_torch_compile: bool = True
def forward(self, x: torch.Tensor) -> tuple:
"""
Forward pass through the MoE layer.
Args:
x: Input tensor of shape [batch_size, seq_len, hidden_size] or [tokens, hidden_size]
Returns:
Tuple of (output, expert_weights) where:
- output: Tensor of same shape as input
- expert_weights: Expert weights for each token [tokens, top_k]
"""
# Get MoE parameters from the wrapped modules
moe_top_k = getattr(self.router, "top_k", 4)
moe_num_experts = getattr(self.experts, "num_experts", 128)
moe_normalize_expert_weights = getattr(
self.experts, "normalize_expert_weights", None
)
# Detect activation type
if hasattr(self.experts, "alpha") and hasattr(self.experts, "limit"):
activation = "swigluoai"
alpha = self.experts.alpha
limit = self.experts.limit
else:
activation = getattr(self.experts, "activation", "silu")
alpha = 1.702
limit = 7.0
# Get weight tensors
if hasattr(self.experts, "gate_up_proj"):
w1 = self.experts.gate_up_proj
is_interleaved = True # GptOss uses interleaved layout
elif hasattr(self.experts, "w1"):
w1 = self.experts.w1
w3 = getattr(self.experts, "w3", None)
if w3 is not None:
w1 = torch.cat([w1, w3], dim=-1)
is_interleaved = False
else:
raise AttributeError("experts module must have 'gate_up_proj' or 'w1' attribute")
if hasattr(self.experts, "down_proj"):
w2 = self.experts.down_proj
elif hasattr(self.experts, "w2"):
w2 = self.experts.w2
else:
raise AttributeError("experts module must have 'down_proj' or 'w2' attribute")
# Get optional bias tensors
w1_bias = getattr(self.experts, "gate_up_proj_bias", None)
w2_bias = getattr(self.experts, "down_proj_bias", None)
# Store original shape
in_shape = x.size()
# Route tokens to experts
logits, expert_weights, expert_indices = route_tokens_cpu(
x,
self.router.weight,
getattr(self.router, "bias", None),
moe_top_k,
moe_num_experts,
moe_normalize_expert_weights,
)
# Reshape input for fused MoE
x_flat = x.view(-1, x.shape[-1])
# Call CPU fused MoE
output = cpu_fused_moe(
hidden_states=x_flat,
w1=w1,
w2=w2,
topk_weights=expert_weights,
topk_ids=expert_indices,
w1_bias=w1_bias,
w2_bias=w2_bias,
activation=activation,
alpha=alpha,
limit=limit,
is_interleaved=is_interleaved,
)
# Restore original shape
output = output.view(in_shape)
return output, expert_weights
# Export classes and functions
__all__ = [
"MegaBlocksMoeMLP",
"cpu_fused_moe",
"route_tokens_cpu",
"swigluoai_activation",
"silu_and_mul_activation",
]
|