xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 07:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[fix][cpu] Use a SwigluOAI impl which supports interleaved gate-up wei (#29273)

Browse Source 
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
This commit is contained in:
Fadi Arafeh 2025-11-25 07:11:11 +00:00 committed by GitHub
parent 64deead719
commit 98caeadd54
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 8 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
vllm/model_executor/layers/fused_moe/cpu_fused_moe.py
View File

@ -6,22 +6,7 @@ import torch
from torch.nn import functional as F from torch.nn import functional as F
from vllm import _custom_ops as ops from vllm import _custom_ops as ops
from vllm.model_executor.layers.activation import SiluAndMul, SwigluOAIAndMul
def silu_and_mul(x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2
return F.silu(x[..., :d]) * x[..., d:]
def swigluoai_and_mul(
x: torch.Tensor, alpha: float = 1.702, limit: float = 7.0
) -> torch.Tensor:
d = x.shape[-1] // 2
gate, up = x[..., :d], x[..., d:]
gate = gate.clamp(max=limit)
up = up.clamp(min=-limit, max=limit)
glu = gate * torch.sigmoid(alpha * gate)
return (up + 1) * glu
def grouped_topk( def grouped_topk(
@ -227,6 +212,11 @@ class CPUFusedMOE:
layer.w13_weight = torch.nn.Parameter(torch.empty(0), requires_grad=False) layer.w13_weight = torch.nn.Parameter(torch.empty(0), requires_grad=False)
layer.w2_weight = torch.nn.Parameter(torch.empty(0), requires_grad=False) layer.w2_weight = torch.nn.Parameter(torch.empty(0), requires_grad=False)
self.act_to_impl = {
"silu": SiluAndMul(),
"swigluoai": SwigluOAIAndMul(),
}
def __call__( def __call__(
self, self,
layer: torch.nn.Module, layer: torch.nn.Module,
@ -246,7 +236,7 @@ class CPUFusedMOE:
apply_router_weight_on_input: bool = False, apply_router_weight_on_input: bool = False,
activation: str = "silu", activation: str = "silu",
) -> torch.Tensor: ) -> torch.Tensor:
assert activation in {"silu", "swigluoai"}, f"{activation} is not supported." assert activation in self.act_to_impl, f"{activation} is not supported."
assert not apply_router_weight_on_input assert not apply_router_weight_on_input
topk_weights, topk_ids = select_experts( topk_weights, topk_ids = select_experts(
hidden_states=x, hidden_states=x,
@ -283,10 +273,7 @@ class CPUFusedMOE:
tokens_for_this_expert = sorted_tokens[start_idx:end_idx] tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
gate_up = layer.gate_up_linear[i](tokens_for_this_expert) gate_up = layer.gate_up_linear[i](tokens_for_this_expert)
if activation == "swigluoai": gate_up = self.act_to_impl[activation].forward_native(gate_up)
gate_up = swigluoai_and_mul(gate_up)
else:
gate_up = silu_and_mul(gate_up)
expert_out = layer.down_linear[i](gate_up) expert_out = layer.down_linear[i](gate_up)
outputs.append(expert_out) outputs.append(expert_out)
start_idx = end_idx start_idx = end_idx