47 lines
1.6 KiB
Python
47 lines
1.6 KiB
Python
"""
|
|
Torch-native implementation for FusedMoE. This is used for torch.compile.
|
|
It is based on https://github.com/pytorch-labs/gpt-fast/blob/32971d3129541c5bfb4f715abc33d1c5f408d204/mixtral-moe/model.py#L204
|
|
"""
|
|
|
|
from typing import Callable, Optional
|
|
|
|
import torch
|
|
from torch.nn import functional as F
|
|
|
|
from sglang.srt.layers.moe.topk import select_experts
|
|
|
|
|
|
def fused_moe_forward_native(
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
use_grouped_topk: bool,
|
|
top_k: int,
|
|
router_logits: torch.Tensor,
|
|
renormalize: bool,
|
|
topk_group: Optional[int] = None,
|
|
num_expert_group: Optional[int] = None,
|
|
custom_routing_function: Optional[Callable] = None,
|
|
correction_bias: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
topk_weights, topk_ids = select_experts(
|
|
hidden_states=x,
|
|
router_logits=router_logits,
|
|
use_grouped_topk=use_grouped_topk,
|
|
top_k=top_k,
|
|
renormalize=renormalize,
|
|
topk_group=topk_group,
|
|
num_expert_group=num_expert_group,
|
|
custom_routing_function=custom_routing_function,
|
|
correction_bias=correction_bias,
|
|
torch_native=True,
|
|
)
|
|
|
|
w13_weights = layer.w13_weight[topk_ids]
|
|
w1_weights, w3_weights = torch.chunk(w13_weights, 2, dim=2)
|
|
w2_weights = layer.w2_weight[topk_ids]
|
|
x1 = torch.einsum("ti,taoi -> tao", x, w1_weights)
|
|
x1 = F.silu(x1)
|
|
x3 = torch.einsum("ti, taoi -> tao", x, w3_weights)
|
|
expert_outs = torch.einsum("tao, taio -> tai", (x1 * x3), w2_weights)
|
|
return torch.einsum("tai,ta -> ti", expert_outs, topk_weights.to(expert_outs.dtype))
|