Updated to latest sonicmoe
Updated vendored quack to 0.6.1
Fixed backward pass for EP sentinel
Breaking changes: requires "nvidia-cutlass-dsl==4.6.0" "apache-tvm-ffi>=0.1.10,<0.2"

SonicMoE

Accelerating Mixture-of-Experts with IO and Tile-aware Optimizations.

SonicMoE is a blazing-fast MoE implementation optimized for NVIDIA Hopper and Blackwell GPUs. It leverages CuTe-DSL and Triton to deliver state-of-the-art performance through IO-aware optimizations.

Paper: arXiv:2512.14080
Source: Dao-AILab/sonic-moe
Vendored dependency: Dao-AILab/quack (v0.6.1)

Requirements

NVIDIA Hopper GPUs (H100, H200) or Blackwell GPUs (GB200, B200, B300)
PyTorch >= 2.7
CUDA 12.9+ (13.0+ for B300)
Python 3.12+

Usage

import torch
from kernels import get_kernel

sonicmoe = get_kernel("axolotl-ai-co/sonic-moe")

from sonicmoe import MoE, KernelBackendMoE
from sonicmoe.enums import ActivationType

moe = MoE(
    num_experts=128,
    num_experts_per_tok=8,
    hidden_size=4096,
    intermediate_size=1536,
    activation_function=ActivationType.SWIGLU,
    add_bias=False,
    std=0.02,
).to(device="cuda", dtype=torch.bfloat16)

x = torch.randn(32768, 4096, device="cuda", dtype=torch.bfloat16)
output, aux_loss = moe(x, kernel_backend_moe=KernelBackendMoE.sonicmoe)

Router variants

Since sonic-moe#39 the functional API exposes both softmax(topk(logits)) (TC default) and topk(softmax(logits)) (Qwen3 style) routing, plus optional top-k probability renormalization:

from sonicmoe import moe_TC_softmax_topk_layer

# Qwen3-style: topk(softmax()) with renormalized probabilities
out, router_logits, expert_freq = moe_TC_softmax_topk_layer(
    x, moe.router.weight,
    moe.c_fc.weight.permute(1, 2, 0), moe.c_fc.bias,
    moe.c_proj.weight.permute(1, 2, 0), moe.c_proj.bias,
    K=8,
    stream_id=torch.cuda.current_stream().cuda_stream,
    is_softmax_over_topk=False,
    norm_topk_probs=True,
)

Weight layout

By default, w1 (gated up-projection) is expected in interleaved layout [gate_0, up_0, gate_1, up_1, ...]. HuggingFace checkpoints (Qwen3, Mixtral, DeepSeek, ...) store gate_up_proj in concatenated layout [gate_0, ..., gate_{I-1}, up_0, ..., up_{I-1}]. Pass concat_layout=True to moe_TC_softmax_topk_layer / moe_general_routing_inputs to consume the concatenated layout directly without a pre-pass permutation.

Vendored Dependencies

This kernel vendors QuACK v0.6.1 for CuTe-DSL grouped GEMM infrastructure (Hopper + Blackwell). The vendored copy is under torch-ext/sonic_moe/quack/. Torch operators are registered through add_op_namespace_prefix, which the build system prefixes with a kernel-unique namespace, so they cannot collide with a user-installed quack-kernels.